Multi-scale convolution residual error active DQN fault detection method

Abstract

The invention relates to a multi-scale convolution residual active DQN fault detection method, which adopts a feature extraction strategy of combining multi-scale convolution with a multi-residual module, and obtains a global feature by fusing features of different receptive fields, so that a sample has higher robustness. For the problem that marking is expensive under the condition of small samples, an active strategy is combined with a deep Q learning network (DQN) method in reinforcement learning to obtain a highest value sample in an unmarked data set, and the highest value sample is updated to a training set after being marked by experts, so that a diagnosis result with higher accuracy is obtained at lower cost. A classification model is constructed through a multi-scale residual network, samples are selected through an active deep Q learning method, and after iteration, the diagnosis model has high accuracy on the premise of small samples and low cost.

Description

Technical Field

[0001] The present invention relates to a method for detecting fault signals of rolling bearings, and specifically to a method for diagnosing different health conditions through a multi-scale residual active DQN network, which can be used for fault diagnosis and detection of mechanical equipment. Background Art

[0002] A suitable fault diagnosis method for bearings can create good conditions for the stable operation of equipment. Effective health monitoring and fault diagnosis can ensure the safety and stability of industrial equipment, and are of great significance for improving the reliability and safety of system operation and reducing the occurrence of catastrophic accidents.

[0003] The bearing fault diagnosis method based on signal processing is the most traditional and widely used method. Its core idea is to extract features from vibration signals, acoustic emission signals, temperature signals, etc. to identify the fault type and degree of the bearing. However, it has defects such as large computational amount, low efficiency, poor interpretability, and difficult optimization. Considering the complexity of the actual service environment of the bearing, the method based on signal processing technology still needs to analyze specific problems specifically, which restricts the actual application range of the fault state recognition technology.

[0004] To solve the defects that cannot be solved by signal analysis, the data-driven diagnosis method is becoming increasingly popular. The data-driven fault diagnosis technology deploys multiple sensors to collect a large amount of operation data of mechanical components, and constructs and optimizes an intelligent model to establish a function mapping relationship between the sensor data and the health state of the equipment, so as to obtain good fault diagnosis results. The intelligent fault diagnosis method based on machine learning algorithms is a classic algorithm in the data-driven fault diagnosis algorithms. This method uses a variety of machine learning strategies and learning methods to establish a mapping model between the measured sensor data and the equipment state, so as to finally realize the health state diagnosis of the equipment. This kind of method better meets the requirements of the big data era, has stronger robustness, can automatically extract features, but still has problems such as relying on a large amount of training data, being difficult to handle multi-condition problems, having poor effects in the case of small sample data, and basically not performing global feature extraction, and the features are not rich enough.

[0005] The deep reinforcement learning method in machine learning can handle multi-condition problems well. It combines the powerful feature extraction ability of deep learning and the decision-making ability of reinforcement learning. A relatively typical one is the deep Q-learning (DQN) network, which mainly focuses on how to take actions in the environment to maximize a certain cumulative reward and learns through interaction with the environment. It needs to select an action at each step, and this action will affect the state of the environment and determine the reward and the new state that the agent will receive next. This way enables the model to solve multi-condition and dynamic environment problems and is widely used in the diagnosis of mechanical equipment.

[0006] In terms of solving the problem of poor generalization of small samples, active learning has the advantages of efficient data utilization and low labeling cost. The basic principle of active learning is to select the most representative samples for labeling according to the prediction results of the model during the model training process, and then use these labeled data to further train the model, which greatly reduces the labeling cost. However, there are still problems such as incomplete feature extraction and limited generalization ability.

[0007] Multi-scale convolution is excellent for extracting global features. It obtains feature information of various receptive fields through different branches and different convolution kernels. After fusion, it obtains very rich global features, which is very beneficial for high-precision diagnosis of rolling bearing faults.

