Fault diagnosis method for heterogeneous data based on transfer learning

Abstract

The invention belongs to the technical field of complex equipment fault diagnosis methods, and particularly relates to a fault diagnosis method for heterogeneous data based on transfer learning. Comprising the following steps: preprocessing collected heterogeneous fault data, including filtering and steady-state feature extraction; determining a classification target domain used for determining a fault type, taking the obtained steady-state heterogeneous data of different dimensions of the fault data source domain and the classification target domain as input, and establishing and training a heterogeneous data feature coding network, so that the heterogeneous data feature dimensions of the fault data source domain and the classification target domain are aligned; constructing a model by using a domain adversarial neural network, and training the model to complete a fault diagnosis task; and further optimizing the model by using the classification target domain heterogeneous data. The method is suitable for various typical fault diagnosis task scenes, and is used for improving the diagnosis capability and robustness of complex equipment under the condition of few target domain fault tags.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of complex equipment fault diagnosis method, and particularly relates to a fault diagnosis method based on transfer learning for heterogeneous data. BACKGROUND

[0002] In order to improve the reliability of equipment operation and the system guarantee ability, in recent years, researchers have proposed various fault diagnosis methods for complex equipment, which are roughly divided into three categories: rule-based method, model-based method and data-driven method. The first two methods rely on a large amount of expert experience and system knowledge, and have weak generalization ability when facing complex and variable operating conditions. However, due to the sensitivity of complex equipment and the complexity of the operating environment, it is difficult to obtain a large amount of complete fault labeled data in actual engineering, which makes it difficult for traditional data-driven methods to achieve accurate diagnosis on target equipment. With the development of sensor technology, artificial intelligence and big data technology have promoted the development of data-driven methods in the field of intelligent fault diagnosis. In particular, deep learning has shown significant advantages in fault diagnosis due to its powerful feature extraction capability. However, it still faces severe challenges in actual engineering deployment. On the one hand, deep learning models usually rely on a large amount of high-quality labeled data for training. However, in complex equipment and high-reliability industrial scenarios, due to the sensitivity of equipment operation, the rarity of faults and the complexity of data collection, it is difficult to obtain sufficient fault labeled data of target equipment, especially in cross-device applications, where fault samples of the target domain are extremely rare. On the other hand, most existing researches focus on single equipment or homogeneous data sources, lacking the generalization ability across devices and operating conditions. When the model needs to be transferred from one device (fault data source domain) to another device with different feature dimensions and data distribution (heterogeneity) (target domain), and the target domain has only a small amount of labeled samples or even no labeled samples, traditional methods often fail to adapt, resulting in a sharp decline in diagnosis performance. Therefore, how to achieve accurate fault diagnosis of chiller under the condition of cross-device and small sample has become a key problem to be solved at present. At the same time, data from different devices or operating conditions have significant differences in feature dimensions and distribution, i.e. data heterogeneity, which further exacerbates the difficulty of model transfer. In addition, the data from different devices or operating conditions have heterogeneous characteristics in feature dimensions and distribution, which further increases the difficulty of cross-domain knowledge transfer. SUMMARY

[0003] The purpose of the present application is to provide a fault diagnosis method based on transfer learning for heterogeneous data, which is suitable for multiple typical fault diagnosis task scenarios, and combines gradient reversal layer domain adversarial training mechanism and semi-supervised learning transfer. Through heterogeneous feature unification at the data level and optimization of fault data source domain model processing, the diagnosis ability and robustness of complex equipment are improved under the condition of a small amount of target domain fault labels.

[0004] To achieve the above objectives, the present invention adopts the following technical solution.

[0005] A fault diagnosis method for heterogeneous data based on transfer learning includes:

[0006] A. Preprocess the collected heterogeneous fault data, including filtering and steady-state feature extraction;

[0007] B. Determine the classification target domain for determining the fault type. Take the obtained fault data from the fault data source domain and the steady-state heterogeneous data of different dimensions of the classification target domain as input, establish and train the heterogeneous data feature encoding network, and align the feature dimensions of the heterogeneous data from the fault data source domain and the classification target domain.

[0008] C. Build a model using a domain adversarial neural network and train the model to complete the fault diagnosis task; further optimize the model using heterogeneous data from the classification target domain.

[0009] Further improvements or specific implementation schemes for the aforementioned fault diagnosis method for heterogeneous data based on transfer learning include a heterogeneous data feature encoding network comprising: a dual-channel sparse autoencoder module, a shared CNN feature extraction and label classification module, and a gradient inversion and domain discrimination module.

