Industrial time-series event analysis methods, equipment, and media based on causality regularization

Abstract

This application discloses a method, device, and medium for industrial time-series event analysis based on causal regularization, relating to the field of artificial intelligence technology. The method includes: multi-scale temporal feature encoding of multivariate time-series data to obtain temporal feature data; latent vector reconstruction of the probability distribution of the temporal feature data to determine reconstruction loss and distance metric loss; obtaining a causal consistency loss function based on a prior causal graph through acyclic constraint balance analysis; performing hyperparameter consistency processing on the causal consistency loss function to determine the overall loss function and obtain a minimized overall loss; fine-tuning the model until convergence using a small number of samples to determine the implicitly labeled CT-VAE model; and inputting the data window into the model to obtain event analysis results and event diagnostic reports. This application solves the technical problems of poor robustness, high identification error, and inability to meet the implicit critical event identification requirements in existing industrial time-series event analysis methods.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, device and medium for industrial time-series event analysis based on causal regularization. Background Technology

[0002] In modern industrial systems, such as mining, power generation, and chemical manufacturing, production processes are continuously monitored by numerous sensors, generating massive amounts of high-dimensional, multi-dimensional time-series data. Accurately and promptly identifying critical events from this complex multi-dimensional time-series data is crucial for ensuring production safety and improving operational efficiency. Critical events include both planned routine operations (such as the "shutdown" switching of a mine's main ventilation fan) and various unplanned equipment failures or abnormal operating conditions.

[0003] Accurate identification of critical events is rarely achievable through a single sensor; it requires analyzing complex, synchronously evolving patterns among multiple variables. Firstly, the relationships between variables are not simple statistical correlations but rather involve physical causal chains. Traditional time series models often overlook this deep causal structure, resulting in insufficient construction of multidimensional time series data and causal dependencies, leading to extremely poor model robustness. Secondly, critical production events, such as fan failures or unexpected malfunctions within the production plan, have extremely low frequencies relative to the total system uptime. The sparsity and class imbalance of critical production events cause standard supervised learning models to heavily favor predicting all situations as normal operation, increasing the difficulty of accurately identifying abnormal events. Thirdly, existing historical maintenance records and critical event logs may contain omissions or errors, leading to noisy and incomplete labels for abnormal events. This label uncertainty makes methods relying entirely on supervised learning infeasible. Fourth, for some critical events, such as unsuccessful reversing, they are defined by the failure of a series of expected actions. These critical events are the absence or deviation of the expected pattern, rather than a simple numerical limit exceeding, i.e., there are no explicit labels (implicit) critical events. Existing technologies cannot meet the requirements for identifying implicit critical events. Summary of the Invention

[0004] This application provides an industrial time-series event analysis method, device, and medium based on causal regularization, which solves the technical problems of poor robustness, high identification error, reliance on supervised learning methods, and inability to meet the requirements of implicit key event identification in the prior art.

[0005] In a first aspect, embodiments of this application provide an industrial time-series event analysis method based on causal regularization. The method comprises: acquiring multivariate time-series data and performing multi-scale time-series feature encoding on the multivariate time-series data to obtain time-series feature data; reconstructing the latent vectors of the probability distribution of the time-series feature data to determine the reconstruction loss and distance metric loss; constructing a prior causal graph and, based on the prior causal graph, obtaining a causal consistency loss function through acyclic constraint balance analysis; performing hyperparameter consistency processing on the causal consistency loss function according to the reconstruction loss and distance metric loss to determine the overall loss function, and performing unsupervised training on the overall loss function to obtain a minimized overall loss; acquiring latent representation data and, based on the latent representation data and the CT-VAE model corresponding to minimizing the overall loss, fine-tuning it with a small number of samples until the model converges to determine an implicitly labeled CT-VAE model; inputting a data window into the implicitly labeled CT-VAE model and performing event diagnosis inference on the data window to obtain event analysis results and an event diagnosis report.

[0006] In one implementation of this application, multi-scale temporal feature encoding is performed on multivariate time series data to obtain temporal feature data. Specifically, this includes: extracting information from the multivariate time series data to obtain dense information temporal feature data; performing short-scale causal dilation convolution on the dense information temporal feature data to determine transient temporal feature data; performing long-scale causal dilation convolution on the dense information temporal feature data to determine persistent temporal feature data; and obtaining temporal feature data based on the transient and persistent temporal feature data.

[0007] In one implementation of this application, latent vector reconstruction of a probability distribution is performed on time-series feature data to determine reconstruction loss and distance metric loss. Specifically, this includes: predicting latent variables from the time-series feature data using a Gaussian distribution to obtain vectors corresponding to the latent variables; wherein the vectors corresponding to the latent variables include: a mean vector and a log-variance vector; reparameter sampling of the vectors corresponding to the latent variables to determine the latent vectors; reconstructing the latent vectors into an original input data window to obtain reconstructed time-series feature data; calculating the reconstruction loss based on the time-series feature data and the reconstructed time-series feature data using a mean squared error function; and determining the distance metric loss based on the time-series feature data and the reconstructed time-series feature data through distribution difference analysis.

[0008] In one implementation of this application, a causal consistency loss function is obtained based on a prior causal graph through acyclic constraint balance analysis. Specifically, this includes: encoding the prior causal graph with knowledge to determine its mandatory edges; obtaining an updated documented causal graph based on the mandatory edges of the prior causal graph through temporal causal discovery analysis; integrating the prior causal graph and the updated documented causal graph to determine the overall causal graph; obtaining a difference loss term based on the overall causal graph through implicit causal graph difference analysis; and balancing the difference loss term with acyclic constraints to obtain the causal consistency loss function.

