Bearing fault diagnosis method, system and electronic equipment

Abstract

The present application discloses a bearing fault diagnosis method, system and electronic device, which relates to the field of fault diagnosis. The first feature extraction network is used to extract the parameter information of the bearing to be diagnosed, and the bearing feature information is obtained. The feature information of the bearing in a known state is used to classify and train the diagnosis model. Finally, the bearing feature information of the bearing to be diagnosed is used to perform model testing to determine whether the bearing to be diagnosed has a fault. The weight parameters of the first feature extraction network are obtained after the initial feature extraction network and the improved least squares discriminator are optimized through adversarial training. The first feature extraction network obtained after parameter optimization improves the accuracy of fault diagnosis of the bearing.

Description

Technical Field

[0001] The present application relates to the field of fault diagnosis, and in particular to a bearing fault diagnosis method, system and electronic equipment. Background Art

[0002] Bearings, as core components of many rotating machines, are prone to various failures in complex working environments and changing working conditions.

[0003] Traditional bearing fault diagnosis methods rely on manual analysis of bearing vibration signals to detect whether the bearing is faulty. However, this diagnostic method is inefficient and cannot guarantee the accuracy of the diagnostic results. Summary of the invention

[0004] In view of this, the present application provides a bearing fault diagnosis method, system and electronic device, and the specific scheme thereof is as follows:

[0005] A bearing fault diagnosis method, comprising:

[0006] Obtain parameter information of the bearing to be diagnosed;

[0007] Extracting bearing feature information from the parameter information of the bearing to be diagnosed based on a first feature extraction network;

[0008] Using the trained diagnostic model to classify the bearing feature information, and determine whether the bearing to be diagnosed has a fault;

[0009] Among them, the trained diagnostic model is: the diagnostic model is trained using the characteristic information of a bearing in a known state, and the weight parameters of the neural network of the diagnostic model are adjusted to obtain the trained diagnostic model; the first feature extraction network weight parameters are obtained by using the output result of the initial feature extraction network as the input information of the improved least squares discriminator, and based on the input information, performing adversarial training and optimization on the initial feature extraction network and the improved least squares discriminator.

[0010] Furthermore, the output layer of the improved least squares discriminator includes a Softmax activation function.

[0011] Furthermore, it also includes:

[0012] Obtain historical data information of bearings;

[0013] Obtaining a training set and a test set from the historical data information, wherein the training set is historical data information of bearings with known fault types, and the test set is data information of bearings with unknown status types;

[0014] Based on the initial feature extraction network, feature extraction is performed on the training set to obtain first feature information, and feature extraction is performed on the test set to obtain second feature information;

[0015] Utilizing the Softmax activation function and the loss function in the improved least squares discriminator, based on the first feature information and the second feature information, the weight parameters in the initial feature extraction network are adjusted to obtain a first feature extraction network.

[0016] Furthermore, the using of the Softmax activation function and the loss function in the improved least squares discriminator, based on the first feature information and the second feature information, adjusting the weight parameters in the initial feature extraction network to obtain a first feature extraction network, includes:

[0017] Based on the Softmax activation function in the improved least squares discriminator, determining a first probability distribution of the first feature information in the improved least squares discriminator and a second probability distribution of the second feature information in the improved least squares discriminator;

[0018] Determine a loss function value based on the first probability distribution and the second probability distribution using a loss function;

[0019] The weight parameters in the initial feature extraction network are adjusted based on the loss function value to obtain a first feature extraction network.

[0020] Furthermore, the using the loss function to determine the loss function value based on the first probability distribution and the second probability distribution includes:

[0021] Determine a first loss function including the first parameter and the second parameter, and a second loss function including the third parameter;

[0022] Using the first loss function, based on the first mean square error and the second mean square error, determining the identification loss value corresponding to the initial feature extraction network, wherein the first mean square error is the mean square error between the first probability distribution and the first parameter, the second mean square error is the mean square error between the second probability distribution and the second parameter, and the identification loss value is used to characterize, in adversarial training, the difference between the third probability distribution of the training set and the expected target distribution, the third probability distribution being the probability distribution of the training set output by the initial feature extraction network and the improved least squares discriminator;

[0023] Utilizing the second loss function, a generation loss value corresponding to the initial feature extraction network is determined based on a third mean square error, wherein the third mean square error is the mean square error between the second probability distribution and the third parameter, and the generation loss value is used to characterize, in adversarial training, the difference between a fourth probability distribution of the test set and the expected target distribution, the fourth probability distribution being the probability distribution of the test set output after passing through the initial feature extraction network and the improved least squares discriminator; the expected target distribution characterizes the expected distribution corresponding to the encoding parameters in adversarial training.

[0024] Further, adjusting the weight parameters in the initial feature extraction network based on the loss function value to obtain a first feature extraction network includes:

[0025] Adjusting weight parameters in the initial feature extraction network based on the loss function value;

[0026] In the process of adjusting the weight parameters of the initial feature extraction network multiple times, determining the loss function value corresponding to the second feature extraction network obtained after each adjustment;

[0027] Based on the loss function value corresponding to the second feature extraction network obtained after each adjustment, a second feature extraction network is determined as the first feature extraction network from the second feature extraction networks obtained after each adjustment.

[0028] Further, based on the loss function value corresponding to the second feature extraction network obtained after each adjustment, determining a second feature extraction network as the first feature extraction network from the second feature extraction networks obtained after each adjustment includes:

[0029] From the loss function values ​​corresponding to the second feature extraction networks obtained after each adjustment, the second feature extraction network corresponding to the minimum loss function value is selected, and the selected second feature extraction network is determined as the first feature extraction network.

[0030] Further, the determining of a first loss function including the first parameter and the second parameter, and a second loss function including the third parameter, includes:

[0031] Determine the target encoding method;

[0032] Encoding the first parameter, the second parameter and the third parameter based on the target encoding method so that the first parameter, the second parameter and the third parameter are two-dimensional vectors;

[0033] A first loss function including the first parameter and the second parameter is determined, and a second loss function including a third parameter is determined.

[0034] A bearing fault diagnosis system, comprising:

[0035] An acquisition unit, used for acquiring parameter information of the bearing to be diagnosed;

[0036] An extraction unit, configured to extract features from the parameter information of the bearing to be diagnosed by using a first feature extraction network to obtain bearing feature information;

[0037] A classification unit, which uses the trained diagnosis model to classify the bearing feature information to determine whether the bearing to be diagnosed has a fault;

[0038] Among them, the trained diagnostic model is: the diagnostic model is trained using the characteristic information of a bearing in a known state, and the weight parameters of the neural network of the diagnostic model are adjusted to obtain the trained diagnostic model; the weight parameters of the first feature extraction network are obtained by using the output result of the initial feature extraction network as the input information of the improved least squares discriminator, and based on the input information, performing adversarial training and optimization on the initial feature extraction network and the improved least squares discriminator.

