Elevator fault diagnosis method based on convolutional neural network

Abstract

The invention discloses an elevator fault diagnosis method based on a convolutional neural network, and the method comprises the steps: constructing an active learning framework based on a query strategy, and enabling the active learning framework to comprise a data management module, a query strategy module and a convolutional neural network model; training an initial convolutional neural network by using a small amount of labeled data; obtaining the prediction probability of the model to all the label-free data sets, and calculating the gradient embedding vector of the model; using a K-means + + algorithm to further screen out a label-free data set with high contribution degree and diversity, performing manual labeling, using the label-free data set with high contribution degree and diversity to re-train the model, and circulating the process until the label-free data set is used up or the model reaches specified test accuracy; according to the method, information implied in unlabeled samples is mined from the aspects of a network model framework, a loss function, a query strategy and the like, importance sorting is performed on the samples which are most difficult to understand in the model, experts are enabled to concentrate on providing the most useful information, and the training cost of model construction is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of elevator fault diagnosis, and in particular to an elevator fault diagnosis method based on a convolutional neural network. Background Art

[0002] As one of the important components in the mechanical system, bearings are responsible for supporting the movement of rotating bodies. Their operating status has a key impact on the performance and reliability of mechanical equipment, especially special equipment such as elevator traction machines. However, due to the harsh working environment, fatigue wear and lack of timely maintenance, rolling bearings often fail and are damaged, resulting in equipment downtime and safety accidents. Therefore, the diagnosis of special equipment bearings is of great research significance. With the continuous advancement of science and technology, many advanced fault diagnosis methods have been applied to bearings. However, traditional deep learning processes often require the construction of large-scale labeled data sets. In actual situations, although it is easy to obtain a large amount of unlabeled data, there are problems such as complex and changeable fault types and difficulty in locating fault signals, resulting in a lack of a large amount of labeled data. Labeling a large amount of data is a time-consuming, laborious and expensive task that requires analysis and judgment by professional engineers or technicians.

[0003] In order to solve the problem of lack of labeled data, active learning can be combined with deep learning to alleviate the pressure of labeling. Active learning focuses on how to discover the implicit information in unlabeled samples from the aspects of network model framework, loss function and query strategy, sort the importance of the samples that are most difficult for the model to understand, skip the samples that are not very helpful to the model, reduce data redundancy and smooth the decision boundary, so that experts can focus on providing the most useful information to reduce the training cost required to build the model. Summary of the invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and provide an elevator fault diagnosis method based on convolutional neural network.

[0005] The objective of the present invention is achieved through the following technical solutions: The present invention provides an elevator fault diagnosis method based on a convolutional neural network, comprising the following steps: Step 1: Build an active learning framework based on query strategy, including data management module, query strategy module and convolutional neural network model; Step 2: Obtain bearing vibration signals under different fault types and different working conditions through an acceleration sensor, and intercept the bearing vibration signals without overlapping through a sliding window to obtain a first bearing vibration signal, classify the first bearing vibration signals according to the health status of the bearing, and randomly divide all the first bearing vibration signals after classification into a first training data set and a first measured data set with the same proportion; Step 3: preprocess the first training data set and the first measured data set through a data management module to obtain a second training data set and a second measured data set, and then use Boolean indexing to divide the second training data set into a labeled data set and an unlabeled data set; Step 4: Obtain a labeled data set from the data management module, input the labeled data set and its corresponding one-hot code label into the convolutional neural network model for the first training of the convolutional neural network model, first perform forward propagation to obtain the training error, and then update the weight of the convolutional neural network model through back propagation and optimizer; Step 5: If the number of first trainings reaches the preset number, a second measured data set is obtained from the data management module, and the second measured data set is used to perform a second training on the convolutional neural network model. The convolutional neural network model is in a test mode during the second training, and a training status report of the convolutional neural network model is obtained; if the number of second trainings does not reach the preset number, the labeled data set and its corresponding unique hot code label are input again into the convolutional neural network model for a third training; the second training includes forward propagation to obtain the prediction probability, and then the prediction probability is input into the data management module; when performing the second training, the convolutional neural network model selects the data set to be learned through the query strategy module; Step 6: Obtain an unlabeled data set from the data management module, forward propagate the unlabeled data set in the test mode of the convolutional neural network model to obtain the prediction probability and the low-dimensional embedding vector, and calculate the gradient embedding vector of the unlabeled data set; Step 7: The gradient embedding vectors are screened by the K-means++ clustering algorithm to screen out diversified high-contribution samples to be labeled, the samples to be labeled are manually labeled, and the manually labeled samples are input into the data management module and marked as labeled data sets; Step 8: Repeat steps 4 to 7 until the accuracy of the convolutional neural network model meets the preset conditions or all samples are labeled; Step 9: After the accuracy of the convolutional neural network model meets the preset conditions or all samples are marked, load the corresponding parameters, input the parameters into the first measured data set, and output the final fault diagnosis result.

