An industrial fault detection method and system based on dynamic drift sensing and diffusion enhancement

Abstract

This invention relates to the field of fault detection technology, and in particular to an industrial fault detection method and system based on dynamic drift sensing and diffusion enhancement. The method includes constructing an unsupervised fault detection DDA-DE model, wherein dynamic drift sensing DDA is used to process industrial data streams to establish a statistical distribution baseline and drift threshold; diffusion-enhanced anomaly detection DE is used to determine initial model parameters and anomaly thresholds; the robustness of the unsupervised fault detection model to concept drift is enhanced based on a diffusion strategy; the proposed industrial-enhanced Mahalanobis distance real-time drift sensing algorithm can more efficiently detect concept drift phenomena, and the parallel design of the drift sensing algorithm and the fault detection classifier achieves collaborative detection of data drift and anomaly events.

Description

Technical Field

[0001] This invention relates to the field of fault detection technology, and in particular to an industrial fault detection method and system based on dynamic drift sensing and diffusion enhancement. Background Technology

[0002] Industrial data stream anomaly detection technology, through real-time analysis of equipment time-series data (such as vibration spectra and temperature fluctuations), has become a core support for predictive maintenance and process optimization in industrial control systems. However, the widespread multi-source data drift phenomenon in industrial settings (such as sensor calibration offsets and gradual changes in feature distribution caused by equipment aging) severely weakens the reliability of detection models. Figure 1 The presentation showcases the time series and statistical characteristics of the FIT101 sensor (used to display influent flow status) in the SWaT dataset. It reveals significant changes in the statistical characteristics between the training and test sets, severely impacting the model's monitoring effectiveness. When implicit drift occurs in the production line data stream, anomaly detection algorithms frequently generate false alarms, trapping maintenance personnel in a vicious cycle of frequent false alarms and ineffective repairs.

[0003] To address the data drift problem, relevant methods can be broadly categorized into two types: one applied in the data preprocessing stage, and the other implemented in the modeling stage. In the modeling stage, the D3R model is highly representative. To address the data drift problem in time series anomaly monitoring, they proposed an anomaly monitoring method based on sequence decomposition. This method utilizes a data-time hybrid concern mechanism to dynamically decompose long-period multivariate time series, thereby eliminating the influence of the trend term on the monitoring results. In the data preprocessing stage, a simple adaptive strategy is typical. This strategy uses statistical estimation methods to remove trend estimates from the time series data, then conducts anomaly monitoring, and uses the detrended sequence to update the model parameters.

[0004] However, the above-mentioned work still faces the following challenges in solving the drift problem in industrial scenarios:

[0005] First, decoupling drift from anomalies is difficult. Existing works often use decomposition to achieve this decoupling. However, the model needs to process multiple subsequences generated by the decomposition simultaneously, leading to a significant increase in computational burden. Ideally, the residuals after decomposition should be white noise, but in practice, they often contain unextracted patterns (such as unidentified short-period fluctuations). When seasonal and trend components are coupled at higher orders, simple decomposition loses interaction information. Furthermore, since drift and anomalies differ significantly in their root causes and manifestations, relying solely on statistical data characteristics to determine real-time drift and anomalies is far from sufficient.

[0006] In industrial scenarios, drift typically stems from gradual changes in the system or environment, while anomalies are often triggered by sudden equipment failures or external attacks. In terms of manifestation, drift appears as a shift in data distribution or a change in temporal patterns, while anomalies manifest as sudden changes in local data points, discontinuous fluctuations, or business logic conflicts. Because the two differ significantly in their root causes and manifestations, analyzing them within a single process makes it difficult for models to effectively distinguish between anomalies and drift. Furthermore, relying solely on statistical data characteristics to determine real-time drift and anomalies is far from sufficient.

[0007] Secondly, the model lacks adaptability. Existing anomaly detection frameworks lack continuous learning mechanisms, making it difficult to capture the dynamic evolution of data distribution in scenarios of concept drift. Specifically, when faced with seasonal fluctuations or long-term system evolution, static model parameters cannot be iteratively updated through domain knowledge fusion, leading to a gradual degradation in detection performance. This limitation stems from the lack of an adaptive parameter optimization mechanism synchronized with time-varying industrial processes.

[0008] Third, fixed thresholds lead to a high false alarm rate. Specifically, because the characteristics of data drift are influenced by multiple factors such as equipment operating status and periodic environmental fluctuations, they exhibit significant time-varying and nonlinear features. When system drift occurs, the abnormal indicators of the monitoring model will show a non-fault-related increase due to distribution shift. Using static thresholds in this situation not only fails to effectively distinguish between drift and real abnormal events but also significantly increases the probability of false alarms under normal operating conditions. This phenomenon highlights the necessity of designing strategies based on dynamic threshold adjustment or adaptive mechanisms to improve the system's robustness to complex operating conditions. Summary of the Invention

[0009] To address the shortcomings of existing technologies in dynamically sensing the coupling relationship between changes in data distribution and abnormal patterns, this invention provides an industrial fault detection method and system based on dynamic drift sensing and diffusion enhancement. Through real-time distance monitoring and an adaptive model incremental learning mechanism, the collaborative detection of data drift and abnormal events is achieved.

[0010] In a first aspect, the present invention provides an industrial fault detection method based on dynamic drift sensing and diffusion enhancement, which adopts the following technical solution:

[0011] An industrial fault detection method based on dynamic drift sensing and diffusion enhancement includes:

[0012] Acquire industrial flow data;

[0013] An unsupervised fault detection model, DDA-DE, is constructed. Dynamic drift sensing DDA is used to process industrial data streams to establish a statistical distribution baseline and drift threshold. Diffusion-enhanced anomaly detection DE is used to determine the initial model parameters and anomaly threshold. The robustness of the unsupervised fault detection model to concept drift is enhanced based on a diffusion strategy.

[0014] The constructed unsupervised fault detection DDA-DE model is optimized.

[0015] Fault detection is performed using the optimized model.

[0016] Furthermore, the construction of the unsupervised fault detection DDA-DE model involves establishing a dual-branch detection architecture based on the essential difference between concept drift (statistical feature changes) and abnormal events (feature correlation disruption), which preserves the feature correlation within the data. The drift detection module employs a sliding window-based covariate offset quantization model, introducing an industrial-enhanced Mahalanobis distance metric to dynamically track changes in feature space distribution in real time. The fault detection branch integrates a diffusion enhancement strategy, enhancing the robustness of distribution shift through a multi-scale feature propagation mechanism while suppressing noise interference through information bottleneck constraints. Based on the detection results of both branches, the model parameters are dynamically calibrated, and the fault detection threshold is optimized through an adaptive threshold adjustment mechanism.

