A method, device, medium and program for training a turbocharger fault diagnosis model

Abstract

The present application belongs to the technical field of diesel engine fault diagnosis, and particularly relates to a turbocharger fault diagnosis model training method, device, medium and program, which can solve the problems of poor model generalization caused by data distribution heterogeneity and domain offset of multiple models of diesel engine turbochargers, solve the problem of low precision rate caused by the scarcity of fault samples and data imbalance, solve the problem of data privacy sensitivity and the inability to centrally train data islands, realize accurate capture of key features of time series fault data and cross-device collaborative modeling, effectively improve the precision rate of fault diagnosis, enhance the generalization ability of the model, and at the same time improve the stability of multi-client model collaborative training.

Description

Technical Field

[0001] This invention belongs to the field of diesel engine fault diagnosis technology, and specifically relates to a training method, equipment, medium and program for a turbocharger fault diagnosis model. Background Technology

[0002] Diesel engine turbochargers are core components of high-power equipment such as locomotives and ships, directly affecting the power output and emission performance of diesel engines. When operating for a long time in high-temperature, high-load, and strong-vibration environments, they are prone to surging, efficiency reduction, bearing wear, and other faults, making efficient and reliable intelligent fault diagnosis technology urgently needed.

[0003] Existing fault diagnosis methods are mainly divided into two categories: model-driven and data-driven. Model-driven methods rely on physical modeling, which is complex, has poor real-time performance, and weak interpretability, making it difficult to adapt to multi-condition and strongly coupled dynamic systems. Data-driven methods are limited by four major bottlenecks: data silos, imbalanced samples, privacy sensitivity, and cross-device domain bias. Specifically, data from different manufacturers and models of turbochargers are scattered, making centralized training impossible; fault samples are scarce, and conventional models are prone to overfitting and have poor generalization ability; operational data involves commercial and user privacy, prohibiting the sharing of raw data; and the data distribution of different devices varies greatly, leading to a significant decrease in diagnostic accuracy when directly trained.

[0004] Federated learning can achieve multi-party collaborative training without sharing the original data, but existing industrial federated learning is mostly for data with the same distribution and has not solved the problem of cross-device domain offset; domain adversarial learning can achieve feature distribution alignment, but it is difficult to integrate with the federated framework to balance privacy protection and cross-domain generalization.

[0005] In summary, there is currently a lack of a diesel engine turbocharger fault diagnosis solution that can simultaneously satisfy privacy protection, cross-device domain adaptation, temporal feature extraction, and robustness to imbalanced samples, which restricts the engineering implementation of intelligent diagnosis in key equipment such as locomotives and ships. Summary of the Invention

[0006] To address the issues of heterogeneous data distribution and domain shift leading to poor model generalization in multi-model diesel engine turbochargers; to resolve the problems of scarce fault samples and imbalanced data resulting in low accuracy; and to address the issues of data privacy sensitivity and data silos preventing centralized training, this invention provides a turbocharger fault diagnosis model training method in its first aspect, comprising: Obtain local training datasets built by multiple clients respectively. Each local training dataset includes time-series training samples and fault data labels corresponding to the time-series training samples, and set domain labels for each time-series training sample to indicate the source of the client. A fault diagnosis model is constructed, which includes a local feature extraction network and a common feature layer connected in sequence. The output of the common feature layer is connected to the fault diagnosis layer and the gradient inversion layer, respectively. The output of the gradient inversion layer is connected to the domain discriminator. The fault diagnosis model is subjected to multiple rounds of cyclic federated training until the convergence condition is met, thus obtaining a global fault diagnosis model. Each round of federal training includes: Based on the time-series training samples of each client and their corresponding fault data labels and domain labels, the fault classification loss and domain discrimination loss are determined respectively. The joint training loss is determined by the weighted sum of the fault classification loss and the domain discrimination loss. The network parameters of the fault diagnosis model are updated based on the joint training loss. The domain discriminator, the local feature extraction network, and the common feature layer are trained adversarially based on the domain discrimination loss. The global model parameters are updated by aggregating the local training results of each client, and the updated global model parameters are then distributed to each client.

