An industrial equipment fault diagnosis method in a data heterogeneous environment based on adversarial federated learning

Abstract

The application discloses an industrial equipment fault diagnosis method based on an adversarial federated learning in a data heterogeneous environment, relates to the technical field of industrial equipment fault diagnosis, and effectively alleviates the influence of data heterogeneity under the premise of guaranteeing the data privacy of participants by introducing a distribution enhancement technology. A two-stage learning mechanism is adopted: firstly, each facility end extracts the potential representation of local data by using an unsupervised deep convolution model, implicit coding of original sensor signals is realized through dimension reduction processing, this step realizes model personalization and ensures data privacy; then, the influence of non-independent and identically distributed data among industrial facilities is suppressed, the data heterogeneity problem is systematically solved, and the robustness of the model performance is ensured. Finally, in the federated aggregation stage, the enhanced representation of other facility ends is fused to cooperatively train a global classification model, so that the data distribution difference is balanced, and the accuracy of the federated model in industrial fault diagnosis is improved.

Description

Technical Field

[0001] This application relates to the field of industrial equipment fault diagnosis technology, specifically to a method for industrial equipment fault diagnosis in a data heterogeneous environment based on adversarial federated learning. Background Technology

[0002] Machine learning algorithms, especially deep learning, have demonstrated remarkable effectiveness in the field of industrial equipment fault diagnosis. However, the performance of these algorithms heavily relies on large-scale, diverse, and high-quality training datasets. In industrial scenarios, data is typically distributed across multiple production facilities, each operating under unique conditions, exhibiting differences in equipment type, sensor configuration, and environmental factors. This inherent diversity leads to significant data heterogeneity, and coupled with practical issues such as privacy, proprietary data restrictions, and data transmission, centralized data collection becomes impractical.

[0003] Federated Learning (FL) offers a decentralized solution for collaborative modeling of distributed datasets while ensuring data privacy and security. This method has rapidly become a crucial technological path for overcoming data sharing barriers in the industrial sector. Patent document CN115438714A discloses a method that continuously monitors and records the operational state characteristics of heterogeneous clients participating in federated learning, constructing a time series of state features based on fuzzy entropy theory to quantify the uncertainty or volatility of client operational states. Patent document CN117808078A discloses a technical solution that downloads feature extractor aggregation parameters and global classifier parameters from the server and then optimizes the local model using gradient descent. Patent document CN116070713A, from the perspective of data imbalance, discloses the introduction of an interpretability mechanism and the use of a structural similarity index for classification evaluation. Based on this, it infers imbalanced clients and adjusts the parameters of imbalanced clients to reduce the gradient distance between their model parameters and the historical global aggregated model parameters.

[0004] However, traditional federated learning typically assumes that the data from each facility satisfies the IID (Independent and Identically Distributed) condition, an assumption that is often difficult to hold in real-world industrial scenarios. In practice, factors such as differences in equipment operating environments, variations in mechanical wear characteristics, and sensor performance drift often lead to data exhibiting non-independent and identically distributed characteristics, which severely weakens the diagnostic performance of federated models. Summary of the Invention

[0005] In order to solve the above-mentioned technical problems, this application proposes the following technical solution:

[0006] In a first aspect, embodiments of this application provide a method for fault diagnosis of industrial equipment in a heterogeneous data environment based on adversarial federated learning, including:

[0007] Collect time-series data on equipment failures and normal operation from each industrial field monitoring device and generate a spectrum diagram;

[0008] Features are extracted from the time-frequency representation (TFR) of the spectrogram, and local feature learning is performed using a shrinking autoencoder (CAE).

[0009] The global generative adversarial network (GAN) is trained using signals obtained from local feature learning, enabling the global GAN ​​to learn the general data distribution for each fault class.

[0010] After the global GAN ​​is trained, its optimized parameters will be distributed to the local GANs in each industrial site. The local GANs will generate synthetic data to enhance the local data distribution based on the bias parameters calculated from the site data summary statistics.

[0011] Under the federated learning framework, a two-dimensional convolutional neural network is used to construct a global classification model, which is then optimized through a learnable activation function to achieve accurate classification of the operating status of industrial equipment.

[0012] The additive secret sharing technology enables the exchange of model parameters between each industrial site and the central server, and synchronously updates the global classification model.

