Industrial abnormal root cause diagnosis method and system based on decoupling causal characterization

Abstract

The invention relates to the technical field of industrial process monitoring, and particularly discloses an industrial abnormal root cause diagnosis method and system based on decoupling causal characterization. The objective of the invention is to solve the technical pain point that the existing anomaly detection technology stops alarming and cannot distinguish correlation and causality, so that root cause positioning is difficult. The core innovation of the invention lies in constructing a causal decoupling variational auto-encoder (CD-VAE) model, and learning a group of low-dimensional, independent and physical potential causal factor representations from high-dimensional industrial time series data. By introducing time sequence causal constraint and decoupling regularization, it is ensured that potential factors correspond to a key causal mechanism in the system. When an anomaly occurs, a fundamental causal factor, instead of a surface-related observation variable, which causes a fault is accurately positioned by calculating a variation score of posterior distribution of a potential factor and tracing a causal path of the potential factor. According to the method, the crossing from anomaly detection to root cause diagnosis is realized, and the operation and maintenance efficiency is remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of industrial process monitoring and fault diagnosis technology, specifically relating to a method and system for diagnosing the root causes of industrial process anomalies based on big data analysis and causal inference. This method is particularly suitable for processing high-dimensional, strongly coupled, and nonlinear industrial time-series data. It can decouple physically meaningful potential causal factors from observed variables and utilize these factors to achieve accurate anomaly root cause localization. It can be widely applied in industrial IoT scenarios such as intelligent manufacturing, process industries, and equipment health management. Background Technology

[0002] As industrial processes become increasingly complex and intelligent, higher demands are placed on their operational safety and reliability. Real-time detection and diagnosis of process anomalies have become crucial for ensuring production safety, improving product quality, and reducing maintenance costs. However, current mainstream industrial anomaly diagnosis methods still have significant limitations, failing to meet the needs for accurate and rapid root cause localization.

[0003] First, statistical and machine learning methods, such as multivariate statistical methods like Principal Component Analysis (PCA) and Partial Least Squares (PLS), and machine learning algorithms like Support Vector Machines (SVM) and Isolation Forest, all focus on learning the observed variables. The correlation pattern under normal operating conditions. The decision function of this type of method is usually in the form of: The basis for anomaly judgment is Whether it is below a preset threshold. However, such models are essentially "black box" or "grey box" structures, and their root cause localization methods (such as contribution graph methods) rely on variables. Model decision Contribution Essentially, this still falls under correlation or sensitivity analysis, rather than causal analysis. Furthermore, in industrial systems with complex coupling characteristics, numerous spurious correlations often exist between variables, leading to the variable with the highest contribution often being a "victim" or accompanying phenomenon of the fault, rather than the root cause of the fault itself. For example, when bearing wear causes increased vibration and temperature as the root cause, the location method based on contribution maps may incorrectly identify temperature as a high-contribution variable, thus misleading maintenance operations.

[0004] Secondly, methods based on causal discovery, such as PC algorithms, Fast Causal Inference (FCI), and LiNGAM, attempt to directly construct causal graph structures between variables from observational data. ,in For variable nodes, For directed edges. However, this type of method has fundamental flaws: on the one hand, its stability and accuracy decrease significantly in high-dimensional, noisy industrial scenarios with latent variables; on the other hand, the granularity of the learned macroscopic causal graph is relatively coarse, making it difficult to reveal the variation of microscopic and unobservable causal factors inside the equipment due to component performance degradation, and such microscopic variations are precisely the root cause of most early failures.

[0005] In summary, existing technologies generally suffer from a cognitive gap that "correlation does not equal causation," and lack an effective technical means to automatically learn from data and decouple potential causal factors, thereby achieving accurate root cause tracing. Summary of the Invention

[0006] The purpose of this invention is to provide an early abnormality root cause diagnosis scheme that can penetrate the appearance and reach the physical essence of the fault, overcoming the shortcomings of the existing technology in that the root cause is not accurate, not fast, and relies on expert experience.

