An industrial equipment fault self-recovery closed-loop disposal method based on monitoring control

Abstract

The application provides an industrial equipment fault self-healing closed-loop treatment method based on monitoring control, and relates to the technical field of distributed program control.The application constructs and compares a physical interaction topology and a logic control topology, quantifies the "topology divergence" between the two as a core health index, and performs networked root cause tracing and utility-optimal predictive rebalancing based on the index.The application can perceive and warn in advance the implicit systemic risks caused by deep physical and logic mismatches, which are difficult to be found by traditional methods;through topology clustering analysis, the accuracy of locating the problem root in complex linkage faults is improved;based on the closed-loop decision mechanism of utility prediction, the safety and economy of the self-healing behavior are ensured, the system stability is restored, and the core production performance is maximized, so that the autonomy, robustness and operation efficiency of the complex industrial system in a dynamic environment are enhanced.

Description

Technical Field

[0001] This invention relates to the field of distributed program control technology, specifically to a closed-loop fault self-healing method for industrial equipment based on monitoring and control. Background Technology

[0002] In modern manufacturing systems, distributed digital control systems, composed of numerous automated units, are the core of achieving efficient and flexible production. As system scale and complexity increase, ensuring the long-term stability and high reliability of collaborative operations among its internal nodes has become a key technological focus in this field. The health of a system depends not only on the physical integrity of individual devices, but more profoundly on the high degree of coordination and matching between digital control commands and the dynamic responses of the physical world.

[0003] However, existing technologies still face several challenges in ensuring the health of such complex systems. Specifically: Existing technologies monitor physical states (vibration, temperature) and logical states (network logs, command flow) as independent data sources. Patent application CN120065949A discloses a system for predicting faults using multi-dimensional data and LSTM algorithms. While this system is valuable in predicting known fault modes, it still relies on the analysis of observable parameters and fails to provide a unified indicator that directly measures the degree of real-time structural tension or "mismatch" between the system's physical dynamic topology and logical control topology. Therefore, it is difficult to perceive deep-seated, cumulative system architecture risks before manifest faults occur.

[0004] When anomalies occur, traditional diagnostic methods often focus their analysis on the node where the signal anomaly is most pronounced. In highly coupled distributed systems, this point is often merely a "symptom" at the end of the fault propagation chain, rather than the "lesion" itself. This diagnostic approach, limited to superficial symptoms, makes subsequent interventions ineffective in addressing the root cause of the problem and easily leads to fault recurrence.

[0005] Self-healing interventions are mostly open-loop responses, lacking the ability to anticipate the consequences of decisions. Existing automated fault response mechanisms execute adjustments based on preset static rules, such as adjusting operating parameters or activating backup equipment. Such "open-loop" interventions cannot assess their potential cascading effects on key performance indicators such as overall system productivity and operating costs before execution, posing a risk of triggering more serious secondary problems due to improper handling. Summary of the Invention

[0006] The purpose of this invention is to provide a self-healing closed-loop fault handling method for industrial equipment based on monitoring and control, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A closed-loop fault self-healing method for industrial equipment based on monitoring and control, comprising the following steps: S1: Obtain the dynamic coupling matrix, which is configured to quantify the physical interaction strength between pairs of nodes in a distributed system. The physical interaction strength is calculated based on the cross-correlation of the vibration signals of the pairs of nodes. S2: Obtain the control influence matrix, which is configured to quantify the degree of logical dependency between each pair of nodes in the plurality of nodes, and the degree of logical dependency is calculated based on the information entropy of the control command flow between each pair of nodes; S3: Perform topological divergence calculation on the dynamic coupling matrix and the control influence matrix to generate a physical-logical mismatch signal; wherein, the physical-logical mismatch signal is configured to characterize the degree of mismatch between the physical interaction topology and the logical control topology of the distributed system; S4: Based on the physical logic mismatch signal, locate the mismatched node pair in the distributed system, where the mismatched node pair is the primary contributor to the physical logic mismatch signal exceeding the preset mismatch threshold; S5: In response to locating the mismatched node pair, generate a rebalancing instruction for adjusting the logical control interaction between the mismatched node pair.

[0008] Compared with existing technologies, the beneficial effects of this invention are as follows: By using signal processing and information theory methods, the invisible physical coupling and logical dependencies are constructed into computable "dynamic coupling matrices" and "control influence matrices," respectively. Subsequently, topological divergence calculation is introduced to mathematically dual these two heterogeneous matrices, thereby generating a "physical-logical mismatch signal" that can directly quantify the global mismatch degree of the system, thus solving the problem of the inability to uniformly quantify systemic risks. Furthermore, this invention analyzes the contribution distribution of this mismatch signal as a "mismatch network graph" and uses community detection and centrality analysis algorithms to conduct in-depth topological analysis within this network, thereby accurately locating the "root cause node pairs" that trigger cascading mismatches, achieving a paradigm shift in diagnosis from "symptoms" to "lesions." In the handling phase, this invention abandons static rules and introduces a predictive decision-making mechanism based on a "system performance proxy model." Before executing any rebalancing operation, this mechanism predicts multiple candidate rebalancing instructions, calculates their comprehensive "utility score" for the future stability and performance of the system, and selects the best one for execution. This constructs a "closed-loop" self-healing logic that balances safety and efficiency, effectively avoiding the risks brought about by blind intervention. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the overall method flow of the present invention; Figure 2 This is a schematic diagram illustrating the execution logic of steps S1 to S3 of the present invention; Figure 3 This is a schematic diagram illustrating the execution logic of steps S4 to S5 of the present invention. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0011] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various elements, but unless otherwise stated, these elements are not limited by these terms. These terms are used only to distinguish one element from another.

