Cross-working-condition fault diagnosis model training method, diagnosis method, system and medium

Abstract

The invention provides a cross-working-condition fault diagnosis model training method, a cross-working-condition fault diagnosis method, a cross-working-condition fault diagnosis system and a medium. The model training method comprises the following steps: acquiring a target data set according to a multi-sensor signal, wherein the target data set comprises a source domain training set, a target domain training set and a target domain test set; then, a fault diagnosis model is built, and model parameters are initialized; training the fault diagnosis model by adopting the source domain training set and the target domain training set to obtain a loss function, updating model parameters based on the loss function, and performing iterative training in the source domain training set and the target domain training set to obtain a target diagnosis model; and further inputting the target domain test set into the target diagnosis model, outputting a fault diagnosis result, and optimizing the target diagnosis model based on the fault diagnosis result. According to the method, the problem of spatial relation errors possibly occurring when the time domain features are independently extracted can be effectively avoided, and the model not only can be well expressed under specific working conditions, but also can well adapt to fault diagnosis tasks under different working conditions.

Description

Technical Field

[0001] The present invention relates to a method for training a fault diagnosis model, in particular to a method for training a cross-condition fault diagnosis model, a diagnosis method, a system and a medium, belonging to the technical field of fault diagnosis. Background Art

[0002] Modern industrial production highly depends on rotating machinery. Faults in rotating machinery may lead to huge economic losses and safety risks.

[0003] For multi-sensor data, the prior art uses two independent modules to extract the time-domain features and frequency-domain features respectively. For example, first use a Long Short-Term Memory Network (LSTM) to extract the time-domain features, and then use a graph neural network to extract the frequency-domain features to model the relationship between multi-sensor data. There are disadvantages: during the process of extracting the time-domain features, errors may occur in the spatial relationship between multiple sensors.

[0004] Based on unsupervised transfer learning for cross-condition fault diagnosis, most of the prior art takes the features before the classifier (fully connected layer) of a deep neural network as the sample features in the source domain and the target domain, calculates the distance between the distributions of the features, and minimizes this distance through a loss function. The disadvantage is that only one layer in the network is extracted for calculation, and the distance between the source domain and the target domain in the feature space cannot be effectively reduced. Summary of the Invention

[0005] In view of this, the present invention provides a method for training a cross-condition fault diagnosis model, a diagnosis method, a system and a medium to solve or alleviate the technical problems existing in the prior art, and at least provide a beneficial option.

[0006] The technical solution of the embodiment of the present invention is implemented as follows: A method for training a cross-condition fault diagnosis model is provided, including the following steps:

[0007] S100: Collect multi-sensor signals to obtain a target data set; wherein, the target data set at least includes a source domain training set, a target domain training set and a target domain test set.

[0008] S200: Build a fault diagnosis model based on a spatio-temporal dynamic graph domain adaptation network, and randomly initialize the model parameters of the fault diagnosis model.

[0009] S300: Use the source domain training set and the target domain training set to train the fault diagnosis model to obtain a loss function, update the model parameters based on the loss function, and iterate a preset number of training rounds in the source domain training set and the target domain training set to obtain a final fault diagnosis model as the target diagnosis model.

[0010] S400: Input the target domain test set into the target diagnosis model to output a fault diagnosis result, evaluate the performance of the target diagnosis model based on the fault diagnosis result, and then optimize the target diagnosis model according to the evaluation result.

[0011] Further preferably: The step S100 includes: simulating various operating conditions of the device to collect multi-sensor signals as time-domain signals, and preprocessing the multi-sensor signals to obtain a target data set.

[0012] Further preferably: The preprocessing at least includes: converting the multi-sensor signals into frequency-domain signals through fast Fourier transform based on the source domain training set and the target domain training set, extracting the low-frequency components in the frequency-domain signals, and normalizing the low-frequency components and the time-domain signals.

[0013] Further preferably: The fault diagnosis model at least includes a time-domain feature extraction module, a frequency-domain feature extraction module, and a classification and domain adaptation module.