[0017] Furthermore, the process of using Dynamic Drift Sensing DDA to process industrial data streams to establish a statistical distribution baseline and drift threshold includes extracting statistical descriptors from historical data using DDA, wherein the baseline distribution parameters include the mean vector of the feature vector. Covariance Matrix This is used to achieve real-time distance calculation, thereby continuously updating the drift threshold; and incremental calculation is used to progressively update the mean and covariance matrix. The incremental calculation process is as follows:

[0018] ,

[0019] ,

[0020] in, This is the mean at the current moment. This is the average value at the previous moment. For the new data point at the current moment, The total number of data points at the current moment. Let be the covariance matrix at the current time. Let covariance be the covariance matrix at the previous time step, and for the eigenvectors... The formula for calculating its distance from the historical distribution is:

[0021] ,

[0022] in , The Hadamard product is used to eliminate scale and unit differences between features of heterogeneous sensors by normalizing the covariance matrix and incorporating it into the Mahalanobis distance. The calculation formula is:

[0023] ,

[0024] For temperature coefficient, Let be the variance of the i-th sensor; finally, the threshold is determined during the training phase, where the decision boundary is dynamically adjusted through real-time updates, calculated based on the mean and standard deviation of the historical distance distribution, using the following formula:

[0025] ,

[0026] in, The drift threshold at the current moment, The average historical distance at the current moment. The standard deviation of the historical distance at the current moment. It is a constant.

[0027] Furthermore, the method of using diffusion-enhanced anomaly detection (DE) to determine initial model parameters and anomaly thresholds includes using statistical features obtained during the training phase to perform dynamic comparative analysis on newly arrived time series. First, stable statistical data and thresholds are calculated based on historical normal data to depict the stable pattern of the system in a healthy state. When a new time window arrives, the industrial-enhanced Mahalanobis distance between the window and the historical distribution is calculated using the same distance calculation process. By introducing weighted and robust adjustments tailored to the characteristics of the industrial system, it exhibits tolerance for fault fluctuations and minor noise. After calculating the distance value, it is compared with a set threshold. If the distance value exceeds the set threshold, i.e. This indicates that the current time window deviates significantly from the historical normal distribution, suggesting that the system behavior may experience statistical drift or abrupt state changes. This indicates that the characteristics of the current window are still within the historical distribution range.

[0028] Furthermore, the robustness enhancement of the unsupervised fault detection model for concept drift based on the diffusion strategy includes iteratively updating the model parameters of the DE through a sliding window input mechanism, deriving the fault threshold based on the calculation of error scores, and including injecting temporal noise through the diffusion process, extracting features from the noise sequence, and calculating the reconstruction error. Specifically, injecting temporal noise through the diffusion process includes, assuming the input of the diffusion module is... The original vector contained therein at a certain moment is represented as The vectors of M-step Gaussian noise pollution can be expressed as follows: And the pollution result obtained in step m is only consistent with Related data The result obtained after m steps of noise pollution is:

[0029] ,

[0030] in Standard Gaussian noise, . This is a noise scheduling parameter that decreases with time step m. ,in It is the noise rate, which increases with time step. The calculation process is as follows:

[0031] ,

[0032] in and These are the preset minimum and maximum noise rates, M is the total number of time steps, and finally, the output of the diffusion strategy is obtained:

[0033] ,

[0034] in, , and The variable values ​​after noise addition by the diffusion module are positionally encoded and represented as follows:

[0035] ,

[0036] in, Represents convolution. This represents the variable value after position encoding. Relative position encoding is represented as:

[0037] ,

[0038] Indicates the position of the moment within the time window. This represents the dimension of the vector embedding. express Even-numbered dimensions in express Odd dimensions in.

[0039] Furthermore, the feature extraction from the noisy sequence includes inputting the word-embedded sequence into a feature extraction structure to extract complex spatiotemporal relationships in the data. The feature extraction module consists of L layers of blocks, where the input to the l-th layer block is... It has a time length as well as Dimension, initial input The embedded noise sequence is obtained in the timing module. The calculation process is expressed as follows:

[0040] ,

[0041] ,

[0042] in, Presentation layer standardization This indicates a multi-head attention mechanism. This represents the temporal hiding representation of the l-th layer. This represents a feedforward network, also obtained in the spatial module. , is represented as:

[0043] ,

[0044] ,

[0045] in This represents the spatial hiding representation of the l-th layer. This indicates transpose; finally, the (l+1)th layer is obtained. The process is as follows:

[0046] ,

[0047] ,

[0048] in This indicates that the symbol is hidden. This indicates a splicing operation.

[0049] Furthermore, the calculation of reconstruction error includes, to achieve the unsupervised training objective, adding a reconstruction module consisting of two fully connected layers at the end of the model to project the output dimension back to the original input dimension. The reconstruction process is described as follows:

[0050] ,

[0051] in express The reconstruction results This represents a fully connected layer; then the model's output is calculated. With the original input The reconstruction error is used as the loss function value and backpropagated back to the neural network to facilitate model parameter updates. Simultaneously, using the reconstruction error as the loss function helps the model find complex correlations between time series and various sensors. The calculation process is as follows:

[0052] ,

[0053] in Represent the Frobenius norm; and generate a reconstruction error score for the sensor values ​​at each time stamp. ,in The inherent distributional differences between faulty and normal data are reflected in the reconstruction error score. Error scores below a certain threshold are considered acceptable. The instances were judged to be normal, exceeding [a certain threshold]. Then it is marked as an exception.

[0054] Furthermore, the model optimization of the constructed unsupervised fault detection DDA-DE model includes configuring the hyperparameters of DDA-DE during training to set the time step m for adding noise in the diffusion module to 2000, and preset minimum and maximum noise rates β. min and β max The values ​​are 0.2 and 1 respectively, the number of layers L in the feature extraction module is 4, and the embedding dimension is... The size of the sliding window is 128. The model uses a 64-fold loss function, employs reconstruction loss, and updates the model using a mini-batch stochastic gradient descent strategy. Each training epoch traverses the entire training set, and the optimizer uses ADAM. During training, the average loss between the training and validation sets is calculated after each epoch to monitor model convergence. To prevent overfitting, dropout is added between transformer layers, and training is terminated early if the validation set error fails to improve after several consecutive epochs (early stopping mechanism). If the validation set reconstruction error fails to decrease for five consecutive epochs, a learning rate decay mechanism is triggered, automatically reducing the learning rate by half to refine the search. The entire training process iterates continuously, allowing the model to gradually learn the normal patterns of the time series. When the network parameters converge and the validation set error is at its lowest, the model weights are saved.

