Abnormal root cause tracing method, device and computer equipment based on multi-causal model integration

Abstract

The application relates to an abnormal root cause tracing method, device and computer equipment based on multi-causal model integration. The method is applied to an industrial equipment and comprises the following steps: acquiring multivariate time series data of the industrial equipment in an abnormal state; training at least three causal inference models with different prediction functions based on the multivariate time series data; determining a unified root cause score vector based on a root cause score vector of each trained causal inference model; and determining a root cause variable based on the unified root cause score vector. The method can improve the root cause identification efficiency and accuracy.

Description

Technical Field

[0001] This application relates to the field of analytical technology, and in particular to a method, apparatus and computer device for tracing the root causes of anomalies using a multi-causal model integration. Background Technology

[0002] In the wave of global manufacturing transformation towards intelligence and digitalization, manufacturing systems are increasingly characterized by interconnected equipment, data-driven processes, and intelligent decision-making. Traditional methods for diagnosing anomalies in manufacturing processes often rely on human experience or simple analysis based on data from a single sensor. These methods have limitations such as slow response, low positioning accuracy, and insufficient correlation analysis, making it difficult to adapt to the characteristics of complex manufacturing systems with multi-variable coupling and dynamic fluctuations. Especially in the field of high-precision manufacturing, small parameter deviations can be amplified step by step through the production chain, leading to product quality defects or unplanned equipment downtime. Traditional methods often fail to quickly locate the source of anomalies, resulting in increased production costs and losses in production efficiency.

[0003] This shows that traditional technologies still suffer from low efficiency and accuracy in root cause identification. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, and computer equipment for tracing the root causes of anomalies by integrating multiple causal models, which can improve the efficiency and accuracy of root cause identification, in order to address the above-mentioned technical problems.

[0005] Firstly, this application provides an anomaly root cause tracing method based on a multi-causal model integration, applied to industrial equipment. The anomaly root cause tracing method includes:

[0006] Acquire multivariate time-series data of the industrial equipment under abnormal conditions;

[0007] Based on the multivariate time series data, at least three causal inference models using different prediction functions are trained;

[0008] Based on the root cause score vector of each of the causal inference models after training, a unified root cause score vector is determined.

[0009] Based on the unified root cause score vector, the root cause variable is determined.

[0010] In one embodiment, the process of acquiring the multivariate time series data includes:

[0011] Acquire historical time-series data of the industrial equipment under abnormal conditions;

[0012] Variables are extracted from the historical time series data to obtain structured time series data;

[0013] The structured time series data is standardized to obtain the multivariate time series data.

[0014] In one embodiment, training at least three causal inference models employing different prediction functions based on the multivariate time-series data includes:

[0015] Based on the multivariate time series data, at least three causal inference models using different prediction functions are pre-trained to obtain multiple pre-trained causal inference models.

[0016] Based on the multivariate time series data and the root cause score vectors of the multiple pre-trained causal inference models, the multiple pre-trained causal inference models are iteratively and interactively optimized.

[0017] In one embodiment, each round of the iterative interaction optimization includes:

[0018] Aggregate the current root cause score vectors of multiple pre-trained causal inference models to obtain a current unified root cause score vector;

[0019] Based on the current unified root cause score vector, construct a root cause consistency constraint term;

[0020] The causal structure of the pre-trained causal inference model is optimized based on the loss function of each model and the root cause consistency constraint to minimize the prediction error.

[0021] In one embodiment, aggregating the current root cause score vectors of multiple pre-trained causal inference models to obtain a current unified root cause score vector includes:

[0022] Based on the current root cause score vectors of multiple pre-trained causal inference models, a first unified root cause score vector is determined.

[0023] The second unified root cause score vector is determined based on the sparse attention mechanism, the first unified root cause score vector, and the root cause score vector of each of the pre-trained causal inference models.

[0024] Based on the projection of the causal matrices of the multiple pre-trained causal inference models into the solution space corresponding to the second unified root cause score vector, the second unified root cause score vector is optimized to obtain the current unified root cause score vector.

[0025] In one embodiment, determining the second unified root cause score vector based on the sparse attention mechanism, the first unified root cause score vector, and the root cause score vector of each of the pre-trained causal inference models includes:

[0026] Based on the first unified root cause score vector, the root cause score vector of each of the pre-trained causal inference models, the nonlinear activation function, and the learnable linear mapping parameters, the root cause residual coefficient of each of the pre-trained causal inference models is calculated.

[0027] The root cause residual coefficients are subjected to sparse normalization to obtain the attention weights of each pre-trained causal inference model;

[0028] A second unified root cause score vector is determined based on the current root cause score vectors of multiple pre-trained causal inference models and the attention weights.

[0029] In one embodiment, optimizing the second unified root cause score vector based on the projection of the causal matrices of the plurality of pre-trained causal inference models onto the solution space corresponding to the second unified root cause score vector to obtain the current unified root cause score vector includes:

[0030] Construct the causal structure solution space based on the second unified root cause fraction vector;

[0031] Project the causal matrix of each pre-trained causal inference model onto the causal structure solution space to obtain the nearest neighbor mapping point of the causal matrix;

[0032] Based on the nearest neighbor mapping point corresponding to each of the pre-trained causal inference models, the second unified root cause score vector is updated in reverse to obtain the current unified root cause score vector.

[0033] Secondly, this application provides a multi-causal model integrated anomaly root cause tracing device for industrial equipment, wherein the multi-causal model integrated anomaly root cause tracing device includes:

[0034] The data collection module is used to acquire multivariate time-series data of the industrial equipment under abnormal conditions;

[0035] The model training module is used to train at least three causal inference models using different prediction functions based on the multivariate time series data.

[0036] The vector determination module is used to determine a unified root cause score vector based on the root cause score vector of each of the causal inference models after training.

[0037] The root cause determination module is used to determine the root cause variable based on the unified root cause score vector.

[0038] Thirdly, this application provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0039] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0040] The aforementioned multi-causal model integrated anomaly root cause tracing method, device, and computer equipment acquire multivariate time-series data of industrial equipment under abnormal conditions. Based on this data, at least three causal inference models employing different prediction functions are trained. A unified root cause score vector is determined based on the root cause score vector of each trained causal inference model. The root cause variable is then determined based on this unified root cause score vector. By constructing a heterogeneous causal modeling capability set covering various causal modes such as linear, nonlinear, and probabilistic models, the models can capture different forms of temporal dependencies between variables. Each model independently outputs a root cause score vector, reflecting its own understanding of the causal structure. By aggregating the scoring results of multiple models to generate a unified root cause score vector, the structural bias and noise interference of single models can be effectively offset, improving scoring stability. Thus, without the need for manually setting rules or prior knowledge, this overcomes the limitations of traditional methods that rely on experience-based judgment and single-point analysis, achieving the technical effect of improving the efficiency and accuracy of root cause identification in complex, coupled, and dynamically fluctuating manufacturing environments. Attached Figure Description

[0041] Figure 1 This is a diagram illustrating the application environment of an anomaly root cause tracing method integrating multiple causal models in one embodiment.

[0042] Figure 2 This is a flowchart illustrating an anomaly root cause tracing method integrating multiple causal models in one embodiment.

[0043] Figure 3 This is a schematic diagram of the model structure in one embodiment;

[0044] Figure 4 This is a schematic diagram of the abnormal root cause tracing results in one embodiment;

[0045] Figure 5 This is a schematic diagram comparing the tracing results of the traditional method in one embodiment;

[0046] Figure 6 This is a structural block diagram of an anomaly root cause tracing device integrating multiple causal models in one embodiment.

[0047] Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0049] The multi-causal model integration anomaly root cause tracing method provided in this application embodiment can be applied to, for example... Figure 1 In the application environment shown, industrial equipment 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or located in the cloud or on other network servers. Server 104 acquires multivariate time-series data of the industrial equipment 102 under abnormal conditions; based on the multivariate time-series data, it trains at least three causal inference models using different prediction functions; based on the root cause score vector of each causal inference model after training, it determines a unified root cause score vector; and based on the unified root cause score vector, it determines the root cause variable. Server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0050] In one embodiment, such as Figure 2 As shown, a method for tracing the root causes of anomalies by integrating multiple causal models is provided, and this method is applied to... Figure 1 Taking server 104 as an example, the anomaly root cause tracing method of multi-causal model integration includes the following steps:

[0051] Step S100: Obtain multivariate time series data of industrial equipment under abnormal conditions.

[0052] The multivariate time-series data can be a set of interrelated numerical sequences continuously collected over time by multiple sensors during the operation of the industrial equipment 102. For example, based on the sensors installed in the industrial equipment 102, this set may include one or more time-series measurements of variables such as temperature, pressure, vibration, and current to reflect the multidimensional state evolution of the equipment. It is understood that the multivariate time-series data can also be other types of variables that the industrial equipment 102 can collect; this embodiment does not limit the specific types.

[0053] It is understandable that abnormal states can be abnormal states output by the industrial equipment's self-test, or abnormal states determined by the user during the use of the industrial equipment because the operating status or results of the equipment do not meet expectations. Taking manufacturing equipment as an example, abnormal states can be output by the manufacturing equipment's self-test, or they can be determined by the user based on abnormal noises emitted by the manufacturing equipment, or they can be determined by the user based on the yield rate of parts or products produced by the manufacturing equipment.

[0054] To acquire multivariate time-series data of industrial equipment under abnormal conditions, for example, server 104 can read sampling data from multiple sensors of industrial equipment 102 within the time range in which the abnormal condition occurs. Furthermore, the multivariate time-series data can be processed with time alignment, normalization, and missing value imputation to ensure the temporal consistency and integrity of the multivariate time-series data.

