Device cross-condition fault prediction method and device based on adversarial decoupling and meta-learning
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-15
- Publication Date
- 2026-08-11
Smart Images
Figure CN122548504A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial equipment health monitoring technology, and in particular to a method and apparatus for predicting equipment faults across operating conditions based on adversarial decoupling and meta-learning. Background Technology
[0002] With the deep integration of next-generation information technology and the Industrial Internet of Things (IIoT), traditional manufacturing is accelerating its digital and intelligent transformation. In key sectors such as energy, aerospace, and modern manufacturing, industrial equipment typically operates under high loads, strong coupling, and complex and variable environments, making it susceptible to multiple factors such as mechanical wear, fatigue degradation, and environmental disturbances. Equipment failure can lead to production line shutdowns and high maintenance costs, or even major safety accidents. Therefore, building a precise equipment health perception and early warning system has become a core task of Predictive Maintenance (PdM). As a core pillar of Industry 4.0, PdM aims to achieve a paradigm shift from "reactive maintenance" to "proactive defense" through condition monitoring and intelligent modeling, thereby significantly reducing the risk of unplanned downtime and improving system operating efficiency.
[0003] In real-world industrial environments, industrial equipment often faces complex and variable operating conditions such as load variations, speed changes, and environmental disturbances. Monitoring data under different operating conditions may exhibit significant differences in amplitude, frequency distribution, and temporal evolution characteristics. This difference is known as distribution drift or domain shift in machine learning. Traditional fault prediction models (i.e., predicting the probability of a fault occurring within a future period) are mostly based on the Independent and Identically Distributed (IID) assumption. Models trained under specific operating conditions often experience a sharp drop in prediction performance when faced with unknown operating conditions or newly deployed equipment. To address this issue, researchers initially introduced transfer learning techniques such as Domain Adaptation (DA), which utilizes target domain data to assist the model in completing inter-domain knowledge transfer. Overall, DA methods can alleviate the distribution difference problem between the training and testing domains to some extent, but they typically require obtaining target domain sample information during the training phase. Therefore, these methods are more suitable for scenarios where the target operating conditions can be obtained in advance or are partially known. However, in real-world industrial applications, models often need to be deployed to unknown operating conditions or even unknown equipment, and target domain data is usually difficult to obtain in advance. To overcome this limitation, Domain Generalization (DG) methods have gradually become a research focus. Unlike Data Algorithm (DA), which relies on target domain data, DG does not use target domain data during the training phase. Instead, it learns domain-invariant features that are well-adapted to unknown domains by mining common patterns among multiple known source domains. Therefore, it is more in line with the deployment conditions of "unknown target operating conditions" in real industrial scenarios. In recent years, DG has gradually become an important research direction for solving fault identification and prediction problems across operating conditions and equipment. From a methodological perspective, existing DG research can be roughly divided into approaches such as representation alignment, causal modeling, feature decoupling, and joint distribution modeling.
[0004] Despite some progress in cross-condition fault prediction research, several shortcomings remain. First, most methods focus on aligning feature distributions or learning domain-invariant representations under different operating conditions, neglecting to adequately model the inherent long- and short-term time dependencies and complex interactions between variables within multivariate time-series data. Second, existing methods often overlook adapting to the dynamic nonlinearity and non-stationarity of time-series features during equipment degradation when dealing with distribution shifts caused by changes in operating conditions, leading to unstable predictive performance under unknown conditions. Therefore, constructing a model that simultaneously considers multivariate time-series feature mining, fault discrimination capabilities, and cross-condition generalization ability remains a key challenge in current research. Summary of the Invention
[0005] To address the above issues, this invention proposes a method and apparatus for predicting equipment faults across operating conditions based on adversarial decoupling and meta-learning. It constructs a cross-operating-condition fault prediction model comprising a multi-scale temporal dynamic graph encoding module, a feature decoupling learning module, and a class-conditional discriminant adversarial module. The multi-scale temporal dynamic graph encoding module couples multi-scale temporal features and variables. The feature decoupling learning module explicitly decomposes the shared representation into fault-invariant and operating-condition-related representations. The class-conditional discriminant adversarial module integrates class prediction probabilities into the domain discriminator, accurately removing operating-condition interference, extracting stable fault features across operating conditions, achieving refined adversarial alignment of categories, avoiding the ambiguity of class boundaries caused by traditional global adversarial methods, and maintaining classification and discriminative capabilities while weakening domain differences. It effectively alleviates data distribution shifts caused by changes in operating conditions and improves the model's generalization ability to unknown operating conditions.
[0006] On the one hand, the equipment cross-operating condition fault prediction method based on adversarial decoupling and meta-learning has the following specific steps:
[0007] S1. Acquire multi-sensor monitoring data of various industrial equipment under different working conditions and preprocess it. Use the known working condition monitoring data in the preprocessed monitoring data as multivariate time series samples of the source domain to construct a training set.
[0008] S2, Construct and train a cross-condition fault prediction model to obtain a trained cross-condition fault prediction model; the cross-condition fault prediction model includes a multi-scale temporal dynamic graph encoding module, a feature decoupling learning module, and a class conditional discriminant adversarial module.
[0009] The multi-scale temporal dynamic graph encoding module includes a multi-scale temporal feature extraction module and a dynamic graph neural network module; the multi-scale temporal samples are processed by the multi-scale temporal feature extraction module to extract local temporal data at different time scales, resulting in multi-scale fusion features; the multi-scale fusion features are processed by the dynamic graph neural network module to extract variable coupling features, resulting in a shared representation;
[0010] The feature decoupling learning module decomposes the shared representation into a fault-invariant representation that characterizes fault discrimination information that represents the stability of the equipment across operating conditions and an operating condition-related representation that retains domain-specific information through two independent linear projection heads.