[0014] Further preferably: The time-domain feature extraction module at least includes a multi-scale convolutional neural network and a spatio-temporal dynamic graph module, and the spatio-temporal dynamic graph module at least includes a first dynamic graph convolutional network and a bidirectional gated recurrent unit; the step S300 at least includes: extracting initial time-domain features from the normalized time-domain signals through the multi-scale convolutional neural network, and performing dimensionality reduction processing on the initial time-domain features; obtaining target time-domain features through the first dynamic graph convolutional network and the bidirectional gated recurrent unit based on the initial time-domain features after dimensionality reduction processing; wherein, the gated unit in the bidirectional gated recurrent unit is replaced with the first dynamic graph convolutional network in advance.

[0015] Further preferably: The frequency-domain feature extraction module at least includes a standard convolutional neural network and a second dynamic graph convolutional network; the step S300 at least includes: extracting initial frequency-domain features from the normalized low-frequency components through the standard convolutional neural network, and performing dimensionality reduction processing on the initial frequency-domain features; obtaining target frequency-domain features through the second dynamic graph convolutional network based on the initial frequency-domain features after dimensionality reduction processing.

[0016] Further preferably: The step S300 further includes: calculating the multi-domain maximum mean discrepancy of the target time-domain features and the target frequency-domain features on the source domain training set and the target domain training set respectively through the classification and domain adaptation module, and adding the multi-domain maximum mean discrepancy to the cross-entropy loss function to obtain a loss function.

[0017] Based on the same inventive concept, the present invention also provides a cross - operating - condition fault diagnosis method, including: collecting the data to be diagnosed of the target device, and inputting the data to be diagnosed into the target diagnosis model to output the final diagnosis result of the target device; wherein, the target diagnosis model is obtained by the above - mentioned cross - operating - condition fault diagnosis model training method.

[0018] Based on the same inventive concept, the present invention also provides a cross - operating - condition fault diagnosis system, which includes: a collection module, configured to collect the data to be diagnosed of the target device and input the data to be diagnosed into the target diagnosis model; and a diagnosis module, configured to perform fault diagnosis on the target device based on the data to be diagnosed through the target diagnosis model and output the final diagnosis result.

[0019] Based on the same inventive concept, the present invention also provides a storage medium, in which a computer program is stored, and wherein the computer program is set to execute the above - mentioned cross - operating - condition fault diagnosis method when running.

[0020] Due to the above - mentioned technical solutions adopted in the embodiments of the present invention, it has the following advantages:

[0021] (1) By constructing a spatio - temporal dynamic graph domain - adaptive network, the present invention simultaneously processes the time - domain and frequency - domain features of multi - sensor data, thereby capturing the state information of the device more comprehensively and effectively avoiding the problem of spatial relationship errors that may occur when extracting time - domain features alone.

[0022] (2) The present invention jointly trains the model with the source - domain training set and the target - domain training set, enabling the model to not only perform well under specific working conditions but also better adapt to the fault diagnosis tasks under different operating conditions, which solves the problem of poor cross - domain adaptation ability caused by extracting features only from a single aspect in traditional methods.

[0023] The above summary is only for the purpose of the specification and is not intended to be limiting in any way. In addition to the above - described illustrative aspects, embodiments and features, further aspects, embodiments and features of the present invention will become apparent by reference to the drawings and the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following - described drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0025] Figure 1Flow chart of the cross - operating - condition fault diagnosis model training method according to the present invention.

[0026] Figure 2 Schematic diagram of the structure after replacing a single GRU according to the present invention.

[0027] Figure 3 Schematic diagram of the experimental results of the first dataset comparison according to the present invention.

[0028] Figure 4 Schematic diagram of the experimental results of the second dataset comparison according to the present invention.

[0029] Figure 5 Schematic diagram of the experimental results of the second dataset comparison after adding - 4dB noise according to the present invention.

[0030] Figure 6 Visualization results on the first dataset of the present application; wherein, Figure 6 (a) is the original data distribution, Figure 6 (b) is the feature under A→B, Figure 6 (c) is the feature under B→A.

