A method for constructing a variable working condition bearing fault diagnosis model and application thereof

Abstract

The application belongs to the technical field of fault diagnosis, and discloses a method for constructing a bearing fault diagnosis model under variable working conditions and application, which comprises the following steps: (1) constructing a bearing fault diagnosis model under variable working conditions, and taking bearing vibration signal data under multiple existing working conditions as source domain data, and taking bearing vibration signal data under a new distribution as target domain data to form a training sample set; (2) inputting the training sample set into the bearing fault diagnosis model under variable working conditions, training the diagnosis model only by using data corresponding to one of the multiple existing working conditions, obtaining an AdaSHAP model under the working condition, and performing first-stage training on the AdaSHAP model; (3) performing second-stage training on the AdaSHAP model, and performing domain self-adaptation on other working condition data, so that the AdaSHAP model is applicable to other working conditions. The application solves the technical problem of low bearing fault diagnosis accuracy of the prior art under variable working condition scenarios.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of fault diagnosis, and more particularly to a method for constructing a variable working condition fault diagnosis model and application thereof. BACKGROUND

[0002] Bearing, as a basic mechanical standard part, mainly plays a role in supporting rotating devices in mechanical devices. Bearing failure accounts for 44% of the total failure of motors, indicating that bearings are parts prone to failure in motors. Motor failures often occur during normal operation. If a high-speed running motor fails, it will not only reduce production efficiency, but also harm the surrounding operators. Therefore, timely warning of bearing failure and judging the failure type have great significance to actual production and life.

[0003] Currently, the bearing diagnosis models based on traditional deep learning mainly include convolutional neural network, autoencoder and deep belief network. As one of the representative algorithms of deep learning, convolutional neural network can automatically extract the required features from raw data. The bearing diagnosis model based on convolutional neural network can realize functions such as bearing health prediction and grease deficiency alarm, which cannot be realized by traditional machine learning. Autoencoder is a common algorithm in unsupervised learning. Autoencoder has the disadvantage of poor robustness to noise. In order to improve the anti-noise ability of the autoencoder network, SDA (stacked denoising autoencoder network) is used. In order to improve the training speed, an autoencoder model based on extreme learning machine (ELM) and a multi-autoencoder network are proposed. The network is composed of a group of autoencoder networks with different activation functions. Each autoencoder network will get a prediction result, and the network will integrate an accurate and stable diagnosis result. Deep belief network is also a common algorithm in unsupervised learning. In addition to the three networks mentioned above, recurrent neural network is also applied to this field, but it is not widely used. In addition, the recently emerging generative adversarial network is also being introduced more and more. Using the data generation characteristics of generative adversarial network, generative adversarial network can be used to solve the class imbalance problem caused by the scarcity of failure samples in the bearing data set.

[0004] However, an important assumption for most neural networks to obtain good results is that the source domain data and the target domain data are identically distributed. However, in actual application scenarios, this assumption is often violated due to changes in the environment for collecting data, resulting in a sharp decline in neural network performance. In order to solve this problem, how to realize domain migration has become a popular research direction in neural networks. SUMMARY

[0005] In view of the above defects or improvement needs of the prior art, the application provides a method for constructing a variable working condition fault diagnosis model and application, which only uses a data set under one working condition to train an AdaSHAP model under the working condition; then, an improved MK-MMD method is used to map the source domain data and the target domain data to a new representation space by using AdaSHAP to compare the data of the two domains, and domain adaptation is performed by using the unlabeled other working condition data, so that the AdaSHAP model is applicable to other working conditions, and a variable working condition bearing fault diagnosis model is established, and the technical problem of low bearing fault diagnosis accuracy of the prior art in the variable working condition scene is solved.