[0055] For example, reading sampled data can be achieved by extracting a timestamped sensor data stream, or by using an edge computing node to perform anomaly tagging retrieval on a locally stored time series database and extracting a multi-dimensional data matrix for the corresponding time period. Other synchronous or asynchronous acquisition methods can also be used, which are not limited in this embodiment.

[0056] Step S200: Based on multivariate time series data, train at least three causal inference models that use different prediction functions.

[0057] The causal inference model can be a computational model that models and quantifies the causal dependencies between variables based on mathematical functions. In this embodiment, the causal inference model can fit linear, nonlinear, and probabilistic causal patterns through prediction functions, thereby mining potential causal structures between variables from data, identifying potential drivers in abnormal propagation paths, and providing causal evidence from multiple model perspectives for root cause localization. For example, prediction functions may include, but are not limited to, linear regression, neural networks, and probabilistic graphical models.

[0058] In one exemplary embodiment, at least three causal inference models employing different prediction functions are trained based on multivariate time-series data. This can be achieved by using multivariate time-series data as input and training each of the three models separately, enabling them to learn temporal dependencies between variables, such as linear lag effects, nonlinear coupling relationships, and probabilistic transition patterns. This constructs a heterogeneous set of causal inference models, covering multiple causal modes and enhancing the adaptability to diverse coupling mechanisms in complex manufacturing systems, allowing the system to cope with different types of anomaly propagation paths. In another exemplary embodiment, at least three causal inference models employing different prediction functions can be jointly trained to improve the consensus among the conclusions of multiple causal inference models and avoid deviations in causal patterns caused by misidentification of prediction functions in individual causal inference models.

[0059] Step S300: Determine a unified root cause score vector based on the root cause score vector of each causal inference model after training.

[0060] The root cause score vector can be a numerical vector output by a single causal inference model, representing the probability of each variable being an anomalous root cause. Each element in the vector can correspond to the causal influence score of a variable, thus providing a root cause priority ranking of variables from the perspective of a single model. In an exemplary embodiment, after each causal inference model has been trained, it can perform causal inference on the input multivariate time-series data and output a causal influence score for each variable. The score reflects the confidence level of that variable as a root cause under the current model structure.

[0061] A unified root cause score vector is a comprehensive score vector generated by aggregating root cause score vectors output by multiple causal inference models. It integrates the causal inference results of different prediction function models and can be used to improve the stability and generalization ability of root cause scoring, reduce the risk of misjudgment due to model structural limitations, and ensure consistency of scoring results under different abnormal scenarios. For example, the unified root cause score vector can align and fuse the root cause score vectors output by each model through weighted averaging, ranking integrals, voting mechanisms, and confidence threshold filtering, thereby eliminating individual model bias. Furthermore, the weights used to aggregate the root cause score vectors in the unified root cause score vector can be dynamically adjusted based on the model's performance on the validation set.

[0062] By determining a unified root cause score vector and adopting a multi-model consensus mechanism, the misjudgment caused by structural bias or overfitting of a single model can be effectively reduced, thereby improving the reliability and robustness of root cause scoring and enabling the system to maintain stable output when faced with noise interference.

[0063] Step S400: Determine the root cause variable based on the unified root cause score vector.

[0064] Root causes can be the initial driving variables identified in abnormal events as causing the system to deviate from its normal state. Understanding is that disturbances in root causes trigger chain reactions within the system's internal coupling relationships, resulting in abnormal states. In this embodiment, by identifying root causes, the source of the anomaly propagation chain can be clearly identified, providing a direct basis for fault intervention and process adjustments, enabling operational strategies to focus on the root cause rather than the symptoms. In an exemplary embodiment, root causes can be sorted according to the scores of variables in a unified root cause score vector, and thresholds can be set to select the variable with the highest score as the final output. This allows for automatic location of the anomaly source from a high-dimensional variable space, avoiding reliance on human experience and subjective judgment bias.

[0065] Taking the source tracing of vibration anomalies in high-precision CNC machine tool spindles as an example, the multi-causal model integrated root cause tracing method in this embodiment can simultaneously collect data from multiple sensors such as temperature, current, vibration acceleration, spindle speed, and coolant pressure during the operation of the CNC machine tool spindle. When a minor vibration anomaly occurs, multivariate time-series data of the abnormal period are automatically extracted and input into multiple causal inference models using different prediction functions. For example, the prediction functions include linear regression models, LSTM models, and dynamic Bayesian network models, which are then trained. Each model outputs its own root cause score vector. In one possible case, the linear model tends to regard vibration acceleration as the root cause, the LSTM model identifies hysteretic changes in coolant pressure as the driving source, and the Bayesian network indicates that current fluctuations are the initial disturbance. By weighted fusion to generate a unified root cause score vector, the variable with the highest score can be determined, and the variable with the highest score is then identified as the root cause variable, which is used as the tracing result.

[0066] This embodiment provides an anomaly root cause tracing method integrating multiple causal models. It acquires multivariate time-series data of industrial equipment under abnormal conditions, trains at least three causal inference models using different prediction functions based on this data, determines a unified root cause score vector based on the root cause score vector of each trained causal inference model, and identifies the root cause variable based on this unified root cause score vector. By constructing a heterogeneous causal modeling capability set covering various causal modes such as linear, nonlinear, and probabilistic models, the models can capture different forms of temporal dependencies between variables. Each model independently outputs a root cause score vector, reflecting its own understanding of the causal structure. By aggregating the scoring results of multiple models to generate a unified root cause score vector, it effectively offsets the structural bias and noise interference of single models, improving scoring stability. Thus, without the need for manually setting rules or prior knowledge, it overcomes the limitations of traditional methods that rely on experience-based judgment and single-point analysis, achieving the technical effect of improving the efficiency and accuracy of root cause identification in complex, coupled, and dynamically fluctuating manufacturing environments.

[0067] In one embodiment, the process of acquiring multivariate time series data includes:

[0068] Acquire historical time-series data of industrial equipment under abnormal conditions;

[0069] Variables are extracted from historical time series data to obtain structured time series data;

[0070] Structured time series data are standardized to obtain multivariate time series data.

[0071] Historical time-series data can be a collection of raw time-series data recorded by sensors during the past operation of industrial equipment. For example, it can include unfiltered, multi-source, heterogeneous, and unstructured measurements. In an exemplary embodiment, historical time-series data can be obtained by reading sensor sampling records that have been stored long-term by the equipment control system or edge gateway. For example, historical time-series data can be unaligned multi-sampling-rate raw logs, long-term operation records containing sensor drift, or hybrid datasets combining manually labeled and automatically collected data.

[0072] In this embodiment, the variables in the historical time-series data can be single measurement dimensions with clear physical meaning obtained from industrial equipment, such as temperature, pressure, vibration amplitude, current intensity, etc. In an exemplary embodiment, variables can be identified from historical time-series data based on sensor type, engineering labels, domain knowledge, etc., and independent measurement channels can be separated. For example, variables can include, but are not limited to, physical measurement variables, derived calculation variables, state indication variables, etc.

[0073] Structured time series data can be a numerical matrix with a unified timestamp and clearly defined variable dimensions, formed after variable extraction and time alignment. Furthermore, by filtering variables in historical time series data, structured time series data can also eliminate invalid or redundant channels, and achieve time axis alignment through interpolation or resampling to form a standardized two-dimensional time series matrix.

[0074] For example, variable extraction from historical time-series data can involve identifying and separating physically meaningful independent measurement channels from raw unstructured logs, and removing irrelevant or duplicate fields. Furthermore, it can be achieved by automatically matching and extracting valid variables based on sensor metadata tags, thereby improving data semantic clarity, reducing model input dimensional redundancy, and enhancing the interpretability and computational efficiency of causal inference.

[0075] Standardization can be performed on the numerical scale of variables in structured time-series data, so that their distribution characteristics participate in subsequent calculations under a unified dimension. For example, standardization can employ one or more methods such as Z-score standardization, Min-Max normalization, and robust scaling to independently perform linear transformations on each variable column, making the mean of each variable 0 or normalizing its range to the interval [0, 1], while preserving the original distribution shape. This can eliminate the interference of dimensional differences on model parameter learning, make the causal influence of different physical quantities comparable in the same numerical space, and avoid high-amplitude variables dominating model training.

[0076] This embodiment provides a method for acquiring multivariate time-series data. By acquiring historical time-series data of industrial equipment under abnormal conditions and extracting variables from the historical time-series data, the semantic clarity of the data can be effectively improved, and the redundancy of model input can be reduced. By standardizing the structured time-series data, the interference of differences in the dimensions of different sensors on model training can be eliminated, thereby effectively improving the signal-to-noise ratio and comparability of the input data, and suppressing false causal relationship misjudgments caused by data noise, asynchronous sampling, or dimensional deviation. Thus, without increasing the complexity of the model, the technical effect of improving the accuracy and precision of root cause identification can be achieved.

[0077] In one embodiment, training at least three causal inference models employing different prediction functions based on multivariate time-series data includes:

[0078] Based on multivariate time series data, at least three causal inference models using different prediction functions are pre-trained to obtain multiple pre-trained causal inference models.

[0079] Based on multivariate time series data and root cause score vectors of multiple pre-trained causal inference models, iterative interactive optimization is performed on multiple pre-trained causal inference models.