[0039] An electronic device, comprising:

[0040] A processor, used to obtain parameter information of a bearing to be diagnosed; use a first feature extraction network to perform feature extraction on the parameter information of the bearing to be diagnosed to obtain bearing feature information; use a trained diagnostic model to classify the bearing feature information to determine whether the bearing to be diagnosed has a fault; wherein the trained diagnostic model is: the diagnostic model is trained using feature information of a bearing in a known state, and the weight parameters of the neural network of the diagnostic model are adjusted to obtain the trained diagnostic model; the weight parameters of the first feature extraction network are obtained by taking the output result of the initial feature extraction network as input information of an improved least squares discriminator, and based on the input information, performing adversarial training optimization on the initial feature extraction network and the improved least squares discriminator;

[0041] The memory is used to store the program required by the processor to execute the above processing process.

[0042] It can be seen from the above technical scheme that the bearing fault diagnosis method, system and electronic device disclosed in the present application obtain parameter information of the bearing to be diagnosed; use the first feature extraction network to extract features from the parameter information of the bearing to be diagnosed to obtain bearing feature information; use the trained diagnostic model to classify the bearing feature information to determine whether the bearing to be diagnosed has a fault; wherein the trained diagnostic model is: the diagnostic model is trained by classifying the diagnostic model using the feature information of the bearing in a known state, optimizing and adjusting the weight parameters of the neural network of the diagnostic model, and obtaining the trained diagnostic model; the first feature extraction network weight parameters are obtained by taking the output result of the initial feature extraction network as the input information of the improved least squares discriminator, and based on the input information, performing adversarial training and optimization on the initial feature extraction network and the improved least squares discriminator. This scheme uses the first feature extraction network to extract features from the parameter information of the bearing to be diagnosed, and uses the diagnostic model to classify and train it to determine whether the bearing to be diagnosed has a fault. The first feature extraction network uses the output result of the initial feature extraction network as the input information of the improved least squares discriminator. Based on the input information, the initial feature extraction network and the improved least squares discriminator are optimized through adversarial training. The accuracy of fault diagnosis of bearings is improved based on the optimized first feature extraction network. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related technologies, the drawings required for use in the embodiments or the related technical descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0044] Figure 1 A flow chart of a bearing fault diagnosis method disclosed in an embodiment of the present application;

[0045] Figure 2 A schematic diagram of data distribution of a Sigmoid activation function disclosed in an embodiment of the present application;

[0046] Figure 3 A schematic diagram of a Softmax activation function output probability distribution disclosed in an embodiment of the present application;

[0047] Figure 4 A flow chart of a bearing fault diagnosis method disclosed in an embodiment of the present application;

[0048] Figure 5 A schematic diagram of a structural block diagram of a feature extraction network training phase disclosed in an embodiment of the present application;

[0049] Figure 6 A schematic diagram of a structural block diagram of diagnosing a bearing fault using a first feature extraction network disclosed in an embodiment of the present application;

[0050] Figure 7a A schematic diagram of a probability distribution between a training set and a test set disclosed in an embodiment of the present application;

[0051] Figure 7b A schematic diagram of the probability distribution between a training set and a test set in the prior art;

[0052] Figure 8 A structural schematic diagram of a bearing fault diagnosis system disclosed in an embodiment of the present application;

[0053] Fig. 9 A schematic diagram of the structure of an electronic device disclosed in an embodiment of the present application. DETAILED DESCRIPTION

[0054] The following describes the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. The terms used in the implementation method section of the present application are only used to explain the specific embodiments of the present application, and are not intended to limit the present application.

[0055] The embodiments of the present application are described below in conjunction with the accompanying drawings. Those skilled in the art will appreciate that, with the development of technology and the emergence of new scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0056] The terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and need not be used to describe a specific order or sequential order. It should be understood that the terms used in this way can be interchangeable under appropriate circumstances, which is only to describe the distinction mode adopted by the objects of the same attributes when describing in the embodiments of the present application. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, so that the process, method, system, product or equipment comprising a series of units need not be limited to those units, but may include other units that are not clearly listed or inherent to these processes, methods, products or equipment.

[0057] The present application discloses a bearing fault diagnosis method, and its flow chart is as follows: Figure 1 As shown, including:

[0058] Step S11, obtaining parameter information of the bearing to be diagnosed;

[0059] Step S12: using the first feature extraction network to extract features from parameter information of the bearing to be diagnosed, to obtain bearing feature information;

[0060] Step S13: classify the bearing feature information using the trained diagnostic model to determine whether the bearing to be diagnosed has a fault.

[0061] Among them, the trained diagnostic model is: the diagnostic model is trained using the characteristic information of a bearing in a known state, and the weight parameters of the neural network of the diagnostic model are optimized and adjusted to obtain the trained diagnostic model; the weight parameters of the first feature extraction network are obtained by taking the output result of the initial feature extraction network as the input information of the improved least squares discriminator, and based on the input information, performing adversarial training and optimization on the initial feature extraction network and the improved least squares discriminator.

[0062] Bearings, as core components of many rotating machines, are prone to various failures in complex working environments and changing working conditions.

[0063] Traditional bearing fault diagnosis methods rely on manual analysis of bearing vibration signals to detect whether the bearing is faulty. However, this diagnostic method is inefficient and cannot guarantee the accuracy of the diagnostic results. With the development of technology, machine learning is gradually used for bearing fault diagnosis. This method usually relies on a large amount of labeled data and feature extraction under specific operating conditions. However, in actual industrial scenarios, the cost of obtaining labeled data is high and there are significant differences in data distribution between different operating conditions, resulting in poor performance of the model under new operating conditions. In order to overcome this problem, transfer learning is widely used in bearing fault diagnosis. However, transfer learning methods based on distance metrics usually rely on specific distribution assumptions and are difficult to capture heterogeneous data with complex distribution characteristics. In addition, although the original least squares adversarial training method achieves transfer learning by minimizing the distribution differences between bearing data samples, it uses the Sigmoid function as the activation function, which is prone to cause the gradient vanishing problem. At the same time, it is difficult to effectively capture the true distribution characteristics in the complex distribution of high-dimensional data.

[0064] Specifically, the original least squares adversarial training method uses the Sigmoid activation function in the output layer of the discriminator. That is, when the discriminator is good enough, the Sigmoid activation function will be in a saturated state, such as Figure 2 As shown, the formula of the Sigmoid activation function is: , z is the input data, such as Figure 2As shown in , when the data input to the discriminator exceeds a certain range, the Sigmoid activation function can only output 0 or 1, and cannot further obtain the accurate distribution state of the input data. In addition, the gradient parameters used by the discriminator for network update, along with the saturation state of the Sigmoid activation function, decrease sharply or disappear as the network depth increases during the back propagation process, which leads to the inability of adversarial training to effectively update the feature extraction network parameters. At this time, the adversarial training fails and cannot further reduce the data distribution difference between the source domain and the target domain in the feature extraction network. Among them, the source domain usually refers to the data distribution range of the training set; the target domain usually refers to the data distribution range of the test set.

