An intelligent positioning monitoring method for industrial control mainboard faults based on twin simulation

Abstract

The application discloses an intelligent fault positioning and monitoring method for an industrial control mainboard based on twin simulation, and relates to the technical field of digital twin simulation, and comprises the following steps: S1, collecting operation monitoring data and health operation benchmarks to form operation state sequences and health benchmark sequences; S2, establishing a health digital twin; S3, comparing the operation state with the health state to determine an abnormality range to be positioned; S4, inputting the combined upstream abnormal state into an improved MambaTS model, and forming candidate abnormal explanation relationships and residual abnormalities through twin simulation; S5, constructing a candidate abnormal explanation directed graph and determining edge weights; S6, executing a Tarjan algorithm to identify strongly connected components and determine conflict relationships; S7, feeding back the conflict relationships, re-simulating the residual abnormalities, and forming a stable abnormal explanation directed acyclic graph and stable residual abnormalities; and S8, determining a fault position and forming a fault positioning and monitoring result. The application can improve fault positioning accuracy and reliability.

Description

Technical Field

[0001] This invention relates to the field of digital twin simulation technology, and in particular to an intelligent fault location and monitoring method for industrial control motherboards based on twin simulation. Background Technology

[0002] Existing industrial control motherboard fault monitoring methods typically collect data on voltage, current, temperature, clock, interface status, and functional unit operation, and then use threshold judgment, rule matching, or fault tree analysis to identify the location of anomalies. While these methods can perform basic status monitoring, when internal functional units are closely interconnected and anomalies propagate continuously along power supply, clock, control, and signal paths, they can easily mistake downstream anomalies caused by upstream faults for independent faults.

[0003] With the development of digital twin technology, some industrial control equipment has begun to use digital twins to compare the actual operating state with the health state in order to assist in anomaly detection and fault location. However, existing methods usually focus on comparing state deviations and lack analysis of continuous replay and explanatory relationships of the anomaly propagation process, making it difficult to distinguish between propagated anomalies that can be explained by upstream anomalies and residual anomalies generated by the functional unit itself.

[0004] Furthermore, when there is cross-influence or cyclical interpretation among multiple abnormal functional units, existing methods are difficult to effectively screen and correct candidate fault relationships, which can easily lead to an expansion of the fault range, repeated location or mislocation, affecting the accuracy of fault location monitoring of industrial control motherboards.

[0005] Therefore, how to provide a method for intelligent fault location and monitoring of industrial control motherboards based on twin simulation is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an intelligent fault location and monitoring method for industrial control motherboards based on twin simulation. This invention employs a combination of a healthy digital twin, an improved MambaTS model, and the Tarjan algorithm. By using twin simulation to determine the degree to which upstream anomalies explain downstream anomalies, and by inheriting, restoring, and re-inferring the hidden state of the remaining anomalies, it achieves the separation of propagated anomalies from local fault anomalies. On this basis, it identifies and eliminates cyclic conflict relationships through a directed graph of candidate anomaly interpretations, gradually stabilizing the fault interpretation results and achieving accurate location of the industrial control motherboard fault. This method has the advantages of strong fault propagation identification capability, low false location rate, and reliable location results.

[0007] A method for intelligent fault location and monitoring of industrial control motherboards based on twin simulation according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect industrial control motherboard operation monitoring data and health operation benchmarks to form an operation status sequence and a health benchmark sequence;

[0009] S2. Establish a health digital twin containing state dependencies based on the health baseline sequence and motherboard functional connections;

[0010] S3. Compare the running state sequence with the corresponding health digital twin, and determine the range of anomalies to be located along the state dependency relationship;

[0011] S4. Extract the combined upstream anomaly state input from the anomaly range to be located to improve the MambaTS model, retain the hidden state before the current functional unit is processed and temporarily store the candidate hidden state, determine the state inheritance based on the reenactment results of the health digital twin, and form the candidate anomaly interpretation relationship and the remaining anomaly.

[0012] S5. Using the abnormal functional units within the range of the anomaly to be located as nodes, construct a directed graph of candidate anomaly interpretations according to the candidate anomaly interpretation relationships, and determine the edge weights based on the remaining anomalies.

[0013] S6. Perform Tarjan's algorithm on the candidate anomaly interpretation directed graph to identify strongly connected components, determine the conflict relationships in the strongly connected components based on the edge weights, and form an anomaly interpretation directed acyclic graph and conflict relationships.

[0014] S7. Feed the conflict relationship back to the improved MambaTS model, revoke the state inheritance corresponding to the conflict relationship and restore the hidden state before processing, re-enact the remaining anomaly, no longer inherit the hidden state corresponding to the revoked conflict relationship, and form a stable anomaly interpretation directed acyclic graph and stable remaining anomaly.

[0015] S8. Based on the interpretation of the directed acyclic graph and the stable residual anomaly, determine the fault location of the industrial control motherboard and form the fault location monitoring results.

[0016] Optionally, S1 specifically includes:

[0017] S11. Collect the operation monitoring data of each functional unit of the industrial control motherboard, and obtain the health operation benchmark when the corresponding functional unit is in a healthy operating state;

[0018] S12. Time-align the operation monitoring data and the health operation benchmark according to the sampling time, and register the operation monitoring data of the same functional unit with the health operation benchmark to the same sampling time;

[0019] S13. Arrange the time-aligned operation monitoring data and health operation benchmarks according to the sampling time sequence to form the operation status sequence and health benchmark sequence.

[0020] Optionally, S2 specifically includes:

[0021] S21. Map the state data in the health baseline sequence to the digital twin state of the corresponding functional unit according to the functional unit to form the health twin state;

[0022] S22. According to the signal transmission direction of the motherboard function connection, associate the output state of the upstream functional unit with the input state of the downstream functional unit to form a state dependency relationship.

[0023] S23. Associate the health twin states of each functional unit according to the state dependency relationship to form a health digital twin.

[0024] Optionally, S3 specifically includes:

[0025] S31. Calculate the absolute difference between the operating status sequence and the corresponding health status of the health digital twin according to the functional unit and sampling time to form the operating deviation sequence;

[0026] S32. The maximum deviation between the health baseline sequence and the corresponding health status of the health digital twin is used as the health deviation limit. Functional units whose operating deviation exceeds the health deviation limit are identified as abnormal functional units.

[0027] S33. Include abnormal functional units and the direct state transmission between abnormal functional units in the state dependency relationship into the scope of abnormalities to be located.

[0028] Optionally, the improved MambaTS model includes an instance normalization and embedding layer, a variable scanning layer, a temporal Mamba encoding layer, and a prediction head;

[0029] The instance normalization and embedding layer performs instance normalization on the corresponding functional unit states in the joint upstream abnormal state and the healthy digital twin, divides the time into blocks along the sampling time sequence, and forms state labels through linear mapping.

[0030] The variable scanning layer arranges the status markers of each functional unit in a staggered manner according to time blocks, forming a variable scanning sequence along time;

[0031] The temporal Mamba coding layer performs selective state-space recursion after linear mapping of the variable scan sequence and adds it to the gated nonlinear mapping result. It retains the hidden state before processing the current functional unit and temporarily stores the candidate hidden state after the current functional unit completes the state update.

[0032] The prediction head maps the encoding results corresponding to the candidate hidden states to the twin replay state of the current functional unit in the health digital twin independently according to the functional unit; compares the twin replay state with the current functional unit's operating state, determines the inheritance of hidden states based on the comparison differences, and forms the candidate anomaly interpretation relationship and the remaining anomalies.

[0033] Optionally, the hidden state inheritance specifically includes:

[0034] When processing along the variable scanning sequence to the current functional unit, the hidden state before entering the current functional unit is saved. The current functional unit's state flag is recursively obtained through selective state space to obtain the candidate hidden state. The candidate hidden state is temporarily stored and not written back as the input hidden state of the next functional unit.

[0035] The prediction head generates a twin complex state based on the candidate hidden state. The twin complex state is compared with the current functional unit's operating state at the same sampling time. The maximum deviation between the health baseline sequence and the corresponding state of the health digital twin is used as the health deviation limit.

[0036] The change in the twin's recreated state relative to the corresponding state of the healthy digital twin is subtracted from the change in the current functional unit's operating state relative to the corresponding state of the healthy digital twin to obtain the remaining state difference. State changes whose amplitude exceeds the health deviation limit in the remaining state difference are identified as remaining anomalies.

[0037] When there are no state changes in the remaining state differences whose amplitude exceeds the health deviation limit, the candidate hidden state is written back as the input hidden state of the next functional unit, and the candidate anomaly interpretation relationship is formed based on the corresponding changes of the joint upstream abnormal state and the actual anomaly of the current functional unit.

