CN119669880A - An intelligent early warning method for equipment status based on danger perception - Google Patents

Abstract

The application provides an intelligent equipment state early warning method based on danger perception, which relates to the technical field of intelligent equipment state early warning and comprises the steps of acquiring equipment operation data, wherein the equipment operation data comprises multi-source heterogeneous data; the method comprises the steps of carrying out feature extraction on multi-source heterogeneous data to obtain unified feature representation, constructing a correlation model between equipment operation states and feature parameters according to the unified feature representation, and obtaining equipment operation data to be analyzed. According to the method, an evaluation index system is automatically generated, comprehensive evaluation is performed by adopting a evidence reasoning network, and an online increment learning and self-adaptive threshold adjustment mechanism is introduced, so that real-time and accurate evaluation of the health state of the equipment is realized.

Description

Intelligent equipment state early warning method based on danger sensing

Technical Field

The invention relates to the field of intelligent early warning of equipment states, in particular to an intelligent early warning method of equipment states based on danger sensing.

Background

In the process of equipment health state evaluation and early warning, technical contradiction of data multidimensional isomerism and domain knowledge complexity is faced. On one hand, the data generated in the running process of the equipment has the characteristics of multi-source isomerism, high dimensionality, nonlinearity and the like, the relevance of different types of data in time and space is complicated, and the characteristic extraction and fusion of the data face a great challenge. On the other hand, the evaluation and early warning of the health state of the equipment are required to integrate the experience knowledge of the field expert, a reasonable evaluation index system is constructed, and qualitative field knowledge and quantitative data analysis results are organically combined to form an interpretable and trusted early warning result. In addition, in practical application, the working condition of the equipment is changeable, the external environment factors are complex and various, and how to adjust the data association analysis model and the early warning model in real time in the dynamically changed environment ensures the accuracy and the instantaneity of the early warning result is also a technical problem to be solved urgently. Meanwhile, the health state evaluation and early warning of the equipment need to consider the full life cycle of the equipment, and how to mine valuable information in massive historical data and construct comprehensive and accurate health state portraits is also a challenging problem.

Disclosure of Invention

The invention provides an intelligent equipment state early warning method based on danger perception, which mainly comprises the following steps:

The method comprises the steps of obtaining equipment operation data, wherein the equipment operation data comprise multi-source heterogeneous data, extracting characteristics of the multi-source heterogeneous data to obtain unified characteristic representation, constructing a correlation model between equipment operation states and characteristic parameters according to the unified characteristic representation, obtaining equipment operation data to be analyzed, analyzing the equipment operation data to be analyzed according to the correlation model to obtain equipment operation state analysis results, and evaluating the current health states of equipment according to the equipment operation state analysis results and a preset equipment health state evaluation index system to obtain equipment health state evaluation results.

Further, the multi-source heterogeneous data is subjected to feature extraction to obtain unified feature representation, the multi-source heterogeneous data is obtained, the multi-source heterogeneous data is processed by adopting a data cleaning and preprocessing method, the processed multi-source heterogeneous data is subjected to dimension reduction by adopting a principal component analysis method to extract key feature parameters, the extracted key feature parameters are input into a self-encoder network, and the unified feature representation of the multi-source heterogeneous data is obtained through unsupervised learning.

Further, the construction of the association model between the equipment operation state and the characteristic parameters according to the unified characteristic representation comprises the steps of obtaining historical equipment operation data, extracting the characteristic parameters of the historical equipment operation data according to the unified characteristic representation, obtaining equipment operation state labels corresponding to the historical equipment operation data, taking the characteristic parameters as input, taking the equipment operation state labels as output, training a support vector machine model, and obtaining the association model between the equipment operation state and the characteristic parameters.

Further, according to the equipment operation state analysis result, the current health state of the equipment is evaluated by combining a preset equipment health state evaluation index system to obtain an equipment health state evaluation result, wherein the index system comprises a plurality of evaluation dimensions and corresponding index items, the index items related to the current equipment operation condition are extracted from the index system by combining the equipment operation state analysis result to serve as evaluation indexes, the health state of each evaluation index is judged according to a threshold value interval and an actual value of each evaluation index, the health states of all the evaluation indexes are synthesized, and the current health state evaluation result of the equipment is obtained by adopting a weighted average method.

Further, the method further comprises the steps of obtaining real-time data in the running process of the equipment, dynamically updating the association model by adopting an online increment learning method to adapt to the change of the working condition of the equipment, predicting the running state of the equipment according to the updated association model, and triggering early warning and generating warning information when a prediction result exceeds a preset threshold value.

Furthermore, before the association model between the running state of the equipment and the characteristic parameters is constructed, the method further comprises the steps of adopting a frequent pattern mining algorithm to find association rules and time sequence patterns in the multi-source heterogeneous data aiming at the unified characteristic representation, and using the association rules and time sequence patterns obtained through mining as priori knowledge to guide the construction of the association model and improve the expression capacity and generalization performance of the model.

Further, after the health status of each evaluation index is synthesized and the current health status evaluation result of the equipment is obtained by adopting a weighted average method, the method further comprises the steps of acquiring equipment operation data, continuously collecting real-time data from various sensors, performing data cleaning and preprocessing to remove abnormal values and noise, comparing the equipment health status evaluation result with a preset health level threshold value to judge the current health level of the equipment, and automatically generating an equipment overhaul priority plan and a maintenance strategy according to the health level and the importance degree of the equipment to guide the development of equipment operation and maintenance work.

Further, a health state evaluation index system is obtained, wherein the health state evaluation index system comprises a plurality of indexes serving as evidence nodes, a hierarchical analysis method is adopted for the index system, the weight of each evidence node is determined, the weight is used as the credibility of the evidence node, pearson correlation coefficients among different evidence nodes are calculated, the correlation coefficients are used as the correlation strength among the evidence nodes, an evidence reasoning network is constructed according to the credibility of the evidence nodes and the correlation strength, and a Bayesian network reasoning algorithm is adopted in the evidence reasoning network to comprehensively analyze the support degree of each evidence node to different health states.

Further, the method comprises the steps of preprocessing equipment operation data, removing abnormal values and noise in the equipment operation data, complementing missing values, setting a sliding time window according to the type and the working characteristics of the equipment, slicing and aggregating the preprocessed equipment operation data in the sliding time window, extracting time domain features and frequency domain features of the aggregated data, wherein the time domain features comprise mean values, variances and peak-to-peak values, the frequency domain features comprise frequency spectrum entropy, carrying out importance analysis on the extracted features by utilizing pearson correlation coefficients and maximum information coefficients, screening out feature subsets with maximum correlation with equipment fault states, adopting a support vector machine, establishing a classification model between the feature subsets and the equipment fault modes, obtaining weight coefficients corresponding to the features in the classification model, determining the features with the weight coefficients larger than preset thresholds as key factors causing equipment faults, calculating root mean square values and kurtosis of vibration signals collected currently on line, if the root mean square values and kurtosis are preset, utilizing pearson correlation coefficients and maximum information coefficients, carrying out importance analysis on the extracted features, screening out a feature subset with the maximum correlation with the equipment fault states, adopting a support vector machine, establishing a classification model between the feature subset and the equipment fault modes, obtaining the weight coefficients corresponding to the characteristics in the classification model, determining the key factors with the weight coefficients to be larger than the preset threshold value, carrying out on-to be the key factors, carrying out on-line monitoring on the equipment operation state, calculating the vibration signals, calculating the root mean square values and kurtosis of the current collected, if the equipment operation state, and the threshold value exceeds the threshold value, and the threshold value is compared with the threshold value, and the threshold value.

The technical scheme provided by the embodiment of the invention can have the following beneficial effects:

The invention discloses an intelligent equipment state early warning method based on danger sensing. Aiming at multi-source isomerism and high-dimensional nonlinear characteristics of equipment operation data, a unified representation model is constructed by adopting a multi-view data fusion technology, and standardized representation of isomerism data is realized. Based on the above, the time sequence and the space relation among the parameters are analyzed by using a space-time correlation mining algorithm, and an implicit mode is found. And on the basis, extracting high-level features by a deep learning method, and constructing a domain knowledge base by combining a knowledge graph technology. According to the invention, an evaluation index system is automatically generated, the comprehensive evaluation is performed by adopting a evidence reasoning network, and an online increment learning and self-adaptive threshold adjustment mechanism is introduced, so that the real-time and accurate evaluation of the health state of the equipment is realized. The method effectively integrates multi-source heterogeneous data, expert knowledge and deep learning technology, improves the comprehensiveness, accuracy and instantaneity of equipment health state assessment, and provides powerful support for equipment predictive maintenance.

