Root fault positioning device and method applied to industrial system

Abstract

The invention provides a root fault positioning scheme applied to an industrial system. In the scheme of the invention, confusion factors corresponding to initialized unobservable equipment in a preset historical time period and causal relationship diagrams corresponding to all the equipment are configured as a model of learnable parameters, and the model is subjected to multi-round iterative training until convergence so as to obtain the causal relationship diagrams of all the equipment; and then obtaining a root fault positioning result of the fault equipment in the industrial system when the fault occurs based on the causal relationship graph corresponding to all the equipment in the industrial system. According to the technical scheme of the invention, the more accurate causal relationship among the devices in the industrial system can be obtained, so that the reliability of the root fault positioning result of the industrial system is improved, and the problem of false alarm or missing report in root fault positioning is avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of root fault localization, in particular, to a root fault localization technology based on machine learning for faults occurring in an industrial system in the field of root fault localization, and more particularly, to a root fault localization device and method applied to an industrial system. BACKGROUND

[0002] Root fault localization technology focuses on identifying the internal causes of industrial system failures, which is an important guarantee for the stability and availability of industrial systems. Root fault localization is widely used in many aspects of industrial systems in many fields, such as telecommunications, IT operations, industrial control processes, etc. However, there are complex component dependencies and interactions in the industrial systems of these fields, and these industrial systems are large and complex, and they are prone to failure, which can result in significant economic losses and a decline in user experience. Therefore, how to get reliable root fault localization results when a fault occurs in an industrial system is the focus of the industry and academia.

[0003] The goal of root fault localization is to provide engineers with reliable and robust root fault localization results when an industrial system fails, so that the fault in the industrial system can be resolved in a timely manner, thereby avoiding meaningless work caused by false positives due to noise. With the development of causal inference and trusted machine learning technology, recent work in the field of root fault localization has focused on building a causal graph to represent the architecture of an industrial system and using causal discovery methods to strengthen the root fault localization process. In the causal graph constructed, nodes represent devices in the industrial system (the observation data of which is device state data, performance indicators, etc. collected by sensors deployed on the devices), and edges represent the causal relationships between devices. In many scenarios, some devices that require human operation in the industrial system are often overlooked or not monitored, and these devices are referred to as unobservable devices. However, existing root fault localization techniques are mostly based on strong causal relationship assumptions, i.e., using machine learning models to obtain a causal graph of all devices in the industrial system without considering the impact of unobservable devices on observable devices, which can overestimate or underestimate the causal relationships between devices, resulting in a false causal graph that further affects the reliability of the root fault localization results of the industrial system, and can result in false positives or false negatives in root fault localization.

[0004] In summary, the existing root fault positioning technology is mostly based on an overstrong causal relationship assumption, that is, a causal relationship graph corresponding to all devices in an industrial system is obtained by using a machine learning model without considering the influence of unobservable devices on observable devices, which makes the model possibly overestimate or underestimate the causal relationship between devices, thereby obtaining a false causal relationship graph to further affect the reliability of the root fault positioning result of the industrial system, and problems such as false alarm or missed alarm in root fault positioning may occur.

[0005] It should be noted that the background technology is only used to introduce the related information of the present application, so as to help understand the technical solutions of the present application, but does not mean that the related information must be the prior art. In the absence of evidence that the related information has been disclosed before the filing date of the present application, the related information should not be regarded as prior art. SUMMARY

[0006] Therefore, the purpose of the present application is to overcome the defects of the prior art, and to provide a root fault positioning device and method applied to an industrial system.

[0007] The purpose of the present application is achieved by the following technical solutions:

[0008] According to a first aspect of the present application, a root fault positioning device applied to an industrial system is provided, the industrial system comprising a plurality of observable devices and a plurality of unobservable devices, and each observable device is deployed with a sensor for collecting observation data, the device comprising: a causal relationship module, configured to use the observation data in a preset historical time period of each observable device as a training sample to perform multi-round iterative training on a preset structured causal model until convergence to obtain a causal relationship graph corresponding to all devices in the industrial system, wherein the preset structured causal model is a model configured with an initialized corresponding confusion factor of the unobservable device in the preset historical time period and a causal relationship graph corresponding to all devices as learnable parameters, and the causal relationship graph indicates the causal relationship between each device in the industrial system; a root fault positioning module, configured to obtain a root fault positioning result of a fault device in the industrial system at the time of fault occurrence based on the causal relationship graph corresponding to all devices in the industrial system.

[0009] In some embodiments of the application, the learnable parameters in the preset structured causal model further include initialized self parameters and weights corresponding to the observation data in each preset historical time period of each observable device, wherein the preset structured causal model takes the observation data in each preset historical time period of each observable device as input, and outputs probability values corresponding to the predicted observation data of each observable device at each time point, and the preset structured causal model includes: a random number generation unit configured to initialize the corresponding confounding factors in each preset historical time period of each unobservable device and a causal relationship graph corresponding to all devices; a sample weight adaptive unit configured to initialize the weights corresponding to the observation data in each preset historical time period of each observable device; a causal relationship unit configured to obtain the predicted observation data of each observable device at each time point based on the self parameters, the observation data in each preset historical time period of each observable device and the corresponding weights, the corresponding confounding factors in each preset historical time period of each unobservable device and the causal relationship graph corresponding to all devices; and a probability output unit configured to obtain the probability values corresponding to the predicted observation data of each observable device at each time point by using a statistical analysis method based on the predicted observation data of each observable device at each time point.

[0010] In some embodiments of the application, the random number generation unit is configured to obtain the initialized corresponding confounding factors in each preset historical time period of each unobservable device and the causal relationship graph corresponding to all devices in the following manner: initializing the mean and variance of the Gaussian distribution to which the corresponding confounding factors in each preset historical time period of each unobservable device conform, and performing Gaussian sampling on the Gaussian distribution to which the corresponding confounding factors in each preset historical time period of each unobservable device conform to obtain the initialized corresponding confounding factors in each preset historical time period of each unobservable device; and initializing the mean and variance of the Gaussian distribution to which the causal relationship between each node conforms, and performing Gaussian sampling on the Gaussian distribution to which the causal relationship between each node conforms to obtain the initialized causal relationship graph corresponding to all devices.

[0011] In some embodiments of the application, the causal relationship between each node includes the existence of an edge between each node and the direction of the edge, and the random number generation unit is further configured to obtain the initialized causal relationship graph corresponding to all devices in the following manner: initializing the mean and variance of the Gaussian distribution to which the existence of an edge between each node conforms, and performing Gaussian sampling on the Gaussian distribution to which the existence of an edge between each node conforms to obtain the initialized existence of an edge between each node; initializing the mean and variance of the Gaussian distribution to which the direction of the edge between each node conforms, and performing Gaussian sampling on the Gaussian distribution to which the direction of the edge between each node conforms to obtain the initialized direction of the edge between each node; and obtaining the initialized causal relationship graph corresponding to all devices based on the initialized existence of an edge between each node and the direction of the edge.