[0006] To achieve the above-mentioned purpose, according to one aspect of the application, a method for constructing a variable working condition fault diagnosis model is provided, which comprises the following steps:

[0007] (1) Constructing a variable working condition fault diagnosis model, and taking bearing vibration signal data under multiple existing working conditions as source domain data and bearing vibration signal data under an unseen working condition as target domain data to constitute a training sample set; wherein the variable working condition fault diagnosis model comprises a basis function decomposition network based on functional data analysis and a bearing fault classification network based on Shapley value, the basis function decomposition network is used to realize a data-driven decoupling process of bearing data, and the bearing fault classification network is used to calculate the influence of each basis function learned by the basis function decomposition network on the fault type prediction result under cross-working condition, so as to dynamically prune the basis function decomposition network;

[0008] (2) Inputting the training sample set into the variable working condition bearing fault diagnosis model, training the variable working condition bearing fault diagnosis model only by using data corresponding to one of the multiple existing working conditions, obtaining an AdaSHAP model under the working condition, and performing first-stage training on the AdaSHAP model;

[0009] (3) Performing second-stage training on the AdaSHAP model, using an improved MK-MMD method to perform domain adaptation on other working condition data in the training sample set, so that the AdaSHAP model is applicable to other working conditions.

[0010] Further, the basis function decomposition network randomly initializes a group of multilayer perceptron networks for fitting basis functions, and each basis function is subjected to inner product operation with input X to obtain a score of the input to the basis function.

[0011] Further, the scores of all basis functions are used as a new representation form of input, and are sent to the bearing fault classification network according to different task types.

[0012] Further, the bearing fault classification network is constructed based on a deep ShapNet network; the deep ShapNet network passes the Shapley value as a feature parameter extracted by the deep ShapNet network between layers through stacking shallow ShapNet; the shallow ShapNet is to obtain the accurate Shapley value of all inputs according to the linear property by covering the number of inputs.

[0013] Further, the first stage of training includes the following steps:

[0014] S21, the base function decomposition network is used to fit the base function beta most suitable for the data distribution under the current working condition by adding a set of multilayer perceptron before the bearing fault classification network;

[0015] S22, the bearing fault classification network is used to calculate the influence of each base function learned by the base function decomposition network on the final judgment result under the cross-condition, so as to prune the base function decomposition network.

[0016] Further, the second stage of training includes the following steps:

[0017] S31, domain adaptation is performed on other working condition data in the training sample set, so that the AdaSHAP model is also applicable to other working conditions;

[0018] S32, using the MK-MMD method, the source domain data and the target domain data are mapped to a new representation space by the AdaSHAP model to compare the data of the two domains, and the difference between the features of the two domains in the new representation space is calculated;

[0019] S33, when training the AdaSHAP model, an intermediate round is set, before the intermediate round, the AdaSHAP model only receives data from the source domain, and no transfer term is added in the loss function, and after the intermediate round, the source domain data and the unlabeled target domain data are input into the AdaSHAP model at the same time, the source domain data output representation space features and prediction results and calculate the classification loss, the target domain data need to get the feature in the representation space and together with the representation space feature of the source domain data to calculate the MK-MMD loss term.

[0020] Further, the calculation formula corresponding to the MK-MMD loss term between the features of the two domains in the new representation space is:

[0021] In the formula, S and T represent the source domain feature and the target domain feature respectively, and G represents the Gaussian kernel function.

[0022] According to another aspect of the present application, a variable working condition bearing fault diagnosis method is provided.

[0023] The present application also provides a variable working condition bearing fault diagnosis system, which comprises a memory and a processor, the memory stores a computer program, and the processor executes the computer program to perform the variable working condition bearing fault diagnosis method as described above.

[0024] The present application also provides a computer readable storage medium, which stores machine executable instructions, the machine executable instructions, when invoked and executed by a processor, cause the processor to implement the variable working condition bearing fault diagnosis model construction method or the variable working condition bearing fault diagnosis method as described above.

[0025] Overall, compared with the prior art, the variable working condition fault diagnosis model construction method and application provided by the present application mainly have the following beneficial effects:

[0026] 1. The constructed variable working condition bearing fault diagnosis model comprises a basis function decomposition network based on functional data analysis and a bearing fault classification network based on Shapley value, the basis function decomposition network adds a set of multilayer perceptron before the bearing fault classification network to fit the basis function most suitable for the current task and data distribution, thereby realizing the data-driven decoupling process of functional data. The bearing fault classification network obtains the influence of each feature on the final result through Shapley value, thereby calculating the influence of each learned basis function on the final judgment result under cross-working condition, and forming the required AdaSHAP model through the combination of the two networks, i.e., establishing the variable working condition bearing fault diagnosis model.