[0009] In one implementation of this application, the causal consistency loss function is subjected to hyperparameter consistency processing based on the reconstruction loss and the distance metric loss to determine the overall loss function. Specifically, this includes: determining a first hyperparameter based on the distance metric loss; determining a second hyperparameter based on the causal consistency loss function; and performing consistency configuration on the first and second hyperparameters to determine the overall loss function.

[0010] In one implementation of this application, based on the latent representation data and the CT-VAE model corresponding to minimizing the overall loss, the implicitly labeled CT-VAE model is determined by fine-tuning it with a small number of samples until the model converges. Specifically, this includes: performing confidence sample discrimination on the latent representation data to determine a high-confidence event sample set; wherein, confidence sample discrimination includes: sample discrimination, confidence discrimination, and sample confidence threshold filtering; based on the high-confidence event sample set, anomaly clusters are obtained by inferring anomalous event data clusters in the latent space; and based on the anomaly clusters, the CT-VAE model is trained until the model converges to determine the implicitly labeled CT-VAE model.

[0011] In one implementation of this application, event diagnosis reasoning is performed on a data window to obtain event analysis results and an event diagnosis report. Specifically, this includes: calculating a reconstruction score for the data window to obtain a reconstruction error coefficient; calculating a potential deviation for the data window to determine the deviation distance between the potential vector corresponding to the data window and the center of the normal data distribution; calculating a causal violation score for the data window to obtain the score of the causal consistency loss term on the data window; performing a weighted sum of the reconstruction error coefficient, deviation distance, and the score of the causal consistency loss term on the data window to determine a total anomaly score; and obtaining event analysis results and an event diagnosis report based on the total anomaly score through event detection and classification.

[0012] In one implementation of this application, after inputting the data window into the implicitly labeled CT-VAE model and performing event diagnosis reasoning on the data window to obtain event analysis results and event diagnosis reports, the method further includes: identifying unknown events in the event analysis results and marking the identified unknown events; performing latent spatial trajectory visualization on the event diagnosis reports to obtain causal graph deviation analysis to determine the visualized event diagnosis reports.

[0013] Secondly, embodiments of this application also provide an industrial time-series event analysis method and apparatus based on causal regularization, characterized in that the apparatus includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to: acquire multivariate time-series data and perform multi-scale time-series feature encoding on the multivariate time-series data to obtain time-series feature data; perform latent vector reconstruction of the probability distribution of the time-series feature data to determine reconstruction loss and distance metric loss; construct a prior causal graph, and based on the prior causal graph... Through acyclic constraint balance analysis, a causal consistency loss function is obtained. Based on the reconstruction loss and distance metric loss, hyperparameter consistency processing is performed on the causal consistency loss function to determine the overall loss function. Unsupervised training is then performed on the overall loss function to minimize the overall loss. Latent representation data is acquired, and based on the latent representation data and the CT-VAE model corresponding to minimizing the overall loss, the model is fine-tuned with a small number of samples until convergence, thus determining the implicitly labeled CT-VAE model. The data window is input into the implicitly labeled CT-VAE model, and event diagnosis inference is performed on the data window to obtain event analysis results and event diagnosis reports.

[0014] Thirdly, embodiments of this application also provide a non-volatile computer storage medium storing computer-executable instructions for an industrial time-series event analysis method based on causal regularization. The computer-executable instructions are characterized by: acquiring multivariate time-series data and performing multi-scale time-series feature encoding on the multivariate time-series data to obtain time-series feature data; reconstructing the latent vectors of the probability distribution of the time-series feature data to determine the reconstruction loss and distance metric loss; constructing a prior causal graph and, based on the prior causal graph, obtaining a causal consistency loss function through acyclic constraint balance analysis; performing hyperparameter consistency processing on the causal consistency loss function according to the reconstruction loss and distance metric loss to determine the overall loss function, and performing unsupervised training on the overall loss function to obtain a minimized overall loss; acquiring latent representation data and, based on the latent representation data and the CT-VAE model corresponding to minimizing the overall loss, fine-tuning it with a small number of samples until the model converges to determine the implicitly labeled CT-VAE model; inputting a data window into the implicitly labeled CT-VAE model and performing event diagnosis inference on the data window to obtain event analysis results and an event diagnosis report.

[0015] This application provides an industrial time-series event analysis method, device, and medium based on causal regularization. Through multi-scale temporal feature encoding, probabilistic latent space analysis and decoding, causal graph regularization of prior causal graphs, feature training using the CT-VAE model, and event diagnostic reasoning based on causal graph bias analysis, it solves the technical problems of poor robustness, high identification error, reliance on supervised learning methods, and inability to meet the requirements of implicit key event identification in existing industrial time-series event analysis. It achieves accurate identification of key production events, improves the interpretability and generalization ability of the industrial time-series event analysis model, reduces the false alarm rate and false negative rate of industrial anomalies, and optimizes the classification and detection effect of known time and location time. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0017] Figure 1 A flowchart of an industrial time-series event analysis method based on causal regularization is provided for embodiments of this application;

[0018] Figure 2 An architecture diagram of an industrial time-series event analysis system based on causal regularization is provided for embodiments of this application;

[0019] Figure 3 A diagram illustrating the architecture of a causal time-series event mining model provided in this application embodiment;

[0020] Figure 4 A flowchart of a CT-VAE model training process is provided for an embodiment of this application;

[0021] Figure 5 A flowchart of event reasoning and diagnosis provided in an embodiment of this application;

[0022] Figure 6 A reversal event recognition diagram provided in an embodiment of this application;

[0023] Figure 7 A continuous failure event recognition diagram provided in an embodiment of this application;

