Search

CN-117290786-B - Fault detection method based on iterative depth time sequence causal discovery

CN117290786BCN 117290786 BCN117290786 BCN 117290786BCN-117290786-B

Abstract

The invention discloses a fault detection method based on depth time sequence causal discovery, which comprises the steps of obtaining observation values of abnormal performance indexes and potential cause indexes and pairing time stamps thereof, structuring the observation values and the time stamps to form uniform time sequences, constructing an iterative depth time sequence causal discovery network, sampling and optimizing a probability causal graph according to a preset causal threshold value by a causal discovery module, fitting a generation model of the time sequences by a time sequence fitting module, training the iterative depth time sequence causal discovery network by using a plurality of causal threshold values, training by using a structured time sequence, adopting a mask loss function, and attributing fault causes according to training results of the causal threshold values. The method has the advantages that the nonlinear time sequence dynamic relation under the complex environment of the factory can be intelligently learned by utilizing the neural network, the fault root cause analysis process can be rapidly and automatically completed, and the fault root cause analysis process can be easily migrated to the strange environment.

Inventors

  • ZHANG GUOQING

Assignees

  • 中工互联(北京)科技集团有限公司

Dates

Publication Date
20260505
Application Date
20231009

Claims (6)

  1. 1. A fault detection method based on iterative depth time sequence causal discovery is characterized by comprising the following steps: s1, acquiring observation values of fault performance indexes and potential cause indexes and pairing time stamps thereof; s2, carrying out time sequence structuring on the observed value and the time stamp obtained in the step S1 to enable the observed value and the time stamp to form a uniform time sequence, wherein missing data are represented by missing indication bits; S3, constructing an iterative deep time sequence causal discovery neural network, wherein the network comprises a causal discovery module and a time sequence fitting module, and alternately learning a time sequence model and a probability causal graph in the training process, wherein the causal discovery module optimizes the probability causal graph according to the time sequence fitting module learned in the current step, and the optimization process is as follows: Setting probability causal graph A matrix of adjacencies is represented, Representing an s-type function for normalizing the parameters to a range of 0-1; Each element of the probability causal graph is used for representing the probability that a certain potential cause performance index has causal influence on an abnormal performance index; the probability causal graph M is sampled by a micro sampler as follows to obtain a mask causal graph S: ; Wherein, the The uniformity represents the uniformity of distribution, Is a pre-designed parameter which changes from large to small according to an exponential function in the training process; Performing binary mask operation on the input potential cause time sequence; Inputting the binary-masked potential cause time sequence into a time sequence fitting module for completing training, wherein the time sequence fitting module is constructed in S3, and then training is carried out in S4 to obtain a time sequence predicted value; according to the following, inputting a time sequence predicted value and an actual observed value, calculating a causal loss function, and then optimizing a probability causal graph M by using an Adam optimizer; wherein lambda represents a causal threshold, Represents a mean square error loss function, o represents a missing indicator bit, Representing an s-type function; Here, the weights of the time sequence fitting module are optimized according to the probability causal graph learned in the current step, and the optimization process is as follows: Constructing a time sequence fitting module by using a long-short time memory network; The probability causal graph M is subjected to Bernoulli sampling to obtain a mask causal graph S', wherein the Bernoulli sampling is performed according to the following formula to obtain a binary result; ; s' after masking the inputted potential cause time sequence, inputting a model; s4, training the iterative depth time sequence causal discovery network constructed in the step S3 by adopting a plurality of causal thresholds and utilizing the structured time sequence in the step S2, wherein a mask loss function is adopted during training, so that the data marked as missing is eliminated; and S5, attributing the fault reasons according to training results of the causal thresholds.
  2. 2. The method for detecting faults based on iterative deep timing causal discovery of claim 1, wherein in step S2, the step of time-series structuring the observed values of the fault performance index and the potential cause index is as follows: Uniformly dividing a plurality of sampling points in a time range to be detected, and filling observed indexes one by one to sampling points closest in time to obtain a data matrix, wherein the indexes comprise fault performance indexes and potential cause indexes at the same time; Constructing a deletion indicating matrix with the same shape, wherein each deletion indicating bit represents whether the corresponding position of the data matrix is deleted or not; And splicing the data matrix and the missing indication matrix to obtain a three-dimensional data tensor.
  3. 3. The fault detection method based on iterative depth time sequence causal discovery of claim 1, wherein the time sequence fitting module comprises three long and short time memory layers, the hidden dimension of each layer is 256, no leakage is introduced, and weight attenuation with a coefficient of 0.0001 is introduced in the training process.
  4. 4. The method for detecting faults based on iterative depth timing causal discovery of claim 1, wherein for fault performance indicators with continuous values, a mean square error is used as a base loss function, and for fault performance indicators with discrete values, cross entropy is used as a base loss function.
  5. 5. The fault detection method based on iterative deep timing causal discovery of claim 1, wherein in step S5, the learned probability causal graphs M are added and divided by the causal threshold number to obtain an average probability causal graph, and the cause of each fault performance index is ranked and attributed by the average probability causal graph.
  6. 6. The method for fault detection based on iterative depth timing causal discovery of claim 5, wherein said causal threshold is set to 5-10.

