Fault-tolerant collaborative control method for automated production line based on AI fault diagnosis

Abstract

The application relates to the technical field of industrial automation control, and particularly discloses an automatic line fault-tolerant collaborative control method based on AI fault diagnosis, which comprises the following steps: a causal diagram database engine is constructed; if a causal diagram transaction is started, a physical verification action to be executed is generated, potential causal coupling risk selection execution paths are simulated and deduced across physical domains, physical verification evaluation conclusion data is output, conflict arbitration operations are executed when multiple causal diagram transactions are concurrently submitted, a to-be-submitted conclusion set is output, a cyclic self-verification detection operation is executed to determine whether there is an error cognitive steady state, if not, the to-be-submitted conclusion set is directly submitted to obtain a final correct cognition, if yes, a cognitive disturbance injection operation is executed on a temporary copy of the causal diagram database engine, a virtual constraint condition is formed, a forced re-reasoning operation is executed, final correct cognition is output, and precise collaborative fault tolerance is executed, so that precise collaborative fault tolerance and safe and reliable adaptive fault diagnosis reasoning are realized.

Description

Technical Field

[0001] This invention relates to the field of industrial automation control technology, specifically to an automated production line fault-tolerant collaborative control method based on AI fault diagnosis. Background Technology

[0002] With the deepening development of intelligent manufacturing and the Industrial Internet, automated production lines are evolving towards high integration, flexibility, and intelligence. Production lines typically contain dozens or even hundreds of process devices spanning multiple physical domains, including mechanical, electrical, thermal, and fluid dynamics. These devices are tightly coupled through material flow, energy flow, and information flow. During production line operation, abnormal conditions in a single device, such as vibration distortion caused by bearing wear, heat accumulation due to motor overcurrent, or pressure instability in hydraulic circuits, often propagate tier by tier along the coupling path between devices, ultimately manifesting as a complex fault phenomenon involving the interweaving and superposition of abnormal signals from multiple devices and multiple physical domains. Therefore, how to quickly and accurately identify the root cause of a fault from multi-source heterogeneous sensing data and formulate safe and effective fault-tolerant control strategies has become a core technical challenge in the intelligent operation and maintenance of production lines.

[0003] Existing technologies include deploying deep learning models on edge nodes for device-level fault feature extraction and diagnostic label generation, as well as constructing production line-level fault causal graphs based on Bayesian networks or causal discovery algorithms for root cause tracing. However, existing solutions still have the following unresolved technical issues:

[0004] First, multiple fault-tolerant logic threads need to concurrently access the cause-effect graph structure and submit diagnostic modification operations. Existing cause-effect graph storage solutions typically use a single-version read-write lock mechanism, which leads to severe access blocking and throughput reduction in concurrent inference scenarios. At the same time, there are conflicts such as node state contradictions, opposite operations, and read-write dependency failures among the diagnostic conclusions made by multiple threads based on different snapshot versions. There is a lack of an effective conflict arbitration mechanism, which makes it easy to write contradictory diagnostic conclusions into the cause-effect graph at the same time, thus destroying the logical consistency of the diagnostic knowledge base.

[0005] Secondly, the fault-tolerant logic thread generates physical verification actions to be executed based on the reasoning conclusions, such as issuing speed adjustment commands and voltage loading commands to the equipment controller. However, due to the complex cross-domain causal chains formed between different physical domains in the production line through mechanical connections, electrical coupling, heat conduction, and fluid transfer, the physical verification actions may spread along the coupling path after execution, causing unpredictable cross-physical domain interference to weak abnormal signals that have not yet been reasonably explained, and even aggravating hidden damage to the equipment. Existing solutions lack a systematic means to simulate and deduce the cross-physical domain causal coupling risks and select safe paths for physical verification actions before execution.

[0006] Thirdly, and more critically, during the continuous learning and reasoning process of data-driven AI diagnostic models on the production line causal graph, overfitting, data bias, or the multiple interpretability of anomalous signals may embed an incorrect fault hypothesis into the causal graph. This hypothesis can then be self-reinforcing through the strength parameters of the causal edges, ultimately forming a highly self-consistent and closed explanatory loop. This closed loop can explain the currently observed major anomalous signals, but it also systematically suppresses other weak anomalous signals that should be of concern, causing them to be noised down into passive results and ignored in the causal graph, thus forming a systematic erroneous cognitive steady state. This erroneous steady state is difficult to automatically identify in conventional diagnostic processes because, from a local logical perspective, each causal edge of the closed loop is self-consistent, but its overall explanatory direction contradicts objective physical causality. Existing technologies lack detection and proactive correction mechanisms for this type of systematic cognitive misjudgment, leading to erroneous diagnostic conclusions that may persist in the causal graph for a long time, misleading subsequent fault-tolerant decisions and causing serious safety hazards.

[0007] To address the aforementioned issues, this invention proposes an automated production line fault-tolerant collaborative control method based on AI fault diagnosis. Summary of the Invention

[0008] The purpose of this invention is to provide an AI-based fault diagnosis-based fault-tolerant collaborative control method for automated production lines. This method aims to solve the problems of conflicting diagnostic conclusions from multiple sources in existing automated production line fault diagnosis, lack of cross-domain risk pre-assessment for physical verification actions, and systemic misjudgments caused by AI diagnostic models falling into an erroneous cognitive steady state due to overconfidence without a proactive correction mechanism.

[0009] The objective of this invention can be achieved through the following technical solution: an automated production line fault-tolerant collaborative control method based on AI fault diagnosis, comprising the following steps:

[0010] A causal graph database engine supporting multi-version concurrency control is constructed. If a causal graph transaction is started, the physical verification actions to be executed are generated. Based on the equipment process mechanism knowledge base, the potential causal coupling risks across physical domains are simulated and deduced. An execution path is selected for the physical verification actions, and physical verification evaluation conclusion data is output.

[0011] Based on the causal graph database engine, when multiple causal graph transactions are committed concurrently, the physical verification and evaluation conclusion data is used as the empirical basis for determining the priority of the causal graph transactions, and a conflict arbitration operation is performed to output a set of conclusions to be committed.

[0012] For the set of conclusions to be submitted, perform a loop self-verification detection operation, including closed-loop self-consistency detection and weak signal root cause suppression detection, to determine whether there is a steady state of erroneous cognition. If not, the set of conclusions to be submitted is directly submitted to obtain the final correct cognition.

[0013] If it exists, combine the erroneous cognitive steady state with the temporary copy of the causal graph database engine to perform a cognitive perturbation injection operation, form virtual constraints, and perform a forced re-inference operation under the virtual constraints to output the final correct cognition and perform precise collaborative fault tolerance.

[0014] Furthermore, the causal graph database engine is constructed as follows:

[0015] A distributed sensing and computing architecture is deployed on the automated production line, including device-level sensing units, device-level edge computing nodes, production line-level edge computing nodes and production line-level data bus. The device-level sensing units collect real-time operating data of each device, and the device-level edge computing nodes run the AI ​​diagnostic label generation module to output device-level AI diagnostic labels and send them to the production line-level edge computing nodes.

[0016] A causal graph database engine is built on production line-level edge computing nodes. A multi-version concurrency control mechanism is used to maintain the global causal graph of the production line. The global causal graph of the production line is a directed cyclic graph. The node set includes equipment nodes, process variable nodes, and AI diagnostic label nodes. The causal edges in the edge set represent the causal relationships between nodes.

[0017] Furthermore, the method for outputting the physical verification and evaluation conclusion data is as follows:

[0018] The causal graph database engine provides a standardized transaction interface. When the fault-tolerant logic thread starts the fault-tolerant inference task, it starts a causal graph transaction, and the operation is recorded in a private operation log.

[0019] A verification action risk reviewer is constructed. When the fault-tolerant logic thread generates a physical verification action to be executed, it submits a review application. The verification action risk reviewer simulates and deduces the physical effect chain based on the equipment process mechanism knowledge base, assesses the potential causal coupling risk across physical domains and calculates a comprehensive risk score. Based on this, it selects the execution path and outputs the physical verification assessment conclusion data.

[0020] Furthermore, the AI ​​diagnostic label generation module incorporates a fault feature extraction network and a fault classifier, employing a joint supervised training method, including:

[0021] Collect historical operational data and corresponding equipment fault tags. Equipment fault tags include fault mode category and fault occurrence timestamp.

[0022] The fault feature extraction network and the fault classifier are cascaded to form an end-to-end trainable structure. The fault feature extraction network adopts the Transformer encoder architecture, and the fault classifier adopts a fully connected layer followed by a softmax normalization layer.