[0007] Furthermore, the construction of the local training dataset includes: Acquire multi-parameter timing monitoring data of the turbocharger; Data cleaning and standardization are performed on multi-parameter time-series monitoring data to obtain standardized feature sequences; The standardized feature sequence is sliced ​​by sliding according to the preset slice length to obtain multiple time-series training samples; Configure corresponding data labels for each time series training sample, and configure corresponding domain labels according to the client of the time series training sample.

[0008] Furthermore, the data cleaning includes: detecting missing values ​​and outliers in the multi-parameter time-series monitoring data; and deleting the corresponding record when the working condition data corresponding to the record containing the missing value meets a preset quantity condition. When the operating condition data does not meet the preset quantity condition and the missing values ​​are scattered, the missing values ​​are filled with the mean or median of the corresponding feature under the corresponding operating condition.

[0009] Furthermore, the standardization process is Z-Score standardization.

[0010] Furthermore, the overlap rate between adjacent time-series training samples is 50%, and the sliding step size is half of the preset slice length.

[0011] Furthermore, the multi-parameter time-series monitoring data includes any number of the following: compressor inlet pressure, compressor outlet pressure, compressor inlet temperature, compressor outlet temperature, turbine inlet pressure, turbine outlet pressure, turbine inlet temperature, turbine outlet temperature, turbocharger speed, and atmospheric pressure.

[0012] Furthermore, based on the local feature extraction network, features are extracted from the time-series training samples of each client to obtain the local features corresponding to each client, and the local features corresponding to each client are input into the common feature layer to output the common features. The common features are input into the fault diagnosis layer, and the classification loss is selected based on the output of the fault diagnosis layer and the corresponding fault data labels. The common features are input into the domain discriminator through a gradient inversion layer, and the domain adaptation loss is determined based on the output of the domain discriminator and the corresponding domain label.

[0013] Furthermore, the local feature extraction network comprises a bidirectional long short-term memory network, a Dropout layer, a multi-head self-attention layer, and a fully connected layer connected in sequence; Bidirectional long short-term memory networks are used to extract temporal correlation features of temporal training samples in the forward and reverse temporal directions; A multi-head self-attention layer is used to assign attention weights to different time steps in temporally related features and obtain local features based on the attention weights; Local features are input into the common feature layer to obtain common features.

[0014] Further, local features are obtained based on the attention weights, including: For each attention head, the multi-head self-attention layer performs a linear mapping of the query matrix, key matrix, and value matrix on the temporal correlation features, determines the attention weight based on the scaled dot product of the query matrix and the key matrix, and weights the value matrix according to the attention weight to obtain the output of the corresponding attention head. The outputs of each attention point are concatenated and linearly transformed to obtain the local features; The common features are input into the fault diagnosis layer to obtain the fault classification result, and the classification loss is calculated based on the fault classification result and the corresponding data label.

[0015] Furthermore, the local features represent potential associations between different time steps of the time-series associated features, and the common features represent general fault features that are independent of the turbocharger domain.

[0016] Further, adversarial training is performed on the domain discriminator, the local feature extraction network, and the common feature layer based on the domain discrimination loss, including: A gradient inversion layer is set between the common feature layer and the domain discriminator; The domain discrimination loss is backpropagated through the gradient inversion layer to the feature layer and the local feature extraction network. The gradient inversion layer is used to multiply the gradient by a negative constant during backpropagation. The domain discriminator is used to accurately identify the domain to which each client input feature belongs, while the local feature extraction network and the common feature layer are used to extract features that prevent the domain discriminator from accurately identifying the domain to which it belongs.

[0017] Furthermore, the local training datasets of each client have the same feature dimensions and feature encoding rules, and the sample spaces of different clients are independent of each other; The original multi-parameter time series monitoring data of each client is stored locally on the respective client, and the original multi-parameter time series monitoring data is not centrally collected during the multiple rounds of federated training.