[0011] The class-condition discriminant adversarial module inputs the fault-invariant representation into the fault classifier and the gradient inversion layer respectively; it then concatenates and fuses the output of the gradient inversion layer with the class prediction probability output by the fault classifier to obtain the class-condition input vector; the class-condition input vector is then input into the domain discriminator to determine the source of the working condition to which the sample belongs.
[0012] S3 uses a trained cross-condition fault prediction model to predict cross-condition faults in industrial equipment and obtains the category prediction probability.
[0013] Preferably, the multi-scale temporal feature extraction module uses several convolutional kernels of different sizes to perform parallel multi-scale convolution on the input multi-time series samples along the time dimension to extract candidate temporal features; it uses a selective convolutional kernel attention mechanism to adaptively fuse candidate temporal features at different scales according to the content of the input multi-time series samples; and it directly superimposes the original input onto the fused multi-scale temporal features through identity mapping and performs residual connection to obtain multi-scale fused features.
[0014] Preferably, the dynamic graph neural network module transforms multi-scale fused features into a dynamic topology with temporal continuity through dynamic graph transformation (DGT); then, it aggregates the information of neighboring nodes at the current time and the information of the same node at the previous time on the dynamic topology through dynamic graph isomorphic network (DyGIN) to obtain a shared representation.
[0015] Preferably, the state labels of the multivariate time-series samples in the training set are set as follows:
[0016] For industrial equipment that has malfunctioned, the time elapsed between each moment and the actual moment of malfunction is calculated as the time elapsed data. When the time elapsed data is less than or equal to the preset equipment warning window length, the status label is set to 1; when the time elapsed data is greater than the preset equipment warning window length, the status label is set to 0.
[0017] For industrial equipment that has not experienced a malfunction, the duration of each moment from the end of the data observation is calculated as the duration data. When the duration data is less than or equal to the preset equipment warning window length, the data for that moment is discarded; when the duration data is greater than the preset equipment warning window length, the status label is set to 0.
[0018] Preferably, the total loss function of the cross-condition fault prediction model includes an orthogonal difference constraint term to reduce information overlap and redundant coupling between the fault-invariant representation and the condition-related representation, expressed as:
[0019] ;
[0020] in, Indicates the orthogonal constraint loss; This represents the transpose of the fault-invariant representation; Indicates operating conditions; This represents the square of the Frobenius norm.
[0021] Preferably, the training process of the cross-condition fault prediction model includes a multi-task loss optimization stage and an embedded meta-learning stage, and the corresponding total loss function is expressed as:
[0022] ;
[0023] in, Represents the total loss function; Indicates supervised classification loss; Indicates loss due to differences in operating conditions; Represents the class-conditional alignment loss; Indicates the orthogonal constraint loss; Indicates the meta-learning loss; , , , and This indicates the weight of each loss.
[0024] Preferably, the meta-learning loss is specifically implemented as follows:
[0025] Calculate the first The average of the fault-invariant representations of all samples in the source domain is obtained as the first... The prototype of each source domain; the global prototype is obtained by averaging the prototypes of all source domains; the first... The squared Euclidean distance between the prototype of the source domain and the global prototype is used to measure the first... Prototype difference degree of each source domain;
[0026] Based on classification loss and prototype difference, the first... The domain difficulty score of each source domain is expressed as:
[0027] ;
[0028] in, Indicates the first Domain difficulty score for each source domain; and For the weighting factor; Indicates the first Supervised classification loss for each source domain; Indicates the first Prototype difference degree of each source domain;
[0029] The source domain with the highest score is selected as the meta-test set, and the remaining source domains are selected as the meta-training set.
[0030] Inputting the meta-training set into the cross-condition fault prediction model yields the meta-training loss. After one inner loop update, the temporary parameters are obtained, represented as follows:
[0031] ;
[0032] in, This represents the temporary parameters after the inner loop is updated; The learning rate for the inner loop; Indicates the parameters of the cross-condition fault prediction model; The model parameters are: Meta-training loss at time;
[0033] The meta-test set is input into the cross-condition fault prediction model to obtain the meta-test loss. The meta-training loss and the meta-test loss are combined to obtain the meta-learning loss, which is expressed as:
[0034]
[0035] in, Indicates the meta-training loss; Indicates the meta-test loss; The weighting coefficients are used to balance the meta-training loss and the meta-test loss.
[0036] Preferably, step S2 further includes: testing and evaluating the trained cross-condition fault prediction model using a test set; the test set is constructed by using the unknown condition monitoring data in the preprocessed monitoring data as the target domain; the evaluation metrics used include the macro precision, macro recall and macro F1 score of the true label and the predicted label corresponding to each test sample.
[0037] Preferably, the preprocessing includes variable filtering, missing value imputation, and normalization.
[0038] On the other hand, the equipment cross-condition fault prediction device based on adversarial decoupling and meta-learning includes the following:
[0039] The sample construction module is used to acquire and preprocess multi-sensor monitoring data of various industrial equipment under different working conditions, and use the known working condition monitoring data in the preprocessed monitoring data as the source domain to construct the training set.
[0040] The model building and training module is used to build and train a cross-working condition fault prediction model to obtain a trained cross-working condition fault prediction model; the cross-working condition fault prediction model includes a multi-scale temporal dynamic graph encoding module, a feature decoupling learning module, and a class conditional discriminant adversarial module.
[0041] The multi-scale temporal dynamic graph encoding module includes a multi-scale temporal feature extraction module and a dynamic graph neural network module. The multivariate temporal samples in the training set are processed by the multi-scale temporal feature extraction module to extract local temporal data at different time scales, resulting in multi-scale fusion features. The multi-scale fusion features are then processed by the dynamic graph neural network module to extract variable coupling features, resulting in a shared representation.
[0042] The feature decoupling learning module decomposes the shared representation into a fault-invariant representation that characterizes fault discrimination information that represents the stability of the equipment across operating conditions and an operating condition-related representation that retains domain-specific information through two independent linear projection heads.