[0012] In some embodiments of the present application, the causality unit is configured to obtain the predicted observation data of each observable device at each time point in the following manner:

[0013]

[0014] i∈(1, d), j∈(1, d), k∈(1, r)

[0015] wherein, represents the predicted observation data of the jth observable device at the tth time point, d is the number of observable devices, r is the number of unobservable devices, f obs,j (·) is an observation factor function, M ij represents the causal relationship between the ith observable device and the jth observable device, represents the nonlinear mapping of the observation data of the ith observable device at the tth time point under the parameter of the causality unit Ω, f conf，j (·) is a confounding factor function, M kj represents the causal relationship between the kth unobservable device and the jth observable device, represents the nonlinear mapping of the confounding factor corresponding to the kth unobservable device at the tth time point under the parameter of the causality unit Ω, f conf，j (·) is a confounding factor function.

[0016] In some embodiments of the present application, the causality module is configured to update the learnable self-parameters in the structured causal model, the confounding factors corresponding to each unobservable device in a preset historical time period of all devices, and the causal relationship graph of all devices according to a first preset loss in each round of iterative training process.

[0017] In some embodiments of the present application, the first preset loss is:

[0018]

[0019] wherein, is the observation fitting likelihood function of the observable device, is the causal relationship graph constraint function, is the confounding factor constraint function corresponding to the unobservable device.

[0020] In some embodiments of the present application,

[0021]

[0022]

[0023]

[0024] wherein d is the number of observable devices, r is the number of unobservable devices, T is a preset historical time period, denotes an expected value under a Gaussian distribution to which M determined according to the existence of edges between devices and the direction of edges is subjected, denotes an expected value under a preset prior distribution p(c t |M), denotes a probability value corresponding to the predicted observation data of the i th observable device at the t th moment obtained under the weight in the structured causal model, the confusion factor c t corresponding to all devices at the t th moment and the causal relationship graph M corresponding to all devices, KL[·] is a divergence function, p(M) denotes a preset prior distribution of the causal relationship graph M corresponding to all devices, q γ，θ (M) denotes a Gaussian distribution to which M determined according to the existence of edges between devices and the direction of edges is subjected, denotes a Gaussian distribution to which the confusion factor c corresponding to the i th unobservable device at the t th moment is subjected under the observation data of all observable devices at the t th moment and the causal relationship graph M corresponding to all devices.

[0025] In some embodiments of the present application, the causal relationship module is configured to: in each round of iterative training process, freeze the weight corresponding to the observation data of each observable device in the preset historical time period in the structured causal model, and perform a preset number of iterative training on the structured causal model, in each iterative training process, update the self parameter in the structured causal model, the confusion factor corresponding to each unobservable device in the preset time period and the causal relationship graph corresponding to all devices according to the second preset loss, and freeze the self parameter in the structured causal model, the confusion factor corresponding to each unobservable device in the preset time period and the causal relationship graph corresponding to all devices obtained after the preset number of iterative training, and perform one training on the structured causal model, and in the training process, update the weight corresponding to the observation data of each observable device in the preset historical time period in the structured causal model according to the second preset function, to obtain the self parameter in the structured causal model, the weight corresponding to the observation data of each observable device in the preset historical time period, the confusion factor corresponding to each unobservable device in the preset time period and the causal relationship graph corresponding to all devices after the iteration training of this round.

[0026] In some embodiments of the present application, the first preset loss is:

[0027]

[0028] wherein, represents the observation likelihood function of the observable device under the weight w corresponding to the observation data of the observable device in the preset historical time period in the structured causal model, represents the causal graph constraint function, represents the confusion factor constraint function corresponding to the unobservable device under the weight w corresponding to the observation data of each observable device in the preset historical time period in the structured causal model.

[0029] In some embodiments of the present application,

[0030]

[0031]

[0032]

[0033] wherein d is the number of observable devices, r is the number of unobservable devices, T is the preset historical time period, w t is the weight corresponding to the observation data of each observable device at time t, represents the expectation value under the Gaussian distribution to which M conforms according to the existence γ and the direction θ of the edge between each device, represents the expectation value under the preset prior distribution p(c t |M), represents the probability value corresponding to the predicted observation data of the i th observable device at time t obtained under the weight Ω in the structured causal model, the confusion factor c t of all unobservable devices at time t, and the causal graph M corresponding to all devices, KL[·] is a divergence function, p(M) represents the preset prior distribution of the causal graph M corresponding to all devices, q γ，θ (M) represents the Gaussian distribution to which M conforms according to the existence γ and the direction θ of the edge between each device, represents the Gaussian distribution to which the confusion factor of the i th unobservable device at time t conforms under the observation data of all observable devices at time t and the causal graph M corresponding to all devices, represents the preset prior distribution of the confusion factor of the i th unobservable device at time t under the causal graph M corresponding to all devices.

[0034] In some embodiments of the present application, the causal graph corresponding to all devices comprises a single-direction edge adjacency matrix and a bidirectional edge adjacency matrix, and the root fault locating module is configured to obtain the root fault of the fault device at the time of the fault in the following manner: calculate the transition probability between each observable device based on the single-direction edge adjacency matrix, and perform a de-biased random walk on all observable devices based on the transition probability between each observable device starting from the fault device to obtain the visit frequency of each observable device within a preset historical time period; sort the probability values corresponding to the predicted observation data of each observable device at the time of the fault to obtain the abnormality score of each observable device at the time of the fault; obtain the fusion abnormal value score of each observable device at the time of the fault based on the abnormality score of each observable device at the time of the fault and the visit frequency of each observable device within the preset historical time period; and sort the fusion abnormal values of each observable device at the time of the fault and obtain the root fault locating result of the fault device in the industrial system at the time of the fault according to a preset standard.

[0035] According to a second aspect of the present application, a root fault locating method applied to an industrial system is provided, the method comprising: in response to the occurrence of a fault, obtaining the observation data of each observable device within a preset historical time period at the time of the fault; and using the device described in the above embodiments to obtain the root fault locating result of the fault device in the industrial system at the time of the fault based on the observation data of each observable device within a preset historical time period at the time of the fault.

[0036] According to a third aspect of the present application, an industrial system is provided, the system comprising:

[0037] a plurality of observable devices and a plurality of unobservable devices, and each observable device is deployed with a sensor for collecting observation data; a device as described in the above embodiments for obtaining the root fault locating result of the fault device in the industrial system at the time of the fault based on the observation data of each observable device within a preset historical time period at the time of the fault.

[0038] Compared with the prior art, the present application has the following advantages:

[0039] The root fault positioning scheme of the application considers the corresponding confusion factors of the unobservable device in the preset historical time period, so that the observation data of the observable device in the preset historical time period is used to perform multi-round iterative training on the structured causal model to convergence, so as to obtain more accurate causal relationship between each device in the industrial system to improve the reliability of the root fault positioning result of the industrial system, and avoid false alarm or missed report problems in root fault positioning. At the same time, the application also reflects the unobservable heterogeneity degree of the observation data with adaptive weight, and adjusts and rewrites the causal relationship discovery process based on the augmented score function (i.e. the second preset loss function), and then obtains the optimal weight of the observation data of the observable device to alleviate the influence of the unobservable heterogeneity of the data, and further obtains more accurate causal relationship between each device in the industrial system to improve the reliability of the root fault positioning result of the industrial system, and avoid false alarm or missed report problems in root fault positioning. BRIEF DESCRIPTION OF DRAWINGS

[0040] The embodiments of the application will be further described below with reference to the accompanying drawings, in which:

[0041] Figure 1 It is a simplified version of the industrial system framework for controlling test schematic diagram;

[0042] Figure 2 It is a root fault positioning method flowchart applied in the industrial system according to the embodiment of the application;

[0043] Figure 3 It is an ablation experiment result schematic diagram of the root fault positioning scheme applied in the industrial system according to the embodiment of the application. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical scheme and advantages of the application more clear, the application will be further described in detail below through specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the application and not to limit the application.