[0038] When there are state changes in the remaining state differences whose amplitude exceeds the health deviation limit, the candidate hidden state is not written back. Instead, the saved hidden state before processing is read, and the corresponding state of the health digital twin is superimposed with the remaining anomaly to form the reconstruction state of the current functional unit. The reconstruction state is then formed by instance normalization and embedding.

[0039] The selective state space recursion of the current functional unit is re-executed using the saved pre-processing hidden state and reconstructed state markers. The regenerated hidden state is written back as the input hidden state of the next functional unit. Candidate anomaly interpretation relationships are formed based on the change in the twin replay state relative to the corresponding state of the healthy digital twin, and the remaining anomalies are retained.

[0040] Optionally, S5 specifically includes:

[0041] S51. Using the abnormal functional units within the range of the anomaly to be located as nodes, point the upstream abnormal functional units in the same candidate anomaly interpretation relationship to the anomaly functional units to be interpreted, forming a directed graph of candidate anomaly interpretation.

[0042] S52. Accumulate the deviation amplitude of the operating state of the explained functional unit relative to the corresponding state of the healthy digital twin and the corresponding residual abnormal amplitude, and use the difference between the two as the edge weight of each directed edge corresponding to the same candidate abnormality explanation relationship.

[0043] Optionally, S6 specifically includes:

[0044] S61. Perform Tarjan's algorithm on the directed graph for candidate anomaly interpretation, and perform a depth-first search according to the direction of the directed edges to identify strongly connected components with mutually reachable nodes.

[0045] S62. In a strongly connected component containing multiple abnormal functional units, compare the edge weights of the directed edges corresponding to the candidate abnormal interpretation relations, determine the candidate abnormal interpretation relations with the smallest edge weight as the conflict relation, and delete all directed edges of the corresponding candidate abnormal interpretation relations.

[0046] S63. Repeatedly identify strongly connected components and delete directed edges corresponding to conflict relationships until the candidate anomaly interpretation directed graph does not contain strongly connected components containing multiple anomaly functional units, forming an anomaly interpretation directed acyclic graph and conflict relationships.

[0047] Optionally, S7 specifically includes:

[0048] S71. Locate the downstream functional unit being interpreted based on the conflict relationship, cancel the corresponding hidden state write-back, and restore the hidden state saved before the downstream functional unit's processing.

[0049] S72. Overlay the corresponding state of the health digital twin with the remaining anomalies, reconstruct the operating state of the downstream functional units, and re-execute the twin reenactment from the restored hidden state.

[0050] S73. Based on the candidate anomaly interpretation relationship formed by the re-replay and the remaining anomalies, update the directed graph of candidate anomaly interpretations and edge weights, and execute the Tarjan algorithm again to update the directed acyclic graph of anomaly interpretations and conflict relationships.

[0051] S74. Repeatedly revoke and reenact newly generated conflict relationships, and do not write back the hidden states corresponding to the revoked conflict relationships, until no new conflict relationships are generated, forming a stable anomaly interpretation directed acyclic graph and a stable residual anomaly.

[0052] Optionally, S8 specifically includes:

[0053] S81. Locate the abnormal function unit that retains the stable remaining abnormal in the directed acyclic graph of stable anomaly interpretation, and determine the corresponding abnormal function unit as the fault location of the industrial control motherboard.

[0054] S82. Extract the state changes of the corresponding functional units in the operating state sequence according to the fault location to form the fault location monitoring results.

[0055] The beneficial effects of this invention are:

[0056] This invention uses a health digital twin to uniformly describe the health status and state dependencies of each functional unit of the industrial control motherboard, and compares the actual operating status with the health status. This allows for a rapid narrowing of the range of anomalies to be located along the functional connections of the motherboard, avoiding the problem of an excessively large range of anomalies due to fault judgment based solely on a single point of deviation. This provides an accurate data foundation for subsequent fault propagation analysis and fault location determination.

[0057] This invention improves the MambaTS model to perform twin replay of joint upstream anomalies. During the current functional unit processing, candidate hidden states are temporarily stored, and the inheritance, recovery, and re-progression of hidden states are controlled based on the replay results. The propagation part that can be explained by upstream anomalies and the remaining anomalies that cannot be explained are separated from the actual anomalies. This reduces the probability of misjudging downstream propagation anomalies as independent faults and improves the accuracy of fault location in multi-level fault propagation.

[0058] Furthermore, this invention utilizes the Tarjan algorithm to identify cyclic interpretation relationships in the directed graph of candidate anomaly interpretations, and combines edge weights to filter out conflict relationships. The conflict results are then fed back to the improved MambaTS model for re-enactment, gradually stabilizing the anomaly interpretation relationships. While preserving the common influence of multiple upstream factors, it eliminates cyclic conflicts and finally determines the fault location based on the stable remaining anomalies. Therefore, it has the beneficial effects of strong fault propagation identification capability, low false location rate, stable and reliable location results, and applicability to online fault monitoring of industrial control motherboards. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0060] Figure 1 This is an overall flowchart of an intelligent fault location and monitoring method for industrial control motherboards based on twin simulation proposed in this invention.

[0061] Figure 2 This is a schematic diagram of the improved MambaTS model of the intelligent fault location and monitoring method for industrial control motherboards based on twin simulation proposed in this invention.

[0062] Figure 3 This is a schematic diagram of the twin simulation-driven anomaly interpretation, resolution, and state rollback closed loop of the intelligent fault location and monitoring method for industrial control motherboards based on twin simulation proposed in this invention. Detailed Implementation

[0063] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0064] refer to Figures 1 to 3 A method for intelligent fault location and monitoring of industrial control motherboards based on twin simulation includes the following steps:

[0065] S1. Collect industrial control motherboard operation monitoring data and health operation benchmarks to form an operation status sequence and a health benchmark sequence;

[0066] S2. Establish a health digital twin containing state dependencies based on the health baseline sequence and motherboard functional connections;

[0067] S3. Compare the running state sequence with the corresponding health digital twin, and determine the range of anomalies to be located along the state dependency relationship;

[0068] S4. Extract the combined upstream anomaly state input from the anomaly range to be located to improve the MambaTS model, retain the hidden state before the current functional unit is processed and temporarily store the candidate hidden state, determine the state inheritance based on the reenactment results of the health digital twin, and form the candidate anomaly interpretation relationship and the remaining anomaly.

[0069] S5. Using the abnormal functional units within the range of the anomaly to be located as nodes, construct a directed graph of candidate anomaly interpretations according to the candidate anomaly interpretation relationships, and determine the edge weights based on the remaining anomalies.

[0070] S6. Perform Tarjan's algorithm on the candidate anomaly interpretation directed graph to identify strongly connected components, determine the conflict relationships in the strongly connected components based on the edge weights, and form an anomaly interpretation directed acyclic graph and conflict relationships.

[0071] S7. Feed the conflict relationship back to the improved MambaTS model, revoke the state inheritance corresponding to the conflict relationship and restore the hidden state before processing, re-enact the remaining anomaly, no longer inherit the hidden state corresponding to the revoked conflict relationship, and form a stable anomaly interpretation directed acyclic graph and stable remaining anomaly.

[0072] S8. Based on the interpretation of the directed acyclic graph and the stable residual anomaly, determine the fault location of the industrial control motherboard and form the fault location monitoring results.

[0073] In this embodiment, S1 specifically includes:

[0074] The functional units of the industrial control motherboard are divided according to the functional boundaries that can independently form input and output states in the motherboard's connection principle. Power transmission, clock transmission, reset control, processing and calculation, storage access, and interface communication are respectively assigned to the corresponding functional units. Operation monitoring data is obtained from the onboard monitoring interface, test interface, and controller acquisition channel. Each data record retains the corresponding functional unit, status variable, sampling time, and sampled value. The healthy operation benchmark is collected by the same model of industrial control motherboard that has been confirmed to be normal by functional testing under the same operating conditions. The status variables in the healthy operation benchmark maintain the same name, unit, and collection object as the operation monitoring data, thereby ensuring that the actual operating state of the same functional unit can establish a one-to-one data foundation with the health state.

[0075] Time alignment uses the sampling time of the operation monitoring data as a common time axis. For each state variable of the same functional unit, the corresponding sampled value in the health operation benchmark is searched. When the same sampling time exists in the health operation benchmark, the corresponding sampled value is directly registered. When there is no corresponding sampled value between adjacent health sampling times for continuous numerical states, linear interpolation is performed with the time interval between the target sampling time and the previous and next health sampling times as the weight. The closer the target time is, the greater the proportion of the health sampled value. For discrete states such as reset level and interface on / off status, the most recent valid health status value before the target sampling time is used. When two health sampled values ​​cannot be obtained at the boundary of the common time range for continuous numerical states, the corresponding boundary sampling time is deleted. In this way, the operation monitoring data and health operation benchmark of the same functional unit at the same sampling time are obtained.