Drawings

Fig. 1 is a flow chart of an intelligent early warning method for equipment state based on danger perception.

Detailed Description

The technical solutions of the present invention will be clearly and completely described in connection with the embodiments, and it is obvious that the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

As shown in fig. 1, an intelligent device state early warning method based on risk awareness in this embodiment may specifically include:

Step S101, constructing a heterogeneous data unified representation model by adopting a multi-view data fusion technology according to multi-source heterogeneous characteristics and high-dimension nonlinear characteristics of equipment operation data, and mapping different sources and different types of data to the same feature space to realize standardized representation and fusion processing of the data.

According to the multi-source heterogeneous characteristics of the equipment operation data, a data preprocessing technology is adopted to clean and convert data of different sources and different types, noise and redundancy in the data are eliminated, and the data quality is improved. Aiming at the high-dimensional nonlinear characteristics of the data, a Principal Component Analysis (PCA) algorithm is adopted to extract the low-dimensional characteristics with the most representation and discrimination from the high-dimensional data, so that the data dimension is reduced, and the calculation complexity is reduced. And constructing a heterogeneous data unified representation model, and mapping data of different sources and different types to the same feature space by adopting a multi-view learning technology to realize standardized representation of the data. Specifically, the data for each view may be feature extracted and compressed using a self-encoder (AutoEncoder), and then feature vectors for different views may be stitched together to form a unified data representation. Based on uniform representation of heterogeneous data, a weighted average data fusion algorithm is adopted to fuse the data representations of multiple view angles, so that a comprehensive data representation is obtained, and the integrity and consistency of the data are improved. Wherein, the weight of each view angle can be determined according to the influence degree of the view angle on the running state of the equipment, and the view angle weight with larger influence is higher. For time series characteristics of equipment operation data, a method of time series analysis, such as sliding window or cyclic neural network (RNN), is adopted to convert the data in the time dimension into characteristic vectors, and the characteristic vectors are input into a machine learning model together with other characteristics. And constructing a classification model of the running state of the equipment by adopting a Support Vector Machine (SVM) algorithm according to the fused data representation and time sequence characteristics, and realizing intelligent early warning and fault diagnosis of the state of the equipment based on dangerous perception. Aiming at the parameter tuning problem in the model training process, an optimal model super-parameter combination is automatically searched in a preset parameter range by adopting a grid search method, so that the generalization performance and the robustness of the model are improved. And evaluating the performance of the model under different parameter combinations by a cross-validation mode, and selecting the parameter with the best performance as the final model parameter. In the process of data acquisition, transmission and storage, security measures such as data encryption and access control are adopted, confidentiality and integrity of data are protected, and data leakage and tampering are prevented. At the same time, compliance of data and protection of user privacy are ensured following relevant data privacy protection specifications and standards, such as GDPR and the like. The trained model is deployed in a production environment, the equipment operation data acquired in real time are predicted and analyzed, and the result is presented to a user through a visual interface, so that the user is assisted in carrying out operation and decision of the equipment, and the operation efficiency and reliability of the equipment are improved.

Illustratively, multisource heterogeneous device operational data, such as from sensors (temperature, pressure, vibration), controllers (switch states, operational parameters), log files (event records, error information), etc., are in different data formats, sampling frequencies, dimensions. In order to effectively use these data, data preprocessing is required. For example, the data unit collected by the temperature sensor is celsius, the data unit collected by the pressure sensor is pascal, and the data unit needs to be converted into a unified unit or subjected to standardized processing, so that the influence of dimension is eliminated. In addition, noise may be present in the data, such as fluctuations in sensor readings, and a filtering algorithm (e.g., kalman filtering) may be required to remove the noise. Meanwhile, redundancy may exist in some data, for example, a plurality of sensors measure the same physical quantity, and feature selection or dimension reduction processing is required to reduce the data redundancy. Therefore, the data quality can be improved, and a foundation is laid for subsequent analysis. The device operational data typically has high dimensional non-linear characteristics. For example, a device may have tens of sensors, with the data collected by each sensor constituting a dimension. There may be complex nonlinear relationships between these data. To reduce computational complexity and avoid dimensional disasters, a Principal Component Analysis (PCA) algorithm may be employed. PCA can project high-dimensional data into a low-dimensional space while preserving the principal variances of the data. For example, reducing the dimension of 50-dimensional sensor data to 5-dimensional while preserving 95% of the data variance reduces both the data dimension and preserves critical information of the data. In order to fuse together different sources, different types of data, a heterogeneous data unified representation model needs to be built. Multiple view learning techniques may be employed to treat data of different sources as different views. For example, the sensor data is one view, the controller data is another view, and the log data is a third view. The data for each view may be feature extracted and compressed with a self-encoder. Assuming that the sensor data is compressed into 10-dimensional feature vectors by the self-encoder, the controller data is compressed into 5-dimensional feature vectors, the log data is compressed into 2-dimensional feature vectors, and the three feature vectors are spliced to form a 17-dimensional unified data representation. By doing so, different types of data can be mapped to the same feature space, so that subsequent fusion and analysis are facilitated. After the unified data representation is obtained, a weighted average data fusion algorithm may be employed. For example, according to expert experience or domain knowledge, it is determined that the degrees of influence of the sensor data, the controller data, and the log data on the running states of the device are 0.6, 0.3, and 0.1, respectively, and then the feature vectors of the three views are multiplied by the corresponding weights and added to obtain a comprehensive 17-dimensional data representation. The weighted fusion mode can embody the importance of different data sources and improve the integrity and consistency of data. The device operational data typically has a time series characteristic. For example, the data collected by the sensors varies over time, forming a time series. A sliding window may be used to convert the time series data into feature vectors. For example, taking the data of 10 time points as a window, calculating the statistical characteristics of the mean, variance, maximum value, minimum value and the like of the data in the window to form a feature vector. Time series features may also be extracted using a Recurrent Neural Network (RNN), such as LSTM, to model the time series data. These time series features are input into the machine learning model along with other features. And constructing a classification model of the running state of the equipment by using the fused data representation and the time sequence characteristics and adopting a Support Vector Machine (SVM) algorithm. for example, according to historical data, the running state of the equipment is divided into three states of normal, abnormal and fault, an SVM classifier is trained, new running data of the equipment is classified, and fault diagnosis is achieved. In the model training process, the model parameters need to be optimized. And automatically searching the optimal model super-parameter combination within a preset parameter range, such as a kernel function type, a penalty coefficient and the like of the SVM by adopting a grid searching method. And evaluating the performance of the model under different parameter combinations by a cross-validation mode, such as 5-fold cross-validation, and selecting the parameter with the best performance as the final model parameter so as to improve the generalization performance and the robustness of the model. In order to ensure data security, security measures such as data encryption and access control are required. For example, the data is stored and transmitted in an encrypted manner, so that the access right to the data is limited, and only authorized users can access the data, and the data leakage and the tampering are prevented. Meanwhile, the compliance of data and the protection of user privacy are ensured by following the related data privacy protection specifications and standards. And finally, deploying the trained model into a production environment, predicting and analyzing the equipment operation data acquired in real time, and presenting the result to a user through a visual interface. For example, the predicted equipment running state, fault type and other information are displayed on a monitoring platform, so that a user is assisted in carrying out operation and decision of equipment, and the running efficiency and reliability of the equipment are improved.

Step S102, analyzing time sequence relation and space dependency relation among operation parameters of equipment by adopting a space-time association mining algorithm based on uniformly represented data, and finding out implicit association modes and evolution rules in the data. Specifically, a sliding time window and spatial correlation analysis are used, and a parameter correlation network is constructed by combining a graph theory algorithm, so that a key space-time mode is identified.