[0013] In one possible implementation, the step of collecting time-series data on equipment failures and normal operation from each industrial field monitoring device and generating a spectrum diagram includes:

[0014] Sampling was performed on the time series of equipment failures and normal operation monitored by each industrial field monitoring device;

[0015] The sampled time-series data is converted into a spectrum, i.e., the absolute value of the Short-Time Fourier Transform (STFT): This is represented as follows:

[0016]

[0017]

[0018] in: It is the actual time. It is runtime. It is the signal to be transformed. It's a window function. It is an even-symmetric window centered at time t. The resulting spectrum provides the time-frequency representation (TFR) of the signal.

[0019] In one possible implementation, each industrial site corresponds to a shrinking autoencoder (CAE), which is trained separately based on the input data corresponding to each category. The CAE consists of an encoder and a decoder. The encoder maps the input data to a latent space to convert the data into a low-dimensional representation. The decoder reconstructs the original signal from the latent representation, ensuring that the encoded features retain the basic information of the input.

[0020] In one possible implementation, the step of extracting features from the time-frequency representation (TFR) of the spectrogram and performing local feature learning through a shrinking autoencoder (CAE) includes:

[0021] For dataset Training instance vectors Define the latent representation and reconstruction of time-frequency representation The definition is as follows:

[0022]

[0023]

[0024] in, It is a weight matrix. It is a bias vector. Indicates the activation function;

[0025] Reconstruction losses The representation is as follows:

[0026]

[0027] A penalty term is added to the reconstruction loss, which is derived from the Frobenius norm of the hidden layer activation Jacobian matrix:

[0028]

[0029] in: It is the square of the Frobenius norm of the Jacobian matrix. and These are the input vectors. The activation value of the b-th hidden unit and the a-th feature value;

[0030] Elements of the Jacobian matrix Defined as:

[0031]

[0032] The total cost function for training a CAE is:

[0033]

[0034] in, To control the relative weights of the penalty items;

[0035] After training, the encoder is decoupled to transform the input TFR from the original dimension to a lower dimension to ensure privacy protection.

[0036] In one possible implementation, the GAN includes a generator and a discriminator. The generator approximates the real data distribution by generating synthetic data that is similar to the real data distribution. The discriminator distinguishes between the actual training data and the synthetic data generated by the generator. As adversarial training progresses, the distribution of the synthetic data generated by the generator gradually converges to the actual data distribution. This process continues until the discriminator can no longer reliably distinguish between real samples and synthetic samples.

[0037] In one possible implementation, the signal obtained by learning local features is used to train a global generative adversarial network (GAN), enabling the global GAN ​​to learn the general data distribution for each fault class, including:

[0038] The global GAN ​​generates feature maps similar to the real conversion signals, using fault labels. As a condition, minimize non-IID gaps;

[0039] In the generator, prior input noise With fault labels The combined representation is then used to update the GAN loss function as follows:

[0040]

[0041] in: and These are the probability distributions from the real data. Probability distribution of the generator The expected value of x in the sample is taken. and These represent the outputs of the generator and discriminator, respectively, which are constrained by fault category labels.

[0042] In one possible implementation, the global classification model is constructed using a two-dimensional convolutional neural network within the federated learning framework, and optimized through a learnable activation function to achieve accurate classification of the operating status of industrial equipment, including:

[0043] Each industrial site maps the summary statistics sent by the server, which includes: the tag distribution of all participating sites;

[0044] Each industrial site uses its locally trained conditional generator to augment the data to balance the class distribution, with the number of augmented instances based on the percentage of the total number of instances ||Xc|| of all classes to be generated. and determine when to use The threshold γ;

[0045] For each site, identify the category with the most data instances and calculate the instance differences for other categories;

[0046] If the difference is less than If the condition is met, then beta-control is used to enhance the category; otherwise, the exact difference is used for enhancement.

[0047] After balancing the data locally, a supervised convolutional model is trained in a federated setting to predict failure conditions.

[0048] In one possible implementation, the exchange of model parameters between each industrial site and the central server via additive secret sharing technology, and the synchronous updating of the global classification model, includes:

[0049] During the federated training process, each local field trains its own model and periodically uploads the model parameters to the central server through additive secret sharing technology.

