Semi-supervised federated learning feature extraction equipment life prediction method

Abstract

The invention belongs to the technical field of equipment life prediction, and more specifically relates to a semi-supervised federated learning feature extraction equipment life prediction method. The method comprises the following steps: generating a randomly initialized unsupervised global federated feature extraction model at a central server, selecting N participating enterprise clients from a client pool by the server, and sending the initialized unsupervised global federated feature extraction model to the participating enterprise clients; n clients are trained at the same time, finally, the trained unsupervised local federated feature extraction models are uploaded to a central server and aggregated, and the trained unsupervised federated feature extraction models are obtained through continuous iteration; and finally, finely adjusting the trained unsupervised federated feature extraction model to realize final mapping of the RUL. The problems that an existing prediction method is high in cost, fault data participating in enterprise monitoring are limited, and privacy leakage exists are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of equipment life prediction, and more specifically, relates to a semi-supervised federated learning feature extraction equipment life prediction method. Background Art

[0002] With the significant development of high-end equipment technology, prediction and health management are widely adopted to ensure the safety of advanced equipment. Among them, the remaining service life prediction is a crucial task, which can clearly indicate the health status of equipment operation, assist enterprises to make timely decisions, implement maintenance or replacement strategies, and minimize the possibility of equipment failure and disaster events. However, the remaining life prediction methods based on expert knowledge or physical models have problems such as high economic cost and low accuracy.

[0003] As the complexity of modern equipment continues to increase, the use of deep learning methods to capture local abstract features and time dependencies of equipment degradation from monitoring data has become a method that has attracted much attention. However, although this type of supervised learning method has achieved some competitive prediction results, preparing sensor data labels requires a high level of expert knowledge, which is costly, and pure unsupervised learning is not suitable for regression tasks. To address these problems, semi-supervised learning architectures have been studied to reduce label costs while improving the effectiveness of life prediction.

[0004] In addition, equipment operation failure data may contain some extremely sensitive corporate privacy data. If the traditional centralized semi-supervised method is used to directly collect and utilize data from different companies, there may be a risk of privacy leakage. The amount of fault monitoring data of a single company is small and insufficient to support high-precision prediction model training. Therefore, in the context of privacy protection, there is still a lack of accurate methods for predicting the remaining life of equipment, which can reduce corporate costs while improving the reliability of equipment prediction.

[0005] Chinese patent document CN114218872A discloses a method for predicting the remaining useful life based on a DBN-LSTM semi-supervised joint model. First, DBN is used to perform data fusion on multi-dimensional time series historical information to extract the health indicator characteristics of the machine; wherein the construction of DBN is divided into two stages, first using a large number of unlabeled data sets for unsupervised learning pre-training, and then using labeled data sets for fine-tuning. The extracted health indicator time series is input into LSTM to calculate the current remaining useful life. Furthermore, two neural networks connected in series are simultaneously supervised trained to update the weights using a labeled sample data set, and the prediction results are evaluated in the validation set and the parameters are adjusted to obtain an optimized model. However, this method requires centralized data collection and training. The company's historical operating data contains privacy information, and centralized training may lead to the leakage of the operating status of key equipment.

[0006] In view of this, the present invention designs a semi-supervised federated learning feature extraction equipment life prediction method. This method uses the federated learning mechanism to ensure the privacy and security of local data of participating enterprises, and realizes the key information mining of original data through a multiple feature extraction architecture to improve the accuracy of equipment remaining life prediction. Summary of the invention

[0007] The present invention aims to overcome at least one defect of the above-mentioned prior art and provide a semi-supervised federated learning feature extraction equipment life prediction method to solve the high cost of existing prediction methods and the problem that participating enterprises have limited monitoring fault data and privacy leakage.