[0080] The pre-trained causal inference model can be an instance of a causal inference model independently trained based on historical multivariate time-series data before iterative interactive optimization. Each model uses a different prediction function to characterize a specific type of variable dependency. In this embodiment, the pre-trained causal inference model can be trained using the original multivariate time-series data as input, under non-interactive conditions, with at least three models having different prediction function structures, converging to a locally optimal causal representation state and outputting a stable but uncoordinated root cause score vector. This provides an initial set of models with differentiated causal cognition capabilities for subsequent iterative interactive optimization, ensuring that the model has a stable representation ability of the main dependency patterns in the data before optimization, and providing a heterogeneous basis for consensus formation among models.

[0081] In this embodiment, the root cause score vector can be a numerical vector output by a pre-trained causal inference model, representing the probability of each variable being an anomalous root cause, and used as a feedback signal between models in iterative interactive optimization. For example, after pre-training, the root cause score vector can be obtained by inputting multivariate time-series data into each model and generating a causal influence score sequence for the corresponding variable using the model's internal scoring module. For instance, the root cause score vector can be, but is not limited to, a score vector based on Granger significance calculation, a score vector based on attention weight aggregation, or a score vector based on posterior probability inference, etc.

[0082] In this embodiment, the iterative interactive optimization can be achieved through a feedback mechanism of root cause score vectors among multiple models, dynamically adjusting the parameters and structural weights of each model in a collaborative learning process. This allows for mutual correction and consensus convergence of causal cognition among models while maintaining the basic model structure. In an exemplary embodiment, the iterative interactive optimization process can involve constructing a scoring consistency loss function using the root cause score vectors output by each pre-trained model as input signals. Then, through gradient backpropagation or weight redistribution mechanisms, the internal parameters are adjusted while maintaining the model structure, causing each model to gradually shift towards the consensus direction of other models.

[0083] In some exemplary embodiments, the Euclidean distance between the root cause score vectors of each model can be calculated as a consistency loss and backpropagated to the internal parameters of each model to reduce the differences. Alternatively, the weight of the model with high consensus can be dynamically increased and the update step size of the divergent model can be reduced based on the ranking consistency of the score vectors of each model. This can achieve knowledge transfer and error cancellation between models, suppress the local optimal bias of individual models, enhance the overall collaborative convergence ability of the global causal structure, improve the stability and generalization of root cause identification, and enhance the robustness of the model in complex industrial scenarios.

[0084] This embodiment provides an anomaly root cause tracing method integrating multiple causal models. It pre-trains at least three causal inference models employing different prediction functions based on multivariate time-series data, resulting in multiple pre-trained causal inference models. Then, iteratively and interactively optimizing these models based on the root cause score vectors of the multivariate time-series data and the multiple pre-trained causal inference models, independently modeling based on different prediction functions during the pre-training phase allows each causal inference model to develop a differentiated causal understanding of the multivariate time-series data. Finally, in the iterative and interactive optimization phase, using the root cause score vector as the sole interaction medium, the models dynamically adjust parameters through score consistency feedback. This enables knowledge transfer and bias cancellation between models, thereby improving the accuracy and precision of root cause identification.

[0085] In one embodiment, each round of iterative interaction optimization includes:

[0086] Aggregate the current root cause score vectors of multiple pre-trained causal inference models to obtain the current unified root cause score vector;

[0087] Construct root cause consistency constraints based on the current unified root cause score vector;

[0088] Based on the loss function and root cause consistency constraint of each pre-trained causal inference model, the causal structure of the pre-trained causal inference model is optimized to minimize the prediction error.

[0089] The current unified root cause score vector can be a comprehensive score vector generated by aggregating the root cause score vectors output by multiple pre-trained causal inference models in each iteration. It can represent the collaborative confidence distribution of each variable as a root cause in the current round. For example, the current unified root cause score vector can be generated by aggregating the root cause score vectors output by each model in the current round after standardization and alignment, using methods such as weighted average or rank sum aggregation.

[0090] Root cause consistency constraints can be regularization terms introduced into the model optimization objective. These terms quantify the deviation between the current model's output root cause score vector and the current unified root cause score vector, forcing the model update direction to align with system consensus. In an exemplary embodiment, root cause consistency constraints can be implemented by calculating a distance metric between the individual model's output root cause score vector and the current unified root cause score vector, such as the L2 norm, Kullback-Leibler divergence, or Spearman rank correlation loss. This metric is then added as a penalty to the loss function of the model pair to quantify the degree to which the model deviates from system consensus and to constrain parameters. Exemplary examples include, but are not limited to, squared penalty terms based on Euclidean distance, probability distribution alignment terms based on KL divergence, and rank loss terms based on ranking differences.

[0091] The loss function can be a mathematical expression that measures the prediction error of a pre-trained causal inference model on multivariate time series data under the current parameters. It is used to guide the local optimization direction of model parameters and serves as a quantitative basis for the model's own fitting ability.

[0092] The causal structure can be a parameterized topology or function mapping within a causal inference model used to characterize the dependencies between variables. For example, it can include adjustable components such as weight matrices, attention mechanisms, and conditional probability tables, and can be dynamically adjusted through the optimization process to adapt to the dual needs of system consensus and local data fitting. For example, the causal structure can be determined by the architecture of the prediction function, such as the coefficient matrix of a linear model, the input layer weights of an LSTM, and the edge weights of a Bayesian network. During optimization, it can be adjusted using gradient descent or weight update rules to simultaneously minimize prediction error and consistency constraints, and gradually converge to a system-level consensus structure.

[0093] The operation of optimizing the causal structure of each pre-trained causal inference model based on its loss function and root cause consistency constraint to minimize prediction error can be achieved by adding the model's own prediction error loss function to the root cause consistency constraint to form a joint optimization objective. This objective is then updated by backpropagating gradients to update the internal causal structure parameters of the model, allowing the model to achieve a balance between fitting the data and conforming to consensus. This enables the joint optimization of local data fitting and global causal consensus, allowing the model to gradually converge to a structure that conforms to the system-level causal laws while maintaining sensitivity to the original data, thus avoiding a single model from getting stuck in local optima or consensus drift.

[0094] This embodiment provides an anomaly root cause tracing method integrating multiple causal models. It aggregates the current root cause score vectors of multiple pre-trained causal inference models to obtain a unified current root cause score vector. Based on this unified root cause score vector, a root cause consistency constraint term is constructed. The causal structure of each pre-trained causal inference model is optimized based on its loss function and the root cause consistency constraint term to minimize prediction error. A system-level consensus benchmark is generated through aggregation, and the consistency constraint term is introduced to jointly optimize the causal structure. This allows the model to gradually converge to a stable structure that conforms to system-level causal laws while maintaining its ability to fit local data. This effectively suppresses inference conflicts and local optimum deviations caused by differences in structural assumptions between heterogeneous models, thereby improving the stability and accuracy of root cause identification.

[0095] In one embodiment, the current root cause score vectors of multiple pre-trained causal inference models are aggregated to obtain a current unified root cause score vector, including:

[0096] Determine the first unified root cause score vector based on the current root cause score vectors of multiple pre-trained causal inference models.

[0097] The second unified root cause score vector is determined based on the sparse attention mechanism, the first unified root cause score vector, and the root cause score vector of each pre-trained causal inference model.

[0098] Based on the projection of the causal matrix of multiple pre-trained causal inference models into the solution space corresponding to the second unified root cause score vector, the second unified root cause score vector is optimized to obtain the current unified root cause score vector.

[0099] The first unified root cause score vector can be an initial consensus score vector generated from the root cause score vectors of multiple pre-trained causal inference models through a basic aggregation method, which can be used to provide an unbiased initial scoring benchmark. For example, the first unified root cause score vector can be directly calculated from the root cause score vectors output by each model using statistical methods, ensuring the initial vector possesses unbiasedness and computational stability. For instance, the first unified root cause score vector can be obtained by taking the arithmetic mean after Z-score standardization of all model score vectors, or by taking the median after sorting the model score vectors as the initial score for each variable, thus obtaining an unbiased and stable initial consensus benchmark.

[0100] The sparse attention mechanism can be a weighted aggregation mechanism that dynamically calculates the semantic relevance between the root cause score vectors of each model and the first unified root cause score vector, and retains only the contributions of a few highly relevant models. This can be used to filter out interference from low-confidence or noisy models, enhancing the sensitivity of the aggregation results to key causal signals. In an exemplary embodiment, the sparse attention mechanism can calculate the cosine similarity or Euclidean distance between each model's root cause score vector and the first unified root cause score vector, apply the Softmax function to generate the original weights, and then reset the weights of low-relevance models to zero through one or more operations such as Top-K truncation and threshold masking, retaining only the most discriminative models for subsequent aggregation. Correspondingly, the second unified root cause score vector can be a high-precision consensus score vector generated by weighted aggregation of the root cause score vectors of highly relevant models after the sparse attention mechanism has filtered the data. This vector carries high-confidence consensus information that has undergone semantic filtering, improving the sensitivity and interpretability of the aggregation results to key causal signals.

[0101] The causal matrix can be a parameterized matrix used within each pre-trained causal inference model to characterize the strength of causal dependencies between variables, and can be used to carry the internal prior knowledge corresponding to the causal structure of the model.

