A machine-cross fault diagnosis method, device, equipment and medium

Abstract

The application discloses a cross-machine fault diagnosis method and device, equipment and medium, and relates to the technical field of mechanical fault diagnosis. The method comprises the following steps: acquiring source domain vibration signal data and target domain vibration signal data; obtaining a source domain training sample set, a target domain labeled training sample set and a target domain test sample set according to the source domain vibration signal data and the target domain vibration signal data; inputting the sample features in the source domain training sample set and the target domain labeled training sample set into a feature extractor, inputting the obtained result into a classifier again, obtaining a classification result; training the feature extractor and the classifier based on the classification result and a health state category, obtaining a trained feature extractor and a trained classifier; inputting the target domain test sample set into the trained feature extractor and the trained classifier, obtaining a plurality of preliminary diagnosis results; and performing fusion processing on the plurality of preliminary diagnosis results, obtaining a target diagnosis result. The method can improve the accuracy of cross-machine fault diagnosis.

Description

Technical Field

[0001] This application relates to the field of mechanical fault diagnosis technology, and in particular to a cross-machine fault diagnosis method, apparatus, equipment and medium. Background Technology

[0002] Rotating machinery is a crucial component of major equipment such as aircraft engines, wind turbines, and high-speed trains, and its health status directly affects the safety of equipment operation. Because most equipment in real-world industrial scenarios operates in a healthy state for extended periods, the number of fault samples, especially labeled samples, is limited. Furthermore, the characteristic distribution differences caused by varying operating conditions, loads, and models among different devices lead to a decrease in diagnostic accuracy when applied to other operating conditions or equipment.

[0003] To address the aforementioned issues, existing technologies primarily employ domain-adaptive transfer learning methods, such as maximum mean difference or adversarial training, to reduce the feature distribution differences between the source and target domains. However, existing methods suffer from significant biases in estimating the target domain distribution under small sample conditions, leading to low accuracy in cross-machine fault diagnosis. Summary of the Invention

[0004] This application provides a method, apparatus, device, and medium for cross-machine fault diagnosis, which can improve the accuracy of cross-machine fault diagnosis.

[0005] To achieve the above objectives, this application adopts the following technical solution: In a first aspect, this application provides a cross-machine fault diagnosis method, the method comprising: Acquire source domain vibration signal data of the first machine bearing and target domain vibration signal data of the second machine bearing. Based on the source domain vibration signal data of the first machine bearing, a source domain training sample set is obtained, which includes labeled samples under various health conditions. Based on the target domain vibration signal data of the second machine bearing, a labeled training sample set and a target domain test sample set are obtained. Input the samples from the source domain training sample set and the target domain labeled training sample set into the feature extractor to extract intermediate features; The intermediate features are input into the classifier to obtain the classification result; Based on the classification results and health status categories, the feature extractor and the classifier are jointly optimized and trained to obtain the trained feature extractor and the trained classifier. The target domain test sample set is input into the trained feature extractor and the trained classifier to obtain multiple preliminary diagnostic results; Multiple preliminary diagnostic results are fused together to obtain the target diagnostic result.

[0006] In some possible implementations, the classification result includes the classification result of the source domain training sample set and the classification result of the target domain labeled training sample set. The step of jointly optimizing and training the feature extractor and the classifier based on the classification result and the health status category to obtain the trained feature extractor and the trained classifier includes: Based on the classification results and health status categories of the source domain training sample set, calculate the source domain cross-entropy loss; based on the classification results and health status categories of the target domain labeled training sample set, calculate the target domain cross-entropy loss; weight the source domain cross-entropy loss, the target domain cross-entropy loss, and the improved Softmax loss to obtain the total loss value; update the network parameters of the feature extractor and the classifier based on the total loss value to obtain the trained feature extractor and the trained classifier.

[0007] In some possible implementations, the process of fusing multiple preliminary diagnostic results to obtain the target diagnostic result includes: The preliminary diagnostic results are transformed into evidence in the form of basic probability allocation; a metric is calculated between any two pieces of evidence; the average divergence of each piece of evidence is calculated based on the metric between all two pieces of evidence; the support of each piece of evidence is calculated based on the average divergence of each piece of evidence; the credibility of each piece of evidence is calculated based on the support of each piece of evidence; a weighted average of all evidence is calculated based on the credibility of each piece of evidence to obtain a weighted average evidence; the weighted average evidence is fused to obtain the target diagnostic result.

[0008] In some possible implementations, a metric is calculated between any two pieces of evidence, including: In the identification framework, select any hypothesis and obtain the first and second confidence values ​​corresponding to the hypothesis from two pieces of evidence; obtain the magnitude difference element value based on the absolute value of the difference between the first and second confidence values ​​and the sum of the absolute values ​​of the first and second confidence values; obtain the direction cosine value based on the ratio of the product of the first and second confidence values ​​and the product of the magnitudes of the confidence vectors of the two pieces of evidence; obtain the metric value between the two pieces of evidence based on the magnitude difference element value and the direction cosine value.

[0009] In some possible implementations, the feature extractor includes multiple cascaded convolutional modules, a global average pooling layer, and a flattening layer. Each convolutional module includes a convolutional layer, a batch normalization layer, and a max pooling layer. The classifier includes a first fully connected layer and a second fully connected layer. The input of the first fully connected layer is connected to the output of the flattening layer, and the output of the first fully connected layer is connected to the input of the second fully connected layer.

[0010] In some possible implementations, a source domain training sample set is obtained based on the source domain vibration signal data of the first machine bearing, including: The source domain vibration signal data of the first machine bearing is sampled to obtain the sampled first signal; the sampled signal is divided into multiple samples, and all samples are used as the source domain training sample set.

[0011] In some possible implementations, based on the target domain vibration signal data of the second machine bearing, a labeled training sample set and a target domain test sample set are obtained, including: The vibration signal data of the target domain of the second machine bearing is sampled to obtain the sampled second signal; the sampled signal is divided into multiple samples, and a portion of the samples under each health state is used as the labeled training sample set of the target domain; the remaining samples are used as the test sample set of the target domain.

