A fault detection method based on iterative deep temporal causal discovery

Abstract

This invention discloses a fault detection method based on deep temporal causal discovery, comprising: acquiring observed values ​​of abnormal performance indicators and potential cause indicators and their paired timestamps; structuring the observed values ​​and timestamps to form a uniform time series; constructing an iterative deep temporal causal discovery network, wherein the causal discovery module samples and optimizes the probabilistic causal graph according to a pre-set causal threshold, and the time series fitting module fits the generated model of the time series; using multiple causal thresholds, the iterative deep temporal causal discovery network is trained using the structured time series, with a mask loss function used during training; and attributing the fault causes based on the training results of multiple causal thresholds. The beneficial effects of this invention are that it can utilize neural networks to intelligently learn nonlinear temporal dynamic relationships in complex factory environments, quickly and automatically complete the fault root cause analysis process, and can be easily transferred to unfamiliar environments.

Description

Technical Field

[0001] This invention belongs to the field of intelligent system fault prediction technology, specifically relating to a fault detection method based on iterative deep temporal causal discovery. Background Technology

[0002] Fault diagnosis technology is crucial for fields such as smart factories, aerospace, and energy management, enabling fault prevention and reducing safety hazards. Neural networks have been widely used in fault detection, including convolutional neural network-based, autoencoder-based, and graph-based methods. Graph-based methods, which can process spatially structured data and mine information from nodes and edges, have achieved breakthroughs in image and video classification and are now being applied to fault detection. Graph-based fault detection methods often construct graphs based on domain-specific knowledge or utilize known node connections; however, relying solely on expert experience or known information is insufficient to fully uncover the potential causal relationships in complex systems, limiting performance and transferability, and reducing application value. Causal discovery can uncover complex causal mechanisms between monitored variables, constructing reliable graph models and improving fault detection performance. However, constraint-based static data causal discovery methods, lacking temporal information, struggle to determine causal directions, yielding only speculative causal graphs rather than consistently reliable ones, thus limiting fault attribution and importance ranking. Through temporal causal discovery, neural networks can be used to intelligently learn nonlinear dynamic relationships in complex factory environments, quickly and automatically completing the root cause analysis process and easily transferring it to unfamiliar environments. Summary of the Invention

[0003] The purpose of this invention is to provide a fault detection method based on iterative deep temporal causal discovery to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a fault detection method based on iterative deep temporal causal discovery, comprising the following steps:

[0005] S1: Obtain the observed values ​​of abnormal performance indicators and potential cause indicators and their paired timestamps;

[0006] S2: The observed values ​​of failure performance indicators and potential cause indicators are structured into time series to form a uniform time series, where missing data is represented by missing data indicators;

[0007] S3: Construct an iterative deep temporal causal discovery network, which alternately learns the temporal model and the probabilistic causal graph M during training. This is accomplished using a causal discovery module and a temporal fitting module. The learning of the temporal model and the learning of the probabilistic causal graph can mutually enhance each other based on their learning processes, thereby improving the final fault detection accuracy.

[0008] S4: Multiple causal thresholds can be pre-set based on experience, and a series of thresholds can be uniformly pre-set. Training is performed using the method in step S2, employing a masked loss function to exclude data marked as missing. This method is suitable for time series with missing values, meaning the missing data is not included in the loss function calculation.

[0009] S5: Attribute the causes of failures based on the training results of multiple causal thresholds.

[0010] Preferably, in step S2, the specific steps for structuring the observed values ​​of the fault performance index and potential cause index into a time series are as follows:

[0011] Multiple sampling points are evenly divided within the time range to be detected. The observed indicators are filled into the sampling points with the closest time to obtain a data matrix. The indicators include fault performance indicators and potential cause indicators.

[0012] Construct missing indicator matrices of the same shape, where each number in the data matrix corresponds to a missing indicator bit at the same position, and each missing indicator bit indicates whether the corresponding position in the data matrix is ​​missing.

[0013] By concatenating the data matrix and the missing indicator matrix, a three-dimensional data tensor is obtained.

[0014] In any of the above schemes, preferably, in step S3, the causal discovery module optimizes the probabilistic causal graph based on the time series fitting module learned in the current step, and the optimization process is as follows:

[0015] Set the probabilistic cause-effect graph M = σ(θ) to represent an adjacency matrix, where σ represents an sigmoid function used to normalize the parameters to the range of 0-1;

[0016] Each element (the result of M) of the probabilistic causal graph represents the probability that a certain potential cause performance index has a causal influence on the abnormal performance index.