[0102] The solution space can be a high-dimensional scoring space defined by the second unified root cause score vector, which can be used to provide a reference benchmark for structural alignment of the causal matrix. Projection can be a linear or nonlinear mapping of the causal matrices of each model in the solution space, aligning their structural orientation with the consensus trend represented by the second unified root cause score vector. For example, projection can be the calculation of the orthogonal projection components of the causal matrix in the solution space direction, thereby adjusting the parameters of the causal matrix to approximate the consensus direction, achieving a physical adaptation of the model's internal structure to the system consensus. This can achieve a leap from score aggregation to structural alignment, making the final unified root cause score vector not only reflect the model's output consensus but also conform to the physical consistency of the model's internal causal structure.

[0103] This embodiment provides an anomaly root cause tracing method integrating multiple causal models. By aggregating the current root cause score vectors of multiple pre-trained causal inference models, a first unified root cause score vector without filtering is generated as the initial consensus. Then, a sparse attention mechanism is introduced to dynamically filter high-discrimination models and construct a semantically refined second unified root cause score vector. The causal matrix of each model is projected onto the solution space defined by the second vector, forcing the internal causal structure of the model to align with the system consensus direction. This ensures that the final unified root cause score vector not only reflects scoring consistency but also carries structural consistency, achieving the dual constraints of scoring consensus and structural consensus. While preserving model heterogeneity, it guides the inference path to converge toward a physically interpretable causal topology, significantly reducing the misjudgment rate and achieving the technical effect of improving the stability and accuracy of root cause identification.

[0104] In one embodiment, determining the second unified root cause score vector based on the sparse attention mechanism, the first unified root cause score vector, and the root cause score vector of each pre-trained causal inference model includes:

[0105] The root cause residual coefficients of each pre-trained causal inference model are calculated based on the first unified root cause score vector, the root cause score vector of each pre-trained causal inference model, the nonlinear activation function, and the learnable linear mapping parameters.

[0106] The root cause residual coefficients are sparsely normalized to obtain the attention weights for each pre-trained causal inference model;

[0107] A second unified root cause score vector is determined based on the current root cause score vectors and attention weights of multiple pre-trained causal inference models.

[0108] The root cause residual coefficient can be a numerical vector characterizing the degree of difference between the root cause score vector output by a single pre-trained causal inference model and a first unified root cause score vector. It can be used to quantify the deviation of each pre-trained causal inference model relative to the overall consensus. In an exemplary embodiment, the root cause residual coefficient can be obtained by subtracting the root cause score vector of each model from the first unified root cause score vector, performing a dimensionality transformation through learnable linear mapping parameters, and then processing it into a nonlinear activation function.

[0109] Nonlinear activation functions are mathematical functions used to perform nonlinear transformations on the root cause residual signal after linear mapping. This allows the residual coefficients to capture asymmetric, threshold-type, or saturation-type model bias characteristics. For example, nonlinear activation functions can improve the discriminative power and robustness of the residual signal by modifying the activation function of the linear unit, preserving significant deviations while suppressing minor fluctuations.

[0110] Learnable linear mapping parameters can be trainable weight matrices used to map the difference between the model root cause score vector and the first unified root cause score vector to a high-dimensional residual space. This allows the system to learn which variables' deviations are more discriminative in a specific manufacturing scenario. For example, learnable linear mapping parameters can be initialized as random matrices and updated along with the model parameters during iterative optimization, thereby achieving differentiated encoding of residual patterns among different variables.

[0111] In an exemplary embodiment, the root cause residual coefficient of each pre-trained causal inference model can be calculated by subtracting the first unified root cause score vector from the product of the causal matrix of each model and the first unified root cause score vector, to obtain a difference vector that expresses the model’s fit to the first unified root cause score vector. After being transformed by learnable linear mapping parameters, the difference vector is input into a nonlinear activation function and outputs the root cause residual coefficient vector of each model. This allows the residual coefficient to accurately reflect the true deviation of the model from specific variables, rather than simple numerical differences, thereby improving the semantic level of the model’s discriminative ability.

[0112] Sparse normalization can be achieved by competitively normalizing the root cause residual coefficients, retaining high-value elements and suppressing low-value elements to make the attention weights sparsely distributed. Correspondingly, the attention weights can be vectors generated by sparse normalization, representing the contribution priority of each pre-trained causal inference model in the current aggregation round. These weights can be used as the basis for weighted aggregation, determining the influence of each model's root cause score vector in generating the second unified root cause score vector. In an exemplary embodiment, the attention weights can be generated using sparse normalization functions such as sparsemax to ensure that only a few models receive significant weights.

[0113] The second unified root cause score vector is generated by weighting and summing the root cause score vectors of each model according to their corresponding attention weights. By calculating the second unified root cause score vector based on the attention weights, it is possible to make the second unified root cause score vector reflect the most reliable causal inference consensus in the system, significantly improving the sensitivity and noise resistance of root cause localization to small perturbations.

[0114] This embodiment provides an anomaly root cause tracing method integrating multiple causal models. It calculates root cause residual coefficients based on a first unified root cause score vector, the root cause score vectors of each pre-trained causal inference model, a nonlinear activation function, and learnable linear mapping parameters. The root cause residual coefficients are then sparsely normalized to obtain attention weights. A second unified root cause score vector is determined based on the root cause score vectors of multiple models and the attention weights. By introducing learnable linear mapping parameters and a nonlinear activation function, semantic discriminative power modeling is performed on the differences between the models and consensus. Sparse normalization enables dynamic filtering of model contributions, and the second unified root cause score vector is generated through weighted aggregation of attention weights. This method transforms the passive acceptance of all model outputs into proactive identification and focus on the most reliable causal inference sources. Without relying on manual rules, it achieves adaptive suppression and high-precision root cause localization of multi-source heterogeneous noise in manufacturing systems. Especially in cascading anomalies caused by minute parameter perturbations, it effectively avoids misjudgments and delays, thereby improving the stability and accuracy of root cause identification.

[0115] In one embodiment, the second unified root cause score vector is optimized based on the projection of the causal matrix of multiple pre-trained causal inference models into the solution space corresponding to the second unified root cause score vector, resulting in the current unified root cause score vector including:

[0116] Construct the causal structure solution space based on the second unified root factor fraction vector;

[0117] Project the causal matrix of each pre-trained causal inference model onto the causal structure solution space to obtain the nearest neighbor mapping point of the causal matrix;

[0118] Based on the nearest neighbor mapping point corresponding to each pre-trained causal inference model, the second unified root cause score vector is updated in reverse to obtain the current unified root cause score vector.

[0119] The causal structure solution space can be a high-dimensional geometric space defined by the second unified root cause score vector. Its direction and scale represent the causal structure trend of the current system-level consensus, and the projectable range of the causal structure is limited by constraints, serving as the target reference coordinate system for projecting the causal matrices of each model. For example, the causal structure solution space can be constructed based on the second unified root cause score vector as the basis vector to include all possible causal structure expressions that are consistent with or approximate with the consensus vector in direction, and satisfy preset constraints.

[0120] Constructing the causal structure solution space can be achieved by using the second unified root cause fraction vector as the reference direction, constructing a high-dimensional subspace using linear algebra methods, and combining it with the constraint matrix as the target space on which all model causal matrices can be projected. This allows the establishment of a geometric reference framework consistent with the system consensus, enabling subsequent model projection operations to have clear optimization objectives, and improving the physical interpretability of the optimization process through structural consistency constraints.

[0121] The nearest neighbor mapping point can be the point that is most similar to the structure of each pre-trained causal inference model after projecting its causal matrix into the causal structure solution space, and satisfies the constraints. It represents the optimal structural expression of the model in the consensus space. For example, the nearest neighbor mapping point can be calculated by performing linear projection on the causal matrix of each model, minimizing the Euclidean distance loss between it and each point in the solution space while satisfying the constraints, and finding the projection point in the solution space that is most consistent with its original structure. This transforms the internal causal structure prior of each model into a quantifiable position in the consensus space, achieving geometric alignment between the model structure and the system consensus.

[0122] Updating the second unified root cause score vector in reverse can be achieved by using the nearest neighbor mapping points of each model as structural guidance information. Through weighted averaging, gradient ascent, or other methods, the direction and magnitude of the second unified root cause score vector are adjusted, and based on minimizing the loss function, it gradually approximates the causal structure commonly accepted by all models. In an exemplary embodiment, this can be achieved by averaging all nearest neighbor mapping points and normalizing the average value as the update direction of the current unified root cause score vector. Alternatively, it can be achieved by calculating the gradient direction between each mapping point and the second unified root cause score vector, and updating the second unified root cause score vector along the gradient ascent direction to maximize structural consistency. This enables a reverse feedback loop from model structure output to global consensus, ensuring that the current unified root cause score vector not only reflects scoring consistency but also the collaborative convergence result of the causal topology within the model. Furthermore, through structural consistency constraints, the accuracy of the representation of the true causal path is improved.

[0123] This embodiment provides an anomaly root cause tracing method based on multi-causal model integration. It constructs a causal structure solution space based on a second unified root cause score vector, projects the causal matrix of each pre-trained causal inference model into this space to obtain nearest neighbor mapping points, and updates the second unified root cause score vector in reverse based on these mapping points. By constructing the causal structure solution space, geometric alignment between the model's prior structure and the system's consensus space can be achieved. The projection operation transforms the model's internal structure into quantifiable positions in the consensus space, and the reverse update operation gradually brings the consensus vector closer to the causal structure region commonly recognized by all models. This improves the accuracy of representing the true causal path, suppresses conflicts and noise caused by differences in model structure, and ensures that the final unified root cause score vector carries a physically interpretable causal structure consensus. This significantly improves the sensitivity to tracing small perturbations, achieving the technical effect of improving the stability and accuracy of root cause identification.