[0023] Using the cross-entropy loss function as the optimization objective, the backpropagation algorithm is used to update the network weight parameters. After training, the fault feature extraction network extracts fault-sensitive features, and the fault classifier outputs the confidence distribution vector and anomaly marker bits for each fault mode.

[0024] Furthermore, the assessment method for potential causal coupling risks across physical domains is as follows:

[0025] From the current main version snapshot of the cause-effect graph database engine, filter AI diagnostic label nodes with true anomaly flags and confidence levels below the anomaly confirmation threshold, as well as process variable nodes whose measured values ​​exceed the normal operating baseline range, and record them as anomaly signals. Anomaly signals with no explanatory path are recorded as unexplained anomaly signals.

[0026] The production line-level data bus caches the most recent real-time running data for a preset duration, calculates statistical features and compares them with steady-state baseline statistical features, and records signals whose statistical features exceed the allowable deviation range but whose corresponding AI diagnostic label node anomaly flag is false as unexplained abnormal signals.

[0027] Using the physical verification action to be performed as the excitation source, it propagates forward along the coupling path in the equipment process mechanism knowledge base, forming a physical effect chain;

[0028] Spatially and temporally match the physical effect chain with the set of unexplained anomalous signals. If the node corresponding to the unexplained anomalous signal appears in the affected node of the physical effect chain and the two belong to different physical domains, calculate the ratio of the expected physical quantity change amplitude to the cross-domain interference tolerance as a single risk score.

[0029] The matched individual risk scores are weighted and accumulated using a weight determined by the coupling strength level to obtain a comprehensive risk score.

[0030] Furthermore, the method for outputting the set of conclusions to be submitted is as follows:

[0031] Deploy a transaction manager, register and obtain a transaction identifier when the fault tolerance logic thread starts a cause-effect graph transaction, and include it in the merge commit cycle when committing.

[0032] The priority of causal graph transactions is determined by the execution path selection results and evidence source types in the physical verification and evaluation conclusion data. The priority of evidence source types, from high to low, is physical verification empirical, virtual verification empirical, and probabilistic graphical model statistical inference.

[0033] Construct an operation dependency graph, identify write-after-read dependencies, write-after-write dependencies, and read-after-write dependencies among operations, and detect node state contradictions, opposite operations on the same causal edge, and read-write dependency failures in the operation dependency graph.

[0034] For detected conflicts, the operation with the higher priority level wins, and the operation with the smaller transaction identifier at the same priority level wins. The winning operation is marked as committable, and the failed operation is marked as conflict rollback.

[0035] After performing a cascading rollback check on the causal graph transaction, the unmarked operations are included in the set of pending conclusions and output.

[0036] Furthermore, the statistical inference method of probabilistic graphical models is as follows:

[0037] Using the current attributes of each node in the global causal graph of the production line as evidence nodes, the global causal graph of the production line is converted into an equivalent Bayesian network. The probability strength parameters of the causal edges are used as the basis for initializing the conditional probability table, and the posterior failure probability of each equipment node is calculated using the connection tree algorithm.

[0038] Device nodes whose posterior failure probability exceeds the inference threshold are selected as candidate root causes of failure, diagnostic conclusions are generated, and the operational basis is marked as statistical inference of the probabilistic graphical model.

[0039] Furthermore, the method for determining whether a erroneous cognitive steady state exists is as follows:

[0040] Extract inference conclusion nodes and inference path information from the set of conclusions to be submitted, and construct a support graph. The support graph contains inference conclusion nodes, intermediate hypothesis nodes and observation signal nodes, and directed edges represent the support relationship from evidence nodes to inference conclusion nodes.

[0041] Starting with the most significant alarm signal, search along the directed edges of the support graph in the forward direction to find a closed path that returns from the inference conclusion node to the most significant alarm signal or its direct predecessor observation signal node. This path is called the closed interpretation loop.

[0042] Calculate the loop self-consistency score for closed interpretation loops. When the score exceeds the self-consistency alarm threshold, it is marked as a highly self-consistent closed interpretation loop.

[0043] In the current snapshot of the global cause-effect graph of the production line, compare the probability of each node as an independent root cause before and after the construction of the support graph. Nodes whose probability of hypothesis before the construction of the support graph is higher than the independent root cause attention threshold, whose probability after construction is lower than the threshold and whose decrease exceeds the decrease of the suppression judgment, and whose probability is not directly explained by the inference conclusion node in the support graph are judged as weak abnormal signals that are suppressed and are collected into a suppression signal set.

[0044] A steady state of erroneous cognition is determined to exist when the following conditions are met simultaneously: there is a highly self-consistent closed explanatory loop, the set of suppression signals is non-empty, and there is a causal path between the inference conclusion node and the suppression signal node in the loop pointing to the suppression signal; otherwise, it is determined not to exist.

[0045] Furthermore, the virtual constraints are formed as follows:

[0046] A temporary graph copy is generated from the consistency snapshot to be committed, and is physically isolated from the major version;

[0047] In the temporary graph copy, locate the inference conclusion node corresponding to the core diagnostic conclusion and the observation signal node corresponding to the most significant alarm signal at present, and extract the causal edge with the highest support strength between the two as the strong correlation edge locked by the closed loop.

[0048] For strongly correlated edges locked in closed loops and directed edges in closed interpretation loops that return from the inference conclusion node to the observation signal node or its direct predecessor observation signal node, a forced cut-off operation is performed, the probability intensity parameter is set to zero and marked as a perturbation cut-off state.

[0049] For each suppressed weak anomalous signal node in the suppressed signal set, the probability of assuming it as an independent root cause is forcibly set as the perturbation injection probability value, and a virtual prior edge is added to the node, marked as the perturbation injection state, forming a virtual constraint condition.

[0050] Furthermore, under virtual constraints, the forced re-reasoning operation is performed to output the final correct understanding in the following way:

[0051] Activate the root cause tracing thread, pass the temporary graph copy access handle, the list of suppressed weak abnormal signal nodes, and the mandatory constraint conditions. The mandatory constraint conditions assume that the device node corresponding to the core diagnostic conclusion is in a healthy state. The root cause tracing thread starts from the suppressed weak abnormal signal node and performs a bidirectional causal search along the directed edges that have not been interrupted. During the expansion process, it calls the equipment process mechanism knowledge base to perform physical consistency verification. The complete causal path that is discovered and passes the self-consistency judgment and physical consistency verification is recorded as the newly discovered causal chain, and a new hypothesis data package is generated.

[0052] After the new hypothesis data package is verified by the action risk reviewer, it is converted into a causal graph operation sequence and encapsulated as a first-priority correction transaction. Operations to be committed that are related to highly self-consistent closed interpretation loops are discarded. The transaction manager merges the operations to be committed that have not been discarded with the correction transaction and writes them into the main version of the causal graph database engine, generating a correct and consistent version as the final correct cognition.

[0053] The beneficial effects of this invention are as follows:

[0054] 1. This invention constructs a causal graph database engine that supports multi-version concurrency control, and combines it with a transaction conflict arbitration mechanism based on multi-level evidence priority. This enables multiple fault-tolerant logic threads to execute causal graph reasoning and modification operations in parallel within independent transactions, effectively solving the technical problems of concurrent access blocking and contradictory diagnostic conclusions. It significantly improves the throughput of production-level causal reasoning and the logical consistency of the diagnostic knowledge base.

[0055] 2. This invention uses a loop self-verification detector to perform closed-loop self-consistency detection and weak signal root cause suppression detection before transaction commit. When an erroneous cognitive steady state is detected, cognitive perturbation is injected into a temporary copy of the causal graph, the strongly correlated edges locked by the closed loop are forcibly cut off, and forced re-inference is performed. This effectively breaks the erroneous self-consistency closed loop formed by the overconfidence of the AI ​​diagnostic model, corrects systematic misjudgments before they are solidified, and greatly improves the accuracy of fault-tolerant decision-making and the safety of production line operation in complex fault scenarios. Attached Figure Description

[0056] The invention will now be further described with reference to the accompanying drawings.

[0057] Figure 1 This is a flowchart of the steps of the automated production line fault-tolerant collaborative control method based on AI fault diagnosis described in the embodiments of the present invention;

[0058] Figure 2 This is a cross-physical domain comprehensive risk score and composition distribution diagram of the physical verification action described in the embodiments of the present invention;

[0059] Figure 3 This is a flowchart illustrating the specific steps in the AI-based fault diagnosis-based automated production line fault-tolerant collaborative control method described in this embodiment of the invention to determine whether an erroneous cognitive steady state exists and to obtain the final correct cognition when no such state exists.