[0018] Furthermore, the local feature extraction network parameters are aggregated and updated based on the local training results of each client, and the updated local feature extraction network parameters are then distributed to each client, including: The parameter increment for each client is determined based on the difference between the local feature extraction network parameters after the end of this round of training and the local feature extraction network parameters issued at the beginning of this round of federated training. Based on the total number of time-series training samples in the local training dataset of each client, a weighted average of the parameter increments is calculated. The parameters of the local feature extraction network are updated based on the sum of the increments of each parameter after weighted averaging.

[0019] In a second aspect, the present invention also provides an electronic device comprising: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the training method of the turbocharger fault diagnosis model described above.

[0020] In a third aspect, the present invention also provides a computer-readable storage medium including a computer program that, when executed on a computer, causes the computer to perform the above-described training method based on a turbocharger fault diagnosis model.

[0021] In a fourth aspect, the present invention also provides a computer program product comprising computer program code, which, when run on a computer, causes the computer to execute the above-described training method based on a turbocharger fault diagnosis model.

[0022] The beneficial effects of this invention are: Privacy protection: No raw data is shared throughout the process; only model parameters are transmitted, meeting industrial data security requirements. Significantly improved accuracy: The F1 score is moderately improved compared to the independently trained baseline model, approaching the upper limit of centralized training; Strong cross-domain generalization: Domain adversarial mechanisms effectively eliminate the differences in the distribution of multiple devices, and the global model can be adapted to different models of turbochargers; Imbalanced robustness: The loss function and attention mechanism significantly improve the identification rate of faulty samples and reduce false negatives / false negatives; Engineering Adaptation: The accuracy of pressure ratio recognition error and speed fluctuation capture has been significantly optimized, making it suitable for actual working conditions on test benches and in vehicles. Attached Figure Description

[0023] Figure 1 The diagram shown is a flowchart of a turbocharger fault diagnosis model training method provided by the present invention. Figure 2 The diagram shown is a structural schematic of the electronic device provided by the present invention. Detailed Implementation

[0024] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0025] Figure 1 The diagram shown is a flowchart of a turbocharger fault diagnosis model training method provided by the present invention. Please refer to it. Figure 1 As shown, the present invention provides a method for training a turbocharger fault diagnosis model, comprising: Obtain local training datasets built by multiple clients respectively. Each local training dataset includes time-series training samples and fault data labels corresponding to the time-series training samples, and set domain labels for each time-series training sample to indicate the source of the client. A fault diagnosis model is constructed, which includes a local feature extraction network and a common feature layer connected in sequence. The output of the common feature layer is connected to the fault diagnosis layer and the gradient inversion layer, respectively. The output of the gradient inversion layer is connected to the domain discriminator. The fault diagnosis model is subjected to multiple rounds of cyclic federated training until the convergence condition is met, thus obtaining a global fault diagnosis model. Each round of federal training includes: Based on the time-series training samples of each client and their corresponding fault data labels and domain labels, the fault classification loss and domain discrimination loss are determined respectively. The joint training loss is determined by the weighted sum of the fault classification loss and the domain discrimination loss. The network parameters of the fault diagnosis model are updated based on the joint training loss. The domain discriminator, the local feature extraction network, and the common feature layer are trained adversarially based on the domain discrimination loss. The global model parameters are aggregated and updated based on the local training results of each client, and the updated global model parameters are then distributed to each client.

[0026] Furthermore, the construction of the local training dataset includes: Acquire multi-parameter timing monitoring data of the turbocharger; Data cleaning and standardization are performed on multi-parameter time-series monitoring data to obtain standardized feature sequences; The standardized feature sequence is sliced ​​by sliding according to the preset slice length to obtain multiple time-series training samples; Configure corresponding data labels for each time series training sample, and configure corresponding domain labels according to the client of the time series training sample.