[0012] Example 1: Please see Figures 1 to 3 The present invention provides a technical solution: A self-healing closed-loop fault handling method for industrial equipment based on monitoring and control includes the following steps: S1: Obtain the dynamic coupling matrix, which is configured to quantify the physical interaction strength between pairs of nodes in a distributed system. The physical interaction strength is calculated based on the cross-correlation of the vibration signals of each pair of nodes. S2: Obtain the control influence matrix. The control influence matrix is ​​configured to quantify the degree of logical dependence between each pair of nodes in multiple nodes. The degree of logical dependence is calculated based on the information entropy of the control command flow between each pair of nodes. S3: Perform topological divergence calculation on the dynamic coupling matrix and control influence matrix to generate a physical-logical mismatch signal; wherein, the physical-logical mismatch signal is configured to characterize the degree of mismatch between the physical interaction topology and the logical control topology of the distributed system; S4: Based on the physical logic mismatch signal, locate the mismatched node pair in the distributed system. The mismatched node pair is the primary contributor to the physical logic mismatch signal exceeding the preset mismatch threshold. S5: In response to locating the mismatched node pair, generate a rebalancing instruction to adjust the logical control interaction between the mismatched node pair.

[0013] It should be noted that: Figure 1The simulation example of a collaborative work unit consisting of multiple automated robotic arms and a shared conveyor belt, displayed in the upper-middle "physical domain," represents the physical entity of the distributed system applied in this invention. The two data streams flowing downwards from this entity—"physical vibration signals" and "logic control commands"—visually reveal the data sources for subsequent analysis. Figure 1 The lower-middle "analysis and control domain" receives the data flow from the physical domain and reveals the closed-loop self-healing process of this invention, which corresponds to step S1: Figure 1 In the internal physical-logical dual-domain topology modeling, the data streams of "physical vibration signals" and "logical control commands" converge to form visual representations of the "dynamic coupling matrix C" and the "control influence matrix H," respectively. This corresponds to steps S1 and S2, where the system acquires a mathematical model that quantifies the intensity of physical interactions and the degree of logical dependence, thus completing the structured data abstraction of the system state. Figure 1 Internal topological divergence calculation and mismatch quantification correspond to S3: Two matrices are compared using an abstract mathematical operator (Δ) to generate a "physical-logical mismatch signal" spectrum with obvious abnormal peaks. This corresponds to step S3, which generates a core indicator that can accurately quantify the degree of mismatch between the physical and logical states of the system through topological divergence calculation. Networked root cause diagnosis and localization correspond to S4: The mismatch signal points to the subsequent flowchart "Networked root cause diagnosis and localization," and the network diagram next to it highlights the core nodes. This corresponds to step S4, which performs in-depth network topology analysis based on the mismatch signal to accurately identify and locate the root cause node pairs that cause system mismatch. Predictive closed-loop self-healing decision-making corresponds to S5: The endpoint of the technology roadmap is "predictive closed-loop self-healing decision-making," from which a clear "rebalancing instruction" arrow points back and acts on the central control unit of the upper physical domain. This corresponds to step S5, where the system generates the optimal adjustment instruction and issues it to the physical devices, ultimately forming a complete and self-consistent technical closed loop from monitoring, analysis, decision-making to execution.

[0014] Further explanation: In S1, the physical interaction strength is calculated based on the cross-correlation of vibration signals between pairs of nodes, including: The normalized cross-correlation spectrum of the vibration signals of each pair of nodes within a preset characteristic frequency band is calculated, and the peak value of the normalized cross-correlation spectrum is taken as the physical interaction intensity.

[0015] Further explanation: Calculating the normalized cross-correlation spectrum and obtaining the peak value specifically includes: performing time-frequency analysis on the vibration signal using continuous wavelet transform to generate time-frequency coefficients; based on the time-frequency coefficients, simultaneously calculating the cross-correlation coefficient characterizing the energy correlation of the signal and the phase lock value characterizing the phase synchronization of the signal; and combining the cross-correlation coefficient and the phase lock value to generate a multidimensional dynamic coupling descriptor characterizing the physical interaction strength.

[0016] Further explanation: In S2, the logical dependency is calculated based on the information entropy of the control command flow between each pair of nodes. This includes: calculating the conditional information entropy of the control command flow from the first node to the second node relative to all control command flows issued by the first node, and using the reciprocal of the conditional information entropy as the logical dependency.

[0017] Further explanation: The calculation of conditional information entropy specifically includes: identifying predefined timing instruction modules composed of multiple instructions from the control instruction flow; calculating the conditional information entropy of instruction type transfers within the timing instruction modules, as well as the timing information entropy of the time interval between occurrences of timing instruction modules; and weighting and combining the conditional information entropy and timing information entropy of instruction type transfers to generate the degree of logical dependency.

[0018] Further explanation: In S3, before calculating the topological divergence of the dynamic coupling matrix and the control influence matrix, the following steps are also included: The dynamic coupling matrix and the control influence matrix are respectively subjected to topological rank sorting and normalization to eliminate the differences in data scale and units between the two matrices; Furthermore, the topological divergence calculation specifically involves using Jensen-Shannon divergence to calculate the difference between the two normalized matrices.

[0019] Further explanation: The Jensen-Shannon divergence (hereinafter referred to as JS divergence value) is used to calculate the difference, specifically including: continuously calculating a series of instantaneous divergence values ​​within a preset time sliding window; obtaining the context state signal characterizing the current global operating condition of the distributed system; adaptively adjusting the weight coefficients of the exponentially weighted moving average filter based on the context state signal; and smoothing the series of instantaneous divergence values ​​using the exponentially weighted moving average filter with adaptive weight coefficients to generate a time-integrated physical-logical mismatch signal characterizing the long-term mismatch trend. The following are specific implementation details: Before the computation process begins in this embodiment, a series of preset configuration parameters are loaded from an externally stored, structured spreadsheet file to achieve technical decoupling between the algorithm logic and the specific application strategy. The key parameters involved in this embodiment are defined as follows: 1.1) The relevant parameters for calculating the physical interaction strength are as follows: The mother wavelet function type, denoted as MWT, logically defines the basis functions of the continuous wavelet transform. It is determined by pre-selecting based on the characteristics of the vibration signal to be analyzed. In this embodiment, the "Morlet" wavelet is chosen, as it has good resolution in both time and frequency, making it suitable for analyzing mechanical vibration signals. The selection of the "Morlet" wavelet as the mother wavelet function type MWT is based on a deep understanding of the physical characteristics of vibration signals from distributed industrial equipment. Choosing the "Morlet" wavelet ensures that the physical interaction intensity calculated in this invention can be directly traced back to mechanical motion at a specific frequency, rather than broad signal fluctuations without clear physical correspondence, thereby improving the accuracy of physical coupling topology inference.