[0012] Secondly, this application provides a cross-machine fault diagnosis device, the device comprising: The acquisition module is used to acquire source domain vibration signal data of the first machine bearing and target domain vibration signal data of the second machine bearing; based on the source domain vibration signal data of the first machine bearing, a source domain training sample set is obtained, which includes labeled samples under various health conditions; based on the target domain vibration signal data of the second machine bearing, a target domain labeled training sample set and a target domain test sample set are obtained. The training module is used to input samples from the source domain training sample set and the target domain labeled training sample set into a feature extractor to extract intermediate features; input the intermediate features into a classifier to obtain a classification result; and perform joint optimization training on the feature extractor and the classifier based on the classification result and the health status category to obtain a trained feature extractor and a trained classifier. The output module is used to input the target domain test sample set into the trained feature extractor and the trained classifier to obtain multiple preliminary diagnostic results; and to fuse the multiple preliminary diagnostic results to obtain the target diagnostic result.

[0013] Thirdly, this application provides a computing device, including a memory and a processor; The memory stores one or more computer programs, the one or more computer programs including instructions; when the instructions are executed by the processor, the computing device performs the method as described in any one of the first aspects.

[0014] Fourthly, this application provides a computer-readable storage medium for storing a computer program for performing the method as described in any one of the first aspects.

[0015] Fifthly, this application provides a computer program product comprising one or more computer instructions, wherein when the computer instructions are executed by a computer, the computer performs the method as described in any one of the first aspects.

[0016] As can be seen from the above technical solution, this application has at least the following beneficial effects: In this application, source domain vibration signal data of a first machine bearing and target domain vibration signal data of a second machine bearing are acquired. Based on the source domain vibration signal data of the first machine bearing, a source domain training sample set is obtained. Based on the target domain vibration signal data of the second machine bearing, a target domain labeled training sample set and a target domain test sample set are obtained. Samples from the source domain training sample set and the target domain labeled training sample set are input into a feature extractor to extract intermediate features. The intermediate features are input into a classifier to obtain classification results. Based on the classification results and health status categories, the feature extractor and classifier are jointly optimized and trained to obtain a trained feature extractor and a trained classifier. The target domain test sample set is input into the trained feature extractor and the trained classifier to obtain multiple preliminary diagnostic results. The multiple preliminary diagnostic results are fused to obtain the target diagnostic result. In existing technical solutions, the domain-adaptive transfer learning method has a large deviation in target domain distribution estimation under small sample conditions, and the information fusion strategy is not reasonable enough, resulting in low accuracy in cross-machine fault diagnosis. As can be seen, this application increases the inter-class feature distance and reduces the intra-class feature dispersion in the angle space by jointly optimizing the improved Softmax loss function in the training, enabling the model to learn discriminative features under small sample conditions. Furthermore, by fusing multiple preliminary diagnostic results, the accuracy of cross-machine fault diagnosis can be effectively improved.

[0017] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0018] Figure 1 A schematic diagram illustrating an application scenario provided in an embodiment of this application; Figure 2 A flowchart illustrating a cross-machine fault diagnosis method provided in this application embodiment; Figure 3 A schematic diagram of the overall model structure of an STLN-IS supervised transfer learning network provided in this application embodiment; Figure 4 The confusion matrix diagram of fault diagnosis results when migrating the CWRU dataset to the IMS dataset under 0hp conditions, provided in the embodiments of this application; Figure 5 The confusion matrix diagram of fault diagnosis results when migrating the CWRU dataset to the IMS dataset under 1hp operating conditions, provided in the embodiments of this application; Figure 6 The confusion matrix diagram of fault diagnosis results for migrating the CWRU dataset to the IMS dataset under 2hp operating conditions provided in the embodiments of this application; Figure 7 The confusion matrix diagram of fault diagnosis results for migrating the CWRU dataset to the IMS dataset under 3hp operating conditions provided in the embodiments of this application; Figure 8 A bar chart comparing the diagnostic accuracy of various fault diagnosis methods provided in the embodiments of this application on the CWRU→IMS cross-machine diagnostic task; Figure 9 A schematic diagram of a cross-machine fault diagnosis device provided in an embodiment of this application; Figure 10 This is a schematic diagram of a computing device provided in an embodiment of this application. Detailed Implementation

[0019] The terms "first," "second," and "third," etc., used in this application specification and accompanying drawings are used to distinguish different objects, not to limit a specific order.

[0020] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0021] Currently, rotating machinery (such as bearings and gears) are key components of major equipment such as aero-engines, wind turbines, and high-speed trains, and their health status directly affects the operational safety of the equipment. Deep learning has already been applied in the field of mechanical fault diagnosis. In actual industrial scenarios, equipment usually operates in a healthy state, and the number of fault samples and labeled samples is limited. At the same time, the operating conditions of equipment vary, and the feature distributions of different devices differ, causing the diagnostic model trained based on data from one operating condition or device to have decreased diagnostic accuracy when applied to other operating conditions or devices.

[0022] Existing methods suffer from significant biases in estimating the target domain distribution under small sample conditions, and their information fusion strategies are not reasonable enough, resulting in low accuracy in cross-machine fault diagnosis.

[0023] In view of this, the present application provides a cross-machine fault diagnosis method. Compared with the prior art, the present application uses both the source domain training sample set and the target domain labeled training sample set to train the feature extractor and classifier, so that the model can learn source domain knowledge and a small amount of target domain labeled information at the same time. Then, by fusing multiple preliminary diagnosis results, the accuracy of cross-machine fault diagnosis can be improved under the condition of using only a small number of target domain labeled samples.

[0024] To make the technical solution of this application clearer and easier to understand, the application scenarios of the technical solution of this application are described below with reference to the accompanying drawings. Figure 1 As shown in the figure, this figure is a schematic diagram of an application scenario provided by an embodiment of this application.