[0124] To more clearly illustrate the technical solution of this application, a detailed embodiment is also provided.

[0125] In the global wave of intelligent and digital transformation of manufacturing, the widespread adoption and application of artificial intelligence and big data technologies are reshaping traditional production models. With the comprehensive advancement of Industry 4.0, manufacturing systems are increasingly characterized by interconnected equipment, data-driven processes, and intelligent decision-making. The massive amounts of sensor data, process parameter data, and equipment status data accumulated during production provide new technological pathways for anomaly diagnosis, handling, and quality control. However, traditional anomaly diagnosis methods in manufacturing processes often rely on manual experience or simple analysis based on single sensor data, which have limitations such as response lag, low positioning accuracy, and insufficient correlation analysis, making them difficult to adapt to the multi-variable coupling and dynamic fluctuation characteristics of complex manufacturing systems. Especially in the field of high-precision manufacturing, minute parameter deviations can be amplified step by step through the production chain, leading to product quality defects or unplanned equipment downtime. Traditional methods often struggle to quickly locate the source of the anomaly, resulting in increased production costs and lost production efficiency. Against this backdrop, root cause analysis technology based on AI and data mining has emerged. Its core value lies in analyzing the information transmission relationships of key parameters in the production process, automatically reflecting the patterns of information propagation, and quickly locating the root causes of anomalies. This breaks through the bottleneck of traditional reliance on experience and improves the automation and timeliness of anomaly diagnosis. From an application perspective, the research and promotion of this technology can significantly reduce the cost of quality control in the manufacturing industry, reduce the time and manpower costs of anomaly handling and investigation, and help manufacturing enterprises achieve intelligent transformation.

[0126] Currently, advanced root cause analysis methods are often based on causal inference methods, reflecting the information transmission path between variables through causal relationships. However, most existing root cause analysis methods infer causal relationships between production process variables based on a single causal model. In reality, causal analysis models are extremely diverse, and different model structures and parameter selections will lead to different causal analysis and root cause analysis conclusions. Taking Granger causality, a typical root cause analysis model, as an example, this type of method reflects causal relationships by using the predictive relationships between variables. It includes a series of specific methods, each employing different predictive models. These different model settings can lead to significantly different final analytical conclusions, presenting a causal heterogeneity problem. As a result, unlike classification and regression tasks that widely adopt ensemble learning paradigms, current root cause analysis methods can only use a single causal inference model in their application, failing to effectively integrate multiple models for ensemble learning to enhance model performance, resulting in insufficient reliability, robustness, and generalization of root cause analysis conclusions.

[0127] Therefore, to address the causal heterogeneity problem in the aforementioned root cause tracing task, there is an urgent need to establish an integrated learning method for causal inference models to effectively integrate causal information from various models and provide an effective intelligent solution for the effective anomaly tracing and handling of manufacturing systems.

[0128] In one embodiment, to address the shortcomings of existing single-root cause tracing methods, a multi-causal model integrated anomaly root cause tracing method is provided. This method can automatically mine consistent causal and root cause information from diverse and heterogeneous causal models, achieving complementary fusion of models to form a unified root cause tracing conclusion. This improves the reliability and robustness of the root cause tracing model and can be applied to actual manufacturing processes, thereby enhancing the efficiency of anomaly handling and maintaining the efficient and stable operation of the manufacturing process. The method includes the following steps:

[0129] Step 1: Collect time-series data of abnormal states in the manufacturing process to form a training set, extract the unified form of different Granger causal inference models, pre-train three or more causal inference models based on the training set to obtain the corresponding heterogeneous causal structure, and use the approximate exponential transformation to obtain the root cause score vector of the causal structure of each method, and initialize the unified root cause vector to the mean of the root cause score vector of the causal structure of each method.

[0130] Collect time-series data x = {x1, x2, …, x} under abnormal conditions in the manufacturing process. N}, consisting of N dimensions A matrix T, where N and T represent the number of variables and the time window length, respectively, x i (1≤i≤N) represents the sequence of the i-th variable. Continuous variables are standardized using Z-scores, which involves removing the mean and dividing by the standard deviation of the variable to eliminate differences in the dimensions of the variables.

[0131] Extract the common form from different Granger causal inference models and establish M (M is an integer greater than or equal to 3) different causal models:

[0132]

[0133] in, A represents the time series prediction result of the j-th inference model; j It is the causal matrix of the j-th model (with dimension N). The value of the element in the u-th row and v-th column of variable x (u and v are any integers between 1 and N, and they are not equal, representing the index of the variable) directly determines the value of variable x. v For x u The predictive strength of (i.e., the u-th and v-th variables) can therefore represent x. v For x u Granger causality; f j Let be the prediction function for the j-th model. This function can be a linear transformation, a fully connected network, a long short-term memory network, a gated recurrent unit, or a Transformer model. Different prediction functions lead to varying structures in the Granger causal inference models, potentially resulting in discrepancies between causal inference and root cause tracing, thus failing to produce a unified conclusion and exhibiting heterogeneity. For example... Figure 3 The diagram shown is a schematic of the model structure.

[0134] The loss functions for each model are:

[0135]

[0136] in The overall loss of the j-th model; Let be the prediction error loss for the j-th model; vec is the matrix vectorization operation; L1 norm-based sparsity constraints are used to achieve sparsification of causal structures in order to filter key causal relationships; denoted as the sparsity constraint coefficient for the j-th model.

[0137] Using training data, by minimizing the loss function Pre-train each model to obtain the causal structure {A1, ..., A} of each model. M The root cause score vector is obtained by using an approximate exponential transformation:

[0138]

[0139] Where P is the approximation order; a higher value results in more accurate calculations. Based on the application effect, a value of 3 for P is appropriate here. k This represents the k-th row vector of any matrix (1≤k≤N); sj Let be the root cause score vector for the j-th model, where the i-th element represents the variable x in the causal structure of the j-th model. i For the amount of causal information output in the entire information dissemination process, the larger its value, the more significant the causal information of variable x in the inference result of the j-th model. i It is more likely to be identified as the root cause.

[0140] The unified root factor vector s is initialized with the mean of the score vectors of each model, resulting in the first unified root factor score vector, i.e.:

[0141]

[0142] Step 2: Establish a root cause sparse attention mechanism, aggregate the root cause score vectors from Step 1 based on the score residuals, update the unified root cause vector, and obtain the second unified root cause score vector.

[0143] Establish a sparse attention mechanism for root cause scores and calculate the root cause residual coefficients for each causal inference model:

[0144]

[0145] Among them, e j Let be the root cause residual coefficient of the j-th model; The parameters are learnable linear mappings; the LeakyReLU function operates as follows:

[0146]

[0147] Then, the root cause attention score for each model is calculated, and the second unified root cause vector s is calculated based on this score.

[0148]

[0149] Where sparsemax is the sparse normalization function.

[0150] Step 3 is as follows: For the M heterogeneous causal structures in Step 1, find their nearest neighbor mapping point in the solution space of the second unified root cause vector, use the mapping point to replace the causal structure of the Granger causal inference model, and update the second unified root cause vector by minimizing the prediction error.

[0151] The unified root cause vector s obtained in step 2 forms a causal structure solution space, where any causal structure A satisfies: s = A T s. Here, we search for the causal structure {A1, ..., A...} output by each causal model. M Find the nearest neighbor in the solution space above, i.e., solve the following optimization problem:

[0152]

[0153] in:

[0154]

[0155]

[0156] s i Represents the variable x in s i The corresponding root factor fraction; Solving this problem using the Lagrange multiplier method yields the following closed-form solution:

[0157]

[0158] in{} 1:N 2 Indicates taking the vector before Each element. Rearranged into a matrix , which is the nearest neighbor of the j-th inference model's causal structure in the solution space formed by s. Subsequently, based on the following optimization objective, the obtained nearest neighbor is used to optimize the second unified root cause vector s, making it more closely match the causal structures of each model:

[0159]

[0160] Step 4: Based on the updated second unified root cause vector from Step 3, establish root cause consistency constraints to optimize each causal inference model, update its causal structure, and return to Step 2. Iterate through training to achieve bidirectional convergence of the causal structure of each model with the second root cause score vector, outputting a unified root cause tracing result. After training, the variable with the highest score in the second unified root cause vector is determined as the root cause variable.

[0161] Based on the updated unified root cause vector s from step 3, a root cause consistency constraint is established to ensure that the heterogeneous causal structures output by each causal inference model point to the same root cause variable. The root cause consistency constraint can be calculated using the following formula:

[0162]

[0163] The loss functions for each causal inference model in this stage are:

[0164]

[0165] Where β is the weighting coefficient.

[0166] The loss function is used to optimize for several rounds to update the causal structure {A1, ..., A} of each causal inference model. MThen return to step 2. Steps 2 through 4 are iteratively trained multiple times until convergence, realizing the causal structure {A1, ..., A} of each model. M The two-way convergence between the root cause score s and the second unified root cause score s. Ultimately, s can serve as the root cause representation after the integration and fusion of causal information from various heterogeneous models, and then s can be used for root cause tracing, where the variable with the largest corresponding root cause score in s is determined as the root cause variable.

[0167] The following will explain how the anomaly root cause tracing method based on a multi-causal model integrated in this embodiment is applied to the Eastman manufacturing process in Tennessee and the active alignment process of an automotive optical module, which are typical representatives of the continuous manufacturing and discrete manufacturing industries, respectively.