[0024] Figure 8 A reverse-failure event recognition diagram provided in this application embodiment;

[0025] Figure 9 This is a schematic diagram of the internal structure of an industrial time-series event analysis device based on causal regularization, provided as an embodiment of this application. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0027] This application provides an industrial time-series event analysis method, device, and medium based on causal regularization. Through multi-scale temporal feature encoding, probabilistic latent space analysis and decoding, causal graph regularization of prior causal graphs, feature training using the CT-VAE model, and event diagnostic reasoning based on causal graph bias analysis, it solves the technical problems of poor robustness, high identification error, reliance on supervised learning methods, and inability to meet the requirements of implicit key event identification in existing industrial time-series event analysis. It achieves accurate identification of key production events, improves the interpretability and generalization ability of the industrial time-series event analysis model, reduces the false alarm rate and false negative rate of industrial anomalies, and optimizes the classification and detection effect of known time and location time.

[0028] The technical solutions proposed in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0029] Figure 1 A flowchart illustrating an industrial time-series event analysis method based on causal regularization, provided for embodiments of this application. Figure 1As shown in the figure, the industrial time-series event analysis method based on causal regularization provided in this application embodiment specifically includes the following steps:

[0030] Step 101: Obtain multivariate time series data and encode the multivariate time series data using multi-scale time series features to obtain time series feature data.

[0031] For example, the encoder's function is to extract data from the input multivariate time series data window. Extract an information-dense temporal feature embedding To simultaneously capture event features at different time scales, this invention preferably employs a multi-scale temporal convolutional network (TCN) as the main structure of the encoder. Causal convolution ensures that at the prediction time point... When outputting, only used The input at or before the specified time point aligns with the unidirectional nature of time series. The core of TCN is Dilated Causal Convolution. Dilated Causal Convolution incorporates dilated convolution into causal convolution, exponentially expanding the receptive field by inserting holes between elements of the convolution kernel. This allows it to capture very long-term dependencies with fewer layers, while avoiding the vanishing / exploding gradient problem common in recurrent neural networks (RNNs). Furthermore, the convolution operations can be computed in parallel, improving model training efficiency.

[0032] Specifically, multi-scale temporal feature encoding is performed on multivariate time series data to obtain temporal feature data, including: extracting information from multivariate time series data to obtain dense information temporal feature data; performing short-scale causal dilation convolution on dense information temporal feature data to determine transient temporal feature data; performing long-scale causal dilation convolution on dense information temporal feature data to determine persistent temporal feature data; and obtaining temporal feature data based on transient and persistent temporal feature data.

[0033] Figure 2 This document presents an architecture diagram of an industrial time-series event analysis system based on causal regularization, as provided in an embodiment of this application.

[0034] In one embodiment, the data acquisition and preprocessing module is responsible for acquiring multivariate time-series data in real time or in batches from multiple sensors (measuring power, current, wind speed, pressure, vibration, temperature, etc.) at the industrial site. The raw data is then preprocessed by cleaning, alignment, missing value handling, and normalization. A sliding window technique is used to divide the continuous data stream into fixed-length, overlapping data windows, which serve as input samples for subsequent models.

[0035] The causal temporal model engine module is responsible for building, training, and storing the causal-temporal variational autoencoder (CT-VAE) model proposed in this invention.

[0036] The event reasoning and diagnostic presentation module is responsible for calling the trained CT-VAE model to perform real-time analysis of new data windows. When an event is detected, this module not only issues an alert but also generates a multi-dimensional diagnostic report, which is presented to operators or engineers through the user interface to assist them in root cause analysis.

[0037] Multi-scale design is achieved by connecting multiple TCN modules with different expansion factor sequences in parallel:

[0038] Short-scale TCN modules, using smaller convolution kernels and smaller dilation rates (e.g., d=1, 2, 4), focus on capturing fast, transient event characteristics, such as current spikes at the moment of motor startup or vibration shocks at the initial stage of equipment failure.

[0039] Long-scale TCN modules, using larger convolutional kernels and larger dilation rates (e.g., d=8, 16, 32), possess a wider receptive field, enabling them to capture the overall contours of events lasting longer, such as a complete shutdown process lasting several minutes. The output features of all scale TCN modules are finally concatenated or summed to form the final temporal feature embedding. This provides high-quality input for subsequent probabilistic modeling and causal analysis.

[0040] Step 102: Reconstruct the latent vector of the probability distribution of the time series feature data to determine the reconstruction loss and distance metric loss.

[0041] For example, a variational autoencoder (VAE) framework is used to learn the probability distribution of the system's normal operating state. Unlike a standard autoencoder, which deterministically maps the input to a point, a VAE maps the input to a probability distribution, making it more robust to input noise and able to generate new, normal-pattern-compliant data by sampling from the latent space.

[0042] Specifically, the process involves reconstructing the latent vectors of the time-series feature data using a probability distribution to determine the reconstruction loss and distance metric loss. This includes: predicting latent variables from the time-series feature data using a Gaussian distribution to obtain vectors corresponding to the latent variables; where the vectors corresponding to the latent variables include: a mean vector and a log-variance vector; reparameterizing the vectors corresponding to the latent variables to determine the latent vectors; reconstructing the latent vectors into the original input data window to obtain reconstructed time-series feature data; calculating the reconstruction loss using a mean squared error function based on the time-series feature data and the reconstructed time-series feature data; and determining the distance metric loss through distribution difference analysis based on the time-series feature data and the reconstructed time-series feature data.

[0043] In one embodiment, firstly, the temporal features output by the TCN encoder are embedded... The data is fed into two separate fully connected layers (or 1x1 convolutional layers) to predict latent variables. The mean vector of the Gaussian distribution it follows Sum of logarithmic variance vector .