[0027] 2. The training of the whole model is divided into two stages: in the first stage, only the data set under one working condition is used for training, and the AdaSHAP model under this working condition is obtained for training; in the second stage, the improved MK-MMD method is used to realize the mapping relationship between the source domain data and the target domain data through the AdaSHAP model to compare the data of the two domains, and the domain adaptation is performed through the unlabeled other working condition data, so that the AdaSHAP model is also applicable to other working conditions.

[0028] 3. In order to realize end-to-end data processing while utilizing the strong representation performance of the deep learning network, a network-in-network technology is introduced, which adds a set of simple multi-layer perceptrons before the bearing fault classification network to fit the base function β that is most suitable for the current task and data distribution, thereby realizing the data-driven decoupling process of the functional data.

[0029] 4. Compared with the traditional bearing fault diagnosis method, the variable working condition bearing fault diagnosis method provided by the present application does not need manual feature extraction, only needs to train the above variable working condition bearing fault diagnosis model, collects corresponding data and applies it to the model, and then the bearing fault diagnosis can be automatically performed, and manual bearing fault discrimination is not needed.

[0030] 5. The model used in the variable working condition bearing fault diagnosis method provided by the present application includes a bearing fault classification network based on the Shapley value, which has the property of quantitatively evaluating the influence of the input on the output, and can help engineering and technical personnel to determine the main influencing factors under the condition of multiple measurement data inputs, thereby more efficiently and accurately performing bearing fault diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 is a flowchart of a variable working condition bearing fault diagnosis model construction method provided by the present application. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical scheme and advantages of the present application clearer and more apparent, the present application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0033] The present application provides a variable working condition bearing fault diagnosis model construction method, which mainly includes the following steps:

[0034] Step one, a variable working condition fault diagnosis model is constructed, at the same time, bearing vibration signal data under multiple existing working conditions are taken as source domain data, bearing vibration signal data under an unobserved working condition is taken as target domain data to form a training sample set; wherein the variable working condition fault diagnosis model comprises a basis function decomposition network based on functional data analysis and a bearing fault classification network based on Shapley value, the basis function decomposition network is used to realize a data-driven decoupling process of bearing data, and the bearing fault classification network is used to calculate the influence of each basis function learned by the basis function decomposition network on the fault type prediction result under cross-working condition, so as to dynamically prune the basis function decomposition network.

[0035] The basis function decomposition network randomly initializes a group of multilayer perceptron networks for fitting basis functions, each basis function performs inner product operation with input X to obtain a score of the input to the basis function, and the calculation formula is:

[0036] c i =∫β i (t)X(t)dt

[0037] The scores of all basis functions together are taken as a new representation form of input, which is sent to a subsequent processing network according to different task types.

[0038] The bearing fault classification network realizes dynamic pruning by relying on a deep ShapNet, the deep ShapNet performs layer-to-layer transmission of Shapley value as a feature parameter extracted by the network by stacking shallow ShapNets. The shallow ShapNet is based on linear properties, and the exact Shapley value of all inputs is calculated by completely covering the number of inputs. The Shapley module is the smallest unit in the ShapNet, and this structure only simply executes the definition formula of the Shapley value, and all combinations affecting the final result are brought into the network for calculation.

[0039] Step two, the training sample set is input into the variable working condition bearing fault diagnosis model, only the data corresponding to one of the existing multiple working conditions is used to train the variable working condition bearing fault diagnosis model, an AdaSHAP model under this working condition is obtained, and the AdaSHAP model is trained in the first stage.

[0040] In this embodiment, the first stage training includes the following steps:

[0041] S21, the basis function decomposition network is used to fit the basis function β most suitable for the data distribution under the current working condition by adding a group of multilayer perceptron networks before the bearing fault classification network.

[0042] S22, the influence of each basis function learned by the basis function decomposition network on the final judgment result under cross-working conditions is calculated by using the bearing fault classification network, thereby pruning the basis function decomposition network, excluding basis functions with small contribution values, and making the basis function decomposition network more focused on domain-invariant features, improving cross-working condition classification performance and speeding up the operation of the basis function decomposition network.

[0043] Step three, the AdaSHAP model is trained in the second stage, and the improved MK-MMD method is used to do domain adaptation on other working condition data in the training sample set, so that the AdaSHAP model is applicable to other working conditions. The AdaSHAP model after the second stage training is the interpretable variable working condition bearing fault diagnosis model to be constructed.