[0055] Furthermore, the fault detection using the optimized model includes calculating the reconstruction error score of the sensor values ​​at each time point based on the model output. ,in The system determines whether a data point is faulty based on both the drift detection results and a threshold: if the reconstruction error score is lower than a preset threshold... If no drift is determined, the data at that time point is considered normal; if the reconstruction error score is greater than or equal to the threshold... If no drift is detected, it is considered a fault; if the reconstruction error score is less than the threshold... If a drift is detected, the data at that time point is considered normal, and the model is incrementally updated; if the reconstruction error score is greater than or equal to the threshold... If drift is detected, the threshold is adjusted. And reassess; the initial assessment method for determining whether a fault has occurred is as follows:

[0056] ;

[0057] If drift is detected, the abnormal threshold is adjusted and reassessed. The adjustment method is as follows:

[0058] ,

[0059] ;

[0060] Finally, if no fault is detected and no concept drift occurs, no record is made; if an anomaly is detected but no concept drift occurs, the fault is recorded and monitoring continues; if no anomaly is detected but concept drift occurs, the DE is incrementally trained to adapt to changes in data distribution; if both anomaly and concept drift are detected, the threshold is adjusted and re-evaluated.

[0061] Secondly, an industrial fault detection system based on dynamic drift sensing and diffusion enhancement includes:

[0062] The data acquisition module is configured to acquire industrial flow data;

[0063] The model building module is configured to build an unsupervised fault detection DDA-DE model, wherein dynamic drift sensing DDA is used to process industrial data streams to establish a statistical distribution baseline and drift threshold; diffusion-enhanced anomaly detection DE is used to determine initial model parameters and anomaly thresholds; and the unsupervised fault detection model is enhanced for concept drift robustness based on a diffusion strategy.

[0064] The model optimization module is configured to optimize the constructed unsupervised fault detection DDA-DE model.

[0065] The fault detection module is configured to perform fault detection using an optimized model.

[0066] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the industrial fault detection method based on dynamic drift sensing and diffusion enhancement.

[0067] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide the industrial fault detection method based on dynamic drift sensing and diffusion enhancement.

[0068] In summary, the present invention has the following beneficial technical effects:

[0069] (1) In solving the problem of data nonstationarity, the dynamic decomposition method D3R based on diffusion reconstruction simply decomposes the data into trend terms and seasonal terms, causing the model to operate simultaneously, resulting in performance degradation and ignoring the linkage between decomposition terms. The real-time drift sensing algorithm based on industrial enhancement Mahalanobis distance proposed in this invention can detect concept drift more efficiently. The parallel design of the drift sensing algorithm and the fault detection classifier realizes the collaborative detection of data drift and abnormal events.

[0070] (2) Most existing time series fault detection methods use fixed thresholds to judge faults, which leads to a high false alarm rate when data drifts due to static parameters or untimely adjustments. This invention uses a threshold adjustment strategy that combines drift perception algorithm and fault detection classifier and a method of real-time updating of model parameters to improve the model's adaptability while effectively reducing the false alarm rate.

[0071] (3) Regarding the model's detection performance, because this invention addresses the non-stationarity problem from three perspectives—decoupling anomalies and drift, adaptive model adjustment, and dynamic threshold adjustment—the model ultimately achieves better detection results. Figure 4 As can be seen, the detection model proposed in this invention is superior to other methods in terms of F1 score, PATE, and AUC, thus proving that this invention has a significant advantage in detection performance. Attached Figure Description

[0072] Figure 1 This is a flowchart of the time-series fault detection method according to Embodiment 1 of the present invention;

[0073] Figure 2 This is a framework diagram of the fault detection model;

[0074] Figure 3 Dividing the dataset into different parts;

[0075] Figure 4 The graph shows the drift and fault detection results for the SWaT dataset;

[0076] Figure 5 The graph shows the drift and fault detection performance under continuous faults in the SWaT dataset.

[0077] Figure 6 This is a three-layer timing diagram. Detailed Implementation

[0078] The present invention will be further described in detail below with reference to the accompanying drawings.

[0079] Example 1

[0080] Reference Figure 1 This embodiment proposes an industrial fault detection method based on dynamic drift sensing and diffusion enhancement, hereinafter referred to as DDA-DE. The specific processing flow includes the following steps:

[0081] S1 Data Acquisition

[0082] This step is the cornerstone of the entire process. Its core function is to systematically collect raw signals from industrial scenarios, providing a high-quality, interpretable data foundation for subsequent modeling. Specific content includes: target definition, industrial protocol access, data acquisition, and secure storage. Its core functions are: supporting real-time decision-making, ensuring data credibility, guaranteeing business continuity, and ultimately extracting key monitoring business data from the industrial control system.

[0083] S1.1 Determine the monitoring targets and parameters

[0084] Identify key monitoring indicators for production line equipment (such as motor vibration frequency, reactor temperature, and pressure threshold), determine the fault type (such as gradual deviation, sudden oscillation, and periodic interruption), and set the acquisition frequency.

[0085] S1.2 Configure Data Source Access

[0086] Connect to the industrial control system to capture signals from key industrial control equipment such as PLCs in real time, deploy edge gateways to collect sensor Modbus / TCP streams, connect to the SCADA system API to obtain process parameter logs, and synchronously configure industrial firewalls to isolate the network and set up communication certificate encryption.

[0087] S1.3 Data Acquisition

[0088] High-frequency sensor data is pushed to edge computing nodes in real time, and batch process data is exported at a fixed frequency through the ODBC interface. A local SD card caching and resume transmission mechanism is enabled for power outage conditions.

[0089] S1.4 Data Storage

[0090] The time-series data streams from each sensor are stored in a database for real-time data analysis by the monitoring system.

[0091] S2 Model Construction

[0092] To address the challenge of scarce and difficult-to-label anomalous data in industrial scenarios, this invention proposes an unsupervised fault detection framework. Its core method involves establishing a baseline distribution using only normal samples during the training phase, while fault detection capability is achieved during the testing phase by identifying distribution shifts in unseen data.

[0093] In constructing the unsupervised fault detection model DDA-DE, we first focus on the complexity and non-stationarity of industrial system operating data, proposing a dual-branch detection architecture to distinguish between "concept drift" and "abnormal events," two phenomena that appear similar but are fundamentally different. Concept drift manifests as a gradual shift in the statistical characteristics of data distribution, while abnormal events are characterized by the sudden disruption of the correlation structure between features. To preserve the inherent feature correlations within the data, the model abandons the traditional single-branch detection approach and instead constructs two parallel but complementary detection paths. This decouples drift and anomalies at the mechanism level, thereby avoiding mutual interference and improving the accuracy and interpretability of detection.

[0094] The core task of the drift detection module is to perceive subtle shifts in data distribution in real time under label-less conditions. To this end, the model introduces a covariate offset quantization framework based on a sliding window. This framework dynamically maintains a historical data window of adjustable length and continuously compares the distribution differences between the current window and the reference window in the feature space. While traditional Mahalanobis distance can measure multidimensional distribution deviations, scale and unit differences between features from heterogeneous sensors affect the results. Therefore, covariance matrix normalization is incorporated into the Mahalanobis distance (defined as a weighted projection of feature bias based on the inverse covariance matrix). This allows the drift detection branch to maintain sensitivity to slight distribution shifts even under complex operating conditions.