[0010] In the dual-channel sparse autoencoder module, a left-right symmetrical dual-channel sparse autoencoder structure is adopted to encode and reconstruct the input from the fault data source domain and the input from the classification target domain, respectively. During the encoding process, sparsity constraints are applied to the hidden layer, introducing a relationship between the average activation of each hidden unit and the target sparsity rate. Divergence penalty term, solving for the total loss of the dual-channel sparse autoencoder. ;

[0011] In the shared CNN feature extraction and label classification module, the latent features output by the dual-channel sparse autoencoder are input into the shared CNN feature extractor, using cross-entropy loss. Minimum value is used as the optimization objective, and a label classifier is used to predict the fault category of the fault data source domain.

[0012] In the gradient inversion and domain discrimination module, the domain discriminator determines the source of features, and the gradient inversion layer enables the feature extractor to generate domain-invariant features. Adversarial training between the domain discriminator and the feature extractor is then used to extract domain-independent general features to achieve domain alignment. The optimization objective of the domain discriminator is... ;

[0013] The joint optimization formula for the heterogeneous data feature encoding network is:

[0014] ;

[0015] F is a feature extractor used to extract high-level feature representations from input data (whether from the source domain or the target domain). The goal of F is to generate domain-invariant features that make it impossible for the domain discriminator D to distinguish which domain they come from.

[0016] C is the feature classifier, used to accurately classify features. It is trained in a supervised manner on source domain data using real labels.

[0017] D stands for Domain Discriminator, used to determine whether the input features come from the source domain or the target domain. The goal of D is to maximize its discriminative power, that is, to distinguish the source of the features as accurately as possible.

[0018] in To share the loss weights of the CNN feature extraction and label classification modules, The loss weights are used for gradient reversal and domain discrimination modules.

[0019] In a further improvement or specific implementation of the aforementioned fault diagnosis method based on transfer learning for heterogeneous data, the dual-channel sparse autoencoder module represents the fault data source domain and the classification target domain using independent encoders and decoders, respectively.

[0020] Fault Data Source Domain Encoder ;

[0021] Fault Data Source Domain Decoder ;

[0022] Classification target domain encoder ;

[0023] Classification target domain decoder ;

[0024] Indicates the fault data source domain input Encoding result, Indicates the faulty data source domain Output the decoding result. Represents the classification target domain Input the encoded result, Indicates the classification target domain Output the decoding result; E represents the encoder, D represents the decoder, and the subscript indicates the data field it belongs to; This is a shared characteristic of the faulty data source domain; Shared features for the classification target domain;

[0025] To ensure that the encoder gives the fault data domain and the classification target domain a unified latent dimension L, an average activation of each latent unit is introduced. With target sparsity Between Divergence penalty term, for a batch The nth sample, the nth The average activation of each hidden unit is ;

[0026] i represents the i-th sample in the batch; j represents the j-th neuron / unit in the latent space (hidden layer);

[0027] This refers to the activation value of the i-th sample on the j-th neuron in the latent space after it has been processed by the encoder;

[0028] The sparse regularization term is

[0029] The total loss of the dual-channel sparse autoencoder is obtained as follows: ;

[0030] This refers to the reconstruction loss; where Control the weights of sparse terms.

[0031] A further improvement or specific implementation of the aforementioned fault diagnosis method based on transfer learning for heterogeneous data involves further extracting deep local patterns and temporal features from the latent features output by the dual-channel sparse autoencoder using a shared CNN feature extractor; a label classifier receives the output of the shared CNN feature extractor and performs label prediction on the fault data source domain samples, using cross-entropy loss. Minimize as the optimization objective, where

[0032] This refers to the total number of training samples in the source domain; k refers to the k-th class. Indicates the true class of the i-th source domain sample;

[0033] This refers to the probability value in the predicted probability vector output by classifier C for the input features, specifically the probability value belonging to the k-th class. C refers to the label classifier. :yes Input, That is, the ultimate high-level shared features; It is the source domain input sample.

[0034] In a further improvement or specific implementation of the aforementioned fault diagnosis method for heterogeneous data based on transfer learning, the filtering and steady-state feature extraction refer to constructing a steady-state detector using a Savitzky-Golay filter and an inverse difference method to identify and extract representative steady-state operating segments.

[0035] Further improvements or specific implementation schemes for the aforementioned fault diagnosis method for heterogeneous data based on transfer learning also include preprocessing of the source data. Specifically, this means achieving signal smoothing through local polynomial fitting, fitting a second-order polynomial using the least squares method within a sliding window, and replacing the window center point with the fitted value. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating a fault diagnosis method for heterogeneous data based on transfer learning.

[0037] Figure 2 This is a structural diagram of a dual-channel sparse autoencoder;

[0038] Figure 3 This is a schematic diagram of a domain adversarial neural network;

[0039] Figure 4 This is a schematic diagram based on the filtering results after preprocessing.