[0076] After time alignment is completed, the operation monitoring data of each functional unit are arranged from early to late according to the sampling time. Within the same sampling time, the correspondence between each state variable and the original data collection object is maintained, and the data of each sampling time are continuously connected to form an operation status sequence. The registered health operation benchmarks are arranged in the same sampling time order to form a health benchmark sequence. The two sequences maintain correspondence at the three levels of functional unit, state variable and sampling time. The formed operation status sequence and health benchmark sequence enter the process of establishing a health digital twin and comparing operation status.

[0077] In this embodiment, S2 specifically includes:

[0078] The health twin state uses functional units as the state mapping unit. It reads the time-aligned health state data from the health baseline sequence and writes the state variables into the digital twin state of the corresponding functional unit at each sampling time. When writing, the state variable name, physical unit, sampling time, and functional unit affiliation are kept consistent with the health baseline sequence. When the same functional unit contains multiple state variables, each state variable occupies an independent state position and retains its original input, internal operation, or output attributes. All state variables at the same sampling time constitute the health twin state of the current functional unit. After completing the mapping of all functional units, a set of health twin states that can correspond to each item in the health baseline sequence is obtained.

[0079] The state dependency relationship is jointly determined by the principle connection of the industrial control motherboard, the interface signal connection, and the already determined input-output correspondence of the functional units. The functional units and signal transmission direction at both ends of each motherboard functional connection are read, and the end that generates the signal is determined as the upstream functional unit, and the end that receives the same signal is determined as the downstream functional unit. Then, the output state of the upstream functional unit is connected to the input state of the downstream functional unit that receives the same physical signal. When multiple upstream functional units transmit different signals to the same downstream functional unit, the output state of each upstream unit is connected to the corresponding downstream input state. Different physical quantities are not merged. In this way, the original multiple upstream input relationship of the industrial control motherboard is preserved and a directional state dependency relationship is formed.

[0080] The health digital twin is composed of health twin states and state dependencies. At each sampling time, the initial values ​​of the digital twin states of each functional unit are first assigned using the health benchmark sequence. Then, the output state of the upstream functional unit is read along the signal transmission direction of the state dependency relationship and written into the corresponding input state of the downstream functional unit. The same transmitted signal only updates the input state position corresponding to that signal. After completing the direct association state update, the current digital twin state of each functional unit is retained as the corresponding health state of the health digital twin. The health benchmark sequence records the actual health operation data, and the health state corresponding to the health digital twin reflects the twin output of the health data after functional connection mapping and state transmission. The two maintain the correspondence between functional units, state variables and sampling times, and serve as the data basis for comparing operating states and calculating health deviation limits.

[0081] In this embodiment, S3 specifically includes:

[0082] The operational deviation sequence compares the operational state sequence corresponding to the completed time with the health state corresponding to the health digital twin. Differences are calculated separately for the same functional unit, the same sampling time, and the same state variable position. For continuous numerical states, the corresponding health state value is subtracted from the operational value and the absolute value is taken. For discrete states, the operational value is recorded as 0 when it matches the health state and as 1 when it does not match. Different state variables retain their respective deviations without cross-dimensional addition. The deviation values ​​are then arranged in the order of the original sampling time to form the operational deviation sequence. Each position in the operational deviation sequence can be associated with a specific functional unit, state variable, and sampling time and enters the abnormal functional unit judgment process.

[0083] The health deviation limits are calculated for each state variable of each functional unit. Within the healthy operating range covered by the health benchmark sequence, the absolute deviation between the measured health status and the corresponding health status of the health digital twin in the same functional unit, the same state variable, and the same sampling time is calculated item by item. Then, the maximum value among all health deviations of the same state variable is taken as the health deviation limit of the corresponding state variable. This preserves the normal modeling deviation between the measured health data and the twin output. When the deviation of any state variable in the operating deviation sequence exceeds its own health deviation limit at any sampling time, the corresponding functional unit is identified as an abnormal functional unit. If there are multiple out-of-bounds states in the same functional unit, it is still retained as an abnormal functional unit. The out-of-bounds state and the corresponding sampling time are used as the state basis for the current abnormal position.

[0084] The anomaly range to be located is composed of the already determined abnormal functional units and the original direct state dependencies between them. Reading the state dependency relationship established above, only the connection where the starting point and ending point both belong to the abnormal functional unit and there is direct signal transmission is retained, thus forming the anomaly range to be located that can represent the abnormal functional unit and its direct propagation direction. Abnormal functional units that do not have direct connections with other abnormal functional units are still retained separately in the anomaly range to be located, to avoid independent local anomalies being excluded due to the lack of adjacent anomaly nodes. The formed anomaly range to be located serves as the input boundary for joint upstream anomaly state extraction and improved MambaTS model processing.

[0085] In this embodiment, the improved MambaTS model includes an instance normalization and embedding layer, a variable scanning layer, a temporal Mamba coding layer, and a prediction head:

[0086] Existing MambaTS models are designed for multivariate time series prediction. Typically, they first perform instance normalization and time block division on each variable, obtain state labels through linear embedding, and then use a variable scanning method along time to interleave the state labels of different variables at the same time position. The encoding part continuously processes the scan sequence by temporal Mamba blocks. The temporal Mamba blocks use selective state space branching to extract temporal dependencies and add the result of this branch to the result of gated nonlinear branching. The prediction end then independently performs linear mapping on each variable to obtain the prediction sequence. In the forward propagation process of the scan sequence, the original model continuously generates the current hidden state from the hidden state of the previous position and continues to use the current hidden state as the historical state of the next position. This continuous recursive method is suitable for long-term series prediction, but when applied to fault propagation and localization of industrial control motherboards, the state changes caused by upstream anomalies and the anomalies generated by the current functional unit itself will continuously accumulate forward along the same hidden state. The model lacks a processing position to determine whether the hidden state should continue to be inherited after confirming the source of the anomaly at the current functional unit position.

[0087] The improved MambaTS model's input consists of the joint upstream abnormal state and the corresponding state of the current functional unit in the health digital twin. The joint upstream abnormal state is obtained from the range of anomalies to be located. For the current functional unit to be processed, the abnormal upstream that directly points to the current functional unit is read from the state dependency relationship. At the same sampling time, the abnormal state variables of each upstream functional unit are extracted and combined along the fixed arrangement order of functional units in the health digital twin. When there is only one abnormal upstream, the upstream state sequence is directly retained. When there are multiple abnormal upstreams, the upstream states are arranged sequentially at the same sampling time while maintaining the original functional unit affiliation of each state variable. The combination result shares the same time axis with the health digital twin state of the current functional unit. The improved MambaTS model consists of instance normalization and embedding layer, variable scanning layer, temporal Mamba encoding layer and prediction head connected sequentially. The state label or encoding result formed by the previous layer is directly used as the input of the next layer. The temporal Mamba encoding layer adds hidden state storage, candidate state temporary storage and state write-back connection, so that the model output can enter the twin replay and hidden state inheritance judgment.

[0088] Within the instance normalization and embedding layer, instance normalization, temporal block partitioning, and linear embedding are performed sequentially. Instance normalization calculates the mean and variance of all sampled values ​​within the current input window for each state variable in the joint upstream abnormal state and the corresponding state of the healthy digital twin. In this implementation, 0.00001 is added as a numerical stability term to the variance, and the square root is taken to obtain the normalization denominator. Each sampled value is subtracted from the mean of the same state variable and then divided by the normalization denominator to obtain the normalized value. At the same time, the mean and normalization denominator of each state variable are saved for the prediction results to restore the original physical scale. After normalization, the continuous sampled values ​​are divided into blocks of equal time length along a common time axis. Each state variable uses the same start and end time of the block. The sampled values ​​in the same block are expanded into an input vector according to the time sequence, and then multiplied by the linear embedding weights item by item and accumulated along the input dimension. The corresponding bias is superimposed to form a state label. The linear embedding weights and biases are obtained by model training. The state label dimension is consistent with the input dimension of the temporal Mamba encoding layer. The training data and the actual monitoring data use the same temporal block partitioning method.