[0044] In the embodiment, the second stage training mainly includes the following steps:

[0045] S31, the AdaSHAP model is also applicable to other working conditions by doing domain adaptation on other working condition data in the training sample set.

[0046] S32, the MK-MMD method is used to map the source domain data and the target domain data to a new representation space by using the AdaSHAP model to compare the data of the two domains. In the new representation space, the features of the two domains will be calculated by the following formula.

[0047]

[0048] In the formula, S and T represent the source domain features and the target domain features respectively, and G represents the Gaussian kernel function.

[0049] S33, when training the AdaSHAP model, an hyperparameter "middle round number" is set. Before the middle round number, the AdaSHAP model only receives data from the source domain, and the transfer term is not added in the loss function. After the middle round number, the source domain data and the unlabeled target domain data are input into the AdaSHAP model at the same time. The source domain data outputs the representation space features and the prediction results and calculates the classification loss. The target domain data needs to get the features in the representation space and calculate the MK-MMD loss term together with the representation space features of the source domain data.

[0050] The application also provides a variable working condition bearing fault diagnosis method. The bearing vibration signal data to be diagnosed is input into the variable working condition bearing fault diagnosis model constructed by the construction method of the variable working condition bearing fault diagnosis model, and the bearing fault diagnosis model performs fault diagnosis based on the received bearing vibration signal data.

[0051] The application further provides a variable working condition bearing fault diagnosis system, which comprises a memory and a processor, the memory stores a computer program, and the processor executes the computer program to execute the variable working condition bearing fault diagnosis method.

[0052] The application further provides a computer readable storage medium, which stores machine executable instructions, the machine executable instructions, when called and executed by a processor, cause the processor to implement the variable working condition bearing fault diagnosis model construction method or the variable working condition bearing fault diagnosis method.

[0053] The application is further described in detail below with specific examples.

[0054] Embodiment 1

[0055] The application embodiment 1 provides a variable working condition bearing fault diagnosis model construction method, frequency domain data in original time series data collected by a sensor is taken as input, a function type data analysis based basis function decomposition network randomly initializes a group of multilayer perception network for fitting basis functions, each basis function is subjected to inner product operation with input X to obtain a score c of the input to the basis function as output.

[0056] Scores of all basis functions are taken together as a new representation form of bearing fault classification network input, and are sent to a basis Shapley value bearing fault classification network according to different task types. A bearing fault classification network based on Shapley value calculates the influence of each learned basis function on the final judgment result under cross working condition, so that the basis function decomposition network can be pruned according to the contribution value, and the basis function with small contribution value is excluded, so that the basis function decomposition network can focus more on domain invariant features to improve the cross working condition classification performance and speed up the operation speed of the basis function decomposition network, and then a variable working condition bearing fault diagnosis model is established.

[0057] Specifically, as shown in Figure 1 The method comprises the following steps:

[0058] S1, a variable working condition bearing fault diagnosis model is built.

[0059] Specifically, the variable working condition bearing fault diagnosis model comprises a function type data analysis based basis function decomposition network and a basis Shapley value bearing fault classification network. The function type data analysis based basis function decomposition network is used to realize a data driven signal decoupling process of bearing data, and the bearing fault classification network based on Shapley value is used to calculate the influence of each basis function learned by the basis function decomposition network on the final judgment result under cross working condition, so that the network can be pruned according to the influence.

[0060] The base function decomposition network based on functional data analysis adopts an AdaFNN network in the present embodiment, which introduces a technology of micro-network within a network in order to realize end-to-end data processing while taking advantage of the strong representation performance of a deep learning network, and fits the base function most suitable for the current task and data distribution, i.e., the base function β most suitable for the data set composed of the vibration signals generated by bearing faults input into the AdaFNN model, by adding a group of multilayer perceptron networks before the bearing fault classification network, so as to realize a data-driven decoupling process of functional data. Each base function is multiplied with the input X, as shown in the following formula, to obtain a score of the input for the base function.

[0061] c i =∫β i (t)X(t)dt

[0062] All base function scores together serve as a new representation form of the input, which is sent to a bearing fault classification network based on Shapley values according to different task types.

