CN111553811A - A method for identifying leakage areas in water supply network based on iterative machine learning - Google Patents

Abstract

A water supply network leakage area identification method based on iterative machine learning belongs to the technical field of water supply network leakage detection. For each iteration, one of the identified leakage regions is selected, k-means clustering is adopted to cluster the leakage regions into two types, all combination types of leakage nodes are used as labels of a random forest classifier model, then, leakage coefficients are added randomly to the nodes of the leakage regions according to the combination types of the leakage nodes so as to generate leakage samples, and the generated leakage samples are adopted to train the classifier model. The model takes the selection of the features into consideration in the training process so as to reduce the feature samples required in model training. Inputting the leakage characteristics into the trained classifier model so as to output the identified leakage region and the number of leakage nodes contained in the leakage region, and repeating the steps until the final identification accuracy is less than 95%, thus finishing the iteration. The method applies the single-label classifier to the identification of the multi-leakage-point area of the water supply network for the first time, and is simple to operate and good in identification effect.

Description

A method for identifying leakage areas in water supply network based on iterative machine learning

technical field

The invention relates to a method for identifying leakage areas of a water supply pipe network based on iterative machine learning, and belongs to the technical field of water supply pipe network leakage detection.

Background technique

The water supply network is an important infrastructure of society and plays an important role in economic development and normal life. Due to the aging and unreasonable design of the water supply pipe network, there are different degrees of leakage problems in the water supply pipe network. Developed countries lose about 15% of their purified water from their water distribution systems every year, while developing countries lose 35% or even up to 60% of their purified water every year. Therefore, accurate location of leakage is one of the key issues to be solved in the leakage rate control of water supply network.

With the advent of the era of intelligent big data, the use of data mining technology to locate the leakage of water supply network has gradually become a hot topic. Leakage research methods based on hydraulic models can be divided into two categories: (1) approximately determine the location of the leaking nodes based on the leakage characteristics of the similarity of adjacent nodes; (2) divide the water supply network into different areas according to the leakage characteristics in advance, to identify the leak area. Directly locating leaking nodes is to first generate leak samples of the water supply network. Since there are similar nodes when locating nodes, the recognition accuracy of the model will be reduced. Generally, the accuracy of the model will be improved by two methods: 1. One is to reduce the number of indistinguishable nodes by merging similar nodes in advance, and the other is to increase the delay mode, that is, to increase the capacity of leaking information by adding time nodes. The method of directly locating the leakage area is to use the clustering method to cluster similar nodes in advance, so as to reduce the number of similar nodes, and use the leakage area as the label of the classifier to identify the leakage area.

Through the analysis of previous research methods, it is found that although the nodes with similar leakage characteristics can be clustered, the recognition accuracy of the classifier can be improved. However, when the clustering method is applied to the water supply network, because the number of clusters is directly determined by the researchers, there is no theoretical basis, and only a single leakage node is considered in the process of localization and identification. However, in real water supply networks, multiple leaks often occur. Therefore, in order to solve this problem, the present invention adopts an iterative machine learning method combining k-means clustering (k=2) and random forest classifier, which provides a theoretical basis for the number of clusters in the clustering of the water supply network, and at the same time Provides a solution for identifying multiple leak areas. For simultaneous leaks, all leak node combination types in the identified regions were analyzed in each iteration. Each leaky combination type is used as a class label for the random forest classifier. The trained random forest classifier is used as the leak area identification model to locate leak areas and identify the number of leaks contained in each leak area.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide an iterative machine learning-based water supply pipe network leakage area identification method, so as to solve the uncertainty of the number of clusters when the clustering method is applied to the water supply pipe network and provide a solution for the area identification of multiple leaks.