[0006] Furthermore, the data management module includes a data preprocessing function, a partitioning strategy and a model training status report; the data preprocessing function is used for data normalization and Gaussian noise data enhancement; the partitioning strategy includes a random initialization strategy and an equal initialization strategy, the random initialization strategy is used to randomly assign the proportion of each fault type in the labeled data set, and the equal initialization strategy is used to use the same partitioning ratio in the labeled data set; the model training status report includes test accuracy, various proportions in labeled samples, confusion matrix and TSNE high-dimensional data dimensionality reduction visualization.

[0007] Furthermore, the convolutional neural network model includes five convolutional layers, five maximum pooling layers and one fully connected layer; the input of the convolutional neural network model is any data set, and the output is the probability value of any data set, and the sum of the probability values ​​is 1.

[0008] Furthermore, the gradient embedding vector includes a low-dimensional embedding vector and a prediction probability of the convolutional neural network model: Assuming that there are m rows and n columns of gradient embedding vectors, the low-dimensional embedding vector The calculation formula is: ;in It refers to the low-dimensional embedding vector output by the last convolutional layer when the number of rows of the gradient embedding vector is m. Refers to the gradient weight corresponding to the gradient embedding vector when the number of columns is n and the number of rows is m; The calculation method is: Assuming there are i rows and j columns of gradient embedding vectors, where 1≤j≤n, 1≤i≤m, the calculation formula is: ;in is the final output of the convolutional neural network model when the number of columns of the gradient embedding vector is j, Refers to the gradient weight corresponding to the gradient embedding vector when the number of columns is i and the number of rows is j.

[0009] Furthermore, the method for screening out diverse high-contribution samples to be labeled comprises the following steps: Step S11, input the gradient embedding vector of the diversified evaluation index of the unlabeled sample into the K-means++ clustering algorithm; Step S12: taking the gradient embedding vector with the largest norm as the first cluster center, and then calculating the distances between other vectors and the cluster center; Step S13: According to the weight Select the next cluster center until the number of cluster centers reaches the number of query samples specified by the query strategy module, and the weight The calculation method is: ; where x represents other samples, Represents the distance between other samples and their nearest cluster center, and U is all unlabeled datasets.

[0010] The beneficial effects of the present invention are: 1) Based on the existing cutting-edge achievements of convolutional neural networks, this invention selects the most representative samples of different fault types under all samples through unsupervised clustering methods, builds the initial model, uses gradient embedding vector as the standard for evaluating sample uncertainty, and uses K-means++ clustering method as the standard for evaluating sample differences. The most important samples are selected for manual labeling, and the model is continuously iterated and trained until the diagnostic accuracy requirements are met.