[0063] The construction and training of the AdaFNN network include the following three parts:

[0064] Base node number selection: reasonable selection of the number of nodes determines the final effect of the AdaFNN, and reasonable base point number enables the base function score to represent the input data well, the network classification accuracy is high, and the waste of computing power and the large number of network parameters are avoided. Since the bearing fault classification network based on Shapley values is connected after the AdaFNN network, the number of nodes is fixed at 2 n (n = 0, 1, 2…).

[0065] Base node network structure determination: considering the input length, various base node network structures are designed, and through the training results, a certain base node network structure is found to make the fluctuation of the network in training minimum, and the network parameters are fewer and the running speed is faster.

[0066] Input type determination: the frequency domain data of the bearing vibration acceleration signal after Fourier transformation.

[0067] A bearing fault classification network based on Shapley values, the data set composed of the vibration signals generated by bearing faults is processed by the AdaFNN network, and the processed data is taken as the characteristic value of the input of the bearing fault classification network, while the bearing fault type represented by the data set composed of the vibration signals generated by bearing faults is taken as the output of the bearing fault classification network. Through the Shapley value, the influence of each feature, i.e., the vibration signals generated by bearing faults on the final result, i.e., the influence size of the bearing fault type represented by the data set composed of the vibration signals generated by bearing faults, can be obtained.

[0068] The ShapNet method is composed of Shapley modules as the minimum unit, which simply executes the definition formula of Shapley value and brings all combinations affecting the final result into the network for calculation.

[0069] To calculate the Shapley value of each input, it is necessary to rely on the linear property of the Shapley value. According to the linear property, the exact Shapley value of all inputs is obtained by completely covering the number of inputs, which is a shallow ShapNet.

[0070] The deep ShapNet is obtained by stacking the shallow ShapNet to pass the feature parameters extracted by the bearing fault classification network as the interlayer transmission of the Shapley value.

[0071] By passing the Shapley value as the interlayer feature, i.e. the deep ShapNet, the network branch with lower Shapley value is inactivated to realize dynamic pruning, thereby improving the running speed of the basis function decomposition network.

[0072] S2, the bearing vibration signal data under multiple existing working conditions is taken as the source domain data, and the bearing vibration signal data under the new working condition is taken as the target domain data to form a training sample set.

[0073] Specifically, the bearing vibration signal data under different existing working conditions is obtained by sensors, i.e. the bearing vibration signal under the bearing inner ring fault, the bearing outer ring fault and the bearing rolling body fault as the source domain data set. The bearing vibration signal data under the new working condition is obtained by sensors, and the bearing vibration signal of the same source domain data of the three types of bearing faults is taken as the target domain data.

[0074] S3, the training sample set is input into the variable working condition bearing fault diagnosis model, only the data set under one working condition is used, i.e. the data set composed of the bearing vibration signal under the bearing inner ring fault, the bearing outer ring fault and the bearing rolling body fault under one working condition of the existing working condition is used for training to obtain the AdaSHAP model under this working condition, and the AdaSHAP model is trained in the first stage. Wherein, AdaSHAP is the fusion of AdaFNN and ShapNet;

[0075] S4, domain adaptation is performed by using the unlabeled other working condition data, i.e. the data set composed of the bearing vibration signal under the bearing inner ring fault, the bearing outer ring fault and the bearing rolling body fault under multiple existing working conditions and the unseen working condition under the new working condition, so that the AdaSHAP model is also applicable to other working conditions, and the improved MK-MMD method is used to train the AdaSHAP model in the second stage. Wherein, MK-MMD is a classic mapping-based transfer learning method.

[0076] Specifically, the basic principle of the mapping-based transfer learning method MK-MMD is to map the source domain data and the target domain data to a new representation space by using the mapping relationship realized by the AdaSHAP.

[0077]

[0078] wherein S and T represent the source domain feature and the target domain feature respectively, and G represents a Gaussian kernel function. In the training of the transfer learning network, an intermediate round number is set. Before the intermediate round number, the AdaSHAP model only receives data from the source domain, and the transfer term is not added to the loss function. After the intermediate round number, the source domain data and the unlabeled target domain data are input into the AdaSHAP model at the same time, the source domain data output the representation space feature and the prediction result, and the classification loss is calculated. The target domain data output the feature in the representation space and calculate the transfer term together with the representation space feature of the source domain data. By adding the feature term, the network is forced to modify the parameters, so that the processed source domain data feature and the target domain data feature tend to be consistent, so that the classifier can accurately classify the input from different working conditions.