The technical scheme adopted by the present invention is: establishing a hydraulic model of the water supply pipe network according to the principle of flow and pressure balance; assuming that there are l leakages at the same time, the leakage characteristic (the difference between the sensor values before and after the leakage) is ΔS _l . The same leakage coefficient C is added to each node of the area to generate a leakage change matrix. According to the leakage change matrix of the identified leakage area, k-means clustering is used to cluster them into two categories; then the leakage events are randomly generated through hydraulic model simulation. , and use the simulation result of the leakage event as the training sample of the classifier; use the sample to train the model of feature selection, and use the feature selection method-Mean Decrease Accuracy (MDA) to calculate the importance of the feature. Sort the features by importance to filter the features; use each leakage combination as the class label of the classifier model, and use the selected features to train the random forest classifier; if the final recognition accuracy of the model (training model ΔS per iteration) _l The product of the recognition accuracy, Acc) is greater than 95%, then the leakage feature ΔS _l is input into the trained random forest classifier, and then the next iteration is performed; if the final recognition accuracy is less than 95%, the iteration is stopped. Outputs the identified leak areas and the number of leak nodes each leak area contains.

The iterative machine learning-based method for identifying leakage areas in water supply network is carried out through the following steps:

(1) There are l nodes leaking at the same time, and the leakage characteristic is ΔS _l ;

每个泄漏区域内存在

个泄漏节点，i＝1,2,...,w；(2) The leakage area identified in the (β-1)th iteration is

exists within each leaked region

leaking nodes, i=1,2,...,w;

泄漏区域

内包含

个泄漏节点；对泄漏区域

内的每个节点添加相同的泄漏系数C从而生成泄漏变化矩阵

(3) Select one of the leakage areas from the w leakage areas

spill area

contains

leaking nodes;

The same leakage coefficient C is added to each node within to generate a leakage change matrix

采用k-means聚类将泄漏区域

聚类为两类，分别为区域

和区域

其余未聚类的(w-1)个区域及其包含的泄漏节点数目不变，则第β次迭代一共有(w+1)个区域；(4) According to the leakage change matrix

Use k-means clustering to classify leaky regions

Clustering is divided into two categories, namely regions

and area

The remaining unclustered (w-1) regions and the number of leaking nodes they contain remain unchanged, then the β-th iteration has a total of (w+1) regions;

个泄漏节点的区域

和区域

其所有的泄漏组合有

种,分别为：0个泄漏节点在

个泄漏节点在区域

1个泄漏节点在

个泄漏节点在区域

个泄漏节点在

0个泄漏节点在区域

因此对于本次迭代，一共有

个不同的标签；(5) Generate the combination type of leaked nodes in the βth iteration; for unclustered (w-1) regions, the number of leaked nodes inside the region remains unchanged.

region of leaking nodes

and area

All its leak combinations have

species, respectively: 0 leaking nodes are in

leaking nodes in the region

1 leaking node in

leaking nodes in the region

leaking nodes in

0 leaking nodes in the region

So for this iteration, there are a total of

different labels;

中选择

个节点，区域

中选择

个节点，对于未聚类的(w-1)个区域，分别从区域

中选择

个节点,

从而产生l个同时泄漏的节点，记为一个泄漏样本；一共产生ε个不同的泄漏样本；泄漏样本的集合称为T，特征的总数为N_PQ,N_P个压力传感器和N_Q个流量传感器，其中N_PQ＝N_P+N_Q；(6) Generate leaky samples of the βth iteration; randomly from the region

choose

nodes, regions

choose

nodes, for unclustered (w-1) regions, respectively from the region

choose

nodes,

As a result, l nodes that leak at the same time are recorded as a leak sample; a total of ε different leak samples are generated; the set of leak samples is called T, and the total number of features is N _PQ , N _P pressure sensors and N _Q flow sensors , where N _PQ =N _P +N _Q ;

(7) Take the hydraulic simulation result of the leakage event as the training sample of the classifier; use MDA to calculate the importance of the classifier model features, and then sort according to the importance of the features, so that the accuracy of the training model does not decrease, from The number of features is deleted in the order from unimportant to important; the calculation of the MDA is divided into the training of the random forest classifier and the calculation of the reduction of the average accuracy of the feature;