[0025] In this application scenario, the terminal is a vibration sensor or data acquisition device deployed at the first and second machine bearings, used to collect source domain vibration signal data of the first machine bearing and target domain vibration signal data of the second machine bearing; the communication network is used to transmit the vibration signal data collected by the terminal to the server; the server is equipped with a cross-machine fault diagnosis model framework. After receiving the vibration signal data, the server uses the source domain vibration signal data as the source domain training sample set, divides the target domain vibration signal data into a target domain labeled training sample set and a target domain test sample set, uses the source domain training sample set and the target domain labeled training sample set to jointly optimize and train the cross-machine fault diagnosis model, and then uses the trained cross-machine fault diagnosis model to perform fault diagnosis on the target domain test sample set, outputting the target diagnosis result.

[0026] To this end, this application uses both the source domain training sample set and the target domain labeled training sample set to train the feature extractor and classifier, enabling the model to learn source domain knowledge and a small amount of labeled information in the target domain simultaneously. Then, by fusing multiple preliminary diagnostic results, the accuracy of cross-machine fault diagnosis can be improved using only a small number of labeled samples in the target domain.

[0027] To make the technical solution of this application clearer and easier to understand, the following describes a cross-machine fault diagnosis method provided by an embodiment of this application, in conjunction with the above application scenarios. Figure 2 As shown, this figure is a flowchart of a cross-machine fault diagnosis method provided in an embodiment of this application. The method includes: S201. Obtain the source domain vibration signal data of the first machine bearing and the target domain vibration signal data of the second machine bearing.

[0028] Source domain vibration signal data and target domain vibration signal data are two data sources in cross-machine fault diagnosis. Source domain vibration signal data refers to vibration signals collected from the bearings of the first machine; this data contains sufficient samples under various health conditions, and each sample is labeled with a health condition category. Target domain vibration signal data refers to vibration signals collected from the bearings of the second machine; only a portion of these samples are labeled with a health condition category, while the remaining samples are not.

[0029] The first and second machine bearings are two different rotating mechanical bearings. The first machine bearing is the source device for the source domain data, which operates under various health conditions and collects sufficient vibration signals. The second machine bearing is the source device for the target domain data; it is the machine bearing to be diagnosed, and only a small number of its vibration signals have health status category labels.

[0030] Multiple health states refer to the different categories of states that a bearing may experience during operation, including normal states and different types of fault states.

[0031] Vibration sensors deployed at the first and second machine bearings collect source domain vibration signal data and target domain vibration signal data, respectively. The collected source domain vibration signal data is transmitted to the server via a communication network. The collected target domain vibration signal data is also transmitted to the server via a communication network.

[0032] S202. Based on the source domain vibration signal data of the first machine bearing, obtain the source domain training sample set. Based on the target domain vibration signal data of the second machine bearing, obtain the target domain labeled training sample set and the target domain test sample set.

[0033] A sample refers to each data unit obtained after sampling and dividing the original vibration signal. Each sample contains a sequence of vibration signal values ​​from consecutive sampling points, and this sequence of values ​​constitutes a complete input data unit.

[0034] A labeled sample is one that has been pre-labeled with its corresponding health status category. All samples in the source domain training sample set are labeled samples. All samples in the target domain labeled training sample set are labeled samples. Samples in the target domain test sample set are not labeled with their health status categories.

[0035] The training and test datasets are two datasets used for model training and testing, respectively. The training dataset is used to train the feature extractor and classifier. The test dataset is used to evaluate the diagnostic performance of the trained feature extractor and classifier.

[0036] In this step, the server samples the received source domain vibration signal data of the first machine bearing to obtain a sampled first signal. The server divides the sampled first signal into multiple samples, each containing a sequence of vibration signal values ​​from consecutive sampling points. The server uses all the divided samples as the source domain training sample set. Each sample in the source domain training sample set is labeled with its corresponding health state category.

[0037] The server samples the received vibration signal data of the second machine bearing in the target domain, obtaining a sampled second signal. The server divides the sampled second signal into multiple samples, each containing a sequence of vibration signal values ​​from consecutive sampling points. From all the divided samples, the server selects a portion of samples under each health state as the labeled training sample set for the target domain. Each sample in the labeled training sample set is labeled with its corresponding health state category. The server uses the remaining samples in the target domain vibration signal data, excluding the labeled training samples, as the target domain test sample set. The samples in the target domain test sample set are not labeled with their health state categories.

[0038] In one embodiment, the source domain vibration signal data is taken from the drive end data of the bearing dataset from Case Western Reserve University, with a sampling frequency of 12kHz, corresponding to four operating conditions: 0hp, 1hp, 2hp, and 3hp loads. To simulate small-sample scenarios in actual engineering, 100 samples of the original vibration signals under 10 healthy states are selected, with each sample containing 1024 sampling points. The server selects 10 samples from the 100 samples under each healthy state as the source domain training sample set, and uses the remaining 90 samples as the source domain test samples. The target domain vibration signal data is taken from the bearing dataset from the University of Cincinnati, with a sampling frequency of 20kHz. The normal state is taken from the early operating stage of bearing 1 in test series 1, the inner ring failure is taken from the failure stage of bearing 3 in test series 1, the outer ring failure is taken from the failure stage of bearing 1 in test series 2, and the rolling element failure is taken from the failure stage of bearing 4 in test series 1. The signal processing method for the target domain vibration signal data is the same as that for the source domain vibration signal data, with each sample containing 1024 sampling points. The server selects a subset of samples from each health state of the vibration signal data in the target domain as the labeled training sample set for the target domain, and uses the remaining samples as the test sample set for the target domain.

[0039] S203. Input the samples from the source domain training sample set and the target domain labeled training sample set into the feature extractor to extract intermediate features; input the intermediate features into the classifier to obtain the classification result; based on the classification result and the health status category, jointly optimize and train the feature extractor and the classifier to obtain the trained feature extractor and the trained classifier.

[0040] A feature extractor is a neural network module used to extract intermediate features from an input vibration signal. The feature extractor consists of multiple cascaded convolutional modules, a global average pooling layer, and a flattening layer. Each convolutional module includes a convolutional layer, a batch normalization layer, and a max-pooling layer. The convolutional layers perform convolution operations on the input signal to extract local features. The batch normalization layer normalizes the output of the convolutional layers. The max-pooling layer downsamples the feature map to reduce its dimensionality. The global average pooling layer performs a global average on the feature map output from the last convolutional module. The flattening layer flattens the pooled feature vector into a one-dimensional vector.