[0089] The variable scanning layer internally completes time block grouping and variable scanning along time. First, it reads the time block number and functional unit affiliation corresponding to each state marker. Within the same time block, it arranges the state markers of each functional unit in the order of the functional units whose abnormal states are fixed in conjunction with the upstream. Then, it connects the state markers corresponding to the current functional unit's health digital twin state. After completing the arrangement of one time block, it moves to the next time block and continues to arrange using the same functional unit order. In this way, the time series that originally belonged to different functional units are rearranged into a variable scanning sequence that unfolds continuously along time. Each scanning position retains the time block number, functional unit affiliation, and the correspondence between state variables. Therefore, when the time Mamba encoding layer processes any position, it can determine which functional unit the current encoding result belongs to and can determine the functional unit corresponding to the next scanning position. Functional units that lack a valid state value at a certain sampling time do not generate a corresponding state marker. The scanning sequence is directly connected to the next valid state marker in that time block, avoiding invalid padding values ​​from entering the selective state space for recursion.

[0090] The temporal Mamba coding layer consists of an input linear mapping, a selective state-space branch, a gated nonlinear branch, result summation, and a hidden state storage and write-back connection. The current state marker in the variable scan sequence first enters two linear mappings. The selective state-space branch uses the current state marker to generate selection parameters and applies them to the discrete state space recursively. During recursion, the hidden state from the previous scan position and the current input are used to calculate the current state. The state transition parameters apply to the previous hidden state, and the input parameters apply to the current linear mapping result. The two parts are accumulated dimension-wise to form the current candidate hidden state. The gated nonlinear branch passes the result of the other linear mapping through a gated nonlinear function. The output is compressed to a limited range and then added dimension by dimension to the encoding result at the corresponding position of the selective state space branch to form the current encoding result. The improved hidden state connection is set between the recursive result of the selective state space and the input of the next scan position. Before entering the current functional unit, the existing hidden state is saved. The hidden state generated by the current recursion is first recorded as a candidate hidden state and kept in a temporary storage position. The candidate hidden state does not directly change the input hidden state of the next scan position before the twin repetition is completed. Whether the state is written back to the comparison result obtained by the twin repetition is controlled. The specific execution process of the remaining exception calculation, the hidden state before recovery processing, and the recursion is further elaborated in the hidden state inheritance part.

[0091] The prediction head employs linear readout with independent settings for each functional unit. The encoding result corresponding to the current functional unit is multiplied by the prediction weight matrix of that functional unit and accumulated along the hidden feature dimension. The output bias is then superimposed to obtain the prediction state at the normalized scale. Each output position corresponds to a state variable of the current functional unit. The rows of the prediction weight matrix correspond to the output state variables, and the columns correspond to the hidden feature dimensions of the time Mamba coding layer. The prediction result is multiplied by the normalized denominator of the same state variable saved in the instance normalization stage, and the corresponding mean is added to restore it to the original physical scale. It is then restored to the corresponding sampling time according to the time block position. This forms a twin replay state that corresponds one-to-one with the current functional unit state of the health digital twin in the three dimensions of functional unit, state variable, and sampling time. The comparison result between the twin replay state and the current functional unit running state enters the hidden state inheritance processing, and this processing forms the candidate anomaly interpretation relationship and the remaining anomaly.

[0092] The training samples for the improved MambaTS model consist of operational records from industrial control motherboards with confirmed states. The input uses a joint upstream state sequence consistent with the actual positioning process and a digital twin state of the current functional unit's health. When the predicted functional unit confirms its normal operation through independent functional testing, its corresponding actual operating state serves as the supervision target, enabling the model to learn the actual response that can be generated when upstream state changes are transmitted to healthy downstream functional units. During training, the input sequence is first generated using the same instance normalization, temporal block partitioning, and variable scanning methods as inference. The prediction head output is compared item by item with the supervision target at the same functional unit, the same state variable, and the same sampling time. The squared difference between the predicted value and the supervision value at each position is averaged to form the training loss. Then, backpropagation is used to update the linear embedding weights and biases, and the linear embedding weights and biases in the temporal Mamba encoding layer. The mapping parameters, selective state space parameters, gating branch parameters, and prediction head weights and biases, as well as the hidden state saving, candidate state temporary storage, and state write-back control, belong to the deterministic forward processing connection and are not set with trainable weights. The Adam optimizer is used to update the trainable parameters, and the two momentum decay parameters are set to 0.9 and 0.999 respectively. A maximum of 10 training rounds are executed. Training ends when the average prediction error on the validation set does not decrease for 3 consecutive rounds, and the model parameters with the lowest average prediction error on the validation set are retained. The temporal block length and hidden feature dimension adopt the fixed structure parameters in the model training configuration. The training set, validation set, and actual monitoring process maintain the same configuration. After each round of training, the average prediction error is calculated on the validation set and the model parameters corresponding to the current lowest error are saved. After training, the parameters with the lowest validation error are used to perform fault localization.

[0093] Compared to the existing MambaTS, the improved MambaTS model adds a hidden state inheritance boundary between the selective state space recursion result of the temporal Mamba coding layer and the next scan position. The temporal hidden state, which was continuously passed by the original model, is transformed into a fault propagation state that is confirmed by twin replay before entering the next scan position. The health state provided by the health digital twin and the model replay state jointly determine whether the current hidden state retains the original propagation source. This internal path change embeds the fault source judgment into the state propagation position of the selective state space recursion, so that upstream propagation anomalies and new anomalies generated by the current functional unit have independent data destinations before entering the next recursion position. At the same time, it retains the hidden state before processing as a state boundary that can be directly recovered when there is conflict feedback. Thus, the temporal Mamba coding layer can simultaneously undertake long temporal state propagation and fault propagation source preservation.

[0094] In this embodiment, hidden state inheritance specifically includes:

[0095] Hidden state inheritance occurs before the selective state space recursion result of the temporal Mamba coding layer is written to the next scan position. When the variable scan sequence processes the current functional unit, it first reads the historical hidden state that entered the current scan position and saves the hidden state along with the current scan position and the functional unit identifier. The dimension of the saved hidden state is consistent with the internal state dimension of the selective state space recursion, and each state position maintains the original recursion channel correspondence. The current functional unit state label enters the selective state space recursion after being linearly mapped by the input. The state transition parameters act on the saved hidden state, and the input parameters act on the mapping result of the current state label. The two parts are accumulated channel by channel to form a candidate hidden state. The candidate hidden state is only saved as a temporary recursion result of the current scan position. It does not overwrite the hidden state before processing, nor is it written to the next scan position. In this way, the state path that was continuously passed immediately after the recursion in the original MambaTS is truncated at the current functional unit position.

[0096] The twin replay reads the current encoding result corresponding to the candidate hidden state from the prediction head and restores it to the original state space of the current functional unit. Each output position inside the prediction head corresponds one-to-one with the monitored state variable of the current functional unit. The output value is multiplied by the normalized denominator of the same state variable saved in the instance normalization stage, and the corresponding mean is added to restore the physical scale. Then it is written back to the same sampling time to obtain the twin replay state. The state corresponding to the health digital twin represents the health output of the current functional unit under healthy operation and predetermined state dependency transmission. The current functional unit operation state represents the actual value collected by the industrial control motherboard. The twin replay state represents the predicted response that can be generated in the current functional unit after the combined upstream abnormal state is propagated through the improved MambaTS. The three states use the same functional unit, state variable and sampling time index. Therefore, the comparison process is performed item by item on the same state variable and the same sampling time.

[0097] The health deviation limits are directly adopted from the maximum normal deviation of each state variable calculated from the health baseline sequence and the corresponding health status of the health digital twin. Each state variable is judged only using its own health deviation limit. For each sampling time of the current functional unit, the actual state change is obtained by subtracting the corresponding state of the health digital twin from the actual operating state. Then, the same health status is obtained by subtracting the twin's recreated state from the recreated state. Subsequently, the recreated state change is subtracted from the actual state change item by item to obtain the remaining state difference. During the calculation, the positive and negative directions of the difference are retained. Positions where the absolute value does not exceed the corresponding health deviation limit are recorded as normal remainders and set to 0. Positions where the absolute value exceeds the corresponding health deviation limit retain the original signed difference. At the same time, the corresponding state variable and sampling time are recorded. These out-of-bounds state changes constitute the remaining anomalies.

[0098] When there are no state changes exceeding the health deviation limit in the remaining state differences, the actual anomaly of the current functional unit can be covered by the state changes obtained from the joint upstream anomaly state through twin replay within the allowable range of health error. At this time, the temporary state of the candidate hidden state is released, and the candidate hidden state is written into the next scan position as the input hidden state. All upstream anomaly functional units corresponding to the joint upstream anomaly state are used as the source of explanation, the current functional unit is used as the object of explanation, and the effective changes of the twin replay state relative to the state corresponding to the healthy digital twin are used as the explanation content to establish the candidate anomaly explanation relationship. When multiple upstream anomaly functional units participate in the same replay, they belong to the same candidate anomaly explanation relationship, thereby maintaining the consistent correspondence between the joint upstream action on the model output and the subsequent anomaly explanation graph.