And constructing a multidimensional space-time data model according to the unified representation of the equipment operation data, and mapping the operation parameters of different equipment into a unified space-time coordinate system. And dynamically dividing time slices by adopting a sliding time window technology, and extracting statistical characteristics, such as mean value, variance and the like, of equipment operation parameters in each time slice to form a characteristic sequence in a time dimension. Based on the spatial correlation analysis method such as the pearson correlation coefficient, the spatial correlation coefficient matrix among different equipment operation parameters is calculated, and the correlation strength of the equipment in the spatial dimension is described. And fusing the characteristic sequence in the time dimension with the correlation coefficient matrix in the space dimension to construct a space-time correlation characteristic matrix, and comprehensively representing the space-time evolution rule of the running state of the equipment. And (3) applying frequent pattern mining algorithms such as an Apriori algorithm and the like to find a parameter combination pattern of frequent co-occurrence in the space-time correlation feature matrix and reveal key space-time correlation rules of equipment operation. And combining graph theory algorithms such as a minimum spanning tree algorithm and the like, taking the equipment operation parameters as nodes, taking the space-time correlation strength as an edge weight, constructing an equipment operation parameter correlation network, and intuitively showing the space-time dependency relationship among the parameters. And identifying closely related parameter clusters in a parameter association network through community discovery algorithms such as Louvain algorithm and the like, revealing a key space-time mode of equipment operation, and providing decision support for equipment operation and maintenance optimization.

Illustratively, the multi-dimensional spatiotemporal data model is constructed to map the operating parameters of different devices into a unified spatiotemporal coordinate system to facilitate analysis of spatiotemporal correlations between the devices. It is assumed that a factory workshop is provided with a plurality of numerical control machine tools, and each machine tool is provided with a plurality of sensors for monitoring parameters such as temperature, pressure, vibration and the like. Each machine tool can be considered as a point in space and each point in time as a scale in the time dimension, thus constructing a three-dimensional space-time coordinate system. To extract the feature sequence in the time dimension, a sliding time window technique is employed. For example, with 10 minutes as a time window, slide every 5 minutes. Within each time window, the statistical characteristics of each parameter of each machine tool, such as mean, variance, etc., are calculated. It is assumed that the temperature of machine A was 25℃in the first time window (0-10 minutes) with a variance of 1℃and the temperature of machine A was 26℃in the second time window (5-15 minutes) with a variance of 0.5 ℃. Thus, a characteristic sequence of the A machine tool temperature parameter in the time dimension can be formed. The spatial correlation analysis is used for calculating a spatial correlation coefficient matrix between different equipment operation parameters and describing the correlation strength of the equipment in the spatial dimension. For example, a pearson correlation coefficient of the temperature of the a machine and the pressure of the B machine is calculated. assuming a correlation coefficient of 0.8, a strong positive correlation exists between the temperature of the machine tool A and the pressure of the machine tool B, namely, when the temperature of the machine tool A is increased, the pressure of the machine tool B is also increased cenderung. By calculating the correlation coefficients between all the equipment operating parameters, a spatial correlation coefficient matrix can be obtained. And fusing the characteristic sequence in the time dimension with the correlation coefficient matrix in the space dimension to construct a space-time correlation characteristic matrix. For example, the temperature mean, variance of the A machine tool in each time window and the correlation coefficient of the temperature of the A machine tool and other equipment parameters are integrated together to form a time-space correlation characteristic vector of the A machine tool. The space-time associated feature vectors of all the devices are combined together to form a space-time associated feature matrix. Frequent pattern mining is used to find parameter combination patterns that co-occur frequently in the spatio-temporal correlation feature matrix. For example, using the Apriori algorithm, it can be found that the combined mode of "a machine tool temperature rise, B machine tool pressure rise, C machine tool vibration increase" frequently occurs. This reveals that there is a spatio-temporal association rule between these three parameters. The space-time dependency relationship between parameters can be intuitively displayed by constructing the equipment operation parameter association network. And constructing a parameter association network by taking the equipment operation parameters as nodes and taking the space-time association strength (such as the absolute value of a correlation coefficient) as an edge weight. For example, if the correlation coefficient between the a machine temperature and the B machine pressure is 0.8, the a machine temperature node and the B machine pressure node are connected in the network with an edge weight of 0.8. The community discovery algorithm is used to identify tightly-associated parameter clusters in a parameter association network. For example, applying the Louvain algorithm, it can be found that there is a parameter cluster in the network that contains A machine temperature, B machine pressure, C machine vibration. This suggests that there is a strong spatiotemporal dependency between these three parameters, which may be commonly affected by some underlying factor. Identifying these key spatiotemporal patterns can help engineers better understand the equipment operation rules for equipment operation and maintenance optimization. for example, if an abnormality is found frequently in a certain parameter cluster, key monitoring and maintenance can be performed on the parameter cluster, so that the reliability and the operation efficiency of the equipment are improved.

And step S103, utilizing the found space-time correlation mode as the input of the feature engineering to guide the construction of a multi-level feature extraction network. The high-level characteristic representation in the data is automatically learned by a deep learning method, including capturing local characteristics by using a convolutional neural network, modeling time sequence dependence of a long-term and short-term memory network, effectively reducing data dimension and improving the efficiency and accuracy of characteristic extraction.

And constructing a characteristic engineering framework according to the space-time correlation mode, and taking the space-time correlation mode as the input of the characteristic engineering. By analyzing the space-time correlation mode, key features in the space-time correlation mode, such as trend, periodicity, spatial position distribution and the like of a time sequence, are extracted and used as the input of a subsequent feature extraction network. A multi-level network structure combining a Convolutional Neural Network (CNN) and a long-short-term memory network (LSTM) is adopted to design a feature extraction network. The local features and the spatial correlation of the data are captured by the convolution operation of the CNN, and the features are extracted and abstracted layer by layer through the multi-layer convolution and pooling operation. Meanwhile, an LSTM module is embedded in the network, the time sequence dependency relationship of the data is modeled through a gating mechanism and a memory unit, and the evolution rule of the data in the time dimension is captured. In the feature extraction process, the high-dimensional data is subjected to dimension reduction processing by adopting dimension reduction technologies such as Principal Component Analysis (PCA) and the like, so that redundancy and noise features are reduced. Meanwhile, feature subsets with the most distinguishability and relevance are selected by using feature selection methods such as mutual information, correlation coefficients and the like, so that the efficiency and the accuracy of feature extraction are improved. And performing end-to-end training optimization on the feature extraction network by using a back propagation algorithm. By defining a suitable loss function, such as Mean Square Error (MSE) or cross entropy loss, the difference between the network output and the real labels is calculated and the network parameters are updated using a gradient descent method, enabling them to adaptively learn and extract high-level feature representations in the data. At the output layer of the feature extraction network, a high-level feature representation of the data is obtained. The extracted high-level features are input into a subsequent task model, such as a classifier or a regression model of a Support Vector Machine (SVM), a random forest and the like, and are used for performing tasks such as classification, prediction or anomaly detection. The performance and the precision of the task can be effectively improved through the automatic learning of the feature extraction network and the extracted high-level features. In practical application, the structure and parameters of the feature extraction network are adjusted and optimized according to the requirements of specific tasks. And selecting the optimal network structure and super-parameter setting by the methods of cross verification and the like so as to achieve the optimal characteristic extraction effect. Meanwhile, a training data set is expanded by utilizing a data enhancement technology such as rotation, translation, scaling and the like, so that the generalization capability and the robustness of the feature extraction network are improved.