[0020] The coupling strength combination weight, denoted as CCW, logically represents the weight assigned to the cross-correlation coefficient and phase lock value when combining them. It is a value between 0 and 1. It is determined through offline experimental calibration, reflecting the relative importance of energy correlation and phase synchronization in the final evaluation. The determination of the parameter coupling strength combination weight CCW is accomplished through the following simulation offline calibration experiment. Specifically, a physical model is established in which the physical coupling strength between two nodes can be accurately and independently measured as the "ground truth value," while its vibration signal can be collected. By systematically changing the coupling strength, an optimal CCW value is sought to maximize the correlation between the "multidimensional dynamic coupling descriptor" calculated in this invention and this "ground truth value." A simulation experimental platform is built consisting of two independent servo motors with identical parameters (as nodes i and j). The two motors are mounted on a shared base made of high-strength alloy material. Crucially, the physical connection between the two motors is achieved through a replaceable "coupling beam" with different stiffness coefficients. At least five coupling beams with different stiffness levels were prepared for the experiment, and their stiffness coefficients had been precisely calibrated in advance through standard materials mechanics tests. High-precision piezoelectric accelerometers were mounted on the housing of each motor along three orthogonal axes. These accelerometers, serving as functional "sensing modules," had the following key performance attributes: a sampling frequency of no less than 51.2 kHz to cover a sufficiently wide analysis bandwidth; and an amplitude linearity error of no more than 0.1% to ensure measurement accuracy. Simultaneously, high-precision torque sensors were mounted on the coupling beams to measure the dynamic torque transmitted from node i to node j in real time. In this embodiment, the objects of nodes i and j are motors, and this measurement is considered the "ground truth" of the coupling strength. The specific implementation process is as follows: The coupling beam with the lowest stiffness level is installed. The motor corresponding to node i is started and allowed to operate stably at a preset series of speeds; in this embodiment, the series of speeds ranges from 10% to 100% of the rated speed, with each 10% increment representing a step. The motor corresponding to node j remains stationary or is subjected to a constant braking load. At each speed increment, simultaneously record at least 60 seconds of vibration timing data from all accelerometers and dynamic torque data from the torque sensor. Repeat the above steps, replacing all coupling beams with different stiffness levels in sequence.

[0021] Data collected under all experimental conditions (different stiffness, different rotational speeds) is iterated. For each condition, the method in step S1 of this invention is applied, but CCW is treated as an undetermined variable. That is, the cross-correlation coefficient sequence and phase-locked value sequence under that condition are calculated. A search space for candidate CCW values ​​is set from 0.01 to 0.99, with a step size of 0.01. For each candidate CCW value, the obtained cross-correlation coefficient and phase-locked value sequences are weighted and combined according to the value to obtain a "candidate integrated coupling strength" sequence. The Pearson correlation coefficient between this "candidate integrated coupling strength" sequence and the "ground true value" synchronously recorded under that condition is calculated. The steps corresponding to "candidate integrated coupling strength sequence and Pearson correlation coefficient" are repeated for all candidate CCW values. The average value of the Pearson correlation coefficients calculated for all experimental conditions is calculated.

[0022] The candidate CCW value that maximizes the average Pearson correlation coefficient under all operating conditions is selected as the final coupling strength combination weight embedded in the configuration of this system. In this embodiment, the CCW value determined through the above experiments is 0.65. The preferred range for the coupling strength combination weight CCW is between 0.6 and 0.75. The technical consideration for this range is that when the CCW value is below 0.6, the model will over-rely on the "phase-locked value," leading to misjudgments of physical coupling relationships with significant energy transfer but normal phase jitter due to flexible connections, thus reducing the recall rate of physical topology identification. Conversely, when the CCW value is above 0.75, the model will over-rely on the "cross-correlation coefficient," leading to excessive sensitivity to spurious correlation energy signals generated by resonance rather than real physical connections, thus increasing the false alarm rate of physical topology identification.

[0023] 2.0) The relevant parameters for calculating logical dependency are as follows: The timing instruction modifier library, denoted as TCML, is a predefined structured list where each item defines an ordered sequence of two or more control instructions with specific business logic meaning. A "Material Grab" modifier is defined as ["Request Grab Permission", "Confirm Permission", "Execute Grab", "Report Completion"]. It is determined by the process engineer according to the system's standard operating procedures (SOPs).

[0024] Entropy combination weight, denoted as ECW, logically means the weight assigned to the instruction type transfer entropy and timing entropy when combining them. It is a value between 0 and 1. By analyzing historical data, the relative contributions of instruction content and instruction timing in representing logical dependencies are evaluated.

[0025] The determination of the combined entropy weights (ECW) is based on a "logic efficiency regression model," which quantifies the relative contribution of instruction content and timing to system performance by analyzing historical operational data. For an efficient distributed control system, the quality of its logical dependencies is ultimately reflected in quantifiable system performance indicators. Therefore, a regression model is established with "instruction type transfer entropy" and "timing information entropy" as independent variables and system key performance indicators (KPIs) as dependent variables. The weights of each entropy value are determined by analyzing its explanatory power for the KPIs. A large-scale dataset is extracted from the system's historical operational logs. This dataset contains at least several thousand complete job cycles, marked as successful or failed. For each job cycle, complete control instruction flow data is extracted, and its corresponding system performance indicator, "total job time" or "energy efficiency," is recorded. This performance indicator serves as the dependent variable in the model. The specific calculation logic is as follows: For each job cycle in the dataset, the mean values ​​of "instruction type transfer entropy" and "timing information entropy" within that cycle are calculated using the method in step S2 of this invention. These two entropy values ​​serve as the two independent variables of the model. A multiple linear regression model is constructed. Specifically, a set of coefficients is sought that minimizes the mean square error between the performance index predicted by the linear combination of the two entropy values ​​and the actual performance index. The model is solved using the standard least squares method. The output of this step is the two original regression coefficients corresponding to "instruction type transfer entropy" and "timing information entropy," respectively. The absolute values ​​of these two original regression coefficients are obtained. The absolute value of the coefficient corresponding to "instruction type transfer entropy" is divided by the sum of the absolute values ​​of the two coefficients. This result is the final determined entropy combination weight ECW. In this embodiment, the ECW value is determined to be 0.4 by analyzing historical data containing job cycles.