[0027] Furthermore, the data cleaning includes: detecting missing values ​​and outliers in the multi-parameter time-series monitoring data; and deleting the corresponding record when the working condition data corresponding to the record containing the missing value meets a preset quantity condition. When the operating condition data does not meet the preset quantity condition and the missing values ​​are scattered, the missing values ​​are filled with the mean or median of the corresponding feature under the corresponding operating condition.

[0028] Furthermore, the standardization process is Z-Score standardization.

[0029] The Z-Score normalization formula used in this embodiment is expressed as follows:

[0030] in, It is the raw data. The mean of the data. The standard deviation of the data Furthermore, the overlap rate between adjacent time-series training samples is 50%, and the sliding step size is half of the preset slice length.

[0031] Furthermore, the multi-parameter time-series monitoring data includes any number of the following: compressor inlet pressure, compressor outlet pressure, compressor inlet temperature, compressor outlet temperature, turbine inlet pressure, turbine outlet pressure, turbine inlet temperature, turbine outlet temperature, turbocharger speed, and atmospheric pressure.

[0032] In this embodiment, the specific implementation steps of the present invention are described in detail below, based on bench test data of 16V280 and 12V240 diesel engine turbochargers.

[0033] Step 1: Data Acquisition Target: 16V280 (maximum speed 39600r / min), 12V240 (maximum speed 27500r / min) turbochargers; Collected parameters: compressor / turbine inlet and outlet temperature and pressure, turbocharger speed, atmospheric pressure, etc. The sampling interval was 1 second. After screening, the following valid samples were obtained: 12,781 groups of 16V280 type and 11,604 groups of 12V240 type. Tags: 0 = healthy, 1 = surge fault, training / test / validation sets are divided into training / test / validation sets according to a 7:2:1 time sequence.

[0034] All turbocharger data is collected and summarized, and the turbocharger performance characteristic value corresponding to each 5-second data point is calculated. To ensure the model's accurate predictive ability in the future, this embodiment adopts a time-aligned data partitioning strategy, dividing the time-series data of each client into training, testing, and validation sets in a 7:2:1 ratio. At the same time, to address the common problem of imbalanced class data distribution in multi-client environments, a stratified sampling mechanism is used to ensure that the label distribution of each dataset is consistent with the original data.

[0035] Furthermore, based on the local feature extraction network, features are extracted from the time-series training samples of each client to obtain the local features corresponding to each client, and the local features corresponding to each client are input into the common feature layer to output the common features. The common features are input into the fault diagnosis layer, and the classification loss is selected based on the output of the fault diagnosis layer and the corresponding fault data labels. The common features are input into the domain discriminator through a gradient inversion layer, and the domain adaptation loss is determined based on the output of the domain discriminator and the corresponding domain label.

[0036] Furthermore, the local feature extraction network comprises a bidirectional long short-term memory network, a Dropout layer, a multi-head self-attention layer, and a fully connected layer connected in sequence; Short-term memory networks are used to extract temporal association features of temporal training samples in both the forward and reverse temporal directions; A multi-head self-attention layer is used to assign attention weights to different time steps in temporally related features and obtain local features based on the attention weights; Local features are input into the common feature layer to obtain common features.

[0037] Furthermore, the local features represent potential associations between different time steps of the time-series associated features, and the common features represent general fault features that are independent of the turbocharger domain.

[0038] In this embodiment, the parameter configurations of each layer in the local feature extraction network are shown in Table 1: Table 1. Model Hyperparameter Configuration

[0039] Further, local features are obtained based on the attention weights, including: For each attention head, the multi-head self-attention layer performs a linear mapping of the query matrix, key matrix, and value matrix on the temporal correlation features, determines the attention weight based on the scaled dot product of the query matrix and the key matrix, and weights the value matrix according to the attention weight to obtain the output of the corresponding attention head. The outputs of each attention point are concatenated and linearly transformed to obtain the local features; The common features are input into the fault diagnosis layer to obtain the fault classification result, and the classification loss is calculated based on the fault classification result and the corresponding data label.