For example, the spatiotemporal correlation pattern may be used as an input to feature engineering to provide powerful data support for subsequent machine learning tasks. The core goal of the feature engineering framework is to extract valuable features from the raw data in order to better understand and model the data. First, analysis of the spatio-temporal correlation pattern is required. For example, assume that temperature, pressure and vibration data acquired by a plurality of sensors on a certain factory pipeline are obtained, and a frequent occurrence mode of "A equipment temperature rise, B equipment pressure rise and C equipment vibration increase" is found through a space-time correlation mining algorithm. Analyzing this pattern, key features can be extracted for time-series trends in A, B, C device parameters (e.g., temperature rise continuously, pressure fluctuation rise, vibration amplitude increase), periodicity (e.g., whether there is a significant day or week period in the parameter change), and spatial location distribution of A, B, C devices (e.g., whether the three devices are located adjacent on the pipeline). These features enable a more detailed characterization of the device operating state and its mutual influence. Next, a feature extraction network is constructed using the extracted key features as input. The network typically employs a multi-layer structure of a combination of CNN and LSTM. CNNs are good at capturing the local features and spatial correlations of data. For example, the CNN may learn the effect that a plant temperature rise spatially has on B plant pressure. Through the multi-layer rolling and pooling operations, the network is able to extract and abstract features layer by layer, such as "temperature rise rate", "pressure fluctuation frequency", "vibration peak value", etc. from the original temperature, pressure, vibration values. LSTM is then used to model the timing dependencies of the data. For example, the LSTM may learn that the a device temperature increases for a period of time before the B device pressure increases, thereby capturing the time evolution law between parameters. In the feature extraction process, in order to reduce data redundancy and noise, a dimension reduction technique such as PCA may be used. For example, assuming 10 features are extracted, but only 3 principal components are found to be able to interpret 95% of the data variance by PCA analysis, the feature dimension can be reduced from 10 to 3, simplifying training of subsequent models. Meanwhile, by adopting a feature selection method such as mutual information or correlation coefficient, the most relevant features of the target task can be further screened out. For example, if the "rate of temperature rise" is found to be most correlated with the equipment failure, it may be input as a primary feature into a subsequent failure prediction model. The training of the feature extraction network uses a back propagation algorithm. For example, if the objective task is to predict a device failure, a mean square error may be defined as a loss function, measuring the difference between the network predicted value and the actual failure signature. The network parameters are updated continuously through a gradient descent method, so that the network can learn the optimal characteristic representation, and the prediction accuracy is improved. After training is completed, the output layer of the feature extraction network will provide a high-level feature representation of the data. These high-level features may be input into various task models. For example, SVM may be used to classify a device failure, or random forest may be used to predict the remaining life of the device. Compared with directly using the original data, the performance and the accuracy of the task model can be remarkably improved by using the extracted high-level features. In practical applications, the feature extraction network needs to be adjusted and optimized according to specific task requirements. For example, a deeper CNN structure may be employed for image data, and the number of hidden units of the LSTM may be adjusted for time series data. The optimal network structure and super-parameter setting can be determined by cross-validation and the like. In addition, the data enhancement technology, such as random clipping or noise addition to the time sequence, can effectively improve the generalization capability and robustness of the network, so that the network can maintain good performance in the face of new data. Summarizing, based on the feature engineering framework of the space-time correlation mode, by combining deep learning technologies such as CNN, LSTM and the like and methods such as PCA, feature selection and the like, high-level features can be effectively extracted from data, and powerful support is provided for various downstream tasks.

And step S104, combining the extracted high-level features, constructing a knowledge base in the field of equipment health state evaluation by adopting a knowledge graph technology, formalizing and representing expert experience as a knowledge graph, and realizing structural representation and reasoning of field knowledge. The knowledge graph contains equipment components, fault types, performance indexes and other entities and relations thereof.

And acquiring various performance index data in the running process of the equipment, and preprocessing, including data cleaning, outlier processing and data normalization, to obtain a standardized index data set. And matching the index names in the standardized index data set with the performance index entity names in the knowledge graph according to the pre-constructed equipment performance knowledge graph to acquire the entity and attribute information thereof related to the current equipment running state. And calculating real-time values of all the performance indexes according to the acquired performance index entity attribute information. And comparing the real-time value of the performance index with a preset normal range threshold value through knowledge reasoning, and judging whether the normal range is exceeded or not. If the performance index exceeds the normal range, inquiring a fault type entity related to the performance index in the knowledge graph. Knowledge reasoning may use rule-based reasoning methods, such as defining a series of IF-THEN rules, triggering the corresponding rules based on the comparison of the real-time value of the performance index with the threshold value, and deducing the likely fault type. And further acquiring equipment component entities possibly causing the faults by combining the associated information of the fault type entities in the knowledge graph to form a preliminary evaluation result of the current health state of the equipment. And taking entity attribute information extracted from the knowledge graph as high-level features, analyzing the high-level features by utilizing a decision tree algorithm, evaluating the importance of each feature by using indexes such as information gain or a base index, and selecting key features which have important influence on the judgment of the health state of the equipment. Comprehensively considering the preliminary evaluation result and the key feature analysis result, and outputting the current health state evaluation report of the equipment. The report is generated in a text format, and the content comprises basic equipment information, general comments of the current health state, possible fault types, equipment components related to the fault, analysis of causes of the fault and the like, so that decision support is provided for subsequent maintenance and repair.

Illustratively, the health management of the device is as if it were a physical examination of the human body. In order to fully understand the health of the device, various performance index data, such as temperature, pressure, vibration, current, etc., need to be collected during the operation of the device. These data are comparable to various physiological indicators of the human body, such as body temperature, blood pressure, heart rate, etc. After the data is acquired, preprocessing is performed first, just as some preparation work is needed before physical examination. Data cleansing is to remove noise and errors in the data, such as to remove abnormal spike or missing values. Outlier processing is performed by correcting or eliminating data that is outside of normal range, such as a sudden temperature surge at a point in time to an impossible value. Data normalization is the conversion of data of different dimensions into the same scale range, e.g., converting temperature values from degrees celsius to degrees fahrenheit. Through these preprocessing steps, a standardized index data set is obtained, which lays a foundation for subsequent analysis. It is assumed that an industrial robot is being monitored, and index data such as joint temperature, motor current, movement speed and the like are collected. The abnormal peak value caused by the sensor fault is removed through data cleaning, abnormal temperature fluctuation caused by external environment interference is corrected through abnormal value processing, and index data of different dimensions are converted into a range of 0 to 1 through data normalization. Next, these data need to be understood by means of a device performance knowledge graph. The knowledge graph is better than the doctor's expert knowledge base, and contains various performance indexes, fault types and association relations among the performance indexes and the fault types of the equipment. And matching the index names in the standardized index data set with the performance index entity names in the knowledge graph, for example, matching the joint temperature with the robot joint temperature entity in the knowledge graph. Thus, the entity and the attribute information thereof related to the running state of the current equipment, such as the normal range of the robot joint temperature entity, the alarm threshold value and the like, can be acquired. For example, the normal range of the "robot joint temperature" entity in the knowledge graph is 20 ℃ to 40 ℃, and the alarm threshold is 45 ℃. By querying the knowledge graph, the type of failure associated with "robot joint temperature", such as "joint overheating", can also be obtained. According to the acquired performance index entity attribute information, the real-time value of each performance index can be calculated. For example, from the data collected by the sensors, a real-time temperature value for each joint of the robot can be calculated. And then, comparing the real-time value of the performance index with a preset normal range threshold value through knowledge reasoning, and judging whether the real-time value exceeds the normal range. For example, the real-time temperature of the robot No. 1 joint is 46 ℃, exceeding the alarm threshold of 45 ℃. Through knowledge reasoning, the joint can be judged to have overheat risk. If a certain performance index exceeds a normal range, inquiring a fault type entity related to the performance index in the knowledge graph. For example, according to a knowledge graph, a "robot joint temperature" exceeding an alarm threshold may result in a "joint overheating" fault. Knowledge reasoning may use rule-based reasoning methods. For example, a rule is defined that if the robot joint temperature exceeds 45 ℃, it is inferred that a "joint overheating" failure may occur. And combining the associated information of the fault type entity in the knowledge graph, further acquiring the equipment component entity possibly causing the fault. For example, a "joint overheating" fault may be associated with a "temperature sensor," "radiator fan," etc. component. Thus, a preliminary evaluation result of the current health state of the device is formed. The entity attribute information extracted from the knowledge graph is used as high-level characteristics, such as 'robot joint temperature', 'motor current', 'movement speed', and the like. And analyzing the high-level features by utilizing a decision tree algorithm, and evaluating the importance of each feature through indexes such as information gain or a base index. For example, the analysis finds that the "robot joint temperature" has the greatest effect on the equipment health status judgment. Comprehensively considering the preliminary evaluation result and the key feature analysis result, and finally outputting the current health state evaluation report of the equipment. The report content includes device basic information, a general evaluation of the current health state, the type of fault that may exist, device components involved in the fault, analysis of the cause of the fault, and the like. For example, the final generated health status assessment report is a device name of industrial robot XYZ-1, a current health status of warning, a possible fault type of joint No.1 overheating, a device component related to the fault of joint No.1 temperature sensor and joint No.1 cooling fan, and a reason analysis that the joint No.1 temperature continuously rises and exceeds an alarm threshold value, which may be caused by cooling fan faults or excessively high ambient temperature.