[0044] To make the sampling process differentiable, thus enabling end-to-end gradient backpropagation training, this invention employs a reparameterization technique. Latent vector Through formula Sampling was performed, among which From the standard normal distribution Random noise in the mid-sample, This indicates element-wise multiplication. For distribution parameters, The difference in labeling can optimize gradient descent.

[0045] Then decoding is performed, receiving the latent vector through a decoder whose structure is symmetric to that of the TCN encoder. And attempt to reconstruct the original input data window. Decoders typically use transposed convolutions to achieve upsampling.

[0046] Finally, for the calculation of the mean squared error function and the analysis of distribution differences, the basic loss function of the model is the standard VAE loss, and the loss function is composed as follows:

[0047] Reconstruction loss Measure the raw input With reconstruction output The difference between them can be expressed using the mean squared error (MSE) function.

[0048] Distance metric loss Measure the potential distribution generated by the encoder Compared with the standard normal prior distribution The differences between them.

[0049] It is important to note the distance metric loss. This term is a regularization term, which promotes the structuring of the latent space. The loss function can be chosen as the KL divergence.

[0050] Step 103: Construct a prior cause-effect graph, and based on the prior cause-effect graph, obtain the causal consistency loss function through acyclic constraint balance analysis.

[0051] For example, by using a differentiable causal consistency loss function, the model is forced to learn data patterns while its internal feature representations must conform to a pre-defined causal structure that represents physical laws. This ensures that the feature representations learned by the model are not only statistically effective but also physically meaningful, thereby directly injecting domain knowledge into the model architecture. It also solves the technical problem that traditional deep learning models operate as "black boxes," with opaque decision-making logic and an inability to provide explanations that conform to physical intuition.

[0052] Specifically, based on the prior causal graph, a causal consistency loss function is obtained through acyclic constraint balance analysis, including: knowledge encoding of the prior causal graph to determine the mandatory edges of the prior causal graph; obtaining an updated documented causal graph based on the mandatory edges of the prior causal graph through temporal causal discovery analysis; integrating the prior causal graph and the updated documented causal graph to determine the overall causal graph; obtaining the difference loss term through implicit causal graph difference analysis based on the overall causal graph; and balancing the difference loss term with acyclic constraints to obtain the causal consistency loss function.

[0053] Figure 3 This is a diagram illustrating the architecture of a causal time-series event mining model provided in an embodiment of this application.

[0054] In one embodiment, firstly, a priori cause-effect graph is constructed. Before model training begins, a directed acyclic graph (DAG) needs to be constructed. , where the node set The edge set corresponds to each variable (i.e., sensor) in a multivariate time series. This represents the direct causal relationship between variables. The construction of the prior cause-effect graph adopts a "human-machine collaboration" model, integrating two sources of knowledge:

[0055] First, deterministic knowledge from domain experts is encoded as mandatory edges in the graph. This knowledge can originate from physical laws (e.g., "increased motor power" leads to "increased fan speed"), equipment design (e.g., "vibration of fan A" is related to "power of fan A"), or operating procedures (e.g., "opening the standby fan damper" is a prerequisite for "starting the standby fan"). These edges have the highest confidence (equivalent to basic knowledge).

[0056] Secondly, time-series causal discovery algorithms (such as DYNOTEARS) are applied to the training dataset to uncover potential, possibly undocumented causal relationships and discover hidden association patterns in the data that experts may overlook.

[0057] The graphs from the two sources are integrated to form the final result. During the integration process, when expert knowledge conflicts with the results of data discovery, the expert knowledge prevails to ensure the accuracy and physical plausibility of the prior cause-effect graph.

[0058] Then, define the causal consistency loss function. Assume the first layer weight matrix of the TCN encoder is... An implicit linear mapping is learned that should approximate the linear structure equation model of the monitored system. Therefore, it is possible to... absolute value Extract an adjacency matrix Its elements Represents the variables learned by the model For variables The strength of the causal influence.

[0059] Introducing a new loss term This is used to penalize the implicit causal graph learned by the model (composed of the adjacency matrix). (Representative) and Prior Cause-and-Effect Graph The structural differences between them. Utilizing the ideas of algorithms such as NOTEARS, the acyclicity constraint of a graph is expressed as a differentiable function. .

[0060] in, It is the number of variables. It is the trace of the matrix.

[0061] The following formula will be used to explain this.

[0062] (1)

[0063] in, a priori causal graph edge set, This is an indicator function.

[0064] It is important to note that the first item The penalty for loops in the graph is the second term, which is a sparsity penalty. Its main components are those in the adjacency matrix. It exists in but does not exist in The edges in the chain (i.e., abnormal causal chains); This is a hyperparameter used to balance acyclic constraints and structural consistency with the prior graph.

[0065] Step 104: Based on the reconstruction loss and distance metric loss, perform hyperparameter consistency processing on the causal consistency loss function to determine the overall loss function, and perform unsupervised training on the overall loss function to obtain the minimized overall loss.

[0066] For example, in order to enable the causal graph regularized model to both reconstruct normal data and reproduce solutions of prior causal knowledge in its internal structure, this application improves the generalization ability of the model and the interpretability of its latent space through hyperparameter consistency processing.

[0067] Specifically, based on the reconstruction loss and the distance metric loss, the causal consistency loss function is subjected to hyperparameter consistency processing to determine the overall loss function, including: determining the first hyperparameter based on the distance metric loss; determining the second hyperparameter based on the causal consistency loss function; and performing consistency configuration on the first and second hyperparameters to determine the overall loss function.

