DE19855745B4 - Method for fault diagnosis and quality assessment (rating) of rotating machines on a series test bench - Google Patents

Abstract

Method for automatic fault identification and rating of rotating machines on a series test bench by evaluation of their emitted sound and processing of the sound signal by means of Fourier analysis or order analysis for determining the feature vector, wherein three evaluation systems, namely a rough diagnosis, a fine diagnosis and a rating system are connected in parallel,
wherein the rough diagnosis system includes a threshold value comparator which compares predefinable threshold values for the individual feature vector components with current component values and the diagnosis is carried out by automatic evaluation of the part list of the test object,
wherein the fine diagnostic system includes an artificial intelligence (AI) based classifier that is trained unsupervised based on the gross diagnostic results and that refines its diagnostic capability through supervised learning,
wherein the rating system includes an AI-based classifier that learns through supervised training based on patterns and the results of the coarse and fine diagnosis,
and where the classification results of the coarse and fine diagnosis and the rating in addition to the associated time and speed signals archive the feature vector and the operating data and for off-line training ...

Description

rotating Machines in the sense of this description are machines that are next to fixed parts also rotating parts such. Gears, bearings and waves included. Particular importance here vehicle transmissions, Electric and internal combustion engines and their components and ancillaries such as. Clutches, turbochargers, differentials. Includes fault diagnosis the detection of a faulty test object, the identification of the faulty component (s) as well as the description of the fault, e.g. Tooth fracture, imbalance. As a rating, the quality assessment of the candidates by Figure designated in a reference scale; a common in the automotive industry Rating scale goes from 1 (= very bad) to 10 (very good).

rotating Parts generate vibrations (structure-borne noise and noises). The appropriate evaluation of these vibrations provides information on quality and possible Manufacturing defects.

at the quality inspection of rotating machines are methods using threshold value deciders known, in which substantially sound level or certain frequency components with manually preset tolerance values are compared. This procedure result in the classification of the candidate as OK or NOK, one detailed error analysis is not performed by the measuring system but must be done by disassembling the test object be determined.

spread alternative quality assessments are the evaluation of the sound emissions of the test specimen by the test personnel or listening of structure-borne sound with stethoscope or ear tube. The examiner conducted However, rating is subjective and poorly reproducible.

Method, those on trainable systems from the environment of the artificial Build intelligence, e.g. neural networks, have so far due their training effort no entry into production for the purpose error identification and rating.

The in Automatisierungsstechnik 39 (1991) 2, pages 43-48, listed methods uses neural networks for classification, but deviates in the approach to train the neural networks sustainably from below suggested Procedure and there is a lack of error diagnostic capability in the sense of identification the mechanical excitation component. The in Picture and Sound Bd. 45 (1992), Issue 11/12, pages 320-326, listed Method describes a method for vibration analysis-based Material testing, also known as sound testing is. Significant differences to the proposed method lie in the procedure for training the classifier, in the result the classification and the respective field of application.

The in metrology and measurement signal processing, K. W. Bonfig, ed. Renningen-Malsheim: expert-Verlag 1996, pages 122-127, listed procedures the order analysis and the measurement setup described there are not The subject matter of the present claims.

The The present invention relates to the realization and the Training of a classifier for fault diagnosis and rating of rotating machines on the series test bench. One after this invention Realized and trained classifier allows a detailed fault diagnosis on the series test bench already with the first examinee, needed so none of the exam previous explicit training phase. Through continuous use while the series test gained experience, the fault diagnosis is refined.

According to the invention Asked task by the specified in the characterizing part of claim 1 Characteristics solved.

advantageous Further developments and refinements of the method are in the dependent claims 2 to 8 indicated. The achievable with the present invention advantages from the description below.

According to the invention the well-known mechanical structure of the specimen and the fixed transmission conditions between rotating parts with known speed signal for fault diagnosis used.