[0045] As described in the background, most of the existing root fault positioning techniques are based on strong causal relationship assumptions, that is, the causal relationship graph of all devices in the industrial system is obtained by using a machine learning model without considering the influence of unobservable devices on observable devices, which may overestimate or underestimate the causal relationship between devices, resulting in a false causal relationship graph that further affects the reliability of the root fault positioning result of the industrial system, and may cause false alarm or missed report problems in root fault positioning.

[0046] Some data engineering techniques (i.e., from the data level) have been adopted in existing research to solve the false alarm or missed alarm problem in root cause localization, such as manually collecting more data of some devices to increase the observable indicators of the devices (i.e., increase the observable data), and extending the cognitive domain by gradually drilling down based on indicators, tracking, and log data. However, in real-world scenarios, there are always unobserved devices in industrial systems, and it is also impossible to obtain all the data at each level of the industrial system. Therefore, it is difficult to completely solve the false alarm or missed alarm problem in root cause localization from the data level.

[0047] The inventors have found that the influence of unobserved devices on observable devices constitutes the confounding factors corresponding to the unobserved devices in the root cause localization process. If the influence of the confounding factors corresponding to the unobserved devices is ignored, false causal relationships between devices in the industrial system may be obtained to further affect the reliability of the root cause localization result of the industrial system, and problems such as false alarm or missed alarm in root cause localization may occur. However, due to the constraints of real-world scenarios and the limitations of statistics, this problem is difficult to be alleviated. On the one hand, the traditional structured causal model does not consider the existence of confounding factors, which may cause model setting bias in real-world scenarios. On the other hand, even if some works introduce additional statistical rules to mine the information of the confounding factors corresponding to the unobserved devices, the huge computational overhead makes such methods unsuitable for root cause localization tasks.

[0048] To solve the above problems, the inventors propose a new root cause localization scheme from the model level, which is also called PORCA (Partially Observed Root Causes Analysis). The scheme reconstructs a structured causal model with strong hypothesis restrictions to avoid potential model setting bias problems. The scheme constructs the model by pre-setting the confounding factors corresponding to each unobserved device in a pre-set historical time period as learnable parameters in the model, so that the pre-set structured causal model is trained for multiple rounds of iteration to convergence by taking the observation data of each observable device in the pre-set historical time period as the training sample, thereby obtaining more accurate causal relationships between devices in the industrial system to further improve the reliability of the root cause localization result of the industrial system and avoid false alarm or missed alarm problems in root cause localization.

[0049] Before the embodiments of the present application are specifically introduced, the basic concepts related to the unobserved devices, unobserved heterogeneity, and false causal relationship graph involved in the present application are introduced in conjunction with the examples of Figure 1 Figure 1 ​As shown in FIG. 1, it shows a simplified version of an industrial system for control test, which includes a water pump, an input valve, a water tank, a chemical tank and a micro actuator, etc. The unobservable device includes a satellite actuator, and the observable device includes the water pump, the input valve, the water tank and the chemical tank.

[0050] From Figure 1 As can be seen from (a), the micro actuator will operate the state of the downstream water pump. In many scenarios, many entities like this related to human operation (i.e. the micro actuator) will be frequently ignored or not monitored during the analysis of the industrial system, and these devices are referred to as unobservable devices. From Figure 1 As can be seen from (a), the occurrence of unobservable heterogeneity such as system failure or potential attack will cause the change of the observation data distribution of the observable device in the industrial system, and these changes are referred to as unobservable heterogeneity.

[0051] Five devices in the industrial control test system, i.e. the water pump, the input valve, the water tank, the chemical tank and the micro actuator, are represented as five nodes, i.e. nodes A, B, C, D and E, from Figure 1 As can be seen from (b), there are confusion factors corresponding to unobservable devices and unobservable heterogeneity in the industrial system, and if the influence of the confusion factors corresponding to unobservable devices and the unobservable heterogeneity is not considered, it will lead to the formation of false causal relationship between nodes during the causal discovery between devices in the industrial system, for example, the causal relationship between nodes B and C is a false causal relationship, which further leads to the false causal relationship graph (i.e. the false discovered causal relationship graph) of the causal relationship graph corresponding to nodes A, B, C, D and E.

[0052] Therefore, in order to fully consider the influence of unobservable devices on observable devices (i.e. consider the confusion factors corresponding to unobservable devices), the present application provides a root fault positioning device applied to an industrial system, the industrial system comprising a plurality of observable devices and a plurality of unobservable devices, and each observable device is deployed with a sensor for collecting observation data, the device comprising: a causal relationship module, configured to perform multi-round iteration training on a preset structured causal model until convergence with the observation data of each observable device in a preset historical time period as a training sample to obtain a causal relationship graph corresponding to all devices in the industrial system, wherein the preset structured causal model is a model configured with the corresponding confusion factors of unobservable devices in a preset historical time period and the causal relationship graph corresponding to all devices as learnable parameters, and the causal relationship graph indicates the causal relationship between devices in the industrial system; a root fault positioning module, configured to obtain a root fault positioning result of a fault device in the industrial system at the time of fault occurrence based on the causal relationship graph corresponding to all devices in the industrial system. Each module in the root fault positioning device of the present application will be described in detail below.

[0053] The various modules of the device of the present application will be described in detail below in conjunction with the accompanying drawings and examples.

[0054] I. Causal relationship module

[0055] The causal relationship module is configured to perform multi-round iterative training on the preset structured causal model using observation data of each observable device in a preset historical time period as a training sample until convergence is achieved to obtain a causal relationship graph corresponding to all devices in the industrial system, wherein the preset structured causal model is a model configured with an initialized corresponding confounder of each unobservable device in the preset historical time period and a causal relationship graph corresponding to all devices as learnable parameters, and the causal relationship graph indicates the causal relationship between each device in the industrial system.

[0056] According to an embodiment of the present application, the observation data of each observable device in a preset historical time period, the corresponding confounder of each unobservable device in a preset time period, and the causal relationship graph corresponding to all devices are subject to a preset joint probability distribution. For example, let x represent the observation data of each observable device in a preset historical time period, c represent the corresponding confounder of each unobservable device in a preset historical time period, and M represent the causal relationship graph corresponding to all devices, then the joint probability distribution of x, c, and M is decoupled according to Bayes' theorem as follows:

[0057]

[0058] where T is the preset historical time period, p(M) is the preset prior distribution of the causal relationship graph M corresponding to all devices, p(x t ,c t |M) represents the joint probability distribution of the observation data set x t and the confounder c t of all unobservable devices at time t under the causal relationship graph M corresponding to all devices.