[0040] Figure 5 This is a graph showing the classification and detection results of the filtered steady-state data;

[0041] Figure 6 It is the loss change curve of the fault data source domain pre-training and target expander warm-up stage;

[0042] Figure 7 It is the dynamic change curve of discriminator loss and adversarial loss during adversarial training;

[0043] Figure 8 This is the loss curve during semi-supervised training;

[0044] Figure 9 It shows the curves of the changes in training loss and validation loss during the fine-tuning phase. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0046] This embodiment uses a chiller unit as an example. During the operation of the chiller unit, different equipment models, operating conditions, and sensor configurations can lead to inconsistent dimensions of monitoring data and significant differences in feature distribution. Directly applying a model trained in the fault data source domain to the target domain often results in decreased diagnostic performance due to distribution shift. To address the problem of multi-type fault diagnosis under cross-equipment and cross-operating condition conditions, this application proposes a dual-channel sparse autoencoder and domain adversarial neural network transfer learning principle based on this method to achieve feature mapping and distribution alignment between the fault data source domain (already labeled chiller unit operating data) and the target domain (new equipment operating data with no or few labels).

[0047] The following describes the specific implementation of this invention in conjunction with specific principles and methods to confirm the accuracy and versatility of the cross-domain fault diagnosis solution of this application. In general, this embodiment uses the Ra-1043 chiller unit public dataset as the fault data source domain. First, the collected chiller unit data is preprocessed, including filtering and steady-state feature extraction. Then, steady-state data from different dimensions of the fault data source domain and the target domain are used as input to establish and train a data heterogeneous feature encoding network, aligning the data feature dimensions of the fault data source domain and the target domain. Next, a domain adversarial neural network is used to build a model, and the model is trained to complete the fault diagnosis task. Subsequently, a small amount of target domain data is used to fine-tune the model, enabling it to achieve good performance in the target domain fault prediction task.

[0048] The overall architecture of the method in this application is as follows: Figure 1 As shown, this application effectively utilizes abundant public datasets of chiller units to accelerate the training process of the target domain chiller unit fault prediction model. By leveraging rich fault data from the target domain, the requirement for target chiller unit fault prediction data during model training can be significantly reduced.

[0049] To address the issue of inconsistent feature dimensions among different chiller units, this embodiment designs a dual-channel sparse autoencoder to encode the input features of the fault data source domain and the target domain separately, mapping them to a unified latent dimension L. Sparse constraints limit the average activation rate of hidden units, avoiding the introduction of redundant features and improving sensitivity to fault modes. In this way, data collected by chiller units under different sensor configurations can be aligned in a unified latent space, facilitating subsequent domain adversarial training.

[0050] Specifically, in cross-equipment fault diagnosis of chiller units, there is a dual heterogeneity between the fault data source domain (high-dimensional sensors, typically with 60+ dimensions) and the target domain (low-dimensional sensors, typically with 20+ dimensions). The fault data source domain has high sensor redundancy (e.g., containing complex parameters such as oil temperature / oil pressure), while the target domain only contains a few basic parameters such as temperature, pressure, and power.

[0051] The above-mentioned differences in the physical meaning of characteristics lead to Traditional single-encoder forced alignment leads to information loss. Therefore, this application proposes a dual-channel sparse autoencoder, which achieves heterogeneous feature unification through independent encoding paths. A symmetrical dual-channel sparse autoencoder structure is employed, with separate encoding paths for each faulty data source domain input. Input from the target domain Encoding and reconstruction are performed, while sparse constraints are applied to the hidden layer to improve the discriminativeness and sparsity of the representation, which is beneficial for subsequent cross-domain alignment and transfer. Figure 2 This is a structural diagram of a dual-channel sparse autoencoder.

[0052] Based on a dual-channel sparse autoencoder, the fault data source domain and the target domain each employ their own encoder and decoder:

[0053]

[0054]

[0055] In specific implementation, this is achieved through selection / design. and Make (Unified Potential Dimensions) This achieves the mapping from different input spaces to a unified latent space. During the mapping process, the conventional reconstruction loss is the mean squared error (MSE): ;

[0056] Reconstruction loss measures the difference between the decoder output (reconstructed data x) of the autoencoder and the original input x. Mean squared error is used as the metric here. This represents the square of the L2 norm, which is the sum of the squares of the differences across all dimensions. Its purpose is to ensure that the latent features z learned by the encoder retain most of the information from the original input, preventing the network from learning a meaningless identity mapping.

[0057] To produce a sparse representation in the hidden layers, an average activation of each hidden unit is introduced. With target sparsity Between Divergence penalty term.