[0008] Therefore, in view of the above technical problems, when diagnosing the fault of a small sample of rolling bearings, it is necessary to deal with the problem of multiple working conditions, the problem of too few samples, the problem of too high sample labeling cost, and the problem of insufficient feature extraction. It is necessary to provide an intelligent fault diagnosis method based on multi-scale residual active DQN with small samples. Summary of the invention

[0009] In view of this, the purpose of the present invention is to use a multi-convolutional residual network to extract the global features of the signal, and comprehensively apply the entropy function, mean cosine function and deep Q learning network of active learning, so as to achieve higher diagnostic accuracy with lower cost and fewer samples.

[0010] In order to achieve the above purpose, the technical solution provided by the embodiment of the present invention is as follows: A multi-scale convolutional residual active DQN fault detection method, characterized in that the method comprises the following steps: Step S1, using a sensor device to input and perform analog / digital conversion to obtain a signal; Step S2: For feature extraction and classifier, a 25×1 wide convolutional layer is first used to preliminarily extract feature information of the vibration signal; Step S3, using convolution kernels of multiple scales, i.e., a multi-scale feature layer to obtain signal features of different receptive fields, obtaining global features after fusion and classifying them through a fully connected layer, and obtaining the diagnosis result at the output layer; Step S4, using the uncertainty criterion and similarity criterion in active learning, using the category prediction probability entropy and the average cosine distance to calculate the state and action respectively, and the reward is the accuracy increase value; Step S5: According to the state and action parameters, the maximum Q value is calculated through the deep learning Q network to obtain a batch of samples corresponding to the maximum value; Step S6: store the current state, action, reward, and the state, action, and tag value after the next model update into the experience playback buffer; Step S7: Based on the obtained high-value batch of samples, perform manual labeling, and then add this batch of samples to the training set to retrain the classifier.

[0011] As a further improvement of the present invention, the step S3 specifically includes: Step S31: The multi-scale feature layer has four branches with different convolutional kernel weights and bias vectors, which are represented by and respectively, where k represents the convolutional kernel size, and k ∈ {1, 3, 5}. The fourth branch applies max pooling MaxP, and each branch has 32 feature map channels; the expression of the multi-scale convolution is as follows: ; Step S32: To reduce the network complexity, an overlapping-max pooling layer is connected later to reduce the feature dimension. The expression is: ; Step S33: To alleviate the problem of model degradation, three residual modules are added, where i ∈ {1, 2, 3}, and different channel sizes are set. The expression is as shown in formula (3); finally, a Flatten layer is used to map to a one-dimensional vector to facilitate connection to the Dense layer to obtain the health status of the rolling bearing: .

[0012] As a further improvement of the present invention, the step S4 specifically includes: Step S41: Initially, randomly select some samples as , calculate the classification prediction probability entropy of this batch of samples; entropy reflects the degree of chaos or disorder of the system. Selecting samples with large entropy for annotation is beneficial to improving the model performance. Calculate the entropy value to obtain the uncertainty degree of the samples, and define it as the state. The expression is: ; Step S42: The samples in are defined as , the initial training samples are defined as , the unlabeled data set is defined as , , . The action consists of three parameters represents entropy, and respectively represent and the average cosine distance between and The average cosine distance between them. The average cosine distance is used to measure the similarity between samples, so as to select the most representative and diverse samples for annotation. The expression is: ; Step S43: Reward the growth value of the classification result accuracy of the representative classifier. When the model iterates to the last time, the flag value done = 1, and in other iteration processes, done = 0.

[0013] As a further improvement of the present invention, the step S5 specifically includes: Input the state and action vectors in S4 into the DQN network. Use the DNN network as the main network. The state vector passes through two dense layers, each dense layer contains 128 units and a "Sigmoid" activation function. The output data is fused with the action vector and passes through the "Linear" activation function to obtain the final output Q value. The corresponding samples of the maximum Q value are manually marked and then added to the training samples , retrain the classifier, and continuously iterate until the requirements are met.