[0059] According to one embodiment of the present application, the learnable parameters in the preset structured causal model further include initialized self parameters and weights corresponding to observation data in a preset historical time period of each observable device, wherein the preset structured causal model takes observation data in a preset historical time period of each observable device as input, and outputs probability values corresponding to predicted observation data of each observable device at each time point, and the preset structured causal model includes: a random number generation unit configured to initialize corresponding confounders in a preset historical time period of each unobservable device and a causal relationship graph corresponding to all devices, wherein the causal relationship graph indicates causal relationships between devices; a sample weight adaptive unit configured to initialize weights corresponding to observation data in a preset historical time period of each observable device; a causal relationship unit configured to obtain predicted observation data of each observable device at each time point based on self parameters, observation data in a preset historical time period of each observable device and corresponding weights, corresponding confounders in a preset historical time period of each unobservable device, and a causal relationship graph corresponding to all devices; and a probability output unit configured to obtain probability values corresponding to predicted observation data of each observable device at each time point by using a statistical analysis method based on predicted observation data of each observable device at each time point. Each unit in the preset structured causal model is described in detail below.(1) Random number generation unit

[0060] The random number generation unit is configured to initialize corresponding confounders in a preset historical time period of each unobservable device and a causal relationship graph corresponding to all devices.

[0061] In order to make corresponding confounders in a preset historical time period of each unobservable device and a causal relationship graph corresponding to all devices have good mathematical properties and be easy to process in the training process of the structured causal model, and thus simplify the calculation process, the posterior distribution of corresponding confounders in a preset historical time period of each unobservable device and a causal relationship graph corresponding to all devices is parameterized as a Gaussian distribution. The initialization method of corresponding confounders in a preset historical time period of each unobservable device and a causal relationship graph corresponding to all devices is described in detail below.

[0062] (1) Initialize corresponding confounders in a preset historical time period of each unobservable device

[0063] According to one embodiment of the present application, the random number generation unit is configured to obtain initialized corresponding confounders in a preset historical time period of each unobservable device in the following manner: by initializing a mean value and a variance of a Gaussian distribution to which the corresponding confounders in a preset historical time period of each unobservable device conform, and performing Gaussian sampling on the Gaussian distribution to which the corresponding confounders in a preset historical time period of each unobservable device conform to obtain the initialized corresponding confounders in a preset historical time period of each unobservable device.

[0064] For example, for the confusion factor c corresponding to the rth unobservable device in the industrial system, the posterior distribution is parameterized as the following Gaussian distribution:

[0065]

[0066] wherein μ k and respectively represent the mean and variance of the Gaussian distribution to which the confusion factor corresponding to the rth unobservable device is subjected.

[0067] (2) Initialize the causal relationship graph corresponding to all devices

[0068] According to an embodiment of the present application, the random number generation unit is configured to obtain the initialized causal relationship graph corresponding to all devices in the following manner: taking each device as a node, initializing the mean and variance of the Gaussian distribution to which the causal relationship between the nodes is subjected, and performing Gaussian sampling on the Gaussian distribution to which the causal relationship between the nodes is subjected to obtain the initialized causal relationship graph corresponding to all devices. According to another embodiment of the present application, the causal relationship between the nodes includes the existence of the edge between the nodes and the direction of the edge, and the random number generation unit is further configured to obtain the initialized causal relationship graph corresponding to all devices in the following manner: initializing the mean and variance of the Gaussian distribution to which the existence of the edge between the nodes is subjected, and performing Gaussian sampling on the Gaussian distribution to which the existence of the edge between the nodes is subjected to obtain the initialized existence of the edge between the nodes; initializing the mean and variance of the Gaussian distribution to which the direction of the edge between the nodes is subjected, and performing Gaussian sampling on the Gaussian distribution to which the direction of the edge between the nodes is subjected to obtain the initialized direction of the edge between the nodes; obtaining the initialized causal relationship graph corresponding to all devices based on the initialized existence of the edge between the nodes and the direction of the edge.

[0069] For example, for the causal relationship graph M corresponding to d observable devices and r unobservable devices in the industrial system, the posterior distribution of the causal relationship between the pth device and the qth device is parameterized in the following form:

[0070] q γ，θ (M pq )~Bern(sigmoid(γ pq )·sigmoid(θ pq ))

[0071]

[0072]

[0073] p e (1, d + r), q e (1, d + r)

[0074] where Bern(·) is a Bernoulli distribution function, sigmoid(·) is a normalization function, γ pq represents the existence of an edge between the pth device and the qth device, θ pq represents the direction of an edge between the pth device and the qth device, and respectively represent the mean and variance of the Gaussian distribution to which the existence of an edge between the pth device and the qth device is subject, and respectively represent the mean and variance to which the direction of an edge between the pth device and the qth device is subject.

[0075] (ii) Sample weight adaptive unit

[0076] The sample weight adaptive unit is configured to initialize the weight corresponding to the observation data of each observable device in a preset historical time period.

[0077] (iii) Causal relationship unit

[0078] The causal relationship unit is configured to obtain the predicted observation data of each observable device at each time based on its own parameters, the observation data and the corresponding weight of each observable device in a preset historical time period, the corresponding confusion factor of each unobservable device in a preset historical time period, and the causal relationship graph corresponding to all devices.

[0079] For example, for an industrial system with d observable devices and r unobservable devices, if it is required to locate the root fault of a fault occurring in the industrial system in a T historical time period, then wherein, represents the observation data of the dth observable device in the T historical time period, wherein, represents the corresponding confusion factor of the rth unobservable device in the T historical time period. According to an embodiment of the present application, the causal relationship unit is configured to obtain the predicted observation data of each observable device at each time in the following manner:

[0080]

[0081] i e (1, d), j e (1, d), k e (1, r)

[0082] wherein, represents the predicted observation data of the jth observable device at the tth time, d is the number of observable devices, r is the number of unobservable devices, f obs，j (·) is an observation factor function, Mij represents the causal relationship between the i-th observable device and the j-th observable device, represents the nonlinear mapping of the observation data of the i-th observable device at time t under the parameter of the causal relationship unit, f conf，j (·) is a confounder function, M kj represents the causal relationship between the k-th unobservable device and the j-th observable device, represents the nonlinear mapping of the confounder corresponding to the k-th unobservable device at time t under the parameter of the causal relationship unit, f conf，j (·) is a confounder function. The functions designed in the formula of the causal relationship unit can all be realized by MLP (i.e., multi-layer perceptron).

[0083] (Four) Probability Output Unit

[0084] The probability output unit is used to obtain the probability value corresponding to the predicted observation data of each observable device at each time based on the predicted observation data of each observable device at each time by using a statistical analysis method.

[0085] In order to reduce the calculation overhead of the present application in training the resulting causal model, the loss function used in the training model process of the present application is an augmented score function, so that in the case where the confounder corresponding to the unobservable device exists, the discovery process of the causal relationship can still be efficiently optimized in the form of gradient descent.