[0031] Figure 7 Visualization results on the second dataset of the present application; wherein, Figure 7 (a) is the original data distribution, Figure 7 (b) is the feature under A→B, Figure 7 (c) is the feature under B→A.

[0032] Figure 8 Flow chart of the cross - operating - condition fault diagnosis method according to the present invention.

[0033] Figure 9 Framework diagram of the cross - operating - condition fault diagnosis system according to the present invention. Detailed Description of the Invention

[0034] In the following, only some exemplary embodiments are briefly described. As those skilled in the art can recognize, the described embodiments can be modified in various different ways without departing from the spirit or scope of the present invention. Therefore, the drawings and description are considered to be exemplary in nature rather than restrictive.

[0035] The embodiments of the present invention will be described in detail below with reference to the drawings.

[0036] As Figure 1 shown, the embodiments of the present invention provide a cross - operating - condition fault diagnosis model training method, including the following steps S100 - S400:

[0037] S100: Collect multi-sensor signals to obtain a target data set; where the target data set includes at least a source domain training set, a target domain training set, and a target domain test set.

[0038] In this embodiment, specifically: the step S100 includes: simulating various operating conditions of the device to collect multi-sensor signals as time-domain signals, and preprocessing the multi-sensor signals to obtain a target data set.

[0039] Taking a rotating machine as an example of the device, simulate the operating environment of the rotating machine through a test bench, collect multi-sensor signals, preprocess these signals, then change the operating conditions of the rotating machine (such as speed, load, etc.), and collect multi-sensor signals under different operating conditions.

[0040] In this embodiment, specifically: the preprocessing at least includes: converting the multi-sensor signals into frequency-domain signals through fast Fourier transform based on the source domain training set and the target domain training set, extracting the low-frequency components in the frequency-domain signals, and normalizing the low-frequency components with the time-domain signals.

[0041] Among them, the normalization is to normalize the time-domain and frequency-domain signals to the range of -1 to 1.

[0042] It should be noted that the range of the normalization can be set and adjusted by those skilled in the art according to actual application requirements, and is not limited thereto.

[0043] S200: Build a fault diagnosis model based on a spatio-temporal dynamic graph domain adaptation network, and randomly initialize the model parameters of the fault diagnosis model.

[0044] Among them, experiments are carried out on a Ubuntu system equipped with an Intel CPU and a GTX 4090 24GB GPU, and the model is built and trained with PyTorch 2.2.1; those skilled in the art can also adjust it according to actual application requirements.

[0045] In this embodiment, specifically: the fault diagnosis model at least includes a time-domain feature extraction module, a frequency-domain feature extraction module, and a classification and domain adaptation module.

[0046] S300: Use the source domain training set and the target domain training set to train the fault diagnosis model to obtain a loss function, update the model parameters based on the loss function, and iterate a preset number of training rounds on the source domain training set and the target domain training set to obtain a final fault diagnosis model as the target diagnosis model.

[0047] Among them, during the training process, the batch size is set to 512, and 400 rounds of iteration are carried out on the training set.

[0048] It should be noted that those skilled in the art can adjust the above training parameters according to actual application requirements and are not limited thereto.

[0049] In this embodiment, specifically: the time-domain feature extraction module at least includes a multi-scale convolutional neural network and a spatio-temporal dynamic graph module, and the spatio-temporal dynamic graph module at least includes a first dynamic graph convolutional network and a bidirectional gated recurrent unit; the step S300 at least includes: extracting initial time-domain features from the normalized time-domain signal through the multi-scale convolutional neural network, and performing dimensionality reduction processing on the initial time-domain features; obtaining target time-domain features through the first dynamic graph convolutional network and the bidirectional gated recurrent unit based on the initial time-domain features after dimensionality reduction processing; wherein, the gated unit in the bidirectional gated recurrent unit is replaced with the first dynamic graph convolutional network in advance.