[0099] When there are state changes in the remaining state differences that exceed the health deviation limit, the candidate hidden state stops writing back and retains the hidden state before processing as the starting point for the recalculation of the current functional unit. The current functional unit reconstructs the state based on the state corresponding to the healthy digital twin. For state variables with remaining anomalies and sampling times, the aforementioned signed remaining anomaly values ​​are added. For positions without remaining anomalies, the state corresponding to the healthy digital twin is directly retained. This results in a reconstructed state that only retains the part of the current functional unit that cannot be explained by the joint upstream anomaly. The reconstruction process no longer writes the upstream propagation changes that have already been explained by the twin replay. Therefore, the local remaining anomalies of the current functional unit and the confirmed upstream explanation part are separated before entering the recalculation.

[0100] When the reconstructed state enters the instance normalization and embedding layer, the previously determined state variable arrangement, time block boundaries, and linear embedding parameters are used. Within the same input time window as the initial processing, the mean and variance of each state variable are recalculated. The square root of the variance after adding 0.00001 is obtained to get the normalization denominator. Instance normalization is completed by subtracting the corresponding mean from each sample value and dividing by the normalization denominator. Then, it is expanded according to the original time block position and multiplied item by item with the same linear embedding weights. The bias is accumulated to form the reconstructed state label. The reconstructed state label is written back to the original scan position of the current functional unit and retains the original time block number and functional unit affiliation. Therefore, the re-recursion will not change the variable scan order. Only the input content of the current functional unit changes from the original state label containing the upstream propagation influence to the reconstructed state label composed of the healthy state and the remaining anomalies.

[0101] When selective state-space recursion is re-executed, the hidden state saved before processing the current functional unit is used as the initial state for recursion, and the result of the reconstructed state label after linear mapping is used as the current input. The state transition parameters continue to act on the saved hidden state, and the input parameters act on the mapping result of the reconstructed state label. The two parts perform the same recursive calculation along the original state-space channel and generate a new hidden state. The new hidden state is directly written into the next scan position as the input hidden state, so that the remaining anomalies become the source of state propagation downstream from the current functional unit again. This round of recursion does not change the candidate anomaly explanation content that has been confirmed by the first twin replay. The candidate anomaly explanation relationship is still determined by the effective change of the state corresponding to the healthy digital twin in the first twin replay state. The remaining anomalies continue to propagate to the next functional unit with the regenerated hidden state, thereby simultaneously retaining the upstream explainable anomaly source and the new anomaly source generated by the current functional unit.

[0102] When there is no upstream anomaly in the current functional unit, the out-of-bounds change of the actual operating state relative to the state corresponding to the healthy digital twin is directly taken as the remaining anomaly. No upstream candidate anomaly interpretation relationship is established. The same healthy state is superimposed with the remaining anomaly to form a reconstructed state before entering the selective state space recursion. When multiple upstream anomalies work together, the joint upstream anomaly states are entered into the model in the order of functional units determined above. Each upstream anomaly functional unit corresponds to the same candidate anomaly interpretation relationship. This avoids splitting the joint propagation effect into independent interpretation sources within the model. The candidate anomaly interpretation relationship formed by the current processing and the remaining anomaly are respectively entered into the candidate anomaly interpretation directed graph construction and the state recursion process of the next functional unit.

[0103] Hidden state inheritance transforms the original continuous state propagation of the selective state-space model into conditional recursion with fault source boundaries. Candidate hidden states only continue to carry the original propagation source if the current anomaly can be explained to the healthy deviation range by existing upstream anomalies. When a residual anomaly occurs, the hidden state before the current functional unit is processed is restored, and the recursive source is re-established by the residual anomaly. Thus, in addition to storing time series historical information, the hidden state also records whether the current propagation source has been verified by twin replay. Upstream propagation components and newly emerging local anomalies will not be fixed into the same recursive source before the next scan position. At the same time, the stored pre-processing state provides an accurate backoff position for cyclic explanation of conflict feedback, enabling the model to maintain the separability and recoverability of fault sources during continuous recursion.

[0104] In this embodiment, S5 specifically includes:

[0105] The directed graph for candidate anomaly interpretation uses the identified abnormal functional units within the range of the anomaly to be located as nodes. The candidate anomaly interpretation relationship directly inherits the aforementioned twin replay results. Each candidate anomaly interpretation relationship retains the set of upstream abnormal functional units that jointly participated in the replay and the current functional unit being interpreted, and uses the sampling interval used in this replay as the effective interval of the relationship. When establishing directed edges, each upstream abnormal functional unit in the same candidate anomaly interpretation relationship is connected to the functional unit being interpreted, and the edge direction is consistent with the anomaly interpretation direction. When multiple upstream abnormal functional units jointly interpret the same functional unit, multiple incoming edges are formed, but these incoming edges continue to belong to the same candidate anomaly interpretation relationship. Therefore, the graph retains the joint action of multiple upstream units of the industrial control motherboard, and can treat the same joint interpretation as a complete object in subsequent conflict handling.

[0106] The weight of the directed edge reflects the amount of abnormal change that has been explained by the corresponding candidate anomaly explanation relation. The calculation range adopts the sampling interval and state variables that actually participated in the twin simulation when the candidate anomaly explanation relation was formed. First, the absolute deviation of the operating state relative to the corresponding state of the healthy digital twin is calculated for each state variable of the explained functional unit. Then, the absolute deviation is divided by the health deviation limit of the same state variable to complete the dimensionless transformation. The remaining anomalies are dimensionless using the health deviation limit of the same state variable and the same sampling time. Within the effective interval of the relation, all dimensionless operating deviations and dimensionless remaining anomalies are summed respectively. Then, the sum of the former cumulative value and the latter cumulative value are subtracted to obtain the explanation amount of the candidate anomaly explanation relation. When the health deviation limit is 0, the smallest adjacent reading difference that the corresponding state variable acquisition channel can distinguish is used as the normalization denominator. This difference is directly obtained by the data resolution capability of the operation monitoring data acquisition interface, thereby maintaining the comparability of the deviations after normalization of different state variables.

[0107] When the same candidate anomaly interpretation relationship contains multiple upstream anomaly functional units, the calculated same interpretation quantity is assigned to all directed edges corresponding to the relationship, so that these edges maintain a consistent interpretation strength in the graph. For a candidate anomaly interpretation relationship formed by a single upstream, the interpretation quantity is directly assigned to the unique directed edge. After completing the node connection and edge weight writing of all candidate anomaly interpretation relationships, a candidate anomaly interpretation directed graph with edge weights is obtained. The nodes, edge directions, candidate anomaly interpretation relationship affiliation, and edge weights in the graph serve as inputs for strongly connected component identification and cyclic conflict relationship determination.

[0108] In this embodiment, S6 specifically includes:

[0109] After the candidate anomaly interpretation directed graph enters the Tarjan algorithm, it takes the unvisited anomaly functional unit as the current search starting point and performs a depth-first search along the directed edges pointing to the interpreted functional units. For each visited node, it records the visit order value, the earliest backtracking order value, and the stack state. The visit order value increases continuously according to the order in which the node first enters the search process. The current node is pushed onto the search stack immediately after entering the search. When searching each outgoing edge of the current node, if the target node has not yet been visited, the search continues recursively towards the target node. After the target node is returned from the search, the smaller value between the current earliest backtracking order value and the target node's earliest backtracking order value is taken to update the current node. If the target node is already in the search stack, the smaller value between the current earliest backtracking order value and the target node's visit order value is taken to complete the update. In this way, the cyclic path that can return to an earlier node along the interpretation direction is kept in the same search chain.

[0110] After all outgoing edges of the current node have been searched, the access order value of the current node is compared with the earliest backtrackable order value. If they are equal, the current node is used as the search root of a strongly connected component. Nodes are popped from the top of the search stack in sequence, and their corresponding push states are cleared, until the current node is popped. All popped nodes form a strongly connected component. The source of the candidate anomaly interpretation relationship and the object being interpreted belong to different functional units. Therefore, the directed graph of candidate anomaly interpretation does not establish directed edges from nodes to themselves. When a strongly connected component contains only one anomaly functional unit, the relevant candidate anomaly interpretation relationship is directly retained. When a strongly connected component contains multiple anomaly functional units, it indicates that these functional units can reach each other along the candidate anomaly interpretation direction. The current interpretation relationship forms a loop and enters the edge weight comparison.