[0040] Further, adversarial training is performed on the domain discriminator, the local feature extraction network, and the common feature layer based on the domain discrimination loss, including: A gradient inversion layer is set between the common feature layer and the domain discriminator; The domain discrimination loss is backpropagated through the gradient inversion layer to the feature layer and the local feature extraction network. The gradient inversion layer is used to multiply the gradient by a negative constant during backpropagation. The domain discriminator is used to accurately identify the domain to which each client input feature belongs, while the local feature extraction network and the common feature layer are used to extract features that prevent the domain discriminator from accurately identifying the domain to which it belongs.

[0041] Furthermore, the local training datasets of each client have the same feature dimensions and feature encoding rules, and the sample spaces of different clients are independent of each other; The original multi-parameter time series monitoring data of each client is stored locally on the respective client, and the original multi-parameter time series monitoring data is not centrally collected during the multiple rounds of federated training.

[0042] In this embodiment, in order to improve the model's ability to capture and classify fault features, a fault diagnosis framework based on a hybrid model is constructed, including: BiLSTM hidden layer 64-dimensional → Dropout (0.5) → multi-head self-attention → fully connected layer 64-dimensional. This framework leverages the memory units and gating structure of LSTM to learn trend and periodic features in sequences, locking in long-term dependencies across time. To avoid processing limitations in capturing complex associations, this embodiment introduces a multi-head attention mechanism. Multiple "heads" learn different attention distributions, assigning different weight combinations to different parts of the time series. Combined with the LSTM model, this further focuses on potential associations between different time steps, capturing global data association analysis and diversity from different subspaces. The formula is expressed as follows:

[0043] In the formula: σ is the activation function, , , This is the weight matrix. Indicates the hidden state in the previous time step. and current input splicing, , , These are the bias terms, used to suppress overfitting. This indicates the sequence number of the acquisition time, and the Dropout mechanism is introduced after the LSTM layer.

[0044] To highlight the contribution of key temporal features, the output features of the LSTM are weighted through a self-attention layer. The weight matrix is ​​obtained by first scaling and adding bias, and then passing through softmax. This weighted output is called the output of each head.

[0045]

[0046] In the formula: To prevent the gradient from vanishing due to excessive variance in the dot product result. For single-head feature dimensions, These respectively represent the query, construct, and value vectors. This represents the normalized activation function.

[0047] Next, the outputs of all attention heads are concatenated to form a long vector, including:

[0048]

[0049] In the formula: For the output of each attention head, For concatenation functions, For learnable weight matrix, For bias terms, This represents a vector concatenation function.

[0050] This embodiment uses the Focal loss function to solve the label imbalance problem. The specific formula is as follows:

[0051]

[0052] Domain adaptation loss uses binary cross-entropy, aiming to distinguish between different client origins, as shown in the following formula:

[0053] In the formula, yi is the domain label of the client, and N represents the number of clients.

[0054] The beneficial effects of the above steps are as follows: by weighting the total loss, the weight of the domain adaptation loss can be reduced to avoid excessive interference from domain adversarial behavior; by using the model to learn fault characteristics that are common to all clients, rather than those specific to a particular client, the problem of large differences in data distribution between different clients can also be alleviated.

[0055] Because the data from each client is affected by the performance differences, operating conditions, and application environments of different turbocharger models, the statistical feature sets of data in different domains are not entirely the same. To address this issue, this embodiment uses domain-adaptive adversarial learning to minimize the distribution difference between the source and target domains, ensuring that the features of different clients have approximately consistent distributions in the common space, thus improving the model's cross-domain generalization ability. In federated learning, data from different heterogeneous diesel engine turbochargers are treated as different domains. Through domain adversarial training, the server's common feature layer learns common fault features independent of the device domain, improving the global model's generalization ability across different turbochargers.