[0014] As a further improvement of the present invention, the step S6 specifically includes: Store the current state, action, reward, and the state, action, and flag value after the next model update into the experience replay buffer. The deep Q-learning network is trained by randomly selecting a small batch of samples from the experience replay buffer. The expression is as follows: ; The network of deep Q-learning is trained using the loss function (7), where e~ε represents a small batch of samples randomly sampled from the model experience replay buffer. The deep Q network is divided into a main network and a target network, which are used to estimate the state-action function and , the two networks have the same structure but different parameters. Every N steps, the parameters of the main network are used to update the parameters of the target network, so as to improve the stability of learning: ; The target Q value is calculated using the target network, represents the reward value, represents the discount factor, and the expression is as follows: ; The update of the main network is achieved through the gradient descent algorithm, and its update process is as follows: .

[0015] Due to the application of the above technical solutions, the present invention has the following advantages compared with the prior art: In order to obtain global features, the present invention proposes a new multi-scale residual CNN network, which can extract rich global discriminant feature information from vibration signals, improve the feature learning ability of the model, and obtain higher diagnostic accuracy.

[0016] In order to solve the problem of too few samples and expensive labeling cost in small sample data, an active deep learning network is proposed. A new state and action strategy is designed by combining entropy and mean cosine distance, so as to obtain high-value samples by utilizing uncertainty and similarity, thereby achieving the purpose of reducing labeling cost.

[0017] By integrating multi-scale residuals and active deep Q learning, the labeling cost is reduced when samples are scarce. At the same time, the global features of the vibration signal are extracted, which enhances the generalization and diagnostic effects of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0019] Figure 1 This is the process of establishing and training the intelligent fault diagnosis model of the present invention; Figure 2 It is the structure diagram and internal structure of the power transmission system dynamics simulation system of Southeast University in Example 1; Figure 3 is a confusion matrix diagram of the experimental results of the invention; Figure 4 It is a graph of accuracy corresponding to the number of iterations of the invention; Figure 5 yes Figure 3 Visualization result diagram. DETAILED DESCRIPTION

[0020] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0021] The present invention will be further described below in conjunction with the accompanying drawings and embodiments. Meanwhile, the enumerated embodiments are only a part of the present invention, rather than all embodiments.

[0022] The method of the present invention includes a model training process and a diagnosis process. The fault diagnosis model is established and trained before formal diagnosis. The labeled samples in the fault diagnosis process can be used as the training data of the fault diagnosis model to continuously update and train the fault diagnosis model, and the growth of the training results is used to update the training parameters.

[0023] The establishment and training process of the intelligent fault diagnosis model is as Figure 1 shown. Through the data sampling system and the bearing fault simulation system, vibration signals of various types of rolling bearings are collected. After the signals are collected, the signals are segmented according to a certain period to obtain samples and establish a data set. The samples are divided into a small number of labeled samples and a large number of unlabeled samples. The labeled samples are further divided into a training set and a test set.

[0024] The classifier consists of a Conv1D layer, an Inception layer, a Maxpooling layer, three residual modules, and a Dense layer. First, the features of the training set and the test set are extracted, the classifier is trained with the training data, and tested with the test data, and finally the classification result is obtained.

[0025] A small number of samples are randomly selected from the unlabeled data set to. After calculation, the state and action are obtained and passed to the recognition agent, which is processed by the DNN to obtain the required batch of samples. The selected batch of samples will be updated to the training data, while the batch of samples in the unlabeled data set will be deleted. At the same time, the interaction experience (s, a, r, s') is saved in the experience replay buffer. Each iteration randomly selects data from the buffer to train the deep Q network. Continuously update the classifier and the training set until the requirements are met.