[0086] According to an embodiment of the present application, the causal relationship module is configured to: in each round of iterative training process, update the learnable self-parameters in the structured causal model, the confounder corresponding to each unobservable device in a preset historical time period of each unobservable device, and the causal relationship graph corresponding to all devices according to a first preset loss. Wherein, the augmented score function is the first preset loss, and the first preset loss is:

[0087]

[0088] Wherein, is the observation fitting likelihood function of the observable device, is the causal relationship graph constraint function, is the confounder constraint function corresponding to the unobservable device. And:

[0089]

[0090]

[0091]

[0092] Wherein, d is the number of observable devices, r is the number of unobservable devices, T is a preset historical time period, represents an expected value under a Gaussian distribution to which M determined according to the existence of edges between devices and the direction of edges obeys, represents an expected value under a preset prior distribution p(c t represents an expected value under a preset prior distribution p(c represents a weight in the structured causal model, a confusion factor c t corresponding to the i th observable device at time t obtained under the causal relationship graph M corresponding to all devices, and a divergence function, p(M) represents a preset prior distribution of the causal relationship graph M corresponding to all devices, q γ,θ (M) represents a Gaussian distribution to which M determined according to the existence of edges between devices and the direction of edges obeys, represents a Gaussian distribution to which the confusion factor c represents a preset prior distribution of the confusion factor c

[0093] It can be known from the above embodiment that, by initializing the confusion factor corresponding to the unobservable device within a preset historical time period in the structured causal model, and using the observation data of the observable device to perform multiple iterative training on the structured causal model, and combining the first preset loss function to constantly update the parameters in the structured causal model, the confusion factor corresponding to each unobservable device within a preset historical time period, and the causal relationship graph corresponding to all devices. Compared with the prior art, which does not consider the influence of unobservable devices on observable devices (the influence of unobservable devices on observable devices constitutes the confusion factor corresponding to the unobservable device in the root fault positioning process), the root fault positioning scheme of the present application considers the confusion factor corresponding to the unobservable device within a preset historical time period, so that the structured causal model is trained using the observation data of the observable device within a preset historical time period for multiple rounds until convergence, thereby obtaining more accurate causal relationship between devices in the industrial system to improve the reliability of the root fault positioning result of the industrial system, and avoid false alarm or missed report problems in root fault positioning.

[0094] In addition, the inventors have found through research that the occurrence of unobservable heterogeneity such as system failure or potential attack will cause the change of the observation data distribution of the observable devices in the industrial system, so that the observation data of the observable devices has the problem of unobservable heterogeneity, so that the machine learning model is difficult to identify the observation data with unobservable heterogeneity, which will also cause false causal relationships between devices in the root fault positioning process to further affect the reliability of the root fault positioning result of the industrial system, and false alarms or false negatives may occur in root fault positioning. However, since the heterogeneity is often invisible to engineers, it is still difficult to alleviate the impact of unobservable heterogeneity in the root fault positioning task. In some previous work practices, manual labeling or rule-based change point detection methods are often used to additionally label information about the heterogeneity of the observation data of the observable devices. However, since the fault types in the real scene are often diverse and the attack methods are carefully designed, the above methods face the problems of excessive resource overhead and lack of flexibility.

[0095] In order to alleviate the impact of unobservable heterogeneity on device observation data, the present application flexibly distinguishes the unobservable heterogeneity of the observation data of the observable devices through the degree of violation of the causal mechanism in the data reconstruction process, that is, after obtaining the probability value corresponding to the predicted observation data of the observable devices through the structured causal model, the degree of unobservable heterogeneity of the observation data is judged based on the probability value, and the unobservable heterogeneity degree of the observation data is reflected by the adaptive weight (that is, a smaller weight is given to a larger probability value, and a larger weight is given to a smaller probability value), and then the causal relationship discovery process between devices in the industrial system is adjusted through the continuous updating of the weight corresponding to the observation data. Based on the adaptive weight, the present application adjusts and rewrites the causal relationship discovery process based on the augmented score function (i.e., the second preset loss function), and then obtains the optimal weight corresponding to the observation data of the observable devices to alleviate the impact of unobservable heterogeneity of the data.

[0096] According to one embodiment of the present application, the causal relationship module is configured to: in each round of iterative training process, freeze the weight corresponding to the observation data of each observable device in a preset historical time period in the structured causal model and perform a preset number of iterative training on the structured causal model, in each iterative training process, update the self parameters in the structured causal model, the corresponding confounding factors of each unobservable device in a preset time period and the causal relationship graph corresponding to all devices according to a second preset loss, and freeze the self parameters in the structured causal model, the corresponding confounding factors of each unobservable device in a preset time period and the causal relationship graph corresponding to all devices obtained after a preset number of iterative training and perform one training on the structured causal model, and in the training process, update the weight corresponding to the observation data of each observable device in a preset historical time period in the structured causal model according to a second preset function, to obtain the self parameters in the structured causal model after this round of iterative training, the weight corresponding to the observation data of each observable device in a preset historical time period, the corresponding confounding factors of each unobservable device in a preset time period and the causal relationship graph corresponding to all devices. According to one embodiment of the present application, the second preset loss is:

[0097]

[0098] wherein, represents the observation fitting likelihood function of the observable device under the weight w corresponding to the observation data of each observable device in a preset historical time period in the structured causal model, represents the causal relationship graph constraint function, represents the confounding factor constraint function corresponding to the unobservable device under the weight w corresponding to the observation data of each observable device in a preset historical time period in the structured causal model. And:

[0099]

[0100]

[0101]

[0102] wherein, d is the number of observable devices, r is the number of unobservable devices, T is a preset historical time period, w t is the weight corresponding to the observation data of each observable device at time t, represents the expected value under the Gaussian distribution to which M conforms according to the existence of the edge between the devices γ and the direction of the edge θ, represents the expected value under the preset prior distribution p(c t |M), Ω, all the confounding factors c corresponding to all the unobservable devices at time t t and the probability value of the i-th observable device at time t under the obtained causal graph M corresponding to all the devices, KL[·] is a divergence function, p(M) represents a preset prior distribution of the causal graph M corresponding to all the devices, q γ，θ (M) represents a Gaussian distribution to which M is subjected according to the existence γ and the direction θ of the edge between the devices, represents a Gaussian distribution to which the confounding factor c corresponding to the i-th unobservable device at time t under the observation data of all the observable devices at time t and the causal graph M corresponding to all the devices is subjected, represents a preset prior distribution of the confounding factor c corresponding to the i-th unobservable device at time t under the causal graph M corresponding to all the devices.

[0103] Further, the optimal weight corresponding to the observation data of each observable device is solved by the following double-layer optimization:

[0104]

[0105]

[0106] wherein τ is a preset control factor, used for controlling the value range of the weight corresponding to the observation data of each observable device, in the above two-layer optimization cycles, the purpose of the inner loop is to minimize the function of the second preset loss to obtain the optimal weight value, and the purpose of the outer loop is to optimize the parameters of the augmented structural causal model.

[0107] According to the relevant prior knowledge, wherein α and ρ are optimized under an augmented Lagrangian framework, the coefficient λ controls the sparsity degree of the edge in the relationship graph of all the devices, h(·) is a discretization constraint function, and wherein The causal graph M corresponding to all the devices includes a single-direction edge adjacency matrix D and a bidirectional edge adjacency matrix B, trace(·) is a trace function, and sum(·) is a sum function.

[0108] It can be known from the above embodiment that, compared with the prior art in which the information about the observation data heterogeneity of the devices is additionally labeled by using a manual labeling manner or a change point detection manner based on rules, the present application reflects the unobservable heterogeneity degree of the observation data by using adaptive weights in the training process of the structured causal model, so that the model can pay more attention to the observation data with higher unobservable heterogeneity. The scheme of the present application can flexibly label the observation data with unobservable heterogeneity, and the resource consumption is small.