[0056] In this embodiment, the backpropagation of the gradient reversal layer is set as follows:

[0057] Domain identifiers primarily distinguish samples originating from the client, thus counteracting the feature extractor.

[0058]

[0059] In the formula: For domain tags, denoted as the predicted probability of the domain discriminator.

[0060] Furthermore, the local feature extraction network parameters are aggregated and updated based on the local training results of each client, and the updated local feature extraction network parameters are then distributed to each client, including: The parameter increment for each client is determined based on the difference between the local feature extraction network parameters after the end of this round of training and the local feature extraction network parameters issued at the beginning of this round of federated training. Based on the total number of time-series training samples in the local training dataset of each client, a weighted average of the parameter increments is calculated. The parameters of the local feature extraction network are updated based on the sum of the increments of each parameter after weighted averaging.

[0061] Furthermore, to avoid the problem of the weights of small sample clients being overwritten by large sample clients and the dilution of parameter updates, this invention uses parameter increments for aggregation, as shown in the following formula:

[0062] In the formula, These are the local model parameters of the k-th client after the t-th round of training. The parameter increment of client k and the aggregation of parameter increments can effectively solve the problem of weight aggregation deviation caused by the difference in data volume among various clients of diesel engine turbochargers.

[0063] To comprehensively evaluate the fault diagnosis performance of this method for diesel engine turbochargers, accuracy, recall, precision, F1 score, and AUC were selected as model evaluation metrics. Recall primarily assesses the model's ability to identify faulty samples, while the F1 score comprehensively considers both precision and recall and is suitable for evaluating models with imbalanced samples. AUC is the area enclosed by the ROC curve and the coordinate axes, with a value ranging from [0.5, 1]. A higher value indicates better fault diagnosis performance.

[0064] To verify the performance of the diesel engine turbocharger fault diagnosis model based on domain adversarial federated learning provided in this invention, three comparative models were set up to observe the differences in performance between different training paradigms: the Baseline model, the FL model, and the Uplimit model. The Baseline model involves each client independently training its BiLSTM-Attention fault diagnosis model without any data sharing or federated collaborative training. The client model with the best performance was selected as the Baseline, which also serves as the basic benchmark for comparison in the experiment. The Uplimit model, on the other hand, involves centralized training of the BiLSTM-Attention fault diagnosis model. All the original private data from all clients were aggregated and uploaded to the server, and a single global model was trained using the merged complete dataset. The performance metrics of each model, representing the theoretical performance ceiling, are compared in Table 2.

[0065] Table 2. Summary of Model Performance Indicators

[0066] As shown in Table 2, the proposed FL model significantly outperforms the Baseline model in all performance metrics. Specifically, compared to the Baseline model, the FL model achieves improvements of 10.34% in accuracy, 0.56% in recall, 51.58% in precision, and 30.6% in F1 score, while the Uplimit model achieves improvements of 11.26%, 10.69%, 53.07%, and 33.58%, respectively. It can be seen that the improvement trends of the FL and Uplimit models are consistent, both showing significant improvements in precision and F1 score. This indicates that both federated learning and centralized learning can effectively utilize collaborative information from multiple clients to alleviate the sample imbalance problem, improve the model's fault diagnosis performance, and effectively avoid safety accidents caused by missed or incorrect fault diagnosis during turbocharger operation. This has significant practical implications for fault diagnosis in engineering applications.

[0067] The difference is that the performance improvement rate of the FL model is slightly lower than that of the Uplimit model. However, the Uplimit model requires the centralized collection of raw data from all clients, which poses a serious risk of data privacy leakage. In contrast, the FL model achieves near-centralized training performance without sharing raw data, thus verifying the effectiveness and practicality of the training method provided by this invention.