[0058] For a batch The nth sample, the nth The average activation of each hidden unit is: ;

[0059] The sparse regularization term is defined as:

[0060]

[0061] Total loss of dual-channel sparse autoencoder

[0062] in By controlling the weights of sparse terms, sparse constraints suppress the activation of redundant features and tend to learn sparse bases that are more discriminative and transferable.

[0063] The domain representation is learned independently through the left and right channels and mapped to the same... The potential space retains the characteristics of each domain while providing data for subsequent domain adversarial operations.

[0064] Define fault data source domain and target domain ;in , To achieve knowledge transfer between the faulty data source domain and the target domain, and to address the issue of identical label spaces but different data distributions, This application achieves the desired feature distribution through feature transformation and model training. This allows the classifier trained on the faulty data domain to maintain high generalization performance on the target domain.

[0065] To achieve the above objectives, this invention employs a domain adversarial neural network, which introduces a gradient inversion layer between the feature extractor and the domain discriminator. During forward propagation, the features remain unchanged, while during backward propagation, the gradient sign is inverted, thereby enabling adversarial training between the feature extractor and the domain discriminator.

[0066] Solution as follows Figure 3 As shown, a dual-channel sparse autoencoder is introduced at the feature extraction end of a classic domain adversarial neural network to complete heterogeneous feature mapping, and combined with... Domain adversarial optimization enables end-to-end training.

[0067] Among them, sharing Feature extractor and label classifier: The latent features of the output are shared Further extraction of deep local patterns and temporal features. Label classifier. take over The output performs label prediction on samples from the faulty data source region, using cross-entropy loss as the optimization objective.

[0068]

[0069] This refers to the total number of training samples in the source domain; in classification loss... In this process, it sums the losses of all source domain samples and then takes the average (1 / This yields the average classification loss for a batch. This makes the loss independent of batch size and more stable.

[0070] This refers to the total number of training samples in the source domain; k refers to the k-th class.

[0071] This represents the true class of the i-th source domain sample.

[0072] It is the probability value of classifier C belonging to the k-th class in the output predicted probability vector for the input features.

[0073] C: Label classifier.

[0074] k: The kth category.

[0075] Meaning: For input features, the probability value of classifier C in the output predicted probability vector that belongs to the k-th class.

[0076] :yes Input. That is, the ultimate advanced shared features.

[0077] It is the source domain input sample.

[0078] Gradient inversion layer ( Domain Discriminator: The discriminator is a binary classification network used to distinguish whether input features come from the faulty data source domain or the target domain. During adversarial training, the domain discriminator engages in an adversarial game with the feature extractors (faulty data source domain compressor and target domain expander). The domain discriminator attempts to accurately determine the source of features, while the feature extractor uses a gradient inversion layer (…). The model learns to generate domain-invariant feature representations, thus "deceiving" the domain discriminator. This adversarial process prompts the model to extract domain-independent general features, thereby achieving domain alignment. The optimization objective of the domain discriminator is:

[0079]

[0080] in These are the shared characteristics of the faulty data source domain and the target domain, respectively.

[0081] Joint optimization formula is

[0082] in , These represent the loss weights for each task.

[0083] F is the feature extractor, typically a deep neural network (such as CNN or ResNet), responsible for extracting high-level feature representations from the input data (whether from the source or target domain). The goal of F is to generate domain-invariant features that the domain discriminator D cannot distinguish which domain they originate from.

[0084] C is the feature classifier, which accurately classifies features and is trained in a supervised manner on source domain data using real labels.

[0085] D stands for Domain Discriminator. It is a binary classification network used to determine whether an input feature comes from the source domain or the target domain. The goal of D is to maximize its discriminative power, that is, to distinguish the source of the feature as accurately as possible.

[0086] By minimizing the classification loss and maximizing the domain discrimination error, the features extracted by the CNN are forced to be both useful for fault identification and difficult to distinguish their domains, thereby achieving alignment between the feature distribution of the fault data source domain and the target domain.

[0087] Data verification:

[0088] To confirm the effectiveness of the proposed solution, experimental studies were conducted on different chiller units under various operating conditions, and fault data from the data source domain and the target domain were collected. The former data collection was based on the ASHRAE-1043 project conducted by Comstock et al. (1999), while the latter was collected in a research project conducted on a certain type of screw chiller unit.