[0095] The fault detection branch focuses on identifying anomalous events that disrupt the intrinsic connections between features. Since previous models are highly sensitive to concept drift and suffer from significantly increased training overhead due to the need to find information bottlenecks, a diffusion strategy is incorporated into the fault detection module. Following this design principle, this invention deliberately avoids the variational autoencoder (VAE) architecture, which generates inherent information bottlenecks, and instead adopts a more advanced Transformer-based framework. The proposed diffusion-based Transformer architecture not only more effectively captures the temporal features in industrial data streams but also significantly reduces training costs by externalizing information bottlenecks.

[0096] The dual-branch detection results are not used in isolation, but rather serve as a dynamic calibration mechanism driving model parameters. The drift detection branch continuously outputs a distribution offset intensity index. When a significant drift is detected, it triggers a progressive update of the reference distribution, preventing the model from failing due to adherence to outdated benchmarks. Simultaneously, it adjusts the fault threshold to prevent false alarms. The two branches run in parallel and are jointly evaluated. The threshold and model parameters are smoothly adjusted as the system's operating state evolves, maintaining low false alarms during stable periods and promptly capturing real anomalies during turbulent periods, thus achieving continuous optimization of detection performance. At this point, the DDA-DE model completes the closed-loop construction from data input to state discrimination, enabling high-precision, adaptive, unsupervised fault detection in complex industrial environments without any labels.

[0097] like Figure 2 As shown, Phase 0 is the data preparation phase, which is the cornerstone of the entire fault detection process. First, the raw industrial time series data pulled from DCS, SCADA, or edge gateways undergoes multiple rounds of cleaning. An outlier removal algorithm based on a combination of physical constraints and statistical rules is used to mark and linearly interpolate significantly distorted points. Then, non-steady-state segments such as start-up, shutdown, maintenance, and load reduction are completely removed according to equipment operating condition labels, retaining only pure normal data under rated operating conditions for training. This prevents the model from mislearning transitional states as normal patterns. For the test set, the complete operating cycle is deliberately retained, including slight drift, gradual degradation, and sudden failures, making subsequent evaluations closer to real production and laying a high-quality data foundation for the unsupervised training and online validation of the DDA-DE model.

[0098] The framework is then divided into two phases: Phase 1: Training Phase; Phase 2: Anomaly and Drift Detection Phase. Phase 1 employs a collaborative training framework, where the Dynamic Drift Awareness (DDA) method and the Diffusion Enhancement (DE) anomaly detection method are trained in parallel and jointly optimized. Specifically, DDA processes the raw data stream to establish a statistical distribution baseline and drift threshold, while DE determines the initial model parameters and anomaly threshold. Phase 2 corresponds to the inference phase, where the integrated DDA-DE framework performs real-time anomaly monitoring. Simultaneously, the drift detection model performs distribution shift analysis in parallel, achieving a collaborative dynamic assessment of data quality. This phase includes dynamic processing mechanisms covering four scenarios:

[0099] (1) Maintain normal operation when no fault or drift is detected.

[0100] (2) When a fault is detected but there is no drift, mark the fault but do not update the model.

[0101] (3) When there is no fault but there is drift, use sliding window data to incrementally update the model parameters.

[0102] (4) When both drift and fault are detected at the same time, adjust the threshold first and then verify the fault; once confirmed, mark the fault and continue the subsequent process.

[0103] S2.1 Dynamic Drift Sensing Method

[0104] It is important to clarify that the fundamental difference between data drift detection and fault detection lies in their respective time scales and data change characteristics. Data drift detection focuses on the gradual changes in data distribution over time, typically manifested as an overall shift in statistical properties (such as mean, variance, and covariance), reflecting the systematic evolution of the data distribution. Fault detection, on the other hand, focuses on sudden, localized deviations in data at specific points in time, typically manifested as isolated points or short sequences inconsistent with the overall pattern or context, reflecting instantaneous anomalous behavior of the data. This fundamental difference in objective provides the theoretical basis and guidance for the structural design of the model in this invention, enabling the model to be optimized for the unique needs of data drift and fault detection, thereby achieving more accurate monitoring and analysis in dynamic industrial environments.

[0105] This invention proposes an online, adaptive, unsupervised drift sensing method (DDA) to process industrial flow data. It employs a fixed sliding window approach and designs an industrial enhanced Mahalanobis distance to measure the differences between data distributions.

[0106] It should be noted that Mahalanobis distance was chosen as the core metric for industrial sensor data analysis in this study primarily due to its significant advantages: First, industrial sensor data typically exhibits multidimensionality and high correlation between features. Mahalanobis distance, by incorporating the covariance matrix, effectively handles dimensional relationships, providing a more accurate measurement of the deviation between the current data point and its historical distribution compared to Euclidean distance—especially avoiding misjudgments caused by dimensional correlation. Second, its scale invariance is crucial for industrial applications—sensor measurements often involve heterogeneous units and scales; Mahalanobis distance eliminates dimensional scaling effects through a built-in normalization mechanism, ensuring robust performance when analyzing multi-source heterogeneous data.

[0107] The practical industrial enhancement scheme proposed in this invention is mainly reflected in two key aspects: First, the distance calculation strategy adopts a sliding window-based method, which can dynamically and continuously update the mean and covariance estimates in real time, significantly improving the system's ability to cope with common asymptotic drift phenomena in industrial environments. Second, addressing the inherent differences in sensor characteristics in actual industrial environments—high-frequency sensors (such as temperature and pressure sensors) are more prone to drift than low-frequency sensors (such as valves)—this research introduces an explicit parameter adjustment mechanism into the original Mahalanobis distance formula to prioritize the processing of drift-prone sensors, thereby improving the applicability and accuracy of the drift detection algorithm in industrial operating scenarios. The implementation steps of the Dynamic Drift Sensing (DDA) method are as follows:

[0108] S2.1.1 Phase 1: Training Statistics and Thresholds

[0109] During the training phase, such as Figure 2 As shown in the first phase, DDA extracts statistical descriptors from historical data. This process involves calculating baseline distribution parameters—that is, the mean vector of the feature vectors. Covariance Matrix This enables real-time distance calculation, thereby continuously updating the drift threshold. Given the massive data volume in real-world production environments, the mean and covariance matrices are updated incrementally through incremental calculations. The incremental calculation process is as follows:

[0110] ,

[0111] ,

[0112] in, This is the mean at the current moment. This is the average value at the previous moment. For the new data point at the current moment, The total number of data points at the current moment. Let be the covariance matrix at the current time. This is the covariance matrix from the previous time step. For the eigenvectors... The formula for calculating its distance from the historical distribution is:

[0113] ,

[0114] in , This represents the Hadamard product (element-wise multiplication). By normalizing the covariance matrix and incorporating it into the Mahalanobis distance (defined as a weighted projection of the eigenvalues ​​based on the inverse covariance matrix), this method effectively eliminates scale and unit differences between features from heterogeneous sensors. The calculation formula is:

[0115] ,

[0116] This is a temperature coefficient used to control the steepness of the function curve. Let be the variance of the i-th sensor.