[0041] A classifier is a neural network module used to classify intermediate features. A classifier consists of a first fully connected layer and a second fully connected layer. The input of the first fully connected layer is connected to the output of the flattened layer, and the output of the first fully connected layer is connected to the input of the second fully connected layer. The second fully connected layer outputs logical values ​​for each category.

[0042] Intermediate features are abstract vector representations obtained after mapping the original vibration signal through a feature extractor. The classification result is the logical value of each category output by the classifier based on the input intermediate features.

[0043] like Figure 3 As shown, the feature extractor and classifier together constitute STLN-IS (Supervised transfer learning network with improved softmax). Figure 3 The overall model structure of STLN-IS is shown. The left side represents the feature extractor, comprising five cascaded convolutional modules followed by a global average pooling layer and a flattening layer. The right side represents the classifier, including a first fully connected layer and a second fully connected layer. Each convolutional module sequentially performs convolution, batch normalization, and max pooling downsampling to progressively extract high-level abstract features from the input vibration signal. The number of convolutional kernels and the size of the output feature maps vary layer by layer to capture signal features at different scales. The global average pooling layer performs a global average of the feature map output from the last convolutional module, and the flattening layer flattens the pooled feature vector into a one-dimensional vector. These intermediate features are then input into the classifier for classification.

[0044] The server combines the source domain training sample set and the target domain labeled training sample set into a joint training set. The server then inputs the samples from the joint training set into the feature extractor. The feature extractor sequentially performs convolution operations, batch normalization, and max pooling downsampling on the input samples to obtain feature maps. The feature extractor then performs global averaging on the feature maps using a global average pooling layer to obtain pooled feature vectors. Finally, the feature extractor flattens the pooled feature vectors into one-dimensional vectors using a flattening layer, which serve as intermediate features.

[0045] The server inputs intermediate features into the classifier. The first fully connected layer of the classifier performs a linear transformation on the intermediate features and processes them through an activation function to obtain the output of the first fully connected layer. The second fully connected layer of the classifier performs a linear transformation on the output of the first fully connected layer, outputting the logical values ​​for each category as the classification result.

[0046] In one embodiment, the feature extractor includes five cascaded convolutional modules. The first convolutional module has a kernel size of 64×1, 64 kernels, a stride of 16, and no padding. It is followed by a batch normalization layer and a ReLU (Rectified Linear Unit) activation function. The max-pooling layer has a window size of 2×1 and a stride of 2, resulting in an output feature map size of 32×192. The second convolutional module has a kernel size of 3×1, 64 kernels, and a stride of 1. It is followed by a batch normalization layer and a ReLU activation function. The max-pooling layer has a window size of 2×1 and a stride of 2, resulting in an output feature map size of 64×142. The third convolutional module has a 3×1 kernel size, 128 kernels, and a stride of 1. It is followed by a batch normalization layer and a ReLU activation function. The max-pooling layer has a 2×1 window size and a stride of 2, resulting in an output feature map size of 128×117. The fourth convolutional module has a 3×1 kernel size, 256 kernels, and a stride of 1. It is followed by a batch normalization layer and a ReLU activation function. The max-pooling layer has a 2×1 window size and a stride of 2, resulting in an output feature map size of 256×104. The fifth convolutional module has a 3×1 kernel size, 512 kernels, and a stride of 1. It is followed by a batch normalization layer and a ReLU activation function. The max-pooling layer has a 2×1 window size and a stride of 2, resulting in an output feature map size of 512×98. Each of the five convolutional modules is followed by a global average pooling layer and a flattening layer. The classifier consists of two fully connected layers. The first fully connected layer has an input dimension of 512 and an output dimension of 512, and uses the ReLU activation function. The second fully connected layer has an input dimension of 512 and an output dimension of 4, and has no activation function, directly outputting the logical values ​​of each category.

[0047] The server calculates the source domain cross-entropy loss based on the classification results and health status categories of the source domain training sample set. The source domain cross-entropy loss measures the difference between the classifier's classification results for the source domain training sample set and the true health status categories of the source domain training sample set.

[0048] The server calculates the target domain cross-entropy loss based on the classification results and health status categories of the target domain labeled training sample set. The target domain cross-entropy loss measures the difference between the classifier's classification results and the true health status categories of the target domain labeled training sample set.

[0049] In one embodiment, the target domain labeled training sample set contains 20 labeled samples for each health state.

[0050] The server performs a weighted sum of the source domain cross-entropy loss, the target domain cross-entropy loss, and the improved Softmax loss to obtain the total loss value. The total loss value can be expressed as:

[0051] in, This is the total loss value. For source domain cross-entropy loss, For the target domain cross-entropy loss, The total number of samples in the source domain training sample set. Cross-entropy loss function measures the degree of difference between the classification output and the true label; For the i1th vibration signal sample in the source domain training sample set, Let i be the true health status label corresponding to the i1th sample in the source domain training sample set. The overall mapping function consisting of the feature extractor and the classifier. The total number of samples in the labeled training sample set for the target domain. For the j-th vibration signal sample in the labeled training sample set of the target domain, Let J be the true health status label corresponding to the j-th sample in the labeled training set for the target domain. For balance coefficient, To improve the Softmax loss.

[0052] The server updates the network parameters of the feature extractor and classifier based on the total loss value. Network parameters refer to the kernel weights and biases of each convolutional layer in the feature extractor, the scaling and translation coefficients of the batch normalization layer, and the weight matrices and bias vectors of the first and second fully connected layers in the classifier. The server calculates the partial derivative of the total loss value with respect to each network parameter in the feature extractor and classifier, obtaining the gradient value for each network parameter. The server subtracts the product of the gradient value and the preset learning rate from the current value of each network parameter to obtain the updated value of that network parameter. The server writes the updated value as the new value of the network parameter into the feature extractor and classifier. In one embodiment, the preset learning rate is 0.001.