The training process of a random forest classifier is as follows:

识别出的泄漏区域

内存在

个泄漏节点，i＝1，2，...，w；定义

l是总的泄漏节点数目；然后根据泄漏区域

的泄漏矩阵

将泄漏区域

聚类为区域

和区域

两部分，泄漏区域组合为

对于包含

个泄漏节点的区域

和区域

其所有的泄漏组合有

种，其他未聚类的(w-1)个区域包含的泄漏节点数目保持不变，则本次迭代有

个随机森林分类器标签，泄漏节点的组合类型分别为0个泄漏节点在区域

个泄漏节点在区域

1个泄漏节点在区域

个泄漏节点在区域

个泄漏节点在区域

0个泄漏节点在区域

随机从区域

中选择

个节点，区域

中选择

个节点，区域

中选择

个节点，其中

从而产生l个同时泄漏的节点，记为一个泄漏样本；假设存在ε个不同的泄漏样本，泄漏样本的集合称为T，特征的总数为N_PQ,其中：N_P个压力传感器和N_Q个流量传感器，N_PQ＝N_P+N_Q)；(a) For the βth iteration, the combination of regions identified in the (β-1)th iteration is

Identified leak area

memory exists

leaky nodes, i = 1, 2, ..., w; definition

l is the total number of leaking nodes; then according to the leaking area

The leakage matrix of

will leak area

cluster into regions

and area

In two parts, the leakage area is combined as

for containing

region of leaking nodes

and area

All its leak combinations have

The number of leaking nodes contained in other unclustered (w-1) regions remains unchanged, then this iteration has

random forest classifier labels, the combination type of leaking nodes is 0 leaking nodes are in the region

leaking nodes in the region

1 leaky node in the region

leaking nodes in the region

leaking nodes in the region

0 leaking nodes in the region

random from area

choose

nodes, regions

choose

nodes, regions

choose

nodes, where

Thus, l simultaneously leaking nodes are generated, which are recorded as a leaking sample; assuming that there are ε different leaking samples, the set of leaking samples is called T, and the total number of features is N _PQ , among which: N _P pressure sensors and N _Q Flow sensor, N _PQ = N _P + N _Q );

(b) Create training subsets T _tr1 ,T _tr2 ,…,T _trM by repeated sampling from leaked samples, where M is the number of classification trees, randomly select training subsets of T _tri from leaked samples T, each The number of training subsets of the classification tree T _tri is ε, which is consistent with the size of the total leaked samples. Therefore, the training subset of the classification tree T _tri will have duplicate samples. The part of the leaked sample T that is not drawn is called out-of-bag (OOB) and is used to evaluate the accuracy of each tree; because the accuracy of the random forest classifier increases as the number of classification trees increases and converges is a constant, so choose the default value M=500;

用来计算每个特征的基尼指数；给定训练子集T_tri和连续特征N_D,D＝1,2,...,m,训练子集T_tri有f类样本，

类别h内有|f_h|样本；特征N_D有r个不同的值；然后，将这些值从小到大排序，并将它们标记为R＝{R¹,R²,…,R^r}，划分点t可以划分N_D为两个子集

和

其中

为包含不大于t的值的样本，

为包含大于t的值的样本，相邻的值分别为R^e和R^e+1；其中在[R^e，R^e+1]中的所有值都有相同的分割结果，所以有(r-1)个分割点是候选分割点；训练子集T_tri和特征点t处的基尼系数为(c) For each node of the training subset T _tri , _1≤i≤500 , randomly select m sub-features from NPQ features to create a classification tree, _1≤m≤NPQ ; the default value of m is

It is used to calculate the Gini index of each feature; given a training subset T _tri and continuous features N _D , D=1,2,...,m, the training subset T _tri has f class samples,