[0026] Furthermore, the entropy combination weight ECW is preferably between 0.35 and 0.5. The core function of ECW is to balance the contributions of "instruction content determinism" (represented by the reciprocal of type transfer entropy) and "instruction rhythm regularity" (represented by the reciprocal of timing entropy) in assessing the degree of logical dependency. When the ECW value is below 0.35, the model will overemphasize the regularity of instruction timing, misjudging some simple but rhythmically fixed heartbeat or polling communications as strong logical dependencies. Conversely, when the ECW value is above 0.5, the model will overemphasize the determinism of instruction content, ignoring key strong logical dependencies existing in asynchronous and event-driven patterns. These instructions have high determinism but appear at irregular times.

[0027] 3.0) The specific parameters related to the calculation of physical-logical mismatch signals include: The divergence calculation time window, denoted as DCW, logically defines the time length for calculating a series of instantaneous divergence values. Specifically, it is set based on the timescale of system state changes, ensuring that the window can smooth out short-term noise while responding promptly to real changes.

[0028] The Context State Mapping Table (CSMT) is a lookup table that defines the mapping relationship between external system operating conditions ("Start-up and Debugging", "Stable Production", "Emergency Shutdown") and the weight coefficients of the exponentially weighted moving average filter. Based on expert experience and risk assessment, appropriate system sensitivities are assigned to different operating conditions. In this embodiment, the weight coefficient values ​​in the CSMT are calculated using the "Operating Risk and Tolerance" model, specifically: The weighting coefficients of the EWMA filter are defined as the system's "mismatch sensitivity." This sensitivity should be proportional to the "risk" of the current operating condition. Higher-risk conditions require a more sensitive system with larger weighting coefficients; lower-risk conditions require a higher system tolerance with smaller weighting coefficients. For each predefined system operating condition, domain experts evaluate two risk factors and quantify them into values ​​between 0 and 1. Expected volatility (EV): Represents the probability of normal, benign fluctuations in the physical-logic mismatch signal under this operating condition. The expected volatility (EV) for the "start-up and commissioning" phase is set to a high 0.9, while the expected volatility (EV) for the "stable production" phase is set to a low 0.1.

[0029] Failure Impact Level (FIS): Characterizes the severity of the consequences if a failure occurs due to mismatch under this operating condition. The expected volatility FIS is set to 1.0 (high) during the "stable production" phase and 0.2 (low) during the "standby" phase.

[0030] The calculation logic for the weighting coefficient is as follows: Subtract the expected volatility (EV) from 1 to obtain the "stability factor". Then, multiply this "stability factor" by the "failure impact level (FIS)". This product is the EWMA weighting coefficient for this operating condition.

[0031] For the "stable production" condition: weight = (1-0.1)×1.0 = 0.9. The system is highly sensitive.

[0032] For the "Startup and Debugging" scenario: Weight = (1-0.9)×0.5 = 0.05. The system has high tolerance.

[0033] These calculated weight values ​​are pre-filled into the Context State Mapping Table (CSMT), which is then directly queried and used during system runtime.

[0034] The EWMA weighting coefficients calculated by the "Operational Risk and Tolerance" model are reasonably limited to a range of 0.05 to 0.95. The technical consideration behind this range is that the weighting coefficients directly determine the smoothness and response speed of the mismatch signal. If the weight is below 0.05, the signal's "memory effect" is too strong, and the system will react sluggishly to newly occurring, real mismatch state changes, potentially missing the optimal intervention opportunity. If the weight is above 0.95, the signal has almost no smoothing effect, becoming indistinguishable from an instantaneous value, reintroducing the problem of excessive sensitivity to short-term noise, thus violating the original intention of using time-delay integration. Therefore, this range ensures that the mismatch signal can effectively filter out high-frequency noise while also responding promptly to continuous state changes.

[0035] 4.1) Optimal implementation of step S1, based on the quantization of physical interaction intensity using the dynamic coherence evolution spectrum: Traditional methods based on STFT and cross-correlation peak values ​​can only provide static, averaged assessments of coupling strength. However, in actual operating conditions, the physical coupling relationship between devices evolves dynamically due to factors such as load changes and structural thermal expansion and contraction. For example, the vibration transmission characteristics of a conveyor belt are completely different under heavy and light loads. Static assessments cannot capture this dynamic nature and may lead to misjudgments at critical moments.

[0036] This invention reveals that a complete physical interaction should include information in two dimensions: energy and phase. The cross-correlation coefficient, representing energy correlation, reflects the "strength" of the coupling, while the phase lock value, representing phase synchronization, reflects the "stability" and "type" of the coupling. Rigid connections exhibit high phase synchronization, while flexible connections show hysteresis and jitter. The core idea of ​​this invention is to construct a more comprehensive description of the coupling state by synchronously monitoring these two dimensions. The specific implementation is as follows: The time-frequency coefficients are obtained. For any two nodes (node ​​i and node j) with synchronously acquired vibration time-series signals, continuous wavelet transform is used for processing. The mother wavelet function type (MWT) of this transform is set to "Morlet" wavelet. The output of this step is a time-frequency coefficient matrix for the two nodes, where each element represents the signal energy and phase at a specific time and frequency point. To ensure that the continuous wavelet transform can accurately focus on the signal components most relevant to the physical interaction of the equipment, the frequency range of its analysis needs to be limited. This is achieved by setting a set of optimal scaling parameters. In this embodiment, the method for determining this set of "scaling parameters" is as follows: a "target characteristic frequency band" related to the key operating frequencies of the equipment (including the motor fundamental frequency and its harmonics) is read from an external configuration file, which is preset by process experts and is defined by the lowest and highest frequencies. The calculation logic for converting the frequency into a scale is executed. The core step of this logic is as follows: For each discrete frequency point within the "target characteristic frequency band," the corresponding scale value that best matches that frequency point is calculated based on the known mathematical relationship between the center frequency of the selected "Morlet" mother wavelet and the sampling rate. The calculation steps are: obtain the signal sampling frequency, the center frequency of the "Morlet" wavelet, and the upper and lower limits of the "target characteristic frequency band." Generate a target frequency sequence that samples at equal intervals between the upper and lower frequency limits. Iterate through each target frequency value in the target frequency sequence, performing a division operation: multiply the center frequency of the mother wavelet by the sampling frequency, and then divide the product by the current target frequency value. The quotient is the scale parameter corresponding to that target frequency. By executing the above steps, a set of scale parameters is generated, and the frequency range corresponding to this set can perfectly cover the preset "target characteristic frequency band." In one embodiment, the signal sampling frequency is 5120Hz, the center frequency of the "Morlet" wavelet is 0.8125Hz, and the preset "target characteristic frequency band" is 100Hz to 500Hz. Based on this, a set of scaling parameters will be calculated to ensure that the analytical energy of wavelet transform is concentrated in the frequency band from 100Hz to 500Hz.