[0060] Figure 4 This is a comparison of the probability of independent root cause hypotheses and the suppression signal detection graph before and after the construction of the support graph as described in the embodiments of the present invention. Detailed Implementation

[0061] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0062] Example 1

[0063] Please see Figure 1 As shown in the embodiments of the present invention, the AI-based fault diagnosis-based fault-tolerant collaborative control method for automated production lines aims to solve the problems of conflicting multi-source concurrent diagnostic conclusions, lack of cross-domain risk pre-assessment for physical verification actions, and systemic misjudgments caused by AI diagnostic models falling into an erroneous cognitive steady state due to overconfidence without an active correction mechanism in existing automated production line fault diagnosis. This achieves accurate collaborative fault tolerance and safe and reliable adaptive fault diagnosis reasoning, specifically including the following steps:

[0064] S1: Build a cause-effect graph database engine that supports multi-version concurrency control. If a cause-effect graph transaction is started, generate the physical verification action to be executed, simulate and deduce the potential causal coupling risk across physical domains based on the equipment process mechanism knowledge base, select the execution path for the physical verification action, and output the physical verification evaluation conclusion data.

[0065] Specifically, a distributed sensing and computing architecture is deployed on the automated production line. The automated production line contains multiple devices. For each device, a device-level sensing unit and a device-level edge computing node are deployed, and a production line-level edge computing node and a production line-level data bus are also deployed.

[0066] The equipment-level sensing unit is used to collect real-time operating data of each device. The real-time operating data includes equipment vibration signals, motor current, spindle speed, temperature field distribution, hydraulic circuit pressure, pneumatic circuit pressure, and internal status words of the device controller. The equipment-level edge computing node runs the AI ​​diagnostic label generation module. The AI ​​diagnostic label generation module has a built-in pre-trained fault feature extraction network and fault classifier. It receives the real-time operating data of the corresponding device at a preset frequency, extracts fault-sensitive features, and outputs the equipment-level AI diagnostic label at the current moment. The equipment-level AI diagnostic label includes the confidence distribution vector of each fault mode of the device and the corresponding abnormal marker bit. The equipment-level edge computing node sends the real-time operating data and the equipment-level AI diagnostic label to the production line-level edge computing node through the production line-level data bus.

[0067] The fault feature extraction network and fault classifier are trained using a joint supervised method: real-time operating data and corresponding equipment fault labels are collected from the automated production line during its historical operation. These labels are determined through manual annotation and post-operation maintenance records, including the fault mode category and fault occurrence timestamp. Using real-time operating data as input samples and equipment fault labels as supervisory signals, the fault feature extraction network and fault classifier are cascaded to form an end-to-end trainable structure. The fault feature extraction network employs a Transformer encoder architecture, while the fault classifier uses a fully connected layer followed by a softmax normalization layer. During training, the cross-entropy loss function is used as the optimization objective. Input samples are fed into the fault feature extraction network in batches with a preset batch size for forward propagation. The cross-entropy loss between the equipment fault label and the output confidence distribution vector is calculated, and the weight parameters of the fault feature extraction network and fault classifier are updated using a backpropagation algorithm. After training, the fault feature extraction network extracts fault-sensitive features from the real-time operating data, and the fault classifier outputs the confidence distribution vector of each fault mode and the corresponding anomaly marker.

[0068] It should be noted that the anomaly flag is used to indicate whether the current device has a confidence level of at least one fault mode that exceeds the preset alarm threshold. A true value indicates that an anomaly exists, and a false value indicates that the device is operating normally.

[0069] A cause-effect graph database engine supporting multi-version concurrency control is built on the production line-level edge computing nodes. The cause-effect graph database engine uses a multi-version concurrency control mechanism to maintain the global cause-effect graph of the production line. The global cause-effect graph of the production line is a directed cyclic graph, denoted as G=(V,E), where G represents the global cause-effect graph of the production line, V represents the set of nodes, and E represents the set of edges.

[0070] Specifically, the node set V contains three types of nodes: equipment nodes, process variable nodes, and AI diagnostic label nodes. Equipment nodes correspond to each piece of equipment in the automated production line. Node attributes include equipment type, equipment number, and current equipment status. Process variable nodes correspond to quantifiable and monitorable process parameters during production line operation. Node attributes include process parameter name, current measurement value, and normal operating condition baseline range. AI diagnostic label nodes correspond to equipment-level AI diagnostic labels generated on each equipment-level edge computing node. The highest confidence level in the confidence distribution vector is recorded as the confidence level of the AI ​​diagnostic label node, and the fault mode corresponding to the confidence level is recorded as the fault mode of the node. Node attributes include fault mode, confidence level, and anomaly flag status.

[0071] Each causal edge in the edge set E represents a causal relationship from a source node to a target node, denoted as e=(v_src,v_tgt,τ,w,d), where e represents the causal edge, v_src represents the source node, v_tgt represents the target node, τ is a time delay parameter, representing the expected time delay from the change in the state of the source node to the observable effect produced in the state of the target node, w is a probability strength parameter, taking a real number in the interval [0,1], representing the correlation strength of the causal relationship corresponding to the causal edge, and d is a direction label, taking a value of positive or negative, indicating that the causal relationship has a positive promoting effect or a negative inhibiting effect.

[0072] It should be noted that the source node represents the node that acts as the cause in a causal relationship, and its state change propagates outward along the causal edge, causing changes in the state of other nodes. The target node represents the node that acts as the result in a causal relationship, and its state change is affected by the effect transmitted by the source node through the causal edge.

[0073] During initialization, the causal graph database engine constructs an initial global causal graph of the production line based on a pre-built equipment process mechanism knowledge base, using expert rules and causal discovery algorithms. During operation, the causal graph database engine supports dynamic addition and deletion of nodes and edges, as well as online updates of attributes.

[0074] The equipment process mechanism knowledge base is constructed as follows: The production line process flow of the automated production line and the engineering design documents, equipment specifications, process flow diagrams, and historical operation and maintenance records of each piece of equipment are collected. The transfer functions between the input and output physical quantities of each piece of equipment are extracted, including the force-displacement-vibration transfer relationship in the mechanical domain, the voltage-current-power transfer relationship in the electrical domain, the temperature-heat flow transfer relationship in the thermal domain, and the pressure-flow transfer relationship in the fluid domain. Based on the flow of materials, energy, and information in the production line process flow, the coupling paths between each piece of equipment are identified, and the areas crossed by each coupling path are marked. The physical domain attributes and coupling strength levels of the equipment are defined. The physical domain attributes include at least one of the mechanical, electrical, thermal, and fluid domains. The coupling strength level is obtained by mapping the coupling coefficients calibrated in the engineering design documents or historical operation and maintenance records, and is divided into three levels: strong coupling, medium coupling, and weak coupling. The transfer function and coupling path are stored in the persistent storage unit of the production line-level edge computing node in the form of structured rules to form an equipment process mechanism knowledge base. When the equipment is modified or the process is changed in the automated production line, the equipment process mechanism knowledge base is updated and iterated by importing the new version of the engineering design documents.

[0075] The cause-effect graph database engine provides a standardized transaction interface. All read and write operations of each fault-tolerant logic thread on the cause-effect graph database engine must be submitted through the transaction interface. The fault-tolerant logic thread refers to an independent execution unit running on the production line-level edge computing node. It is used to perform fault-tolerant reasoning tasks, including fault root cause tracing and fault-tolerant decision reasoning, on the global cause-effect graph of the production line for AI diagnostic label nodes, and to generate the physical verification actions to be executed based on the fault-tolerant reasoning tasks.

[0076] When each fault-tolerant logic thread starts a fault-tolerant inference task, it initiates a new causal graph transaction through the transaction interface. The causal graph transaction represents the consistent snapshot version requested by the fault-tolerant logic thread from the causal graph database engine through the standardized transaction interface. During the execution of the causal graph transaction, all operations of the fault-tolerant logic thread on the global causal graph of the production line are recorded in the private operation log of the causal graph transaction and are not directly written to the main version of the causal graph database engine.

[0077] It should be noted that physical verification actions refer to operations that require the equipment controller to send excitation signals to the actual physical equipment in the production line, change the equipment operating parameters, or trigger the equipment to execute a specific sequence of actions.

[0078] A verification action risk reviewer is built. The verification action risk reviewer runs on the production line-level edge computing node and interacts with the cause-effect graph database engine through an internal inter-process communication mechanism.

[0079] When any fault-tolerant logic thread generates a physical verification action to be executed, it submits a review application to the verification action risk reviewer. The review application includes: a complete description of the physical verification action to be executed, the target device number, the expected duration, the name of the expected stimulus parameter and the stimulus amplitude.

[0080] After receiving a review application, the verification action risk reviewer performs a simulation and deduction based on the equipment process mechanism knowledge base to simulate potential causal coupling risks across physical domains.