[0089] The fault data source domain adopted chiller unit operating data from the ASHRAE-1043 project. Seven typical chiller faults with different severities (SLs) were studied using a 316kW indoor centrifugal chiller unit based on tetrafluoroethane (R-134a). In the test facility, both the evaporator and condenser were installed as water-cooled flooded heat exchangers using a shell-and-tube design. The original system was tuned by installing additional water and steam loops to simulate different operating conditions. The system was first introduced into a fault-free state, and each fault was studied under laboratory conditions at a nearly constant ambient temperature. Then, the corresponding faults were artificially induced, with their severity gradually increasing throughout the test series. A predefined test sequence was defined using evaporator outlet water temperature, condenser inlet water temperature, and calculated evaporator cooling rate as control variables, resulting in 27 operating states. These operating states were sequential, with each fault satisfying a steady-state criterion before proceeding to the next state.

[0090] The following is a list of faults used in this embodiment:

[0091] Table 1 Fault Data Source Domain Fault List

[0092]

[0093] For the target domain dataset, the system consists of two parts: the experimental platform and the data acquisition system.

[0094] The main experimental equipment of the cooling station test bench is a screw chiller unit, employing a return water temperature control strategy. The chiller unit consists of a compressor, condenser, expansion valve, evaporator, piping accessories, and electrical control box, all mounted on a common base. The data acquisition system is connected to the cooling station test bench via sensors for temperature and pressure. The data acquisition system comprises both hardware and software. The hardware includes data acquisition equipment and sensors for temperature, humidity, pressure, flow rate, and electrical parameters.

[0095] In addition to load control, the system also incorporates fault simulation devices, such as bypass valves between the compressor's suction and discharge pipes, regulating valves on the refrigerant pipeline, cooling water flow regulating valves, and chilled water flow regulating valves. Parameter signals from all measuring points in the system communicate with a PC via a data acquisition card, enabling continuous data sampling, real-time display, and recording. The software component mainly includes data display, data processing, and data storage.

[0096] Due to the characteristics of the target chiller unit's operation, the data collected is under rated operating conditions: chilled water outlet temperature controlled at 7 degrees Celsius, cooling water return temperature at 20 degrees Celsius, and compressor capacity adjustment using automatic regulation. After the unit is powered on, the control system performs a self-check. Once the self-check is successful, turn the "Mode" switch to "Automatic," set the desired target temperature value on the touchscreen, and press the automatic start button. The compressor contactor engages according to the reduced-pressure start logic, the compressor running indicator light illuminates, and the unit adjusts its energy according to the control logic set in the program, using the chilled water temperature feedback as a reference. The unit's operating status and parameters can be viewed on the touchscreen.

[0097] The specific fault data collection process is as follows:

[0098] For a condenser water flow rate reduction fault, set the opening of the electric two-way valve for cooling water (100% is fully open) on the control screen. Observe the cooling water flow meter and adjust the flow rate to 90%, 80%, and 70% of 20 m³ / h. After each flow rate adjustment, run the unit for a period of time to stabilize, then observe and record the data changes.

[0099] If the evaporator water flow is reduced, adjust the opening of the refrigerant water electric two-way valve on the control screen (100% is fully open). Observe the refrigerant water flow meter and adjust the flow rate to 90%, 80%, and 70% of 12 m³ / h. After each flow adjustment, run the unit for a period of time to stabilize, then observe and record the data changes.

[0100] In case of overcharge of refrigerant, refrigerant should be charged at specific locations in stages: 3 kg (10%), 6 kg (20%), and 9 kg (30%) of R134A refrigerant. After each overcharge, the unit should be run for a period of time to stabilize, and the data changes should be observed and recorded.

[0101] Insufficient refrigerant charge: After completely evacuating the unit, charge the refrigerant in stages: 18 kg (60%), 22.5 kg (80%), and 27 kg (90%) of R134A refrigerant. After each overcharge, run the unit for a period of time to stabilize, then observe and record the data changes.

[0102] If there is a non-condensable gas malfunction, perform a non-condensable gas charging procedure, charging in stages with 3 kg (10% of the refrigerant charge) and 6 kg (20%) of nitrogen. After each charging, the unit should be run for a period of time to stabilize, and the data changes should be observed and recorded.

[0103] Table 2 below is a fault simulation table:

[0104] Table 2 Target Domain Fault List

[0105]

[0106] The variables collected by the sensor are shown in Table 3 below:

[0107] Table 3. Variables collected on the target domain experimental platform

[0108]

[0109] Since the operation of a chiller unit includes stages such as startup, steady state, and shutdown, the raw data exhibits strong non-stationarity. To improve the stability and robustness of the diagnostic model, a steady-state detector is constructed using a Savitzky-Golay (SG) filter and the inverse difference (BD) method to identify and extract representative steady-state operating segments.