[0011] 2) The present invention explores the implicit information in unlabeled samples from the aspects of network model framework, loss function and query strategy, and ranks the importance of samples that are most difficult for the model to understand, allowing experts to focus on providing the most useful information to reduce the training cost of building the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 Flow chart of a method for diagnosing elevator faults based on a convolutional neural network according to an embodiment of the present invention; Figure 2 It is a schematic diagram of the accuracy of different methods under different query times according to an embodiment of the present invention; Figure 3 A schematic diagram of the number of queries required for different methods of the embodiments of the present invention to reach an accuracy of 80%, 90%, 95% and 100%; Figure 4 A schematic diagram of sample ratios of random sampling under different query times according to an embodiment of the present invention; Figure 5 A schematic diagram of sample ratios of K-means cluster sampling under different query times in an embodiment of the present invention; Figure 6 A schematic diagram of sample ratios of the lowest confidence sampling under different query times according to an embodiment of the present invention; Figure 7 A schematic diagram of sample ratios of BAGE sampling under different query times according to an embodiment of the present invention; Figure 8 is a sample ratio of an elevator fault diagnosis method based on a convolutional neural network under different query times according to an embodiment of the present invention; Fig. 9 This is a TSNE dimensionality reduction graph of the test data after 10 queries under random sampling in an embodiment of the present invention. DETAILED DESCRIPTION

[0013] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0014] Exemplarily, the flowchart of the present invention is as follows Figure 1 As shown, a method for elevator fault diagnosis based on convolutional neural network includes the following steps: Step 1: Build an active learning framework based on query strategy, including data management module, query strategy module and convolutional neural network model; Step 2: Obtain bearing vibration signals under different fault types and different working conditions through an acceleration sensor, and intercept the bearing vibration signals without overlapping through a sliding window to obtain a first bearing vibration signal, classify the first bearing vibration signals according to the health status of the bearing, and randomly divide all the first bearing vibration signals after classification into a first training data set and a first measured data set with the same proportion; Step 3: preprocessing the first training data set and the first measured data set through a data management module, wherein the preprocessing includes normalizing and enhancing the data set through a data preprocessing function to obtain a second training data set and a second measured data set, and then using a Boolean index to divide the second training data set into a labeled data set and an unlabeled data set; Step 4: Obtain a labeled data set from the data management module, input the labeled data set and its corresponding one-hot code label into the convolutional neural network model for the first training of the convolutional neural network model, first perform forward propagation to obtain the training error, and then update the weight of the convolutional neural network model through back propagation and optimizer; Step 5: If the number of first trainings reaches the preset number, a second measured data set is obtained from the data management module, and the second measured data set is used to perform a second training on the convolutional neural network model. The convolutional neural network model is in a test mode during the second training, and a training status report of the convolutional neural network model is obtained; if the number of second trainings does not reach the preset number, the labeled data set and its corresponding unique hot code label are input again into the convolutional neural network model for a third training; the second training includes forward propagation to obtain the prediction probability, and then the prediction probability is input into the data management module; when performing the second training, the convolutional neural network model selects the data set to be learned through the query strategy module; Step 6: Obtain an unlabeled data set from the data management module, forward propagate the unlabeled data set in the test mode of the convolutional neural network model to obtain the prediction probability and the low-dimensional embedding vector, and calculate the gradient embedding vector of the unlabeled data set; Step 7: The gradient embedding vectors are screened by the K-means++ clustering algorithm to screen out diversified high-contribution samples to be labeled, the samples to be labeled are manually labeled, and the manually labeled samples are input into the data management module and marked as labeled data sets; Step 8: Repeat steps 4 to 7 until the accuracy of the convolutional neural network model meets the preset conditions or all samples are labeled; Step 9: After the accuracy of the convolutional neural network model meets the preset conditions or all samples are marked, load the corresponding parameters, input the parameters into the first measured data set, and output the final fault diagnosis result.