There are |f _h | samples in class h; feature N _D has r distinct values; then, sort these values from small to large and label them as R = {R ¹ ,R ² ,...,R ^r }, The dividing point t can divide _ND into two subsets

and

in

is a sample containing values not greater than t,

For the samples containing values greater than t, the adjacent values are ^Re and Re ⁺¹ respectively; in which all values in [ ^Re , Re ⁺¹ ] have the same segmentation result, so there is (r- 1) The split points are candidate split points; the Gini coefficient at the training subset T _tri and the feature point t is

表示随机选取的样本被误分类的概率；基尼指数越大，样本被错误分类的可能性越大；选择t的最小的Gini指数和相应的特征进行分割，然后，构造每个分支来重复上述过程；Where: In the formula, p _f represents the probability that the sample belongs to class f,

Represents the probability of a randomly selected sample being misclassified; the larger the Gini index, the greater the probability of the sample being misclassified; select the smallest Gini index of t and the corresponding feature for segmentation, and then construct each branch to repeat the above process ;

(d) Each classification tree contributes to the recognition accuracy of the training subset T _tri , and the average recognition accuracy of 500 classification trees is the recognition accuracy of the training model

The calculation process of MDA is as follows:

For each classification tree T _tri of the random forest classifier model, 1≤i≤500, use the OOB data to calculate the OOB error, denoted as OOB _b1 ; randomly scramble the out-of-bag sample data at the feature F, 1≤F≤N _PQ , and calculate the out-of-bag error again, denoted as OOB _b2 ; assuming that there are 500 trees in the forest, the importance of feature F is expressed as:

个泄漏节点的组合类型作为分类器模型的类别标签，利用所选特征训练随机森林分类器；如果模型第β次迭代的最终识别准确率Acc^β，即每次迭代的训练模型识别准确率的乘积

大于95％，则将泄漏特征输入训练好的随机森林分类器，然后进行下一次迭代；如果最终识别准确率小于95％，则停止迭代；输出识别的泄漏区域和每个泄漏区域包含的泄漏节点数目。(8) will

The combination type of the leaked nodes is used as the class label of the classifier model, and the random forest classifier is trained with the selected features; if the final recognition accuracy Acc ^β of the βth iteration of the model is the product of the recognition accuracy of the training model for each iteration

If it is greater than 95%, input the leak feature into the trained random forest classifier, and then proceed to the next iteration; if the final recognition accuracy rate is less than 95%, stop the iteration; output the identified leak area and the leak nodes contained in each leak area number.

The present invention has the following beneficial effects:

This iterative machine learning-based method for identifying leaking areas in water supply network, for each iteration, select one of the identified leaking areas, use k-means clustering to cluster them into two categories, and classify all the combined types of leaking nodes. As the label of the random forest classifier model, then according to the combination type of the leaked nodes, the leak coefficient is randomly added to the nodes in the leak area to generate leak samples, and the generated leak samples are used to train the classifier model. In the process of training, the model considers the selection of features to reduce the feature samples required for model training. Input the leak feature into the trained classifier model to output the identified leak area and the number of leak nodes contained therein, and repeat the above steps until the final identification accuracy rate is less than 95%, and the iteration ends. The number of leaked regions identified and the leaky nodes they contain is thus output. When a single node leaks, the binary clustering solves the uncertainty of the number of clusters when the clustering method is applied to the leak detection of the water supply network, provides a theoretical basis for the number of clusters, and reduces the required trial calculation The number of times; the technology of combining clustering and single-label classifier is applied to the identification and positioning of the leaking area with multiple leaking points, and the leaking area and the number of leaking nodes it contains can be identified.

Description of drawings

FIG. 1 is a flow chart of the method of the present invention.

Figure 2 is a hydraulic model diagram of the water supply network.

Figure 3 shows the location of the pressure and flow measurement points.