[0081] Specifically, from the current main version snapshot of the cause-effect graph database engine, all AI diagnostic label nodes are traversed, and AI diagnostic label nodes with true anomaly flags and confidence levels lower than the preset anomaly confirmation threshold are selected. Process variable nodes whose current measured values ​​exceed the normal operating condition baseline range are also selected and recorded as anomaly signals. If there is no anomaly explanation path corresponding to the anomaly signal in the cause-effect graph database engine, it is recorded as an unexplained anomaly signal.

[0082] The anomaly interpretation path is defined as a directed causal link that starts from any node identified as the root cause of the fault, traces forward along the direction of the causal edge, and can completely connect all the abnormal signals of downstream nodes caused by the root cause of the fault.

[0083] Real-time running data of the most recent preset duration is continuously cached on the production line-level data bus in a sliding window manner. In the real-time running data cached in the sliding window, statistical characteristics within the current sliding window are calculated for each type of real-time running data of each device node. The statistical characteristics include mean, variance, kurtosis, and skewness. Steady-state baseline statistical characteristics of the same type of real-time running data of the corresponding device during the historical normal operation period are retrieved from the persistent storage unit of the production line-level edge computing node. The steady-state baseline statistical characteristics are the long-term mean and allowable deviation range of the corresponding statistical characteristics during the historical normal operation period. The statistical characteristics within the current sliding window are compared with the steady-state baseline statistical characteristics. If the statistical characteristics within the current sliding window exceed the allowable deviation range, but the abnormal flag bit of the AI ​​diagnostic tag node corresponding to the device is false, then a weak symptom signal is detected and recorded as an unexplained abnormal signal.

[0084] Simulate and deduce the entire physical effect chain of physical verification actions based on the equipment process mechanism knowledge base;

[0085] Specifically, the physical verification action to be performed is used as the excitation source. It propagates forward along all possible coupling paths in the equipment process mechanism knowledge base. The expected physical effects of the physical verification action on each equipment and process variable in the automated production line are calculated to form a physical effect chain. The physical effect chain includes the affected nodes, expected physical quantity change direction and change amplitude estimate for each propagation step.

[0086] To assess the potential causal coupling risk across physical domains, all unexplained anomalous signals are compiled into an unexplained anomalous signal set. The propagation path of the physical effect chain is then spatially and temporally matched with the unexplained anomalous signal set. For each unexplained anomalous signal in the set, it is checked whether the corresponding equipment or process variable appears in the affected node of the physical effect chain. If so, it is further determined whether the physical domain of the anomalous signal and the physical domain of the effect generated at the corresponding node of the physical effect chain belong to different physical domains. If they belong to different physical domains, the ratio of the expected change amplitude of the physical quantity at the corresponding node of the physical effect chain to the pre-calibrated cross-domain interference tolerance is calculated and recorded as a single risk score. If the single risk score exceeds 1, it is determined that there is a potential causal coupling risk across physical domains.

[0087] The single risk scores of all matched cross-physical domain potential causal coupling risks are weighted and accumulated, where the weight coefficients are determined by the coupling strength level marked in the equipment process mechanism knowledge base for the corresponding coupling path, and the comprehensive risk score is calculated.

[0088] The verification action risk reviewer selects an execution path for the proposed physical verification action based on the comprehensive risk score and outputs physical verification assessment conclusion data, which includes: physical verification action identifier, comprehensive risk score, and execution path selection result;

[0089] The execution path selection result is determined according to the following rules: If the comprehensive risk score exceeds the preset risk threshold, the execution path is prohibited. The physical verification assessment conclusion data also includes risk explanation information, which explains which coupling paths the proposed physical verification action may use to cause cross-physical domain interference to which unexplained abnormal signals. After receiving the prohibited execution path, the fault-tolerant logic thread uses a high-fidelity digital twin model to perform pure virtual verification. The high-fidelity digital twin model is a high-precision simulation model that is synchronized with the physical entity in real time in the virtual space of the automated production line. It can simulate the excitation and response process of the physical verification action and output virtual verification evidence. The physical verification assessment conclusion data also includes virtual verification evidence.

[0090] If the overall risk score does not exceed the preset risk threshold, the permitted execution path is output, the fault-tolerant logic thread executes the physical verification action, and real-time running data is obtained on each device of the automated production line as physical verification evidence. The physical verification evaluation conclusion data also includes physical verification evidence.

[0091] For example, to visually demonstrate the risk level and risk source of each candidate physical verification action, the verification action risk reviewer will visualize the comprehensive risk score results according to the physical domain composition, such as... Figure 2 As shown, Figure 2This is a comprehensive risk score and composition distribution diagram of physical verification actions across physical domains. Figure 2 The comprehensive risk scores of eight typical physical verification actions, such as PC-001 speed step and PC-005 current limiting, are presented in the form of a stacked bar chart. The black, dark gray, light gray, and white bars represent the risk contributions of the mechanical, electrical, thermal, and fluid domains, respectively. The horizontal dashed line in the chart indicates the preset risk threshold (1.00). As can be seen from the chart, the comprehensive risk scores of PC-001, PC-005, and PC-007 all significantly exceed the threshold, mainly due to the cross-physical domain coupling between the mechanical and electrical domains. Therefore, the risk reviewer will output prohibited execution paths. Actions with lower scores, such as PC-003 and PC-007, are marked as permitted execution paths, providing a clear quantitative basis for whether to introduce a high-fidelity digital twin model for pure virtual verification in the fault-tolerant logic thread.

[0092] S2: Based on the causal graph database engine, when multiple causal graph transactions are committed concurrently, the physical verification and evaluation conclusion data is used as the empirical basis for determining the priority of causal graph transactions, and conflict arbitration operation is performed to output the set of conclusions to be committed.

[0093] Specifically, a transaction manager is deployed on the production line-level edge computing nodes. The transaction manager is a component of the cause-effect graph database engine and is responsible for coordinating the concurrent access of all fault-tolerant logic threads to the global cause-effect graph of the production line. When each fault-tolerant logic thread starts a cause-effect graph transaction, it registers with the transaction manager through a standardized transaction interface and obtains a globally unique transaction identifier. The transaction identifier is monotonically incremented by the transaction manager according to the order of registration time.

[0094] Define the internal structure of a causal graph transaction as an operation sequence, where each operation op_i is a quadruple, denoted as op_i=(operation type, target node or target edge, operation parameter, operation basis), i=1,2,...,k; denote the causal graph transaction T as T={op_1,op_2,...,op_k};

[0095] It should be noted that the range of values ​​for the operation type includes: adding a node, deleting a node, modifying node attributes, adding a causal edge, deleting a causal edge, and modifying edge attributes. The target node or target edge is used to uniquely specify the node or causal edge in the global causal graph of the production line to which this operation is applied. The operation parameters are the specific data required to execute this operation. When the operation type is to modify node attributes, the operation parameters include the attribute name and the modified attribute value. When the operation type is to modify edge attributes, the operation parameters include at least one of the following: time delay parameter, probability intensity parameter, and direction marker, and its modified value. The operation basis is used to record the type of evidence source from which this operation is derived. The evidence source type includes physical verification, virtual verification, and statistical inference from probabilistic graphical models.

[0096] When the fault-tolerant logic thread commits a causal graph transaction, it sends a commit request to the transaction manager. The transaction manager includes all currently active causal graph transactions in a merge commit cycle. The merge commit cycle is triggered by the transaction manager at regular intervals according to the preset commit cycle duration, or in advance when the number of active causal graph transactions reaches the preset transaction backlog threshold.

[0097] During the merge commit cycle, the transaction manager extracts the execution path selection results and evidence source types of the physical verification actions corresponding to each fault-tolerant logic thread from the physical verification evaluation conclusion data, as empirical evidence for determining the priority of causal graph transactions.

[0098] The correspondence between the types of evidence sources and their priorities is arranged from high to low as follows: the first priority corresponds to physical verification and empirical evidence, the second priority corresponds to virtual verification and empirical evidence, and the third priority corresponds to statistical inference of probabilistic graphical models. Statistical inference of probabilistic graphical models refers to the diagnostic conclusion derived by Bayesian network reasoning using the probability intensity parameters and time delay parameters already present in the global causal graph of the production line, and the diagnostic conclusion has not been confirmed by physical verification and empirical evidence or virtual verification and empirical evidence.

[0099] The diagnostic conclusion is derived through Bayesian network inference as follows: the current node attributes of each node in the global causal graph of the production line are used as evidence nodes. The global causal graph of the production line is regarded as a directed cyclic graph structure and converted into an equivalent Bayesian network representation. The probability strength parameters of each causal edge are used as the initialization basis of the conditional probability table. The connection tree algorithm is used to calculate the posterior failure probability of each other equipment node under the condition of the evidence node. Equipment nodes whose posterior failure probability exceeds the preset inference threshold are taken as candidate root causes of failure. The candidate root causes of failure and their corresponding posterior failure probabilities are used as diagnostic conclusions. The operation basis of the diagnostic conclusion is marked as statistical inference of the probabilistic graphical model.