[0065] The rapid development of artificial intelligence based on deep learning has enabled the fault diagnosis of rotating machinery with bearings as important components to develop towards intelligence and automation. Bearing fault diagnosis models based on deep learning usually rely on labeled data to train the model to achieve automatic diagnosis, but in actual industry, it is difficult to obtain all labeled fault data to train the model, and even if the same fault occurs in bearings of different sizes or the same bearings under different working conditions, the parameters have domain offsets, that is, the data distribution is inconsistent, which makes it difficult to establish a universal bearing fault diagnosis model.

[0066] Based on this, this scheme uses the initial feature extraction network and the improved least squares discriminator to perform transfer learning of adversarial training, and optimizes the initial feature extraction network based on the output results of the least squares discriminator to obtain the weight parameters of the first feature extraction network, thereby realizing the determination of the first feature extraction network, and using the first feature extraction network to extract the bearing feature information in the parameter information of the bearing to be diagnosed, and using the diagnostic model to classify the bearing feature information to determine whether the bearing to be diagnosed has a fault. By improving the Softmax activation function in the least squares discriminator, the gradient vanishing caused by the Sigmoid activation function used in the original least squares is avoided, which leads to adversarial failure, inability to accurately obtain data distribution, and instability based on cross entropy adversarial problems, and more effective domain-invariant features are established in the feature extraction network to improve the accuracy of fault diagnosis.

[0067] The improved least squares discriminator includes the Softmax activation function, which is a vector activation function mainly used in the output layer of the neural network to provide the probability of each category, converting the vector or original output into a probability distribution so that the sum of the output elements is 1. For a classification problem with N categories, the Softmax activation function will output N values, and the sum of these values ​​is 1, and each value represents the probability of the corresponding category, which enables the Softmax activation function to clearly give the distribution of input data in different categories;

[0068] In addition, the Softmax activation function can consider the relative size of each input value when calculating the probability, and determine the probability of the corresponding category according to the relative size of each element, such as Figure 3 The figure shows the principle diagram of the Softmax activation function, where the formula of the Softmax activation function is:

[0069]

[0070] Among them, z represents the input data, i represents the i-th element in the input data, j represents the index in the sum operation of all elements in the input data, and j is used to traverse all elements in the input data to achieve the sum. The input data z passes through the formula of the Softmax activation function to obtain the output probability p, which is the data probability distribution obtained after the input data z is nonlinearly transformed by the Softmax activation function, so as to be used for subsequent adversarial training.

[0071] like Figure 3 The formula shown in: The main function of the Softmax activation function is to convert the input vector (usually an unnormalized score or log probability) into a multidimensional probability distribution, that is, to obtain a probability distribution consisting of multiple output probabilities p, where each element in the probability distribution represents the probability that the input belongs to the corresponding category, the sum of all elements in the probability distribution is equal to 1, and each element is in the interval [0, 1].

[0072] The output value of the Softmax activation function in the improved least squares discriminator is used as the output value of the discriminator, and the output value of the improved least squares discriminator is used to adjust the initial feature extraction network.

[0073] In adversarial training, the feature extraction network (generator) extracts features from real data, generates new data samples based on the features, and makes the generated data samples similar to the real data, and outputs the generated data samples and the real data; the improved least squares discriminator obtains the output result of the feature extraction network, judges the authenticity of the obtained data, that is, judges whether the obtained data is real data or generated data samples. The authenticity result judged by the improved least squares discriminator will be fed back to the feature extraction network to optimize the feature extraction network so that the generated data distribution is closer to the real data, and at the same time, improves the accuracy of the judgment of the improved least squares discriminator.

[0074] Transfer learning is the use of knowledge learned on one task to solve another related but different problem. Transfer learning can reduce the training time and computing resource requirements while improving the performance of new tasks.

[0075] By combining adversarial training and transfer learning, using transfer learning to adjust the pre-trained model, and using adversarial training to further improve the generalization ability of the model, through multiple iterative training, both the feature extraction network and the discriminator can achieve optimal performance.

[0076] It is determined that the feature extraction network that achieves the best performance is the first feature extraction network. At this time, when there is a bearing to be diagnosed, the parameter information of the bearing to be diagnosed can be obtained, and the parameter information of the bearing to be diagnosed is subjected to feature extraction by the first feature extraction network to obtain the bearing feature information. The bearing feature information is classified using the diagnostic model to determine whether the bearing to be diagnosed has a fault, thereby ensuring the accuracy of fault diagnosis.

[0077] Among them, the parameter information of the bearing to be diagnosed may include at least one of the following: the bearing vibration signal of the bearing to be diagnosed, the bearing sound print signal of the bearing to be diagnosed, the bearing acceleration of the bearing to be diagnosed, the temperature of the bearing to be diagnosed, the bearing speed of the bearing to be diagnosed and other information, so as to ensure that the information obtained about the bearing to be diagnosed is more accurate and ensure that the fault diagnosis of the bearing to be diagnosed is more precise.

[0078] The bearing fault diagnosis method disclosed in this embodiment obtains parameter information of the bearing to be diagnosed; uses the first feature extraction network to extract features from the parameter information of the bearing to be diagnosed to obtain bearing feature information; uses the trained diagnostic model to classify the bearing feature information to determine whether the bearing to be diagnosed has a fault; wherein the trained diagnostic model is: the diagnostic model is trained by classifying the diagnostic model using feature information of a bearing in a known state, optimizing and adjusting the weight parameters of the neural network of the diagnostic model, and obtaining the trained diagnostic model; the weight parameters of the first feature extraction network are obtained by taking the output result of the initial feature extraction network as the input information of the improved least squares discriminator, and based on the input information, performing adversarial training and optimization on the initial feature extraction network and the improved least squares discriminator. The first feature extraction network obtained after optimization improves the accuracy of fault diagnosis of the bearing.

[0079] This embodiment discloses a bearing fault diagnosis method, and its flow chart is as follows: Figure 4 As shown, including:

[0080] Step S41, obtaining historical data information of the bearing; obtaining a training set and a test set in the historical data information, wherein the training set is the historical data information of the bearing with known fault types, and the test set is the data information of the bearing with unknown status types;

[0081] Step S42: extracting features from the training set based on the initial feature extraction network to obtain first feature information, and extracting features from the test set to obtain second feature information;

[0082] Step S43, using the Softmax activation function and the loss function in the improved least squares discriminator, based on the first feature information and the second feature information, adjusting the weight parameters in the initial feature extraction network to obtain a first feature extraction network;

[0083] Step S44, obtaining parameter information of the bearing to be diagnosed;

[0084] Step S45, extracting bearing feature information from parameter information of the bearing to be diagnosed based on the first feature extraction network;

[0085] Step S46: Classify the bearing feature information using the trained diagnostic model to determine whether the bearing to be diagnosed has a fault.

[0086] The training of feature extraction network can be achieved based on the migration method.