Figure 4 shows the training process of the random forest classifier.

FIG. 5 is a process of identifying the leaking area containing leaky node 101 .

Figure 6 is the identification of the leakage area of a single node: (a) the leakage area determined by the traditional method; (b) the leakage area determined by the iterative machine learning method.

Figure 7 Identification process of the region containing leaky nodes 65 and 281

Figure 8 Identification process for the region containing leaky nodes 132 and 406

Figure 9 is the identification of two simultaneous leakage regions: (a) the leakage region containing nodes 65 and 281; (b) the leakage region containing nodes 132 and 406.

Detailed ways

The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention.

As shown in Figure 1, the hydraulic model of the water supply pipe network is established according to the principle of flow and pressure balance; assuming that there are l nodes leaking at the same time, the leakage characteristic (the difference between the sensor values before and after the leak) is ΔS _l . The same leakage coefficient C is added to each node of the area to generate a leakage change matrix. According to the leakage change matrix of the identified leakage area, k-means clustering is used to cluster them into two categories; then the leakage events are randomly generated through hydraulic model simulation. , and use the simulation result of the leakage event as the training sample of the classifier; use the sample to train the model of feature selection, sort the features according to the importance of the features to filter the features; use each leakage combination as the class label of the classifier model , using the selected features to train the random forest classifier; if the final recognition accuracy of the model (the product of the recognition accuracy of the training model at each iteration, Acc) is greater than 95%, then the leakage feature ΔS _l is input into the trained random forest classifier , output the leak area and the number of leak nodes contained in each leak area, and then proceed to the next iteration; if the final recognition accuracy rate is less than 95%, stop the iteration.

Example

The first step is to establish the hydraulic model of the water supply pipe network according to the principle of flow and pressure balance. The present invention adopts EPANET to establish the hydraulic model. As shown in Figure 2, it consists of a reservoir (1), a reservoir (2), 375 nodes (3) and 469 pipe sections (4). As shown in Figure 3, the measuring points are composed of 3 flowmeters (Q1-Q3) and 18 pressure measuring points (P1-P18). Assuming that the sensor measurement value is destroyed by a uniform zero-mean Gaussian error, its amplitudes are 1.5% of the mean of the residuals of the sensor samples. The basic water demand is 148L/s. The maximum water demand is 162.8L/s, and the minimum water demand is 118.4L/s. Since the leakage amount is too small, the sensor will not change, while the leakage amount is too large for residents to find out first, so the present invention selects the leakage coefficient between 0.5-1.2.

In the second step, the present invention assumes that there are two types of leakage in the water supply pipe network, one is a single leakage node, and the other is a leakage of two nodes.

The third step, when a leak alarm occurs in the water supply network, the initial iteration starts from the entire water supply network;

The fourth step is to perform cluster division of the leaked area, selection of features and training of the multi-level random forest classifier model. The random forest classification training model is shown in FIG. 4 .

single node leak

When a 4L/s leak occurs at node 101, only one node in the leak area identified in the previous iteration is selected each time to generate leak samples.

One of the identified leak areas is selected, and a leak coefficient of C=0.8 is randomly added to the nodes in the leak area to generate a leak change matrix. K-means was used to cluster the leaked regions into two parts.

In each iteration, 30 leak samples are randomly generated within the leak coefficient range for each node of the leak region identified in the last iteration. Therefore, in the first iteration, the total number of nodes is 375, and there are 11250 independent leak samples. As the number of iterations increases, the area of candidate leaky regions decreases, and the number of leaky samples used for model training decreases.

With feature selection, you can reduce the number of features needed to identify leaks at each iteration. For example, iteration 2 in FIG. 5 requires a maximum of 9 sensors, and the entire iteration requires 10 different sensors, and the number of sensors is smaller than the total number of sensors.