[0053] The server recalculates the classification result based on the updated network parameters. It then recalculates the source domain cross-entropy loss, target domain cross-entropy loss, and improved Softmax loss based on the recalculated classification result. The server weights and sums these recalculated losses to obtain a new total loss value. The server determines whether the difference between the new total loss value and the total loss value calculated in the previous iteration is less than a preset threshold. If the difference is less than the preset threshold, the server determines that the total loss value has converged, stops the iteration, and uses the feature extractor and classifier corresponding to the current network parameters as the trained feature extractor and classifier. If the difference is greater than or equal to the preset threshold, the server continues to update the network parameters based on the new total loss value and repeats the above iterative process. In one embodiment, the preset threshold is 0.001.

[0054] Based on the cross-machine fault diagnosis method described in the above embodiments, the method simultaneously optimizes the source domain classification loss, target domain classification loss, and improved Softmax loss during joint training. This allows the feature extractor to learn various health status features of the source domain bearings during training, and also learns the feature distribution of the target domain bearings guided by a small number of labeled samples from the target domain. The improved Softmax loss forces an increase in the angular distance between features of different categories and a decrease in the dispersion between features of the same category in the angular space. When the total loss value converges, the intermediate features extracted by the trained feature extractor have a distribution characteristic of being dispersed between classes and compact within classes in the angular space. The trained classifier can accurately classify the target domain test samples based on these intermediate features.

[0055] S204. Input the target domain test sample set into the trained feature extractor and the trained classifier to obtain multiple preliminary diagnostic results.

[0056] The server inputs the target domain test sample set into the trained feature extractor. The trained feature extractor performs convolution, batch normalization, max pooling downsampling, global average pooling, and flattening operations on each sample in the target domain test sample set to extract the intermediate features corresponding to each sample. The trained feature extractor outputs the extracted intermediate features to the trained classifier.

[0057] The server inputs intermediate features into the trained classifier. The first fully connected layer of the trained classifier performs a linear transformation on the intermediate features and processes them through an activation function to obtain the output of the first fully connected layer. The second fully connected layer of the trained classifier performs a linear transformation on the output of the first fully connected layer, outputting logical values ​​for each category. Based on the logical values ​​of each category output by the second fully connected layer, the server determines the health status category of each test sample, obtaining multiple preliminary diagnostic results.

[0058] In one embodiment, a diagnostic model is constructed based on four operating conditions in the source domain, and each sample in the test sample set of the target domain is diagnosed to obtain four sets of preliminary diagnostic results for each sample. Each set of preliminary diagnostic results corresponds to a health status category determination result under one operating condition.

[0059] S205. The multiple preliminary diagnostic results are fused to obtain the target diagnostic result.

[0060] The server transforms the preliminary diagnostic results into evidence in the form of Basic Probability Assignments (BPAs). Each piece of evidence contains a confidence value assigned to each hypothesis in the identification framework based on the preliminary diagnostic result. For any preliminary diagnostic result, the server normalizes the result by applying a normalization process to transform it into a Basic Probability Assignment. This Basic Probability Assignment assigns a confidence value to each hypothesis in the identification framework, and the sum of the confidence values ​​of all hypotheses equals 1. The identification framework is a set of multiple mutually exclusive and complete hypotheses. In this embodiment, the hypotheses in the identification framework include normal states and different types of failure states. A hypothesis is a single element in the identification framework, representing a possible health state category. Evidence is a confidence assignment vector in the form of Basic Probability Assignments, containing a confidence value assigned to each hypothesis in the identification framework.

[0061] The server calculates a metric between any two pieces of evidence. Before calculating the metric, the server determines the number of hypotheses contained in the identification framework of the two pieces of evidence, denoted as 'a'. For the i2-th hypothesis in the identification framework, the server obtains the first and second confidence values ​​corresponding to the i2-th hypothesis from the two pieces of evidence.

[0062] The server obtains the magnitude difference element value based on the absolute value of the difference between the first and second confidence values, and the sum of the absolute values ​​of the first and second confidence values; it obtains the direction cosine value based on the ratio of the product of the first and second confidence values ​​and the product of the magnitudes of the confidence vectors of the two pieces of evidence; the server obtains the original metric value between the two pieces of evidence based on the magnitude difference element value and the direction cosine value, the calculation formula of which is:

[0063] in, The original measurement value between the first and second pieces of evidence; The first reliability value, This is the second reliability score. The basic probability allocation for the primary evidence. The basic probability allocation for the second piece of evidence. It is the L2 norm. To identify the i2th hypothesis in the framework, which corresponds to a bearing health state, 'a' represents the total number of hypotheses contained in the identification framework.

[0064] The server calculates the metric between any two pieces of evidence based on the original metric between them, using the following formula:

[0065] in, This is a measure between the first and second pieces of evidence. As the primary evidence and its own original measurement value, This is the second piece of evidence and its original measurement value.

[0066] Next, a divergence metric matrix is ​​constructed based on the metric values ​​between all pairs of evidence. The rows and columns of the divergence metric matrix correspond to individual pieces of evidence, and each element in the matrix represents the metric value between the evidence in the corresponding row and the evidence in the corresponding column. The divergence metric matrix can be represented as:

[0067] in, For the divergence metric matrix, This is the metric between the first piece of evidence and the i-th piece of evidence. This is the measure between the first piece of evidence and the j-th piece of evidence. Let i be the measure between the i-th piece of evidence and the first piece of evidence. Let i be the measure between the i-th piece of evidence and the n-th piece of evidence. This is the measure between the nth piece of evidence and the first piece of evidence. Let n be the metric between the nth piece of evidence and the ith piece of evidence, where n is the total number of pieces of evidence to be fused.

[0068] The server calculates the average divergence of each piece of evidence based on the divergence metric matrix. For the i-th piece of evidence, the server takes the arithmetic mean of all elements in the row vector corresponding to that evidence, excluding the diagonal elements, to obtain the average divergence of the i-th piece of evidence. Average divergence It can be represented as:

[0069] The server calculates the support score for each piece of evidence based on its average divergence. The server then takes the reciprocal of the average divergence of each piece of evidence to obtain its support score. Support It can be represented as:

[0070] The server calculates the credibility of each piece of evidence based on its support level. The server then divides the support level of each piece of evidence by the sum of the support levels of all other evidence to obtain the credibility of each piece of evidence. Credibility It can be represented as:

[0071] in, Let be the support level of the k-th piece of evidence.