And S10.5, automatically generating an equipment health state assessment index system according to the constructed domain knowledge graph and combining the equipment operation working conditions and performance requirements. Specifically, key evaluation indexes are identified through map traversal and relation reasoning, and a multi-dimensional evaluation system is formed by utilizing a hierarchical structure in the map, so that a quantitative evaluation method comprehensively considering multi-dimensional characteristics of equipment is formed.

And constructing a knowledge graph related to the equipment according to the domain knowledge, wherein the knowledge graph contains multidimensional information such as equipment component parts, operation conditions, performance parameters and the like, and a complete equipment knowledge representation is formed. And forming a classification system of the equipment health evaluation index through an ontology hierarchical structure in the knowledge graph, and dividing the evaluation index into different levels such as a component level, a subsystem level, a system level and the like. Traversing the knowledge graph by adopting a breadth-first search algorithm, identifying key nodes highly related to the operation working condition and the performance requirement of the equipment, and taking the key nodes as key evaluation indexes. And extracting attribute nodes directly connected with each key evaluation index from the knowledge graph as influencing factors of the index. And acquiring real-time operation data of the equipment, and mapping the data to corresponding attribute nodes in the knowledge graph through a pre-established mapping table. And setting a health state threshold value for each attribute node, and judging the health state of each index by comparing the real-time data with the threshold value. Based on key evaluation indexes, traversing the knowledge graph through a depth-first search algorithm, taking the evaluation indexes as starting points, mining attribute nodes connected with the intermediate nodes, and taking the attribute nodes as influencing factors of the health state of the equipment to form a more comprehensive evaluation index system. And determining the weight of each index by adopting a hierarchical analysis method according to the ontology hierarchical relationship among all evaluation indexes in the knowledge graph. Firstly, an index pairwise comparison matrix is constructed, and the values of elements in the matrix are determined in an expert scoring mode. And then calculating a feature vector corresponding to the maximum feature value of the matrix, and normalizing the feature vector to be used as index weight. And constructing a weighted average model as a comprehensive evaluation model, multiplying the real-time data of each evaluation index by the corresponding weight, and summing to obtain the quantization score of the current health state of the equipment. And judging whether the equipment is in a health state currently according to a preset health state score threshold value. And when the score is lower than the early warning threshold value, the system automatically sends out early warning prompt to indicate that the equipment has fault risk.

For example, when building a knowledge graph related to a device, it is first necessary to define entities such as component parts, operation conditions, performance parameters, and the like of the device, and to define relationships between them. For example, for an air conditioning system, the entities in the knowledge graph may include components such as compressors, condensers, evaporators, and the like, as well as their operating temperature, pressure, and other performance parameters. Through the body hierarchy, the evaluation indexes can be classified into a component level, a subsystem level and a system level, so that a multi-level evaluation system is formed. Such classification helps to systematically understand and analyze the health status of the device. When traversing the knowledge graph by adopting the breadth-first search algorithm, key nodes closely related to the running state of the equipment can be effectively identified. Taking an air conditioning system as an example, if the operating temperature of the compressor is found to be a critical factor affecting system performance, then this index is selected as the key evaluation index. Next, by extracting attribute nodes that are directly related to the compressor temperature, such as coolant temperature and ambient temperature, it is possible to further analyze how these factors affect the operating efficiency of the compressor. When the real-time monitoring equipment operates, the acquired data is required to be mapped to corresponding attribute nodes in the knowledge graph. For example, if the real-time data shows that the operation temperature of the compressor is 50 degrees and the normal operation temperature threshold set in the knowledge map is 30 to 40 degrees, it may be judged that the compressor is at risk of overheating. This comparison can immediately reveal potential equipment problems, thereby taking appropriate maintenance measures. By means of the depth-first search algorithm, other attribute nodes related to the indirect can be explored from a key evaluation index. This helps build a more comprehensive device health assessment system. For example, indirect factors such as power stability, compressor age, etc., may need to be considered in addition to factors that directly affect compressor temperature. The analytic hierarchy process provides a systematic approach in determining the weights of the assessment indicators. First, a matrix with index comparison is constructed, and element values in the matrix are determined through expert scoring. And calculating the maximum eigenvalue of the matrix and the eigenvector corresponding to the maximum eigenvalue, and normalizing the eigenvector to obtain the weight of each index. These weights reflect the relative importance of the different metrics in the device health assessment. And finally, multiplying and summing the real-time data of each evaluation index with the corresponding weight by constructing a weighted average model, so as to obtain the current health state quantization score of the equipment. For example, if the weights for compressor temperature, power stability, and compressor age are 0.5, 0.3, and 0.2, respectively, and the current real-time data are 50 degrees, good, and medium, respectively, the composite score may indicate the overall health of the device. When the score is lower than a set early warning threshold, the system automatically sends out early warning to prompt maintenance personnel to pay attention to possible equipment faults.

And S106, constructing a evidence reasoning network by adopting a evidence reasoning method aiming at the generated evaluation index system. And taking expert knowledge and a data analysis result as evidence nodes, calculating the association relation and the credibility between evidences by using the Derster-Sha Fu theory, and carrying out joint reasoning to generate a health state assessment result comprehensively considering qualitative and quantitative factors.

And constructing a health state evaluation index system according to expert knowledge and data analysis results, and taking each index as an evidence node. And determining the weight of each evidence node by adopting an analytic hierarchy process in a way of expert scoring, and taking the weight as the credibility of the evidence node. And calculating the correlation between different evidence nodes by using the pearson correlation coefficient as the correlation strength between the evidence nodes. And constructing the evidence reasoning network by using the network analysis software Gephi according to the credibility and the association strength of the evidence nodes. In the evidence reasoning network, a Bayesian network reasoning algorithm is used for comprehensively analyzing the support degree of each evidence node to different health states. And (3) calculating posterior probabilities of various health states through Bayesian network reasoning, wherein the health state with the highest probability is the health state in which the evaluation object is most likely to be, and taking the health state as a result of health state evaluation. And according to the health state evaluation result, analyzing and explaining the current health level of the evaluation object, and providing targeted health management advice to guide the evaluation object to improve the health state.