[0017] The probabilistic causal graph M is sampled by a differentiable sampler according to the following formula to obtain the masked causal graph S:

[0018]

[0019] Where g = -log(-log(u)), u ~ Uniform(0, 1), where Uniform represents a uniform distribution, and τ is a pre-designed parameter that changes from large to small according to an exponential function during training. This parameter needs to be determined empirically or through trial and error, but there are some commonly used preset values, such as changing from 0.1 to 0.01.

[0020] Perform a binary masking operation on the input potential cause time series;

[0021] The potential cause time series after binary masking is input into the time series fitting module after training. The input is to the time series fitting module, which is first built in S3 and then trained in S4 to obtain the time series prediction value.

[0022] Input the time series fitting module with fixed weights to obtain the time series prediction value;

[0023] The causal loss function is calculated according to the following formula, and then the probabilistic causal graph M is optimized using the Adam optimizer;

[0024]

[0025] Where λ represents the causality threshold. σ represents the mean squared error loss function, o represents the missing indicator bit, and σ represents the sigmoid function.

[0026] In any of the above schemes, it is preferred that, in step S3, the weights of the time series fitting module are optimized based on the probabilistic causal graph learned in the current step, and the optimization process is as follows:

[0027] A time series fitting module is constructed using a long short-term memory network;

[0028] The probabilistic causal graph M is passed through a Bernoulli sampler to obtain a masked causal graph S', where the Bernoulli sampler is a sampler that performs sampling according to the following formula to obtain a binary result;

[0029] s′ ij ~Ber(m ij (3),

[0030] After S' performs a masking operation on the input potential cause time series, it is input into the model, which is the time series fitting module.

[0031] After calculating the prediction loss function (2), the model is optimized using the Adam optimizer.

[0032] In any of the above schemes, the designed neural network contains three long short-term memory layers, each with a hidden dimension of 256, without introducing any omissions, and introduces a weight decay with a coefficient of 0.0001 during training.

[0033] In any of the above schemes, the preferred approach is to use the mean squared error as the basic loss function for numerically continuous fault performance indicators, and to use the cross-entropy as the basic loss function for numerically discrete fault performance indicators.

[0034] In any of the above schemes, it is preferred that, in step S5, the learned probability causal graphs M are summed and then divided by the number of causal thresholds to obtain an average probability causal graph, which can be better sorted and attributed.

[0035] By using the average probability causal graph, the causes of each failure performance index are sorted and attributed. The training result is a matrix (average probability causal graph) of size, which is the number of potential cause indexes multiplied by the number of failure performance indexes. For the failure performance index we want to analyze, we take out a column of the corresponding matrix and sort it from largest to smallest according to the causal probability. The order of the corresponding potential cause indexes is the sorting result.

[0036] In any of the above schemes, it is preferred that the causal threshold is set to 5-10.

[0037] The technical effects and advantages of this invention are as follows: The fault detection method based on iterative deep temporal causal discovery employs a causal discovery module and a temporal fitting module. The causal discovery module samples and optimizes the probabilistic causal graph according to a pre-set causal threshold, while the temporal fitting module fits the time series generation model. It can utilize neural networks to intelligently learn the nonlinear temporal dynamic relationships in complex factory environments, quickly and automatically complete the fault root cause analysis process, and can be easily transferred to unfamiliar environments. Attached Figure Description

[0038] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0039] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be noted that these descriptions are for the purpose of aiding understanding the present invention, but do not constitute a limitation thereof. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0040] This invention provides, for example Figure 1 The fault detection method shown includes the following steps: (Iterative deep temporal causal discovery method)

[0041] S1: Obtain the observed values ​​of abnormal performance indicators and potential cause indicators and their paired timestamps, and determine the number of sampling points according to the frequency of the collected time and the estimated rate of causal effect.

[0042] S2: The observed values ​​of fault performance indicators and potential cause indicators are structured into time series to form a uniform time series. Missing data is represented by missing indicator bits. For example, if a bit of the missing indicator matrix is ​​1, it means that the data at the corresponding position has not been observed.

[0043] The specific steps for structuring the observed values ​​of failure performance indicators and potential cause indicators into time series are as follows:

[0044] For the time range to be detected, the sampling points are evenly divided, and the observed indicators are filled into the sampling points with the closest time to obtain the data matrix.