[0087] Specifically, when it is necessary to train the feature extraction network, historical data information of the bearing can be obtained first. The historical data information of the bearing may include historical data information that is known to characterize whether the bearing is faulty, and data information that is not sure whether it can characterize the bearing fault, or historical data information that is known to characterize what kind of bearing fault it is, and data information that is not sure whether it can characterize the bearing fault, that is, historical data information with known data labels and data information with unknown data labels.

[0088] The historical data information of the bearing obtained is divided into different data sets, that is, the historical data information of known fault types is determined as the training set, that is, the source domain, and the data information of unknown state types is determined as the test set, that is, the target domain.

[0089] Afterwards, the initialized feature extraction network is used to perform feature extraction on the training set and the test set, respectively. Feature extraction is performed on the training set to obtain first feature information, and feature extraction is performed on the test set to obtain second feature information.

[0090] In order to apply the bearing condition diagnosis experience learned from the training set to the test set, adversarial training is performed on the output features of the feature extraction network to reduce the data distribution differences (i.e., domain offset) between the training set and the test set in the feature extraction network, and to establish domain-invariant features. That is, to eliminate the data differences in bearing parameters caused by mechanical equipment load, speed, etc., and to extract universal features from bearing parameters under different working conditions.

[0091] Adversarial training is performed using the Softmax activation function and loss function in the improved least squares discriminator. Based on the first feature information and the second feature information, the weight parameters in the initial feature extraction network are optimized and adjusted to obtain the first feature extraction network.

[0092] Specifically, based on the Softmax activation function in the improved least squares discriminator, a first probability distribution of the first feature information in the improved least squares discriminator and a second probability distribution of the second feature information in the improved least squares discriminator are determined; using the loss function, based on the first probability distribution and the second probability distribution, a loss function value is determined; based on the loss function value, the weight parameters in the initial feature extraction network are adjusted to obtain a first feature extraction network.

[0093] After obtaining the first feature information and the second feature information, the diagnostic model feature extraction network of the neural network transmits the first feature information and the second feature information to the improved least squares discriminator. The output layer of the improved least squares discriminator includes a Softmax activation function, which can accurately obtain the first probability distribution corresponding to the first feature information and the second probability distribution corresponding to the second feature information, wherein the first probability distribution is the Softmax binary probability distribution obtained by the training set in the improved least squares discriminator. , the second probability distribution is the Softmax binary probability distribution obtained by the test set in the improved least squares discriminator .

[0094] The formula is as follows:

[0095]

[0096] in, is the output feature of the test set after the feature extraction network, Output features of the test set after the feature extraction network , the probability value of the first column and the i-th row in the binary distribution probability y predicted by the Softmax activation function in the improved least squares discriminator, Output features representing the test set of the improved least squares discriminator output The corresponding entire Softmax probability distribution, the right side of the equation represents the Softmax activation function Perform nonlinear mapping. Among them, That is .

[0097] After obtaining the first probability distribution and the second probability distribution, a loss value is determined according to a set loss function, and the neural network optimizer adjusts the weight parameters in the initial feature extraction network based on the loss value to optimize the initial feature extraction network.

[0098] The above process is a continuous iterative process until the feature extraction network reaches the optimum and obtains the first feature extraction network, so that the features extracted by the first feature extraction network can be more conducive to extracting common features from bearing parameters under various working conditions, and further improve the recognition accuracy of the diagnostic model, which is more conducive to predicting the operating status of the bearing. After obtaining the first feature extraction network, the first feature extraction network is used to extract features from the parameter information of the bearing to be diagnosed, and the extracted features are input into the diagnostic model for classification training or testing to achieve the diagnosis of bearing faults.

[0099] Specifically, the structural diagram of the feature extraction network training phase can be shown as follows: Figure 5 As shown, until the feature extraction network reaches the optimum, a first feature extraction network is obtained. The structural block diagram of using the first feature extraction network to diagnose bearing faults can be shown as follows Figure 6 shown.

[0100] In this embodiment, the training stage of the feature extraction network will have two loss functions, namely the first loss function and the second loss function. The first loss function is the identification loss of the improved least squares discriminator for the training set in adversarial training; the second loss function is the generation loss of the improved least squares discriminator for the test set in adversarial training.

[0101] Specifically, it can be:

[0102] Determine a first loss function including a first parameter and a second parameter, and a second loss function including a third parameter; use the first loss function to determine the identification loss value corresponding to the initial feature extraction network based on the first mean square error and the second mean square error, wherein the first mean square error is a first probability distribution and the mean square error between the first parameter b, and the second mean square error is the second probability distribution The mean square error between the second parameter a and the training set is used to characterize the difference between the third probability distribution of the training set and the expected target distribution in adversarial training. The third probability distribution is the probability distribution of the training set after the initial feature extraction network and the improved least squares discriminator output. The second loss function is used to determine the generation loss value corresponding to the initial feature extraction network based on the third mean square error, where the third mean square error is the second probability distribution. The mean square error between the fourth probability distribution and the third parameter c generates a loss value used to characterize the difference between the fourth probability distribution of the test set and the expected target distribution in adversarial training; the fourth probability distribution is the probability distribution of the test set output after the initial feature extraction network and the improved least squares discriminator, and the expected target distribution represents the expected distribution corresponding to the encoding parameters in adversarial training.

[0103] Among them, in the bearing fault diagnosis method disclosed in this embodiment, in adversarial training, the feature extraction network can also be used as a generator, and the generator is used to generate new data samples so that the distribution of the new data samples is closer to the real data, ensuring that the new data samples are similar to the real data, wherein the generator has encoding parameters when generating data, that is, the identification loss value and the generation loss value are loss values ​​related to the encoding parameters of the generator.

[0104] Among them, the first loss function can be:

[0105]

[0106] The second loss function can be:

[0107]

[0108] represents minimizing the identification loss of the improved least squares discriminator for the training set, and accordingly, represents minimizing the generation loss of the improved least squares discriminator for the test set, where Represents the data feature distribution of the source domain The average expectation of the mean square error between and the corresponding encoding parameter b, Represents the data feature distribution of the target domain The average expectation of the mean square error between and the corresponding coding parameter b or c.

[0109] Minimize the identification loss of the improved least squares discriminator for the training set, that is, the first probability distribution Close to the first parameter b, while making the second probability distribution Close to the second parameter a; minimize the generation loss of the improved least squares discriminator for the test set, even if the second probability distribution Close to the third parameter c. To simultaneously satisfy the minimization of the above two loss functions, it is necessary to continuously and alternately optimize and update the weight parameters of the feature extraction network so that the probability distribution of the training set and the test set obtained through the feature extraction network constantly changes, that is, the improved least squares discriminator is used to compete with the feature extraction network through adversarial training to achieve the purpose of blurring the data distribution boundary of the training set and the test set, so that the distribution of the training set and the test set is as close as possible, so as to reduce the data distribution differences caused by working conditions, loads and other conditions.