[0079] Those skilled in the art will readily understand that the above description is only the preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for constructing a bearing fault diagnosis model under variable operating conditions, characterized in that, The method includes the following steps: (1) Construct a fault diagnosis model under varying working conditions. At the same time, use bearing vibration signal data under multiple existing working conditions as source domain data and bearing vibration signal data under new working conditions as target domain data to form a training sample set. The fault diagnosis model under varying working conditions includes a basis function decomposition network based on functional data analysis and a bearing fault classification network based on Shapley value. The basis function decomposition network is used to realize the decoupling process of data-driven bearing data. The bearing fault classification network is used to calculate the influence of each basis function learned by the basis function decomposition network on the fault type prediction result under cross-working conditions, thereby dynamically pruning the basis function decomposition network. (2) Input the training sample set into the variable working condition bearing fault diagnosis model, and use only the data corresponding to one of the existing multiple working conditions to train the variable working condition bearing fault diagnosis model to obtain the AdaSHAP model under this working condition, and perform the first stage training of the AdaSHAP model; wherein, the AdaSHAP model is a fusion of AdaFNN and ShapNet. (3) The AdaSHAP model is trained in the second stage. The improved MK-MMD method is used to perform domain adaptation on other working conditions in the training sample set, so that the AdaSHAP model can be applied to other working conditions at the same time. The basis function decomposition network randomly initializes a set of multilayer perceptron networks to fit basis functions. Each basis function is inner-producted with the input X to obtain a score for the input on that basis function. The scores of all basis functions are combined as a new representation input to the classification network, and are fed into the bearing fault classification network according to different task types. The bearing fault classification network is constructed based on a deep ShapNet network. The deep ShapNet network uses stacked shallow ShapNet layers to pass the Shapley value as a feature parameter extracted by the deep ShapNet network between layers. The shallow ShapNet layer calculates the accurate Shapley value of all inputs by completely covering the number of inputs based on the linear property. The second phase of training includes the following steps: S31, perform domain adaptation using other working condition data in the training sample set so that the AdaSHAP model can also be applied to other working conditions; S32, using the MK-MMD method, uses the AdaSHAP model to map source domain data and target domain data to a new representation space to compare the data of the two domains and calculates the measurement difference between the features of the two domains in the new representation space. S33. When training the AdaSHAP model, an intermediate round number is set. Before the intermediate round number, the AdaSHAP model only receives data from the source domain and no transfer term is added to the loss function. After the intermediate round number, the source domain data and the unlabeled target domain data are simultaneously input into the AdaSHAP model. The source domain data outputs the representation space features and prediction results and calculates the classification loss. The target domain data needs to obtain the features in the representation space and calculate the MK-MMD loss term together with the representation space features of the source domain data. The formula for calculating the MK-MMD loss term between features of two domains in the new representation space is as follows: In the formula, S and T represent the source domain features and target domain features, respectively, and G represents the Gaussian kernel function.

2. The method for constructing a bearing fault diagnosis model under variable operating conditions as described in claim 1, characterized in that: The first phase of training includes the following steps: S21, the basis function decomposition network is used to fit the basis function β that best suits the data distribution under the current working condition by adding a set of multilayer perceptrons before the bearing fault classification network; S22, the bearing fault classification network is used to calculate the influence of each basis function learned by the basis function decomposition network on the final judgment result under cross-working conditions, thereby pruning the basis function decomposition network and excluding basis functions with small contribution values, so that the basis function decomposition network focuses more on domain-invariant features.

3. A method for diagnosing bearing faults under varying operating conditions, characterized in that: The diagnostic method inputs the bearing vibration signal data to be diagnosed into a bearing fault diagnosis model constructed using the construction method of the bearing fault diagnosis model under variable operating conditions according to any one of claims 1-2. The bearing fault diagnosis model performs fault diagnosis based on the received bearing vibration signal data.

4. A bearing fault diagnosis system under variable operating conditions, characterized in that: The fault diagnosis system includes a memory and a processor. The memory stores a computer program, and the processor executes the variable operating condition bearing fault diagnosis method according to claim 3 when executing the computer program.

5. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores machine-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the method for constructing a variable-condition bearing fault diagnosis model as described in any one of claims 1-2 or the method for diagnosing variable-condition bearing faults as described in claim 3.

Images