When a 4L/s leak occurs at node 101, the iterative process is shown in Figure 5. The recognition accuracy of each sub-iteration is 99.99%, 99.82%, 99.25%, and 97.35%, respectively. The final recognition of the fourth iteration is accurate The rate is 99.99%×99.82%×99.25%×97.35%=96.44%. For the fifth sub-iteration, the accuracy is 96.67%, and the final recognition accuracy is 96.44% × 96.67% = 93.22% < 95%, then the iteration is terminated; as shown in Fig. 6(b), determine the leaking node 101(5) in the zone (Z1). The total number of iterations is 4.

Compared with the traditional method (the total number of clusters is given in advance, and the final number of clusters is determined by the trial algorithm), the number of experimental calculations of the traditional method is obviously more than that of the method in this paper, and the present invention has a more specific analysis of the leakage area. According to the traditional method, if the recognition accuracy rate is 95% as the training model standard, the total number of identifiable regions is 18. If the traditional method starts from 1 and performs the clustering trial calculation in increments of 1 area at a time, the total number of calculations is 17. But for iterative method. It can be seen from Table 2 that the number of iterations is 4, and the number of iterations is less than the number of trials of the traditional algorithm. Although the traditional method also has a good effect on the identification of the leakage area, the identified leakage area (Z2) containing the leakage node 101(5) (as shown in FIG. ) is larger. In contrast, the method proposed in this paper not only reduces the total number of trial computations for the iterative method, but also reduces the area of identifiable regions by eliminating leak-free regions in each iteration.

Two nodes leak at the same time

A two-node leak combination occurs at nodes 65 and 281, the leak flow at node 65 is 3.6L/s, and the leak flow at node 281 is 2.9L/s; another leak combination occurs at nodes 132 and 406, and the leak at node 132 The flow is 3.0 L/s and the leakage flow at node 406 is 4.7 L/s.

For each iteration, one of the identified leakage regions is selected, and a leakage coefficient of C=0.8 is randomly added to the nodes in the leakage region to generate a leakage change matrix. K-means clustering was used to cluster the leaked regions into two parts.

Each leak node combination yields 30 leak factor combinations within the leak factor range. Likewise, the first iteration randomly generated 11,250 independent leak samples. The sample size at each iteration decreases gradually as the number of nodes in the leaky region increases. For example, if the leak area identified in the last iteration had 375 nodes, then 11,250 leak samples would be generated, and a proportional algorithm, if the leak area identified in the last iteration had 50 nodes, would yield 1,500 leak samples leaked samples. The specific identification process of the iterative method for simultaneous leakage of two nodes is shown in FIG. 7 and FIG. 8 . Figures 7 and 8 show different recognition results.

For simultaneous leakage of two nodes, feature selection can also reduce unnecessary features when training two models with simultaneous leakage in each iteration, but as the number of iterations increases, the effect relative to single-node leakage decreases . As shown in Figure 7, iteration 1 requires 5 sensors and iteration 4 requires 13 sensors, as shown in Figure 8, iteration 1 requires 5 sensors, and iteration 4 requires 12 sensors. For the entire iteration, Figure 7 shows that 13 different sensors are required for the entire iteration, while Figure 8 shows that 17 different sensors are required for the entire iteration. The total number of sensors required for the combination of closely spaced adjacent leaks is less than the total number of sensors required for the combination with the more distant leaks. Although the number of selected sensors is all smaller than the total number of sensors, the effect is reduced compared to single node leakage.

As shown in Figure 7, when the nodes 65 and 281 leak, the recognition accuracy rates of the first four sub-iterations are 99.41%, 98.83%, 98.57%, and 99.76%, respectively, and the final recognition accuracy is 99.41%× 98.83%×98.57%×99.76%=96.61%. For the fifth sub-iteration, the accuracy was 97.22%, the final recognition accuracy was 96.61% × 97.22% = 93.92% < 95%, the iteration was stopped; the leak area and the number of leaks within the leak area were determined.