[0050] Among them, first, the original multi-sensor signals are concatenated along the channel dimension, and then a multi-scale convolutional neural network (CNN) is used to extract features and perform dimensionality reduction. The multi-scale CNN consists of multiple branches, each branch uses different-sized convolutional kernels, and finally the outputs of each branch are summed. The extracted features are adjusted through a channel attention mechanism and then input into the proposed spatio-temporal dynamic graph module (Spatial-Temporal Dynamic Graph Module, STDGM). STDGM is constructed by a dynamic graph convolutional network (Dynamic Graph Convolutional Network, DGCN) and a bidirectional gated recurrent unit (Bidirectional Gated Recurrent Unit, BiGRU).

[0051] Specifically: the multi-sensor data of rotating machinery has dynamic correlations in the spatial dimension. However, most existing fault diagnosis methods use a pre-defined adjacency matrix to model the relationship between multi-sensor data, and the adjacency matrix remains unchanged throughout the training process. The present invention introduces DGCN, whose graph structure can be adaptively updated during the training process, and the formula is as follows:

[0052] where, A t ∈R N×N represents the adjacency matrix at time t, N is the number of nodes, that is, the number of output channels of the multi-scale CNN. E t ∈R N×d represents the node spatial relationship feature at time t, where d is the dimension of the node spatial relationship feature. Matrix is E tThe transpose. The Softmax and ReLU functions are non-linear activation functions. The initial node spatial relationship features are sampled from a uniform distribution and updated through backpropagation.

[0053] After obtaining the adjacency matrix through calculation, the graph convolutional network (GCN) is used to propagate and aggregate the neighbor information of each node, as shown in the following formula: O t = A t X t W + b.

[0054] Among them, O t ∈ R N×H represents the output at time step t, and H is the dimension of the hidden layer. X t ∈ R N×F represents the node features at time step t, where F is the feature dimension of each node. W ∈ R F×H and b ∈ R N×H represent the weight matrix and the bias vector respectively.

[0055] The proposed STDGM replaces the gating unit in BiGRU with DGCN, which can simultaneously extract the temporal dependencies and spatial relationships of multi-sensor signals. The structure after replacing a single GRU is as Figure 2 shown. By combining each GRU in BiGRU with DGCN, STDGM can be obtained.

[0056] The formula for combining a single GRU and DGCN is as follows:

[0057] R t = σ(A t X t W r + b r )

[0058] Z t = σ(A t X t W z + b z )

[0059]

[0060]

[0061] Among them, the subscript t represents the time step t. R t ∈ R N×H represents the reset gate, Z t ∈ R N×H represents the update gate, Tanh is a non-linear activation function, σ is the sigmoid function, and ⊙ represents element-wise multiplication. H t-1 is the hidden state of the previous time step, is the candidate hidden state. W r ,W z ,b r and b z are the weight matrix and bias vector of the reset gate and update gate respectively.

[0062] Denote as the output of the forward GRU. Similarly, the output can be calculated through the backward GRU Add and to obtain the final output target time domain feature of the spatio-temporal dynamic graph module (STDGM)

[0063]

[0064] In this embodiment, specifically: the frequency domain feature extraction module includes at least a standard convolutional neural network and a second dynamic graph convolutional network; the step S300 includes at least: extracting initial frequency domain features from the normalized low-frequency components through the standard convolutional neural network, and performing dimensionality reduction processing on the initial frequency domain features; obtaining target frequency domain features through the second dynamic graph convolutional network based on the initial frequency domain features after dimensionality reduction processing.

[0065] Among them, features are extracted using a standard CNN and dimensionality reduction is performed through a pooling operation; then spatial information is extracted through a dynamic graph convolutional network (DGCN); the frequency domain branch and the spatio-temporal domain branch share the same dynamic graph structure; finally, the features are pooled and converted to the specified dimension.

[0066] It should be noted that the first dynamic graph convolutional network and the second dynamic graph convolutional network are only used to distinguish the DGCNs in the time domain feature extraction module and the frequency domain feature extraction module, and have no other meaning.

[0067] In this embodiment, specifically: the step S300 further includes: calculating the multi-domain maximum mean discrepancy of the target time domain feature and the target frequency domain feature on the source domain training set and the target domain training set respectively through the classification and domain adaptation module, and adding the multi-domain maximum mean discrepancy to the cross-entropy loss function to obtain a loss function.