[0068] To evaluate and verify the effectiveness and contribution of different modules in practical applications, four ablation experiments were designed using the aforementioned dataset, hyperparameters, and evaluation metrics to compare the performance of different models. These experiments included the complete model presented in this paper, the attention-free model, the domain-free adversarial model, and the basic model. Table 3 shows the performance results of the ablation experiments. Table 3. Comparison of Performance Indicators of Ablation Test Models

[0069] As shown in Table 3, when the attention mechanism was removed, the model's F1 score decreased by 1.71% and AUC decreased by 3.54%. This is because the attention mechanism can focus on the key time steps of the fault occurrence and highlight the contribution of fault features; while the model without the attention mechanism only averages the features of all time steps, losing the key temporal information of the fault, resulting in a decrease in diagnostic performance.

[0070] After removing the domain adversarial module, the model's F1 score decreased by 5.03%, and its AUC decreased by 5.57%. This indicates that the domain adversarial module can effectively eliminate the differences in data distribution across multiple clients, enabling the global model to adapt to the domain characteristics of different devices; however, removing the domain adversarial module cannot solve the domain offset problem, resulting in weak generalization ability of the model on cross-device test sets.

[0071] When both the attention mechanism and the domain adversarial module are removed, the F1 score decreases by 11.2% compared to the complete model, which verifies the complementarity of the two core modules in this paper: the attention mechanism improves the ability to extract local fault features, and the domain adversarial module improves the domain adaptability of global features. The combination of the two achieves the optimal solution for fault diagnosis performance.

[0072] Figure 2 The diagram shown is a structural schematic of the electronic device provided by the present invention. Please refer to... Figure 2 As shown, the electronic device may include: a processor 310, a communications interface 320, a memory 330, and a communication bus 340, wherein the processor 310, the communications interface 320, and the memory 330 communicate with each other via the communication bus 340. The processor 310 can call logical instructions in the memory 330 to execute the aforementioned training method for the turbocharger fault diagnosis model, the method including: Obtain local training datasets built by multiple clients respectively. Each local training dataset includes time-series training samples and fault data labels corresponding to the time-series training samples, and set domain labels for each time-series training sample to indicate the source of the client. A fault diagnosis model is constructed, which includes a local feature extraction network and a common feature layer connected in sequence. The output of the common feature layer is connected to the fault diagnosis layer and the gradient inversion layer, respectively. The output of the gradient inversion layer is connected to the domain discriminator. The fault diagnosis model is subjected to multiple rounds of cyclic federated training until the convergence condition is met, thus obtaining a global fault diagnosis model. Each round of federal training includes: Based on the time-series training samples of each client and their corresponding fault data labels and domain labels, the fault classification loss and domain discrimination loss are determined respectively. The joint training loss is determined by the weighted sum of the fault classification loss and the domain discrimination loss. The network parameters of the fault diagnosis model are updated according to the joint training loss. The domain discriminator, the local feature extraction network, and the common feature layer are subjected to adversarial training based on the domain discrimination loss. The global model parameters are aggregated and updated based on the local training results of each client, and the updated global model parameters are then distributed to each client.

[0073] Furthermore, the logical instructions in the aforementioned memory 330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0074] Although the invention has been described in considerable detail and particularly with regard to several of the described embodiments, it is not intended to limit itself to any of these details or embodiments or any particular embodiment, thereby effectively covering the intended scope of the invention. Furthermore, the invention has been described above with respect to embodiments foreseeable by the inventors in order to provide a useful description, and non-substantial modifications to the invention that have not yet been foreseen may still represent equivalent modifications.