[0109] To obtain a more accurate causal relationship graph corresponding to all devices in the industrial system, thereby improving the accuracy of root fault positioning results and avoiding false alarms or missed reports in root fault positioning as much as possible.

[0110] II. Root fault positioning module

[0111] The root fault positioning module is configured to obtain a root fault positioning result of a fault device in the industrial system at the time of fault occurrence based on a causal relationship graph corresponding to all devices in the industrial system.

[0112] According to an embodiment of the present application, the causal relationship graph corresponding to all devices includes a one-way edge adjacency matrix and a two-way edge adjacency matrix, and the root fault positioning module is configured to obtain the root fault of the fault device at the time of fault occurrence in the following manner: calculate the transition probability between each observable device based on the one-way edge adjacency matrix, and perform a de-biased random walk on all observable devices based on the transition probability between each observable device starting from the fault device to obtain the visit frequency of each observable device within a preset historical time period; sort the probability value corresponding to the predicted observation data of each observable device at the time of fault occurrence to obtain the abnormality score of each observable device at the time of fault occurrence; obtain the fusion abnormal value score of each observable device at the time of fault occurrence based on the abnormality score of each observable device at the time of fault occurrence and the visit frequency of each observable device within the preset historical time period; and sort the fusion abnormal value of each observable device at the time of fault occurrence and obtain the root fault positioning result of the fault device in the industrial system at the time of fault occurrence according to a preset standard.

[0113] According to an embodiment of the present application, the root fault positioning module is further configured to obtain the transition probability between each observable device in the following manner:

[0114]

[0115] wherein φ is a preset hyperparameter, is the transpose of the one-way edge adjacency matrix, represents the causal relationship between the i-th device and the j-th device in the transpose matrix of the one-way edge adjacency matrix.

[0116] According to the causal Markov decomposition theorem, there is Therefore, the abnormality score at the node level can be analyzed using the causal mechanism with autonomy, and the probability value corresponding to the predicted observation data of the observable device is used as the abnormality score at the node level. ​According to one embodiment of the present application, the root fault locating module is further configured to obtain the abnormality score of each observable device at the time of fault occurrence by:

[0117]

[0118] wherein, is the probability value corresponding to the predicted observation data of the i-th observable device T in the preset historical time period, and RANK(·) is a ranking function.

[0119] According to one embodiment of the present application, the root fault locating module is further configured to obtain the fusion anomaly score of each observable device at the time of fault occurrence by:

[0120]

[0121] wherein, represents the normalized access frequency of the i-th observable device in the preset historical time period, represents the normalized abnormality score of the i-th observable device at the time of fault occurrence, and ψ is a preset parameter.

[0122] As can be seen from the above embodiments, by initializing the confounding factors corresponding to the unobservable devices in the preset historical time period and the weights of the observation data of each observable device in the preset historical time period in the structured causal model, and using the observation data of the observable devices in the preset historical time period to perform multiple iterative training on the structured causal model until convergence, and in the process of multiple rounds of iteration, the second preset loss function is used to update the self-parameters in the structured causal model, the weights of the observation data of each observable device in the preset historical time period, the confounding factors corresponding to each unobservable device in the preset time period, and the causal relationship graphs corresponding to all devices by alternating update to obtain more accurate causal relationships between devices in the industrial system to further improve the reliability of the root fault locating result of the industrial system and avoid false alarm or missed report problems in root fault locating.

[0123] After the root fault locating device is constructed by the above embodiments, according to one embodiment of the present application, the present application proposes a root fault locating method applied in an industrial system, the method comprising: in response to the occurrence of a fault, obtaining the observation data of each observable device in a preset historical time period at the time of fault occurrence; using the device in the above embodiments to obtain the root fault locating result of the fault device in the industrial system at the time of fault occurrence based on the observation data of each observable device in the preset historical time period at the time of fault occurrence. The implementation process of obtaining the root fault locating result of the fault device in the industrial system at the time of fault occurrence by using the device in the above embodiments is described in detail below.

[0124] AsFigure 2 As shown, the present application divides the implementation process of the root fault positioning result of the faulty equipment in the industrial system at the time of fault occurrence into:

[0125] (1) Taking the observation data of each observable equipment in a preset historical time period as the training sample, the preset structured causal model is trained for multiple rounds of iteration until convergence to obtain the causal relationship graph corresponding to all the equipment in the industrial system.

[0126] The causal relationship module is configured to perform multiple rounds of iteration training on the structured causal model using the observation data x of the observable equipment in a preset historical time period until convergence, and in each round of iteration training, the augmented score function is used to update the learnable self parameters in the structured causal model, the weights corresponding to the observation data of each observable equipment in a preset historical time period, the confusion factors corresponding to each unobservable equipment in a preset time period, and the causal relationship graph corresponding to all the equipment. Among them, the updating method is divided into two kinds. One is that in each round of iteration training, the first preset loss is used to update the learnable self parameters in the structured causal model, the confusion factors corresponding to each unobservable equipment in a preset historical time period, and the causal relationship graph corresponding to all the equipment. The other is that in each round of iteration training, the weights corresponding to the observation data of each observable equipment in a preset historical time period in the structured causal model are frozen, and the structured causal model is trained for a preset number of times. In each iteration training process, the self parameters in the structured causal model, the confusion factors corresponding to each unobservable equipment in a preset time period, and the causal relationship graph corresponding to all the equipment are updated according to the second preset loss. After the preset number of iterations, the self parameters in the structured causal model, the confusion factors corresponding to each unobservable equipment in a preset time period, and the causal relationship graph corresponding to all the equipment are frozen, and the structured causal model is trained once. In this training process, the weights corresponding to the observation data of each observable equipment in a preset historical time period in the structured causal model are updated according to the second preset function, so as to obtain the self parameters in the structured causal model after this round of iteration training, the weights corresponding to the observation data of each observable equipment in a preset historical time period, the confusion factors corresponding to each unobservable equipment in a preset time period, and the causal relationship graph corresponding to all the equipment.

[0127] From Figure 2 , it can be seen that θ MSB is the learnable self parameters in the structured causal model, the confusion factors corresponding to each unobservable equipment in a preset time period, and the causal relationship graph corresponding to all the equipment. θ HAS is the weights corresponding to the observation data of each observable equipment in a preset historical time period in the structured causal model.

[0128] (2) based on the causal graph corresponding to all devices in the industrial system, the root fault positioning result of the fault device in the industrial system at the time of fault occurrence is obtained.

[0129] The root fault positioning module is configured to obtain the visit frequency of each observable device in a preset historical time period and the abnormality score of each observable device at the time of fault occurrence from the node level fault ranking based on the causal graph corresponding to all devices obtained after training in (1) using the method of debiased random walk, and further fuse the visit frequency of each observable device in the preset historical time period and the abnormality score of each observable device at the time of fault occurrence to obtain the fusion abnormal value score of each observable device at the time of fault occurrence. Finally, the fusion abnormal values of each observable device at the time of fault occurrence are sorted and the root fault positioning result of the fault device in the industrial system at the time of fault occurrence is obtained according to the preset standard.