[0007] The essence of this invention lies in avoiding direct root cause competition at the level of observed variables. Instead, it maps the observation space to a low-dimensional, decoupled latent causal space through a carefully designed deep generative model. Furthermore, this invention assumes that each dimension in this latent space corresponds to an independent, key causal factor driving the operation of an industrial process. On the other hand, the nature of a failure is defined as a non-negligible shift in the probability distribution of one or more such causal factors.

[0008] Compared with the prior art, the present invention can produce the following beneficial effects:

[0009] First, this invention systematically applies causal decoupling representation learning to industrial root cause diagnosis, realizing a paradigm shift in analysis from "relevant appearances" to "causal kernels".

[0010] Secondly, in terms of diagnostic accuracy, the fault mechanism is directly located by decoupling the potential factors, which effectively avoids the interference of pseudo-root causes and makes the diagnostic results have clear physical meaning, thereby greatly improving the accuracy of operation and maintenance decisions.

[0011] Furthermore, in terms of early warning capabilities, since variations in potential causal factors usually precede significant anomalies in macroscopic observation indicators, this method can capture such early and subtle changes, enabling early warning of faults.

[0012] Finally, in terms of automation and intelligence, this method significantly reduces the reliance on detailed domain knowledge and expert experience, achieving end-to-end automated root cause analysis, and is applicable to large-scale and complex industrial systems.

[0013] According to the technical solution of the present invention, the method includes the following steps:

[0014] First, in the data preprocessing and causal prior knowledge representation steps, multivariate time-series data of the industrial process are collected. Standardization, missing value handling, and noise reduction are performed.

[0015] Furthermore, domain knowledge (such as known causal constraints) is formalized into a priori causal adjacency matrix. ,in This represents the number of potential causal factors. If the factors are known... yes The direct cause is... .

[0016] Secondly, in the causal decoupling variational autoencoder model construction step, the encoder A bidirectional long short-term memory network is used as input, with observation window data as input and potential causal factors as output. The posterior distribution parameters.

[0017] On the other hand, decoder A time series generation model, based on and optional context variables This reconstructs the data for the entire observation window.

[0018] Furthermore, in the causal decoupling regularization process, the following key regularization terms are introduced into the loss function of the standard VAE to force... Each dimension is independent and possesses a causal temporal structure. Specifically, it includes the total correlation (TC) penalty term. This promotes independence. Among them... Product with its marginal distribution The KL divergence between them.

[0019] Temporal Causal Consistency Constraint Using the prior matrix Construct a causal mask matrix It is applied to the evolution of latent factors. yes The The line-defined mask ensures It is only affected by its causal parent node. The constraint is parameterized by a neural network and is implemented by embedding a mask in a graph neural network (GNN) or a Transformer.

[0020] Furthermore, during the model training phase, maximum likelihood estimation is employed to optimize the model parameters using historical data under normal operating conditions. and The total loss function is ,in To reconstruct the loss, Let KL divergence be the KL divergence. For the standard normal prior, , , Hyperparameters used to control the weights of each item.

[0021] Finally, in the anomaly detection and root cause analysis phase, anomaly detection is performed by online computation of input data. Reconstruction probability Alternatively, an Evidence Lower Bound (ELBO) can be used, triggering an anomaly alarm when the value falls below a preset threshold. Root cause tracing includes:

[0022] Latent factor variation calculation, obtaining the posterior distribution of outlier data points. And calculate each causal factor The variation fraction A scoring method based on Wasserstein distance is used: ,in It is the baseline distribution estimated on the normal validation set;

[0023] Causal path activation analysis, targeting factors with high variance scores, utilizes a pre-trained time-series causal model. By tracing back its causal parent nodes over a period of time, analyzing the abnormal propagation path, and identifying the initial source of mutation; Attached Figure Description

[0024] Figure 1 System overall architecture diagram;

[0025] Figure 2 Schematic diagram of the causal decoupling variational autoencoder model structure;

[0026] Figure 3 Schematic diagram of anomaly detection and root cause tracing process; Detailed Implementation

[0027] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that the following embodiments are intended to facilitate understanding of the present invention and are not intended to limit it in any way.