Illustratively, constructing a health state assessment index system first requires explicit assessment of the subject. Assuming that the evaluation object is an industrial robot, it is desired to evaluate its health state. Expert knowledge and data analysis results show that the health state of the robot is mainly influenced by the following indexes of joint temperature, vibration amplitude, working current, working time and maintenance record. These indicators constitute evidence nodes. The analytic hierarchy process is an effective method for determining the weight of each evidence node. Inviting a plurality of robot specialists, comparing the indexes in pairs, and scoring. For example, if the expert considers that the joint temperature is slightly more important than the vibration amplitude, the importance score given to the joint temperature with respect to the vibration amplitude is 3. By collecting scores from all experts, a comparison matrix can be constructed. The weights of the indexes, such as the weight of the joint temperature, the vibration amplitude, the working current, the working time and the maintenance record are calculated to be 0.3, 0.2, 0.25, 0.15 and 0.1 respectively. These weights represent the degree of importance of the individual indicators in assessing the health of the robot. The pearson correlation coefficient may be used to measure the strength of association between different evidence nodes. For example, by analyzing the historical data, a strong positive correlation between joint temperature and vibration amplitude was found, with a pearson correlation coefficient of 0.8. This suggests that an increase in joint temperature is often accompanied by an increase in vibration amplitude. Similarly, correlation coefficients between other evidence nodes, such as between operating current and operating time, between vibration amplitude and operating current, etc., may be calculated. Evidence inference networks can be constructed using network analysis software such as Gephi. And taking each index as a node, taking the correlation among the indexes as the weight of the edge, and visualizing the weight of the index as the size of the node. The size of the nodes reflects the importance of the indexes, and the thickness of the edges reflects the association strength between the indexes. The network diagram can intuitively display the relation among the indexes, and is helpful for understanding the influence factors of the health state of the robot. Based on the constructed evidence reasoning network, a Bayesian network reasoning algorithm can be used for health status assessment. The bayesian network is a probabilistic graph model that can efficiently handle uncertainty information. First, the health of the robot needs to be defined. For example, the health status may be divided into three states of "healthy", "sub-healthy" and "faulty". And then, according to expert knowledge and data analysis results, determining the support degree of each evidence node to different health states. For example, joint hyperthermia may support a "failure" condition, while regular maintenance records support a "health" condition. The posterior probabilities of various health states can be calculated through Bayesian network reasoning. Assuming that the probability of the "healthy" state is calculated to be 0.7, the probability of the "sub-healthy" state is calculated to be 0.25, and the probability of the "fault" state is calculated to be 0.05, based on the current monitoring data. Since the probability of the "healthy" state is the greatest, the result of the evaluation is that the robot is currently in the "healthy" state. And finally, according to the health state evaluation result, analyzing and explaining and providing health management advice. For example, robots are currently in a "healthy" state, but considering that the joint temperature is slightly raised, it is recommended to strengthen the heat dissipation and pay attention to the change in vibration amplitude. If the evaluation is in a "sub-health" state, further checks are recommended to find potential problems. If the evaluation result is in a fault state, the machine is required to be stopped immediately for maintenance. Through the steps, a robot health state assessment system based on evidence reasoning can be constructed. The system comprehensively considers a plurality of indexes, utilizes a Bayesian network reasoning algorithm, can accurately evaluate the health state of the robot, and provides corresponding health management suggestions. This helps to improve reliability and life of the robot, reducing maintenance costs.

And step S107, constructing an evidence reasoning network by adopting network analysis software according to the credibility value and the association strength of the evidence nodes, comprehensively analyzing the support degree of each evidence node to different health states through a Bayesian network reasoning algorithm, obtaining posterior probabilities of various health states, and determining the health state in which an evaluation object is most likely to be in so as to obtain a health state evaluation result.

And constructing an evidence reasoning network according to the credibility value and the association strength of the evidence nodes. And converting the constructed evidence reasoning network into a corresponding Bayesian network model. Aiming at the constructed Bayesian network model, adopting a Bayesian network reasoning algorithm to comprehensively analyze the support degree of each evidence node on different health states. And obtaining posterior probabilities corresponding to various health states according to analysis results of the Bayesian network reasoning algorithm. The posterior probabilities of the various health states are ordered in descending order. And determining the most probable health state of the evaluation object by comparing the ordered health state posterior probabilities. Comparing the posterior probability of the most probable health state with a preset threshold, and outputting the health state as a health state evaluation result if the posterior probability exceeds the preset threshold; and if the pre-set threshold is not exceeded, selecting the rest non-analyzed evidence nodes according to the pre-defined evidence node priority. Adding the selected unanalyzed evidence nodes into the existing evidence reasoning network, and updating the Bayesian network model by recalculating the conditional probability table. And re-executing the Bayesian network reasoning algorithm by using the updated Bayesian network model until the output condition is met or all evidence nodes are analyzed. And outputting the health state with the highest final posterior probability as a health state evaluation result of the evaluation object.

Illustratively, in practical applications, constructing a evidence inference network and converting to a bayesian network model is a common method to evaluate the state of a complex system, such as the state of health evaluation of an industrial robot. First, by collecting operational data about the robot, such as joint temperature, vibration amplitude, etc., these data are input as evidence nodes into the evidence inference network. The credibility of each evidence node is evaluated according to historical data and expert experience. For example, if the historical data shows that an anomaly in joint temperature is generally indicative of an impending failure of the robot, then the reliability of the joint temperature node will be set higher. Next, a Bayesian network model is constructed based on the correlations between the evidence nodes. For example, if the analysis finds a high correlation between joint temperature and vibration amplitude, the two nodes will be connected by one edge in a bayesian network. The weight of an edge reflects the strength of the association between two nodes, and this weight can be determined by calculating the pearson correlation coefficients of the two nodes. After the Bayesian network model is constructed, a Bayesian network reasoning algorithm is used for analyzing the support degree of each evidence node to different health states. For example, by network reasoning, it can be found that the combination of high joint temperature and high vibration amplitude greatly supports the assumption that the robot is in a "faulty" state. From these analyses, the posterior probabilities of the robot being in a healthy, sub-healthy and faulty state can be calculated. By ordering the posterior probabilities, the most likely health of the robot can be determined. If the posterior probability of the "fault" state is highest and exceeds a preset threshold, then this state can be directly used as the health assessment result of the robot. If the probability of none of the states exceeds the threshold, then more evidence nodes, such as operating current and maintenance records, may need to be considered, added to the existing Bayesian network, and the conditional probability table recalculated. By the method, the Bayesian network model can be gradually perfected until a reliable health state evaluation result is obtained. The Bayesian network-based reasoning method not only can provide quantitative health state assessment, but also can adapt to the change of the running state of the robot by continuously introducing new evidence nodes, thereby realizing dynamic health management and preventive maintenance strategies.

And step S108, based on the evaluation result, acquiring equipment operation data in real time by adopting an online incremental learning technology, and dynamically updating a data association analysis model and an early warning model. The method is continuously suitable for the change of the working condition of equipment through a sliding time window and a forgetting factor mechanism. Meanwhile, an adaptive threshold adjustment mechanism is introduced, and an exponential weighted moving average method is used for dynamically adjusting the early warning threshold according to the historical data and the current running state of the equipment, so that the sensitivity and the reliability of early warning are improved.

And acquiring equipment operation data. Real-time data from various sensors including temperature, pressure, current, voltage, vibration frequency and the like are continuously collected, data cleaning and preprocessing are performed, abnormal values and noise are removed, and data quality is guaranteed. And carrying out online incremental modeling on the data by using the preprocessed data and adopting a recursive least square method, updating a correlation model among the operating parameters of the equipment in real time, and identifying the association relation among different parameters and potential fault modes. Based on the updated data association model, the parameters of the fault early warning model are dynamically adjusted by using the online sequence extreme learning machine, so that the potential equipment faults can be predicted more accurately. The online sequence extreme learning machine can realize quick online learning by randomly generating hidden layer parameters and fixing and only updating output weights. By adopting a sliding time window mechanism, only the equipment operation data in the last period of time is reserved, so that the current equipment working condition is better reflected, and the influence of historical data on a model is reduced. The length of the time window may be set according to the characteristics of the device and the period of the data change, and is typically several hours to several days. A forgetting factor mechanism is introduced, so that higher weight is given to new data, the weight of old data is reduced, and the model can adapt to the change of the working condition of equipment more quickly. The forgetting factor can be adjusted according to the data change speed of the device, and is generally 9 to 1. And dynamically adjusting a fault early warning threshold value by using an exponential weighted moving average method according to the historical data and the current running state of the equipment. The calculation formula is that the current threshold value=alpha×the predicted value of the current moment + (1-alpha) x the last moment threshold value, wherein alpha is a smoothing coefficient, and the value is generally between 0.5 and 2. The threshold value needs to be set by comprehensively considering the safety margin, the maintenance cost, the false alarm rate and other factors of the equipment. If the current fault prediction value exceeds the dynamic early warning threshold value, early warning information is generated, and relevant personnel are informed to take measures. If the predicted value is lower than the threshold value, the equipment is considered to be normal in operation, and on-line monitoring and model updating are continued.