[0100] The transaction manager performs conflict arbitration operations, which include building an operation dependency graph, performing conflict detection, and performing conflict arbitration in sequence.

[0101] Specifically, the operation dependency graph is constructed as follows: The transaction manager traverses all causal graph transactions to be committed within the current merge commit cycle, extracts all operations from the operation sequence of each causal graph transaction, and constructs an operation dependency graph with each operation as an operation node. The operation dependency graph is a directed acyclic graph, denoted as D=(O,Dep), where D represents the operation dependency graph, O represents the set of operation nodes containing all operation nodes, and Dep is the set of directed edges representing the dependency relationships between operation nodes.

[0102] It should be noted that there are three types of dependencies: write-after-read dependency, which means that one operation reads the attribute of a node or edge modified by another operation; write-after-write dependency, which means that two operations modify the same attribute of the same node or edge; and read-after-write dependency, which means that one operation determines its own modification content based on the attribute value of a node or edge it reads.

[0103] The conflict detection is performed as follows: The transaction manager identifies various conflict relationships in the operation dependency graph, including conflicting node state markers, conflicting operations on the same causal edge, and read / write dependency failures.

[0104] Specifically, if two operations belong to different cause-effect graph transactions, and both operations are of the type of modifying node attributes, the target node is the same device node, the modified node attributes are both device states, and the modified device states are different, it is determined that there is a node state contradiction and conflict between the two operations.

[0105] If two operations belong to different causal graph transactions, and the target edge of the two operations is the same edge, and the operation types are respectively deleting a causal edge and adding a causal edge or modifying edge attributes, or the operation types of the two operations are both modifying edge attributes, and the direction marker of the same edge is modified to positive and negative respectively, it is determined that there is a conflict between the two operations on the same causal edge.

[0106] If there are write-after-read or read-after-write dependency edges in the operation dependency graph, and the source operation and target operation of the dependency chain belong to different cause-effect graph transactions, check whether the dependency relationship is inconsistent due to the two cause-effect graph transactions being based on different snapshot versions of the cause-effect graph database engine. If the snapshot version of the transaction where the source operation is located is later than the snapshot version of the transaction where the target operation is located, then the read data on which the target operation is based has been modified by the source operation, and it is determined that there is a read-write dependency failure conflict.

[0107] All identified conflict relationships are recorded as a conflict set. Each record in the conflict set contains the conflict type, the transaction identifier and operation identifier involved in the conflict, and the identifier of the node or edge involved in the conflict.

[0108] The method for executing conflict arbitration is as follows: For each conflict in the conflict set, when the two operations involved in the conflict are at different priority levels, the operation with the higher priority level wins; when the two operations involved in the conflict are at the same priority level, the transaction identifiers of the causal graph transactions to which the two operations belong are further compared, and the operation with the smaller transaction identifier wins, thus obtaining the arbitration result. According to the arbitration result, the winning operation is marked as committable, and the failing operation is marked as conflict rollback.

[0109] Traverse all causal graph transactions. For causal graph transactions where all operations are marked as committable, include all operations in the operation sequence into the set of conclusions to be committed. For causal graph transactions where at least one operation is marked as conflict rollback, check if there are any operations in the causal graph transaction that are not marked as conflict rollback. If so, further check if there is a dependency relationship in the operation dependency graph between the operations not marked as conflict rollback and the operations marked as conflict rollback. If a dependency relationship exists, mark all unmarked operations that depend on the conflict rollback operations as cascading rollbacks. Include the remaining unmarked operations after the cascading rollback determination into the set of conclusions to be committed, and determine the causal graph transaction as a partial rollback transaction. Obtain the latest graph snapshot updated by the committed transaction as the new inference basis and enter the retry process. If, after the cascading rollback determination, there are no operations that can be included into the set of conclusions to be committed, determine the causal graph transaction as a full rollback transaction, obtain the latest graph snapshot as the new inference basis, and enter the retry process.

[0110] The collected and outputted set of conclusions to be submitted includes all committable operations to be applied to the main version of the causal graph database engine after conflict arbitration within this merge submission cycle, the transaction identifier of each committable operation, the operation basis of each committable operation, and the corresponding priority level.

[0111] S3: For the set of conclusions to be submitted, perform a loop self-verification detection operation including closed interpretation loop self-consistency detection and weak signal root cause suppression detection to determine whether there is an erroneous cognitive steady state. If not, the set of conclusions to be submitted is directly submitted to obtain the final correct cognition.

[0112] like Figure 3 As shown, the specific steps to determine whether there is a erroneous cognitive steady state and to obtain the final correct cognition when it does not exist are as follows;

[0113] Specifically, a loop self-verification detector is deployed on the production line-level edge computing node. The transaction manager sends a self-verification detection request to the loop self-verification detector. The self-verification detection request contains a set of conclusions to be submitted. After receiving the self-verification detection request, the loop self-verification detector performs loop self-verification detection operations, including closed interpretation loop self-consistency detection and weak signal root cause suppression detection.

[0114] Specifically, the method for testing the self-consistency of a closed-loop interpretation is as follows;

[0115] Construct a support graph for the current set of conclusions to be submitted. Extract all operations to be submitted from the set of conclusions to be submitted. Filter out operations whose operation type is modifying node attributes and whose modified node attribute is device status, and operations whose operation type is modifying edge attributes and whose modified edge attribute is probability intensity parameter. For each filtered operation, use the node or causal edge modified by the operation as the inference conclusion node, and use the operation basis as the evidence source marker. Extract the inference path information recorded in the conflict detection process of the causal graph transaction where the operation is located. The inference path information is the node sequence and edge sequence in the global causal graph of the production line that the operation passes through from the initial observed signal to the final diagnostic conclusion.

[0116] Using all extracted inference conclusion nodes as the root node set, and all nodes and causal edges contained in the inference path information as the skeleton, a support graph is constructed. The support graph is a directed acyclic graph, denoted as S_support=(N,K), where S_support represents the support graph, N represents the node set, which includes three types: inference conclusion nodes, intermediate hypothesis nodes, and observation signal nodes. Inference conclusion nodes correspond to the equipment fault status or causal edge strength change in the operation to be submitted. Intermediate hypothesis nodes correspond to the intermediate variables or transitional diagnostic hypotheses introduced in the inference path information. Observation signal nodes correspond to the abnormal signals in the AI ​​diagnostic label nodes or process variable nodes that trigger the inference path information. K is the set of directed edges, including each directed edge representing the support relationship from the evidence node to the inference conclusion node. The direction of the directed edge is from the evidence node to the inference conclusion node. The attributes of the directed edge include support strength, which is taken from the probability strength parameter corresponding to the directed edge in the global causal graph of the production line.

[0117] For each inference conclusion node, extract the key evidence edge that it directly depends on. The key evidence edge is defined as a directed edge in the support graph that points to the inference conclusion node and whose support strength exceeds the preset key evidence threshold.

[0118] The search for closed-loop explanations starts with the most significant alarm signal. The most significant alarm signal is determined as follows: traverse all observation signal nodes in the support graph, obtain the current confidence level and anomaly flag bit of each observation signal node in the global cause-effect graph of the production line, filter out observation signal nodes with true anomaly flag bits, and select the observation signal node with a confidence level exceeding the preset significant alarm threshold. The one with the highest confidence level is taken as the most significant alarm signal. If there are multiple observation signal nodes with the same confidence level and all exceeding the preset significant alarm threshold, the one with the longest duration of the anomaly flag bit output by the AI ​​diagnostic label generation module is taken as the most significant alarm signal.

[0119] In the support graph, starting from the most significant alarm signal, we search forward along the direction of the directed edge to find if there is a path that can reach the inference conclusion node. Starting from the inference conclusion node, we continue to search along the direction of the directed edge of the support graph to find if there is another path or an existing causal edge in the global causal graph of the production line, and return to the most significant alarm signal or the direct predecessor observation signal node of the most significant alarm signal. This is called a closed path.

[0120] If any closed path exists, starting from the most significant alarm signal, passing through at least one reasoning conclusion node, and finally returning to the most significant alarm signal itself or its direct predecessor observation signal node, then it is determined that a closed interpretation loop has been found.

[0121] The closed interpretation loops found are evaluated for self-consistency. Specifically, the closed interpretation loops are labeled with attributes including: loop length (the number of directed edges in the loop); mean support strength (the arithmetic mean of the support strengths of all directed edges in the loop); and loop self-consistency score (the ratio of the product of the support strengths of all directed edges in the loop to the loop length). When the loop self-consistency score exceeds a preset self-consistency alarm threshold, the closed interpretation loop is marked as a highly self-consistent closed interpretation loop.