[0015] Specifically, the data management module includes a data preprocessing function, a partitioning strategy and a model training status report; the data preprocessing function is used for data normalization and Gaussian noise data enhancement; the partitioning strategy includes a random initialization strategy and an equal initialization strategy, the random initialization strategy is used to randomly assign the proportion of each fault type in the labeled data set, and the equal initialization strategy is used to use the same partitioning ratio in the labeled data set; the model training status report includes test accuracy, the proportion of each type in the labeled samples, the confusion matrix and TSNE high-dimensional data dimensionality reduction visualization.

[0016] Specifically, the convolutional neural network model includes five convolutional layers, five maximum pooling layers and one fully connected layer; the input of the convolutional neural network model is any data set, and the output is the probability value of any data set, and the sum of the probability values ​​is 1.

[0017] Specifically, the gradient embedding vector includes a low-dimensional embedding vector and a prediction probability of the convolutional neural network model: Assuming there are m rows and n columns of gradient embedding vectors, the low-dimensional embedding vector The calculation formula is: ;in It refers to the low-dimensional embedding vector output by the last convolutional layer when the number of rows of the gradient embedding vector is m. Refers to the gradient weight corresponding to the gradient embedding vector when the number of columns is n and the number of rows is m; The calculation method is: Assuming there are i rows and j columns of gradient embedding vectors, where 1≤j≤n, 1≤i≤m, the calculation formula is: ;in is the final output of the convolutional neural network model when the number of columns of the gradient embedding vector is j, Refers to the gradient weight corresponding to the gradient embedding vector when the number of columns is i and the number of rows is j.

[0018] Specifically, the method for screening out diverse high-contribution samples to be labeled includes the following steps: Step S11, input the gradient embedding vector of the diversified evaluation index of the unlabeled sample into the K-means++ clustering algorithm; Step S12: taking the gradient embedding vector with the largest norm as the first cluster center, and then calculating the distances between other vectors and the cluster center; Step S13: According to the weight Select the next cluster center until the number of cluster centers reaches the number of query samples specified by the query strategy module, and the weight The calculation method is: ; where x represents other samples, Represents the distance between other samples and their nearest cluster center, and U is all unlabeled datasets.

[0019] Exemplarily, when this method is used, data acquisition is first performed: an acceleration sensor is arranged on a rolling bearing test bench, and experiments are performed under motor loads of 0 to 3 horsepower and fault diameters of 7, 14 and 21 mils, respectively. The vibration signal during operation is collected at the drive end, and the bearing health status is divided into ten categories, as shown in Table 1 for details.

[0020] Table 1 Bearing data details

[0021] After that, data preprocessing and slicing are performed: the collected samples are divided into data sets through sliding window sampling, the sliding windows have no overlap, and the length of each sample is 2048. There are 200 samples in each of the ten different fault types, including normal ones, of which 10 are labeled samples, 140 are unlabeled samples, and 50 are used as test sets.

[0022] Then the model is constructed: the model is constructed based on the Pytorch deep learning framework. The specific parameters of the model network in this embodiment are shown in Table 2.

[0023] Table 2 Specific parameter settings of the network structure of the method of the embodiment of the present invention

[0024] Then the model training is carried out: the network model uses Adam optimizer to update parameters, the initial learning rate is set to 0.001, the batch size is set to 512, and the training is performed for 100 rounds. 10 samples are queried each time. After the query, the model is retrained until the test accuracy or the number of labeled samples reaches the specified value.

[0025] Then the model is tested: the convolutional neural network classifier model parameters are loaded, the test set in the dataset is input into the trained model, and the fault identification results are output.

[0026] Finally, the beneficial effects are analyzed: In order to verify the effectiveness of the present invention, the embodiment selects four query strategies, namely, random sampling (RS), least confidence (LC), K-means sampling (KS), and BAGE (Batch Active learning by Gradient Embeddings) using only gradient embedding vectors, as comparative methods for experiments, and each method is repeated 10 times. RS is randomly selected from unlabeled samples, LC is to select the sample with the lowest prediction probability in the unlabeled samples, KS is to perform unsupervised clustering in the unlabeled samples to select the most diverse samples, and BAGE is to use only gradient embedding vectors without further performing K-means++ clustering ablation experiments. The experimental comparison results are shown in Table 3.