[0050] Central server The parameters of each site are weighted and averaged to update the global classification model synchronously.

[0051] In one possible implementation, the central server... The parameters of each site are weighted and averaged to update the global classification model synchronously, including:

[0052] First, local parameters are used to generate fragments in the industrial field. Discriminator parameters It is split into K partitions:

[0053]

[0054] The shards are then distributed to the central server and other industrial sites. Send to the central server Distribute to other K−1 sites (i≥2);

[0055] When setting security parameters, the server only calculates the sum of the first fragment:

[0056]

[0057] Local aggregation of remaining fragments in each industrial site:

[0058]

[0059] Finally, global parameter restoration is performed, and the local aggregation results are broadcast to all industrial sites. Output the final global parameters:

[0060] .

[0061] Secondly, embodiments of this application provide an industrial equipment fault diagnosis system based on adversarial federated learning in a data heterogeneous environment, including:

[0062] The acquisition module is used to collect time-series data on equipment failures and normal operation from each industrial field monitoring device and generate a spectrum diagram.

[0063] The feature learning module is used to extract features from the time-frequency representation (TFR) of the spectrogram and perform local feature learning through a shrinking autoencoder (CAE).

[0064] The training module is used to train a global generative adversarial network (GAN) using signals obtained from local feature learning, so that the global GAN ​​learns the general data distribution for each fault class.

[0065] The data augmentation module is used to distribute the optimized parameters of the global GAN ​​to the local GANs in each industrial site after the global GAN ​​training is completed. The local GANs generate synthetic data to enhance the local data distribution based on the bias parameters calculated from the field data summary statistics.

[0066] The model building module is used to build a global classification model using a two-dimensional convolutional neural network under the federated learning framework. It is optimized through a learnable activation function to achieve accurate classification of the operating status of industrial equipment.

[0067] The model update module is used to exchange model parameters between each industrial site and the central server through additive secret sharing technology, and to synchronously update the global classification model.

[0068] In this embodiment, by introducing distributional augmentation technology, the impact of data heterogeneity is effectively mitigated while ensuring the data privacy of participating parties. A two-stage learning mechanism is adopted: First, each facility uses an unsupervised deep convolutional model to extract latent representations of its local data, achieving implicit encoding of the original sensor signals through dimensionality reduction. This step achieves both model personalization and ensures data privacy. Subsequently, by suppressing the influence of non-independent and identically distributed data among industrial facilities, the data heterogeneity problem is systematically addressed, ensuring the robustness of model performance. Finally, in the federated aggregation stage, the augmented representations from other facilities are fused for global classification model collaborative training, thereby balancing data distribution differences and improving the accuracy of the federated model in industrial fault diagnosis. Attached Figure Description

[0069] Figure 1 A flowchart illustrating a method for diagnosing industrial equipment faults in a heterogeneous data environment based on adversarial federated learning, provided for an embodiment of this application;

[0070] Figure 2 A schematic diagram illustrating the non-independent and identically distributed characteristics of data from different locations provided in the embodiments of this application;

[0071] Figure 3 An implementation framework diagram for collaborative training and updating of the global classification model using the federated learning framework provided in this application embodiment;

[0072] Figure 4 This is a schematic diagram of an industrial equipment fault diagnosis system based on adversarial federated learning in a heterogeneous data environment, which is provided as an embodiment of this application. Detailed Implementation

[0073] The present solution will now be described in conjunction with the accompanying drawings and specific embodiments.

[0074] See Figure 1 The industrial equipment fault diagnosis method based on adversarial federated learning in a heterogeneous data environment provided in this embodiment includes:

[0075] S101 collects time-series data on equipment failures and normal operation from each industrial field monitoring device and generates a spectrum diagram.

[0076] In this embodiment, time-series data on equipment failures and normal operation obtained from monitoring equipment at each industrial site are collected, and then a spectrum diagram, i.e., the absolute value of the STFT, is generated. The STFT of the time-series signal is represented as follows:

[0077]

[0078] in It's a window function that takes the absolute value of STFT:

[0079]

[0080] in: It is the actual time. It is runtime. It is the signal to be transformed. It's a window function. It is an even-symmetric window centered at time t. The resulting spectrum provides the time-frequency representation (TFR) of the signal.