[0037] Within a sliding time window, the Pearson correlation coefficient, which characterizes the instantaneous cross-correlation coefficient, is calculated for the energy components of the two time-frequency coefficient matrices within a preset characteristic frequency band. This result characterizes the degree of synchronous change in the vibration energy of the two nodes within this time window.

[0038] Within the same time window, the phase information of the two time-frequency coefficient matrices in the characteristic frequency band is extracted. The instantaneous phase difference between the two phase sequences is calculated, and the concentration of this phase difference distribution within one period is statistically analyzed. The concentration is the phase lock value, with a value range of [0,1]. The closer the phase lock value is to 1, the stronger the phase synchronization.

[0039] Obtain the calculated cross-correlation coefficient and the calculated phase-lock value. Obtain the preset coupling strength combination weight (CCW). Multiply the cross-correlation coefficient by the coupling strength combination weight (CCW), and simultaneously multiply the phase-lock value by (1 minus the coupling strength combination weight (CCW)). Then add the two products to obtain the comprehensive coupling strength scalar. Simultaneously, calculate the variance of the phase-lock value over a past period, using it as the coupling stability factor. Finally, combine the comprehensive strength scalar and the coupling stability factor into a two-dimensional vector, outputting it as a "multidimensional dynamic coupling descriptor" between nodes i and j. The magnitude of this multidimensional dynamic coupling descriptor is used as the element value to fill the dynamic coupling matrix.

[0040] 5.1) Optimal implementation of step S2, based on the quantification of logical dependencies in a time-weighted directed information flow paradigm: Traditional conditional information entropy methods only focus on the type distribution of instructions, ignoring the temporal relationships between instructions. A tightly coupled instruction pair like "request-response," and two independent broadcast instructions, appear to have similar entropy values ​​in traditional methods, but the reality of this logical dependency is that time information is not considered. This invention reveals that the logical dependency between nodes is not only reflected in "what instruction is issued," but also in "when it is issued" and "the sequential relationship between instructions." Its core idea stems from viewing discrete instruction sequences as a "language," whose "grammar" (instruction order) and "rhythm" (time interval) also carry rich information. The specific implementation is as follows: The instruction sequence within a sliding window is matched against a pre-defined timing instruction modifier library (TCML). If an instruction subsequence completely matches a modifier in the library, the subsequence is marked as a complete "modifier event"; this allows for real-time monitoring of the control instruction flow between nodes.

[0041] The instruction type transfer entropy is calculated, specifically for the identified "motif event," calculating the conditional information entropy of the type transfer from one instruction to the next within that event. This entropy value reflects the determinism of the instruction flow within a specific business logic.

[0042] Calculate the timing information entropy, specifically the time interval between two consecutive "motif events" sent from the same source node to the same target node. Discretize these time intervals into bins and calculate the Shannon entropy of their probability distribution. This entropy value reflects the rhythm and predictability of business logic occurrence. Obtain the calculated instruction type transfer entropy and timing information entropy. Obtain the preset entropy combination weight (ECW). Take the reciprocals of both entropy values, then sum them according to the ECW to obtain the final logical dependency degree, which is used to fill the control influence matrix.

[0043] 6.1) A preferred implementation of step S3, based on adaptive time-delay integral and context-gated mismatch signal calculation: Directly calculating the instantaneous JS divergence value, while reflecting the current mismatch state, is overly sensitive to instantaneous disturbances in the system (including network jitter and brief mechanical shocks), easily generating unnecessary false alarms. Furthermore, it cannot distinguish between "benign, temporary mismatches" and "malignant, continuously worsening mismatches" during the system startup phase. This invention reveals that systemic architectural risks are a slow, cumulative process, not an instantaneous event. Therefore, the evaluation of mismatch signals should have a "memory effect." Its core idea stems from the aim of filtering out short-term fluctuations and revealing long-term trends. The tolerance for "mismatch" should differ under different operating conditions, thus introducing the concept of "context adaptation."

[0044] Calculate the instantaneous divergence sequence. For two matrices normalized by topological rank sorting, calculate continuously within a preset divergence calculation time window (DCW) with a small step size to obtain a series of instantaneous JS divergence values. Obtain the context state and adjust the weights. Obtain a "context state signal" representing the current global operating condition from the upper-level MES or control system. The system queries the preset context state mapping table (CSMT) to find the weight coefficient value of the exponentially weighted moving average (EWMA) filter corresponding to the operating condition. This includes: a larger weight coefficient of 0.8 for the "stable production" mode, indicating that the system is highly sensitive to the current state; and a smaller weight coefficient of 0.2 for the "start-up and debugging" mode, tolerating greater fluctuations. Input the obtained series of instantaneous divergence values ​​into the EWMA filter. The weight coefficient of this filter is "dynamically determined by the exponentially weighted moving average (EWMA) filter based on the context state mapping table (CSMT)". The calculation logic of the filter is: the current output value equals the current input value multiplied by the weight coefficient, plus the output value of the previous time step multiplied by (1 minus the weight coefficient). The final output value of the EWMA filter is the "time-integrated physical-logic mismatch signal," which is characterized as the system mismatch index SDI. This signal is smooth and reflects long-term trends, and is used in subsequent steps S4 and S5.