[0043] The class-condition discriminant adversarial module inputs the fault-invariant representation into the fault classifier and the gradient inversion layer respectively; it then concatenates and fuses the output of the gradient inversion layer with the class prediction probability output by the fault classifier to obtain the class-condition input vector; the class-condition input vector is then input into the domain discriminator to determine the source of the working condition to which the sample belongs.
[0044] The fault prediction module is used to predict faults in industrial equipment across operating conditions using a trained cross-operating-condition fault prediction model, and to obtain the predicted probability of each category.
[0045] Compared with the prior art, the present invention has the following beneficial effects:
[0046] (1) The present invention explicitly decomposes the shared representation into fault-invariant representation and operating condition-related representation through the feature decoupling learning module, and introduces orthogonal difference constraints. The model can accurately remove operating condition interference and extract fault features that are stable across operating conditions, thereby enhancing the reliability and robustness of the prediction results.
[0047] (2) This invention utilizes a class conditional discriminant adversarial module to integrate class prediction probability into the domain discriminator, thereby achieving refined adversarial alignment of classes, avoiding the blurring of class boundaries caused by traditional global adversarial methods, and maintaining classification and discrimination capabilities while weakening domain differences;
[0048] (3) The present invention adopts a meta-learning strategy based on domain difficulty awareness, dynamically selects the source domain that is most difficult to generalize as the meta-test set, and guides the model to learn a more generalizable parameter update direction through meta-training-meta-testing alternating optimization, which significantly improves the adaptability and prediction accuracy of unknown working conditions.
[0049] (4) The cross-condition fault prediction model constructed in this invention combines multi-scale spatiotemporal feature encoding, feature decoupling learning, class condition discrimination adversarial and domain difficulty-aware meta-learning, which can effectively alleviate the data distribution shift caused by changes in working conditions and improve the model's generalization ability to unknown working conditions. Attached Figure Description
[0050] The present invention will now be described in further detail with reference to the accompanying drawings;
[0051] Figure 1 This is a flowchart of the equipment cross-operating condition fault prediction method based on adversarial decoupling and meta-learning according to an embodiment of the present invention;
[0052] Figure 2 This is a model architecture diagram of the equipment cross-operating condition fault prediction method based on adversarial decoupling and meta-learning according to an embodiment of the present invention;
[0053] Figure 3 This is a schematic diagram of the feature decoupling learning module of the equipment cross-operating condition fault prediction method based on adversarial decoupling and meta-learning in an embodiment of the present invention.
[0054] Figure 4 This is a schematic diagram of the class condition discrimination adversarial module of the equipment cross-operating condition fault prediction method based on adversarial decoupling and meta-learning in an embodiment of the present invention;
[0055] Figure 5 This is a schematic diagram of the domain difficulty-aware meta-learning module in the device cross-operating condition fault prediction method based on adversarial decoupling and meta-learning according to an embodiment of the present invention.
[0056] Figure 6 This is a structural block diagram of a device for predicting equipment failures across operating conditions based on adversarial decoupling and meta-learning, according to an embodiment of the present invention. Detailed Implementation
[0057] The present invention will be further described below through specific embodiments.
[0058] like Figure 1 As shown, the specific steps of the equipment cross-operating condition fault prediction method based on adversarial decoupling and meta-learning are as follows:
[0059] S1: Acquire multi-sensor monitoring data of various industrial equipment under different operating conditions and preprocess the data. Use the known operating condition monitoring data in the preprocessed monitoring data as multivariate time-series samples of the source domain to construct a training set.
[0060] Acquire historical feature data from multiple sensors for various industrial equipment under different operating conditions. This historical feature data includes the operational status information of the industrial equipment.
[0061] Acquire historical feature data from multiple sensors of various industrial equipment under different operating conditions, specifically including:
[0062] Using a uniform fixed sampling frequency Operational monitoring data of all industrial equipment under various operating conditions are collected at equal intervals to obtain historical characteristic data of the industrial equipment. Various working conditions include One device.
[0063] Set a unified observation cutoff time. For in For equipment that has previously failed, the actual time of failure is used as the end time of the equipment's data sequence; for equipment that has already failed, the actual time of failure is used as the end time of the data sequence. Equipment that had not previously experienced a malfunction, This serves as the end point of the equipment's data sequence, allowing for the acquisition of historical characteristic data for each industrial device, including operating conditions. The following equipment It contains N variables and has a length of Multivariate time series data can be represented as:
[0064] ;
[0065] in, , , Indicates working conditions Lower device In the At the [time]th moment The observations from each sensor.
[0066] Historical feature data is preprocessed to obtain preprocessed data, specifically including:
[0067] To improve model training efficiency and reduce interference from redundant information and invalid variables, variable filtering is performed on the historical feature data, resulting in the final data containing the following number of variables: ( ),Right now Specifically, we first extract the common features of each device under different operating conditions to ensure the consistency of feature dimensions among different devices. Based on this, we further screen the candidate features for effectiveness, removing variables with a null value ratio exceeding 50%, a zero value ratio exceeding 80%, and variables whose values remain unchanged throughout the entire operation. This yields the data after variable filtering.
[0068] To address the issue of missing values in the features of the data after variable selection, a step-by-step imputation strategy is adopted to process the data from each device. First, forward imputation is performed on the time series data to fill in the current missing items using valid observations from historical time points. Then, for any remaining missing values after forward imputation, backward imputation is performed to fully utilize valid information from subsequent time points. If a small number of missing values still exist after both forward and backward imputation, the mean of the feature is used for imputation. This yields the imputed data.
[0069] Based on the filled data, global maximum-minimum normalization is performed on the characteristic data of each channel according to the operating conditions to obtain normalized data, specifically including:
[0070] Because different monitored variables have different physical meanings and dimensions, directly inputting the raw data into the model may cause variables with larger numerical ranges to dominate during training, thus affecting the model's learning performance on other variables. To eliminate the impact of dimensional differences and improve the stability and convergence speed of model training, the retained variables are normalized according to the operating conditions. After normalization, the values of each variable are mapped to a unified interval, thereby enhancing the comparability between different features.