[0081] S102, extract features from the time-frequency representation (TFR) of the spectrogram and perform local feature learning through a shrinking autoencoder (CAE).

[0082] To extract TFR features from the signal obtained from S101, this embodiment designs a shrinking autoencoder (CAE) for each industrial site. This CAE is trained separately based on the input data corresponding to each category, rather than being trained collectively on the input data of all categories. The CAE consists of an encoder and a decoder. The encoder maps the input data to a latent space, effectively converting the data into a low-dimensional representation. Because the autoencoder is trained independently for each fault category, the resulting latent space captures the unique distribution characteristics of each category. The decoder then reconstructs the original signal from the latent representation, ensuring that the encoded features retain the basic information of the input. Specifically, for the dataset... Training instance vectors Potential representation and reconstructed TFR The definition is as follows:

[0083]

[0084]

[0085] in, It is a weight matrix. It is a bias vector. Represents the activation function. Reconstruction loss. The representation is as follows:

[0086]

[0087] To enhance robustness, this invention incorporates a penalty term into the reconstruction loss, which is derived from the Frobenius norm of the hidden layer activation Jacobian matrix:

[0088]

[0089] in: It is the square of the Frobenius norm of the Jacobian matrix. and These are the input vectors. The activation value of the b-th hidden unit and the a-th feature value;

[0090] Elements of the Jacobian matrix Defined as:

[0091]

[0092] The total cost function for training a CAE is:

[0093]

[0094] in, The relative weights of the penalty terms are used to control the weights. After training, the encoder is decoupled to transform the input TFR from the original dimension d to a lower dimension. To ensure privacy protection.

[0095] S103, the signal obtained by local feature learning is used to train the global generative adversarial network (GAN), so that the global GAN ​​learns the general data distribution of each fault class.

[0096] The signal after S102 transformation is used to train a global GAN ​​to handle non-IID distributions. In this embodiment, the GAN mainly consists of two key components: a generator and a discriminator. For ease of description, it is assumed that the training dataset is represented as follows: ,in Indicates features, This represents the corresponding label. The goal of the generator is to generate data that matches the real data distribution. Extremely similar synthetic data To approximate the true data distribution ,Right now Conversely, the discriminator The task is to distinguish between actual training data and synthetic data generated by the generator.

[0097] In this embodiment, GAN is used to model the conditional distribution of transformed vibration signals in different industrial sites. By generating feature maps similar to the actual transformed signals, fault labels are generated. As a condition, minimize the non-IID gap.

[0098] Both the generator and discriminator of a GAN are designed based on fault category labels. In the generator, the prior input noise... With fault labels The combined representations are then formed. The GAN loss function is updated as follows:

[0099]

[0100] in: and These are the probability distributions from the real data. Probability distribution of the generator The expected value of x in the sample is taken. and These represent the outputs of the generator and discriminator, respectively, which are constrained by fault category labels.

[0101] S104 After the global GAN ​​training is completed, its optimized parameters will be distributed to the local GANs in each industrial site. The local GANs will generate synthetic data to enhance the local data distribution based on the deviation parameters calculated from the site data summary statistics.

[0102] See Figure 2 The characteristics and distributions differ across different data sources and locations. Assume the industrial data for the k-th facility is from a local distribution. Independent and identically distributed (IID) data sampled from the middle. Under the ideal IID assumption, the distribution of all client data should satisfy... (For any k). The global data distribution can then be represented as a weighted mixed distribution:

[0103]

[0104] However, in real-world scenarios, it is almost impossible to satisfy the independent and identically distributed (IID) assumption. That is, there exists at least one k such that... This makes fault detection and prediction extremely challenging when multiple customers are collaboratively training a global classification model. To quantify the degree of non-independent and identically distributed faults, let... and Represents the global loss function and local loss function The average value. In this embodiment, the heterogeneity measure is defined as:

[0105]

[0106] If the client's data distribution is independent and identically distributed (IID), then as the data sample size increases, It will converge to zero. Conversely, in the non-independent identically distributed (non-IID) setting, The value remains non-zero, reflecting the degree of heterogeneity among the client data distributions. The data distributions of all participants do not satisfy the independent and identically distributed assumption. Under these conditions, the standard FedAvg algorithm fails to converge satisfactorily because of the weighted local average. They could be very different.