[0068] In one embodiment, a model overall loss function is constructed, which is interpreted using the following formula.

[0069] (2)

[0070] in, and It is a hyperparameter used to balance reconstruction accuracy, the smoothness of the latent space, and consistency with physical knowledge.

[0071] By minimizing the total loss function, the model is guided to find a solution that can both reconstruct normal data well and reproduce prior causal knowledge in its internal structure. This transforms the model from a simple "pattern recognizer" into a simplified, data-driven "physical simulator." The latent space learned by the model is thus more interpretable, as different latent dimensions may naturally correspond to different physical subsystems, achieving meaningful feature decoupling.

[0072] Step 105: Obtain latent representation data, and based on the latent representation data and the CT-VAE model corresponding to minimizing the overall loss, fine-tune the model with a small number of samples until it converges, and determine the implicitly labeled CT-VAE model.

[0073] For example, existing historical maintenance records and critical event logs may contain omissions or errors. For instance, an operational failure such as "unsuccessful reversal" is unlikely to be formally recorded. This uncertainty in labeling makes fully supervised learning methods infeasible, necessitating the use of weakly supervised or unsupervised paradigms capable of handling imperfect labels. Due to the unreliability of historical event labels, this application uses a CT-VAE model, employing a progressive strategy from unsupervised to weakly supervised learning. By fine-tuning the model with a small number of samples until convergence, the implicit labeling of the CT-VAE model is determined, solving the problems of label noise and event sparsity commonly found in industrial data.

[0074] Specifically, based on the latent representation data and the CT-VAE model corresponding to minimizing the overall loss, the implicitly labeled CT-VAE model is determined by fine-tuning it with a small number of samples until the model converges. This includes: performing confidence sample discrimination on the latent representation data to determine the high-confidence event sample set; wherein, confidence sample discrimination includes: sample discrimination, confidence discrimination, and sample confidence threshold filtering; based on the high-confidence event sample set, abnormal clusters are obtained by inferring abnormal event data clusters in the latent space; based on the abnormal cluster clusters, the CT-VAE model is trained until the model converges, thus determining the implicitly labeled CT-VAE model.

[0075] Figure 4 A flowchart of a CT-VAE model training process is provided for an embodiment of this application.

[0076] In one embodiment, it is assumed that the majority of data in the dataset represents the normal operating state of the system. Under this assumption, the CT-VAE model is trained unsupervised on the entire training set. The training objective is to minimize the total loss function defined in Part II. .

[0077] The purpose of unsupervised training is to enable the model to first learn the basic dynamic characteristics and inherent physical causal structure of the system under normal operating conditions, build a deep understanding of "what is normal", and lay the foundation for subsequent event recognition.

[0078] Incomplete and noisy historical event logs in industrial scenarios are considered a "weak label set". In order to filter out truly valuable high-confidence samples from this weak label set, this application adopts a filtering mechanism inspired by adversarial training to filter high-confidence samples.

[0079] First, a simple discriminator network is trained, whose task is to distinguish the latent representations of two classes of samples. One type is the potential representation from a large amount of unlabeled data (which is likely normal data), and the other type is the potential representation from log events.

[0080] During training, samples that are labeled as events but whose underlying representation z is difficult for the discriminator to distinguish from normal data are likely to be normal operating conditions that have been mislabeled, or minor anomalies with very inconspicuous features. These samples have low confidence.

[0081] Then, the output of the discriminator is used to evaluate the degree of anomaly or confidence of each log event sample, and those samples with scores below a certain threshold are filtered out, resulting in a small but high-quality set of high-confidence event samples.

[0082] Finally, the pre-trained CT-VAE model is fine-tuned based on a high-confidence event sample set. The training objective is to add an additional discriminative or comparative loss term to the original loss function.

[0083] This loss term originates from the idea of ​​Open-Set Recognition, which aims to push the representations of known, labeled anomalous events (such as "system crash") away from the center (usually the origin) of the normal data cluster in the latent space.

[0084] Furthermore, a center loss or contrastive loss can be added to penalize the latent representation of known outliers for being too close to the center of normal data clusters.

[0085] After fine-tuning, the model can not only detect "abnormalities" (i.e. deviations from normal clusters), but also clearly "identify" specific, known event types that form their own unique clusters in the latent space that are far from normal clusters.

[0086] When the model discovers new or unknown anomalies in subsequent applications, these new samples can be added to the high-confidence event set after being confirmed and labeled by human experts, for use in the next round of model iteration and fine-tuning.

[0087] Step 106: Input the data window into the implicitly labeled CT-VAE model, and perform event diagnosis reasoning on the data window to obtain event analysis results and event diagnosis reports.

[0088] For example, existing technologies rely on a single reconstruction error for event diagnosis reasoning, which cannot accurately describe the event state. This application uses a combination of three complementary parts to form an anomaly score for event diagnosis reasoning, which realizes the visualization of multi-dimensional event states and improves the richness of system state description.

[0089] Specifically, event diagnostic reasoning is performed on the data window to obtain event analysis results and event diagnostic reports. This includes: calculating a reconstruction score for the data window to obtain a reconstruction error coefficient; calculating a potential deviation for the data window to determine the deviation distance between the potential vector corresponding to the data window and the center of the normal data distribution; calculating a causal violation score for the data window to obtain the score of the causal consistency loss term on the data window; performing a weighted summation of the reconstruction error coefficient, deviation distance, and the score of the causal consistency loss term on the data window to determine the total anomaly score; and obtaining event analysis results and event diagnostic reports based on the total anomaly score through event detection and classification.