[0071] Based on the filled data, each feature is normalized using a normalization model. The normalization formula is as follows: ;
[0072] in For working conditions Lower device exist Time characteristics The normalized value, For working conditions Lower device exist Time characteristics The original observations, For centralized working conditions of the data Internal characteristics The minimum value, For centralized working conditions of the data Internal characteristics The maximum value, i.e., the normalized value, is calculated as (eigenvalue - minimum eigenvalue) / (maximum eigenvalue - minimum eigenvalue). The normalized value will fall within the range of [0,1]. In this way, each eigenvalue is scaled to a similar range, making it easier for the model to process.
[0073] Based on the duration of each time point relative to the actual equipment failure time or the end time of data observation, status labels are set for each time point in the normalized data of industrial equipment. Specifically, these include:
[0074] The given device warning window length is For industrial equipment that has malfunctioned, the time elapsed between each moment and the actual malfunction is calculated based on the normalized data of the equipment. In the formula For working conditions Lower device Actual downtime (for the faulty device, = ), For working conditions Lower device The current moment when this operational status characteristic data is collected. For working conditions Lower device The time elapsed since the actual failure.
[0075] when When this happens, the status label is set to 1, that is... This indicates that the device will be available in the future. If a fault occurs within a certain time period, the data at that moment is considered faulty data.
[0076] when When this happens, the status label is set to 0, that is... This indicates that the device will be available in the future. If there is no failure within a certain time period, the data at that moment is considered healthy data. For working conditions Lower device The status label corresponding to the historical operating status characteristics data of industrial equipment at the current moment.
[0077] For industrial equipment that has not yet malfunctioned, the duration of the distance from the last moment in the data for each moment is calculated based on the normalized data of the industrial equipment. Among these, In the formula For working conditions Lower device Data length, For working conditions Lower device The current moment when this operational status characteristic data is collected. For working conditions Lower device The time elapsed since the last moment of the device data.
[0078] when For health devices, since it is difficult to determine the state of their data during this period, the data at that moment is discarded.
[0079] when When this happens, the status label is set to 0, that is... This indicates that the device will be available in the future. If there is no failure within a certain time period, the data at that moment is considered healthy data. For working conditions Lower device The status label corresponding to the historical operating status feature data of industrial equipment at the current moment.
[0080] For labeled normalized data, an equal-interval sliding window segmentation strategy is used to generate samples: the sequence moves at equal intervals within a fixed-length time window to extract multiple equal-length subsequences, specifically including:
[0081] For the data in the dataset with pre-defined status labels, construct samples for each device using a sliding window approach. Let the window length be... The sliding step size is The warning window length is Then, depending on the operating conditions... Extracted from the original sequence of device k Time samples can be represented as:
[0082] ;
[0083] in, .
[0084] Then, by sliding the window in steps The working condition can be obtained by translating the original sequence. Lower device The corresponding sample set is:
[0085] ;
[0086] The sample label is then defined as:
[0087] ;
[0088] Label 1 represents a fault sample, i.e., the equipment in the future prediction window. A malfunction occurred; label 0 indicates a normal sample, meaning the device will fail within the future prediction window. The system is still operating normally.
[0089] So, regarding the working conditions Lower device This allows for the further construction of a sample-label pair set:
[0090] ;
[0091] in, Indicates working conditions Lower device Length of the original multivariate time series data.
[0092] Therefore, the entire dataset can be represented as:
[0093] ;
[0094] ;
[0095] in, Indicates working conditions Total number of devices. This represents the total number of operating conditions.
[0096] The data after sliding window sampling is divided into training and testing sets according to the operating conditions, specifically including:
[0097] In cross-condition scenarios, different conditions are treated as different domains. Let's assume it includes all... The set of various working conditions is denoted as:
[0098] ;
[0099] Subsequently, the operating condition space is divided into two mutually exclusive subsets: the source domain operating condition set. With target domain condition set ,satisfy:
[0100] ;
[0101] ;
[0102] Among them, source domain operating condition set Includes Various operating conditions, target domain operating condition set Includes The following working conditions are met:
[0103] ;
[0104] Source domain operating condition set The training set consists of known operating conditions, and the target domain operating condition set is also included. If the unknown operating conditions constitute the test set, then:
[0105] ;
[0106] ;
[0107] During the training phase, only source domain data can be accessed. Target domain data It is not visible during training and is only used during the testing phase to test the model's ability to generalize to unknown operating conditions.
[0108] S2, construct and train a cross-condition fault prediction model to obtain a trained cross-condition fault prediction model.
[0109] See Figure 2 As shown, based on the training set, a cross-operating-condition equipment fault prediction model based on adversarial decoupling and meta-learning is trained to establish a nonlinear mapping relationship between input features and data labels, specifically including:
[0110] The equipment cross-operating-condition fault prediction model based on adversarial decoupling and meta-learning mainly consists of a multi-scale temporal dynamic graph encoding module, a feature decoupling learning module, a class-conditional discriminant adversarial module, and a domain difficulty-aware meta-learning module. First, in the feature extraction stage, multivariate temporal samples are input into the multi-scale temporal dynamic graph encoding module for deep spatiotemporal representation learning. This module maps the original data into a shared representation containing temporal evolution and variable coupling information, providing a feature foundation for subsequent cross-operating-condition generalization learning. Second, based on the shared representation, a decoupled feature projection structure is constructed, decomposing the features into two parts: fault-invariant representation and operating-condition-related representation. The fault-invariant representation is mainly used to characterize fault discrimination information that is stable across operating conditions; the operating-condition-related representation is used to retain domain-specific information caused by factors such as load changes, speed fluctuations, and environmental disturbances. Through this decoupled modeling, the model can selectively extract fault representations that are insensitive to operating conditions, avoiding global feature confusion. Then, to avoid the "over-alignment" in the global feature space caused by traditional domain adversarial methods leading to blurred category discrimination boundaries, a class-conditional discriminant adversarial mechanism is introduced. The input to the domain discriminator fuses the class prediction probabilities of samples with the fault-invariant representation, enabling the adversarial process to proceed under class conditions, thus maintaining class separability while achieving cross-domain alignment. Finally, to improve the model's adaptability to unseen domain shifts, a domain difficulty-aware meta-learning strategy is adopted. This mechanism dynamically evaluates the generalization difficulty based on the classification loss of each source domain and the domain prototype difference, prioritizing the most difficult source domains for generalization as the meta-test set. By alternately optimizing inner and outer layers in the difficult domains, the model's generalization performance to operational shifts is enhanced.