[0122] After completing the closed-loop self-consistency test, the loop self-verification detector performs weak signal root cause suppression test.

[0123] Specifically, in the current snapshot version of the global cause-effect graph of the production line, all process variable nodes and AI diagnostic label nodes are traversed to extract nodes whose hypothesis probability as an independent root cause changes before and after the support graph is constructed. The hypothesis probability as an independent root cause is defined as follows: in the global cause-effect graph of the production line, with the node as the source node and any equipment node as the sink node, the posterior probability of the node as the root cause explaining all current abnormal signals is calculated by the causal path search algorithm.

[0124] For each node traversed, compare the probability of the node as an independent root cause before the support graph is constructed with the probability of the node as an independent root cause after the support graph is constructed and the inference conclusion nodes in the support graph are substituted into the global causal graph of the production line as known conditions.

[0125] If the following two conditions are met, the node is determined to be a suppressed weak abnormal signal;

[0126] Condition 1: Before the support graph is built, the probability of a node being assumed to be an independent root cause is higher than the preset independent root cause attention threshold. After the support graph is built, the probability of a node being assumed to be an independent root cause is lower than the independent root cause attention threshold, and the decrease exceeds the preset suppression judgment decrease.

[0127] Condition 2: The node is not directly explained by any inference conclusion node through a directed edge in the support graph. That is, there is no directed path from any inference conclusion node to the node, or although there is a directed path, the support direction on the directed path is from the node to the inference conclusion node. This means that the node is interpreted by the closed interpretation loop as a passive result of the fault rather than an active root cause.

[0128] All nodes that are identified as having suppressed weak abnormal signals are grouped into a suppression signal set.

[0129] For example, to quantify the suppressive effect of closed-loop explanation reasoning on potential true root cause signals, the loop self-verification detector compares the probabilities of independent root cause hypotheses before and after support graph construction, such as... Figure 4 As shown, Figure 4 This is a comparison of the independent root cause hypothesis probabilities before and after support graph construction, along with the suppression signal detection graph. It shows the changes in hypothesis probabilities for multiple equipment nodes and process variable nodes, such as vibration node V_VIB_03, temperature node V_TEMP_07, and AI diagnostic label nodes, before and after support graph construction (dark diamond nodes). The horizontal dashed line in the graph represents the preset independent root cause attention threshold (0.60). The comparative data clearly shows that when the inference conclusion node in the support graph is submitted, the independent root cause hypothesis probability of most nodes drops sharply: for example, V_V... Both the spindle vibration node (IB_03) and the heating temperature node (V_TEMP_07) dropped from above the attention threshold to around 0.20. Meanwhile, although the probability of the conveyor current node (V_CURR_02) was as high as nearly 0.90 before construction, the connection and state mapping in the figure show that it is affected by the reverse interpretation of the closed interpretation loop. If such nodes have a significantly reduced probability and are not directly interpreted in the support figure, they are judged as suppressed weak anomalous signals, providing key target location basis for the subsequent cognitive disturbance injection module to perform forced re-inference and break the erroneous cognitive steady state.

[0130] After completing the closed-loop self-consistency test and the weak signal root cause suppression test, the loop self-verification detector determines whether there is a erroneous cognitive steady state. The erroneous cognitive steady state is determined by simultaneously satisfying the following three conditions.

[0131] Condition A: At least one closed interpretation loop marked as highly self-consistent was found;

[0132] Condition B: The set of suppressed signals is not empty, that is, there is at least one suppressed weak abnormal signal;

[0133] Condition C: The inference conclusion nodes contained in the highly self-consistent closed interpretation loop have a causal relationship path with the suppressed weak anomalous signals in the suppressed signal set, and the direction of the causal relationship path in the global causal graph of the production line is marked from the inference conclusion nodes to the suppressed weak anomalous signals, indicating that the suppressed weak anomalous signals are interpreted as passive results or noise by the diagnostic conclusions in the closed interpretation loop.

[0134] If the three conditions are not met simultaneously, it is determined that there is no erroneous cognitive steady state. The cyclic self-verification detector outputs a detection pass conclusion and directly submits the set of conclusions to be submitted. After receiving the detection pass conclusion, the transaction manager applies all committable operations in the set of conclusions to be submitted to the main version of the cause-effect graph database engine in sequence, generates a new main version number, forms a new steady state version of the production line global cause-effect graph, and obtains the final correct cognition.

[0135] If all three conditions are met, it is determined that there is a erroneous cognitive steady state. The cyclic self-verification detector outputs a failure conclusion, refuses to directly submit the set of conclusions to be submitted as the final correct cognition, and suspends the submission operation of this merge submission cycle.

[0136] S4: If it exists, combine the erroneous cognitive steady state with the temporary copy of the causal graph database engine to perform a cognitive perturbation injection operation, form virtual constraints, and perform a forced re-inference operation under the virtual constraints to output the final correct cognition and perform precise collaborative fault tolerance.

[0137] Specifically, a cognitive disturbance execution module is deployed on the production line-level edge computing node. The cognitive disturbance execution module receives the detection failure conclusion and accompanying information output by the cyclic self-verifying detector when it determines that there is an erroneous cognitive steady state. The accompanying information includes: the complete path information of a highly self-consistent closed interpretation loop, the set of suppressed weak abnormal signals, the node identifier of each suppressed weak abnormal signal, and the core diagnostic conclusion. The core diagnostic conclusion represents the inference conclusion node in the highly self-consistent closed interpretation loop that plays the role of interpreting the most significant alarm signal and whose equipment status is marked as faulty.

[0138] After receiving the detection failure conclusion and accompanying information, the cognitive perturbation execution module performs a cognitive perturbation injection operation on a temporary copy of the causal graph database engine, based on the erroneous cognitive steady state.

[0139] Specifically, a temporary copy is created: The cognitive perturbation execution module requests the causal graph database engine to copy a temporary graph copy, denoted as G_dist, from the latest consistency snapshot of the production global causal graph that is currently pending but not yet written to the main version. The temporary graph copy G_dist is physically isolated from the main version of the causal graph database engine, and any modification to the temporary graph copy G_dist will not affect the main version.

[0140] In the temporary graph copy G_dist, the cognitive disturbance execution module locates the inference conclusion node corresponding to the core diagnostic conclusion, denoted as v_core; locates the observation signal node corresponding to the most significant alarm signal, denoted as v_alarm; and extracts the causal edge with the highest support strength on the directed path from v_alarm to v_core from the complete path information of the highly self-consistent closed interpretation loop, denoted as e_lock, which is defined as a strongly correlated edge locked by the closed loop.

[0141] The cognitive perturbation execution module performs a forced cut-off operation on the temporary graph copy G_dist: temporarily modifies the probability strength parameter w of the strongly correlated edge e_lock that is locked in the closed loop to zero, and marks the state of the causal edge as the perturbation cut-off state. The cognitive perturbation execution module traverses all directed edges in the complete path information of the highly self-consistent closed interpretation loop that return from v_core to v_alarm or v_alarm's direct predecessor observation signal node, and performs the same forced cut-off operation on each directed edge, temporarily modifies the probability strength parameter to zero and marks it as the perturbation cut-off state, so as to completely break the complete loop structure of the highly self-consistent closed interpretation loop.

[0142] The cognitive perturbation execution module traverses each node of the suppressed weak anomalous signal in the suppressed signal set, denoted as v_suppressed. For each v_suppressed, the cognitive perturbation execution module forcibly sets the node's hypothesis probability as an independent root cause to a preset perturbation injection probability value in the temporary graph copy G_dist. The perturbation injection probability value is determined as follows: the maximum value between the node's hypothesis probability as an independent root cause before the support graph is constructed and the preset independent root cause forced boosting benchmark value is taken. The independent root cause forced boosting benchmark value is a pre-calibrated constant, and its value is higher than the independent root cause attention threshold.

[0143] In the temporary graph copy G_dist, the cognitive disturbance execution module adds a virtual prior edge for each v_suppressed, denoted as e_virtual=(v_suppressed,v_sink,τ_default,w_virtual,forward). Here, v_sink is any equipment node or process variable node that v_suppressed can affect in the equipment process mechanism knowledge base. If there is no clear causal path in the equipment process mechanism knowledge base, v_sink is the equipment node or process variable node where the abnormal signal is first observed. τ_default is a preset default time delay parameter. w_virtual is the virtual prior strength, which is the product of the disturbance injection probability value and the preset virtual prior strength coefficient. The state of the virtual prior edge is marked as the disturbance injection state.