[0072] The server calculates a weighted average of all evidence based on the credibility of each piece of evidence. Alternatively, the server uses the credibility of each piece of evidence as a weight to sum all the evidence, resulting in a weighted average. Weighted average evidence It can be represented as:

[0073] The server performs multiple fusions of the weighted average evidence to obtain the target diagnostic result. The server then performs multiple Dempster combination rules fusions of the weighted average evidence with itself, the number of fusions being one less than the number of pieces of evidence, to obtain the target diagnostic result, expressed as:

[0074] in, For the target diagnostic results, This is the Dempster combination rule fusion operator. This is the index for the number of fusions; the total number of fusions is the total number of pieces of evidence minus 1.

[0075] In one embodiment, the server diagnoses the target domain test sample set under four different operating conditions, obtaining four preliminary diagnostic results. The server then transforms each of these preliminary diagnostic results into four pieces of evidence. The server calculates the pairwise metrics between the four pieces of evidence, constructing a 4×4 divergence metric matrix. Based on the divergence metric matrix, the server calculates the average divergence, support, and credibility of each piece of evidence. The server then performs a weighted average of the four pieces of evidence based on their credibility, obtaining a weighted average evidence. Finally, the server fuses the weighted average evidence with itself using three Dempster combination rules to obtain the target diagnostic result.

[0076] In one embodiment, cross-machine fault diagnosis experiments were conducted using the publicly available CWRU (Case Western Reserve University) bearing dataset and the IMS (Intelligent Maintenance Systems, University of Cincinnati) bearing dataset. Drive-end data with a sampling frequency of 12kHz was selected from the CWRU dataset, corresponding to four operating conditions: 0hp, 1hp, 2hp, and 3hp loads. A total of 100 samples of raw vibration signals were selected for each of the 10 healthy states, with each signal sample containing 1024 sampling points. Ten samples for each healthy state were used for training, and 90 samples were used for testing. The IMS dataset had a sampling frequency of 20kHz. Normal state data was taken from the early operating phase of bearing 1 in test series 1; inner ring fault data was taken from the failure phase of bearing 3 in test series 1; outer ring fault data was taken from the failure phase of bearing 1 in test series 2; and rolling element fault data was taken from the failure phase of bearing 4 in test series 1. The signal processing method was consistent with that of the CWRU dataset. Ten source domain training samples and 90 source domain test samples were used. The test samples of the target domain IMS dataset were used as cross-machine diagnostic objects. The input signal length was 1024 sampling points.

[0077] The results of the fused diagnostic analysis are compared with existing methods such as DDC (Deep Domain Confusion), DCORAL (Deep Correlation Alignment), DANN (Domain-Adversarial Neural Network), MCD (Maximum Classifier Discrepancy), FTNN (Fine-Tuning Neural Network), JDA (Joint Distribution Adaptation), and DDTLN (Deep Domain Transfer Learning Network) as shown in Table 1.

[0078] Table 1: Comparison Results

[0079] As shown in Table 1, the accuracy rates of the DDC method are 60.81%, DCORAL method 68.31%, DANN method 58.37%, MCD method 60.63%, FTNN method 58.05%, JDA method 61.53%, and DDTLN method 85.38%, while the accuracy rate of the method proposed in this application reaches 90.23%. The accuracy rate of the method proposed in this application is higher than all existing methods compared, verifying the effectiveness of the method provided in cross-machine fault diagnosis. This information fusion method not only improves the accuracy of fault diagnosis but also reduces the problem of incorrect health status judgments that may occur in traditional neural network models.

[0080] In one embodiment, the server fuses cross-machine fault diagnosis results from the CWRU dataset to the IMS dataset under four different operating conditions. The diagnosis results before fusion are as follows: Figures 4-7 As shown, Figure 4 The confusion matrix corresponding to CWRU to IMS under 0hp operating conditions. Figure 5 The confusion matrix corresponding to CWRU to IMS under 1hp operating conditions. Figure 6 The confusion matrix corresponding to CWRU to IMS under 2hp operating conditions. Figure 7 The confusion matrix corresponding to CWRU to IMS under 3hp conditions. The comparison results after fusion with other methods are as follows: Figure 8 As shown. From Figure 8As can be seen, the method provided in this application outperforms methods such as DDC, DCORAL, DANN, MCD, FTNN, JDA, and DDTLN in terms of diagnostic accuracy. This information fusion method not only improves the accuracy of fault diagnosis but also reduces the problem of incorrect health status judgment that may occur in traditional neural network models.

[0081] In one embodiment, the basic probability distribution of outer ring faults in a cross-machine fault diagnosis experiment from the CWRU dataset to the IMS dataset is shown in Table 2.

[0082] Table 2: Basic Probability Distribution of CWRU→IMS Outer Ring Faults

[0083] Table 2 presents the basic probability allocation of outer ring faults under four operating conditions. The column for outer ring faults in the table shows the maximum confidence value for each operating condition, representing the model's confidence in the outer ring fault. As can be seen from Table 2, with the change of operating conditions, the model's confidence value for outer ring faults increases from 0.37 to 0.86, and the fused confidence value is 0.64, indicating that the method provided in this application has a stable ability to identify outer ring faults.

[0084] Based on the above, the cross-machine fault diagnosis method provided in this application obtains source domain vibration signal data of the first machine bearing and target domain vibration signal data of the second machine bearing. It then obtains a source domain training sample set based on the source domain vibration signal data, and a target domain labeled training sample set and a target domain test sample set based on the target domain vibration signal data. The source domain training sample set and the target domain labeled training sample set are combined into a joint training set for joint optimization training of the feature extractor and classifier. The trained feature extractor and classifier are then used to diagnose the target domain test sample set to obtain multiple preliminary diagnostic results. Finally, these multiple preliminary diagnostic results are fused to obtain the target diagnostic result, thus solving the problem of low accuracy in cross-machine fault diagnosis under small sample conditions.