[0111] The model's input is a tensor ,in For batch size, For the number of channels, This represents the number of nodes (generally corresponding to the number of variables). This represents the number of time steps (i.e., the length of each subsequence).
[0112] The multi-scale temporal dynamic graph encoding module mainly consists of two parts: a multi-scale temporal feature extraction module and a dynamic graph neural network module. The dynamic graph neural network module further includes a dynamic graph transform (DGT) and a dynamic graph isomorphism network (DyGIN), which is used to extract joint spatiotemporal feature representations from the input samples. Let the encoder be denoted as... Then the shared representation can be written as:
[0113] ;
[0114] in, Represents sample vector Shared feature representation, This indicates encoder parameters.
[0115] If we further break it down into two stages: multi-scale temporal convolutional modeling and dynamic graph modeling, it can be written as:
[0116] ;
[0117] ;
[0118] in, This represents a multi-scale temporal convolutional encoder. It represents multi-scale temporal features; This represents a dynamic graph encoder, responsible for modeling dynamic dependencies between variables; and These are the parameters for the corresponding modules.
[0119] First, multi-scale temporal convolutional branches are used to extract temporal patterns under different receptive fields. Then, dynamic graph structures are used to model the coupling relationship between variables over time, and finally, a sample-level shared representation is obtained. This provides the basic input for subsequent feature decoupling learning and cross-domain generalization training.
[0120] The multi-scale temporal feature extraction module draws inspiration from Selective Kernel Network (SKNet) to adaptively extract feature representations at different time scales from multivariate temporal data. This module integrates parallel multi-scale convolution, selective convolutional kernel attention mechanisms, and residual connection strategies. It extracts candidate temporal features in parallel along the time dimension using multiple convolutional kernels of varying sizes, and adaptively fuses temporal representations at different scales based on the input sample content. Simultaneously, the residual connection strategy directly superimposes the original input onto the fused multi-scale features through an identity mapping, ensuring that the original temporal information is not lost during feature transformation. This enhances the model's ability to represent complex temporal dependencies and more effectively models short-term fluctuations and long-term evolutionary patterns in multivariate temporal data.
[0121] The Dynamic Graph Neural Network (DGT) module does not rely on a predefined static graph structure. Instead, it automatically learns the implicit dependencies between variables from the data and describes their dynamic changes as device states evolve. The module mainly consists of two parts: DGT and DyGIN. DGT aims to capture the shifts and propagations of correlation patterns between adjacent time slices, reconstructing the originally discrete graph sequence into a dynamic topology with temporal continuity, thereby characterizing the evolution of variable dependencies. DyGIN performs cross-time-sequence information aggregation on the dynamic topology, fusing node states from different time steps, explicitly modeling the collaborative degradation patterns between variables over time, and ultimately achieving the processing of multivariate time-series data.
[0122] refer to Figure 3 As shown, the feature decoupling learning module maps features into two orthogonal subspaces using two independent linear projection heads: fault-invariant representations. Operating condition related indications ,in, Used to carry fault degradation semantics that are stable across operating conditions, serving as the core input for the subsequent fault classifier and class-conditional domain adversarial module; Then, time-varying factors closely related to operating conditions are stripped away and retained to assist in modeling differences in operating conditions and training the base domain discriminator. Formally, the feature decoupling process can be represented as:
[0123] ;
[0124] ;
[0125] in, and These represent two independent projection heads.
[0126] Based on this decoupling structure, the output of the fault classifier is:
[0127] ;
[0128] The domain information branch can be written as:
[0129] ;
[0130] in, Indicates the fault classification header, This indicates the operating condition identification head.
[0131] To minimize information overlap and redundant coupling between the two types of sub-representations and ensure the thoroughness of the decoupling process, an orthogonal difference constraint is introduced. This constraint encourages the fault-invariant space and the condition-related space to be as orthogonal as possible in terms of geometric topology, thereby achieving the deconstruction of the feature dimensions. Let the feature matrices corresponding to the current batch of samples be as follows: and Then the orthogonal constraint can be expressed as:
[0132] ;
[0133] in, This represents the square of the Frobenius norm. By introducing this constraint, the model can effectively suppress the information aliasing between fault features and operating condition features, making the fault-invariant representation purer, thereby significantly improving the convergence stability and generalization performance of the subsequent class-conditional domain adversarial process.
[0134] See Figure 4 As shown, the core idea of the class-conditional discriminative adversarial module is to push the adversarial process down from the global feature space to the class-conditional distribution space. By introducing prediction probability feedback, the model can explicitly maintain the separability between classes while aligning cross-domain features.
[0135] In the input of the domain discriminator, not only is fault-invariant representation used... It also integrates the current class prediction probability of the sample. This allows the domain adversarial process to proceed under category conditions. Let the category probabilities output by the fault classifier be:
[0136] ;
[0137] In order to make the domain discriminator It has the ability to perceive category semantics and represent fault-invariant representations. With predicted probability Perform feature fusion to construct a class-conditional input vector:
[0138] ;
[0139] in, This represents a vector concatenation operation. In this way, the decision criteria of the domain discriminator will no longer be based solely on the original features, but rather on the "feature distribution under a specific fault category".