[0107] In this embodiment, after the global GAN ​​completes training, the model learns the common distribution patterns of vibration signals. Each local GAN ​​generates new samples based on its own aggregated data statistics, effectively mitigating the differences caused by non-independent and identically distributed signals.

[0108] S105 uses a two-dimensional convolutional neural network to build a global classification model under the federated learning framework. It is optimized by a learnable activation function to achieve accurate classification of the operating status of industrial equipment.

[0109] After training the global GAN, this embodiment trains a global classification model for the federated fault prediction task. Prior to training, each industrial site maps aggregated statistics sent by the server. This includes the label distribution of all participating sites (e.g., fault type and severity). Each site uses its locally trained conditional generator to augment the data to balance the class distribution.

[0110] The number of enhanced instances is based on two parameters: It defines the total number of instances of all categories to be generated ||X c The percentage of ||; and a threshold. It determines when to apply. For each site, identify the category with the most data instances and calculate the instance differences for other categories. If the difference is less than... Then use Boost that category. Otherwise, boost using exact differences. After balancing the data locally, train a supervised convolutional model in a federated setting to predict failure conditions.

[0111] S106, The model parameters of each industrial site and the central server are exchanged through additive secret sharing technology, and the global classification model is updated synchronously.

[0112] See Figure 3 In an industrial setting, assuming N facilities collaboratively train a global classification model using a federated learning framework, its distributed optimization problem can be formally represented as:

[0113]

[0114] in Indicates the first The weight of each facility, and satisfying , Each facility Having a dataset Its samples are derived from the distribution of specific fields. Used for estimating facilities The local objective function for empirical risk is expressed as:

[0115]

[0116] in, It is the specified loss function.

[0117] Typically, the formula for distributed optimization problems involves applying probabilities in each communication round t. A subset of facilities is randomly selected, and optimization is performed locally through multiple iterations of optimization algorithms such as stochastic gradient descent (SGD). The central server then assigns the weights of the current global classification model. Distribute to all clients. Then, each facility models its local application. The above will be updated E>1 times, as follows:

[0118]

[0119] in, Indicates iteration The learning rate at that time. Finally, the central server aggregates and updates the local model. To form a new global classification model .

[0120] During federated training, each local field trains its own GAN, periodically sharing model parameters through additive secret sharing technology. and Uploaded to the central server. The server then... The parameters of each site are weighted and averaged to update the global classification model synchronously.

[0121] First, local parameters are used to generate fragments on-site. Discriminator parameters It is split into K partitions:

[0122]

[0123] Then the fragments are distributed to the central server and other sites. Send to the central server Distribute to the other K−1 sites (i≥2).

[0124] Next, a set of security parameters is performed; the server only calculates the sum of the first fragment.

[0125]

[0126] Local aggregation of remaining fragments for each site (k-calculation):

[0127]

[0128] Finally, global parameter restoration is performed, and all on-site broadcasts are aggregated locally. Final global parameters:

[0129]

[0130] In the framework of federated conditional generative adversarial networks, additive secret sharing technology is deeply integrated with the algorithm through the above methods to ensure the privacy of parameter exchange in the fault diagnosis scenario of industrial rotating equipment.

[0131] Corresponding to the above embodiment of the method for diagnosing industrial equipment faults in a heterogeneous data environment based on adversarial federated learning, this application also provides an embodiment of an industrial equipment fault diagnosis system in a heterogeneous data environment based on adversarial federated learning.

[0132] See Figure 4 This embodiment of the industrial equipment fault diagnosis system 20 based on adversarial federated learning in a heterogeneous data environment includes:

[0133] The acquisition module 201 is used to acquire time series data of equipment failure and normal operation obtained by each industrial field monitoring device and generate a spectrum diagram.

[0134] Feature learning module 202 is used to extract features from the time-frequency representation (TFR) of the spectrogram and perform local feature learning through a shrinking autoencoder (CAE).

[0135] Training module 203 is used to train a global generative adversarial network (GAN) using signals obtained from local feature learning, so that the GAN learns the general data distribution for each fault class.

[0136] The data augmentation module 204 is used to distribute the optimized parameters of the global conditional GAN ​​to the local GANs in each industrial site after the global conditional GAN ​​training is completed. The local GANs generate synthetic data to enhance the local data distribution based on the bias parameters calculated from the field data summary statistics.