[0008] The detailed technical scheme of the present invention is as follows:

[0009] A semi-supervised federated learning feature extraction equipment life prediction method, the method comprising:

[0010] S1. Generate a randomly initialized unsupervised global federated feature extraction model Φ on the central server t ,The server selects N participating enterprise clients from the client pool and sends the initialized unsupervised global federated feature extraction model to them;

[0011] The unsupervised global federated feature extraction model integrates the convolutional autoencoder CAE into the Wasserstein generative adversarial network-gradient penalty WGAN-GP as a multiple unsupervised feature learning architecture constructed by the generator to extract high-level representation information from the raw data;

[0012] S2, N clients are trained simultaneously. On each client participating in the training, its local model is first initialized to the unsupervised global federated feature extraction model sent by the server. Then, the processed local sensor monitoring data is used to train the client's unsupervised local federated feature extraction model. The training is repeated until the preset local training rounds M or the error threshold and other termination conditions are met. Finally, the trained unsupervised local federated feature extraction model is uploaded to the central server.

[0013] S3. After the training of N enterprise clients is completed, the parameters of the unsupervised local federated feature extraction model are uploaded to the central server and aggregated; then the next round of federated learning training begins, and the newly formed unsupervised global federated feature extraction model is sent to the participating N enterprise clients again for a new round of federated training iterations until the preset maximum federated learning training round P or the error threshold is reached, and the training is terminated to obtain the trained unsupervised federated feature extraction model;

[0014] S4, in the fine-tuning stage, for enterprise clients with equipment prediction needs, the trained unsupervised federated feature extraction model is used to map local sensor monitoring data into high-level abstract features; then, the timing degradation trend in key information is further extracted through two consecutive GRU network layers; finally, the output of the GRU network is passed into two fully connected layers to achieve the final mapping to the remaining useful life RUL.

[0015] Preferably, according to the present invention, the WGAN-GP is composed of two adversarial models: a CAE generator G and a discriminator D, wherein the CAE generator G is used to learn the true distribution of data and generate samples that can successfully deceive the discriminator D. The formula is as follows:

[0016]

[0017] In formula (1), Represents the loss function of the generator participating in WGAN-GP, express is the data distribution P generated from the generator G g The data points generated by sampling are represents the output of the discriminator for the sampled data points, and E represents the expected value of the specified distribution sample;

[0018] The discriminator D is used to estimate the probability that the training sample belongs to the real data rather than the data generated by the CAE generator G, and correctly identify the generated samples from the real data and the generator;

[0019] The generator G and the discriminator D are optimized alternately with each other.

[0020] Preferably, the processed local sensor monitoring data refers to, first, normalizing the selected sensor monitoring data, selecting the minimum and maximum normalization method, and unifying the local sensor monitoring data into the range of [-1, 1] to eliminate the influence of different dimensions on the prediction results. The formula is as follows:

[0021]

[0022] In formula (2), represents the normalized value, x j,k represents the raw data of the jth data point of the kth sensor, and are the minimum and maximum values ​​of the kth sensor respectively;

[0023] Then, in order to make full use of the local original multivariate time series data, the sliding window method is used to process the normalized data. The sliding window is a two-dimensional matrix [x window ,x feature], where x window Indicates the size of the sliding window, x feature represents the number of selected input features, and the sliding window step size is set to 1 to maximize data usage.

[0024] Preferably, the training of the client unsupervised local model comprises the following specific steps:

[0025] S21, the processed data enters the CAE generator G in WGAN-GP as the original data, completes the data reconstruction, and generates samples that can successfully deceive the discriminator D;

[0026] S22, the real data and the samples generated by the generator enter the discriminator D, and the discriminator D identifies the real data and the samples generated by the generator;

[0027] S23. Update client model parameters using RMSprop optimizer

[0028] Preferably, the data reconstruction is first encoded in CAE to realize the projection of the original data to a lower dimension, and the formula is as follows:

[0029]

[0030] In formula (3), ε represents the feature after dimensionality reduction, is a nonlinear activation function, is the weight between the i-th neuron in layer n and the j-th neuron in layer n+1; in addition, represents the data x corresponding to the i-th matrix of network layer n, and is the bias term in the encoding layer;

[0031] Then the decoding is performed. The decoding process is to obtain the reconstructed data from the encoded feature ε

[0032]

[0033] In formula (4), T represents transposition, Represents the bias term of the decoding layer;

[0034] Finally, the reconstruction error loss of CAE is calculated, and the loss function L CAE as follows:

[0035]

[0036] In formula (5), ||.||1 represents the L1 norm, M is the total number of local samples, and x k Represents the original data, Represents the data reconstructed by CAE. Since CAE is also the generator of WGAN-GP, The training goal of CAE is to minimize the loss function. During this process, a set of feature representations ε can be obtained through the encoding process of CAE.