[0140] Therefore, the output of the domain discriminator is:
[0141] ;
[0142] During training, the feature extractor and the domain discriminator engage in a minimax game. This method introduces a gradient reversal layer (GRL) to implement this adversarial process: during forward propagation, the GRL keeps the signal unchanged; during backward propagation, the GRL automatically multiplies the gradient by a negative scalar. The corresponding class-conditional adversarial loss function Defined as:
[0143] ;
[0144] in, This is a real domain tag.
[0145] By minimizing The domain discriminator is optimized, while GRL (Global Representation Language) prompts the feature projection head to generate a representation that the discriminator cannot distinguish from the source of the operating condition. This mechanism effectively avoids the "over-alignment" problem caused by traditional global adversarial approaches, ensuring that the model can accurately lock robust features with class discriminative power while strengthening cross-operating condition consistency.
[0146] This embodiment also includes a domain difficulty-aware meta-learning strategy, see [link to relevant documentation]. Figure 5 As shown, the domain difficulty-aware meta-learning strategy dynamically evaluates the generalization difficulty of each source domain in each iteration and prioritizes the "most difficult to generalize" source domain as the meta-test set, thereby forcing the model to learn more challenging domain-invariant features. The estimation of domain difficulty considers two factors: first, the magnitude of the classification loss in the current domain. If a sample in a certain domain has a large classification loss under the current model, it indicates that the model's fitting and generalization ability for that domain is relatively weak; second, the offset between the domain prototype and the global prototype.
[0147] Let the first The prototype of each source domain in the fault-invariant representation space is:
[0148] ;
[0149] Therefore, the global prototype of all source domain samples is:
[0150] ;
[0151] Then the first can be defined The prototype difference degree of each domain is defined as:
[0152] ;
[0153] By combining classification loss and prototype difference, the first... The domain difficulty score for each domain is:
[0154] ;
[0155] in, and The weighting coefficients are used. The source domain with the highest score is selected as the meta-test set, and the remaining source domains are selected as the meta-training set.
[0156] Let the model parameters be The loss on the meta-training set is After one inner loop update, the temporary parameters are obtained:
[0157] ;
[0158] in, This is the learning rate for the inner loop. Then, the updated parameters... Apply to the meta-test set and calculate its meta-test loss. The final meta-learning loss can be written as:
[0159] ;
[0160] in, To balance the weighting coefficients of the two losses, by minimizing The model can gradually learn more cross-domain stable parameter update directions, thereby improving its generalization ability to unknown target conditions.
[0161] Based on the above modules, a unified joint optimization objective function was constructed, covering four dimensions: classification and discrimination, representation decoupling, adversarial alignment, and meta-learning generalization. Firstly, the fault classifier uses fault-invariant representations... Calculate the supervised classification loss:
[0162] ;
[0163] Secondly, operating condition related indications Supervised learning is performed using domain information headers to explicitly preserve operating condition differences; the loss is:
[0164] ;
[0165] Then, class-conditional discriminant adversarial loss is applied to the fault-invariant representation. Apply orthogonal decoupling constraints between the operating condition / fault branches. Furthermore, it incorporates a domain-difficulty-aware meta-learning loss. Therefore, the total loss function of the model can be expressed as:
[0166] ;
[0167] in, These represent the weighting coefficients of each loss term.
[0168] During training, the model employs the following optimization strategies: First, the backbone encoder and feature decoupling learning module are jointly trained on multiple source domain samples; then, adversarial constraints are imposed on the fault-invariant representation through a class-conditional domain discriminator, while the domain information header is used to enhance the ability of the condition-related representation to preserve domain information; in each training epoch, a meta-test set is dynamically selected based on the current domain difficulty score, and a primary meta-learning parameter update is performed; finally, direct testing is conducted on completely unseen external conditions (target domain) to evaluate the model's cross-condition generalization performance.
[0169] Based on the training set, a cross-condition fault prediction model for equipment is trained using adversarial decoupling and meta-learning. Specifically, end-to-end parameter learning is performed using the training set: in each training epoch, the generalization difficulty of each source domain is dynamically evaluated, and the source domain with the highest difficulty is selected as the meta-test set, while the remaining source domains are used as the meta-training set. Mini-batch gradient updates are performed on the meta-training set to obtain temporary parameters, and then the meta-test loss is calculated on the meta-test set. The model's total loss function includes: fault classification loss, supervised loss of condition-related representations, class-conditional discriminative adversarial loss, orthogonal constraint loss, and meta-learning loss (a weighted sum of meta-training loss and meta-test loss). Gradients are calculated through backpropagation and the model parameters are updated using an optimizer to establish a nonlinear mapping relationship between input features and state labels.
[0170] During training, a fixed validation set is not pre-set. Early stopping is determined based on the trend of the meta-test loss (training stops when the meta-test loss no longer decreases for several consecutive epochs), and the learning rate is dynamically adjusted. After training, the model with the best average performance across all source domains or the lowest meta-test loss is selected as the optimal model weights. After training is complete and meta-learning tuning and model selection are performed, the final model weights and optimal checkpoints are saved. Simultaneously, complete experimental metadata, including hyperparameter settings, random seeds, and various evaluation reports, is recorded to ensure reproducibility of the experiment during final evaluation on an independent test set (target domain) and to meet auditing and traceability requirements.
[0171] The trained cross-condition fault prediction model is tested and evaluated using a test set (target domain operating condition data) until it meets the requirements. Otherwise, the model hyperparameters or network structure are adjusted and the model is retrained until it meets the requirements. The evaluation metrics employ the macro-average method, suitable for imbalanced tasks. The model's suitability is determined by its macro-precision, macro-recall, and macro-F1 score, specifically including:
[0172] The normalized test set is input into the trained fault prediction model, and the model provides a corresponding prediction label for each sample in the test set.
[0173] Based on the real label corresponding to each test sample With predictive labels The macro-precision, macro-recall, and macro-F1 score are calculated. The macro-average method is suitable for class imbalanced classification tasks because it assigns equal weight to each class during calculation, thus reflecting the recognition performance of the minority class more fairly.