Description

Fault detection method based on iterative depth time sequence causal discovery Technical Field The invention belongs to the technical field of intelligent system fault prediction, and particularly relates to a fault detection method based on iterative deep timing causal discovery. Background The fault diagnosis technology is important to the fields of intelligent factories, aerospace, energy management and the like, and can prevent faults and reduce potential safety hazards. While neural networks have been widely used in the field of fault detection, such as convolutional neural network-based fault detection methods, automatic encoder-based fault detection methods, and graph-based fault detection methods. The method based on the graph can process the space structured data, mine the information of nodes and edges, break through in image and video classification, and is also applied to fault detection. The fault detection method based on the graph is often based on specific domain knowledge to construct the graph or utilizes known node connection to construct the graph, however, potential causal relations in a complex system are difficult to fully mine only by expert experience or known information, performance and mobility are limited, and application value is reduced. The causal discovery can excavate complex causal mechanism among the monitoring variables, build a reliable graph model and improve fault detection performance. However, the constraint-based static data causal discovery method is difficult to judge causal directions due to lack of time information, only a suspected causal graph can be obtained, a reliable causal graph cannot be obtained stably, and attribution and importance ordering of faults are limited. Through time sequence causal discovery, nonlinear dynamic relations in complex environments of factories can be intelligently learned by utilizing a neural network, so that a fault root analysis process can be rapidly and automatically completed, and the fault root analysis process can be easily migrated to a strange environment. Disclosure of Invention The invention aims to provide a fault detection method based on iterative depth time sequence causal discovery, so as to solve the problems in the background technology. In order to achieve the above purpose, the invention provides a fault detection method based on iterative depth time sequence causal discovery, which comprises the following steps: S1, acquiring observation values of abnormal performance indexes and potential cause indexes and pairing time stamps thereof; S2, carrying out time sequence structuring on observed values of fault performance indexes and potential cause indexes to form a uniform time sequence, wherein missing data are represented by missing indication bits; S3, constructing an iterative deep time sequence causal discovery network, alternately learning a time sequence model and a probability causal graph M in a training process, and adopting a causal discovery module and a time sequence fitting module to complete the training, wherein the learning of the time sequence model and the learning of the probability causal graph) can be mutually enhanced based on the learning process of the other party, and the final fault detection accuracy is improved; s4, adopting a plurality of causal thresholds, wherein the causal thresholds can be preset empirically, and a series of thresholds can be preset uniformly. Training by using the step S2 method, wherein a mask loss function is adopted during training, so that the data marked as missing is eliminated, and the method can adapt to a time sequence with missing values, namely, the missing part of data does not participate in loss function calculation; and S5, attributing the fault reasons according to training results of the causal thresholds. Preferably, in step S2, the specific steps of time-series structuring the observed values of the fault performance index and the potential cause index are as follows: Uniformly dividing a plurality of sampling points in a time range to be detected, and filling the observed indexes one by one to the sampling point with the closest time to obtain a data matrix, wherein the indexes comprise fault performance indexes and potential cause indexes; Constructing a deletion indicating matrix with the same shape, wherein each number of the data matrix corresponds to one deletion indicating bit at the same position, and each deletion indicating bit represents whether the corresponding position of the data matrix is deleted or not; And splicing the data matrix and the missing indication matrix to obtain a three-dimensional data tensor. In any of the foregoing solutions, preferably, in step S3, the cause and effect discovery module optimizes the probability cause and effect graph according to the time sequence fitting module learned in the current step, where the optimizing process is: Setting a probability causal graph m=σ (θ) to represent a adjacency