[0027] Table 3 Diagnostic accuracy of each method under different query times

[0028] The accuracy diagram of different methods under different query times is as follows Figure 2 As shown in the figure, it can be seen that under the same number of queries, the BADGE method has a higher accuracy than other query strategies, and is only slightly lower than the K-means clustering sampling method when there is a lack of samples in the early stage. The number of queries required for different methods to reach 80%, 90%, 95% and 100% accuracy is shown in the figure. Figure 3 As shown; it can be seen that compared with other methods, the method in this paper can reach the specified accuracy faster and greatly reduce the number of samples required to build the fault diagnosis model.

[0029] In order to further analyze the selection characteristics of different query strategies, the present invention plots the proportion of samples of different categories in the training set under different query strategies at different query times, such as Figures 4 to 8 As shown, B007 refers to the fault type of a slight fault of the rolling element (7mil); B014 refers to the fault type of a moderate fault of the rolling element (14mil); B021 refers to the fault type of a heavy fault of the rolling element (21mil); IR007 refers to the fault type of a slight fault of the inner ring (7mil); IR014 refers to the fault type of a slight fault of the inner ring (14mil); IR021 refers to the fault type of a slight fault of the inner ring (21mil); OR007@6 refers to the fault type of a slight fault of the center outer ring (7mil); OR014@6 refers to the fault type of a slight fault of the center outer ring (14mil); OR021@6 refers to the fault type of a slight fault of the center outer ring (21mil); normal refers to the normal state; it can be seen that the traditional random sampling method Figure 4 and K-means sampling method Figure 5 , the sample ratio will eventually tend to be balanced. Although sample balance is generally more beneficial for model building, it is not actually suitable for fault diagnosis environments. The TSNE dimensionality reduction results of the test set after 10 queries under random sampling are as follows: Fig. 9 As shown, Fig. 9 The horizontal axis is TSNE dimension 1, and the vertical axis is TSNE dimension 2. It can be seen that aliasing occurs in some areas, which indicates that the difficulty of distinguishing different fault types and the data quality are also different. For the data set used in this experiment, it is difficult to distinguish between moderate faults of rolling elements, heavy faults of rolling elements, and slight faults of inner rings. The schematic diagram of the minimum confidence method is shown in Figure 6 The schematic diagram of BAGE sampling is shown in Figure 7 The schematic diagram of this method is shown in Figure 8 As shown in the figure, this feature is captured and the sample ratio change of this method is more stable.

[0030] The above is only a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art shall not deviate from the spirit and scope of the present invention, and shall be within the scope of protection of the claims attached to the present invention.