[0068] Among them, the classification and domain adaptation module is the classifier, which is composed of two fully connected layers and Softmax, and outputs the probability of each class. The classification loss uses the cross-entropy loss function.

[0069] For domain adaptation, the present invention proposes multi-domain MK-MMD (i.e., multi-domain maximum mean discrepancy), which can reduce the distance between the source domain and the target domain in the feature space from multiple angles.

[0070] Specifically, after extracting the spatio-temporal domain features and frequency domain features, the MK-MMD distances of the two types of features on the source domain and the target domain are calculated respectively, and added to the calculation result of the cross-entropy loss function as the final loss function.

[0071] As the backpropagation minimizes the loss and updates the model parameters, the model can map the samples in the source domain and the target domain to the same spatio-temporal domain feature space and frequency domain feature space, thus realizing cross-condition fault diagnosis. The formula for multi-domain MK-MMD is as follows:

[0072]

[0073] where the superscript st represents the spatio-temporal domain, the superscript freq represents the frequency domain, H k represents the Reproducing Kernel Hilbert Space (RKHS), φ represents the function mapping to the RKHS, s i and t j represent the samples from the source domain and the target domain respectively, and N and M represent the number of samples in the source domain and the target domain respectively. and represent the MK-MMD losses of the spatio-temporal features and frequency domain features between the source domain and the target domain respectively. Adding and together, the total multi-domain MK-MMD loss

[0074] Finally, the total loss function consists of two parts, such as L = L c + λL MK-MMD , the first part L c is the cross-entropy loss for classification, and the second part is the proposed multi-domain MK-MMD loss. λ is the weight factor.

[0075] S400: Input the target domain test set into the target diagnosis model to output the fault diagnosis result, and evaluate the performance of the target diagnosis model based on the fault diagnosis result, and then optimize the target diagnosis model according to the evaluation result.

[0076] Among them, use the trained model to make inferences on the test set constructed in the target domain, calculate the accuracy rate and F1 score, and evaluate the performance of the model; the optimizer uses Adam, and the initial learning rate is set to 0.001.

[0077] If the accuracy rate is low, it means that the classification effect of the model in the target domain is not ideal; if the F1 score is low, it means that the model performs poorly in the case of class imbalance.

[0078] If the model converges slowly or overfits, the learning rate can be adjusted. Usually, the learning rate decay strategy can be tried.

[0079] Furthermore, when the present invention is working:

[0080] Two datasets are used. The first dataset includes gearboxes and bearings, with a total of ten states, which is a ten-classification problem, and there are two different working conditions. The rotational speed and load are set to 20 Hz - 0 V and 30 Hz - 2 V respectively; 20 Hz - 0 V is denoted as working condition A, and 30 Hz - 2 V is denoted as working condition B. The second dataset includes gearboxes, with a total of five states, including the normal state and four types of root crack faults with different degrees, which is a five-classification problem, and there are two different working conditions: the rotational speeds are 900 r / min and 1200 r / min. The rotational speed of 900 r / min is denoted as working condition A, and 1200 r / min is denoted as working condition B.

[0081] Among them, A→B means that A is the source domain and B is the target domain, and B→A means that B is the source domain and A is the target domain.

[0082] To evaluate the generalization ability of the proposed model and the effectiveness of the multi-domain MK-MMD strategy, comparisons are made with different models under cross-working conditions. Feature extraction models are compared, including CNN, CNN+LSTM, DAGCN, and different transfer learning strategies: JAN, DANN, CDANN, CDANN+E. Each method includes a feature extraction module and a domain adaptation technique. JAN, DANN, CDANN, and CDANN+E all use a multi-layer standard CNN as the feature extractor, and its structure is the same as that of the CNN in the comparison method.

[0083] The results on the first dataset are as Figure 3 shown, and the best results are in bold; it can be seen that the proposed STDGDAN achieves the best effect.