[0045] Construct a missing indicator matrix of the same shape, where each missing indicator bit represents whether the corresponding position in the data matrix is ​​missing;

[0046] By concatenating the data matrix and the missing indicator matrix, a three-dimensional data tensor is obtained.

[0047] S3: Construct an iterative deep temporal causal discovery network, which alternately learns the temporal model and the probabilistic causal graph M during training. This is accomplished using a causal discovery module and a temporal fitting module. In the initial stage of training, the causal discovery module is not trained, and only the temporal fitting module is optimized.

[0048] The causal discovery module optimizes the probabilistic causal graph based on the time series fitting module learned in the current step. The optimization process is as follows:

[0049] Set the probabilistic cause-effect graph M = σ(θ) to represent an adjacency matrix, where σ represents an sigmoid function used to normalize the parameters to the range of 0-1;

[0050] Each element of the probabilistic cause-effect graph represents the probability that a certain potential cause performance index has a causal effect on an abnormal performance index.

[0051] The probabilistic causal graph M is sampled by a differentiable sampler according to the following formula to obtain the masked causal graph S:

[0052]

[0053] Where g = -log(-log(u)), u ~ Uniform(0, 1), where Uniform represents a uniform distribution, and τ is a pre-designed parameter that changes exponentially from large to small during training;

[0054] Perform a binary masking operation on the input potential cause time series;

[0055] Input the time series fitting module with fixed weights to obtain the time series prediction value;

[0056] The causal loss function is calculated according to the following formula, and then the Adam optimizer is used to optimize the probabilistic causal graph M. No acyclic constraints are imposed on the probabilistic causal graph M, and the causal relationship can only flow forward in time.

[0057]

[0058] Where λ represents the causality threshold. σ represents the mean squared error loss function, o represents the missing indicator bit, and σ represents the sigmoid function.

[0059] The parameter θ is used as the optimization variable of the causal discovery module, and after being calculated by the sigmoid function, it represents the probability causal matrix.

[0060] The neural network weights are optimized based on the probabilistic causal graph learned in the current step. The designed neural network contains three long short-term memory layers, each with a hidden dimension of 256. No leakage is introduced, and a weight decay coefficient of 0.0001 is introduced during training. The optimization process is as follows:

[0061] A time series fitting module is constructed using a long short-term memory network;

[0062] The probabilistic causal graph M is passed through a Bernoulli sampler to obtain a masked causal graph S', where the Bernoulli sampler is a sampler that performs sampling according to the following formula to obtain a binary result;

[0063] s′ ij ~Ber(m ij (3),

[0064] S' is used to mask the input potential cause time series before inputting it into the model;

[0065] After calculating the prediction loss function (2), the model is optimized using the Adam optimizer.

[0066] For numerically continuous fault performance indicators, mean squared error is used as the basic loss function; for numerically discrete fault performance indicators, cross-entropy is used as the basic loss function.

[0067] S4: Use 10 causal thresholds and train using the method in step S2, employing a mask loss function during training;

[0068] S5: Based on the training results of 10 causal thresholds, attribute the cause of the failure. For each learned probability causal graph M, sum them up and then divide by the number of causal thresholds to obtain the average probability causal graph.

[0069] The causes of each failure performance index are ranked and attributed using an average probability cause-effect graph.

[0070] This method employs nonlinear Granger causality theory: that is, causality expressed by the following formula:

[0071] f j (x t-τ：t-1，1 ,...,x′ t-τ：t-1，i , ..., x t-τ：t-1 ,N)≠f j (xt-τ：t-1，1 , ..., x t-τ：t-1，i , ..., x t-τ，t-1，N );

[0072] Let τ denote the maximum time lag. In this paper, we focus on discovering causal relationships from X to Y. For a dynamic system, in time series i, when time series x... i When past values ​​of a time series i help predict the future state of a label Y, then the time series i has a Granger causal relationship with the future outcome Y. The standard definition of Granger causality applies to linear relationships but can be extended to nonlinear relationships.

[0073] Nonlinear Granger causality has strong compatibility with neural networks. Considering the universal approximation capability of neural networks, component-wise multilayer perceptrons or long short-term memory networks can be used to fit functions with causal relationships.

[0074] Although Granger causality is not necessarily true causality, it has been shown to support causal conclusions under the assumption of no transient effects and potential confounding factors.

[0075] During the training phase, the time series data collected over a long period of time is structured and used as the training set to train the neural network. The computing platform used for network training is an RTX 3090, and the duration of a single training session is approximately 10 minutes.