[0174] The macro precision, macro recall, and macro F1 score are combined to determine whether the model meets the requirements. If the model meets the requirements, a pre-trained fault prediction model based on a multi-scale time-series dynamic graph neural network algorithm is obtained; otherwise, the model is retrained.
[0175] The specific calculation rules for macro precision, macro recall, and macro F1 score are as follows:
[0176] (1) Construct a confusion matrix and calculate the basic counting terms.
[0177] For each category ( , Calculated as the total number of categories:
[0178] True positive ( ): Reality as a category And it was predicted to be a category The number of samples;
[0179] False positives ( ): True is non-category But it was predicted as a category The number of samples;
[0180] False negative ( ): Reality as a category But it was predicted to be non-class. The number of samples;
[0181] True negative ( ): True is non-category And predicted to be non-class The number of samples.
[0182] (2) Calculation formula for indicators (for each category).
[0183] For each category Calculate precision, recall, and F1 score.
[0184] The formula for calculating the accuracy is: .
[0185] The formula for calculating recall is: .
[0186] The formula for calculating the F1 score is: .
[0187] (3) Formula for calculating macro average index.
[0188] The formula for calculating macro precision is: .
[0189] The formula for calculating macro recall is: .
[0190] The formula for calculating the macro F1 value is: .
[0191] S3 uses a trained cross-condition fault prediction model to predict cross-condition faults in industrial equipment and obtains the category prediction probability.
[0192] like Figure 6 As shown, the present invention also discloses a device for predicting equipment faults across operating conditions based on adversarial decoupling and meta-learning, comprising:
[0193] The sample construction module 601 is used to acquire multi-sensor monitoring data of various industrial equipment under different working conditions and preprocess it, and use the known working condition monitoring data in the preprocessed monitoring data as the source domain to construct a training set.
[0194] The model building and training module 602 is used to build and train a cross-working condition fault prediction model to obtain a trained cross-working condition fault prediction model; the cross-working condition fault prediction model includes a multi-scale temporal dynamic graph encoding module, a feature decoupling learning module, and a class conditional discriminant adversarial module.
[0195] The multi-scale temporal dynamic graph encoding module includes a multi-scale temporal feature extraction module and a dynamic graph neural network module. The multivariate temporal samples in the training set are processed by the multi-scale temporal feature extraction module to extract local temporal data at different time scales, resulting in multi-scale fusion features. The multi-scale fusion features are then processed by the dynamic graph neural network module to extract variable coupling features, resulting in a shared representation.
[0196] The feature decoupling learning module decomposes the shared representation into a fault-invariant representation that characterizes fault discrimination information that represents the stability of the equipment across operating conditions and an operating condition-related representation that retains domain-specific information through two independent linear projection heads.
[0197] The class-condition discriminant adversarial module inputs the fault-invariant representation into the fault classifier and the gradient inversion layer respectively; it then concatenates and fuses the output of the gradient inversion layer with the class prediction probability output by the fault classifier to obtain the class-condition input vector; the class-condition input vector is then input into the domain discriminator to determine the source of the working condition to which the sample belongs.
[0198] The fault prediction module 603 is used to perform cross-operating condition fault prediction of industrial equipment using a trained cross-operating condition fault prediction model to obtain the category prediction probability.
[0199] The specific implementation of the equipment cross-condition fault prediction device based on adversarial decoupling and meta-learning is the same as that of the equipment cross-condition fault prediction method based on adversarial decoupling and meta-learning, and will not be described again in this embodiment.
[0200] The above are merely specific embodiments of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial modifications made to the present invention using this concept shall be considered as infringing upon the protection scope of the present invention.
Claims
1. A device cross-operation fault prediction method based on adversarial decoupling and meta-learning, characterized in that, Includes the following steps: S1. Acquire multi-sensor monitoring data of various industrial equipment under different working conditions and preprocess it. Use the known working condition monitoring data in the preprocessed monitoring data as multivariate time series samples of the source domain to construct a training set. S2, Construct and train a cross-condition fault prediction model to obtain a trained cross-condition fault prediction model; the cross-condition fault prediction model includes a multi-scale temporal dynamic graph encoding module, a feature decoupling learning module, and a class conditional discriminant adversarial module. The multi-scale temporal dynamic graph encoding module includes a multi-scale temporal feature extraction module and a dynamic graph neural network module; the multi-scale temporal samples are processed by the multi-scale temporal feature extraction module to extract local temporal data at different time scales, resulting in multi-scale fusion features; the multi-scale fusion features are processed by the dynamic graph neural network module to extract variable coupling features, resulting in a shared representation; The feature decoupling learning module decomposes the shared representation into a fault-invariant representation that characterizes fault discrimination information that represents the stability of the equipment across operating conditions and an operating condition-related representation that retains domain-specific information through two independent linear projection heads. The class-condition discriminant adversarial module inputs the fault-invariant representation into the fault classifier and the gradient inversion layer respectively; it then concatenates and fuses the output of the gradient inversion layer with the class prediction probability output by the fault classifier to obtain the class-condition input vector; the class-condition input vector is then input into the domain discriminator to determine the source of the working condition to which the sample belongs. S3 uses a trained cross-condition fault prediction model to predict cross-condition faults in industrial equipment and obtains the category prediction probability.
2. The device cross-operation fault prediction method based on adversarial decoupling and meta-learning according to claim 1, characterized in that, The multi-scale temporal feature extraction module uses several convolution kernels of different sizes to perform parallel multi-scale convolution on the input multi-dimensional temporal samples along the time dimension to extract candidate temporal features. By utilizing the selective convolutional kernel attention mechanism, candidate temporal features at different scales are adaptively fused based on the content of the input multi-scale temporal samples. The original input is directly superimposed onto the fused multi-scale temporal features through identity mapping and residual connection is performed to obtain multi-scale fused features.