[0168] In one specific embodiment, the root cause analysis case data of the Tennessee Eastman manufacturing process is used to illustrate the application of this embodiment. The Tennessee Eastman dataset was developed and constructed by Eastman Corporation. It simulates the typical characteristics of a continuous manufacturing process and various failure cases to evaluate the effectiveness of the manufacturing process monitoring and diagnosis algorithm. The manufacturing process package covers five main units: reactor, condenser, gas-liquid separator, circulating compressor, and stripping tower. The dataset measures 41 process variables and 12 operational variables. By introducing different abnormal disturbances, 21 failure cases are formed, each containing 960 sample points. Here, the IDV (4) case from the dataset is selected to illustrate the effectiveness of this embodiment. Specifically, during the manufacturing process, an abnormal step change occurs in the reactor cooling water inlet temperature, affecting the reactor temperature x9 and the reactor cooling water outlet temperature x. 21 However, at the reactor cooling water outlet temperature x 21 With reactor cooling water flow rate x 51 There is a controller that adjusts x 51 x was eliminated 21 The abnormal changes. Ultimately, x9 and x 21 Keep it normal, while x 51 A step change occurs, therefore x 51 These were identified as root causes. The key failure variables in this case include {x2, x3, x5, x6, x7, x8, x...} 15 , x 16 ,x 18 , x 19 , x 20 , x 23 , x 29 , x 40 , x 41 , x 47 , x 48 , x 49 , x 50 , x51 The time-series data following these variable failures are collected here for modeling. The method in this specific embodiment includes the following steps:

[0169] Step 1: Collect time-series data of abnormal states in the manufacturing process to form a training set, extract the unified form of different Granger causal inference models, pre-train three or more causal inference models based on the training set to obtain the corresponding heterogeneous causal structure, and use the approximate exponential transformation to obtain the root cause score vector of the causal structure of each method, and initialize the unified root cause vector to the mean of the root cause score vector of the causal structure of each method.

[0170] Collect time-series data x = {x1, x2, …, x} under abnormal conditions in the manufacturing process. N}, consisting of N dimensions The matrix T, where N and T represent the number of variables and the length of the time window, respectively. Let N represent the sequence of the i-th variable. Continuous variables are standardized using Z-scores, which involves removing the mean and dividing by the standard deviation to eliminate differences in the variables' dimensions. Here, N and T are 20 and 10, respectively.

[0171] Extract the common form from different Granger causal inference models and establish M different causal models, where M is set to 4:

[0172]

[0173] in, A represents the time series prediction result of the j-th inference model; j It is the causal matrix of the j-th model (with dimension N). The value of the element in the u-th row and v-th column of variable x (u and v are any integers between 1 and N, and they are not equal, representing the index of the variable) directly determines the value of variable x. v For x u The predictive strength of (i.e., the u-th and v-th variables) can therefore represent x. v For x u Granger causality; f j Let be the prediction function for the j-th model. The four models selected here are linear transformation, fully connected network, long short-term memory, and Transformer.

[0174] The loss functions for each model are:

[0175]

[0176] in The overall loss of the j-th model; Let be the prediction error loss for the j-th model; vec is the matrix vectorization operation; L1 norm-based sparsity constraints are used to achieve sparsification of causal structures in order to filter key causal relationships; Here, the sparsity constraint coefficients for the j-th model are uniformly selected as 0.3, 0.2, 0.3, and 0.6 for the four models, respectively.

[0177] Using training data, by minimizing the loss function Pre-train each model to obtain the causal structure {A1, ..., A} of each model. M The root cause score vector is obtained by using an approximate exponential transformation:

[0178]

[0179] Where P is the approximation order; a higher value results in more accurate calculations. Based on the application effect, a value of 3 for P is appropriate here. k This represents the k-th row vector of any matrix (1≤k≤N); s j Let be the root cause score vector for the j-th model, where the i-th element represents the variable x in the causal structure of the j-th model. i For the amount of causal information output in the entire information dissemination process, the larger its value, the more significant the causal information of variable x in the inference result of the j-th model. i It is more likely to be identified as the root cause of the failure.

[0180] The unified root factor vector s is initialized with the mean of the score vectors of each model, resulting in the first unified root factor score vector, i.e.:

[0181]

[0182] Step 2: Establish a root cause sparse attention mechanism, aggregate the root cause score vectors from Step 1 based on the score residuals, update the unified root cause vector, and obtain the second unified root cause score vector.

[0183] Establish a sparse attention mechanism for root cause scores and calculate the root cause residual coefficients for each causal inference model:

[0184]

[0185] Among them, e j Let be the root cause residual coefficient of the j-th model; The parameters are learnable linear mappings; the LeakyReLU function operates as follows:

[0186]

[0187] Then, the root cause attention score for each model is calculated, and the second unified root cause vector s is updated accordingly:

[0188]

[0189] Where sparsemax is the sparse normalization function.

[0190] Step 3 is as follows: For the M heterogeneous causal structures in Step 1, find their nearest neighbor mapping point in the solution space of the second unified root cause vector, use the mapping point to replace the causal structure of the Granger causal inference model, and update the second unified root cause vector by minimizing the prediction error.

[0191] The unified root cause vector s obtained in step 2 forms a causal structure solution space, where any causal structure A satisfies: s = A T s. Here, we search for the causal structure {A1, ..., A...} output by each causal model. M Find the nearest neighbor in the solution space above, i.e., solve the following optimization problem:

[0192]

[0193] in:

[0194]

[0195]

[0196] s i Represents the variable x in s i The corresponding root factor fraction; Solving this problem using the Lagrange multiplier method yields the following closed-form solution:

[0197]

[0198] in{} 1:N 2 Indicates taking the vector before Each element. Rearranged into a matrix , which is the nearest neighbor of the j-th inference model's causal structure in the solution space formed by s. Subsequently, based on the following optimization objective, the obtained nearest neighbor is used to optimize the second unified root cause vector s, making it more closely match the causal structures of each model:

[0199]

[0200] Step 4: Based on the updated second unified root cause vector from Step 3, establish root cause consistency constraints to optimize each causal inference model, update its causal structure, and return to Step 2. Iteratively train to achieve bidirectional convergence of the causal structure of each model with the second root cause score vector, outputting a unified root cause tracing result. After training, the variable with the highest score in the second unified root cause vector is determined as the root cause variable of the fault.

[0201] Based on the updated second unified root cause vector s from step 3, a root cause consistency constraint is established to ensure that the heterogeneous causal structures output by each causal inference model point to the same root cause variable. The root cause consistency constraint can be calculated using the following formula:

[0202]

[0203] The loss functions for each causal inference model in this stage are:

[0204]

[0205] Where β is the weighting coefficient, which is chosen to be 0.5 here.

[0206] The loss function is used to optimize for several rounds to update the causal structure {A1, ..., A} of each causal inference model. M Then return to step 2. Steps 2 through 4 are iteratively trained multiple times until convergence, realizing the causal structure {A1, ..., A} of each model. M The two-way convergence between the root cause score s and the second unified root cause score s. Ultimately, s can serve as the root cause representation after the integration and fusion of causal information from various heterogeneous models, and then s can be used for root cause tracing, where the variable with the largest corresponding root cause score in s is determined as the fault root cause variable.

[0207] The multi-causal model integration anomaly root cause tracing method provided in this embodiment yields the following root cause tracing results for this fault dataset: Figure 4 As shown in the figure, the unified root cause vector s given in this embodiment is illustrated. It can be seen that the variable x... 51 The corresponding root cause score is the highest and significantly higher than the scores of other variables, indicating that this embodiment can effectively integrate multiple heterogeneous causal inference models and provide accurate source tracing conclusions.

[0208] Compared to the four existing root cause analysis methods, which are MPGE & RootRank (Song P, Zhao C, Huang B. MPGE and RootRank: A sufficient root cause characterization and quantification framework for industrial process faults[J]. Neural Networks, 2023, 161: 397-417.), Neural Granger (Tank A, Covert I, FotiN, et al. Neural granger causality[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 44(8): 4267-4279.), Lasso Granger (Arnold A, Liu Y, Abe N. Temporal causal modeling with graphical granger methods[C] / / Proceedings of the 13th ACM SIGKDD international conference on Knowledgediscovery and data mining. 2007: 66-75.), Symbolic Transfer Entropy (Staniek M, Lehnertz) K. Symbolic transfer entropy[J]. Physical review letters, 2008, 100(15):158-101.), where MPGE & RootRank are similar to this embodiment, tracing root causes based on root cause scores. Other methods use causal graphs to find root variables, that is, if a variable is the cause of other variables but not the result of any other variable, it is considered a root variable. It can be seen that none of the comparison methods can successfully trace the correct root variable x. 51 Among them, although the MPGE & RootRank model gives x 51 The root cause score is relatively high, but due to the high score of variable x2, it cannot give the correct root cause identification result. The causal structure given by the neuro-Grange and Lasso-Grange methods is more complex, with bidirectional and circular structures, making it difficult to identify the root cause variable. The symbolic transfer entropy method, on the other hand, uses variable x... 40 With x 41The root cause was misidentified. It's worth noting that all four comparison methods use a single model structure and cannot integrate different causal models to provide a comprehensive conclusion on the origin of the problem; therefore, their accuracy in tracing the origin is weaker than the method in this embodiment. This experimental conclusion demonstrates the effectiveness of this embodiment in integrating and fusing heterogeneous causal models.