[0110] Specifically, adjusting the weight parameters in the initial feature extraction network based on the loss function value to obtain the first feature extraction network can be as follows: adjusting the weight parameters in the initial feature extraction network based on the loss function value, and in the process of adjusting the weight parameters of the initial feature extraction network multiple times, determining the loss function value corresponding to the second feature extraction network obtained after each adjustment; based on the loss function value corresponding to the second feature extraction network obtained after each adjustment, determining a second feature extraction network as the first feature extraction network from the second feature extraction networks obtained after each adjustment.

[0111] The weight parameters of the initial feature extraction network are adjusted to obtain the first feature extraction network. This process requires multiple adjustments. Based on the initial feature extraction network, the first feature extraction network cannot be directly obtained by a single adjustment. The adjustment of the weight parameters of the feature extraction network is determined based on the loss function value. After the weight parameters of the feature extraction network are determined, the loss function value can be calculated. After the loss function value is obtained, it can be further determined based on the loss function value whether the weight parameters of the current feature extraction network need to be adjusted. After the weight parameters of the feature extraction network are adjusted, the loss function value calculated using the adjusted feature extraction network meets the conditions, and the adjustment of the weight parameters of the feature extraction network is stopped.

[0112] For example, the weight parameters of the initial feature extraction network are determined, and the loss function value corresponding to the initial feature extraction network is calculated through the loss function. After obtaining the loss function value, the weight parameters of the initial feature extraction network are adjusted based on the loss function value. After adjustment, the second feature extraction network obtained by the first adjustment is obtained. At this time, the loss function value corresponding to the second feature extraction network obtained by the first adjustment is calculated through the loss function. After obtaining the loss function value, the weight parameters of the second feature extraction network are adjusted based on the loss function value. After adjustment, the second feature extraction network obtained by the second adjustment is obtained, and so on. When the conditions are met, the first feature extraction network is determined.

[0113] Among them, the condition is satisfied: the second feature extraction network corresponding to the minimum loss function value is obtained, at this time, the second feature extraction network corresponding to the minimum loss function value can be determined as the first feature extraction network. That is: from the loss function values ​​corresponding to the second feature extraction networks obtained after each adjustment, the second feature extraction network corresponding to the minimum loss function value is selected, and the selected second feature extraction network is determined as the first feature extraction network.

[0114] The loss function value may include: identifying a loss value and generating a loss value.

[0115] Determine the total number of times k for adjusting the weight parameters of the feature extraction network, that is, adjust the weight parameters of the feature extraction network. When the number of adjustments reaches k, the weight parameters of the feature extraction network are no longer adjusted, but the adjusted feature extraction network corresponding to the smallest loss function value is selected from the k adjustments, and it is determined as the first feature extraction network, where k is a positive integer, and k can be specifically set according to actual needs, and there is no restriction here.

[0116] From the feature extraction networks obtained from the k-times adjustment of the weight parameters, the adjusted feature extraction network corresponding to the minimum loss function value is selected as the first feature extraction network. In this process, the feature extraction network obtained by each adjustment is used to obtain the corresponding identification loss value and generation loss value, and the identification loss value and the generation loss value are compared respectively, so as to select the feature extraction network with the smallest identification loss value and generation loss value. For example, in the 5-times adjustment process, 5 feature extraction networks are obtained respectively, and the identification loss value and the generation loss value corresponding to these 5 feature extraction networks are calculated respectively. If it is determined that the identification loss value of the feature extraction network obtained after the 3rd adjustment is less than the identification loss value of the feature extraction network obtained after the 1st, 2nd, 4th and 5th adjustments, and the generation loss value of the feature extraction network obtained after the 3rd adjustment is less than the generation loss value of the feature extraction network obtained after the 1st, 2nd, 4th and 5th adjustments, then the feature extraction network obtained after the 3rd adjustment is determined as the first feature extraction network.

[0117] That is, after adjusting the weight parameters in the feature extraction network for the nth time, the second feature extraction network is obtained, and the identification loss value obtained by using the second feature extraction network is minimized. At the same time, the generation loss value obtained by using the second feature extraction network is minimized. At this time, the minimization of the identification loss and the minimization of the generation loss are achieved. The current second feature extraction network is determined as the first feature extraction network to be used as the optimal feature extraction network, where n is a positive integer, and n is less than or equal to k.

[0118] Furthermore, since the improved least squares discriminator uses the output of the Softmax activation function as output, that is, the output of the improved least squares discriminator is a probability distribution, in order to match the output of the improved least squares discriminator to smoothly calculate the probability distribution, it can be determined that the first parameter, the second parameter and the third parameter are encoded using a target encoding method, rather than using the loss function of the original least squares adversarial training method to directly set the first parameter, the second parameter and the third parameter to specific target encoding values ​​0 or 1.

[0119] That is: determine the target encoding method; encode the first parameter, the second parameter and the third parameter based on the target encoding method so that the first parameter, the second parameter and the third parameter are two-dimensional vectors; determine a first loss function including the first parameter and the second parameter, and determine a second loss function including the third parameter, so that the first parameter and the second parameter in the first loss function, and the third parameter in the second loss function are two-dimensional vectors.

[0120] Specifically, the target encoding method can make the first parameter, the second parameter and the third parameter all two-dimensional vectors, so that when calculating the loss function, the probability distribution can be calculated corresponding to the corresponding parameters. Specifically, the first parameter is a two-dimensional vector set with the first column being 1 and the second column being 0; the second parameter is a two-dimensional vector set with the first column being 0 and the second column being 1; the third parameter is a two-dimensional vector set with the first column being 1 and the second column being 0.

[0121] Right now:

[0122]

[0123]

[0124]

[0125] That is, the first parameter b is equal to the third parameter c, while the second parameter a is different from the first parameter b and the third parameter c.

[0126] Based on the classification principle of Softmax, the target encoding method is used to encode the first parameter, the second parameter and the third parameter, so that the second parameter is equivalent to 0, and the first parameter and the third parameter are equivalent to 1. Correspondingly, the first loss function and the second loss function can be equivalent to:

[0127]

[0128]

[0129] The optimization goal of the first loss function and the second loss function is to minimize the distribution difference between the test set and the training set. Since the second parameter a and the third parameter c are not equal, and the second parameter a is equivalent to 0, and the third parameter c is equivalent to 1, the first loss function and the second loss function are actually to make the probability distribution of the training set play a game between maximum and minimum to force the feature extraction network to establish domain invariant features.

[0130] When the feature extraction network is optimized based on the identification loss value of the improved least squares discriminator for the training set and the generation loss value of the improved least squares discriminator for the test set in the adversarial training so that the features output by the feature extraction network achieve the desired generation effect, that is, the identification loss value of the improved least squares discriminator for the training set is minimized, and the generation loss value of the improved least squares discriminator for the test set is minimized, then, according to the principle of the original least squares adversarial training, the optimization target expression of the improved least squares discriminator can be derived, which can be:

[0131]

[0132] According to the expression of the improved least squares discriminator mentioned above, the feature extraction network and the improved least squares discriminator compete with each other through adversarial training, which is equivalent to using the Pearson divergence to measure and minimize the distribution difference between the training set (source domain) and the test set (target domain), thereby achieving distribution alignment of different data sets.