[0028] First, refer to Figure 1 The overall architecture block diagram of an industrial anomaly root cause diagnosis system based on decoupled causal representation according to an embodiment of the present invention.

[0029] S101 Data Preprocessing Module: Responsible for preparing high-quality data for subsequent modeling. Its specific implementation methods include:

[0030] Data Acquisition and Integration: Collect multi-dimensional time-series observation data from historical databases, real-time data platforms (such as SCADA and DCS), or sensor network interfaces in industrial settings. ,in For the number of time points, The dimension of the observed variable (such as temperature, pressure, vibration amplitude, etc.).

[0031] Data cleaning and reduction: Preprocessing of the raw data. Z-score normalization is used to eliminate the influence of different physical dimensions; linear interpolation or K-nearest neighbor (KNN) based algorithms are used to address missing data; wavelet transform or moving average filters are used for data denoising to improve the signal-to-noise ratio.

[0032] Causal Prediction Identification and Representation: This module is also responsible for integrating domain knowledge. It formalizes expert experience, process mechanism diagrams, or known device dependencies into a priori causal adjacency matrix. ,in This represents the preset number of potential causal factors. If the factors are known... yes The direct cause is... This matrix will be fed into the model training module S102 as key prior knowledge.

[0033] The S102 model training module is responsible for building and optimizing the causal decoupled variational autoencoder (CD-VAE) model. (Reference) Figure 2 Its model structure and training process are as follows:

[0034] S201 Encoder: Encoder It is implemented using a bidirectional long short-term memory (Bi-LSTM) network structure. Its input is a length of... Standardized observation data window The output is a potential causal factor. The posterior distribution parameters, i.e., the mean. and diagonal variance .

[0035] Specifically, assuming the potential space dimension is Then the posterior distribution is modeled as a diagonal Gaussian distribution. .

[0036] S202 Decoder: Decoder Designed as a temporal generative model, for example using an LSTM or Transformer decoder. It receives latent variables sampled from the encoder. and optional context variables (such as operating condition mode labels), and reconstruct the data for the entire observation window.

[0037] Causal decoupling regularization: The core of this implementation lies in the regularization term introduced on the standard VAE loss function.

[0038] Total Relevance (TC) Penalty This regularization term is used to minimize the latent factors. Mutual information among the dimensions promotes their statistical independence. In practice, a mini-batch weighted sampling strategy is used to estimate the aggregated posterior distribution. Its marginal distribution product The KL divergence between them.

[0039] Temporal Causal Consistency Constraint This constraint utilizes the prior causal adjacency matrix obtained from module S101. Specifically, a causal mask matrix is ​​constructed. Its row vector Depend on The The row definition. During the evolution of the latent factors, this constraint is implemented through a nonlinear state-space model (e.g., embedding this mask in a graph neural network (GNN) or a masked Transformer) to ensure... Only affected by its causal parent node (by The impact of (specified).

[0040] Model optimization: During the model training phase, historical datasets under normal operating conditions are used to optimize the model parameters. Optimization is performed. The optimization objective is to minimize the total loss function. .in, To reconstruct the loss, Let KL divergence be the KL divergence. These are hyperparameters. The training process uses the Adam optimizer with stochastic gradient descent until the model converges.

[0041] S103 Online Monitoring and Traceability Module: This module is responsible for deploying the trained model in the production environment to achieve real-time monitoring and diagnosis.

[0042] refer to Figure 3 Anomaly detection and root cause tracing process flowchart:

[0043] S301 Anomaly Detection: During online monitoring, new real-time data windows... The lower bound of evidence ELBO or reconstruction probability is calculated using the trained CD-VAE model.