[0090] Furthermore, after inputting the data window into the implicitly labeled CT-VAE model and performing event diagnosis reasoning on the data window to obtain event analysis results and event diagnosis reports, the method also includes: identifying unknown events in the event analysis results and marking the identified unknown events; performing latent spatial trajectory visualization on the event diagnosis reports to obtain causal graph deviation analysis to determine the visualized event diagnosis reports.

[0091] Figure 5 This is a flowchart of event reasoning and diagnosis provided for an embodiment of this application.

[0092] Figure 6 This is a diagram for identifying a system reversal event, provided as an embodiment of this application.

[0093] Figure 7 This is a continuous failure event recognition diagram provided in an embodiment of this application.

[0094] Figure 8 This is a diagram illustrating a reversing failure event, provided as an embodiment of this application.

[0095] In one embodiment, after a trained CT-VAE model is deployed, for each new input data window... Perform the following reasoning and diagnostic steps:

[0096] First, reconstruct the fraction. Calculate the reconstruction error. High reconstruction error indicates that the current data pattern is not seen in normal data and is a traditional anomalous signal.

[0097] Potential bias Calculate the latent vector corresponding to the current data window. The distance from the center of the normal data distribution (usually the origin). The greater the distance, the more the system state deviates from the normal at the level of abstract understanding in the model.

[0098] Causal violation of fractions Calculation, i.e., causal consistency loss term The value in the current data window. A high score indicates that the system's current behavior violates known physical causal laws, and can capture some cases where the reconstruction error may not be obvious but is physically and logically incorrect.

[0099] The three scores mentioned above can be weighted and summed to obtain a total anomaly score. Combinations of different scores can provide more nuanced insights; for example, high... But low This could represent a very subtle, yet physically illogical, early failure.

[0100] Then, when the total anomaly score exceeds a preset threshold adjusted using the validation set, the system determines that an event has occurred. Subsequently, the location of this event in the latent space is analyzed. To classify them.

[0101] For known event classifications, if If a location is close to the cluster center of a known event category (such as "unsuccessful reversal") learned during the fine-tuning phase, it is classified as that specific event.

[0102] For the identification of unknown events, if If a data cluster is located significantly away from the center of a normal data cluster but not close to any known event cluster, it is marked as a "novel or unknown anomaly" and a suggestion is made that human experts need to be involved in the analysis.

[0103] For generating diagnostic reports, the system first performs reconstruction error analysis. By visualizing the point-by-point reconstruction error of each sensor channel, the system can immediately highlight the variables that deviate the most from the normal mode, providing the first clue for problem localization.

[0104] Plot the latent vectors representing the system state during the event. By observing the shape of the trajectory in a two-dimensional or three-dimensional reduced space (such as sharp deviation, slow drift, or oscillation between different states), the dynamic characteristics of the event can be determined and compared with the characteristic trajectories of known events.

[0105] Then, feature attribution methods such as SHAP (SHapley Additive exPlanations) or input gradients are used to analyze the sources of the total outlier score.

[0106] Finally, perform causal graph bias analysis to analyze the dynamic causal relationships (i.e., adjacency matrix) inferred from the model within the event occurrence window. Prior causal graph of the baseline The comparison is made and significant deviations are highlighted to visualize the event diagnostic report.

[0107] Furthermore, this deviation includes: broken links, where the expected causal relationship fails to materialize. For example, the diagnostic report indicates that the power of fan B has reached its rated value, but this does not lead to the expected increase in airflow; and spurious links, where new, unexpected causal relationships emerge, such as ambient temperature beginning to significantly affect bearing temperature, a relationship that does not exist under normal operating conditions.

[0108] The dashed lines indicate that the fan equipment is not running, the solid lines indicate that it is running, and the gray areas indicate the detected parts.

[0109] The above are embodiments of the method proposed in this application. Based on the same inventive concept, embodiments of this application also provide an industrial time-series event analysis device based on causal regularization, the structure of which is as follows: Figure 9 As shown.

[0110] Figure 9 This is a schematic diagram of the internal structure of an industrial time-series event analysis device based on causal regularization, provided as an embodiment of this application. Figure 9 As shown, the device includes:

[0111] At least one processor 901;

[0112] And a memory 902 that is communicatively connected to at least one processor;

[0113] The memory 902 stores instructions executable by at least one processor, which are executed by at least one processor 901 to enable at least one processor 901 to:

[0114] This process involves acquiring multivariate time series data and performing multi-scale temporal feature encoding to obtain temporal feature data. The temporal feature data undergoes latent vector reconstruction based on probability distributions to determine reconstruction loss and distance metric loss. A prior causal graph is constructed, and based on this graph, a causal consistency loss function is obtained through acyclic constraint balance analysis. Hyperparameter consistency processing is applied to the causal consistency loss function based on the reconstruction loss and distance metric loss to determine the overall loss function. Unsupervised training is then performed on the overall loss function to minimize the overall loss. Latent representation data is acquired, and based on the latent representation data and the CT-VAE model corresponding to minimizing the overall loss, the model is fine-tuned with a small number of samples until convergence, thus determining the implicitly labeled CT-VAE model. A data window is input into the implicitly labeled CT-VAE model, and event diagnosis inference is performed on the data window to obtain event analysis results and event diagnosis reports.