[0037] Preferably, according to the present invention, during the training of the unsupervised local model, the reconstruction error and the discriminator error of the generator are calculated, and the total loss function of the CAE generator G is:

[0038]

[0039] In formula (6), L G (x i ) represents the total loss function of the multiple feature extraction generator of the present invention, express is the data distribution P generated from the generator G g The generated data points are used in ; E represents the expected value of the sample of the specified distribution;

[0040] The calculation formula of the discriminator error is as follows:

[0041]

[0042] In formula (7), L D represents the discriminator error, L real (x) represents the loss calculation for the real sample, L penalty (x,x * ) represents the gradient penalty term, which is used to constrain the discriminator to satisfy 1-Lipschitz continuity, x~P r Indicates that x is from the real data distribution P r The sampled points; D(x) represents the output of the discriminator for the real data points; Indicates that at the real data point x~P r and generate data points The distribution obtained by uniform sampling on the straight line is used to calculate the gradient penalty; x * represents the distribution obtained from uniform sampling The sampling points in Denotes the discriminator output D(x * ) for input x * The gradient norm of is the Euclidean norm of the gradient; λ is the weight hyperparameter of the gradient penalty, which controls the influence of the gradient penalty on the total loss.

[0043] Preferably, according to the present invention, the aggregation process uses the FedAvg algorithm to average the received local model weights and form a new global model, namely:

[0044]

[0045] In formula (8), Φ t+1 represents the central server model parameters of the t+1th federated learning iteration training, N represents the number of clients participating in the training, represents the client model parameters of the ith client and the tth iteration, t∈(1,P), and FedAvg represents the federated averaging algorithm.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] (1) The present invention establishes a new semi-supervised federated learning feature framework in the field of equipment remaining life prediction, which solves the problem of limited fault monitoring data of participating enterprises and reduces the risk of data leakage. It allows participating enterprises to locally utilize unlabeled degradation data through unsupervised federated learning to obtain advanced feature extraction pre-trained models. For enterprises with prediction needs, a small number of labels are prepared in the fine-tuning stage, and finally the mapping of pre-trained model captured features to equipment remaining life estimation is realized.

[0048] (2) The present invention utilizes the fusion architecture of autoencoder and generative adversarial network to extract key feature information of local original monitoring data and improve the accuracy of equipment remaining life prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flow chart of the semi-supervised federated learning feature extraction equipment life prediction method described in the present invention.

[0050] Figure 2 It is a central server-client architecture diagram of the semi-supervised federated learning feature extraction prediction model described in the present invention.

[0051] Figure 3 It is a comparison chart between the actual RUL and the predicted RUL of the first subset FD001 in the C-MAPSS data set in an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The present disclosure is further described below in conjunction with the accompanying drawings and embodiments.

[0053] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present disclosure belongs.

[0054] Embodiment 1,

[0055] This embodiment provides a semi-supervised federated learning feature extraction device life prediction method, such as Figure 1 As shown, the method includes:

[0056] S1. Generate a randomly initialized unsupervised global model Φ on the central server t ,The unsupervised global federated feature extraction model fuses the convolutional autoencoder CAE into the Wasserstein generative adversarial network - gradient penalty WGAN-GP as a generator to construct a multiple unsupervised feature learning architecture to extract high-level representation information from the raw data.,The server selects N participating enterprise clients from the client pool and sends the,initialized unsupervised global federated feature extraction model to them.

[0057] S2, N clients are trained simultaneously. On each client participating in the training, first initialize its local model to the global unsupervised model sent by the server, and then use local sensors to monitor data The unsupervised model is trained and iterated continuously until the preset local training rounds M or error threshold and other termination conditions are completed, and finally the trained unsupervised local federated feature extraction model is uploaded to the central server.