[0044] Furthermore, this statistic is compared with a preset threshold based on historical quantiles of normal data. When this value remains below the threshold, the system triggers an anomaly alarm.

[0045] S302 Root Cause Tracing: Once an anomaly is detected, the root cause tracing process is initiated immediately.

[0046] Latent factor variation calculation: Identifying outlier data points Corresponding posterior distribution For each potential causal factor Calculate its distribution relative to the normal baseline distribution. The variation fraction .

[0047] This embodiment preferably uses the Wasserstein distance ( Measuring distance: .in, It is the marginal distribution of each factor estimated on an independent normal validation set.

[0048] Causal path activation analysis: on the fraction of variation The data is sorted to identify the one or more key factors with the highest scores. Furthermore, using a trained temporal causal model, the causal parent nodes of these key factors are traced back within a past time window to analyze the propagation chain of anomalies, thereby determining the root cause factor that initially caused the mutation. .

[0049] Result delivery: The located root cause factors The relevant information is then passed to the result output module S104.

[0050] S104 Result Output Module: This module is the interactive interface between the system and the user, responsible for presenting the diagnostic results clearly and intuitively.

[0051] Report Generation: The module receives root cause information from module S103 and automatically generates a structured root cause diagnosis report. The report content is formatted as follows: Root Cause Factors: (For example, the physical meaning of 'bearing wear'), variation fraction: Confidence level: (This confidence level can be calculated based on statistical tests or the uncertainty of the model output.)

[0052] Visualization: Furthermore, through human-machine interface (HMI), web dashboard or operation and maintenance management platform, key information such as root cause reports, anomaly propagation path diagrams, and potential factor variation trends can be visualized, and operation and maintenance personnel can be provided with historical case query and decision support functions.

[0053] In summary, this implementation method, through four clearly defined and closely connected modules, systematically achieves penetrating analysis of the physical essence of industrial process failures, resulting in accurate, rapid, and automated root cause diagnosis.

[0054] Although the invention has been described by way of examples, the invention is not limited to the above-described examples, and various modifications and variations can be made thereto without departing from the scope of the invention.

Claims

1. A method for diagnosing the root causes of industrial anomalies based on decoupled causal representation, characterized in that, Includes the following steps: Collect multivariate time-series observation data of industrial processes; A causal decoupled variational autoencoder model is constructed and trained, which maps the observed data to a low-dimensional latent space and obtains a set of decoupled latent causal factor representations. Based on the trained model, online data is monitored, and when an anomaly is detected, the variance fraction of the posterior distribution of each potential causal factor relative to the normal baseline distribution is calculated. Based on the magnitude of the variance fraction, the fundamental causal factor leading to the anomaly is located.

2. The method according to claim 1, characterized in that, The training loss function of the causal decoupled variational autoencoder model includes a total correlation penalty term to minimize the dependencies between potential causal factors.

3. The method according to claim 2, characterized in that, The total relevance penalty term is estimated using a mini-batch weighted sampling strategy.

4. The method according to claim 1, characterized in that, The causal decoupling variational autoencoder model incorporates temporal causal constraints, which are implemented through a mask matrix based on prior causal knowledge to constrain the evolution of potential causal factors over time.

5. The method according to claim 1, characterized in that, The variance fraction is calculated using any one of the statistical distance measures: Wasserstein distance, KL divergence, or JS divergence.

6. The method according to claim 1, characterized in that, After locating the fundamental causal factor, the next step includes tracing back the causal parent node of that factor to analyze the abnormal propagation path.

7. A system for implementing the method of any one of claims 1-6, characterized in that, include: The data preprocessing module is used to clean and standardize the raw time-series data; The model training module is used to build and train the causal decoupling variational autoencoder model. The online monitoring and tracing module is used to detect anomalies in real time and execute root cause tracing algorithms; The results output module is used to generate and display root cause diagnosis reports.

Images