[0137] The model building module 205 is used to build a global classification model using a two-dimensional convolutional neural network under the federated learning framework. It is optimized by a learnable activation function to achieve accurate classification of the operating status of industrial equipment.

[0138] The model update module 206 is used to exchange model parameters between each industrial site and the central server through additive secret sharing technology, and to update the global classification model synchronously.

[0139] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects have an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, and c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0140] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. The protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for fault diagnosis of industrial equipment in a heterogeneous data environment based on adversarial federated learning, characterized in that, include: Collect time-series data on equipment failures and normal operation from each industrial field monitoring device and generate a spectrum diagram; Each industrial site corresponds to a shrinking autoencoder (CAE). The CAE is trained separately based on the input data corresponding to each category. The CAE consists of an encoder and a decoder. The encoder maps the input data to a latent space to convert the data into a low-dimensional representation. The decoder reconstructs the original signal from the latent representation, ensuring that the encoded features retain the basic information of the input. Features are extracted from the time-frequency representation (TFR) of the spectrogram, and local feature learning is performed using a shrinking autoencoder (CAE), including: For dataset Training instance vectors Define the latent representation and reconstruction of time-frequency representation The definition is as follows: in: It is a weight matrix. It is a bias vector. Indicates the activation function; Reconstruction losses The representation is as follows: A penalty term is added to the reconstruction loss, which is derived from the Frobenius norm of the hidden layer activation Jacobian matrix: in: It is the square of the Frobenius norm of the Jacobian matrix. and These are the input vectors. The activation value of the b-th hidden unit and the a-th feature value; Elements of the Jacobian matrix Defined as: The total cost function for training a CAE is: in, To control the relative weights of the penalty items; After training, the encoder is decoupled to transform the input TFR from the original dimension to a lower dimension to ensure privacy protection. The global generative adversarial network (GAN) is trained using signals obtained from local feature learning, enabling the global GAN ​​to learn the general data distribution for each fault class. After the global GAN ​​is trained, its optimized parameters will be distributed to the local GANs in each industrial site. The local GANs will generate synthetic data to enhance the local data distribution based on the bias parameters calculated from the site data summary statistics. Under the federated learning framework, a two-dimensional convolutional neural network is used to construct a global classification model, which is then optimized through a learnable activation function to achieve accurate classification of the operating status of industrial equipment. The additive secret sharing technology enables the exchange of model parameters between each industrial site and the central server, and synchronously updates the global classification model.

2. The method for fault diagnosis of industrial equipment in a heterogeneous data environment based on adversarial federated learning according to claim 1, characterized in that, The process of collecting time-series data on equipment failures and normal operation from each industrial field monitoring device and generating a spectrum diagram includes: Sampling was performed on the time series of equipment failures and normal operation monitored by each industrial field monitoring device; The sampled time-series data is converted into a spectrum, i.e., the absolute value of the Short-Time Fourier Transform (STFT): This is represented as follows: in: It is the actual time. It is runtime. It is the signal to be transformed. It's a window function. It is an even-symmetric window centered at time t. The resulting spectrum provides the time-frequency representation (TFR) of the signal.

3. The method for fault diagnosis of industrial equipment in a heterogeneous data environment based on adversarial federated learning according to claim 1, characterized in that, The GAN includes a generator and a discriminator. The generator approximates the real data distribution by generating synthetic data that is similar to the real data distribution. The discriminator distinguishes between the actual training data and the synthetic data generated by the generator. As adversarial training progresses, the distribution of synthetic data generated by the generator gradually converges to the actual data distribution. This process continues until the discriminator can no longer reliably distinguish between real samples and synthetic samples.

4. The method for fault diagnosis of industrial equipment in a heterogeneous data environment based on adversarial federated learning according to claim 3, characterized in that, The signal obtained through local feature learning is used to train a global generative adversarial network (GAN), enabling the global GAN ​​to learn the general data distribution for each fault class, including: The global GAN ​​generates feature maps similar to the real conversion signals, using fault labels. As a condition, minimize non-IID gaps; In the generator, prior input noise With fault labels The combined representation is then used to update the GAN loss function as follows: in: and These are the probability distributions from the real data. Probability distribution of the generator The expected value of x in the sample is taken. and These represent the outputs of the generator and discriminator, respectively, which are constrained by fault category labels.