[0115] Some embodiments of this application provide corresponding to Figure 1 A non-volatile computer storage medium based on causal regularization for industrial time-series event analysis, storing computer-executable instructions, wherein the computer-executable instructions are configured as follows:

[0116] This process involves acquiring multivariate time series data and performing multi-scale temporal feature encoding to obtain temporal feature data. The temporal feature data undergoes latent vector reconstruction based on probability distributions to determine reconstruction loss and distance metric loss. A prior causal graph is constructed, and based on this graph, a causal consistency loss function is obtained through acyclic constraint balance analysis. Hyperparameter consistency processing is applied to the causal consistency loss function based on the reconstruction loss and distance metric loss to determine the overall loss function. Unsupervised training is then performed on the overall loss function to minimize the overall loss. Latent representation data is acquired, and based on the latent representation data and the CT-VAE model corresponding to minimizing the overall loss, the model is fine-tuned with a small number of samples until convergence, thus determining the implicitly labeled CT-VAE model. A data window is input into the implicitly labeled CT-VAE model, and event diagnosis inference is performed on the data window to obtain event analysis results and event diagnosis reports.

[0117] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for IoT devices and media are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0118] The systems, media, and methods provided in this application are one-to-one correspondences. Therefore, the systems and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the systems and media will not be repeated here.

[0119] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0120] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0121] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0122] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0123] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0124] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0125] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0126] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0127] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for analyzing industrial time-series events based on causal regularization, characterized in that, The method includes: Multivariate time-series data are acquired from multiple sensors in the industrial field, and multi-scale time-series feature encoding is performed on the multivariate time-series data to obtain time-series feature data; wherein, the multiple sensors are used to measure power, current, wind speed, pressure, vibration, and temperature; The temporal feature data is reconstructed using a potential vector of probability distribution to determine the reconstruction loss and distance metric loss. A prior cause-effect graph is constructed, and based on the prior cause-effect graph, a causal consistency loss function is obtained through acyclic constraint balance analysis. Based on the reconstruction loss and the distance metric loss, the causal consistency loss function is subjected to hyperparameter consistency processing to determine the overall loss function, and the overall loss function is subjected to unsupervised training to obtain the minimized overall loss; Acquire latent representation data, and based on the latent representation data and the CT-VAE model corresponding to minimizing the overall loss, fine-tune the model with a small number of samples until it converges, and determine the implicitly labeled CT-VAE model; The data window is input into the implicitly labeled CT-VAE model, and event diagnosis reasoning is performed on the data window to obtain event analysis results and event diagnosis reports; Based on the aforementioned prior cause-effect graph, a causal consistency loss function is obtained through acyclic constraint balance analysis, specifically including: The prior causal graph is knowledge-encoded to determine the forced edges of the prior causal graph; Based on the forced edges of the prior causal graph, an updated documented causal graph is obtained through temporal causal discovery analysis. The prior cause-effect graph is integrated with the updated documented cause-effect graph to determine the overall cause-effect graph; Based on the overall cause-effect graph, the difference loss term is obtained through implicit cause-effect graph difference analysis; The difference loss term is balanced with an acyclicity constraint to obtain the causal consistency loss function; Based on the latent representation data and the CT-VAE model corresponding to minimizing the overall loss, the implicitly labeled CT-VAE model is determined by fine-tuning it with a small number of samples until the model converges. Specifically, this includes: The potential representation data is subjected to confidence sample discrimination to determine a high-confidence event sample set; wherein, the confidence sample discrimination includes: sample discrimination, confidence discrimination, and sample confidence threshold filtering; Based on the high-confidence event sample set, anomaly clusters are obtained by inferring from the anomalous event data clusters in the latent space; Based on the abnormal cluster clustering, the CT-VAE model is trained until the model converges, and the implicitly labeled CT-VAE model is determined.

2. The industrial time-series event analysis method based on causal regularization according to claim 1, characterized in that, Multi-scale time series feature encoding is performed on the multivariate time series data to obtain time series feature data, specifically including: Information is extracted from the multivariate time series data to obtain dense information time series feature data; Short-scale causal dilated convolution is performed on the dense information temporal feature data to determine transient temporal feature data; Long-scale causal dilated convolution is performed on the dense information temporal feature data to determine the continuous temporal feature data; The time series feature data is obtained based on the transient time series feature data and the continuous time series feature data.

3. The industrial time-series event analysis method based on causal regularization according to claim 1, characterized in that, The temporal feature data is reconstructed using a latent vector of probability distribution to determine the reconstruction loss and distance metric loss, specifically including: Gaussian distribution latent variable prediction is performed on the time series feature data to obtain the vector corresponding to the latent variable; wherein, the vector corresponding to the latent variable includes: mean vector and log-variance vector; Reparameter sampling is performed on the vectors corresponding to the latent variables to determine the latent vectors; The latent vectors are reconstructed into the original input data window to obtain reconstructed temporal feature data; Based on the aforementioned time-series feature data and the reconstructed time-series feature data, the reconstruction loss is determined by calculating using the mean square error function; Based on the aforementioned temporal feature data and reconstructed temporal feature data, the distance metric loss is determined through distribution difference analysis.

4. The industrial time-series event analysis method based on causal regularization according to claim 1, characterized in that, Based on the reconstruction loss and the distance metric loss, the causal consistency loss function undergoes hyperparameter consistency processing to determine the overall loss function, specifically including: Based on the distance metric loss, determine the first hyperparameter; The second hyperparameter is determined based on the causal consistency loss function. The first hyperparameter and the second hyperparameter are configured in a consistent manner to determine the overall loss function.

5. The industrial time-series event analysis method based on causal regularization according to claim 1, characterized in that, The data window is subjected to event diagnostic reasoning to obtain event analysis results and event diagnostic reports, specifically including: The data window is reconstructed to calculate a reconstruction score to obtain the reconstruction error coefficient. The potential deviation is calculated for the data window to determine the deviation distance between the potential vector corresponding to the data window and the center of the normal data distribution. A causal violation score is calculated on the data window to obtain the score of the causal consistency loss term on the data window; The scores of the reconstruction error coefficient, deviation distance, and causal consistency loss term on the data window are weighted and summed to determine the total anomaly score; Based on the total anomaly score, the event analysis results and the event diagnosis report are obtained through event detection and classification.