[0045] 7.1) The overall calculation process of steps S1 to S3 in this embodiment connects the above three modules in series, forming a complete closed loop. Within one calculation cycle, the system performs the following operations in parallel: Each distributed node acquires vibration signals and listens to control command streams. The local or regional master node performs calculations to generate the latest dynamic coupling matrix C. Simultaneously, it performs calculations to generate the latest control influence matrix H.

[0046] The dynamic coupling matrix C and the control influence matrix H are input into step S3. Specifically, the two matrices are topologically ordered and normalized, and then the divergence calculation of adaptive time-delay integral and context-gated control is performed, finally outputting a stable, reliable, time-integrated physical-logical mismatch signal. This signal is then used for mismatch localization in step S4 and rebalancing instruction generation in step S5, thereby completing a complete and intelligent fault self-healing closed-loop handling process.

[0047] To decouple the core algorithm from the specific application strategy of this invention, and to ensure the configurability and ease of debugging of the technical solution, in the specific implementation path of this invention, all configurable runtime configuration parameters are read through a standardized "configuration interface". The data source of this interface is a "data storage module" (containing a non-transitory computer-readable storage medium on which a structured text file such as a configuration file in JSON or YAML format is stored). The following is a further explanation of steps S1 to S3 above: The "time-integrated physical-logic mismatch signal" ultimately generated by the technical solution of this invention is also known as the "System Mismatch Index (SDI)". This signal is a normalized scalar, and its output value range is strictly limited to the interval [0,1]. The boundary of this value range has clear and crucial technical implications: When the system mismatch index (SDI) approaches 0, it indicates a high degree of coordination and consistency between the physical interaction topology and the logical control topology of the distributed system. In this state, the probability distributions of the dynamic coupling matrix C and the control influence matrix H are similar after topological rank sorting and normalization. Physically, there are tightly coupled node pairs (connected by highly rigid structures or through frequent transmission of vibrational energy), and logically, they also exhibit a strong dependency (interacting through high-frequency, regular, and deterministic control commands). This state is an ideal representation of a healthy and stable system operation, indicating a perfect match between the system's physical structure design and the upper-level control strategy, high collaborative efficiency, and low potential structural risks.

[0048] When the system mismatch index (SDI) approaches 1, it indicates a severe mismatch or divergence between the physical interaction topology and the logical control topology of the distributed system. In this state, the probability distributions of the two matrices differ significantly. Technically, this manifests in the following typical abnormal operating conditions: Physical anomalies dominate: Two nodes with no strong logical dependency (sparse and random control commands) exhibit an abnormally high physical coupling strength. This directly points to potential physical faults, such as a loose component causing abnormal vibration transmission, or bearing wear leading to heat and vibration conduction across nodes.

[0049] Logical anomalies dominate: Two physically decoupled or weakly coupled nodes exhibit an abnormally high degree of logical dependency. This directly points to potential faults at the control or network level, such as a node's software getting stuck in an error loop, causing it to send a large number of redundant instructions to unrelated nodes, or network routing anomalies causing communication traffic to concentrate on unexpected paths. Therefore, when SDI approaches 1, it clearly indicates a deep-seated structural risk across the physical and control layers, requiring further fault diagnosis and self-healing closed-loop management.

[0050] The final output of this invention, the System Mismatch Index (SDI), is directly influenced by a series of key intermediate parameters. With other input parameters remaining constant, an increase in the cross-correlation coefficient or phase-lock value between any pair of nodes will monotonically increase the value of its corresponding multidimensional dynamic coupling descriptor. This will increase the value of the corresponding element in the dynamic coupling matrix C, thereby altering the overall topology of the dynamic coupling matrix C. If the control influence matrix H does not possess equally strong logical dependencies at the corresponding positions, the difference in the normalized probability distributions of the two matrices will increase, leading to an increase in the Jensen-Shannon divergence value, ultimately causing a monotonically increasing SDI. Therefore, these two parameters are positively correlated with the final SDI. This positive correlation reflects the failure evolution laws of the physical world. In industrial equipment, unexpected, localized enhancements in physical coupling (abnormal contact due to loose components or increased friction due to lubrication failure) are manifested through the energy (characterized by the cross-correlation coefficient) and synchronicity (characterized by the phase-lock value) of the vibration signal. This embodiment can accurately capture the initial signs of such physical anomalies and transform them into the growth of mismatch signals. If such a physical anomaly occurs in a place where there should logically be no strong interaction, it will be identified as a high-risk mismatch.

[0051] Regarding the "conditional information entropy of instruction type transfer" and "timing information entropy" in S2, the "degree of logical dependence" is negatively correlated with information entropy. Therefore, under the condition that other input parameters remain unchanged, an increase in the "conditional information entropy of instruction type transfer" or "timing information entropy" between any pair of nodes will monotonically decrease the value of its corresponding "degree of logical dependence." This will make the value of the corresponding element in the control influence matrix H smaller. If the dynamic coupling matrix C has strong physical coupling at the corresponding position, the distribution difference between the two matrices will increase, ultimately causing a monotonically increasing system mismatch index SDI. Therefore, these two information entropy parameters are positively correlated with the final SDI. This positive correlation reveals the failure mode of the logic control system. Information entropy is a measure of the degree of uncertainty or disorder in a system. In a healthy control system, the instruction flow should be highly deterministic and regular in both content and timing. When defects occur, network congestion occurs, or node controllers fail, the external manifestation is often a chaotic and disordered instruction flow. This embodiment transforms the increase in "chaos" at the logical level into a decrease in "logical dependency," ultimately manifesting as an increase in mismatch signals. This enables the present invention not only to monitor physical faults but also to more acutely perceive "soft" faults in the control system, achieving unified quantification of cross-domain faults.

[0052] Regarding the "context state signal" in S3; the "context state signal" itself does not directly participate in the calculation of the instantaneous divergence value, but it nonlinearly affects the dynamic response characteristics of the final output SDI by adaptively adjusting the weight coefficients of the EWMA filter. Specifically: when the context state is "stable production", the corresponding weight coefficient is set to a relatively high 0.9, close to 1. This makes SDI extremely sensitive to newly generated instantaneous divergence values, enabling it to quickly increase the SDI value in response to anomalies.