[0084] The results on the second dataset are as Figure 4 shown. It can be seen that on this dataset, except for CNN, CNN+LSTM, and DANN, the cross-working condition fault diagnosis accuracies of other models are all close to 100%. Therefore, in order to make a better comparison, Gaussian noise of -4 dB is added to the original second dataset, and the diagnosis results are as Figure 5 shown.

[0085] Figure 5 Among them, the proposed model (dark purple) has the highest accuracy, proving that the model can perform cross-working condition fault diagnosis and has noise robustness at the same time.

[0086] In addition, it should be noted that in this invention, two datasets are used as examples. The multi-sensor signals in the datasets are input into the network to extract features, and then compared with the original datasets, so as to reflect the functions and effects of this invention. Specifically, the extracted spatio-temporal domain features and frequency domain features are concatenated together as the final features of each sample.

[0087] The visualization results on the two datasets are respectively as Figure 6 and Figure 7 shown. In these two figures, each blue data point represents a sample under working condition A, and each red data point represents a sample under working condition B. Figure 6 shows the visualization result on the first dataset, where Figure 6 (a) shows the distribution of the original data. It can be seen from Figure 6 (a) that the distributions of working conditions A and B hardly overlap. Figure 6 (b) and Figure 6 (c) respectively show the sample features extracted by the proposed model under cross-working conditions A→B and B→A. It can be seen from Figure 6 (b) that the degree of overlap of the samples is relatively high, while Figure 6 (c) shows less overlap, which is consistent with the results in Figure 3 , where the accuracy rate of A→B is 72.73% and the accuracy rate of B→A is 63.93%. Figure 7 shows the visualization result on the original second dataset. In Figure 7 (b) and Figure 7 (c), the distributions of working conditions A and B almost completely overlap, which is consistent with the results in Figure 4 , where the accuracy rate of A→B is 100.00% and the accuracy rate of B→A is 99.92%. These results further verify the effectiveness of the proposed multi-domain MK-MMD.

[0088] Based on the same inventive concept, as Figure 8 shown, this invention also provides a cross-working condition fault diagnosis method, including: collecting the data to be diagnosed of the target device, and inputting the data to be diagnosed into the target diagnosis model to output the final diagnosis result of the target device; wherein, the target diagnosis model is obtained by the above-mentioned cross-working condition fault diagnosis model training method.

[0089] Among them, taking the fan in the rotating machinery as an example of the target device, the data to be diagnosed is collected from the multi-sensors installed on the fan (such as vibration sensors, temperature sensors, pressure sensors, etc.), and the collected data is preprocessed, such as denoising, normalization and other operations, to ensure the data quality; and then the preprocessed data is further divided into a dataset to be diagnosed, including time domain signals and frequency domain signals.

[0090] Furthermore, load the already trained target diagnostic model. This model is constructed based on a spatio-temporal dynamic graph domain adaptation network, has strong time-domain and frequency-domain feature extraction capabilities, and can adapt to cross-condition fault diagnosis. The target diagnostic model processes the data to be diagnosed, and the model outputs the final diagnostic result of the target device, including the operating state of the device (such as normal, minor fault, severe fault, etc.) and possible fault types (such as bearing wear, imbalance, looseness, etc.). Among them, the output result can be in the form of a probability distribution or a clear class label, depending on the design of the model.

[0091] Based on the same inventive concept, as Figure 9 shown, the present invention also provides a cross-condition fault diagnosis system. The system includes: a collection module for collecting the data to be diagnosed of the target device and inputting the data to be diagnosed into the target diagnostic model; and a diagnosis module for performing fault diagnosis on the target device based on the data to be diagnosed through the target diagnostic model and outputting the final diagnostic result.

[0092] Among them, the cross-condition fault diagnosis system adopts the same inventive concept as the cross-condition fault diagnosis method. The specific implementation manner of the diagnosis system is the same as that of the diagnosis method and will not be elaborated here.

[0093] Based on the same inventive concept, the present invention also provides a storage medium. A computer program is stored in the storage medium, wherein the computer program is set to execute the above-mentioned cross-condition fault diagnosis method when running.