[0110] As a preprocessing module, signal smoothing is achieved through local polynomial fitting. Mathematically, this involves fitting a second-order polynomial (polyorder=2) using the least squares method within a sliding window (15 points, corresponding to 1.5 minutes of data), replacing the window center point with the fitted value. The choice between window length and polynomial order requires a trade-off between noise suppression and feature preservation: increasing the window size enhances smoothing but weakens abrupt changes (such as temperature jumps caused by faults); lower-order polynomials (2nd-3rd order) preserve the signal profile while avoiding overfitting. Compared to traditional moving average filtering, SG filtering offers superior shape preservation—it retains peak and edge information (such as early fault fluctuations of ±0.5℃), avoiding feature loss due to over-smoothing.

[0111] The result of filtering is as followsFigure 4 As shown.

[0112] The BD method refers to a geometrically weighted steady-state detection algorithm based on statistical feature analysis using a sliding window. Its time window length is set to 180 seconds. ), sampling interval 10 seconds ( ), through exponentially decaying weights (weighting coefficients) To enhance sensitivity to recent data, the weighted standard deviations of the evaporator inlet (TEI), outlet (TEO), and condenser inlet (TCI) temperatures are calculated within a sliding window. If all three standard deviations are below a threshold of 0.1℃, the system is considered to have entered a steady state. The filtered steady-state data serves as input for subsequent feature learning and classification. The detection results are as follows: Figure 5 As shown.

[0113] Table 4 below shows the datasets of the fault data source domain and target domain chiller units obtained after processing, with each type of fault data divided into 3 categories.

[0114] Table 4 shows the normal and fault operation data used for verification.

[0115]

[0116] The specific model training process is as follows:

[0117] Phase 1: Pre-training of the dual-channel sparse autoencoder: The dual-channel sparse autoencoder is trained independently on both the faulty data source domain and the target domain data, optimizing the reconstruction loss and sparsity constraints to map the two domain data to a unified latent dimension. This stage does not involve label information; it only involves unsupervised feature learning.

[0118] Phase 2, Joint Training of Domain Adversarial Networks: Fix or fine-tune some parameters of the dual-channel sparse autoencoder and connect it with a shared CNN feature extractor to construct a complete domain adversarial network;

[0119] The model in this embodiment is trained using the Adam optimizer, whose update rule is based on the estimation of the first and second moments of the gradient, enabling adaptive learning rate adjustment in non-convex optimization problems.

[0120] Initial learning rate: η =

[0121] First-order moment estimation coefficient: β1 = 0.9;

[0122] Second-order moment estimation coefficient: β2 = 0.999;

[0123] Smoothing term: ϵ= ;

[0124] StepLR learning rate decay is employed, reducing the learning rate to 0.5 every 20 epochs. This strategy ensures a relatively fast convergence speed in the early stages of training, while gradually reducing the learning rate in the later stages to improve convergence accuracy and stability.

[0125] Hyperparameter settings:

[0126] Potential dimensions: 64;

[0127] Sparsity ρ: 0.05, sparsity penalty coefficient β = 0.1;

[0128] The classification loss weight λcls = 1.0;

[0129] The adversarial loss weight λadv = 0.1;

[0130] Training epochs: 100.

[0131] Other model implementation details:

[0132] All input features are normalized before training (Min-Maxscaling to the [0,1] interval).

[0133] To avoid overfitting, Dropout (dropout rate of 0.5) and BatchNormalization are introduced after the CNN convolutional layers;

[0134] After training, the classification accuracy, macro-average F1 score, and confusion matrix are evaluated using the target domain test set.

[0135] Results Analysis

[0136] In the results analysis, i.e., the performance evaluation of the method of this application, this embodiment selects multi-dimensional quantitative indicators to comprehensively reflect the model characteristics, including:

[0137] A. Basic indicator accuracy, which measures the overall classification accuracy.

[0138] B. The F1 score effectively mitigates evaluation bias caused by data imbalance by harmonizing precision and recall.

[0139] C. The Confusion Matrix presents the classification details of each category in a visual form, revealing the model's misclassification patterns in a specific category.

[0140] Combining these three elements allows for both a grasp of overall performance and the identification of local optimization directions.

[0141] The training loss experiment systematically analyzed the loss dynamics of each stage of faulty data source domain pre-training, adversarial training, semi-supervised learning, and fine-tuning, comprehensively verifying the effectiveness of the proposed staged domain adaptation strategy.