[0117] The final step is to determine the threshold during the training phase. The decision boundary is dynamically adjusted in real time, and this update is calculated based on the mean and standard deviation of the historical distance distribution. The calculation formula is as follows:

[0118] ,

[0119] in, The drift threshold at the current moment, The average historical distance at the current moment. The standard deviation of the historical distance at the current moment. It is a constant (used to control the degree of deviation of the threshold from the mean).

[0120] Phase 2 of S2.1.2: Real-time Drift Detection

[0121] During the inference phase, the drift threshold is no longer updated. Instead, statistical features obtained during the training phase are used to perform dynamic comparative analysis on newly arriving time series. Specifically, stable statistical data and thresholds are first calculated based on historical normal data. These parameters depict the stable pattern of the system in a "healthy" state. When a new time window arrives, the industrial enhanced Mahalanobis distance between this window and the historical distribution is calculated using the same distance calculation process. This distance not only considers the correlation between feature dimensions but also introduces weighted and robust adjustments tailored to the characteristics of the industrial system, making it more tolerant of fault fluctuations and minor noise. The drift detection result is shown in the following formula:

[0122] ,

[0123] After calculating the distance value, it is compared with a pre-set threshold. If the distance value exceeds the threshold, that is... This indicates a significant deviation between the current time window and the historical normal distribution, suggesting that the system behavior may have experienced statistical drift or abrupt state changes. This often signifies changes in industrial equipment, production processes, or sensing systems, triggering a drift alarm. Conversely, if If the current window's features remain within the historical distribution range, its changes can be considered normal fluctuations, and the system is in a stable operating state. Through this mechanism, drift detection during the inference phase maintains both timeliness and stability, providing a basis for decision-making in subsequent fault alarms and incremental training of the fault detection model.

[0124] S2.2 Diffusion-enhanced fault detection

[0125] The core objective of the diffusion strategy is to enhance the model's robustness to concept drift while reducing training overhead caused by information bottlenecks. Following this design principle, this invention deliberately avoids the variational autoencoder (VAE) architecture, which inherently creates information bottlenecks, and instead adopts a more advanced Transformer-based framework. The proposed diffusion-based Transformer architecture not only more effectively captures the temporal features in industrial data streams but also significantly reduces training costs by externalizing information bottlenecks. Furthermore, to address the challenge of scarce labeled anomaly data in real-world production scenarios, this invention employs unsupervised learning methods for fault detection, thereby better adapting to the data characteristics of industrial environments. The original time-series data is input into the diffusion-based fault model (DA). Figure 2 As shown, the DA model mainly includes a diffusion enhancement module, a word embedding module, a feature extraction module, and a reconstruction module. The specific architecture and data processing steps are as follows:

[0126] S2.2.1 Phase 1: Training Parameters and Thresholds

[0127] During the training phase, such as Figure 2 As shown in the first stage, the model parameters of the Deconstruction (DE) are iteratively updated using a sliding window input mechanism. Subsequently, the fault threshold is derived based on the calculation of the error score. The DE consists of three steps: injecting temporal noise through a diffusion process, extracting features from the noise sequence, and calculating the reconstruction error. The entire process is as follows: Figure 3 As shown.

[0128] Time-domain noise is injected through a diffusion process: the input to the diffusion module is... The original vector contained therein at a certain moment can be represented as: The vectors of M-step Gaussian noise pollution can be expressed as follows: And the pollution result obtained in step m is only consistent with Related. Data The result obtained after m steps of noise pollution is:

[0129] ,

[0130] in Standard Gaussian noise, . This is a noise scheduling parameter that controls the degree of noise addition, and it typically decreases with time step m. ,in It is the noise rate, which increases with time step. The calculation process is as follows:

[0131] ,

[0132] in and Here, represents the preset minimum and maximum noise rates, and M is the total number of time steps. Therefore, we can obtain the output of the diffusion strategy:

[0133] ,

[0134] in, , and .

[0135] Position encoding: The variable values ​​after noise addition by the diffusion module are position encoded, as shown in the following formula:

[0136] ,

[0137] in, Represents convolution. This represents the variable value after position encoding. The relative position encoding is described by the following formula:

[0138] ,

[0139] Indicates the position of the moment within the time window. This represents the dimension of the vector embedding. express Even-numbered dimensions in express Odd dimensions in.

[0140] Feature extraction from noisy sequences: The sequence after word embedding is input into the feature extraction structure to extract complex spatiotemporal relationships in the data. The feature extraction module consists of L layers of blocks, where the input of the l-th layer block is... It has a time length as well as Dimension, initial input This represents the embedded noise sequence, which we obtain in the timing module. The calculation process can be expressed as follows:

[0141] ,

[0142] ,

[0143] in, Presentation layer standardization This indicates a multi-head attention mechanism. This represents the temporal hiding representation of the l-th layer. This represents a feedforward network. Similarly, in the spatial module, we obtain... The calculation process can be expressed as follows:

[0144] ,

[0145] ,

[0146] in This represents the spatial hiding representation of the l-th layer. This represents the transpose. Finally, we obtain the (l+1)th layer. The process is as follows:

[0147] ,

[0148] ,

[0149] in This indicates that the symbol is hidden. This indicates a splicing operation.

[0150] To achieve the goal of unsupervised training, a reconstruction module consisting of two fully connected layers is added to the end of the model, aiming to project the output dimension back to the original input dimension. The reconstruction process can be described as follows:

[0151] ,

[0152] in express The reconstruction results This indicates a fully connected layer.

[0153] Reconstruction loss calculation: Calculate the output of the model. With the original input The reconstruction error is used as the loss function value and backpropagated back to the neural network to facilitate model parameter updates. Simultaneously, using the reconstruction error as the loss function helps the model find complex correlations between time series and various sensors. The calculation process is as follows:

[0154] ,

[0155] in This represents the Frobenius norm.

[0156] S2.2.2 Phase Two: Fault Assessment Methods

[0157] The model generates a reconstruction error score for the sensor values ​​at each time stamp. ,in .