[0058] In step S2, for each enterprise client participating in the training, take the i-th client as an example: Since the physical characteristics measured by each sensor monitoring data are different, the selected monitoring variables need to be normalized, and the minimum and maximum normalization method is selected to unify the local sensor monitoring data into the range of [-1,1] to eliminate the influence of different dimensions on the prediction results. The formula is as follows:

[0059]

[0060] In formula (1), represents the normalized value, x j,k represents the raw data of the jth data point of the kth sensor, and are the minimum and maximum values ​​of the kth sensor, respectively.

[0061] Then, in order to make full use of the local original multivariate time series data, the sliding window method is used to process the normalized data. The sliding window is a two-dimensional matrix [x window ,x feature ], where x window Indicates the size of the sliding window, x feature represents the number of selected input features, and the sliding window step size is set to 1 to maximize data usage.

[0062] The processed data is passed to the unsupervised feature extraction model for iterative training. In the unsupervised model, CAE participates in WGAN-GP as a generator to complete the extraction of important information of local data and improve the accuracy of remaining useful life (RUL) prediction.

[0063] The encoding process in CAE realizes the projection of the original data to a lower dimension, and its formula is as follows:

[0064]

[0065] In formula (2), ε represents the feature after dimensionality reduction, is a nonlinear activation function, is the weight between the i-th neuron in network layer n and the j-th neuron in network layer n+1; in addition, represents the data x corresponding to the i-th matrix of network layer n, and is the bias term in the encoding layer.

[0066] Then the decoding is performed. The decoding process is to obtain the reconstructed data from the encoded projection ε

[0067]

[0068] In formula (3), T represents transposition, Represents the bias term of the decoding layer.

[0069] Finally, the reconstruction error loss of CAE is calculated, and the loss function L CAE as follows:

[0070]

[0071] In formula (4), ||.||1 represents the L1 norm, M is the total number of local samples, and x k Represents the original data, Represents the data reconstructed by CAE. Since CAE is also the generator of WGAN-GP, The training goal of CAE is to minimize the loss function. During this process, a set of feature representations ε can be obtained through the encoding process of CAE.

[0072] CAE, as the generator of WGAN-GP, participates in the unsupervised training process of WGAN-GP. WGAN-GP is an improved version based on GAN, which consists of two adversarial models: the generator G, which is used to learn the true distribution of data; the discriminator D, which is used to estimate the probability that the training sample belongs to the real data rather than the data generated by G. G and D compete with each other to enhance performance. The generator G is trained to learn the distribution p on the real data x. g , which represents the noise variable p from the input z The mapping from (z) to the data space is denoted by G(z; θ g ), where θ g Denotes the parameters of the function G. Discriminator D(x; θ d) outputs a single scalar, representing that x comes from real data rather than generated data p g The goal of the discriminator D is to correctly identify samples from real data and the generator. On the contrary, the goal of the generator G is to generate samples that can successfully deceive the discriminator, that is:

[0073]

[0074] In formula (5), Represents the loss function of the generator participating in WGAN-GP, express is the data distribution P generated from the generator G g The data points generated by sampling are Denotes the output of the discriminator for the sampled data points, and E denotes the expected value of the specified distribution sample. The generator and the discriminator are optimized alternately.

[0075] In the training process of WGAN-GP, CAE is used as a generator, and the generated data is not only used to calculate its own reconstruction error, but also to confuse the discriminator. Therefore, the total loss function of the generator G in the unsupervised model is:

[0076]

[0077] In formula (6), L G (x i ) represents the total loss function of the multiple feature extraction generator of the present invention, express is the data distribution P generated from the generator G g The generated data points are used in ; E represents the expected value of a sample from the specified distribution.

[0078] In addition, the discriminator error needs to be calculated:

[0079]

[0080] In formula (7), L D represents the discriminator error, L real (x) represents the loss calculation for the real sample, L penalty (x,x * ) represents the gradient penalty term, which is used to constrain the discriminator to satisfy 1-Lipschitz continuity, x~P r Indicates that x is from the real data distribution P r The sampled points; D(x) represents the output of the discriminator for the real data points; Indicates that at the real data point x~P r and generate data points The distribution obtained by uniform sampling on the straight line is used to calculate the gradient penalty; x* represents the distribution obtained from uniform sampling The sampling points in Denotes the discriminator output D(x * ) for input x * The gradient norm is the Euclidean norm of the gradient; λ is the weight hyperparameter of the gradient penalty, which controls the impact of the gradient penalty on the total loss. Based on this multiple feature extraction structure, it is possible to capture the deep intrinsic information of the original monitoring data and the changing trend of the remaining life of the equipment.