[0026] The features of the present invention are further described below through experiments and comparative examples: The example is the fault detection of the Southeast University rolling bearing data set. The data set provided by Southeast University comes from a mechanical simulator for power transmission system design. The test bench (see details Figure 2 ) consists of a planetary gearbox (F1), a reduction gearbox (F2), a motor (F3), a speed controller, and a brake adjustment motor. It is equipped with six vibration sensors (model: 608A11), with a frequency range of 0.5 Hz to 10 kHz, a measurement range of ±50 g, and an accuracy of 100 mV / g. The positions of these sensors in the DDS system are as Figure 2 (B) shown. The changes in speed and load can be achieved through the speed controller and the brake adjustment motor respectively. In order to truly simulate different operating conditions, different speed-load configurations are regarded as multiple different tasks. Different fault states include normal state, outer ring fault, inner ring fault, rolling element fault, and compound fault.

[0027] Table 1 lists the relevant data details, including the fault type, rotational speed, and load.

[0028]

[0029] For experimental verification, the data under the working conditions of 20 Hz - 0 N*m were selected, as shown in Table 2. The vibration signals under different health conditions of the bearing were segmented into independent samples. Each sample contains 2048 data points and is segmented every 1000 points. A total of 1000 samples were obtained for each category, of which 200 samples were used as the test set.

[0030]

[0031] Fault diagnosis was performed on the selected vibration signals, with the initial labeled samples being 10 for each category, as Figure 3 、 4 、5 shows, which are respectively the confusion matrix of the experimental results of the present invention, the accuracy corresponding to the number of iterations, and the result visualization. Figure 3 It shows that the overall diagnostic consistency of the present invention is very good. Except for a few mispredicted healthy samples, the others are predicted correctly. Figure 4 It shows that as the number of labeled samples increases, the diagnostic accuracy continuously improves, basically stabilizing above 95% when around 100 samples, and already above 99% when around 150 samples. Figure 5 It is the clustering visualization graph of the experimental results. The clustering effect is very outstanding. The five faults are isolated from each other, and only a very small number of samples are misclustered.

[0032] The present invention is respectively compared with SVM, 1DCNN, ResNet combined with the information entropy and margin sampling of active learning. The comparison results of the present invention and other network models using the same data for experiments are as follows in the table.

[0033]

[0034] Three criteria, namely accuracy (Acc), bias accuracy (bAcc), and macro F1-score (Macro-F1), were used for evaluation. Accuracy refers to the proportion of correctly predicted samples in the total samples. The mathematical expressions of bAcc and Macro-F1 are as follows.

[0035] ; .

[0036] Except that the training effect of SVM is very poor, the accuracy rates of other networks have achieved good results. Since SVM mainly serves binary classification and has poor performance for multi-classification problems. The training effects of both 1DCNN for these two categories have reached over 90%, and relatively good feature extraction effects have been achieved. However, it is still slightly inferior compared with the ResNet network. The accuracy rates of ResNet for these two categories of classification have basically reached over 98%, and richer deep features are extracted. And the present invention has better effects than other networks, with a classification accuracy rate of over 99.5%, and even reached 100% accuracy rate in multiple experiments.

[0037] As can be seen from the analysis process and application examples, the present invention is very effective for the intelligent diagnosis of bearing faults. Its characteristics of reducing costs and being able to extract global features determine that this method can be effectively applied to the fault diagnosis of rotating machinery.

[0038] The present invention discloses a rolling bearing fault diagnosis method based on small-sample multi-scale residual active deep Q learning, which inputs through a sensing device and performs analog-to-digital conversion to obtain a signal , extracts global features of the signal, and selects high-value samples for marking, including the following steps: for feature extraction and classifier, first use a 25×1 wide convolution to preliminarily extract the feature information of the vibration signal; use convolutional kernels of multiple scales to obtain signal features with different receptive fields, and after fusion, obtain global features and perform classification to obtain the diagnosis result of this time; use the uncertainty criterion and similarity criterion in active learning, and respectively use the category prediction probability entropy and the average cosine distance to calculate the state and action, and the reward is the growth value of the classification accuracy rate; according to the state and action parameters, calculate the maximum Q value through the deep learning Q network to obtain a batch of samples corresponding to the maximum value; store the current state, action, reward, and the state, action, and marked value after the next model update in the experience replay buffer; after manually marking the samples, add them to the training set to retrain the classifier. By extracting global features through multi-scale convolution and actively marking samples by DQN, it enriches features and reduces costs at the same time, and is very suitable for the intelligent fault diagnosis of rolling bearings.