[0144] After the cognitive perturbation injection operation is completed, a set of virtual constraints are formed in the temporary graph copy G_dist. The virtual constraints include: the constraint that the probability strength parameter of all causal edges marked as perturbation cut-off states is zero, and the constraints of all virtual prior edges marked as perturbation injection states and their virtual prior strength.

[0145] Under virtual constraints, the cognitive perturbation execution module performs a forced re-inference operation;

[0146] The cognitive perturbation execution module activates a dedicated root cause tracing thread, which is a type of fault-tolerant logic thread. The cognitive perturbation execution module passes the re-inference task parameters to the root cause tracing thread, including: the access handle of the temporary graph copy G_dist, the list of node identifiers of all v_suppressed nodes in the suppression signal set, and the mandatory constraint condition. The mandatory constraint condition is: during the re-inference process, it is assumed that the device node corresponding to the core diagnostic conclusion is in a healthy state, that is, the device state of the device node is fixed as healthy, and no inference operation is allowed to modify it to a fault.

[0147] The root cause tracing thread starts the causal explanation path search on the temporary graph copy G_dist, using each v_suppressed in the set of suppressed weak abnormal signals as the starting point of inference and the mandatory constraints as boundary constraints.

[0148] The causal explanation path search adopts a bidirectional causal search algorithm: starting from v_suppressed, along the directed edges in the temporary graph replica G_dist that are not marked as perturbation cut-off states, breadth-first expansion is performed according to the forward or reverse direction of the direction marking. Forward expansion searches for downstream anomalies caused by v_suppressed along the direction of the causal edge, and reverse expansion searches for upstream causes of v_suppressed along the reverse direction of the causal edge. During the search process, at each expansion step, the equipment process mechanism knowledge base is called to perform physical consistency verification on the currently searched causal path. The physical consistency verification includes: checking whether the physical domain attributes between the nodes on the current causal path are compatible, checking whether the timing relationship of the delay parameters conforms to the physical causality law, and checking whether the direction of change of the physical quantity of each node on the path is consistent with the promoting or inhibiting effect represented by the direction marking. If the current causal path fails the physical consistency verification, further expansion of the path is terminated.

[0149] If, during the search process, the root cause tracing thread successfully discovers a complete causal path that starts from v_suppressed, passes through several intermediate nodes, and finally explains one or more other observed signal nodes in a manner that conforms to physical consistency checks, and there are no causal edges marked as perturbation cut-off states on the complete causal path, then the root cause tracing thread records the complete causal path as a newly discovered causal chain.

[0150] The newly discovered causal chain must simultaneously satisfy the following self-consistency criteria: there are no logical contradictions within the subgraph formed by all nodes and causal edges on the newly discovered causal chain; the newly discovered causal chain is not included in a highly self-consistent closed explanatory loop; and the set of anomalous signals that the newly discovered causal chain can explain intersects with the set of anomalous signals where the most significant alarm signal is located.

[0151] For each newly discovered causal chain that satisfies the self-consistency determination condition, the root cause tracing thread generates a new hypothesis data packet. The new hypothesis data packet includes: the complete node sequence and edge sequence of the newly discovered causal chain, the list of abnormal signals explained by the newly discovered causal chain, the preliminary estimate of the probability intensity parameter of each causal edge on the newly discovered causal chain, and the verification record of physical consistency check.

[0152] The cognitive perturbation execution module submits all new hypothesis data packets generated by the root cause tracing thread to the verification action risk reviewer for verification. For each new hypothesis data packet, the verification action risk reviewer uses the physical verification action implied by the key causal edge on the newly discovered causal chain as the incentive source, simulates the entire physical effect chain, and evaluates whether it generates potential causal coupling risk across physical domains with the signals in the unexplained abnormal signal set. If the comprehensive risk score does not exceed the preset risk threshold, the verification action risk reviewer outputs a verification pass conclusion; otherwise, it outputs a verification fail conclusion.

[0153] The cognitive disturbance execution module collects all verified new hypothesis data packets and converts them into corresponding causal graph operation sequences. The causal graph operation sequences include: adding causal edges that do not yet exist on the newly discovered causal chain in the global causal graph of the production line, and setting their time delay parameters, probability strength parameters, and direction markers; updating the hypothesis probability of the node corresponding to the suppressed weak abnormal signal as an independent root cause to the posterior probability value supported by the newly discovered causal chain; and correcting the equipment status of the equipment node corresponding to the core diagnostic conclusion to healthy.

[0154] The cognitive disturbance execution module encapsulates the causal graph operation sequence into a correction transaction and sets the priority level of the correction transaction to the first priority. The cognitive disturbance execution module sends an instruction to the transaction manager to discard all pending operations related to the highly self-consistent closed interpretation loop within the current merge commit cycle, including operations that mark the device status of the core diagnostic conclusion as faulty, and operations that enhance the probability strength parameters of strongly correlated edges locked by the closed loop.

[0155] After receiving the correction transaction and discard instructions, the transaction manager executes the final merge commit: discarding the pending operations related to the highly self-consistent closed interpretation loop, and performing conflict detection and merging on the committable operations in the set of pending conclusions that have not been discarded during this merge commit cycle, together with the causal graph operation sequence in the correction transaction, and writing them into the main version of the causal graph database engine to generate a new correct and consistent version of the production line global causal graph as the final correct cognitive output.

[0156] Based on the final correct understanding, precise collaborative fault tolerance is implemented. Based on the updated device status markings of each device node in the correct and consistent version, as well as the newly established causal paths in the global causal graph of the production line, precise collaborative fault tolerance instructions are issued to each device controller. Precise collaborative fault tolerance instructions include: executing isolation or power reduction operation instructions for devices identified as faulty, removing the abnormal markings and restoring normal operating parameters for devices identified as healthy, and executing preventive monitoring enhancement instructions for related devices located on the newly discovered causal chain.

[0157] The technical solution of this invention is as follows: A causal graph database engine supporting multi-version concurrent control is constructed. If a causal graph transaction is initiated, for the physical verification action to be executed, the potential causal coupling risk across physical domains is simulated and deduced based on the equipment process mechanism knowledge base. An execution path is selected for the physical verification action, and physical verification evaluation conclusion data is output. Based on the causal graph database engine, when multiple causal graph transactions are submitted concurrently, the physical verification evaluation conclusion data is used as the empirical basis for determining the priority of the causal graph transaction. A conflict arbitration operation is performed, and a set of conclusions to be submitted is output. For the set of conclusions to be submitted, a loop self-verification detection operation including closed interpretation loop self-consistency detection and weak signal root cause suppression detection is performed to determine whether there is an erroneous cognitive steady state. If not, the set of conclusions to be submitted is directly submitted to obtain the final correct cognition. If it exists, a cognitive disturbance injection operation is performed on a temporary copy of the causal graph database engine in combination with the erroneous cognitive steady state to form virtual constraints. Under the virtual constraints, a forced re-inference operation is performed to output the final correct cognition, and precise collaborative fault tolerance is performed.

[0158] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. An automated production line fault-tolerant collaborative control method based on AI fault diagnosis, characterized by: Includes the following steps: A causal graph database engine supporting multi-version concurrency control is constructed. If a causal graph transaction is started, the physical verification actions to be executed are generated. Based on the equipment process mechanism knowledge base, the potential causal coupling risks across physical domains are simulated and deduced. An execution path is selected for the physical verification actions, and physical verification evaluation conclusion data is output. Based on the causal graph database engine, when multiple causal graph transactions are committed concurrently, the physical verification and evaluation conclusion data is used as the empirical basis for determining the priority of the causal graph transactions, and a conflict arbitration operation is performed to output a set of conclusions to be committed. For the set of conclusions to be submitted, perform a loop self-verification detection operation, including closed-loop self-consistency detection and weak signal root cause suppression detection, to determine whether there is a steady state of erroneous cognition. If not, the set of conclusions to be submitted is directly submitted to obtain the final correct cognition. If it exists, combine the erroneous cognitive steady state with the temporary copy of the causal graph database engine to perform a cognitive perturbation injection operation, form virtual constraints, and perform a forced re-inference operation under the virtual constraints to output the final correct cognition and perform precise collaborative fault tolerance.

2. The automated production line fault-tolerant collaborative control method based on AI fault diagnosis according to claim 1, characterized in that: The cause-effect graph database engine is constructed as follows: A distributed sensing and computing architecture is deployed on the automated production line, including device-level sensing units, device-level edge computing nodes, production line-level edge computing nodes and production line-level data bus. The device-level sensing units collect real-time operating data of each device, and the device-level edge computing nodes run the AI ​​diagnostic label generation module to output device-level AI diagnostic labels and send them to the production line-level edge computing nodes. A causal graph database engine is built on production line-level edge computing nodes. A multi-version concurrency control mechanism is used to maintain the global causal graph of the production line. The global causal graph of the production line is a directed cyclic graph. The node set includes equipment nodes, process variable nodes, and AI diagnostic label nodes. The causal edges in the edge set represent the causal relationships between nodes.