[0053] When the context state is "start-up debugging," the corresponding weighting coefficient is set to a low value of 0.05, close to 0. This makes SDI exhibit strong smoothing and suppression capabilities against fluctuations in instantaneous divergence values, tolerating normal, temporary mismatches during the startup phase. Therefore, the relationship between this signal and SDI is not a simple positive / negative correlation, but rather a gating relationship that adjusts its "sensitivity" and "inertia." This gating mechanism is a technological advancement that distinguishes this invention from traditional static threshold monitoring systems. In industrial systems, the boundary between "normal" and "abnormal" changes dynamically at different operating stages. During stable production, any minor mismatch could be a precursor to a major failure, requiring the system to have "zero tolerance" alertness. However, during start-up, shutdown, or operating condition switching phases, asynchronous dynamic responses of system components are physically inevitable. If high-sensitivity monitoring is still used at this time, a large number of meaningless false alarms will be generated. This invention, by introducing a context state signal, endows the monitoring algorithm with "contextual intelligence," improving the signal-to-noise ratio of fault monitoring and the reliability of decision-making.

[0054] Further explanation: In S4, the location of mismatched node pairs based on physical-logical mismatch signals includes: The mismatch contribution matrix, generated during the topological divergence calculation and representing the contribution of each node to the mismatch, is used as part of the physical-logic mismatch signal; and, By identifying the peak elements in the mismatch contribution matrix, the node pairs corresponding to the peak elements are taken as mismatch node pairs.

[0055] Further explanation: Identifying peak elements to locate mismatched node pairs specifically includes: processing the mismatch contribution matrix into a weighted network graph; performing a community detection algorithm on the weighted network graph to identify one or more mismatched topology clusters composed of strong mismatched node pairs; and within each mismatched topology cluster, calculating the network centrality of each node to identify the root cause node; and identifying the root cause node and its neighboring node with the largest mismatch contribution value within the mismatched topology cluster as the root cause mismatched node pair that caused the cluster mismatch.

[0056] Further explanation: In S5, the rebalancing command is configured to adaptively select at least one of the following adjustment strategies based on the mismatch type of the mismatched node pair: If the mismatch type is physically strong coupling but logically weak dependency, then increase the synchronization frequency of control commands between mismatched node pairs; If the mismatch type is physically weakly coupled but logically strongly dependent, then reduce the communication priority of control commands between the mismatched node pairs.

[0057] Further explanation: The adaptive selection of adjustment strategies specifically includes: generating a set of candidate rebalancing instructions covering different adjustment granularities for root cause mismatch node pairs; using a pre-trained system performance proxy model to predict the comprehensive impact of executing each candidate rebalancing instruction on preset key system performance indicators and future physical logic mismatch signals; calculating a utility score for each candidate rebalancing instruction based on the comprehensive impact; and selecting and generating the candidate rebalancing instruction with the highest utility score as the final rebalancing instruction.

[0058] Further explanation: Before S3, it also includes: Obtain confidence scores for the dynamic coupling matrix and the control influence matrix; Furthermore, in S3, if the confidence score of any matrix is ​​lower than the preset confidence threshold, the contribution of the low confidence score matrix is ​​masked when performing topological divergence calculation.

[0059] Further explanation: Masking the contribution of the low-confidence scoring matrix specifically includes: converting the confidence scores of the dynamic coupling matrix and the control influence matrix into their respective information entropy weights; during the topological divergence calculation, using the information entropy weights to weight and fuse the dynamic coupling matrix and the control influence matrix to generate a confidence-weighted fusion matrix; and performing subsequent processing on the confidence-weighted fusion matrix, thereby dynamically reducing the contribution of the low-confidence scoring matrix to the final mismatch signal. The following is a detailed implementation description of the above: Before the calculation process of this embodiment begins, the system loads a series of preset configuration parameters from an externally stored, structured spreadsheet file. The key parameters involved in this embodiment are defined as follows: 8) The parameters related to mismatch localization and self-healing are as follows: 8.1) Community Detection Resolution Parameter, denoted as CDR: Its logical meaning is to control the granularity of the community detection algorithm. A higher value tends to create smaller, more compact clusters. It is determined by presetting based on the system node size and typical failure modes to achieve a balance between "location accuracy" and "avoiding over-segmentation." The value of the "Community Detection Resolution Parameter CDR" directly determines the granularity of identifying mismatched topology clusters. A value that is too low will cause faults from different sources to be incorrectly merged into a large cluster, reducing location accuracy; a value that is too high will over-segment a highly correlated fault cluster into multiple fragments, making it impossible to identify its systemic root cause. Therefore, an optimal CDR value needs to be determined; the specific determination steps are as follows: The core idea behind determining the optimal CDR value in this embodiment is to treat it as an optimization problem. The goal of the optimization is to maximize the matching degree between the "mismatched topology clusters" identified by the CDR value and the set of associated nodes in historically known "real failure events" confirmed by operations and maintenance personnel. This embodiment uses a computational experiment method based on historical data backtracking verification to find the optimal parameters by quantifying this matching degree.

[0060] 8.11) From the historical operation and maintenance logs and monitoring database of the distributed system, select a set of historical events with clear fault labels. Each historical event should contain two parts: first, the "mismatch contribution matrix D" within the time window preceding the event; second, the "set of real faulty nodes" directly related to the event, determined manually or through post-event analysis. This dataset constitutes the calibrated "standard answer".

[0061] 8.12) Based on the total number of system nodes and network topology characteristics, a reasonable search range for CDR candidate values ​​is defined. In this embodiment, in a system with 100 nodes, this range is set to a discrete set from 0.1 to 2.0 with a step size of 0.1. Each candidate CDR value is traversed. For each value, it is applied to all "mismatch contribution matrices D" in the validation dataset, and a community detection algorithm is executed to obtain a set of "predicted mismatch topology clusters". For each historical event, its "set of real faulty nodes" is compared with the corresponding "predicted mismatch topology cluster". The Jaccard similarity coefficient is used as a matching metric. The calculation logic of this metric is: divide the number of common nodes in the two sets by the total number of all unique nodes in the two sets. The resulting value is between 0 and 1, where 1 indicates a perfect match. For each candidate CDR value, the average Jaccard similarity coefficient obtained on the entire validation dataset is calculated. Finally, the CDR value that maximizes the average Jaccard similarity coefficient is selected as the "community detection resolution parameter" for the final deployment of this system. In this embodiment, the parameter value determined through the above process is 1.2. The technical consideration behind this value is that it is a value that has been verified on historical data and can achieve the best balance between "avoiding underreporting of systemic risks" and "avoiding false positives about irrelevant nodes".