[0039] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be regarded as limiting the claimed rights.

[0040] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative way of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A multi-scale convolutional residual active DQN fault detection method, characterized in that: The method comprises the following steps: Step S1: Use the sensor device to input and perform analog / digital conversion to obtain a signal ; Step S2: For feature extraction and classifier, a 25×1 wide convolutional layer is first used to preliminarily extract feature information of the vibration signal; Step S3, using convolution kernels of multiple scales, i.e., a multi-scale feature layer to obtain signal features of different receptive fields, obtaining global features after fusion and classifying them through a fully connected layer, and obtaining the diagnosis result at the output layer; Step S4, using the uncertainty criterion and similarity criterion in active learning, using the category prediction probability entropy and the average cosine distance to calculate the state and action respectively, and the reward is the accuracy increase value; Step S5: According to the state and action parameters, the maximum Q value is calculated through the deep learning Q network to obtain a batch of samples corresponding to the maximum value; Step S6: store the current state, action, reward, and the state, action, and tag value after the next model update into the experience playback buffer; Step S7: perform manual labeling based on the obtained high-value batch samples, and then add this batch of samples to the training set to retrain the classifier.

2. The multi-scale convolutional residual active DQN fault detection method according to claim 1 is characterized in that: The step S3 specifically includes: Step S31, the multi-scale feature layer has four branches, each with different convolution kernel weights and bias vectors, respectively. and Indicates that k represents the convolution kernel size, where k∈{1,3,5}. The fourth branch applies the maximum pooling MaxP, and each branch has 32 feature mapping channels; Step S32: To reduce network complexity, an overlap-max pooling layer is added to reduce feature dimension. Step S33: To alleviate the model degradation problem, three residual modules are added , where i∈{1,2,3}, setting different channel sizes; finally, the Flatten layer is used to map it to a one-dimensional vector, which is convenient for connecting to the Dense layer to obtain the health status of the rolling bearing.

3. The multi-scale convolutional residual active DQN fault detection method according to claim 1, characterized in that: The step S4 specifically includes: Step S41: Initially, some samples are randomly selected as , calculate the classification prediction probability entropy of the batch of samples; entropy reflects the degree of chaos or disorder of the system. Selecting samples with large entropy for labeling is beneficial to improving model performance. The entropy value is calculated to obtain the uncertainty of the sample, which is defined as the state; Step S42: The sample in is defined as , the initial training sample is defined as , the unlabeled dataset is defined as , actions use three parameters , , constitute, represent The entropy of and Respectively and The mean cosine distance between and The average cosine distance between samples. The average cosine distance is used to measure the similarity between samples, so as to select the most representative and diverse samples for annotation; Step S43, the reward represents the growth value of the accuracy of the classification result of the classifier. When the model iterates to the last time, the mark value done = 1, and done = 0 in other iterations.

4. The multi-scale convolutional residual active DQN fault detection method according to claim 1, characterized in that: The step S5 specifically includes: The state and action vectors in S4 are input into the DQN network. The DNN network is used as the main network. The state vector passes through two dense layers, each of which contains 128 units and a "Sigmoid" activation function. The output data is fused with the action vector and the final output Q value is obtained through the "Linear" activation function. The corresponding samples with the maximum Q value are manually marked and added to the training samples. , retrain the classifier and iterate continuously until the requirements are met.

5. The multi-scale convolutional residual active DQN fault detection method according to claim 1, characterized in that: The step S6 specifically includes: The current state, action, reward, and the state, action, and tag value after the next model update are stored in the experience replay buffer. The deep Q learning network is trained by randomly selecting small batches of samples in the experience replay buffer.

Images