[0130] In order to verify the effectiveness of the root fault positioning scheme applied in the industrial system proposed in the above embodiments of the application, the inventors use Simulation (i.e. a simulation data set generated according to a fault mechanism), CRACs (i.e. a data set collected by temperature sensors in an IT data center), and SWaT (i.e. a data set collected by sensors and actuators in an intelligent water treatment plant) to perform the following experiments:

[0131] I. Comparative experiment

[0132] (1) Evaluation index

[0133] PR@5 represents the accuracy rate of the top 5 prediction results of the structured causal model, PR@Avg represents the average value of the accuracy rate of the top 5 prediction results, i.e. the value of (PR@1+PR@2+PR@3+PR@4+PR@5) / 5, and RankScore represents the weighted sum value of the accuracy rate of the 5 prediction results. The higher the above evaluation index, the better the performance of the structured causal model. AUC represents the area under the curve calculated by regarding the existence of each edge as a classification problem. The larger the index, the better the performance. SHD (Structural Hamming Distance) represents the number of different edges of the learned causal graph corresponding to all devices and the real causal relationship (or the causal graph corresponding to all devices annotated according to expert experience) corresponding to all devices. The smaller the index, the better the performance.

[0134] (2) Comparative experiment results

[0135] Table 1 shows the comparative experimental results of root fault location using the scheme of the present invention and existing technologies. Existing technologies such as CloudRanger, MicroCause, AutoMAP, and CORAL do not consider the influence of obfuscation factors and unobservable heterogeneity within a preset historical time period for unobservable devices. RCD and RCD* consider the influence of unobservable heterogeneity, while FCI* considers the influence of obfuscation factors within a preset historical time period for unobservable devices. As can be seen from Table 1, the accuracy of root fault location using the scheme of the present invention is superior to that of existing technologies.

[0136] Table 1

[0137]

[0138] As shown in Table 2, it presents the experimental results comparing the causal relationship graphs of all devices obtained by training the structured causal model using the scheme of the present invention with those of the prior art. As can be seen from Table 2, the causal relationship graphs of all devices obtained by using the scheme of the present invention are more consistent with the real situation or expert experience.

[0139] Table 2

[0140]

[0141] II. Ablation Experiment Results

[0142] like Figure 3 As shown, the ablation experiment results of the proposed solution are illustrated. Here, w / o DC represents the root fault location accuracy without considering the confusion factor corresponding to the unobservable devices within a preset historical time period, and w / o SH represents the AUC of the causal relationship graph for all devices without considering the unobservable heterogeneity problem. The robustness of the proposed solution is tested by changing the number of unobservable devices and the proportion of heterogeneity in the observations. The performance of the proposed solution (i.e., PORCA) becomes more stable as the number of unobservable devices increases. Even when the number of heterogeneous observations is less than 20%, the proposed solution still brings a significant performance improvement.

[0143] Compared with the root fault positioning technology in the prior art which does not consider the corresponding confounding factors of the unobservable equipment in the industrial system within a preset time period, the root fault positioning scheme considers the corresponding confounding factors of the unobservable equipment within a preset historical time period, so that the structured causal model is trained by multiple rounds of iteration using the observation data of the observable equipment within the preset historical time period until convergence, thereby obtaining more accurate causal relationships between the equipment in the industrial system to improve the root fault positioning result reliability of the industrial system and avoid false alarm or missed report problems in root fault positioning. Meanwhile, the unobservable heterogeneity of the observation data is reflected by the adaptive weight, and the causal relationship discovery process based on the augmented score function (i.e., the second preset loss function) is adjusted and rewritten, and then the optimal weight of the observation data of the observable equipment is obtained to alleviate the influence of the unobservable heterogeneity of the data, and further obtain more accurate causal relationships between the equipment in the industrial system to improve the root fault positioning result reliability of the industrial system and avoid false alarm or missed report problems in root fault positioning.

[0144] In practical applications, the root fault positioning schemes of the above embodiments are used in industrial systems in many fields, such as telecommunications, IT operation and maintenance, industrial control processes, etc. The device obtained by the above embodiments can be configured in different types of industrial systems to obtain the root fault positioning result of the fault equipment in the industrial system when the fault occurs.

[0145] It should be noted that although the above describes the steps in a specific order, it does not mean that the steps must be performed in the above specific order, in fact, some of the steps can be performed concurrently or even in a changed order, as long as the desired function can be achieved.

[0146] The present application can be a system, a method, and / or a computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions embodied therewith, wherein the computer readable program instructions are used to cause a processor to implement the various aspects of the present application.

[0147] Computer-readable storage media can be tangible devices that hold and store instructions for use by an instruction execution device. Computer-readable storage media can be, for example, including but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof.

[0148] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A root fault location device applied in an industrial system, the industrial system comprising multiple observable devices and multiple unobservable devices, wherein each observable device is equipped with a sensor for collecting observation data, characterized in that, The device includes: The causal relationship module is used to perform multiple rounds of iterative training on a preset structured causal model using observation data within a preset historical time period of each observable device as training samples until convergence, so as to obtain a causal relationship graph corresponding to all devices in the industrial system. The preset structured causal model is a model configured with the initial confusion factor corresponding to the preset historical time period of the unobservable device and the causal relationship graph corresponding to all devices as learnable parameters. The causal relationship graph indicates the causal relationship between devices in the industrial system. The root fault location module is used to obtain the root fault location result of the faulty device in the industrial system when the fault occurs, based on the cause-effect graph corresponding to all devices in the industrial system.

2. The apparatus according to claim 1, characterized in that, The learnable parameters in the preset structured causal model also include initialized self-parameters and weights corresponding to the observation data within a preset historical time period for each observable device. The preset structured causal model takes the observation data within a preset historical time period for each observable device as input and outputs the probability value corresponding to the predicted observation data at each moment for each observable device. The preset structured causal model includes: The random number generation unit is used to initialize the confusion factor corresponding to each unobservable device within a preset historical time period and the causal relationship graph corresponding to all devices; The sample weight adaptive unit is used to initialize the weights corresponding to the observation data within a preset historical time period for each observable device. The causal relationship unit is used to obtain the predicted observation data of each observable device at each moment based on its own parameters, the observation data and corresponding weights of each observable device within a preset historical time period, the confusion factor of each unobservable device within a preset historical time period, and the causal relationship graph of all devices. The probability output unit is used to obtain the probability value corresponding to the predicted observation data of each observable device at each time step using statistical analysis methods based on the predicted observation data of each observable device at each time step.

3. The apparatus according to claim 2, characterized in that, The random number generation unit is configured to obtain the initialization of the confusion factor corresponding to each unobservable device within a preset historical time period and the causal relationship graph corresponding to all devices in the following manner: Initialize the mean and variance of the Gaussian distribution that the confusion factor of each unobservable device follows within a preset historical time period, and perform Gaussian sampling on the Gaussian distribution that the confusion factor of each unobservable device follows within a preset historical time period to obtain the initialized confusion factor of each unobservable device within a preset historical time period. Treat each device as a node, initialize the mean and variance of the Gaussian distribution that the causal relationships between nodes follow, and perform Gaussian sampling on the Gaussian distribution that the causal relationships between nodes follow to obtain the causal relationship graph corresponding to all devices.