[0081] Then, the client model parameters are updated using the RMSprop optimizer After completing local training, the updated model parameters are sent to the server.

[0082] The operations for the remaining clients are the same as described above.

[0083] S3. After each enterprise client is trained, the unsupervised federated feature extraction model parameters are uploaded to the central server and aggregated to form a new federated learning model for the next iteration. The aggregation process uses the FedAvg algorithm to average the received model weights and form a new global model:

[0084]

[0085] In formula (8), Φ t+1 represents the central server model parameters of the t+1th federated learning iteration training, N represents the number of clients participating in the training, represents the client model parameters of the ith client and the tth iteration, t∈(1,P), and FedAvg represents the federated averaging algorithm.

[0086] Subsequently, the aggregated global federated feature extraction model is continuously sent to N participating enterprise clients, and degraded features are extracted from local sensor data in the next training cycle. When all unsupervised collaborative training is completed, the obtained unsupervised federated feature extraction model can obtain key information from sensor data. S3 is repeated until the preset training rounds or error thresholds are met, and important information of local raw data is fully extracted.

[0087] The central server-client architecture process of the semi-supervised federated learning feature extraction prediction model is as follows: Figure 2 shown.

[0088] S4, fine-tuning stage, for enterprise clients with equipment prediction needs, use the trained unsupervised feature extraction model to map local monitoring data into high-level abstract features, and then further extract the timing degradation trend in key information through two consecutive GRU network layers. Finally, the output of the GRU network is passed to two fully connected layers to achieve the final mapping to the remaining service life RUL. Repeat S4 until the preset training rounds or error thresholds are completed, thereby achieving more accurate prediction performance.

[0089] Experimental example

[0090] Furthermore, experimental tests were conducted, and the experimental configurations and experimental results of the present invention include the following:

[0091] A. Dataset and Model Settings

[0092] The present invention focuses on using the U.S. commercial modular aviation propulsion system simulation data set C-MAPSS to achieve equipment life prediction. The prediction performance on this benchmark data set can verify the effectiveness of the proposed model. The C-MAPSS data set is a benchmark data set for the equipment life prediction model, simulating the degradation process of a turbofan engine. It contains four sub-datasets FD001-FD004, which are obtained under different operating conditions. In each sub-dataset, there is a training subset and a test subset, covering multivariate time series data generated under 21 sensors and 3 operating condition settings (including flight altitude, Mach number, and throttle lever angle). Different operating conditions significantly affect engine performance, so sensor measurements may vary under different operating scenarios. In addition, these engines have different degrees of initial wear and manufacturing differences, which is normal. Each subset contains a different number of turbofan engines, and the operating cycle of each engine is also different, so the data values ​​of each subset total 26 columns, including the number of engines and the operating cycle column.

[0093] Since the input data comes from different sensors, two one-dimensional convolution kernels [16×1] with a stride of 1 are first used in the CAE encoder to extract degradation information from the monitoring data of each sensor. Then, two two-dimensional convolution kernels [12×12] with a stride of 2 are used to further obtain the relationship between adjacent elements in the space of the data sample. In order to maintain the same dimension between the original data and the generated data, two deconvolution layers are stacked in the CAE decoder. In addition, the discriminator adopts the same architecture as the CAE encoder, except that Tanh is used as the activation function in the last CNN layer.

[0094] B. Training Settings and Data Segmentation

[0095] The default settings in the experiment are as follows: In order to keep the dimension of the data generated by WGAN-GP unchanged, the sliding window size is set to 32, that is, x window =32, the number of input features is 24, that is, x feature =24. The training data set does not provide the real equipment remaining life label. This patent uses a piecewise linear degradation model to construct a training label for each operating cycle of the turbofan engine, and sets the maximum remaining life to 120 cycles. All experiments are performed using an Intel(R) Core(TM) i7-1260P CPU@2.10GHz, 16GB of memory, and Microsoft Windows 11 operating system. The software used includes TensorFlow 2.1.4 and Python 3.7.