[0094] As mentioned above, the above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of various changes or substitutions, and these should all be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A cross-operating condition fault diagnosis model training method, characterized in that: The following steps are involved: S100: Collect multi-sensor signals to obtain a target data set; wherein the target data set at least includes a source domain training set, a target domain training set, and a target domain test set; S200: building a fault diagnosis model based on a spatiotemporal dynamic graph domain adaptive network, and randomly initializing model parameters of the fault diagnosis model; S300: training the fault diagnosis model using the source domain training set and the target domain training set to obtain a loss function, updating the model parameters based on the loss function, and iterating a preset number of training rounds on the source domain training set and the target domain training set to obtain a final fault diagnosis model as a target diagnosis model; S400: Inputting the target domain test set into the target diagnosis model to output a fault diagnosis result, and evaluating the performance of the target diagnosis model based on the fault diagnosis result, and then optimizing the target diagnosis model according to the evaluation result.

2. The cross-operating condition fault diagnosis model training method according to claim 1 is characterized in that: The step S100 includes: Various operating conditions of the equipment are simulated to collect multi-sensor signals as time domain signals, and the multi-sensor signals are preprocessed to obtain a target data set.

3. The cross-operating condition fault diagnosis model training method according to claim 2 is characterized in that: The pre-processing at least comprises: Based on the source domain training set and the target domain training set, the multi-sensor signal is converted into a frequency domain signal by fast Fourier transform, and the low-frequency component in the frequency domain signal is extracted, and the low-frequency component is normalized with the time domain signal.

4. The cross-operating condition fault diagnosis model training method according to claim 3 is characterized in that: The fault diagnosis model at least includes a time domain feature extraction module, a frequency domain feature extraction module and a classification and domain adaptation module.

5. The cross-operating condition fault diagnosis model training method according to claim 4 is characterized in that: The time domain feature extraction module at least includes a multi-scale convolutional neural network and a spatiotemporal dynamic graph module, and the spatiotemporal dynamic graph module at least includes a first dynamic graph convolutional network and a bidirectional gated recurrent unit; the step S300 at least includes: Extracting initial time domain features from the normalized time domain signal through the multi-scale convolutional neural network, and performing dimensionality reduction processing on the initial time domain features; Based on the initial time domain features after dimensionality reduction processing, the target time domain features are obtained through the first dynamic graph convolutional network and the bidirectional gated recurrent unit; wherein the gating unit in the bidirectional gated recurrent unit is replaced in advance with the first dynamic graph convolutional network.

6. The cross-operating condition fault diagnosis model training method according to claim 5 is characterized in that: The frequency domain feature extraction module at least includes a standard convolutional neural network and a second dynamic graph convolutional network; the step S300 at least includes: Extracting initial frequency domain features from the normalized low-frequency components through the standard convolutional neural network, and performing dimensionality reduction processing on the initial frequency domain features; Based on the initial frequency domain features after dimensionality reduction processing, the target frequency domain features are obtained through the second dynamic graph convolutional network.

7. The cross-operating condition fault diagnosis model training method according to claim 6 is characterized in that: The step S300 further includes: The multi-domain maximum mean difference of the target time domain features and the target frequency domain features on the source domain training set and the target domain training set is calculated respectively by the classification and domain adaptation module, and the multi-domain maximum mean difference is added to the cross entropy loss function to obtain the loss function.

8. A cross-operating condition fault diagnosis method, characterized in that: include: Collecting the to-be-diagnosed data of the target device, and inputting the to-be-diagnosed data into the target diagnosis model to output the final diagnosis result of the target device; Wherein, the target diagnosis model is obtained by the cross-operating condition fault diagnosis model training method as described in any one of claims 1-7.

9. A system using the cross-operating condition fault diagnosis method according to claim 8, characterized in that: The system comprises: A collection module, used for collecting the to-be-diagnosed data of the target device and inputting the to-be-diagnosed data into the target diagnosis model; The diagnosis module is used to perform fault diagnosis on the target device based on the data to be diagnosed through the target diagnosis model and output a final diagnosis result.

10. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the cross-operating condition fault diagnosis method according to claim 8 when running.

Images