As shown in Figure 8, when the nodes 132(9) and 406(8) leak, the accuracies of the first seven sub-iterations are 99.41%, 99.52%, 99.64%, 99.78%, 98.88%, 99.36%, 99.67%, and the final recognition accuracy is 96.32%. The recognition accuracy rate of the 8th iteration is 97.12%, then the final recognition accuracy rate is 96.32%×97.12%=93.55%<95%, the iteration stops; the leak area and the number of leaks in each leak area are determined.

As shown in FIG. 9(a), when leakage occurs at nodes 65(6) and 281(7), the leakage area identified by the present invention is Z3. As shown in FIG. 9(b), when node 132(9) ) and 406(8) when leakage occurs, the leakage areas identified by the present invention are Z4 and Z5. For nodes 65(6) and 281(7), it is difficult to distinguish between the two simultaneously leaking nodes due to their close distance. And for nodes 132(9) and 406(8), the leakage characteristics of nodes 132(9) and 406(8) are more obvious and more obvious than nodes 65(6) and 281(7) due to the greater distance between them. It is easy to distinguish, so nodes 132(9) and 406(8) have more iterations.

The invention proposes a method for identifying the leakage area of a water supply pipe network based on iterative machine learning. In each iteration, k-means is used to cluster one of the identified leak areas into two parts, and then the combination type of all leak nodes in this iteration is determined, and the combination type of each leak node is used as a random forest classifier A class label of the training model is trained, and then leak coefficients are randomly added to the nodes in the leak area according to the combination type of leak nodes to generate leak samples, and the generated leak samples are used to train the classifier model. If the trained model satisfies the iterative criteria, input the leaky features for the next iteration, if not, the iteration ends, and output the leaky area and the number of leaky nodes contained in each leaky area. The method is applied to a water supply network and its performance is evaluated. The results show that the method can better identify the leaking area and the number of leaking nodes it contains, and improve the efficiency and accuracy of leak detection.

Claims (1)

1. A method for identifying leakage areas of water supply network based on iterative machine learning, which is characterized by: The method proceeds through the following steps: (1) There are l nodes leaking at the same time, and the leakage characteristic is ΔS _l ; (2) The leakage area identified in the (β-1)th iteration is

exists within each leaked region

leaky nodes, i=1,2,...,w; (3) Select one of the leakage areas from the w leakage areas

spill area

contains

leaking nodes;

The same leakage coefficient C is added to each node within to generate a leakage change matrix

(4) According to the leakage change matrix

Use k-means clustering to classify leaky regions

Clustering is divided into two categories, namely regions

and area

The remaining unclustered (w-1) regions and the number of leaking nodes they contain remain unchanged, then the β-th iteration has a total of (w+1) regions; (5) Generate the combination type of leaked nodes in the βth iteration; for unclustered (w-1) regions, the number of leaked nodes inside the region remains unchanged.

region of leaking nodes

and area

All its leak combinations have

species, respectively: 0 leaking nodes are in

leaking nodes in the region

1 leaking node in

leaking nodes in the region

leaking nodes in

0 leaking nodes in the region

So for this iteration, there are a total of

different labels; (6) Generate leaky samples of the βth iteration; randomly from the region

choose

nodes, regions

choose

nodes, for unclustered (w-1) regions, respectively from the region

choose

nodes,

As a result, l nodes that leak at the same time are recorded as a leak sample; a total of ε different leak samples are generated; the set of leak samples is called T, and the total number of features is N _PQ , N _P pressure sensor and N _Q flow sensor, where N _PQ =N _P +N _Q ; (7) Take the hydraulic simulation result of the leakage event as the training sample of the classifier; use MDA to calculate the importance of the classifier model features, and then sort according to the importance of the features, so that the accuracy of the training model does not decrease, from The number of features is deleted in the order from unimportant to important; the calculation of the MDA is divided into the training of the random forest classifier and the calculation of the reduction of the average accuracy of the feature; The training process of a random forest classifier is as follows: (a) The leak area identified in the (β-1)th iteration is