1 contains a block diagram for the construction of the universal fault diagnosis system. In the "Sensor" module, the measured data are recorded.The system measures not only structure-borne noise and / or noise, but also the speed or several speed signals as well as operating data, the measurement setup may vary depending on the application Compression pressure in internal combustion engines The measured data at the output of the "Sensors" module are digitized. The module "Signal preprocessing" processes the digitized measurement data for a detailed signal analysis NEN are time signal segmentation, windowing and filtering of the vibration signals.

in the Block "feature calculation" are from the preprocessed Data extracts those features that are significant to the classification are. The feature vector then becomes the feature vector thus determining the significant information for discrimination the different test specimen qualities and contains the faulty components. Operating data, speed ranges and load conditions can Be part of the feature vector.

Of the Feature vector is determined by parallel classifiers, the Blocks "rough diagnostics", "fine diagnostics" and "rating" processes are described below. For each measurement, the sensor signals, Feature vectors and classification results in an archive for Documentation of the inspection process and to train the systems "enemy diagnostics" and "rating".

Should As provided here rotating machinery to be tested, it is appropriate method to use for order analysis. A method based on digital Resampling is in Groppe, Jonuscheit, Strama, Thomä: "order analysis", metrology and Measurement signal processing, MESSCOMP 1996, expert Verlag, 1996, described. At constant speeds, instead of order analysis can also Fast Fourier transformation can be used.

The Meaning of order analysis for the downstream classification is that an immediate and speed independent Relationship between the order lines and the mechanical structure of the test piece given is. This is the example below for a transmission explains: On the drive shaft sits a gear with 23 teeth, in which a second gear with 44 teeth should intervene.

2 shows a section of the associated Ordnungsspektrum relative to the drive shaft. Order 1 corresponds to the revolution of the drive shaft. The order 23 is the order of meshing. Gear Z23 and gear Z44 contribute to the total energy of the meshing order, their respective shares are not to be separated. Due to imbalance on the gears Z23 and Z44 arise to the order of the meshing sidebands. Crucial for the fault diagnosis is that the sidebands are clearly attributable to the mechanical causers. At a distance of integer multiples of ± 1 order from the meshing order are the sideband lines of gear Z23, at a distance of ± 23/44-stel order are the sideband lines of gear Z44. The order analysis requires a sufficiently high resolution to separate the sidebands of gears Z23 and Z44.

The Order analysis must therefore have a sufficiently high resolution, around the sidebands of the to separate rotating components. Order resolution and Speed accuracy has a lasting effect on the informative value. The existing ones Harmonics, e.g. the meshing order, can also be used for the classification be used.

3 shows the function blocks of the classifier for rough diagnostics. A threshold value comparator compares the feature vector component by component against a predetermined upper and lower threshold. The order lines that are above or below the respective thresholds are analyzed in the block "rough diagnosis" for their mechanical causes.This knowledge is available to the block "rough diagnosis" by evaluating the unique and available parts list of the test object. The parts list contains all components of the test object with the relevant design features such as gear type, number of teeth, bearing type. The measuring system automatically determines those orders that are constructive, ie excited, during production startup, production change and BOM changes, and stores these in a reference table. In the case of the rough diagnosis, the relation between the deviating order lines and their mechanical exciters can thus be established. The connection between order resolution and pathogen identification must be taken into account.

the Threshold decision makers go through the upper and lower thresholds given an expert. The expert can be the result of the rough diagnosis use to adjust the thresholds. The diagnostic ability The threshold decision maker is and remains the context of BOMs and the derivable assignment of excited orders limited to the respective mechanical agents. The result of the rough diagnosis is stored in the archive for documentation and further processing stored in the "Advanced diagnostics" block.

The feature vector is evaluated parallel to the processing in the module "rough diagnosis" also in the module "fine diagnostics". 4 In essence, "fine diagnostics" includes a classifier based on Artificial Intelligence methods, hereafter referred to as an AI-based classifier. The AI-based classifier is trainable and distinguishes itself according to the invention in that it can be Knowledge partly (by the expert) unsupervised and partially supervised (by the expert) learns. The training concept is explained below.

To the The start of production is the AI-based classifier "Wissenfrei." The AI system is learning while the series production permanently from the "rough diagnosis" .This training done by the expert unmonitored by using the "rough diagnosis" results.