[0096] C. Comparison of baseline and experimental results

[0097] Compare to baseline

[0098] The comparison benchmarks include the prediction results of five experiments that have performed well in the past four years using C-MAPSS as the benchmark dataset. These experiments cover a range of model types, including unsupervised models, ensemble models, and semi-supervised models.

[0099] Experimental results, such as Figure 3 As shown in the figure, as the remaining life of the equipment decreases, the prediction accuracy of the model continues to improve, and the predicted value is closer to the true value, ensuring the accuracy of equipment status monitoring.

[0100] The error comparison between the model of the present invention and the algorithm tested on the same training set is shown in Table 1:

[0101] Table 1 Comparison of prediction score function and RMSE with recent advanced methods in the FD001 subset

[0102]

[0103] By comparison, it can be seen that the prediction error of the semi-supervised federated learning feature extraction equipment life prediction method proposed in the present invention is at the lowest level, regardless of RMSE or scoring function. The prediction error RMSE is 8.7% lower than the previous optimal variational autoencoder VAE, and the scoring function is 10.1% lower than the optimal bidirectional GRU: BiGRU-AS among all previous algorithms. The comparison shows that the proposed method captures higher-level abstract features of the original data through multiple feature extraction models under the framework of federated learning privacy protection, which contains more useful degradation information of the original data, laying the foundation for the subsequent supervised model to predict the remaining life. At the same time, due to the existence of the federated learning mechanism, local data does not need to be uploaded to the central server to prevent privacy data leakage.

Claims

1. A semi-supervised federated learning feature extraction equipment life prediction method, characterized in that: The method comprises the following steps: S1. Generate a randomly initialized unsupervised global federated feature extraction model Φ on the central server t ,The server selects N participating enterprise clients from the client pool and sends the initialized unsupervised global federated feature extraction model to them; The unsupervised global federated feature extraction model integrates the convolutional autoencoder CAE into the Wasserstein generative adversarial network-gradient penalty, i.e., the multiple unsupervised feature learning architecture constructed as the generator in WGAN-GP; S2, N clients are trained simultaneously. On each client participating in the training, its local model is first initialized to the unsupervised global federated feature extraction model sent by the server. Then, the processed local sensor monitoring data is used to train the client's unsupervised local federated feature extraction model. The training is repeated until the preset local training rounds M or the error threshold and other termination conditions are met. Finally, the trained unsupervised local federated feature extraction model is uploaded to the central server. S3. After the training of N enterprise clients is completed, the parameters of the unsupervised local federated feature extraction model are uploaded to the central server and aggregated; then the next round of federated learning training begins, and the newly formed unsupervised global federated feature extraction model is sent to the participating N enterprise clients again for a new round of federated training iterations until the preset maximum federated learning training round P or the error threshold is reached, and the training is terminated to obtain the trained unsupervised federated feature extraction model; S4, in the fine-tuning stage, for enterprise clients with equipment prediction needs, the trained unsupervised federated feature extraction model is used to map local sensor monitoring data into high-level abstract features; then, the timing degradation trend in key information is further extracted through two consecutive GRU network layers; finally, the output of the GRU network is passed into two fully connected layers to achieve the final mapping to the remaining useful life RUL.

2. According to claim 1, a semi-supervised federated learning feature extraction equipment life prediction method is characterized in that: The WGAN-GP consists of two adversarial models: a CAE generator G and a discriminator D. The CAE generator G is used to learn the true distribution of data and generate samples that can successfully deceive the discriminator D. The formula is as follows: In formula (1), Represents the loss function of the generator participating in WGAN-GP, express is the data distribution P generated from the generator G g The data points generated by sampling are represents the output of the discriminator for the sampled data points, and E represents the expected value of the specified distribution sample; The discriminator D is used to estimate the probability that the training sample belongs to the real data rather than the data generated by the CAE generator G, and correctly identify the generated samples from the real data and the generator; The generator G and the discriminator D are optimized alternately with each other.