6. The industrial time-series event analysis method based on causal regularization according to claim 1, characterized in that, After inputting the data window into the implicitly labeled CT-VAE model and performing event diagnostic inference on the data window to obtain event analysis results and event diagnostic reports, the method further includes: The event analysis results are used to identify unknown events, and the identified unknown events are marked as events. The event diagnosis report is visualized using a potential spatial trajectory to obtain a causal graph bias analysis, which is used to determine the visualized event diagnosis report.

7. An industrial time-series event analysis method and equipment based on causal regularization, characterized in that, The device includes: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to: Multivariate time-series data are acquired from multiple sensors in the industrial field, and multi-scale time-series feature encoding is performed on the multivariate time-series data to obtain time-series feature data; wherein, the multiple sensors are used to measure power, current, wind speed, pressure, vibration, and temperature; The temporal feature data is reconstructed using a potential vector of probability distribution to determine the reconstruction loss and distance metric loss. A prior cause-effect graph is constructed, and based on the prior cause-effect graph, a causal consistency loss function is obtained through acyclic constraint balance analysis. Based on the reconstruction loss and the distance metric loss, the causal consistency loss function is subjected to hyperparameter consistency processing to determine the overall loss function, and the overall loss function is subjected to unsupervised training to obtain the minimized overall loss; Acquire latent representation data, and based on the latent representation data and the CT-VAE model corresponding to minimizing the overall loss, fine-tune the model with a small number of samples until it converges, and determine the implicitly labeled CT-VAE model; The data window is input into the implicitly labeled CT-VAE model, and event diagnosis reasoning is performed on the data window to obtain event analysis results and event diagnosis reports; Based on the aforementioned prior cause-effect graph, a causal consistency loss function is obtained through acyclic constraint balance analysis, specifically including: The prior causal graph is knowledge-encoded to determine the forced edges of the prior causal graph; Based on the forced edges of the prior causal graph, an updated documented causal graph is obtained through temporal causal discovery analysis. The prior cause-effect graph is integrated with the updated documented cause-effect graph to determine the overall cause-effect graph; Based on the overall cause-effect graph, the difference loss term is obtained through implicit cause-effect graph difference analysis; The difference loss term is balanced with an acyclicity constraint to obtain the causal consistency loss function; Based on the latent representation data and the CT-VAE model corresponding to minimizing the overall loss, the implicitly labeled CT-VAE model is determined by fine-tuning it with a small number of samples until the model converges. Specifically, this includes: The potential representation data is subjected to confidence sample discrimination to determine a high-confidence event sample set; wherein, the confidence sample discrimination includes: sample discrimination, confidence discrimination, and sample confidence threshold filtering; Based on the high-confidence event sample set, anomaly clusters are obtained by inferring from the anomalous event data clusters in the latent space; Based on the abnormal cluster clustering, the CT-VAE model is trained until the model converges, and the implicitly labeled CT-VAE model is determined.

8. A non-volatile computer storage medium based on a causal regularization-based industrial time-series event analysis method, storing computer-executable instructions, characterized in that, The computer-executable instructions are set as follows: Multivariate time-series data are acquired from multiple sensors in the industrial field, and multi-scale time-series feature encoding is performed on the multivariate time-series data to obtain time-series feature data; wherein, the multiple sensors are used to measure power, current, wind speed, pressure, vibration, and temperature; The temporal feature data is reconstructed using a potential vector of probability distribution to determine the reconstruction loss and distance metric loss. A prior cause-effect graph is constructed, and based on the prior cause-effect graph, a causal consistency loss function is obtained through acyclic constraint balance analysis. Based on the reconstruction loss and the distance metric loss, the causal consistency loss function is subjected to hyperparameter consistency processing to determine the overall loss function, and the overall loss function is subjected to unsupervised training to obtain the minimized overall loss; Acquire latent representation data, and based on the latent representation data and the CT-VAE model corresponding to minimizing the overall loss, fine-tune the model with a small number of samples until it converges, and determine the implicitly labeled CT-VAE model; The data window is input into the implicitly labeled CT-VAE model, and event diagnosis reasoning is performed on the data window to obtain event analysis results and event diagnosis reports; Based on the aforementioned prior cause-effect graph, a causal consistency loss function is obtained through acyclic constraint balance analysis, specifically including: The prior causal graph is knowledge-encoded to determine the forced edges of the prior causal graph; Based on the forced edges of the prior causal graph, an updated documented causal graph is obtained through temporal causal discovery analysis. The prior cause-effect graph is integrated with the updated documented cause-effect graph to determine the overall cause-effect graph; Based on the overall cause-effect graph, the difference loss term is obtained through implicit cause-effect graph difference analysis; The difference loss term is balanced with an acyclicity constraint to obtain the causal consistency loss function; Based on the latent representation data and the CT-VAE model corresponding to minimizing the overall loss, the implicitly labeled CT-VAE model is determined by fine-tuning it with a small number of samples until the model converges. Specifically, this includes: The potential representation data is subjected to confidence sample discrimination to determine a high-confidence event sample set; wherein, the confidence sample discrimination includes: sample discrimination, confidence discrimination, and sample confidence threshold filtering; Based on the high-confidence event sample set, anomaly clusters are obtained by inferring from the anomalous event data clusters in the latent space; Based on the abnormal cluster clustering, the CT-VAE model is trained until the model converges, and the implicitly labeled CT-VAE model is determined.

Images