3. The automated production line fault-tolerant collaborative control method based on AI fault diagnosis according to claim 2, characterized in that: The method for outputting physical verification and evaluation conclusion data is as follows: The causal graph database engine provides a standardized transaction interface. When the fault-tolerant logic thread starts the fault-tolerant inference task, it starts a causal graph transaction, and the operation is recorded in a private operation log. A verification action risk reviewer is constructed. When the fault-tolerant logic thread generates a physical verification action to be executed, it submits a review application. The verification action risk reviewer simulates and deduces the physical effect chain based on the equipment process mechanism knowledge base, assesses the potential causal coupling risk across physical domains and calculates a comprehensive risk score. Based on this, it selects the execution path and outputs the physical verification assessment conclusion data.

4. The automated production line fault-tolerant collaborative control method based on AI fault diagnosis according to claim 2, characterized in that: The AI ​​diagnostic label generation module incorporates a fault feature extraction network and a fault classifier, employing a joint supervised training method, including: Collect historical operational data and corresponding equipment fault tags. Equipment fault tags include fault mode category and fault occurrence timestamp. The fault feature extraction network and the fault classifier are cascaded to form an end-to-end trainable structure. The fault feature extraction network adopts the Transformer encoder architecture, and the fault classifier adopts a fully connected layer followed by a softmax normalization layer. Using the cross-entropy loss function as the optimization objective, the backpropagation algorithm is used to update the network weight parameters. After training, the fault feature extraction network extracts fault-sensitive features, and the fault classifier outputs the confidence distribution vector and anomaly marker bits for each fault mode.

5. The automated production line fault-tolerant collaborative control method based on AI fault diagnosis according to claim 3, characterized in that: The assessment method for the potential causal coupling risk across physical domains is as follows: From the current main version snapshot of the cause-effect graph database engine, filter AI diagnostic label nodes with true anomaly flags and confidence levels below the anomaly confirmation threshold, as well as process variable nodes whose measured values ​​exceed the normal operating baseline range, and record them as anomaly signals. Anomaly signals with no explanatory path are recorded as unexplained anomaly signals. The production line-level data bus caches the most recent real-time running data for a preset duration, calculates statistical features and compares them with steady-state baseline statistical features, and records signals whose statistical features exceed the allowable deviation range but whose corresponding AI diagnostic label node anomaly flag is false as unexplained abnormal signals. Using the physical verification action to be performed as the excitation source, it propagates forward along the coupling path in the equipment process mechanism knowledge base, forming a physical effect chain; Spatially and temporally match the physical effect chain with the set of unexplained anomalous signals. If the node corresponding to the unexplained anomalous signal appears in the affected node of the physical effect chain and the two belong to different physical domains, calculate the ratio of the expected physical quantity change amplitude to the cross-domain interference tolerance as a single risk score. The matched individual risk scores are weighted and accumulated using a weight determined by the coupling strength level to obtain a comprehensive risk score.

6. The automated production line fault-tolerant collaborative control method based on AI fault diagnosis according to claim 1, characterized in that: The method for outputting the set of conclusions to be submitted is as follows: Deploy a transaction manager, register and obtain a transaction identifier when the fault tolerance logic thread starts a cause-effect graph transaction, and include it in the merge commit cycle when committing. The priority of causal graph transactions is determined by the execution path selection results and evidence source types in the physical verification and evaluation conclusion data. The priority of evidence source types, from high to low, is physical verification empirical, virtual verification empirical, and probabilistic graphical model statistical inference. Construct an operation dependency graph, identify write-after-read dependencies, write-after-write dependencies, and read-after-write dependencies among operations, and detect node state contradictions, opposite operations on the same causal edge, and read-write dependency failures in the operation dependency graph. For detected conflicts, the operation with the higher priority level wins, and the operation with the smaller transaction identifier at the same priority level wins. The winning operation is marked as committable, and the failed operation is marked as conflict rollback. After performing a cascading rollback check on the causal graph transaction, the unmarked operations are included in the set of pending conclusions and output.

7. The automated production line fault-tolerant collaborative control method based on AI fault diagnosis according to claim 6, characterized in that: The statistical inference method of probabilistic graphical models is as follows: Using the current attributes of each node in the global causal graph of the production line as evidence nodes, the global causal graph of the production line is converted into an equivalent Bayesian network. The probability strength parameters of the causal edges are used as the basis for initializing the conditional probability table, and the posterior failure probability of each equipment node is calculated using the connection tree algorithm. Device nodes whose posterior failure probability exceeds the inference threshold are selected as candidate root causes of failure, diagnostic conclusions are generated, and the operational basis is marked as statistical inference of the probabilistic graphical model.

8. The automated production line fault-tolerant collaborative control method based on AI fault diagnosis according to claim 1, characterized in that: The method to determine whether a erroneous cognitive steady state exists is as follows: Extract inference conclusion nodes and inference path information from the set of conclusions to be submitted, and construct a support graph. The support graph contains inference conclusion nodes, intermediate hypothesis nodes and observation signal nodes, and directed edges represent the support relationship from evidence nodes to inference conclusion nodes. Starting with the most significant alarm signal, search along the directed edges of the support graph in the forward direction to find a closed path that returns from the inference conclusion node to the most significant alarm signal or its direct predecessor observation signal node. This path is called the closed interpretation loop. Calculate the loop self-consistency score for closed interpretation loops. When the score exceeds the self-consistency alarm threshold, it is marked as a highly self-consistent closed interpretation loop. In the current snapshot of the global cause-effect graph of the production line, compare the probability of each node as an independent root cause before and after the construction of the support graph. Nodes whose probability of hypothesis before the construction of the support graph is higher than the independent root cause attention threshold, whose probability after construction is lower than the threshold and whose decrease exceeds the decrease of the suppression judgment, and whose probability is not directly explained by the inference conclusion node in the support graph are judged as weak abnormal signals that are suppressed and are collected into a suppression signal set. A steady state of erroneous cognition is determined to exist when the following conditions are met simultaneously: there is a highly self-consistent closed explanatory loop, the set of suppression signals is non-empty, and there is a causal path between the inference conclusion node and the suppression signal node in the loop pointing to the suppression signal; otherwise, it is determined not to exist.

9. The automated production line fault-tolerant collaborative control method based on AI fault diagnosis according to claim 1, characterized in that: The way to form virtual constraints is as follows: A temporary graph copy is generated from the consistency snapshot to be committed, and is physically isolated from the major version; In the temporary graph copy, locate the inference conclusion node corresponding to the core diagnostic conclusion and the observation signal node corresponding to the most significant alarm signal at present, and extract the causal edge with the highest support strength between the two as the strong correlation edge locked by the closed loop. For strongly correlated edges locked in closed loops and directed edges in closed interpretation loops that return from the inference conclusion node to the observation signal node or its direct predecessor observation signal node, a forced cut-off operation is performed, the probability intensity parameter is set to zero and marked as a perturbation cut-off state. For each suppressed weak anomalous signal node in the suppressed signal set, the probability of assuming it as an independent root cause is forcibly set as the perturbation injection probability value, and a virtual prior edge is added to the node, marked as the perturbation injection state, forming a virtual constraint condition.

10. The automated production line fault-tolerant collaborative control method based on AI fault diagnosis according to claim 9, characterized in that: The way to output the final correct understanding by performing forced re-reasoning under virtual constraints is as follows: Activate the root cause tracing thread, pass the temporary graph copy access handle, the list of suppressed weak abnormal signal nodes, and the mandatory constraint conditions. The mandatory constraint conditions assume that the device node corresponding to the core diagnostic conclusion is in a healthy state. The root cause tracing thread starts from the suppressed weak abnormal signal node and performs a bidirectional causal search along the directed edges that have not been interrupted. During the expansion process, it calls the equipment process mechanism knowledge base to perform physical consistency verification. The complete causal path that is discovered and passes the self-consistency judgment and physical consistency verification is recorded as the newly discovered causal chain, and a new hypothesis data package is generated. After the new hypothesis data package is verified by the action risk reviewer, it is converted into a causal graph operation sequence and encapsulated as a first-priority correction transaction. Operations to be committed that are related to highly self-consistent closed interpretation loops are discarded. The transaction manager merges the operations to be committed that have not been discarded with the correction transaction and writes them into the main version of the causal graph database engine, generating a correct and consistent version as the final correct cognition.

Images