[0158] The inherent distributional difference between faulty and normal data is reflected in the reconstruction error score, with faulty instances scoring significantly higher than normal instances. Based on this discriminative characteristic, the framework implements a threshold detection mechanism: if the error score falls below a certain threshold... The instances were judged to be normal, exceeding [a certain threshold]. If the data is abnormal, it is marked as such. This threshold determination method is not only simple and efficient, but also effectively distinguishes abnormal data from normal data, providing reliable technical support for real-time anomaly detection in industrial scenarios.

[0159] S3. Model Training Optimization

[0160] During training, the hyperparameters of DDA-DE were configured as follows: the time step m for adding noise in the diffusion module was set to 2000, and the preset minimum and maximum noise rates β were set. min and β max Specifically, the feature extraction module has 4 layers L and an embedding dimension of [missing information]. The size of the sliding window is 128. The batch size is 64, the loss function is the reconstruction loss introduced in S2.2.1, the entire model is updated using a mini-batch stochastic gradient descent strategy, the batch size is 64, the entire training set is traversed in each round of training, the optimizer is ADAM, the learning rate is 1e-4, and the training cycle is 20.

[0161] During training, the average loss between the training and validation sets is calculated after each epoch to monitor model convergence. To prevent overfitting, a dropout of 0.2 is added between transformer layers, and training is terminated early if the validation set error fails to improve after several consecutive epochs—an early stopping mechanism. If the validation set reconstruction error does not decrease for five consecutive epochs, a learning rate decay mechanism is triggered, automatically reducing the learning rate by half to refine the search. To further ensure training stability, gradient pruning is performed, truncating gradients when the norm exceeds a threshold of 5.

[0162] The entire training process is iterative, and the model gradually learns the normal patterns of the time series. When the network parameters converge and the validation set error is minimized, the model weights are saved. The final model can effectively reconstruct normal sequences during the inference phase, but it will exhibit a large reconstruction error for fault sequences. This characteristic will be used in the subsequent fault detection phase to determine the existence of fault points.

[0163] S4 Model Deployment

[0164] The model outputs the reconstruction error score of the sensor values ​​at each time point. ,in This invention determines whether a data point is faulty based on the drift detection result and a threshold in step S2: if the reconstruction error score is lower than a preset threshold... If no drift is determined in step S2, then the data at that time point is considered normal; if the reconstruction error score is greater than or equal to the threshold... If no drift is determined in step S2, it is considered a fault; if the reconstruction error score is less than the threshold... If a drift is determined to have occurred in step S2, then the data at that time point is considered normal, and an incremental model update is performed; if the reconstruction error score is greater than or equal to the threshold... If a drift is determined to have occurred in step S2, then the threshold is adjusted. And re-evaluate.

[0165] S4.1 makes a preliminary assessment to determine if a fault has occurred. The specific assessment method is as follows:

[0166] ,

[0167] S4.2 If a drift is determined to have occurred in step S2, the abnormal threshold needs to be adjusted and re-evaluated. The adjustment method is as follows:

[0168] ,

[0169]

[0170] S4.3 Final Adjustments

[0171] If no fault is detected and no concept drift occurs, no record is made; if an anomaly is detected but no concept drift occurs, the fault is recorded and monitoring continues; if no anomaly is detected but concept drift occurs, the DE is incrementally trained to adapt to changes in data distribution; if both anomaly and concept drift are detected, the threshold is adjusted and re-evaluated according to S4.2.

[0172] S4.4 Repeat S4.1-4.3 for each sample to be tested to achieve fault detection of attack behavior in the data.

[0173] Experimental verification

[0174] To verify the effectiveness of DDA-DE, we primarily evaluated it based on two real-world datasets: (1) the Safe Water Treatment (SWaT) dataset, collected from an industrial water treatment plant producing filtered water, recording 11 days of operational data (7 days of normal operation and 4 days of continuous attack), containing labeled data from 51 sensors and actuators to distinguish between normal and abnormal behavior. (2) the Water Distribution (WADI) dataset, an extension of SWaT, integrating the three major processes of water treatment, storage, and distribution networks to simulate a complete real-world water management cycle. This dataset covers 16 days of operational data (14 days of normal operation and 2 days of attack), recording data from 123 sensors and actuators. Notably, this study excluded the Mars Science Laboratory (MSL) rover and the Soil Moisture Active Passive (SMAP) satellite datasets—datasets collected by NASA to monitor the status of sensors and actuators on Mars rovers—due to inherent recorded defects that prevented rigorous analysis. To optimize computational efficiency and verify the effectiveness of the method, continuous variables in the experimental datasets were selectively retained during model construction. The training set was repartitioned into training and validation sets in an 8:2 ratio. The adjusted SWAT dataset contains 24 dimensions, with 396,000 samples in the training set, 99,000 samples in the validation set, 449,919 samples in the test set, and 156,915 samples in the validation set. Outliers account for 12.14% of the test set. The WADI dataset contains 59 dimensions, with 627,656 training samples, 156,915 validation samples, and 172,803 test samples. Outliers account for 5.75% of the test set. See Table 1 for details of the adjusted dataset specifications.

[0175] Table 1. Dataset Details

[0176] Dataset Dimension attack training set Validation set test set Failure rate (%) SWAT 24 41 396000 99000 449919 12.14 WADI 59 15 627656 156915 172803 5.75

[0177] This invention selects recent research results based on deep learning as model benchmarks, namely D3R and test-time tuning methods. The baseline methods are described below:

[0178] D3R: A dynamic decomposition method based on diffusion reconstruction, which decomposes the original non-stationary sequence into trend and stationary terms through a hierarchical decoupling mechanism, and combines diffusion model to reconstruct highly noise-contaminated data to enhance the ability to extract essential features.

[0179] DE*: This indicates that the modeling stage is carried out using the DE model designed in this paper, and at the same time, the test-time adjustment and detrending strategy proposed by Kim et al. is combined. This strategy is to use statistical estimation methods to remove the trend estimate of time series data, and then carry out online monitoring and use the detrended series to update the model parameters.

[0180] The performance evaluation of the model and baseline model in this invention employs F1 score, PATE, and AUC based on correlation. The F1 score is obtained by combining the average directional distance between predicted anomalies and real events to calculate precision (P), and the average directional distance between real events and predicted anomalies to determine recall (R), and then calculating the F1 score accordingly. Proximity-based time series anomaly evaluation (PATE) is an evaluation metric that combines the temporal correlation between predicted and anomaly intervals, and considers the anomaly interval buffer zone through proximity weighting. The area under the curve (AUC) in precision-recall (PR) analysis is formally defined as the integral of the PR curve; this metric quantifies the overall performance of the fault detection model at different thresholds.