Illustratively, acquiring device operational data is the first step in fault pre-warning. For example, in a chemical plant, data such as temperature and pressure of a reaction kettle, current and vibration of a stirring motor are required to be monitored. These data are collected in real time by various sensors mounted on the reactor and transmitted to a data center via an industrial control network. In order to ensure the data quality, the collected data needs to be cleaned and preprocessed. For example, abnormal values due to sensor malfunction and disturbances due to environmental noise are removed. The data may also be smoothed to reduce the effects of data fluctuations. Data cleaning and preprocessing can improve accuracy and reliability of the model. With the preprocessed data, the data can be modeled in online increments using a recursive least squares approach. For example, a model of the relationship between reactor temperature and pressure may be established. Because the running state of the reaction kettle can change along with time, model parameters need to be updated in real time so as to better reflect the current equipment working condition. The online incremental modeling can avoid repeated calculation of all historical data, so that the calculation efficiency is improved. Meanwhile, the model can be continuously adjusted according to new data, and the adaptability of the model is improved. Based on the updated data correlation model, parameters of the fault early warning model can be dynamically adjusted using line sequence extreme learning. For example, according to a relation model of temperature and pressure, the temperature change trend of the reaction kettle in a future period of time can be predicted. The online sequence extreme learning machine is an efficient online learning algorithm, and can realize rapid online learning by randomly generating hidden layer parameters and fixing and only updating output weights, so that the online sequence extreme learning machine can adapt to the change of equipment working conditions more rapidly and find potential faults in time. To better reflect the current device operating conditions, a sliding time window mechanism is required. For example, only the last 24 hours of autoclave run data was retained. Thus, the influence of the historical data on the model can be reduced, and the model is focused on the current equipment state. The length of the time window needs to be set according to the characteristics of the device and the data change period. For example, the time window may be set shorter for faster changing devices and longer for slower changing devices. The forgetting factor mechanism is introduced to give higher weight to new data and reduce the weight of old data. For example, a forgetting factor of 0.9 may be set, which means that the latest data has the greatest effect on the model, while older data has a progressively smaller effect on the model. The forgetting factor mechanism can enable the model to adapt to the change of the working condition of the equipment more quickly, so that the prediction accuracy of the model is improved. According to the historical data and the current running state of the equipment, an exponential weighted moving average method can be used for dynamically adjusting the fault early warning threshold. For example, the temperature early warning threshold value can be dynamically adjusted according to the deviation between the predicted value and the actual value of the temperature of the reaction kettle in the past period of time. If the current temperature predicted value exceeds the dynamic early warning threshold value, early warning information is generated, and relevant personnel are informed of taking measures. For example, if the predicted reactor temperature is about to exceed the upper safety limit, the system may automatically alert the operator to lower the reactor temperature. The dynamic adjustment of the threshold value can effectively avoid false alarm and missing alarm caused by the change of the working condition of the equipment. For example, assume that the temperature of the reaction vessel continues to rise, exceeding a set threshold. The system can send out early warning information to remind an operator to check whether the cooling system of the reaction kettle works normally. If the cooling system fails, an operator can take measures in time, such as starting a standby cooling system or reducing the operation load of the reaction kettle, so as to prevent safety accidents caused by overhigh temperature of the reaction kettle. If the vibration frequency of the reaction kettle suddenly rises and exceeds a preset threshold value, the system also sends out early warning information to remind an operator to check whether the stirring motor of the reaction kettle fails. The dynamic early warning threshold value can be adjusted according to the real-time running state of the equipment, so that the accuracy and timeliness of early warning are improved, the false alarm rate is reduced, unnecessary shutdown and maintenance are avoided, and meanwhile, the potential faults can be found in time, and the occurrence of accidents is prevented.

Step S109, guaranteeing data quality through data cleaning and preprocessing, identifying parameter association relation and potential fault modes by adopting a recursive least square method, and setting sliding time window length according to equipment characteristics and data period to obtain data which more accurately reflects the current equipment working condition.

By cleaning and preprocessing the original data, removing abnormal values and noise data, complementing the missing values, and carrying out standardization and normalization processing on the data, the data quality and usability are improved. After the data preprocessing is completed, a proper sliding time window is set according to the working characteristics and the data acquisition period of different types of equipment, the data are sliced and aggregated in the time window, and the time domain and frequency domain characteristics reflecting the current working condition of the equipment, such as mean value, variance, peak-to-peak value, spectrum entropy and the like, are extracted. And (3) carrying out importance analysis and screening on the extracted features by using indexes such as pearson correlation coefficient, maximum information coefficient and the like, and selecting a feature subset with the maximum correlation with the equipment fault state. And a classification model between the equipment parameters and the fault modes is established by adopting a support vector machine SVM, and the super parameters of the model are optimized through grid search and cross verification, so that the accuracy of fault diagnosis is improved. In the model training process, the weight coefficient of each feature is recorded, and the larger the weight is, the larger the contribution of the feature to classification decision is indicated, so that the intrinsic mechanism of fault occurrence is facilitated to be understood. When the model is applied, whether the working condition of the equipment is obviously changed is judged by calculating statistical indexes such as root mean square value, kurtosis and the like of the vibration signal of the equipment. If the index exceeds the preset threshold, the working condition is considered to be changed, the length of the sliding time window is dynamically shortened, and the sensitivity and the instantaneity of fault diagnosis are improved. And performing incremental updating on the fault diagnosis model by adopting an online sequential extreme learning machine OS-ELM algorithm. And when new equipment operation data is collected, automatically adding a training set, updating the network connection weight, adapting to the dynamic change of the equipment working condition, and continuously improving the model performance. And predicting the degradation trend of the health state of the equipment by carrying out trend extrapolation and similarity analysis on the historical operation data of the equipment, and estimating the residual service life. When the predicted remaining life is lower than a preset threshold, the factors such as the importance degree, the maintenance cost and the fault result of the equipment are combined, an optimal maintenance scheme is generated by utilizing a multi-objective optimization algorithm, the replacement and maintenance strategies of spare parts are formulated, the reliability of the equipment is ensured, the maintenance cost is minimized, and the economical efficiency of equipment management is improved.

Data cleaning and preprocessing, for example, are the basis for fault pre-warning and diagnosis. The original data may have problems of abnormal values caused by sensor faults, interference caused by environmental noise, data missing and the like. For example, a reactor temperature sensor in a chemical plant occasionally has peak readings far outside of normal range, which may be caused by sensor failure or external disturbances. To improve the data quality, these outliers need to be culled. Common methods include statistical-based methods, such as the 3 sigma criterion, that consider data that exceeds the mean by a factor of + -3 standard deviations as outliers. In addition, a model-based method, such as smoothing the data using Kalman filtering, may be used to remove noise interference. For missing values, interpolation, mean filling, or model-based prediction methods may be employed for completion. The data normalization and normalization can eliminate the influence of the dimensions among different features, for example, the data of the different dimensions such as temperature, pressure and the like are converted into dimensionless data, so that the overlarge influence of certain features on the model is avoided, and the accuracy and the stability of the model are improved. For different types of equipment, a proper sliding time window needs to be set according to the working characteristics and the data acquisition period. For example, for rotating machinery, such as motors, pumps, etc., where the vibration signal changes rapidly, a shorter time window, e.g., a few seconds or minutes, may be provided. While for some slowly changing equipment, such as reaction kettles, tanks etc., a longer time window may be provided, for example a few hours or days. And slicing and aggregating the data in a time window, and extracting time domain and frequency domain features reflecting the current working condition of the equipment. The time domain features include mean, variance, peak-to-peak value, root-mean-square value, kurtosis, etc., reflecting the amplitude and fluctuation of the signal. The frequency domain features include frequency spectrum, frequency spectrum entropy, etc., reflecting the frequency content and energy distribution of the signal. These features can effectively describe the operating state of the device and provide basis for fault diagnosis. Feature selection is one of the key steps in fault diagnosis. By carrying out importance analysis and screening on the extracted features, redundant features and irrelevant features can be removed, the complexity of the model is reduced, and the diagnosis efficiency and accuracy are improved. For example, the pearson correlation coefficient may be used to measure the linear correlation between features and fault conditions, with features of higher correlation being selected. The maximum information coefficient can measure the nonlinear correlation between the feature and the fault state, and the importance of the feature is reflected more comprehensively. Assuming that the average value, variance and spectral entropy of the temperature of the reaction kettle are found to have higher correlation with the fault state by analysis, and other features such as pressure, stirring motor current and the like have lower correlation with the fault state, the average value, variance and spectral entropy of the temperature can be selected as feature subsets for constructing a fault diagnosis model. A Support Vector Machine (SVM) is a commonly used fault diagnosis classification model. It can effectively cope with high-dimensional data and non-linearity problems. Through grid search, cross verification and other methods, the hyper-parameters of the SVM model, such as kernel function type, penalty coefficient and the like, can be optimized, and the generalization capability and the diagnosis accuracy of the model are improved. In the model training process, the weight coefficient of each feature is recorded, the contribution degree of different features to fault diagnosis can be analyzed, for example, the weight coefficient of the temperature mean value is larger, the fact that the contribution of the temperature mean value to fault diagnosis is larger is explained, and the understanding of the inherent mechanism of fault occurrence is facilitated. When the model is applied, the working condition of the equipment needs to be monitored in real time. For example, by calculating statistical indexes such as root mean square value, kurtosis and the like of the vibration signals of the equipment, whether the working condition of the equipment is changed significantly is judged. If the index exceeds the preset threshold, the working condition is considered to be changed, and the model parameters are required to be dynamically adjusted or corresponding measures are required to be taken. For example, the length of the sliding time window can be shortened, and the sensitivity and the real-time performance of fault diagnosis can be improved, so that the abnormal state of the device can be captured more quickly. An online sequential extreme learning machine (OS-ELM) is an efficient online learning algorithm suitable for processing streaming data and updating models in real time. The OS-ELM algorithm may automatically add new data to the training set and update the network connection weights whenever new device operational data is collected, thereby enabling the model to adapt to dynamic changes in device operating conditions and constantly improving model performance. by trend extrapolation and similarity analysis of the historical operating data of the device, the degradation trend of the health state of the device can be predicted, and the remaining service life can be estimated. For example, a time series analysis method, such as ARIMA model, may be used to predict trends in key parameters of the device to determine whether the device is in a degraded state. When the predicted remaining life is lower than a preset threshold, factors such as importance degree, maintenance cost and fault result of equipment are required to be considered, and a reasonable overhaul scheme is formulated. For example, a multi-objective optimization algorithm can be utilized, so that maintenance cost is minimized and economical efficiency of equipment management is improved on the premise of ensuring equipment reliability. for example, if the predicted remaining life of a critical device is about to be below a safety threshold, an inspection plan may be scheduled in advance to avoid production interruption due to equipment failure, resulting in greater economic loss.