4. The apparatus according to claim 3, characterized in that, The causal relationships between nodes include the existence and direction of edges between nodes. The random number generation unit is also configured to obtain the causal relationship graph corresponding to all devices in the following manner: Initialize the mean and variance of the Gaussian distribution that the existence of edges between each node follows, and perform Gaussian sampling on the Gaussian distribution that the existence of edges between each node follows to obtain the initial existence of edges between each node. Initialize the mean and variance of the Gaussian distribution that the directions of the edges between each node follow, and perform Gaussian sampling on the Gaussian distribution that the directions of the edges between each node follow to obtain the initial directions of the edges between each node. Based on the existence and direction of edges between the initialized nodes, a causal relationship graph corresponding to all initialized devices is obtained.

5. The apparatus according to claim 2, characterized in that, The causal relationship unit is configured to obtain the predicted observation data for each observable device at each time step in the following manner: in, Let d represent the predicted observation data of the j-th observable device at time t, d be the number of observable devices, r be the number of unobservable devices, and f be the predicted observation data of the j-th observable device at time t. obs,j (·) is the observation factor function, M ij This represents the causal relationship between the i-th observable device and the j-th observable device. f represents the nonlinear mapping of the observation data of the i-th observable device at time t with parameter Ω in the causal unit. conf,j (·) is the confusion factor function, M kj This represents the causal relationship between the k-th unobservable device and the j-th observable device. f represents the nonlinear mapping of the confusion factor at time t for the k-th unobservable device with parameter Ω in the causal unit. conf,j (·) is the confusion factor function.

6. The apparatus according to claim 4, characterized in that, The causal relationship module is configured as follows: During each round of iterative training, the learnable parameters of the structured causal model, the confusion factor corresponding to the preset historical time period for each unobservable device, and the causal relationship graph corresponding to all devices are updated according to the first preset loss.

7. The apparatus according to claim 6, characterized in that, The first preset loss is: in, The likelihood function is fitted to the observations of the observable device. For the constraint function of the causal relationship graph, This is the confusion factor constraint function for unobservable devices.

8. The apparatus according to claim 7, characterized in that, Where d represents the number of observable devices, r represents the number of unobservable devices, and T represents the preset historical time period. This represents the expected value of M, which is determined by the existence γ and direction θ of the edges between the devices, under a Gaussian distribution. This indicates that under the pre-defined prior distribution p(c) t Expected value under |M), The weights Ω and the confusion factors c for all unobservable devices at time t in the structured causal model are represented. t The probability value corresponding to the predicted observation data of the i-th observable device at time t is obtained under the causal relationship graph M corresponding to all devices, where KL[·] is the divergence function, p(M) represents the preset prior distribution of the causal relationship graph M corresponding to all devices, and q γ,θ (M) represents the Gaussian distribution that M follows, determined by the existence γ of edges between devices and the direction θ of the edges. The confusion factor of the i-th unobservable device at time t, under the causal relationship graph M of all observable devices at time t, follows a Gaussian distribution. Let represent the preset prior distribution of the confusion factor corresponding to the i-th unobservable device at time t under the causal relationship graph M corresponding to all devices.

9. The apparatus according to claim 4, characterized in that, The causal relationship module is configured as follows: In each round of iterative training, the weights corresponding to the observation data of each observable device within a preset historical time period in the structured causal model are frozen, and the structured causal model is trained iteratively a preset number of times. In each round of iterative training, the self-parameters of the structured causal model, the confusion factor corresponding to each unobservable device within a preset time period, and the causal relationship graph corresponding to all devices in the structured causal model are updated according to the second preset loss. After the preset number of iterative trainings, the self-parameters of the structured causal model, the confusion factor corresponding to each unobservable device within a preset time period, and the causal relationship graph corresponding to all devices in the structured causal model are frozen, and the structured causal model is trained once. In this training, the weights corresponding to the observation data of each observable device within a preset historical time period in the structured causal model are updated according to the second preset function, so as to obtain the self-parameters of the structured causal model after this round of iterative training, the weights corresponding to the observation data of each observable device within a preset historical time period, the confusion factor corresponding to each unobservable device within a preset time period, and the causal relationship graph corresponding to all devices.

10. The apparatus according to claim 9, characterized in that, The second preset loss is: in, This represents the observation fit likelihood function of each observable device in a structured causal model, under the weight w corresponding to the observation data within a preset historical time period. Represented as a causal relationship graph constraint function. This represents the confusion factor constraint function for unobservable devices under the weight w corresponding to the observation data of each observable device within a preset historical time period in the structured causal model.

11. The apparatus according to claim 10, characterized in that, Where d is the number of observable devices, r is the number of unobservable devices, T is the preset historical time period, and w t Let each observable device have a weight corresponding to the observation data at time t. This represents the expected value of M, which is determined by the existence γ and direction θ of the edges between the devices, under a Gaussian distribution. This indicates that under the pre-defined prior distribution p(c) t Expected value under |M), The weights Ω and the confusion factors c for all unobservable devices at time t in the structured causal model are represented. t The probability value corresponding to the predicted observation data of the i-th observable device at time t is obtained under the causal relationship graph M corresponding to all devices, where KL[·] is the divergence function, p(M) represents the preset prior distribution of the causal relationship graph M corresponding to all devices, and q γ,θ (M) represents the Gaussian distribution that M follows, determined by the existence γ of edges between devices and the direction θ of the edges. The confusion factor of the i-th unobservable device at time t, under the causal relationship graph M of all observable devices at time t, follows a Gaussian distribution. Let represent the preset prior distribution of the confusion factor corresponding to the i-th unobservable device at time t under the causal relationship graph M corresponding to all devices.

12. The apparatus according to claim 1, characterized in that, The causal relationship graph corresponding to all devices includes a one-way edge adjacency matrix and a two-way edge adjacency matrix, and the root fault location module is configured to obtain the root fault of the faulty device at the time of the fault occurrence in the following manner: The transition probability between each observable device is calculated based on the one-way edge adjacency matrix. Starting from the faulty device, a bias-reduced random walk is performed on all observable devices based on the transition probability between each observable device to obtain the access frequency of each observable device within a preset historical time period. The probability values ​​corresponding to the predicted observation data of each observable device at the time of failure are sorted by failure to obtain the anomaly score of each observable device at the time of failure. The fusion anomaly score of each observable device at the time of failure is obtained based on the anomaly score of each observable device at the time of failure and the access frequency of each observable device within a preset historical time period. The fused outliers of each observable device at the time of failure are sorted and the root fault location results of the faulty device in the industrial system are obtained according to the preset standard.

13. A root fault location method applied to industrial systems, characterized in that, The method includes: In response to the occurrence of a fault, acquire observation data within a preset historical time period when the fault occurs for each observable device; The device described in any one of claims 1-12 is used to obtain the root fault location result of the faulty equipment in the industrial system at the time of the fault based on the observation data within a preset historical time period when the fault occurs in each observable equipment.

14. An industrial system, characterized in that, The system includes: Multiple observable devices and multiple unobservable devices, with sensors deployed on each observable device for collecting observational data; The apparatus as described in any one of claims 1-12 is used to obtain the root fault location result of the faulty equipment in the industrial system at the time of the fault based on the observation data within a preset historical time period when the fault occurs in each observable equipment.

15. A computer-readable storage medium, characterized in that, It stores a computer program that can be executed by a processor to implement the steps of the method of claim 13.

Images