The AI-based classifier can detect additional errors through further training, which are not recognized by the module rough diagnosis, eg because the mechanical cause of the error is not clear (see example to 2 ) or the error is in non-excited orders; non-excited orders are orders that have no direct reference to an excited mechanical component. In addition to identifying the damaged component, the AI-based classifier can also learn to distinguish types of errors on the same component, such as gear breakages and imbalances, or speed-independent errors or statistically occurring errors as they can occur in combustion processes. This property for detailed error indication is understood here as the fine diagnostic capability of the AI-based classifier.

Around this fine diagnostic ability to promote the expert the opportunity an (archived) test result a particular interpretation, e.g. a special component error, assigned.

Becomes a mistake on the test object it is common the disassembly of the test object necessary, to determine the cause of the error clearly. In the meantime will be the series test continued. It is therefore important that the test results and associated feature vectors and / or measurement signals, speed signals and / or operating data archived and for one latter Evaluation of the test piece available are.

is an error can be uniquely identified, the feature vector and the correct diagnosis is assigned to the measuring signal, the expert evaluates the archived data. The rating will also be archived.

Of the AI-based classifiers train based on their training data typically off-line. Each class is a limited number of patterns available in the archive. After completing the training, the corresponding system parameters on the measuring system on the series test system copied the system parameters for a later one Traceability of the test process archived.

The Availability of the classification result of the module "rough diagnosis" is essential for the training of the "enemy diagnostics" system Trainings of the AI based classifier are e.g. in neural Networks clusters for equivalent classes determined (compare Kolb: Tutorial Intelligent Systems - Neuronal Networks and Statistical Methods of Pattern Recognition, MEDAV, 1998) or in case-based reasoning methods similarity measures (cf. for example, e.g. Thank God, Frühwirth, Horn: Expert Systems, Springer Verlag, 1990). A typical problem When training AI-based classifiers, the description of the Cluster expansion; This problem is especially in a relative given a small number of training patterns for different classes. The "rough diagnosis" results will be used to describe the cluster boundaries or similarity measures. With this strategy, the majority of production, namely the production of faultless products, through an unmonitored Learning process for to use the training.

The classification system "Rating" is comparable to the functionality of the classifier "Feiniagnostik" according to 4 , The ratings of the rating classifier are evaluated by the expert, and the errors as identified in the "rough diagnosis" and "fine diagnostics" modules are assigned rating values and also used for the training, the selected samples and their rating are archived. After completion of the training, the corresponding system parameters are copied to the measuring system on the series test system and the system parameters are archived for later traceability of the test process.

The resulting classification results are juxtaposed and complete yourself. The results of "rough diagnosis" and "fine diagnostics" can be used for mutual plausibility check be used. The "Rating" result can be specified become.

Claims (8)

Method for automatic fault identification and rating of rotating machines on a series test bench by evaluation of their emitted sound and processing of the sound signal by means of Fourier analysis or order analysis for determining the feature vector, wherein three evaluation systems, namely a rough diagnosis, a fine diagnosis and a rating system are connected in parallel, wherein the rough diagnosis system includes a threshold value comparator which compares predefinable threshold values for the individual feature vector components with current component values and the diagnosis by automatic evaluation of the pieces The candidate system includes an Artificial Intelligence (AI) based classifier that is trained unmonitored on the basis of the gross diagnostic results and that refines its diagnostic capability through supervised learning, the rating system including an AI based classifier based on supervised training learn the patterns and the results of coarse and fine diagnosis, and where the classification results of coarse and fine diagnosis and rating in addition to the associated time and speed signals, the feature vector and the operating data can be archived and used for off-line training of the classification systems. The method of claim 1, wherein the feature vector also contains operating data. A method according to claim 1 and claim 2, characterized characterized in that as an AI-based classifier for the fine diagnosis and / or rating Neural Networks. A method according to claim 1 and claim 2, characterized characterized in that as a Ki-based classifier for the fine diagnosis and / or rating methods of case-based closing. The method of claim 1 to 4, wherein the classification system only the rough diagnosis and the fine diagnosis supported. The method of claim 1 to 5, wherein the speed is constant and instead of order analysis a Fourier transform is used. The method of claim 1 to 6, wherein only the Time signal that archives speed signals and classification results become, The method of claim 1 to 6, wherein only the Feature vector and the classification results are archived.