[0181] Table 2 Comparison of experimental results

[0182]

[0183] Each method's unique parameters follow the optimal configuration outlined in its corresponding paper. As shown in the figure, compared to the current state-of-the-art methods, experimental results demonstrate that the DDA-DE model outperforms the state-of-the-art D3R model on the SWAT dataset with a 2.46% improvement in F1 score (0.7747→0.7993), and achieves a 3.7% improvement on the WADI dataset (0.7017→0.7387). It is noteworthy that the WADI dataset typically limits model performance due to its high-dimensional features, low-frequency sampling, and covert attack patterns, while the dense sampling and explicit attack features employed by SWAT are more conducive to the performance of existing algorithms. Compared to the test-time tuning strategy proposed by Kim et al., this invention effectively models the coupling relationship between trend and seasonal components and continuously captures evolutionary trends during deployment, resulting in a 4.96% improvement in F1 score on the SWAT dataset (0.7497→0.7993) and a 2.6% improvement on the WADI dataset (0.7127→0.7387).

[0184] In the PATE metric, which emphasizes time dependence, the inherent advantages of the process design and Transformer architecture of this invention enabled DDA-DE to exhibit the best performance, with an improvement of 24.46% (0.6001→0.8447). On the WADI dataset, DDA-DE achieved the best score of 0.3088, while DA* achieved the second best score of 0.1924. Under the PATE metric, all models performed worse than the SWAT benchmark on the WADI dataset. Notably, the performance degradation of DDA-DE exceeded 50%, which is mainly attributed to the inherent complexity of WADI as a large-scale water supply system. The spatial dispersion and temporal sparsity of WADI anomalous events exacerbate time series noise and amplify runtime variability, thereby fundamentally weakening the model's ability to capture comprehensive temporal proximity.

[0185] DDA-DE achieved the highest AUC scores (88.37% and 57.69%) on both the SWAT and WADI datasets. This performance gap underscores the higher demands placed on modeling capabilities by the dynamic and high-dimensional characteristics of complex systems in WADI, further exacerbating the challenges of fault detection. These findings establish enhancing the spatiotemporal modeling capabilities of complex industrial systems as a core direction for our future research.

[0186] This invention explains the research motivation from a visual perspective. Figure 4 and Figure 5 The drift detection normalization results for all SWaT and WADI datasets are presented systematically. It is worth noting that although both datasets originate from similar industrial processes, they exhibit distinctly different data characteristics—we believe this difference is the key determinant of the performance differences between the models on these two benchmark datasets.

[0187] Figure 4 Figures 5 and 6 respectively demonstrate the normalized detection instances of the DDA-DE framework on the SWaT dataset when faults occur and when there are consecutive faults, revealing its dual-threshold adaptive mechanism: when the thresholds of drift detection and fault detection outputs exceed the limit at the same time, the system implements dynamic threshold recalibration; while in the scenario where the drift detection signal exceeds the preset threshold but is not accompanied by an abnormal signal, incremental model updates are selectively triggered.

[0188] Figure 6 The architecture of this paper is illustrated by a three-layer time series diagram: the bottom layer is the original time series (including labeled drift points and outliers), the middle layer is the drift detection output (including drift confidence curves and threshold trigger markers), and the top layer is the fault detection scoring curve (including dynamic threshold lines that are adjusted in real time according to drift results).

[0189] The drift detection results show that drift is triggered at the beginning of the test set, indicating a significant distributional shift between the test set and the normal points in the training set. This clearly demonstrates the non-stationarity of the time series data—its statistical properties drift over time. Furthermore, the underlying plot reveals that the fault detection score does not react sensitively to drift events.

[0190] Example 2

[0191] This embodiment provides an industrial fault detection system based on dynamic drift sensing and diffusion enhancement, including:

[0192] The data acquisition module is configured as follows:

[0193] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned industrial fault detection method based on dynamic drift sensing and diffusion enhancement.

[0194] A terminal device includes a processor and a computer-readable storage medium, the processor being configured to implement various instructions; the computer-readable storage medium being configured to store multiple instructions adapted for loading and execution by the processor of the aforementioned industrial fault detection method based on dynamic drift sensing and diffusion enhancement.

[0195] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for industrial fault detection based on dynamic shift perception and diffusion enhancement, characterized by, The method comprises the following steps: Acquiring industrial flow data, wherein the industrial flow data comprises equipment vibration spectrum and temperature fluctuation; The unsupervised fault detection DDA-DE model is constructed, wherein dynamic shift perception DDA is used to process industrial data flow to establish a statistical distribution baseline and a shift threshold, specifically, statistical descriptors are extracted from historical data by using DDA, baseline distribution parameters including a mean vector of feature vectors are calculated and a covariance matrix to realize real-time distance calculation, so as to continuously update the shift threshold; and the gradual update of the mean and the covariance matrix is realized by incremental calculation; The initial model parameters and the abnormal threshold are determined by diffusion enhancement anomaly detection DE, specifically, the statistical features obtained in the training stage are used for dynamic comparison analysis of the newly arrived time series, wherein, first, the stable statistical data and the threshold are calculated based on the historical normal data, which is used to describe the stable mode of the system in the healthy state, when a new time window arrives, the industrial enhancement Mahalanobis distance between the window and the historical distribution is calculated through the same distance calculation process, and the tolerance to fault fluctuations and slight noise is obtained by introducing the weighting and robust adjustment for the characteristics of the industrial system; after the distance value is calculated, the distance value is compared with the set threshold, if the distance value exceeds the set threshold, that is , it indicates that the deviation of the current time window from the historical normal distribution is significant, which indicates that the system behavior may have statistical drift or state mutation, if , it indicates that the characteristics of the current window are still within the range of the historical distribution. Robustness enhancement of concept drift of an unsupervised fault detection model based on diffusion strategy; by establishing a double-branch detection architecture based on the difference between concept drift and abnormal events, for preserving the feature correlation within the data, including a drift detection module and a fault detection module, wherein the drift detection module adopts a covariate shift quantification model based on a sliding window, and by introducing an industrial enhanced Mahalanobis distance metric, it dynamically tracks the feature space distribution changes in real time; the fault detection module integrates diffusion enhancement strategy, enhances the distribution shift robustness through multi-scale feature propagation mechanism, and suppresses noise interference with the help of information bottleneck constraint, dynamically calibrates the model parameters based on the detection results of the two branches, and optimizes the fault detection threshold through an adaptive threshold adjustment mechanism; Model optimization is performed on the constructed unsupervised fault detection DDA-DE model; Fault detection is performed by using the optimized model.

2. The method as claimed in claim 1, wherein, The incremental calculation process is: ， ， where, is the mean of the current time, is the mean of the previous time, is the new data point of the current time, is the total number of data points of the current time, is the covariance matrix of the current time, is the covariance matrix of the previous time, for the eigenvector The formula for calculating the distance of the eigenvector from the historical distribution is: ， wherein , denotes the Hadamard product, which eliminates the scale and unit differences between heterogeneous sensor features by normalizing the covariance matrix into Mahalanobis distance; The calculation formula of is: ， is the temperature coefficient, is the variance of the ith sensor; finally threshold determination is performed in the training phase, where the decision boundary is dynamically adjusted by real-time update, which is calculated based on the mean and standard deviation of the historical distance distribution, the calculation formula is: ， wherein, is a drift threshold value for the current time instant, is a historical distance mean for the current time instant, is a historical distance standard deviation for the current time instant, is a constant.