[0111] Conflict determination in strongly connected components uses candidate anomaly interpretation relationships as the complete comparison object. First, the relationship affiliation identifier of the directed edges that make up the strongly connected component is read. Then, the interpretation amount of each directed edge, which has been calculated and written in the previous text, is obtained as the edge weight. When the same candidate anomaly interpretation relationship contains multiple upstream anomaly functional units, the corresponding directed edges have the same edge weight and are all classified into the same relationship. Therefore, the edge weight of the candidate anomaly interpretation relationship is recorded only once during the comparison. The edge weights of each candidate anomaly interpretation relationship in the strongly connected component are arranged from smallest to largest. The smallest edge weight indicates that the relationship can explain the least amount of anomaly change after deducting the remaining anomalies from the total anomalies of the explained functional units. The candidate anomaly interpretation relationship is determined as the conflict relationship in this round, and all directed edges belonging to this relationship are deleted. When there are the same smallest edge weight, the candidate anomaly interpretation relationship is selected as the conflict relationship in this round according to the order in which it was formed in the twin replay process. This ensures that the deletion object in each round is unique and can determine the graph update result in the next round.

[0112] After a conflict relationship is deleted, the remaining nodes and directed edges are used to reconstruct the current candidate anomaly interpretation directed graph. The access order value, the earliest backtrackable order value, the search stack, and the push state are reinitialized. Then, the same depth-first search and strongly connected component identification are performed from the unvisited nodes. At least one candidate anomaly interpretation relationship that has formed a cycle is deleted in each round. The number of directed edges corresponding to the cycle relationship decreases with each round until there are no strongly connected components containing multiple anomaly functional units in the current round's identification results. At this point, no anomaly functional unit can return to itself along the interpretation direction through several directed edges. The remaining directed graph is determined to be an anomaly interpretation directed acyclic graph. The complete candidate anomaly interpretation relationships deleted in each round together form a conflict relationship set and enter the hidden state revocation and replay process of the improved MambaTS model.

[0113] After the Tarjan algorithm completes the identification of strongly connected components, the edge weight comparison is only performed between the candidate anomaly interpretation relations corresponding to the current strongly connected component. First, complete candidate anomaly interpretation relations that can jointly explain the same downstream functional unit are retained. Then, the candidate anomaly interpretation relation with the smallest edge weight is deleted from the candidate anomaly interpretation relations that form a cycle, and all directed edges corresponding to the relation are deleted simultaneously. Thus, the cycle resolution targets the mutual conflict between interpretation relations, rather than splitting a single directed edge within the same joint upstream interpretation. This allows the interpretation results of multiple upstream anomalies acting together on the same functional unit to be completely preserved. At the same time, relations with weaker explanatory power and participating in the cycle are identified as conflict relations, forming conflict objects that can be directly fed back to the hidden state inheritance position of the improved MambaTS model.

[0114] In this embodiment, S7 specifically includes:

[0115] The feedback of conflict relationships uses the state information synchronously saved when the aforementioned candidate anomaly interpretation relationship is formed as the basis for rollback. When each candidate anomaly interpretation relationship is formed, the set of upstream anomaly functional units participating in the interpretation, the downstream functional unit being interpreted, the scanning position of the downstream functional unit in the variable scanning sequence, the hidden state saved before entering the scanning position, and the position where the hidden state is finally written back at the scanning position are recorded synchronously. The final written-back state is the candidate hidden state in the remaining anomaly branch or the hidden state recursively generated by the remaining anomaly branch. After the Tarjan algorithm determines that the candidate anomaly interpretation relationship is a conflict relationship, it matches the corresponding state information of the upstream anomaly functional unit set and the interpreted functional unit in the conflict relationship, locates the hidden state write-back position of the downstream functional unit, cancels the state inheritance formed at the position, and reads the hidden state saved before the downstream functional unit is processed. When the conflict relationship contains multiple upstream anomaly functional units, the complete candidate anomaly interpretation relationship is rolled back once.

[0116] After the conflict relationship is revoked, the twin complex state and the corresponding healthy digital twin state saved when the candidate anomaly explanation relationship was formed are read first. The state change of the twin complex state relative to the healthy digital twin state is calculated at the same functional unit, the same state variable, and the same sampling time. This state change is then reassigned to the remaining anomaly currently retained by the corresponding downstream functional unit. If a remaining anomaly already exists at the same state variable and sampling time, it is accumulated according to the signed state change to obtain the updated remaining anomaly. The updated remaining anomaly serves as the anomaly source for re-entering the model. The healthy state at the same functional unit, state variable, and sampling time is read from the healthy digital twin. The updated signed remaining anomaly is added at the corresponding position to form the reconstructed running state. The instance normalization method, time block boundary, linear embedding parameters, and variable scanning position determined earlier are used to form the reconstructed state label. The recovered pre-processing hidden state is used as the recursive starting point of the current scanning position to perform selective state space recursion. The recursion parameters and prediction head parameters remain at the fixed values ​​after training. Thus, the hidden state and twin complex state are regenerated.

[0117] The scope of the replay starts from the downstream functional unit scan position where the state inheritance was revoked. The scan positions within the range of the anomaly to be located are reprocessed along the original variable scan order. Scan results that have been completed before the revoked position and have not been affected by the conflict relationship in this round are retained. When reprocessing each scan position, the twin replay comparison, residual state difference calculation and hidden state inheritance methods determined above are still used. The hidden state regenerated at the current scan position is written to the next scan position after the inheritance requirements are met. If there is a residual anomaly, a new anomaly propagation starting point is formed again from the current functional unit. This ensures that the state change generated by a rollback can be continuously propagated to the affected subsequent scan positions along the original recursive direction.

[0118] After the replay is completed, the directed edges and edge weights corresponding to the original candidate anomaly interpretation relationships within the reprocessing range are updated. If the original relationship has lost the corresponding replay result due to state rollback, the original edge record is deleted. The newly formed candidate anomaly interpretation relationships are re-established with directed edges pointing from the upstream anomaly functional unit to the explained functional unit. The operational deviation of the explained functional unit and the corresponding remaining anomaly are accumulated in the same dimensionless manner as described above. The difference between the two accumulated values ​​is used as the new edge weight. The directed edges and edge weights outside the reprocessing range that have not undergone state changes remain unchanged. After the graph update is completed, the Tarjan algorithm is executed again to identify strongly connected components containing multiple anomaly functional units. New conflict relationships are generated within the strongly connected components according to the edge weight comparison method determined above. The current round of anomaly interpretation directed acyclic graph and remaining anomalies are obtained again.

[0119] Revoked conflict relationships are kept in an unrecoverable state through the exclusion relationship set. The exclusion relationship set records the upstream abnormal functional unit set that has been identified as conflicting and the corresponding explained functional unit. Each time a new conflict relationship is generated, this set of corresponding information is written into the set without deleting existing records. After replaying to form candidate abnormal explanation relationships, the upstream abnormal functional unit set and explained functional unit in the candidate relationship are compared with the exclusion relationship set item by item. If they are consistent, the corresponding hidden state is not written back. Instead, the hidden state before the current functional unit is processed is read directly and the reconstructed state and recursive result are formed along the remaining abnormal path. Candidate explanation relationships that do not hit the exclusion relationship set continue to enter the normal hidden state inheritance judgment. In this way, the explanation path that has been excluded by Tarjan conflict resolution is continuously isolated from subsequent recursion.

[0120] The closed-loop termination is determined by the fact that no new conflict relationships are generated in the current round of Tarjan processing. Within a fixed range of anomalies to be located and a fixed sampling window, the number of anomaly functional units, variable scanning positions, and the number of candidate anomaly interpretation relationships that can be formed are all finite values. In each round, as long as a new conflict relationship is generated, at least one new relationship record is added to the exclusion relationship set. Relationships already written into the set will not be restored to their hidden state in subsequent rounds. Therefore, the exclusion interpretation paths are increased unidirectionally during the feedback process and will not be repeatedly restored between different rounds. When no new conflict relationships are found after the replay and graph update are completed, the current anomaly interpretation directed acyclic graph will no longer change due to conflict cancellation. The current graph is determined as a stable anomaly interpretation directed acyclic graph. At the same time, the remaining anomalies retained by each anomaly functional unit are determined as stable remaining anomalies, and the fault location determination process begins.

[0121] This feedback process changes the previously non-existent connection between the graph relation determination result and the inheritance of the hidden state within MambaTS. The conflicts identified by the Tarjan algorithm no longer remain at the level of deleting edges in the graph structure, but directly locate the hidden state within the model that has been written back when the conflict interpretation is formed and trigger state recovery. This allows the global loop interpretation to modify the local recursive source in reverse. At the same time, through the set of irreversible exclusion relations, a conflict determination is solidified into the inheritance boundary in the next round of forward propagation. This connects the model's internal time recursion and the global consistency correction of the anomaly interpretation graph into the same closed loop, so that the source of error propagation will not re-enter the hidden state in the replay. At the same time, the remaining anomaly is retained as a new local propagation starting point, so that the stable anomaly interpretation directed acyclic graph and the stable remaining anomaly have a consistent model state source.