exists within each leaked region

leaky nodes, i=1,2,...,w; definition

l is the total number of leaking nodes; then according to the leaking area

The leakage matrix of

will leak area

cluster into regions

and area

In two parts, the leakage area is combined as

for containing

region of leaking nodes

and area

All its leak combinations have

The number of leaking nodes contained in other unclustered (w-1) regions remains unchanged, then this iteration has

random forest classifier labels, the combination type of leaking nodes is 0 leaking nodes are in the region

leaking nodes in the region

1 leaking node in

leaking nodes in the region

leaking nodes in the region

0 leaking nodes in the region

random from area

choose

nodes, regions

choose

nodes, regions

choose

nodes, where

Thus, l simultaneously leaking nodes are generated, which are recorded as a leaking sample; assuming that there are ε different leaking samples, the set of leaking samples is called T, and the total number of features is N _PQ , among which: N _P pressure sensors and N _Q Flow sensor, N _PQ =N _P +N _Q ; (b) Create training subsets T _tr1 ,T _tr2 ,…,T _trM by repeated sampling from leaked samples, where M is the number of classification trees, randomly select training subsets of T _tri from leaked samples T, each The number of training subsets of the classification tree T _tri is ε, which is consistent with the size of the total leaked samples. Therefore, the training subset of the classification tree T _tri will have duplicate samples. The part of the leaked sample T that is not drawn is called OOB, which is used to evaluate the accuracy of each tree; since the accuracy of the random forest classifier increases with the number of classification trees and tends to a constant, the default valueM=500; (c) For each node of the training subset T _tri , _1≤i≤500 , randomly select m sub-features from NPQ features to create a classification tree, _1≤m≤NPQ ; the default value of m is

It is used to calculate the Gini index of each feature; given a training subset T _tri and continuous features N _D , D=1,2,...,m, the training subset T _tri has f class samples,

There are |f _h | samples in class h; feature N _D has r distinct values; then, sort these values from small to large and label them as R = {R ¹ ,R ² ,...,R ^r }, The dividing point t can divide _ND into two subsets

and

in

is a sample containing values not greater than t,

For the samples containing values greater than t, the adjacent values are ^Re and Re ⁺¹ respectively; in which all values in [ ^Re , Re ⁺¹ ] have the same segmentation result, so there is (r- 1) The split points are candidate split points; the Gini coefficient at the training subset T _tri and the feature point t is

Where: In the formula, p _f represents the probability that the sample belongs to class f,

Represents the probability of a randomly selected sample being misclassified; the larger the Gini index, the greater the probability of the sample being misclassified; select the smallest Gini index of t and the corresponding feature for segmentation, and then construct each branch to repeat the above process ; (d) Each classification tree contributes to the recognition accuracy of the training subset T _tri , so the average recognition accuracy of 500 classification trees is the recognition accuracy of the trained model

The calculation process of MDA is as follows: For each classification tree T _tri of the random forest classifier model, 1≤i≤500, use the OOB data to calculate the OOB error, denoted as OOB _b1 ; randomly scramble the out-of-bag sample data at the feature F, 1≤F≤N _PQ , and calculate the out-of-bag error again, denoted as OOB _b2 ; assuming that there are 500 trees in the forest, the importance of feature F is expressed as:

(8) will

The combination type of the leaked nodes is used as the class label of the classifier model, and the random forest classifier is trained with the selected features; if the final recognition accuracy Acc ^β of the βth iteration of the model is the product of the recognition accuracy of the training model for each iteration

If it is greater than 95%, input the leak feature into the trained random forest classifier, and then proceed to the next iteration; if the final recognition accuracy rate is less than 95%, stop the iteration, and output the identified leak area and the leak nodes contained in each leak area. number.

Images