3. The method as claimed in claim 2, wherein the method is based on dynamic shift perception and diffusion enhancement for industrial fault detection. The robustness of the unsupervised fault detection model based on the diffusion strategy concept drift is enhanced, including iterative updating of the model parameters of DE through a sliding window input mechanism, deriving a fault threshold based on error score calculation, including injecting time series noise through a diffusion process, extracting features from the noise sequence, and calculating reconstruction error, wherein injecting time domain noise through a diffusion process includes setting the input of the diffusion module as wherein the original vector at a certain time is represented as , the vector contaminated by M-step Gaussian noise can be represented as , and the contamination result obtained at the mth step is only related to , and the data The result obtained after m-step noise contamination is: ， wherein is a standard Gaussian noise, , is a noise schedule parameter, decreasing with time step m; ， wherein is the noise rate, incremented with time step, The calculation process is: ， where and are preset minimum and maximum noise rates, M is the total number of time steps, and the output of the diffusion strategy is finally obtained: ， wherein, , and ; the variable value after the diffusion module is added with noise, and is represented as: ， wherein, denotes a convolution, denotes a variable value that is position encoded, denotes a relative position encoding, denoted as: ， denotes the position of the time instant in the time window, denotes the dimension of the vector embedding, denotes the even dimensions in denotes the odd dimensions in 4. The method as claimed in claim 3, wherein the method is based on dynamic shift perception and diffusion enhancement for industrial fault detection. The feature extraction from the noise sequence includes inputting the sequence after the word embedding into a feature extraction structure for extracting complex space-time correlation in the data, wherein the feature extraction module is composed of L layers of blocks, wherein the input of the block module of the lth layer is , has a time length and dimension, the initial input represents the embedded noise sequence, and the time sequence module obtains , and the calculation process is represented as: ， ， wherein, denotes layer normalization, denotes multi-head attention mechanism, denotes the time-step hidden representation of the l-th layer, denotes a feed-forward network, again obtained in the spatial module is denoted as: ， ， wherein denotes the spatially hidden representation of the l-th layer, denotes the transpose; finally resulting in the l+1-th layer is obtained by the process: ， ， wherein represents a hidden representation, represents a concatenation operation.

5. The method for industrial fault detection based on dynamic shift perception and diffusion enhancement as claimed in claim 4, wherein, The reconstruction error is calculated, including adding a reconstruction module composed of two fully connected layers at the end of the model to project the output dimension back to the original input dimension to realize the unsupervised training goal, and the reconstruction process is expressed as: ， wherein denotes the reconstruction result of denotes a fully connected layer; then the output result of the model is calculated and the reconstruction error with the original input as a loss function value, and is back-propagated back to the neural network, facilitating the updating of the model parameters; at the same time, the reconstruction error is used as a loss function, facilitating the model to find the complex correlation between the time sequence and each sensor, and the calculation process is: ， wherein denotes the Frobenius norm; and generate a reconstruction error score for each time-stamped sensor value wherein the intrinsic distribution difference between data with faults and normal data is embodied in the reconstruction error score, and error scores below a threshold are determined to be normal, and scores exceeding are flagged as abnormal.

6. The method as claimed in claim 5, wherein the method is based on dynamic shift perception and diffusion enhancement for industrial fault detection. The model optimization of the constructed unsupervised fault detection DDA-DE model includes setting the time step m of the noise added in the diffusion module of the DDA-DE to 2000, setting the preset minimum and maximum noise rates and , the number of layers L in the feature extraction module is 4, the embedding dimension is 128, the size of the sliding window is 64, the loss function adopts reconstruction loss, the model adopts a small batch stochastic gradient descent strategy for updating, each round of training traverses the complete training set, the optimizer adopts ADAM, and the average loss of the training set and the validation set is calculated after each epoch is completed in the training process to monitor the convergence of the model; in order to prevent overfitting, dropout is added between the transformer layers, and early termination is adopted when the validation set error does not improve for a certain number of consecutive rounds, that is, an early stopping mechanism is adopted; If it is found during training that the reconstruction error of the validation set does not decrease for 5 consecutive epochs, the learning rate decay mechanism is triggered to automatically reduce the learning rate by half to refine the search; the whole training process is continuously iterated, and the model gradually learns the normal mode of the time series; when the parameters of the network converge and the validation set error is the lowest, the model weight is saved.

7. The method as claimed in claim 6, wherein the method is based on dynamic shift perception and diffusion enhancement for industrial fault detection. The fault detection using the optimized model includes reconstructing error scores for sensor values at each time point according to the model output , wherein , according to the drift detection result and a threshold value, it is determined whether the data point is faulty: if the reconstructing error score is lower than a preset threshold value and it is determined that no drift occurs, it is determined that the data at the time point is normal. if the reconstruction error score is greater than or equal to the threshold value and it is determined that no drift occurs, then it is determined that a fault has occurred; if the reconstruction error score is less than the threshold value and it is determined that drift has occurred, then it is determined that the data at the time point is normal, and the model is incrementally updated; if the reconstruction error score is greater than or equal to the threshold value and it is determined that drift has occurred, then the threshold value is adjusted and re-evaluation is performed; wherein the manner of determining whether a fault has occurred is as follows: ； If it is judged that drift occurs, the abnormal threshold is adjusted and re-evaluated, and the adjustment method is: ， ; finally, if no anomaly is detected and no concept drift occurs, nothing is logged; if an anomaly is detected but no concept drift occurs, a fault is logged and monitoring continues; if no anomaly is detected but concept drift occurs, the DE is incrementally trained to adapt to the change in data distribution; if both anomaly and concept drift are detected, the thresholds are adjusted and re-evaluated.

8. An industrial fault detection system based on dynamic shift-awareness and diffusion enhancement, performing an industrial fault detection method based on dynamic shift-awareness and diffusion enhancement as claimed in claim 1, characterized by, The method comprises the following steps: The model construction module is configured to construct an unsupervised fault detection DDA-DE model, wherein industrial data streams are processed by a dynamic drift perception DDA to establish a statistical distribution baseline and a drift threshold; diffusion enhancement anomaly detection DE is used to determine the initial model parameters and the abnormal threshold; and robustness enhancement of concept drift of the unsupervised fault detection model based on diffusion strategy; The model optimization module is configured to optimize the constructed unsupervised fault detection DDA-DE model; The fault detection module is configured to perform fault detection by using the optimized model.

Images