Claims

1. A method for training a fault diagnosis model for a turbocharger, characterized in that, include: Obtain local training datasets built by multiple clients respectively. Each local training dataset includes time-series training samples and fault data labels corresponding to the time-series training samples, and set domain labels for each time-series training sample to indicate the source of the client. A fault diagnosis model is constructed, which includes a local feature extraction network and a common feature layer connected in sequence. The output of the common feature layer is connected to the fault diagnosis layer and the gradient inversion layer, respectively. The output of the gradient inversion layer is connected to the domain discriminator. The fault diagnosis model is subjected to multiple rounds of cyclic federated training until the convergence condition is met, thus obtaining a global fault diagnosis model. Each round of federal training includes: Based on the time-series training samples of each client and their corresponding fault data labels and domain labels, the fault classification loss and domain discrimination loss are determined respectively. The joint training loss is determined by the weighted sum of the fault classification loss and the domain discrimination loss. The network parameters of the fault diagnosis model are updated according to the joint training loss. The domain discriminator, the local feature extraction network, and the common feature layer are subjected to adversarial training based on the domain discrimination loss. The local feature extraction network parameters are aggregated and updated based on the local training results of each client, and the updated local feature extraction network parameters are then distributed to each client.

2. The training method for the turbocharger fault diagnosis model according to claim 1, characterized in that, The construction of the local training dataset includes: Acquire multi-parameter timing monitoring data of the turbocharger; Data cleaning and standardization are performed on multi-parameter time-series monitoring data to obtain standardized feature sequences; The standardized feature sequence is sliced ​​by sliding according to the preset slice length to obtain multiple time-series training samples.

3. The training method for the turbocharger fault diagnosis model according to claim 1, characterized in that, Based on the local feature extraction network, features are extracted from the time-series training samples of each client to obtain the local features corresponding to each client. The local features corresponding to each client are then input into the common feature layer to output the common features. The common features are input into the fault diagnosis layer, and the classification loss is selected based on the output of the fault diagnosis layer and the corresponding fault data labels. The common features are input into the domain discriminator through a gradient inversion layer, and the domain adaptation loss is determined based on the output of the domain discriminator and the corresponding domain label.

4. The training method for the turbocharger fault diagnosis model according to claim 3, characterized in that, The local feature extraction network comprises a bidirectional long short-term memory network, a Dropout layer, a multi-head self-attention layer, and a fully connected layer connected in sequence. A bidirectional long short-term memory network is used to extract the temporal correlation features of the temporal training samples in the forward and reverse temporal directions; A multi-head self-attention layer is used to assign attention weights to different time steps in the temporal correlation features and obtain local features based on the attention weights; Local features are input into the common feature layer to obtain common features; The local features represent the potential correlations between different time steps of the time-series correlation features, and the common features represent general fault features that are independent of the turbocharger domain.

5. The training method for the turbocharger fault diagnosis model according to claim 4, characterized in that, The adversarial training of the domain discriminator against the local feature extraction network and the common feature layer based on the domain discriminant loss includes: A gradient inversion layer is set between the common feature layer and the domain discriminator; The domain discrimination loss is backpropagated to the common feature layer and the local feature extraction network via the gradient inversion layer. The gradient inversion layer is used to multiply the gradient by a negative constant during backpropagation. The domain discriminator is used to accurately identify the domain to which each client input feature belongs, while the local feature extraction network and the common feature layer are used to extract features that prevent the domain discriminator from accurately identifying the domain to which it belongs.

6. The training method for the turbocharger fault diagnosis model according to claim 5, characterized in that, The step of aggregating and updating the local feature extraction network parameters based on the local training results of each client, and then distributing the updated local feature extraction network parameters to each client, includes: The parameter increment for each client is determined based on the difference between the local feature extraction network parameters after the end of this round of training and the local feature extraction network parameters issued at the beginning of this round of federated training. Based on the total number of time-series training samples in the local training dataset of each client, a weighted average of the parameter increments is calculated. The parameters of the local feature extraction network are updated based on the sum of the increments of each parameter after weighted averaging.

7. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the training method for the turbocharger fault diagnosis model as described in any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, The system includes a computer program that, when executed on a computer, causes the computer to perform the training method for the turbocharger fault diagnosis model according to any one of claims 1-6.

9. A computer program product, characterized in that, The computer program product includes computer program code, which, when run on a computer, causes the computer to execute the training method for the turbocharger fault diagnosis model according to any one of claims 1-6.

Images