[0142] Figure 6 The loss variation curves for the faulty data source domain pre-training and target extender warm-up phases are derived from... Figure 6 As can be seen, the pre-training loss of the fault data source domain (blue curve) significantly decreased from the initial 3.7 to 1.1 (a reduction of 70.3%), exhibiting a typical exponential decay trend, indicating that the model can effectively learn the fault feature representation of the fault data source domain. Notably, the pre-training loss of the target expander (orange curve) remained consistently around 1.0 (fluctuating within ±0.15). This low-loss stability indicates a high compatibility between the target domain feature distribution and the fault data source domain, meaning the target expander only needs to perform shallow feature transformations to achieve domain adaptation. The comparison of the two curves demonstrates the effectiveness of the phased training strategy adopted in this embodiment: the fault data source domain pre-training completes deep feature extraction first, while the target expander achieves domain alignment initialization with minimal computational cost, laying a solid foundation for subsequent adversarial training.

[0143] Figure 7 To illustrate the dynamic relationship between discriminator loss and adversarial loss during adversarial training, by Figure 7 As can be seen, the discriminator loss (blue curve) decreased significantly from the initial value of 1.6 to 0.6 (a decrease of 62.5%), showing a rapid initial decrease followed by a slower decrease. This indicates that the discriminator rapidly improves its domain discrimination ability in the early stages of training and gradually stabilizes in the later stages. Notably, the adversarial loss (orange curve) remained consistently stable around 0.8 (fluctuating within ±0.05). This stability indicates that the generator consistently and stably generates domain-invariant feature representations during training. The comparison of the two curves reveals a typical adversarial training equilibrium: the continuous decrease in discriminator loss indicates its increasing discrimination ability, while the stability of the adversarial loss proves that the generator can effectively deceive the discriminator, allowing the two modules to reach a dynamic balance. This loss change pattern demonstrates the effectiveness of the gradient inversion layer (GRL) and adversarial training mechanism used in this embodiment, laying a solid foundation for subsequent domain adaptation tasks.

[0144] Figure 8 This represents the loss value during semi-supervised training. Figure 8As can be seen, the loss value continuously decreased from the initial 0.055 to 0.015 (a decrease of 72.7%), showing a clear convergence trend. Specifically, the loss value decreased rapidly in the first 10 training epochs (from 0.055 to 0.030), indicating that the model could quickly learn effective feature representations; the rate of decrease slowed down during epochs 10-30 (from 0.030 to 0.020), indicating that the model entered a stable optimization phase; finally, it stabilized in epochs 30-50 (maintaining around 0.015), verifying the effectiveness of the semi-supervised training strategy. This convergence characteristic indicates that: 1) the quality of pseudo-labels continuously improves during the training process; 2) the effective utilization of unlabeled data enhances the model's generalization ability; and 3) no significant overfitting occurred during the training process. The smooth decrease in the loss curve proves that the consistency regularization and confidence screening mechanisms used in this embodiment effectively stabilized the semi-supervised learning process.

[0145] Figure 9 It is the curve showing the change between training loss and validation loss during the fine-tuning phase, from Figure 9 As can be seen, the training loss (blue curve) rapidly decreased from its initial value of 5.8 to around 0.3 (a reduction of 94.8%), and essentially converged after 40 epochs, indicating that the model can effectively learn the features of the target domain. The validation loss (orange curve) simultaneously decreased from 4.2 to 0.6 (a reduction of 85.7%), maintaining a similar downward trend to the training loss, indicating that the model has good generalization ability. It is noteworthy that the two curves maintained a small gap throughout (maximum difference < 0.5), without significant divergence, proving that the fine-tuning process did not produce overfitting. This convergence characteristic indicates that the learning rate decay strategy (decreasing by a factor of 0.1 every 10 epochs) and weight regularization (L2=0.01) effectively stabilized the optimization process, enabling the model to achieve stable convergence within 50 epochs, providing a reliable guarantee for subsequent practical applications.

[0146] In summary, the exponential decay loss (70.3% reduction) of the faulty data source domain pre-training and the stable loss of the target expander (fluctuation ±0.15) confirm the synergy between feature extraction and domain adaptation; the equilibrium state of the discriminator loss decreasing by 62.5% and the adversarial loss remaining stable (fluctuation ±0.05) during adversarial training reveals the mechanism by which the gradient reversal layer achieves domain-invariant features; the 72.7% loss reduction in semi-supervised training and the synchronous convergence of training / validation losses during the fine-tuning stage (difference <0.5) further demonstrate the role of pseudo-label enhancement and regularization strategies in improving generalization ability.

[0147] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A fault diagnosis method for heterogeneous data based on transfer learning, characterized in that, include: A. Preprocess the collected heterogeneous fault data, including filtering and steady-state feature extraction; B. Determine the classification target domain for determining the fault type. Take the obtained fault data from the fault data source domain and the steady-state heterogeneous data of different dimensions of the classification target domain as input, establish and train the heterogeneous data feature encoding network, and align the feature dimensions of the heterogeneous data from the fault data source domain and the classification target domain. C. Build a model using a domain adversarial neural network and train the model to complete the fault diagnosis task; further optimize the model using heterogeneous data from the classification target domain.