[0122] In this embodiment, S8 specifically includes:

[0123] Stable residual anomalies are used to determine the fault location at the functional unit level of the industrial control motherboard. All abnormal functional units in the directed acyclic graph of stable anomaly interpretation are read, and the functional units that still retain stable residual anomalies after feedback convergence are identified as fault locations. When multiple functional units retain stable residual anomalies, they form corresponding fault locations, thereby obtaining the functional unit level location results corresponding to the power supply, clock, reset, processing, storage, or interface functional areas.

[0124] The fault location monitoring result is formed by the determined fault location and the corresponding change in operating status. For each fault location, the status data of the same functional unit in the operating status sequence within the sampling interval corresponding to the stable remaining abnormality are returned. The status variables and their operating values ​​that have undergone abnormal changes are extracted in the original sampling time order, so that the fault location can continuously correspond to the actual monitored abnormal state. Multiple fault locations maintain their own status change sequences and together form the fault location monitoring result of this industrial control motherboard. Thus, the final monitoring information that can clearly reflect the functional unit where the fault occurred and the corresponding abnormal state change is output.

[0125] Example 1:

[0126] To verify the feasibility of this invention in practice, it was applied to the overall testing of an industrial control motherboard in an industrial control equipment manufacturing company. The tested motherboard included functional units such as power management, clock, reset, processor, memory, communication interface, and input / output control. These functional units had clear dependencies on upstream power supply, clock synchronization, and control signals. In actual testing, it was found that after abnormal power ripple, clock skew, or reset signal occurred, the processor, memory, and communication interface often simultaneously exhibited state deviations. Traditional threshold detection easily misclassified multiple downstream abnormalities as independent faults. When different upstream abnormalities acted on the same downstream functional unit, the fault range was easily expanded, and repeated fault localization was common. Therefore, this scenario effectively demonstrated the problem that this invention aims to solve: the difficulty in distinguishing between propagating abnormalities and localized faults.

[0127] In practical applications, operational monitoring data such as voltage, current, clock frequency, reset level, bus response time, and interface status of each functional unit are first collected. A healthy operating benchmark is established using a healthy motherboard that has passed functional testing. The health status of each functional unit is associated with the motherboard signal transmission relationship to form a health digital twin. During operation, the real-time status is compared with the health digital twin to first determine the abnormal functional unit and the range of anomalies to be located. Then, the MambaTS model is improved by extracting the interrelated joint upstream anomaly state input from the anomaly range. The model saves the hidden state before processing the current functional unit. After completing the selective state space recursion, the candidate hidden state is temporarily stored. The current functional unit state is reconstructed through the health digital twin, and state changes that can be explained by upstream anomalies are deducted from the actual anomalies. If there are no obvious remaining anomalies, the candidate hidden state is inherited. If there are remaining anomalies, the hidden state before processing is restored. The current functional unit state is reconstructed with the health state and remaining anomalies, and the recursion is restarted, thereby separating the propagated anomalies from the new local anomalies. Subsequently, a directed graph is established based on the candidate anomaly interpretation relationships. The Tarjan algorithm is used to identify cyclic interpretation relationships, and conflicting relationships are deleted according to the interpretation edge weights. The conflict results are fed back to the improved MambaTS model for re-enactment, and finally the fault functional units and monitoring status corresponding to the stable remaining anomalies are output.

[0128] To verify the positioning effect, in one embodiment, 480 fault samples were selected for verification after manual re-inspection and board-level testing. These included 168 single-function unit faults, 196 upstream faults causing multi-level propagation anomalies, and 116 composite faults caused by two or more anomaly sources. The fault types covered power supply anomalies, clock anomalies, reset anomalies, memory access anomalies, and communication interface anomalies. The traditional threshold and rule positioning method, the method based solely on the state deviation of the health digital twin, the original MambaTS model method, and the method of this invention were tested. The results of manual re-inspection were used as the fault location determination benchmark. The fault location accuracy, propagation anomaly misjudgment rate, composite fault identification rate, and average positioning time were statistically analyzed. The comparison results are shown in Table 1 below.

[0129] Table 1 Comparison of Multi-level Fault Propagation and Location Effects of Industrial Control Motherboards

[0130] Comparison indicators Traditional threshold and rule-based localization methods Health Digital Twin State Deviation Method Original MambaTS model method Method of the present invention Fault location accuracy / % 81.5 87.3 90.6 96.0 False positive rate of abnormal propagation / % 18.8 12.1 9.4 3.8 Composite fault identification rate / % 72.4 79.3 84.5 93.1 Average positioning time / s 3.7 3.1 2.8 2.2

[0131] As can be seen from the data in Table 1 above, the method of the present invention exhibits good overall performance in four indicators: fault location accuracy, propagation anomaly misjudgment rate, composite fault identification rate, and average location time. The fault location accuracy of the traditional threshold and rule-based location method is 81.5%, and the propagation anomaly misjudgment rate reaches 18.8%, indicating that it mainly relies on local thresholds and fixed rules for judgment, making it difficult to distinguish between downstream anomalies caused by upstream fault propagation and functional unit faults themselves. The health digital twin state deviation method improves the fault location accuracy to 87.3% and reduces the propagation anomaly misjudgment rate to 12.1% by introducing a health state comparison. However, due to the lack of continuous explanation of the anomaly propagation process, the composite fault identification rate is still only 79.3%. After the original MambaTS model uses temporal state modeling, the fault location accuracy is further improved to 90.6%, and the composite fault identification rate reaches 84.5%. However, during the continuous recursion of hidden states, propagation information from different fault sources may still be mixed. The method of this invention achieves a fault location accuracy of 96.0%, reduces the propagation anomaly misjudgment rate to 3.8%, achieves a composite fault identification rate of 93.1%, and shortens the average location time to 2.2 seconds. Fault location accuracy represents the proportion of samples correctly identifying faulty functional units out of the total number of fault samples; propagation anomaly misjudgment rate represents the proportion of anomalies generated by upstream fault propagation that are incorrectly identified as independent faults out of the total number of propagation anomaly samples; composite fault identification rate represents the proportion of samples correctly identifying all faulty functional units when there are two or more fault sources; and average location time represents the time required from anomaly data input to outputting the fault location result. This demonstrates that the present invention can improve the location accuracy under complex fault conditions while maintaining high processing efficiency.

[0132] This embodiment establishes a health digital twin containing the health status of functional units and their state dependencies. It then compares the actual operating status of the industrial control motherboard with the status dependencies to narrow down the range of anomalies to be located. Based on this, it inputs the upstream anomaly status into an improved MambaTS model. Before the current functional unit is recursively processed, hidden states are saved and candidate hidden states are temporarily stored. The health digital twin is used to re-encode state changes in the current anomaly that can be explained by upstream anomalies. Any changes exceeding the health deviation limit are retained as residual anomalies. If residual anomalies exist, the hidden states before processing are restored to reconstruct the state and re-execute the recursion, thereby reducing the mixing of different fault sources during continuous state propagation. Furthermore, the Tarjan algorithm is used to identify strongly connected components in the directed graph of candidate anomaly interpretations. Conflicting relationships in cyclic interpretations are deleted based on edge weights, and the conflict results are fed back to the improved MambaTS model for re-encodement, stabilizing the candidate anomaly interpretation relationships after multiple rounds of feedback. Therefore, this embodiment can more accurately distinguish between propagation anomalies caused by upstream faults and local anomalies retained by the functional units themselves when multiple functional units such as power supply, clock, reset and communication experience abnormal responses simultaneously, thus narrowing the actual fault location range and providing a clearer location basis for online detection, maintenance and troubleshooting and fault tracing of industrial control motherboards.