[0133] Among them, a is the second parameter and b is the first parameter.

[0134] Using the solution disclosed in this embodiment to perform adversarial training, the optimization goal of adversarial training can be as follows:

[0135]

[0136] in, Represents the Pearson divergence. The optimization goal of adversarial training is actually and Performs the Pearson divergence measure of distributional differences.

[0137] The scheme disclosed in this embodiment sets the activation function of the improved least squares discriminator to the Softmax activation function. At the same time, the encoding method of the first parameter, the second parameter and the third parameter in the improved loss function can improve the shortcomings of poor distribution measurement ability and inability to continuously train when the traditional least squares is used as an adversarial loss. It can perform more accurate adversarial training on the training set and the test set, establish better domain-invariant features (data probability density distribution), and achieve efficient diagnostic experience transfer. In the adversarial training between the training set and the test set in the feature extraction network and the improved least squares discriminator, the distribution difference between the training set and the test set is minimized, so that the features extracted by the feature extraction network can be accurately classified, thereby ensuring the accuracy of bearing fault diagnosis.

[0138] like Figure 7a As shown, it is a schematic diagram of the probability distribution between the training set and the test set obtained by the adversarial training method disclosed in this embodiment, wherein the ordinate is the probability value, and the abscissa is the feature value in the parameter information of the bearing extracted by the feature extraction network; in addition, as Figure 7b The figure is a schematic diagram of the probability distribution between the training set and the test set obtained by the adversarial training method in the prior art, wherein the ordinate is the probability value and the abscissa is the feature value in the bearing parameter information extracted by the feature extraction network. Figure 7b As shown, Figure 7a In the schematic diagram of the probability distribution shown in , the distribution difference between the training set and the test set is smaller, and there is a greater migration advantage, which shows that the adversarial training method is more accurate and has a higher accuracy in diagnosing bearing fault migration.

[0139] The bearing fault diagnosis method disclosed in this embodiment, when training the feature extraction network, divides the obtained historical data information of the bearing into a training set and a test set, and performs feature extraction on each set, and inputs the extracted features into an improved least squares discriminator including a Softmax activation function to determine the first probability distribution of the features corresponding to the training set and the second probability distribution of the features corresponding to the test set, and uses the loss function and the first probability distribution and the second probability distribution to determine the loss function value, so as to adjust the weight parameters in the feature extraction network based on the loss function value, thereby optimizing the feature extraction network, achieving more accurate adversarial training on the data distribution of the training set and the test set, establishing better domain-invariant features, and achieving efficient migration of diagnostic experience.

[0140] This embodiment discloses a bearing fault diagnosis system, and its structural diagram is shown in FIG. Figure 8 As shown, including:

[0141] An acquisition unit 81 , an extraction unit 82 and a classification unit 83 .

[0142] Wherein, the obtaining unit 81 is used to obtain parameter information of the bearing to be diagnosed;

[0143] The extraction unit 82 is used to use the first feature extraction network to perform feature extraction on the parameter information of the bearing to be diagnosed, so as to obtain bearing feature information;

[0144] The classification unit 83 is used to classify the bearing feature information using the trained diagnosis model to determine whether the bearing to be diagnosed has a fault;

[0145] Among them, the trained diagnostic model is: the diagnostic model is trained using the characteristic information of a bearing in a known state, and the weight parameters of the neural network of the diagnostic model are optimized and adjusted to obtain the trained diagnostic model; the first feature extraction network weight parameters are obtained by using the output result of the initial feature extraction network as the input information of the improved least squares discriminator, and based on the input information, the initial feature extraction network and the improved least squares discriminator are subjected to adversarial training and optimization.

[0146] Furthermore, the output layer of the improved least squares discriminator includes a Softmax activation function.

[0147] Furthermore, the bearing fault system disclosed in this embodiment may also include:

[0148] An optimization unit is used to obtain historical data information of a bearing; obtain a training set and a test set from the historical data information, wherein the training set is historical data information of a bearing of a known fault type, and the test set is data information of a bearing of an unknown state type; based on an initial feature extraction network, feature extraction is performed on the training set to obtain first feature information, and feature extraction is performed on the test set to obtain second feature information; using a Softmax activation function and a loss function in an improved least squares discriminator, based on the first feature information and the second feature information, weight parameters in the initial feature extraction network are adjusted to obtain a first feature extraction network.

[0149] Furthermore, the optimization unit is used to:

[0150] Based on the Softmax activation function in the improved least squares discriminator, determine the first probability distribution of the first feature information in the improved least squares discriminator and the second probability distribution of the second feature information in the improved least squares discriminator; use the loss function to determine the loss function value based on the first probability distribution and the second probability distribution; adjust the weight parameters in the feature extraction network based on the loss function value to obtain the first feature extraction network.

[0151] Furthermore, the optimization unit is used to:

[0152] Determine a first loss function including a first parameter and a second parameter, and a second loss function including a third parameter; use the first loss function to determine the identification loss value corresponding to the initial feature extraction network based on the first mean square error and the second mean square error, wherein the first mean square error is the mean square error between the first probability distribution and the first parameter, the second mean square error is the mean square error between the second probability distribution and the second parameter, and the identification loss value is used to characterize the difference between the third probability distribution of the training set and the expected target distribution in adversarial training, and the third probability distribution is the probability distribution of the training set output after the initial feature extraction network and the improved least squares discriminator; use the second loss function to determine the minimum generation loss value based on the third mean square error, wherein the third mean square error is the mean square error between the second probability distribution and the third parameter, and the generation loss value is used to characterize the difference between the fourth probability distribution of the test set and the expected target distribution in adversarial training; the fourth probability distribution is the probability distribution of the test set output after the initial feature extraction network and the improved least squares discriminator; the expected target distribution characterizes the expected distribution corresponding to the encoding parameters in adversarial training.

[0153] Furthermore, the optimization unit is used to:

[0154] The weight parameters in the initial feature extraction network are adjusted based on the loss function value; in the process of adjusting the weight parameters of the initial feature extraction network multiple times, the loss function value corresponding to the second feature extraction network obtained after each adjustment is determined; based on the loss function value corresponding to the second feature extraction network obtained after each adjustment, a second feature extraction network is determined as the first feature extraction network from the second feature extraction networks obtained after each adjustment.

[0155] Furthermore, the optimization unit is used to:

[0156] From the loss function values ​​corresponding to the second feature extraction network obtained after each adjustment, the second feature extraction network corresponding to the minimum loss function value is selected, and the selected second feature extraction network is determined as the first feature extraction network.

[0157] Furthermore, the optimization unit is used to:

[0158] Determine a target encoding method; encode the first parameter, the second parameter and the third parameter based on the target encoding method so that the first parameter, the second parameter and the third parameter are two-dimensional vectors; determine a first loss function including the first parameter and the second parameter, and determine a second loss function including the third parameter.