2. The fault diagnosis method for heterogeneous data based on transfer learning according to claim 1, characterized in that, The heterogeneous data feature encoding network specifically includes: a dual-channel sparse autoencoder module, a shared CNN feature extraction and label classification module, and a gradient inversion and domain discrimination module; In the dual-channel sparse autoencoder module, a left-right symmetrical dual-channel sparse autoencoder structure is adopted to encode and reconstruct the input from the fault data source domain and the input from the classification target domain, respectively. During the encoding process, sparsity constraints are applied to the hidden layer, introducing a relationship between the average activation of each hidden unit and the target sparsity rate. Divergence penalty term, solving for the total loss of the dual-channel sparse autoencoder. ; In the shared CNN feature extraction and label classification module, the latent features output by the dual-channel sparse autoencoder are input into the shared CNN feature extractor, using cross-entropy loss. Minimum value is used as the optimization objective, and a label classifier is used to predict the fault category of the fault data source domain. In the gradient inversion and domain discrimination module, the domain discriminator determines the source of features, and the gradient inversion layer enables the feature extractor to generate domain-invariant features. Adversarial training between the domain discriminator and the feature extractor is then used to extract domain-independent general features to achieve domain alignment. The optimization objective of the domain discriminator is... ; The joint optimization formula for the heterogeneous data feature encoding network is: ； F is a feature extractor used to extract high-level feature representations from input data (whether from the source domain or the target domain). The goal of F is to generate domain-invariant features that make it impossible for the domain discriminator D to distinguish which domain they come from. C is the feature classifier, used to accurately classify features. It is trained in a supervised manner on source domain data using real labels. D stands for Domain Discriminator, which is used to determine whether the input features come from the source domain or the target domain. The goal of D is to maximize its discriminative power, that is, to distinguish the source of the features as accurately as possible. in To share the loss weights of the CNN feature extraction and label classification modules, The loss weights are used for gradient reversal and domain discrimination modules.

3. The fault diagnosis method for heterogeneous data based on transfer learning according to claim 2, characterized in that, In the dual-channel sparse autoencoder module, the fault data domain and the classification target domain are represented by independent encoders and decoders, respectively. Fault Data Source Domain Encoder ; Fault Data Source Domain Decoder ; Classification target domain encoder ; Classification target domain decoder ; Indicates the fault data source domain input Encoding result, Indicates the faulty data source domain Output the decoding result. Represents the classification target domain Input the encoded result, Indicates the classification target domain Output the decoding result; E represents the encoder, D represents the decoder, and the subscript indicates the data field it belongs to; This is a shared characteristic of the faulty data source domain; Shared features for the classification target domain; To ensure that the encoder gives the fault data domain and the classification target domain a unified latent dimension L, an average activation of each latent unit is introduced. With target sparsity Between Divergence penalty term, for a batch The nth sample, the nth The average activation of each hidden unit is ; i represents the i-th sample in the batch; j represents the j-th neuron / unit in the latent space (hidden layer); This refers to the activation value of the i-th sample on the j-th neuron in the latent space after it has been processed by the encoder; The sparse regularization term is 4. Obtain the total loss of the dual-channel sparse autoencoder: ; This refers to the reconstruction loss; where Control the weights of sparse terms.

5. The fault diagnosis method for heterogeneous data based on transfer learning according to claim 3, characterized in that, The latent features output by the dual-channel sparse autoencoder are further extracted by a shared CNN feature extractor to obtain deep local patterns and temporal features. A label classifier receives the output of the shared CNN feature extractor and performs label prediction on samples from the faulty data source region, using cross-entropy loss. Minimize as the optimization objective, where ; This refers to the total number of training samples in the source domain; k refers to the k-th class. Indicates the true class of the i-th source domain sample; This refers to the probability value in the predicted probability vector output by classifier C for the input features, specifically the probability value belonging to the k-th class; C refers to the label classifier. :yes Input, That is, the ultimate high-level shared features; It is the source domain input sample.

6. The fault diagnosis method for heterogeneous data based on transfer learning according to claim 1, characterized in that, The filtering and steady-state feature extraction refer to the construction of a steady-state detector using a Savitzky-Golay filter and an inverse differential method to identify and extract representative steady-state operating segments.

7. The fault diagnosis method for heterogeneous data based on transfer learning according to claim 5, characterized in that, It also includes preprocessing of the source data, specifically, smoothing the signal through local polynomial fitting, fitting a second-order polynomial using the least squares method within a sliding window, and replacing the window center point with the fitted value.

Images