[0209] In another specific embodiment, real data from a vehicle-mounted optical module production line is used to construct a dataset around the core process of active alignment (AA) to illustrate the application of this embodiment. AA is the core process for assembling high-precision optical modules. This process dynamically adjusts the relative positions of components such as lenses and sensors through real-time optical feedback to ensure that the module's full-field-of-view imaging performance meets the required standards. Five field-of-view light tubes (center, upper left, upper right, lower left, and lower right) of the AA equipment are selected as acquisition nodes. The dataset covers the following dimensions: horizontal / vertical angle of the light tube, fan angle and fan angle ratio, R / G / B values, and the module's X / Y / Z axis coordinates and X / Y direction rotation angle, totaling 40 variables. Optical module manufacturing belongs to the discrete manufacturing industry, and each product needs to perform an independent AA operation. In this example, abnormal equipment position deviation causes a very slight tilt in the X-direction rotation angle, ultimately resulting in a decrease in product pass rate. The experiment collected AA data from 263 defective products under this abnormal condition, hoping that each causal inference model can identify the X-direction rotation angle as the root cause.

[0210] The use of this embodiment includes the following steps:

[0211] Step 1: Collect time-series data of abnormal states in the manufacturing process to form a training set, extract the unified form of different Granger causal inference models, pre-train three or more causal inference models based on the training set to obtain the corresponding heterogeneous causal structure, and use the approximate exponential transformation to obtain the root cause score vector of the causal structure of each method, and initialize the first unified root cause vector to the mean of the root cause score vector of the causal structure of each method.

[0212] Collect time-series data x = {x1, x2, …, x} under abnormal conditions in the manufacturing process. N}, consisting of N dimensions A matrix T, where N and T represent the number of variables and the time window length, respectively, x i (1≤i≤N) represents the sequence of the i-th variable. Z-score standardization is used for continuous variables, i.e., removing the mean and dividing by the standard deviation to eliminate differences in the dimensions of the variables. Here, N and T are 40 and 12, respectively.

[0213] Extract the common form from different Granger causal inference models and establish M different causal models, where M is set to 4:

[0214]

[0215] in, A represents the time series prediction result of the j-th inference model; j It is the causal matrix of the j-th model (with dimension N). The value of the element in the u-th row and v-th column of variable x (u and v are any integers between 1 and N, and they are not equal, representing the index of the variable) directly determines the value of variable x. v For x u The predictive strength of (i.e., the u-th and v-th variables) can therefore represent x. v For x u Granger causality; f j Let be the prediction function for the j-th model. The four models selected here are linear transformation, fully connected network, long short-term memory, and Transformer.

[0216] The loss functions for each model are:

[0217]

[0218] in The overall loss of the j-th model; Let be the prediction error loss for the j-th model; vec is the matrix vectorization operation; L1 norm-based sparsity constraints are used to achieve sparsification of causal structures in order to filter key causal relationships; Here, the sparsity constraint coefficients for the j-th model are uniformly selected as 0.35, 0.15, 0.4, and 0.5 for the four models, respectively.

[0219] Using training data, by minimizing the loss function Pre-train each model to obtain the causal structure {A1, ..., A} of each model. M The root cause score vector is obtained by using an approximate exponential transformation:

[0220]

[0221] Where P is the approximation order; a higher value results in more accurate calculations. Based on the application effect, a value of 3 for P is appropriate here. k This represents the k-th row vector of any matrix (1≤k≤N); s j Let be the root cause score vector for the j-th model, where the i-th element represents the variable x in the causal structure of the j-th model. i For the amount of causal information output in the entire information dissemination process, the larger its value, the more significant the causal information of variable x in the inference result of the j-th model. i It is more likely to be judged as the root cause of product defects.

[0222] The unified root factor vector s is initialized with the mean of the score vectors of each model, resulting in the first unified root factor score vector, i.e.:

[0223]

[0224] Step 2: Establish a root cause sparse attention mechanism, aggregate the root cause score vectors from Step 1 based on the score residuals, update the unified root cause vector, and obtain the second unified root cause score vector.

[0225] Establish a sparse attention mechanism for root cause scores and calculate the root cause residual coefficients for each causal inference model:

[0226]

[0227] Among them, e j Let be the root cause residual coefficient of the j-th model; The parameters are learnable linear mappings; the LeakyReLU function operates as follows:

[0228]

[0229] Then, the root cause attention score of each model is calculated, and the second unified root cause vector s is obtained accordingly:

[0230]

[0231] Where sparsemax is the sparse normalization function.

[0232] Step 3 is as follows: For the M heterogeneous causal structures in Step 1, find their nearest neighbor mapping point in the solution space of the unified root cause vector, use the mapping point to replace the causal structure of the Granger causal inference model, and update the second unified root cause vector by minimizing the prediction error.

[0233] The second unified root cause vector s obtained in step 2 forms a causal structure solution space, where any causal structure A satisfies: s=A T s. Here, we search for the causal structure {A1, ..., A...} output by each causal model. M Find the nearest neighbor in the solution space above, i.e., solve the following optimization problem:

[0234]

[0235] in:

[0236]

[0237]

[0238] s i Represents the variable x in si The corresponding root factor fraction; Solving this problem using the Lagrange multiplier method yields the following closed-form solution:

[0239]

[0240] in{} 1:N 2 Indicates taking the vector before Each element. Rearranged into a matrix , which is the nearest neighbor of the j-th inference model's causal structure in the solution space formed by s. Subsequently, based on the following optimization objective, the obtained nearest neighbor is used to optimize the second unified root cause vector s, making it more closely match the causal structures of each model:

[0241]

[0242] Step 4: Based on the updated second unified root cause vector from Step 3, establish root cause consistency constraints to optimize each causal inference model, update its causal structure, and return to Step 2. Iteratively train to achieve bidirectional convergence of the causal structure of each model with the second root cause score vector, outputting a unified root cause tracing result. After training, the variable with the highest score in the second unified root cause vector is determined as the root cause variable leading to product non-compliance.

[0243] Based on the updated second unified root cause vector s from step 3, a root cause consistency constraint is established to ensure that the heterogeneous causal structures output by each causal inference model point to the same root cause variable. The root cause consistency constraint can be calculated using the following formula:

[0244]

[0245] The loss functions for each causal inference model in this stage are:

[0246]

[0247] Where β is the weighting coefficient, which is chosen as 0.4 here.

[0248] The loss function is used to optimize for several rounds to update the causal structure {A1, ..., A} of each causal inference model. M Then return to step 2. Steps 2 through 4 are iteratively trained multiple times until convergence, realizing the causal structure {A1, ..., A} of each model. M The two-way convergence between the root cause score s and the second unified root cause score s. Ultimately, s can serve as the root cause representation after the integration and fusion of causal information from various heterogeneous models, and then s can be used for root cause tracing, where the variable with the largest corresponding root cause score in s is determined as the root cause variable.

[0249] Step 5: The statistical model analyzes the data for each product, and the variable that is identified most frequently as the root cause of product nonconformity is identified as the final root cause.

[0250] The multi-causal model integration anomaly root cause tracing method in this embodiment, in the root cause tracing results of 263 non-conforming product datasets, inferred the root cause of non-conformity as the X-direction rotation angle for 143 products, with an accuracy rate of 54.37%. It can reveal the real reasons for product non-conformity relatively accurately, which helps to solve problems as soon as possible and improve yield. Compared to four existing root cause analysis methods, namely MPGE & RootRank (Song P, Zhao C, Huang B. MPGE and RootRank: A sufficient root cause characterization and quantification framework for industrial process faults[J]. Neural Networks, 2023, 161: 397-417.), Neural Granger (Tank A, Covert I, Foti N, et al. Neural granger causality[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 44(8): 4267-4279.), Lasso Granger (Arnold A, Liu Y, Abe N. Temporal causal modeling with graphical granger methods[C] / / Proceedings of the 13th ACM SIGKDD international conference on Knowledge discovery and data mining. 2007: 66-75.), and Symbolic transfer entropy (Staniek M, Lehnertz K. Symbolic transfer entropy[J]. Physicalreview letters, 2008, 100(15): 158101.), such as Figure 5 As shown, the accuracy rates of the four comparison methods are 42.96%, 40.68%, 33.84%, and 29.65%, respectively. It can be seen that this embodiment has higher root cause tracing accuracy among all methods.

[0251] This embodiment provides an anomaly root cause tracing method integrating multiple causal models, addressing the challenge of causal heterogeneity and difficulty in integrated analysis caused by the influence of different causal inference model structures. It proposes an ensemble learning scheme for heterogeneous causal models. This method fully explores the potentially consistent causal and root cause indication information in different causal inference models, achieving complementary fusion of multiple models in root cause tracing. Compared to traditional single-source tracing methods, it enriches causal information, possesses strong reliability and generalization, and points to a new direction for the research of anomaly problem tracing methods. Specifically:

[0252] (1) By extracting the unified form of different Granger causal inference models, namely, the prediction results obtained by nonlinear transformation of the time series data information of each variable and the causal matrix, it is possible to establish the general correlation between different model structures, and provide a formal basis for information fusion and integrated root cause tracing of each causal inference model.

[0253] (2) By constructing a sparse root cause attention and root cause solution space nearest neighbor mapping point computing mechanism, it is possible to integrate multiple heterogeneous causal models (greater than or equal to 3) to output a unified root cause score vector and indicate common root cause variables. It is possible to mine unified root cause indication information from multiple heterogeneous causal graphs (greater than or equal to 3), avoiding the limitation that each model independently gives different analysis conclusions and is difficult to integrate reasoning. The reliability of root cause tracing conclusions is improved by using an integrated learning architecture.