The foregoing is only a preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art, who is within the scope of the present invention, should make equivalent substitutions or modifications according to the technical scheme of the present invention and the inventive concept thereof, and should be covered by the scope of the present invention.

Claims (9)

1. A device status intelligent early warning method based on danger perception, characterized by comprising: Acquiring device operation data, wherein the device operation data includes multi-source heterogeneous data; Extracting features from the multi-source heterogeneous data to obtain a unified feature representation; According to the unified feature representation, a correlation model between the equipment operation status and the feature parameters is constructed; Acquire the device operation data to be analyzed, and analyze the device operation data to be analyzed according to the association model to obtain the device operation status analysis result; According to the equipment operation status analysis result, combined with a preset equipment health status evaluation index system, the current health status of the equipment is evaluated to obtain an equipment health status evaluation result. 2. The device status intelligent early warning method based on danger perception according to claim 1 is characterized in that the feature extraction of the multi-source heterogeneous data to obtain a unified feature representation includes: Acquire the multi-source heterogeneous data, and process the multi-source heterogeneous data using data cleaning and preprocessing methods; For the processed multi-source heterogeneous data, the principal component analysis method is used to reduce the dimension and extract key feature parameters; The extracted key feature parameters are input into the autoencoder network, and a unified feature representation of the multi-source heterogeneous data is obtained through unsupervised learning. 3. The device state intelligent early warning method based on danger perception according to claim 1 is characterized in that the association model between the device operation state and the characteristic parameters is constructed according to the unified characteristic representation, comprising: Acquire historical equipment operation data, and extract characteristic parameters of the historical equipment operation data according to the unified characteristic representation; Obtaining the equipment operation status label corresponding to the historical equipment operation data; The characteristic parameters are used as input and the equipment operation status labels are used as output to train a support vector machine model to obtain an association model between the equipment operation status and the characteristic parameters. 4. The device status intelligent early warning method based on danger perception according to claim 1 is characterized in that the current health status of the device is evaluated based on the analysis result of the device operation status and in combination with a preset device health status evaluation index system to obtain the device health status evaluation result, including: Obtain a preset equipment health status evaluation indicator system, the indicator system including multiple evaluation dimensions and corresponding indicator items; In combination with the equipment operation status analysis result, index items related to the current equipment operation condition are extracted from the index system as evaluation indicators; For each evaluation indicator, determine the health status of the indicator based on its threshold range and actual value; The health status of each evaluation indicator is comprehensively considered and the weighted average method is adopted to obtain the current health status evaluation result of the equipment. 5. The device state intelligent early warning method based on danger perception according to claim 1, characterized in that the device state intelligent early warning method based on danger perception further comprises: Acquire real-time data during the operation of the equipment, and dynamically update the association model using an online incremental learning method to adapt to changes in equipment operating conditions; The equipment operating status is predicted based on the updated correlation model. When the prediction result exceeds the preset threshold, an early warning is triggered and an alarm message is generated. 6. The device state intelligent early warning method based on danger perception according to claim 1 is characterized in that before constructing the association model between the device operation state and the characteristic parameters, it also includes: Based on the unified feature representation, a frequent pattern mining algorithm is used to discover association rules and time series patterns in multi-source heterogeneous data; The mined association rules and time series patterns are used as prior knowledge to guide the construction of association models. 7. The device status intelligent early warning method based on danger perception according to claim 4 is characterized in that after the health status of each evaluation index is comprehensively evaluated by a weighted average method to obtain the current health status evaluation result of the device, it further comprises: Obtain equipment operation data; Continuously collect real-time data from various sensors, and perform data cleaning and preprocessing to remove outliers and noise; Compare the equipment health status assessment result with the preset health level threshold to determine the current health level of the equipment; According to the health level and importance of equipment, equipment maintenance priority plan and maintenance strategy are automatically generated to guide the development of equipment operation and maintenance work. 8. The device status intelligent early warning method based on danger perception according to claim 4 is characterized in that obtaining a health status evaluation index system comprises: Obtain multiple indicators as evidence nodes; For the indicator system, a hierarchical analysis method is used to determine the weight of each of the evidence nodes, and the weight is used as the credibility of the evidence node; Calculating the Pearson correlation coefficient between different evidence nodes, and using the correlation coefficient as the correlation strength between the evidence nodes; Constructing an evidence reasoning network according to the credibility of the evidence nodes and the association strength; In the evidence reasoning network, a Bayesian network reasoning algorithm is used to comprehensively analyze the support of each evidence node for different health states. 9. The device status intelligent early warning method based on danger perception according to claim 7, characterized in that the acquisition of device operation data comprises: Preprocessing the equipment operation data to remove outliers and noise and to complete missing values; According to the type and working characteristics of the equipment, a sliding time window is set, and the pre-processed equipment operation data is sliced and aggregated within the sliding time window; For the aggregated data, extract its time domain features and frequency domain features, wherein the time domain features include mean, variance and peak-to-peak value, and the frequency domain features include spectrum entropy; The Pearson correlation coefficient and maximum information coefficient are used to analyze the importance of the extracted features and select the feature subset with the greatest correlation with the equipment fault status; Using a support vector machine, a classification model between the feature subset and the equipment failure mode is established; Obtaining a weight coefficient corresponding to each feature in the classification model, and determining a feature whose weight coefficient is greater than a preset threshold as a key factor causing equipment failure; Monitor the equipment operating status online and calculate the RMS value and kurtosis of the currently collected vibration signal; If the RMS value and kurtosis exceed the preset range, it is determined that the equipment operating condition has changed significantly, and the length of the sliding time window is dynamically shortened; Adding newly collected operation data of the equipment to the training set, and incrementally updating the classification model using an online sequential extreme learning machine algorithm; Predict the remaining useful life of the equipment based on the updated classification model; If the remaining service life is lower than a preset threshold, an equipment maintenance plan is generated through a multi-objective optimization algorithm based on factors such as equipment importance, maintenance cost and failure consequence.