[0076] In the initial training phase, all elements of the probability causality matrix are initialized to 0.5;

[0077] To mitigate the overfitting problem of the network, the training process will be terminated if the prediction performance metrics do not improve within a specified period.

[0078] For each batch, data is extracted from the structured time series. Then, predictions are made at each time point by iteratively utilizing the previously collected data.

[0079] To handle data from the initial time and those lacking historical time points, a learnable parameter is configured in the network to derive the initial hidden state, which serves as a substitute for unavailable predecessor data.

[0080] This fault detection method based on iterative deep temporal causal discovery, during its use,

[0081] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A fault detection method based on iterative deep temporal causal discovery, characterized in that: Includes the following steps: S1: Obtain the observed values ​​of failure performance indicators and potential cause indicators and their paired timestamps; S2: Perform time series structuring on the observations and timestamps obtained in step S1 to form a uniform time series. Missing data is represented by missing indicator bits. S3: Construct an iterative deep temporal causal discovery neural network. This network includes a causal discovery module and a temporal fitting module, which alternately learn the temporal model and the probabilistic causal graph during training. The causal discovery module optimizes the probabilistic causal graph based on the temporal fitting module learned in the current step. The optimization process is as follows: Set up a probability cause-effect graph Represent an adjacency matrix. This represents an sigmoid function, used to normalize parameters to the range of 0-1; Each element of the probabilistic cause-effect graph represents the probability that a certain potential cause performance index has a causal effect on an abnormal performance index. The probabilistic causal graph M is sampled by a differentiable sampler according to the following formula to obtain the masked causal graph S: ； in, Uniform indicates a uniform distribution. It is a pre-designed parameter that changes exponentially from large to small during training; Perform a binary masking operation on the input potential cause time series; The potential cause time series after binary masking is input into the time series fitting module after training. This "time series fitting module" is first built in S3 and then trained in S4 to obtain the time series prediction value. According to the following formula, input the time series predicted values ​​and actual observed values, calculate the causal loss function, and then use the Adam optimizer to optimize the probabilistic causal graph M; Where λ represents the causal threshold, This represents the mean squared error loss function, and 'o' represents the missing indicator bit. Represents an sigmoid function; Here, the weights of the time series fitting module are optimized based on the probabilistic causal graph learned in the current step. The optimization process is as follows: A time series fitting module is constructed using a long short-term memory network; The probabilistic causal graph M is passed through a Bernoulli sampler to obtain a masked causal graph S', where the Bernoulli sampler is a sampler that performs sampling according to the following formula to obtain a binary result; ； S' is used to mask the input potential cause time series before inputting it into the model; S4: Using multiple causal thresholds, the iterative deep temporal causal discovery network constructed in step S3 is trained using the structured time series from step S2. During training, a mask loss function is used to exclude data that is marked as missing. S5: Attribute the causes of failures based on the training results of multiple causal thresholds.

2. The fault detection method based on iterative deep temporal causal discovery according to claim 1, characterized in that: In step S2, the step of structuring the observed values ​​of the failure performance index and potential cause index into a time series is as follows: Multiple sampling points are evenly divided within the time range to be detected. The observed indicators are filled into the sampling points with the closest time to obtain a data matrix. The indicators include both fault performance indicators and potential cause indicators. Construct a missing indicator matrix of the same shape, where each missing indicator bit represents whether the corresponding position in the data matrix is ​​missing; By concatenating the data matrix and the missing indicator matrix, a three-dimensional data tensor is obtained.

3. The fault detection method based on iterative deep temporal causal discovery according to claim 1, characterized in that: The time-series fitting module contains three long short-term memory layers, each with a hidden dimension of 256, without introducing any omissions, and introduces a weight decay with a coefficient of 0.0001 during training.

4. The fault detection method based on iterative deep temporal causal discovery according to claim 1, characterized in that: For numerically continuous fault performance indicators, mean squared error is used as the basic loss function; for numerically discrete fault performance indicators, cross-entropy is used as the basic loss function.

5. The fault detection method based on iterative deep temporal causal discovery according to claim 1, characterized in that: In step S5, the learned probability causal graphs M are summed and then divided by the number of causal thresholds to obtain the average probability causal graph; the causes of each failure performance index are sorted and attributed using the average probability causal graph.

6. The fault detection method based on iterative deep temporal causal discovery according to claim 5, characterized in that: The causal threshold is set to 5-10.

Images