[0254] (3) By designing root cause consistency constraints and using root cause vectors as a medium, different causal models can achieve information complementarity from multiple perspectives, effectively realize information fusion and causal model integration, so that the reasoning process of each causal model is no longer isolated, but can learn from each other, promote information complementarity and fusion of each causal model in the root cause tracing task, and improve the accuracy of causal reasoning.

[0255] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0256] Based on the same inventive concept, this application also provides an apparatus for implementing the method described above. The solution provided by this apparatus is similar to the solution described in the above method; therefore, specific limitations in one or more apparatus embodiments provided below can be found in the limitations of the method described above, and will not be repeated here.

[0257] In one embodiment, such as Figure 6 As shown, this application provides an anomaly root cause tracing device integrating multiple causal models, applied to industrial equipment. The anomaly root cause tracing device integrating multiple causal models includes:

[0258] The data collection module 100 is used to acquire multivariate time-series data of industrial equipment under abnormal conditions.

[0259] The model training module 200 is used to train at least three causal inference models using different prediction functions based on multivariate time series data.

[0260] The vector determination module 300 is used to determine a unified root cause score vector based on the root cause score vector of each causal inference model after training.

[0261] The root cause determination module 400 is used to determine the root cause variable based on a unified root cause score vector.

[0262] In one embodiment, the data collection module 100 is further configured to:

[0263] Acquire historical time-series data of the industrial equipment under abnormal conditions;

[0264] Variables are extracted from the historical time series data to obtain structured time series data;

[0265] The structured time series data is standardized to obtain the multivariate time series data.

[0266] In one embodiment, the model training module 200 is further configured to:

[0267] Based on the multivariate time series data, at least three causal inference models using different prediction functions are pre-trained to obtain multiple pre-trained causal inference models.

[0268] Based on the multivariate time series data and the root cause score vectors of the multiple pre-trained causal inference models, the multiple pre-trained causal inference models are iteratively and interactively optimized.

[0269] In one embodiment, each round of the iterative interaction optimization includes:

[0270] Aggregate the current root cause score vectors of multiple pre-trained causal inference models to obtain a current unified root cause score vector;

[0271] Based on the current unified root cause score vector, construct a root cause consistency constraint term;

[0272] The causal structure of the pre-trained causal inference model is optimized based on the loss function of each model and the root cause consistency constraint to minimize the prediction error.

[0273] In one embodiment, the model training module 200 is further configured to:

[0274] Based on the current root cause score vectors of multiple pre-trained causal inference models, a first unified root cause score vector is determined.

[0275] The second unified root cause score vector is determined based on the sparse attention mechanism, the first unified root cause score vector, and the root cause score vector of each of the pre-trained causal inference models.

[0276] Based on the projection of the causal matrices of the multiple pre-trained causal inference models into the solution space corresponding to the second unified root cause score vector, the second unified root cause score vector is optimized to obtain the current unified root cause score vector.

[0277] In one embodiment, the model training module 200 is further configured to:

[0278] Based on the first unified root cause score vector, the root cause score vector of each of the pre-trained causal inference models, the nonlinear activation function, and the learnable linear mapping parameters, the root cause residual coefficient of each of the pre-trained causal inference models is calculated.

[0279] The root cause residual coefficients are subjected to sparse normalization to obtain the attention weights of each pre-trained causal inference model;

[0280] A second unified root cause score vector is determined based on the current root cause score vectors of multiple pre-trained causal inference models and the attention weights.

[0281] In one embodiment, the model training module 200 is further configured to:

[0282] Construct the causal structure solution space based on the second unified root cause fraction vector;

[0283] Project the causal matrix of each pre-trained causal inference model onto the causal structure solution space to obtain the nearest neighbor mapping point of the causal matrix;

[0284] Based on the nearest neighbor mapping point corresponding to each of the pre-trained causal inference models, the second unified root cause score vector is updated in reverse to obtain the current unified root cause score vector.

[0285] Each module in the aforementioned multi-causal model integrated anomaly root cause tracing device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0286] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 7 As shown, the computer device includes a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements an anomaly root cause tracing method integrating a multi-causal model. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0287] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0288] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the anomaly root cause tracing method of multi-causal model integration of any of the above embodiments:

[0289] Acquire multivariate time-series data of the industrial equipment under abnormal conditions;

[0290] Based on the multivariate time series data, at least three causal inference models using different prediction functions are trained;

[0291] Based on the root cause score vector of each of the causal inference models after training, a unified root cause score vector is determined.

[0292] Based on the unified root cause score vector, the root cause variable is determined.

[0293] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the anomaly root cause tracing method of multi-causal model integration of any of the above embodiments:

[0294] Acquire multivariate time-series data of the industrial equipment under abnormal conditions;

[0295] Based on the multivariate time series data, at least three causal inference models using different prediction functions are trained;

[0296] Based on the root cause score vector of each of the causal inference models after training, a unified root cause score vector is determined.

[0297] Based on the unified root cause score vector, the root cause variable is determined.

[0298] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0299] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0300] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0301] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for tracing the root causes of anomalies using a multi-causal model integration, characterized in that, Applied to industrial equipment, the anomaly root cause tracing method integrating the multi-causal model includes: Acquire multivariate time-series data of the industrial equipment under abnormal conditions; the multivariate time-series data includes one or more time-series measurements of temperature, pressure, vibration, and current; Based on the multivariate time-series data, at least three causal inference models employing different prediction functions are trained. The training includes: determining a first unified root cause score vector based on the current root cause score vectors of multiple pre-trained causal inference models; determining a second unified root cause score vector based on a sparse attention mechanism, the first unified root cause score vector, and the root cause score vector of each of the pre-trained causal inference models; and optimizing the second unified root cause score vector based on the projection of the causal matrix of the multiple pre-trained causal inference models onto the solution space corresponding to the second unified root cause score vector to obtain the current unified root cause score vector. Based on the root cause score vector of each of the causal inference models after training, a unified root cause score vector is determined. Based on the unified root cause score vector, the root cause variable is determined.

2. The anomaly root cause tracing method based on multi-causal model integration according to claim 1, characterized in that, The process of acquiring the multivariate time series data includes: Acquire historical time-series data of the industrial equipment under abnormal conditions; Variables are extracted from the historical time series data to obtain structured time series data; The structured time series data is standardized to obtain the multivariate time series data.

3. The anomaly root cause tracing method based on multi-causal model integration according to claim 1, characterized in that, The training of at least three causal inference models employing different prediction functions based on the multivariate time-series data includes: Based on the multivariate time series data, at least three causal inference models using different prediction functions are pre-trained to obtain multiple pre-trained causal inference models. Based on the multivariate time series data and the root cause score vectors of the multiple pre-trained causal inference models, the multiple pre-trained causal inference models are iteratively and interactively optimized.

4. The anomaly root cause tracing method based on multi-causal model integration according to claim 3, characterized in that, Each round of iterative interaction optimization includes: Aggregate the current root cause score vectors of multiple pre-trained causal inference models to obtain a current unified root cause score vector; Based on the current unified root cause score vector, construct a root cause consistency constraint term; The causal structure of the pre-trained causal inference model is optimized based on the loss function of each model and the root cause consistency constraint to minimize the prediction error.

5. The anomaly root cause tracing method based on multi-causal model integration according to claim 1, characterized in that, The step of determining the second unified root cause score vector based on the sparse attention mechanism, the first unified root cause score vector, and the root cause score vector of each of the pre-trained causal inference models includes: Based on the first unified root cause score vector, the root cause score vector of each of the pre-trained causal inference models, the nonlinear activation function, and the learnable linear mapping parameters, the root cause residual coefficient of each of the pre-trained causal inference models is calculated. The root cause residual coefficients are subjected to sparse normalization to obtain the attention weights of each pre-trained causal inference model; A second unified root cause score vector is determined based on the current root cause score vectors of multiple pre-trained causal inference models and the attention weights.

6. The anomaly root cause tracing method based on multi-causal model integration according to claim 1, characterized in that, The step of optimizing the second unified root cause score vector based on the projection of the causal matrices of the multiple pre-trained causal inference models into the solution space corresponding to the second unified root cause score vector to obtain the current unified root cause score vector includes: Construct the causal structure solution space based on the second unified root cause fraction vector; Project the causal matrix of each pre-trained causal inference model onto the causal structure solution space to obtain the nearest neighbor mapping point of the causal matrix; Based on the nearest neighbor mapping point corresponding to each of the pre-trained causal inference models, the second unified root cause score vector is updated in reverse to obtain the current unified root cause score vector.

7. An anomaly root cause tracing device integrating multiple causal models, characterized in that, The anomaly root cause tracing device, which integrates multiple causal models and is applied to industrial equipment, includes: The data collection module is used to acquire multivariate time-series data of the industrial equipment under abnormal conditions; the multivariate time-series data includes one or more time-series measurements of temperature, pressure, vibration, and current. The model training module is used to train at least three causal inference models employing different prediction functions based on the multivariate time-series data. The training further includes: determining a first unified root cause score vector based on the current root cause score vectors of multiple pre-trained causal inference models; determining a second unified root cause score vector based on a sparse attention mechanism, the first unified root cause score vector, and the root cause score vectors of each pre-trained causal inference model; and optimizing the second unified root cause score vector based on the projection of the causal matrices of the multiple pre-trained causal inference models into the solution space corresponding to the second unified root cause score vector to obtain the current unified root cause score vector. The vector determination module is used to determine a unified root cause score vector based on the root cause score vector of each of the causal inference models after training. The root cause determination module is used to determine the root cause variable based on the unified root cause score vector.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.

Images