[0085] The above text combined Figures 1 to 8 The cross-machine fault diagnosis method provided in the embodiments of this application has been described in detail. The apparatus and equipment provided in the embodiments of this application will be described below with reference to the accompanying drawings.

[0086] This application also provides a cross-machine fault diagnosis device, such as... Figure 9 As shown in the figure, this is a schematic diagram of a cross-machine fault diagnosis device provided in an embodiment of this application. The device includes: The acquisition module 301 is used to acquire source domain vibration signal data of the first machine bearing and target domain vibration signal data of the second machine bearing; based on the source domain vibration signal data of the first machine bearing, a source domain training sample set is obtained, which includes labeled samples under various health conditions; based on the target domain vibration signal data of the second machine bearing, a target domain labeled training sample set and a target domain test sample set are obtained. Training module 302 is used to input samples from the source domain training sample set and the target domain labeled training sample set into a feature extractor to extract intermediate features; input the intermediate features into a classifier to obtain a classification result; and perform joint optimization training on the feature extractor and the classifier based on the classification result and the health status category to obtain a trained feature extractor and a trained classifier. The output module 303 is used to input the target domain test sample set into the trained feature extractor and the trained classifier to obtain multiple preliminary diagnostic results; and to fuse the multiple preliminary diagnostic results to obtain the target diagnostic result.

[0087] In some possible implementations, the training module 302 is specifically configured to: calculate the source domain cross-entropy loss based on the classification results and health status categories of the source domain training sample set; calculate the target domain cross-entropy loss based on the classification results and health status categories of the target domain labeled training sample set; perform a weighted summation of the source domain cross-entropy loss, the target domain cross-entropy loss, and the improved Softmax loss to obtain a total loss value; and update the network parameters of the feature extractor and the classifier based on the total loss value to obtain the trained feature extractor and the trained classifier.

[0088] In some possible implementations, the output module 303 is specifically used to convert the preliminary diagnostic result into evidence in the form of a basic probability allocation; calculate the metric between any two pieces of evidence; calculate the average divergence of each piece of evidence based on the metric between all two pieces of evidence; calculate the support of each piece of evidence based on the average divergence of each piece of evidence; calculate the credibility of each piece of evidence based on the support of each piece of evidence; perform a weighted average of all evidence based on the credibility of each piece of evidence to obtain a weighted average evidence; and fuse the weighted average evidence to obtain the target diagnostic result.

[0089] In some possible implementations, the output module 303 is specifically used to select any hypothesis in the recognition framework, obtain a first confidence value and a second confidence value corresponding to the hypothesis from two pieces of evidence; obtain a magnitude difference element value based on the absolute value of the difference between the first confidence value and the second confidence value, and the sum of the absolute values ​​of the first confidence value and the second confidence value; obtain a direction cosine value based on the ratio of the product of the first confidence value and the second confidence value, and the product of the magnitudes of the confidence vectors of the two pieces of evidence; and obtain a metric value between the two pieces of evidence based on the magnitude difference element value and the direction cosine value.

[0090] In some possible implementations, the feature extractor includes multiple cascaded convolutional modules, a global average pooling layer, and a flattening layer. Each convolutional module includes a convolutional layer, a batch normalization layer, and a max pooling layer. The classifier includes a first fully connected layer and a second fully connected layer. The input of the first fully connected layer is connected to the output of the flattening layer, and the output of the first fully connected layer is connected to the input of the second fully connected layer.

[0091] In some possible implementations, the acquisition module 301 is specifically used to sample the source domain vibration signal data of the first machine bearing to obtain the sampled first signal; the sampled signal is divided into multiple samples, and all samples are used as the source domain training sample set.

[0092] In some possible implementations, the acquisition module 301 is specifically used to sample the target domain vibration signal data of the second machine bearing to obtain the sampled second signal; divide the sampled signal into multiple samples, use a portion of the samples under each health state as the target domain labeled training sample set; and use the remaining samples as the target domain test sample set.

[0093] The cross-machine fault diagnosis apparatus according to the embodiments of this application can correspondingly execute the method described in the embodiments of this application, and the other operations and / or functions of each module / unit of the cross-machine fault diagnosis apparatus are respectively for implementing Figure 2 For the sake of brevity, the corresponding processes of each method in the illustrated embodiments will not be described in detail here.

[0094] This application also provides a computing device. For example... Figure 10 As shown in the figure, this is a schematic diagram of a computing device provided in an embodiment of this application. The computing device 400 includes a bus 401, a processor 402, a communication interface 403, and a memory 404. The processor 402, the memory 404, and the communication interface 403 communicate with each other via the bus 401.

[0095] Bus 401 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 10 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0096] Processor 402 can be any one or more of the following processors: central processing unit (CPU), graphics processing unit (GPU), microprocessor (MP), or digital signal processor (DSP).

[0097] The communication interface 403 is used for communication with external devices. For example, if the computing device is a first switch, the communication interface 403 can be used for communication between the first switch and a first user terminal, or for communication between the first switch and a second switch.

[0098] Memory 404 may include volatile memory, such as random access memory (RAM). Memory 404 may also include non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid state drive (SSD).

[0099] The memory 404 stores executable code, which the processor 402 executes to perform the aforementioned cross-machine fault diagnosis method.

[0100] Specifically, in achieving Figure 9 In the case of the illustrated embodiment, and Figure 9 When the modules or units of the cross-machine fault diagnosis device described in the embodiments are implemented in software, the following steps are performed: Figure 9 The software or program code required for the functions of each module / unit can be partially or wholly stored in memory 404. Processor 402 executes the program code corresponding to each unit stored in memory 404 to perform the aforementioned cross-machine fault diagnosis method.

[0101] This application also provides a computer-readable storage medium. The computer-readable storage medium can be any available medium that a computing device can store, or a data storage device such as a data center containing one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive). The computer-readable storage medium includes instructions that instruct the computing device to perform the aforementioned cross-machine fault diagnosis method.