[0159] The bearing fault diagnosis system disclosed in this embodiment is implemented based on the bearing fault diagnosis method disclosed in the above embodiment, which will not be described in detail here.

[0160] The bearing fault diagnosis system disclosed in this embodiment obtains parameter information of the bearing to be diagnosed; uses the first feature extraction network to extract features from the parameter information of the bearing to be diagnosed to obtain bearing feature information; uses the trained diagnostic model to classify the bearing feature information to determine whether the bearing to be diagnosed has a fault; wherein the trained diagnostic model is: the diagnostic model is trained using feature information of a bearing in a known state, and the weight parameters of the neural network of the diagnostic model are optimized and adjusted to obtain the trained diagnostic model; the weight parameters of the first feature extraction network are obtained by taking the output result of the initial feature extraction network as the input information of the improved least squares discriminator, and based on the input information, the initial feature extraction network and the improved least squares discriminator are trained and optimized in adversarial manner. This scheme uses the first feature extraction network to extract features from the parameter information of the bearing to be diagnosed, and uses the diagnostic model to classify and train it to determine whether the bearing to be diagnosed has a fault. The weight parameters of the first feature extraction network are obtained by taking the output result of the initial feature extraction network as the input information of the improved least squares discriminator, and based on the input information, the initial feature extraction network and the improved least squares discriminator are trained and optimized in adversarial manner, and the first feature extraction network obtained after optimization improves the accuracy of fault diagnosis of the bearing.

[0161] This embodiment discloses an electronic device, and its structural diagram is as follows: Fig. 9 As shown, including:

[0162] Processor 91 and memory 92.

[0163] The processor 91 is used to obtain parameter information of the bearing to be diagnosed; the first feature extraction network is used to extract features from the parameter information of the bearing to be diagnosed to obtain bearing feature information; the bearing feature information is classified using the trained diagnostic model to determine whether the bearing to be diagnosed has a fault; the trained diagnostic model is: the diagnostic model is classified and trained using feature information of a bearing in a known state, and the weight parameters of the neural network of the diagnostic model are optimized and adjusted to obtain the trained diagnostic model; the weight parameters of the first feature extraction network are obtained by taking the output result of the initial feature extraction network as input information of the improved least squares discriminator, and performing adversarial training and optimization on the initial feature extraction network and the improved least squares discriminator based on the input information;

[0164] The memory 92 is used to store the programs required by the processor to execute the above processing procedures.

[0165] The electronic device disclosed in this embodiment is implemented based on the bearing fault diagnosis method disclosed in the above embodiment, which will not be described in detail here.

[0166] The electronic device disclosed in this embodiment obtains parameter information of a bearing to be diagnosed; uses a first feature extraction network to extract features from the parameter information of the bearing to be diagnosed to obtain bearing feature information; uses a trained diagnostic model to classify the bearing feature information to determine whether the bearing to be diagnosed has a fault; wherein the trained diagnostic model is: the diagnostic model is trained using feature information of a bearing in a known state, and the weight parameters of the neural network of the diagnostic model are optimized and adjusted to obtain a trained diagnostic model; the weight parameters of the first feature extraction network are obtained by taking the output result of the initial feature extraction network as input information of the improved least squares discriminator, and based on the input information, the initial feature extraction network and the improved least squares discriminator are trained and optimized in adversarial manner. This solution uses the first feature extraction network to extract features from the parameter information of the bearing to be diagnosed, and uses the diagnostic model to classify and train it to determine whether the bearing to be diagnosed has a fault. The weight parameters of the first feature extraction network are obtained by taking the output result of the initial feature extraction network as input information of the improved least squares discriminator, and based on the input information, the initial feature extraction network and the improved least squares discriminator are trained and optimized in adversarial manner, and the first feature extraction network obtained after optimization improves the accuracy of fault diagnosis of the bearing.

[0167] The embodiment of the present application also provides a readable storage medium on which a computer program is stored. The computer program is loaded and executed by a processor to implement the various steps of the above-mentioned bearing fault diagnosis method. The specific implementation process can refer to the description of the corresponding part of the above-mentioned embodiment, which will not be repeated in this embodiment.

[0168] The present application also proposes a computer program product or a computer program, which includes a computer instruction, and the computer instruction is stored in a computer-readable storage medium. The processor of the electronic device reads the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the electronic device executes the method provided in various optional implementations of the above-mentioned bearing fault diagnosis method or bearing fault diagnosis system. The specific implementation process can refer to the description of the above-mentioned corresponding embodiment, and will not be repeated.

[0169] It should also be noted that the device embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed over multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, in the drawings of the device embodiments provided by the present application, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines.

[0170] Through the description of the above implementation mode, the technicians in the field can clearly understand that the present application can be implemented by means of software plus necessary general hardware, and of course, it can also be implemented by special hardware including special integrated circuits, special CPUs, special memories, special components, etc. In general, all functions completed by computer programs can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be various, such as analog circuits, digital circuits or special circuits. However, for the present application, software program implementation is a better implementation mode in more cases. Based on such an understanding, the technical solution of the present application is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a readable storage medium, such as a computer floppy disk, a U disk, a mobile hard disk, a ROM, a RAM, a disk or an optical disk, etc., including a number of instructions to enable a computer device (which can be a personal computer, a training device, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0171] In the above embodiments, all or part of the embodiments may be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments may be implemented in the form of a computer program product.

[0172] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from a website site, a computer, a training device, or a data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode to another website site, computer, training device, or data center. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device, a data center, etc. that includes one or more available media integrations. The available medium may be a magnetic medium, (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)), etc.

Claims

1. A bearing fault diagnosis method, characterized in that: include: Obtain parameter information of the bearing to be diagnosed; Using a first feature extraction network, feature extraction is performed on parameter information of the bearing to be diagnosed to obtain bearing feature information; Using the trained diagnostic model to classify the bearing feature information, and determine whether the bearing to be diagnosed has a fault; The trained diagnostic model is obtained by training the diagnostic model using feature information of a bearing in a known state and adjusting the neural network weight parameters of the diagnostic model; the weight parameters of the first feature extraction network are obtained by taking the output result of the initial feature extraction network as input information of an improved least squares discriminator, and performing adversarial training optimization on the initial feature extraction network and the improved least squares discriminator based on the input information; Wherein, the output layer of the improved least squares discriminator includes a Softmax activation function; Among them, it also includes: Obtain historical data information of bearings; Obtaining a training set and a test set from the historical data information, wherein the training set is historical data information of bearings with known fault types, and the test set is data information of bearings with unknown status types; Performing feature extraction on the training set based on the initial feature extraction network to obtain first feature information, and performing feature extraction on the test set to obtain second feature information; Using the Softmax activation function and the loss function in the improved least squares discriminator, based on the first feature information and the second feature information, adjusting the weight parameters in the initial feature extraction network to obtain a first feature extraction network; The method of using the Softmax activation function and the loss function in the improved least squares discriminator to adjust the weight parameters in the initial feature extraction network based on the first feature information and the second feature information to obtain a first feature extraction network includes: Based on the Softmax activation function in the improved least squares discriminator, determining a first probability distribution of the first feature information in the improved least squares discriminator and a second probability distribution of the second feature information in the improved least squares discriminator; Determine a loss function value based on the first probability distribution and the second probability distribution using a loss function; The weight parameters in the initial feature extraction network are adjusted based on the loss function value to obtain a first feature extraction network.