3. The method for predicting equipment life by semi-supervised federated learning feature extraction according to claim 1, characterized in that: The processed local sensor monitoring data refers to, first, normalizing the selected sensor monitoring data, selecting the minimum and maximum normalization method, and unifying the local sensor monitoring data into the range of [-1, 1] to eliminate the influence of different dimensions on the prediction results. The formula is as follows: In formula (2), represents the normalized value, x j,k represents the raw data of the jth data point of the kth sensor, and are the minimum and maximum values ​​of the kth sensor respectively; Then, a sliding window is used to process the normalized data. The sliding window is a two-dimensional matrix [x window ,x feature ], where x window Indicates the size of the sliding window, x feature represents the number of selected input features, and the sliding window step size is set to 1 to maximize data usage.

4. The method for predicting equipment life by semi-supervised federated learning feature extraction according to claim 2, characterized in that: The specific steps of training the client-side unsupervised local federated feature extraction model are as follows: S21, the processed data enters the CAE generator G in WGAN-GP as the original data, completes the data reconstruction, and generates samples that can successfully deceive the discriminator D; S22, the real data and the samples generated by the generator enter the discriminator D, and the discriminator D identifies the real data and the samples generated by the generator; S23. Update client model parameters using RMSprop optimizer 5. The method for predicting equipment life by semi-supervised federated learning feature extraction according to claim 4, characterized in that: The data reconstruction is to first encode in CAE to realize the projection of the original data to a lower dimension, and the formula is as follows: In formula (3), ε represents the feature after dimensionality reduction, is a nonlinear activation function, is the weight between the i-th neuron in layer n and the j-th neuron in layer n+1; in addition, represents the data x corresponding to the i-th matrix of network layer n, and is the bias term in the encoding layer; Then the decoding is performed. The decoding process is to obtain the reconstructed data from the encoded feature ε In formula (4), T represents transposition, Represents the bias term of the decoding layer; Finally, the reconstruction error loss of CAE is calculated, and the loss function L CAE as follows: In formula (5), ||.||1 represents the L1 norm, M is the total number of local samples, and x k Represents the original data, Represents the data reconstructed by CAE. Since CAE is also the generator of WGAN-GP, The training goal of CAE is to minimize the loss function. During this process, a set of feature representations ε can be obtained through the encoding process of CAE.

6. The method for predicting equipment life by semi-supervised federated learning feature extraction according to claim 5, characterized in that: During the training process of the client-side unsupervised local federated feature extraction model, the reconstruction error and discriminator error of the generator are calculated, and the total loss function of the CAE generator G is: In formula (6), L G (x i ) represents the total loss function of the multiple feature extraction generator of the present invention, express is the data distribution P generated from the generator G g The generated data points are used in ; E represents the expected value of the sample of the specified distribution; The calculation formula of the discriminator error is as follows: In formula (7), L D represents the discriminator error, L real (x) represents the loss calculation for the real sample, L penalty (x,x * ) represents the gradient penalty term, which is used to constrain the discriminator to satisfy 1-Lipschitz continuity, x~P r Indicates that x is from the real data distribution P r The sampled points; D(x) represents the output of the discriminator for the real data points; Indicates that at the real data point x~P r and generate data points The distribution obtained by uniform sampling on the straight line is used to calculate the gradient penalty; x * represents the distribution obtained from uniform sampling The sampling points in Denotes the discriminator output D(x * ) for input x * The gradient norm of is the Euclidean norm of the gradient; λ is the weight hyperparameter of the gradient penalty, which controls the influence of the gradient penalty on the total loss.

7. The method for predicting equipment life by semi-supervised federated learning feature extraction according to claim 1, characterized in that: The aggregation process uses the FedAvg algorithm to average the received local model weights and form a new global model, namely: In formula (8), Φ t+1 represents the central server model parameters of the t+1th federated learning iteration training, N represents the number of clients participating in the training, represents the client model parameters of the ith client and the tth iteration, t∈(1,P), and FedAvg represents the federated averaging algorithm.

Images