5. The method for fault diagnosis of industrial equipment in a heterogeneous data environment based on adversarial federated learning according to claim 1, characterized in that, The method described above employs a two-dimensional convolutional neural network to construct a global classification model within the federated learning framework, and optimizes it using a learnable activation function to achieve accurate classification of the operating status of industrial equipment. This includes: Each industrial site maps the summary statistics sent by the server, which includes: the tag distribution of all participating sites; Each industrial site uses its locally trained conditional generator to augment the data to balance the class distribution, with the number of augmented instances based on the percentage of the total number of instances ||Xc|| of all classes to be generated. and determine when to use The threshold γ; For each site, identify the category with the most data instances and calculate the instance differences for other categories; If the difference is less than If the condition is met, then beta-control is used to enhance the category; otherwise, the exact difference is used for enhancement. After balancing the data locally, a supervised convolutional model is trained in a federated setting to predict failure conditions.

6. The method for fault diagnosis of industrial equipment in a heterogeneous data environment based on adversarial federated learning according to claim 1, characterized in that, The method of exchanging model parameters between various industrial sites and the central server through additive secret sharing technology, and synchronously updating the global classification model, includes: During the federated training process, each local field trains its own model and periodically uploads the model parameters to the central server through additive secret sharing technology. Central server The parameters of each site are weighted and averaged to update the global classification model synchronously.

7. The method for fault diagnosis of industrial equipment in a heterogeneous data environment based on adversarial federated learning according to claim 6, characterized in that, The central server The parameters of each site are weighted and averaged to update the global classification model synchronously, including: First, local parameters are used to generate fragments in the industrial field. Discriminator parameters It is split into K partitions: The shards are then distributed to the central server and other industrial sites. Send to the central server Distribute to other There are 2 stations, i ≥ 2; When setting security parameters, the server only calculates the sum of the first fragment: Local aggregation of remaining fragments in each industrial site: Finally, global parameter restoration is performed, and the local aggregation results are broadcast to all industrial sites. Output the final global parameters: 。 8. A fault diagnosis system for industrial equipment in a heterogeneous data environment based on adversarial federated learning, characterized in that, include: The acquisition module is used to collect time-series data on equipment failures and normal operation from each industrial field monitoring device and generate a spectrum diagram. Each industrial site corresponds to a shrinking autoencoder (CAE). The CAE is trained separately based on the input data corresponding to each category. The CAE consists of an encoder and a decoder. The encoder maps the input data to a latent space to convert the data into a low-dimensional representation. The decoder reconstructs the original signal from the latent representation, ensuring that the encoded features retain the basic information of the input. The feature learning module is used to extract features from the time-frequency representation (TFR) of the spectrogram and perform local feature learning through a shrinking autoencoder (CAE), including: For dataset Training instance vectors Define the latent representation and reconstruction of time-frequency representation The definition is as follows: in: It is a weight matrix. It is a bias vector. Indicates the activation function; Reconstruction losses The representation is as follows: A penalty term is added to the reconstruction loss, which is derived from the Frobenius norm of the hidden layer activation Jacobian matrix: in: It is the square of the Frobenius norm of the Jacobian matrix. and These are the input vectors. The activation value of the b-th hidden unit and the a-th feature value; Elements of the Jacobian matrix Defined as: The total cost function for training a CAE is: in, To control the relative weights of the penalty items; After training, the encoder is decoupled to transform the input TFR from the original dimension to a lower dimension to ensure privacy protection. The training module is used to train a global generative adversarial network (GAN) using signals obtained from local feature learning, so that the global GAN ​​learns the general data distribution for each fault class. The data augmentation module is used to distribute the optimized parameters of the global GAN ​​to the local GANs in each industrial site after the global GAN ​​training is completed. The local GANs generate synthetic data to enhance the local data distribution based on the bias parameters calculated from the field data summary statistics. The model building module is used to build a global classification model using a two-dimensional convolutional neural network under the federated learning framework. It is optimized through a learnable activation function to achieve accurate classification of the operating status of industrial equipment. The model update module is used to exchange model parameters between each industrial site and the central server through additive secret sharing technology, and to synchronously update the global classification model.

Images