[0102] This application also provides a computer program product comprising one or more computer instructions. When the computer instructions are loaded and executed on a computing device, all or part of the processes or functions described in this application are generated.

[0103] The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, or data center to another website, computer, or data center via wired (e.g., coaxial cable, fiber optic) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0104] When the computer program product is executed by a computer, the computer performs any of the aforementioned cross-machine fault diagnosis methods. The computer program product can be a software installation package; when any of the aforementioned cross-machine fault diagnosis methods is required, the computer program product can be downloaded and executed on the computer.

[0105] The descriptions of the processes or structures corresponding to the above figures each have their own emphasis. For parts of a process or structure that are not described in detail, please refer to the relevant descriptions of other processes or structures.

[0106] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be covered within the scope of protection of this application.

Claims

1. A cross-machine fault diagnosis method, characterized in that, The method includes: Acquire source domain vibration signal data of the first machine bearing and target domain vibration signal data of the second machine bearing. Based on the source domain vibration signal data of the first machine bearing, a source domain training sample set is obtained, which includes labeled samples under various health conditions. Based on the target domain vibration signal data of the second machine bearing, a labeled training sample set and a target domain test sample set are obtained. Input the samples from the source domain training sample set and the target domain labeled training sample set into the feature extractor to extract intermediate features; The intermediate features are input into the classifier to obtain the classification result; Based on the classification results and health status categories, the feature extractor and the classifier are jointly optimized and trained to obtain the trained feature extractor and the trained classifier. The target domain test sample set is input into the trained feature extractor and the trained classifier to obtain multiple preliminary diagnostic results; Multiple preliminary diagnostic results are fused together to obtain the target diagnostic result.

2. The method according to claim 1, characterized in that, The classification results include the classification results of the source domain training sample set and the classification results of the target domain labeled training sample set. Based on the classification results and the health status category, the feature extractor and the classifier are jointly optimized and trained to obtain the trained feature extractor and the trained classifier, including: Based on the classification results of the source domain training sample set and the health status category of the source domain training sample set, calculate the source domain cross-entropy loss; Based on the classification results of the labeled training sample set in the target domain and the health status category of the labeled training sample set in the target domain, calculate the target domain cross-entropy loss; The source domain cross-entropy loss, the target domain cross-entropy loss, and the improved Softmax loss are weighted and summed to obtain the total loss value. The network parameters of the feature extractor and the classifier are updated based on the total loss value to obtain the trained feature extractor and the trained classifier.

3. The method according to claim 1, characterized in that, The process of fusing multiple preliminary diagnostic results to obtain the target diagnostic result includes: Evidence that transforms the preliminary diagnostic results into a basic probability allocation form; Calculate the metric between any two pieces of evidence; Calculate the average divergence of each piece of evidence based on the metric between all two pieces of evidence; Calculate the support level of each piece of evidence based on the average divergence of each piece of evidence; Calculate the credibility of each piece of evidence based on its level of support; Based on the credibility of each piece of evidence, a weighted average of all the evidence is calculated to obtain a weighted average of the evidence. The weighted average evidence is fused to obtain the target diagnostic result.

4. The method according to claim 3, characterized in that, Calculate the metric between any two pieces of evidence, including: Select any hypothesis within the identification framework and obtain the first and second confidence values ​​corresponding to the hypothesis from two pieces of evidence; The magnitude difference element value is obtained based on the absolute value of the difference between the first confidence value and the second confidence value, and the sum of the absolute values ​​of the first confidence value and the second confidence value. The direction cosine value is obtained by the ratio of the product of the first confidence value and the second confidence value, and the product of the magnitudes of the confidence vectors of the two pieces of evidence. The metric between the two pieces of evidence is obtained based on the difference in magnitude and the cosine of direction.

5. The method according to claim 1, characterized in that, The feature extractor includes multiple cascaded convolutional modules, a global average pooling layer, and a flattening layer. Each convolutional module includes a convolutional layer, a batch normalization layer, and a max pooling layer. The classifier includes a first fully connected layer and a second fully connected layer. The input of the first fully connected layer is connected to the output of the flattened layer, and the output of the first fully connected layer is connected to the input of the second fully connected layer.

6. The method according to claim 1, characterized in that, Based on the source domain vibration signal data of the first machine bearing, a source domain training sample set is obtained, including: The source domain vibration signal data of the first machine bearing is sampled to obtain the sampled first signal; The sampled signal is divided into multiple samples, and all samples are used as the source domain training sample set.

7. The method according to claim 1, characterized in that, Based on the vibration signal data of the target domain of the second machine bearing, a labeled training sample set and a test sample set for the target domain are obtained, including: The vibration signal data of the target domain of the second machine bearing is sampled to obtain the sampled second signal; The sampled signal is divided into multiple samples, and a portion of the samples under each health state is used as the labeled training sample set for the target domain. Use the remaining samples as the test sample set for the target domain.

8. A cross-machine fault diagnosis device, characterized in that, The device includes: The acquisition module is used to acquire source domain vibration signal data of the first machine bearing and target domain vibration signal data of the second machine bearing; based on the source domain vibration signal data of the first machine bearing, a source domain training sample set is obtained, which includes labeled samples under various health conditions; based on the target domain vibration signal data of the second machine bearing, a target domain labeled training sample set and a target domain test sample set are obtained. The training module is used to input samples from the source domain training sample set and the target domain labeled training sample set into a feature extractor to extract intermediate features; input the intermediate features into a classifier to obtain a classification result; and perform joint optimization training on the feature extractor and the classifier based on the classification result and the health status category to obtain a trained feature extractor and a trained classifier. The output module is used to input the target domain test sample set into the trained feature extractor and the trained classifier to obtain multiple preliminary diagnostic results; and to fuse the multiple preliminary diagnostic results to obtain the target diagnostic result.

9. A computing device, characterized in that, Including memory and processor; The memory stores one or more computer programs, the one or more computer programs including instructions; when the instructions are executed by the processor, the computing device performs the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program for performing the method as described in any one of claims 1 to 7.

Images