3. The device cross-operation fault prediction method based on adversarial decoupling and meta-learning according to claim 1, characterized in that, The dynamic graph neural network module transforms multi-scale fused features into a dynamic topology with temporal continuity through dynamic graph transformation (DGT); then, it aggregates the information of neighboring nodes at the current time and the information of the same node at the previous time on the dynamic topology through dynamic graph isomorphic network (DyGIN) to obtain a shared representation.
4. The device cross-operation fault prediction method based on adversarial decoupling and meta-learning according to claim 1, characterized in that, The state labels for the multivariate time-series samples in the training set are set as follows: For industrial equipment that has malfunctioned, the time elapsed between each moment and the actual moment of malfunction is calculated as the time elapsed data. When the time elapsed data is less than or equal to the preset equipment warning window length, the status label is set to 1; when the time elapsed data is greater than the preset equipment warning window length, the status label is set to 0. For industrial equipment that has not experienced a malfunction, the duration of each moment from the end of the data observation is calculated as the duration data. When the duration data is less than or equal to the preset equipment warning window length, the data for that moment is discarded; when the duration data is greater than the preset equipment warning window length, the status label is set to 0.
5. The equipment cross-condition fault prediction method based on adversarial decoupling and meta-learning according to claim 1, characterized in that, The total loss function of the cross-condition fault prediction model includes an orthogonal difference constraint term to reduce information overlap and redundant coupling between fault-invariant representations and condition-related representations, expressed as: ； in, Indicates the orthogonal constraint loss; This represents the transpose of the fault-invariant representation; Indicates operating conditions; This represents the square of the Frobenius norm.
6. The equipment cross-condition fault prediction method based on adversarial decoupling and meta-learning according to claim 5, characterized in that, The training process of the cross-condition fault prediction model includes a multi-task loss optimization stage and an embedded meta-learning stage. The corresponding total loss function is expressed as follows: ； in, Represents the total loss function; Indicates supervised classification loss; Indicates loss due to differences in operating conditions; Represents the class-conditional alignment loss; Indicates the orthogonal constraint loss; Indicates the meta-learning loss; , , , and This indicates the weight of each loss.
7. The equipment cross-condition fault prediction method based on adversarial decoupling and meta-learning according to claim 6, characterized in that, The meta-learning loss is specifically implemented as follows: Calculate the first The average of the fault-invariant representations of all samples in the source domain is obtained as the first... The prototype of each source domain; the global prototype is obtained by averaging the prototypes of all source domains; the first... The squared Euclidean distance between the prototype of the source domain and the global prototype is used to measure the first... Prototype difference degree of each source domain; Based on classification loss and prototype difference, the first... The domain difficulty score of each source domain is expressed as: ； in, Indicates the first Domain difficulty score for each source domain; and For the weighting factor; Indicates the first Supervised classification loss for each source domain; Indicates the first Prototype difference degree of each source domain; The source domain with the highest score is selected as the meta-test set, and the remaining source domains are selected as the meta-training set. Inputting the meta-training set into the cross-condition fault prediction model yields the meta-training loss. After one inner loop update, the temporary parameters are obtained, represented as follows: ； in, This represents the temporary parameters after the inner loop is updated; The learning rate for the inner loop; Indicates the parameters of the cross-condition fault prediction model; The model parameters are: Meta-training loss at time; The meta-test set is input into the cross-condition fault prediction model to obtain the meta-test loss. The meta-training loss and the meta-test loss are combined to obtain the meta-learning loss, which is expressed as: in, Indicates the meta-training loss; Indicates the meta-test loss; The weighting coefficients are used to balance the meta-training loss and the meta-test loss.
8. The equipment cross-condition fault prediction method based on adversarial decoupling and meta-learning according to claim 1, characterized in that, S2 further includes: testing and evaluating the trained cross-condition fault prediction model using a test set; the test set is constructed by using the unknown condition monitoring data in the preprocessed monitoring data as the target domain; the evaluation metrics used include the macro precision, macro recall and macro F1 score of the true label and the predicted label corresponding to each test sample.
9. The equipment cross-condition fault prediction method based on adversarial decoupling and meta-learning according to claim 1, characterized in that, The preprocessing includes variable selection, missing value imputation, and normalization.
10. A device for predicting equipment faults across operating conditions based on adversarial decoupling and meta-learning, characterized in that, Including the following: The sample construction module is used to acquire and preprocess multi-sensor monitoring data of various industrial equipment under different working conditions, and use the known working condition monitoring data in the preprocessed monitoring data as the source domain to construct the training set. The model building and training module is used to build and train a cross-working condition fault prediction model to obtain a trained cross-working condition fault prediction model; the cross-working condition fault prediction model includes a multi-scale temporal dynamic graph encoding module, a feature decoupling learning module, and a class conditional discriminant adversarial module. The multi-scale temporal dynamic graph encoding module includes a multi-scale temporal feature extraction module and a dynamic graph neural network module. The multivariate temporal samples in the training set are processed by the multi-scale temporal feature extraction module to extract local temporal data at different time scales, resulting in multi-scale fusion features. The multi-scale fusion features are then processed by the dynamic graph neural network module to extract variable coupling features, resulting in a shared representation. The feature decoupling learning module decomposes the shared representation into a fault-invariant representation that characterizes fault discrimination information that represents the stability of the equipment across operating conditions and an operating condition-related representation that retains domain-specific information through two independent linear projection heads. The class-condition discriminant adversarial module inputs the fault-invariant representation into the fault classifier and the gradient inversion layer respectively; it then concatenates and fuses the output of the gradient inversion layer with the class prediction probability output by the fault classifier to obtain the class-condition input vector; the class-condition input vector is then input into the domain discriminator to determine the source of the working condition to which the sample belongs. The fault prediction module is used to predict faults in industrial equipment under different operating conditions using a trained cross-operating-condition fault prediction model, and to obtain the predicted probability of each category.