[0133] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent fault location and monitoring of industrial control motherboards based on twin simulation, characterized in that, Includes the following steps: S1. Collect industrial control motherboard operation monitoring data and health operation benchmarks to form an operation status sequence and a health benchmark sequence; S2. Establish a health digital twin containing state dependencies based on the health baseline sequence and motherboard functional connections; S3. Compare the running state sequence with the corresponding health digital twin, and determine the range of anomalies to be located along the state dependency relationship; S4. Extract the combined upstream anomaly state input from the anomaly range to be located to improve the MambaTS model, retain the hidden state before the current functional unit is processed and temporarily store the candidate hidden state, determine the state inheritance based on the reenactment results of the health digital twin, and form the candidate anomaly interpretation relationship and the remaining anomaly. S5. Using the abnormal functional units within the range of the anomaly to be located as nodes, construct a directed graph of candidate anomaly interpretations according to the candidate anomaly interpretation relationships, and determine the edge weights based on the remaining anomalies. S6. Perform Tarjan's algorithm on the candidate anomaly interpretation directed graph to identify strongly connected components, determine the conflict relationships in the strongly connected components based on the edge weights, and form an anomaly interpretation directed acyclic graph and conflict relationships. S7. Feed the conflict relationship back to the improved MambaTS model, revoke the state inheritance corresponding to the conflict relationship and restore the hidden state before processing, re-enact the remaining anomaly, no longer inherit the hidden state corresponding to the revoked conflict relationship, and form a stable anomaly interpretation directed acyclic graph and stable remaining anomaly. S8. Based on the interpretation of the directed acyclic graph and the stable residual anomaly, determine the fault location of the industrial control motherboard and form the fault location monitoring results.

2. The intelligent fault location and monitoring method for industrial control motherboards based on twin simulation according to claim 1, characterized in that, S1 specifically includes: S11. Collect the operation monitoring data of each functional unit of the industrial control motherboard, and obtain the health operation benchmark when the corresponding functional unit is in a healthy operating state; S12. Time-align the operation monitoring data and the health operation benchmark according to the sampling time, and register the operation monitoring data of the same functional unit with the health operation benchmark to the same sampling time; S13. Arrange the time-aligned operation monitoring data and health operation benchmarks according to the sampling time sequence to form the operation status sequence and health benchmark sequence.

3. The intelligent fault location and monitoring method for industrial control motherboards based on twin simulation according to claim 1, characterized in that, S2 specifically includes: S21. Map the state data in the health baseline sequence to the digital twin state of the corresponding functional unit according to the functional unit to form the health twin state; S22. According to the signal transmission direction of the motherboard function connection, associate the output state of the upstream functional unit with the input state of the downstream functional unit to form a state dependency relationship. S23. Associate the health twin states of each functional unit according to the state dependency relationship to form a health digital twin.

4. The intelligent fault location and monitoring method for industrial control motherboards based on twin simulation according to claim 1, characterized in that, S3 specifically includes: S31. Calculate the absolute difference between the operating status sequence and the corresponding health status of the health digital twin according to the functional unit and sampling time to form the operating deviation sequence; S32. The maximum deviation between the health baseline sequence and the corresponding health status of the health digital twin is used as the health deviation limit. Functional units whose operating deviation exceeds the health deviation limit are identified as abnormal functional units. S33. Include abnormal functional units and the direct state transmission between abnormal functional units in the state dependency relationship into the scope of abnormalities to be located.

5. The intelligent fault location and monitoring method for industrial control motherboards based on twin simulation according to claim 1, characterized in that, The improved MambaTS model includes an instance normalization and embedding layer, a variable scanning layer, a temporal Mamba coding layer, and a prediction head. The instance normalization and embedding layer performs instance normalization on the corresponding functional unit states in the joint upstream abnormal state and the healthy digital twin, divides the time into blocks along the sampling time sequence, and forms state labels through linear mapping. The variable scanning layer arranges the status markers of each functional unit in a staggered manner according to time blocks, forming a variable scanning sequence along time; The temporal Mamba coding layer performs selective state-space recursion after linear mapping of the variable scan sequence and adds it to the gated nonlinear mapping result. It retains the hidden state before processing the current functional unit and temporarily stores the candidate hidden state after the current functional unit completes the state update. The prediction head maps the encoding results corresponding to the candidate hidden states to the twin replay state of the current functional unit in the health digital twin independently according to the functional unit; compares the twin replay state with the current functional unit's operating state, determines the inheritance of hidden states based on the comparison differences, and forms the candidate anomaly interpretation relationship and the remaining anomalies.

6. The intelligent fault location and monitoring method for industrial control motherboards based on twin simulation according to claim 5, characterized in that, The hidden state inheritance specifically includes: When processing along the variable scanning sequence to the current functional unit, the hidden state before entering the current functional unit is saved. The current functional unit's state flag is recursively obtained through selective state space to obtain the candidate hidden state. The candidate hidden state is temporarily stored and not written back as the input hidden state of the next functional unit. The prediction head generates a twin complex state based on the candidate hidden state. The twin complex state is compared with the current functional unit's operating state at the same sampling time. The maximum deviation between the health baseline sequence and the corresponding state of the health digital twin is used as the health deviation limit. The change in the twin's recreated state relative to the corresponding state of the healthy digital twin is subtracted from the change in the current functional unit's operating state relative to the corresponding state of the healthy digital twin to obtain the remaining state difference. State changes whose amplitude exceeds the health deviation limit in the remaining state difference are identified as remaining anomalies. When there are no state changes in the remaining state differences whose amplitude exceeds the health deviation limit, the candidate hidden state is written back as the input hidden state of the next functional unit, and the candidate anomaly interpretation relationship is formed based on the corresponding changes of the joint upstream abnormal state and the actual anomaly of the current functional unit. When there are state changes in the remaining state differences whose amplitude exceeds the health deviation limit, the candidate hidden state is not written back. Instead, the saved hidden state before processing is read, and the corresponding state of the health digital twin is superimposed with the remaining anomaly to form the reconstruction state of the current functional unit. The reconstruction state is then formed by instance normalization and embedding. The selective state space recursion of the current functional unit is re-executed using the saved pre-processing hidden state and reconstructed state markers. The regenerated hidden state is written back as the input hidden state of the next functional unit. Candidate anomaly interpretation relationships are formed based on the change in the twin replay state relative to the corresponding state of the healthy digital twin, and the remaining anomalies are retained.

7. The intelligent fault location and monitoring method for industrial control motherboards based on twin simulation according to claim 1, characterized in that, S5 specifically includes: S51. Using the abnormal functional units within the range of the anomaly to be located as nodes, point the upstream abnormal functional units in the same candidate anomaly interpretation relationship to the anomaly functional units to be interpreted, forming a directed graph of candidate anomaly interpretation. S52. Accumulate the deviation amplitude of the operating state of the explained functional unit relative to the corresponding state of the healthy digital twin and the corresponding residual abnormal amplitude, and use the difference between the two as the edge weight of each directed edge corresponding to the same candidate abnormality explanation relationship.

8. The intelligent fault location and monitoring method for industrial control motherboards based on twin simulation according to claim 1, characterized in that, S6 specifically includes: S61. Perform Tarjan's algorithm on the directed graph for candidate anomaly interpretation, and perform a depth-first search according to the direction of the directed edges to identify strongly connected components with mutually reachable nodes. S62. In a strongly connected component containing multiple abnormal functional units, compare the edge weights of the directed edges corresponding to the candidate abnormal interpretation relations, determine the candidate abnormal interpretation relations with the smallest edge weight as the conflict relation, and delete all directed edges of the corresponding candidate abnormal interpretation relations. S63. Repeatedly identify strongly connected components and delete directed edges corresponding to conflict relationships until the candidate anomaly interpretation directed graph does not contain strongly connected components containing multiple anomaly functional units, forming an anomaly interpretation directed acyclic graph and conflict relationships.

9. The intelligent fault location and monitoring method for industrial control motherboards based on twin simulation according to claim 1, characterized in that, Specifically, S7 includes: S71. Locate the downstream functional unit being interpreted based on the conflict relationship, cancel the corresponding hidden state write-back, and restore the hidden state saved before the downstream functional unit's processing. S72. Overlay the corresponding state of the health digital twin with the remaining anomalies, reconstruct the operating state of the downstream functional units, and re-execute the twin reenactment from the restored hidden state. S73. Based on the candidate anomaly interpretation relationship formed by the re-replay and the remaining anomalies, update the directed graph of candidate anomaly interpretations and edge weights, and execute the Tarjan algorithm again to update the directed acyclic graph of anomaly interpretations and conflict relationships. S74. Repeatedly revoke and reenact newly generated conflict relationships, and do not write back the hidden states corresponding to the revoked conflict relationships, until no new conflict relationships are generated, forming a stable anomaly interpretation directed acyclic graph and a stable residual anomaly.

10. The intelligent fault location and monitoring method for industrial control motherboards based on twin simulation according to claim 1, characterized in that, S8 specifically includes: S81. Locate the abnormal function unit that retains the stable remaining abnormal in the directed acyclic graph of stable anomaly interpretation, and determine the corresponding abnormal function unit as the fault location of the industrial control motherboard. S82. Extract the state changes of the corresponding functional units in the operating state sequence according to the fault location to form the fault location monitoring results.

Images