2. The method according to claim 1, characterized in that The using the loss function to determine the loss function value based on the first probability distribution and the second probability distribution includes: Determine a first loss function including the first parameter and the second parameter, and a second loss function including the third parameter; Using the first loss function, based on the first mean square error and the second mean square error, determining the identification loss value corresponding to the initial feature extraction network, wherein the first mean square error is the mean square error between the first probability distribution and the first parameter, the second mean square error is the mean square error between the second probability distribution and the second parameter, and the identification loss value is used to characterize the difference between the third probability distribution of the training set and the expected target distribution in adversarial training, the third probability distribution being the probability distribution of the training set output by the initial feature extraction network and the improved least squares discriminator; Utilizing the second loss function, a generation loss value corresponding to the initial feature extraction network is determined based on a third mean square error, wherein the third mean square error is the mean square error between the second probability distribution and the third parameter, and the generation loss value is used to characterize the difference between the fourth probability distribution of the test set and the expected target distribution in adversarial training; the fourth probability distribution is the probability distribution of the test set output after passing through the initial feature extraction network and the improved least squares discriminator; the expected target distribution characterizes the expected distribution corresponding to the encoding parameters in adversarial training.

3. The method according to claim 1, characterized in that The step of adjusting the weight parameters in the initial feature extraction network based on the loss function value to obtain a first feature extraction network includes: Based on the loss function value, adjusting the weight parameters in the initial feature extraction network; In the process of adjusting the weight parameters of the initial feature extraction network multiple times, determining the loss function value corresponding to the second feature extraction network obtained after each adjustment; Based on the loss function value corresponding to the second feature extraction network obtained after each adjustment, a second feature extraction network is determined as the first feature extraction network from the second feature extraction networks obtained after each adjustment.

4. The method according to claim 3, characterized in that: The step of determining a second feature extraction network as the first feature extraction network from the second feature extraction networks obtained after each adjustment based on the loss function value corresponding to the second feature extraction network obtained after each adjustment includes: From the loss function values ​​corresponding to the second feature extraction networks obtained after each adjustment, the second feature extraction network corresponding to the minimum loss function value is selected, and the selected second feature extraction network is determined as the first feature extraction network.

5. The method according to claim 2, characterized in that: The determining of a first loss function including a first parameter and a second parameter, and a second loss function including a third parameter, comprises: Determine the target encoding method; Encoding the first parameter, the second parameter and the third parameter based on the target encoding method so that the first parameter, the second parameter and the third parameter are two-dimensional vectors; A first loss function including the first parameter and the second parameter is determined, and a second loss function including a third parameter is determined.

6. A bearing fault diagnosis system, characterized in that: include: An acquisition unit, used for acquiring parameter information of the bearing to be diagnosed; An extraction unit, configured to extract features from the parameter information of the bearing to be diagnosed by using a first feature extraction network to obtain bearing feature information; A classification unit, used to classify the bearing feature information using the trained diagnosis model to determine whether the bearing to be diagnosed has a fault; The trained diagnostic model is obtained by training the diagnostic model using feature information of a bearing in a known state and adjusting weight parameters of a neural network of the diagnostic model; the weight parameters of the first feature extraction network are obtained by taking the output result of the initial feature extraction network as input information of an improved least squares discriminator, and performing adversarial training optimization on the initial feature extraction network and the improved least squares discriminator based on the input information; the output layer of the improved least squares discriminator includes a Softmax activation function; Among them, historical data information of the bearing is obtained; a training set and a test set in the historical data information are obtained, wherein the training set is the historical data information of the bearing of known fault type, and the test set is the data information of the bearing of unknown state type; based on the initial feature extraction network, feature extraction is performed on the training set to obtain first feature information, and feature extraction is performed on the test set to obtain second feature information; using the Softmax activation function and loss function in the improved least squares discriminator, based on the first feature information and the second feature information, the weight parameters in the initial feature extraction network are adjusted to obtain the first feature extraction network; The method of using the Softmax activation function and the loss function in the improved least squares discriminator to adjust the weight parameters in the initial feature extraction network based on the first feature information and the second feature information to obtain a first feature extraction network includes: determining a first probability distribution of the first feature information in the improved least squares discriminator and a second probability distribution of the second feature information in the improved least squares discriminator based on the Softmax activation function in the improved least squares discriminator; using the loss function to determine a loss function value based on the first probability distribution and the second probability distribution; and adjusting the weight parameters in the initial feature extraction network based on the loss function value to obtain a first feature extraction network.

7. An electronic device, characterized in that: include: A processor is used to obtain parameter information of a bearing to be diagnosed; use a first feature extraction network to extract features from the parameter information of the bearing to be diagnosed to obtain bearing feature information; use a trained diagnostic model to classify the bearing feature information to determine whether the bearing to be diagnosed has a fault; wherein the trained diagnostic model is: the diagnostic model is trained using feature information of a bearing in a known state, and the weight parameters of the neural network of the diagnostic model are adjusted to obtain the trained diagnostic model; the weight parameters of the first feature extraction network are: the output result of the initial feature extraction network is used as the input information of the improved least squares discriminator, and based on the input information, the initial feature extraction network and the improved least squares discriminator are subjected to adversarial training optimization. ; wherein the output layer of the improved least squares discriminator includes a Softmax activation function; wherein the historical data information of the bearing is obtained; the training set and the test set in the historical data information are obtained, wherein the training set is the historical data information of the bearing with a known fault type, and the test set is the data information of the bearing with an unknown state type; based on the initial feature extraction network, feature extraction is performed on the training set to obtain the first feature information, and feature extraction is performed on the test set to obtain the second feature information; using the Softmax activation function and the loss function in the improved least squares discriminator, based on the first feature information and the second feature information, the weight parameters in the initial feature extraction network are adjusted to obtain the first feature extraction network; The method of using the Softmax activation function and the loss function in the improved least squares discriminator to adjust the weight parameters in the initial feature extraction network based on the first feature information and the second feature information to obtain a first feature extraction network includes: determining a first probability distribution of the first feature information in the improved least squares discriminator and a second probability distribution of the second feature information in the improved least squares discriminator based on the Softmax activation function in the improved least squares discriminator; using the loss function to determine a loss function value based on the first probability distribution and the second probability distribution; and adjusting the weight parameters in the initial feature extraction network based on the loss function value to obtain a first feature extraction network. The memory is used to store the program required by the processor to execute the above processing process.

Images