Claims

1. A method for elevator fault diagnosis based on convolutional neural network, characterized in that: The following steps are involved: Step 1: Build an active learning framework based on query strategy, including data management module, query strategy module and convolutional neural network model; Step 2: Obtain bearing vibration signals under different fault types and different working conditions through an acceleration sensor, and intercept the bearing vibration signals without overlapping through a sliding window to obtain a first bearing vibration signal, classify the first bearing vibration signals according to the health status of the bearing, and randomly divide all the first bearing vibration signals after classification into a first training data set and a first measured data set with the same proportion; Step 3: preprocess the first training data set and the first measured data set through a data management module to obtain a second training data set and a second measured data set, and then use Boolean indexing to divide the second training data set into a labeled data set and an unlabeled data set; Step 4: Obtain a labeled data set from the data management module, input the labeled data set and its corresponding one-hot code label into the convolutional neural network model for the first training of the convolutional neural network model, first perform forward propagation to obtain the training error, and then update the weight of the convolutional neural network model through back propagation and optimizer; Step 5: If the number of first trainings reaches the preset number, a second measured data set is obtained from the data management module, and the second measured data set is used to perform a second training on the convolutional neural network model. The convolutional neural network model is in a test mode during the second training, and a training status report of the convolutional neural network model is obtained; if the number of second trainings does not reach the preset number, the labeled data set and its corresponding unique hot code label are input again into the convolutional neural network model for a third training; the second training includes forward propagation to obtain the prediction probability, and then the prediction probability is input into the data management module; when performing the second training, the convolutional neural network model selects the data set to be learned through the query strategy module; Step 6: Obtain an unlabeled data set from the data management module, forward propagate the unlabeled data set in the test mode of the convolutional neural network model to obtain the prediction probability and the low-dimensional embedding vector, and calculate the gradient embedding vector of the unlabeled data set; Step 7: The gradient embedding vectors are screened by the K-means++ clustering algorithm to screen out diversified high-contribution samples to be labeled, the samples to be labeled are manually labeled, and the manually labeled samples are input into the data management module and marked as labeled data sets; Step 8: Repeat steps 4 to 7 until the accuracy of the convolutional neural network model meets the preset conditions or all samples are labeled; Step 9: After the accuracy of the convolutional neural network model meets the preset conditions or all samples are marked, load the corresponding parameters, input the parameters into the first measured data set, and output the final fault diagnosis result.

2. The elevator fault diagnosis method based on convolutional neural network according to claim 1, characterized in that: The data management module includes a data preprocessing function, a partitioning strategy and a model training status report; the data preprocessing function is used for data normalization and Gaussian noise data enhancement; the partitioning strategy includes a random initialization strategy and an equal initialization strategy, the random initialization strategy is used to randomly assign the proportion of each fault type in the labeled data set, and the equal initialization strategy is used to use the same partitioning ratio in the labeled data set; the model training status report includes test accuracy, various proportions in labeled samples, confusion matrix and TSNE high-dimensional data dimensionality reduction visualization.

3. The elevator fault diagnosis method based on convolutional neural network according to claim 1, characterized in that: The convolutional neural network model includes five convolutional layers, five maximum pooling layers and one fully connected layer; the input of the convolutional neural network model is any data set, and the output is the probability value of any data set, and the sum of the probability values ​​is 1.

4. The elevator fault diagnosis method based on convolutional neural network according to claim 1 is characterized in that: The gradient embedding vector includes the low-dimensional embedding vector and prediction probability of the convolutional neural network model: Assuming there are m rows and n columns of gradient embedding vectors, the low-dimensional embedding vector The calculation formula is: ;in It refers to the low-dimensional embedding vector output by the last convolutional layer when the number of rows of the gradient embedding vector is m. Refers to the gradient weight corresponding to the gradient embedding vector when the number of columns is n and the number of rows is m; The calculation method is: Assuming there are i rows and j columns of gradient embedding vectors, where 1≤j≤n, 1≤i≤m, the calculation formula is: ;in is the final output of the convolutional neural network model when the number of columns of the gradient embedding vector is j, Refers to the gradient weight corresponding to the gradient embedding vector when the number of columns is i and the number of rows is j.

5. The elevator fault diagnosis method based on convolutional neural network according to claim 1 is characterized in that: The method for screening out diverse high-contribution samples to be labeled comprises the following steps: Step S11, input the gradient embedding vector of the diversified evaluation index of the unlabeled sample into the K-means++ clustering algorithm; Step S12: taking the gradient embedding vector with the largest norm as the first cluster center, and then calculating the distances between other vectors and the cluster center; Step S13: According to the weight Select the next cluster center until the number of cluster centers reaches the number of query samples specified by the query strategy module, and the weight The calculation method is: ; Where x represents other samples, Represents the distance between other samples and their nearest cluster center, and U is all unlabeled datasets.

Images