[0062] The centrality metric type, denoted as CMT, defines the network centrality calculation method used to identify root cause nodes. Optional values ​​include "between-number centrality" or "degree centrality," etc. It is determined by configuring the root cause type to be identified. In this embodiment, the root cause types include "transportation hub" and "connection core."

[0063] The choice of "Centrality Metric Type (CMT)" depends on the network topology characteristics and expected main failure modes of the distributed system to be diagnosed. Its selection logic is solidified into a deterministic set of rules: Rule 1: If the logical control topology of the system is represented by the "control influence matrix H" and presents a "star" or "bus" structure, in which a few core nodes communicate with a large number of edge nodes, then CMT is configured as "degree centrality". In this type of topology, the root cause of failure is often the failure of the core "hub" node, and its most significant feature is high connectivity.

[0064] Rule 2: If the system's logical control topology exhibits a "mesh" or "distributed peer-to-peer" structure, with complex and redundant interaction paths between nodes, then the CMT is configured with "between-centrality." The technical consideration is that in such topologies, the root cause of failures often lies in the "bridge" or "bottleneck" nodes located on multiple critical communication paths. Their failure can lead to system fragmentation or decreased communication efficiency, and between-centrality is precisely an indicator used to measure the "bridge" role a node plays in the network. During system initialization, the topological characteristics of its control influence matrix H are first analyzed, specifically including calculating the diameter and average path length of the network graph. Then, the corresponding CMT is selected and configured according to the aforementioned rule set.

[0065] To achieve adaptive configuration of the "Centrality Metric Type (CMT)", the following automated decision-making process is executed during the initialization phase. First, obtain the "Control Influence Matrix H". For topology analysis, it needs to be converted into an undirected, unweighted graph: construct a zero matrix of the same size as the Control Influence Matrix H. If the value of element Hi1j1 in the Control Influence Matrix H is greater than a preset "connection strength threshold" (in this embodiment, the connection strength threshold is 0.2, used to filter out weak, non-dominant logical dependencies), then set both Hi1j1 and Hj1i1 positions in the new matrix to 1. Next, calculate the "average path length" based on this unweighted graph. The calculation logic is as follows: for each pair of different nodes in the graph, calculate the shortest path length between them, i.e., the minimum number of edges required; then, sum all these shortest path lengths to obtain a total; finally, divide this sum by the total number of pairs of different nodes in the graph. After obtaining the "average path length," the following decision rule is applied: A "topology type determination threshold" is obtained. This threshold is determined by performing the above calculations on known "star" and "mesh" industrial network topologies and taking the critical point of their average path length distribution; in this embodiment, it is 4.5. If the calculated "average path length" is less than this topology type determination threshold, it indicates that the network nodes are tightly connected and the paths are short, exhibiting "mesh" topology characteristics. In this case, CMT is automatically configured as "between-centrality." Conversely, if it is greater than or equal to this topology type determination threshold, it indicates that the network structure is relatively sparse and hierarchical, exhibiting "star / bus" topology characteristics. In this case, CMT is configured as "degree-centrality."

[0066] 8.2) System Performance Proxy Model, denoted as SPSM: This is a lightweight computational model pre-trained using machine learning on historical system operating data. Logically, it receives candidate rebalancing instructions as input and can quickly output predicted values ​​for future key system performance indicators (output, energy consumption). It is determined by offline training and storage in system memory. The construction and training process of the System Performance Proxy Model SPSM is as follows: A Multilayer Perceptron (MLP) is chosen as the basic architecture of the model. The input layer is designed to receive a normalized vector containing all information describing the current system state and candidate rebalancing instructions. The output layer outputs a vector containing predictions of future Key Performance Indicators (KPIs) and the System Mismatch Index (SDI). The input vector specifically includes: the current SDI value, the mismatch type encoding of the root cause mismatch node pair, the current physical coupling strength and logical dependency of the root cause mismatch node pair, and the type encoding and strength parameters of the candidate rebalancing instructions. The output vector specifically includes: the predicted average output over the next minute, the predicted average energy consumption per unit, and the predicted peak SDI over the next minute. In this embodiment, the MLP model includes one input layer, three hidden layers, and one output layer. The number of neurons in the three hidden layers are 64, 32, and 16, respectively. All hidden layers use the Modified Linear Unit (ReLU) as the activation function, and the output layer uses a linear activation function.

[0067] The dataset used to train the model meets the following key characteristics: The data source is historical monitoring and control logs covering at least 12 months of system operation, including normal, abnormal, and various manual / automatic adjustment operations. The data sampling frequency is no less than 1Hz. All time series data must be windowed and normalized using the Z-score method before being input into the model to eliminate dimensional differences and accelerate model convergence. All obvious sensor noise or missing data segments in the dataset have been imputed using cubic spline interpolation. Each sample in the dataset consists of a "state-action" pair and an "outcome" vector. The "state-action" pair is the model input, and the "outcome" vector is the true value that the model needs to predict.

[0068] The model training process uses Mean Squared Error (MSE) as the loss function, aiming to minimize the MSE between the model's predicted values ​​and the true results in the dataset. The Adam optimizer is used for iterative updates of the model parameters, with an initial learning rate set to 0.001. An early-stopping strategy is employed during training, monitoring the model's performance on an independent validation set. If the loss on the validation set no longer decreases after 10 consecutive training epochs, training is stopped to prevent overfitting. The model's network structure, weights, and bias parameters are fixed and stored in a file. During step S5, this file is loaded, instantiating an executable SPSM.

[0069] 8.3) Utility Function Weight Set, denoted as UFWS: This set contains a set of weight coefficients, each corresponding to a different degree of improvement in key system performance indicators and mismatch signals. Logically, it defines the preference between "improving performance" and "reducing risk" when making self-healing decisions. This is specifically configured by system operation and maintenance experts based on current production goals. Production goals include maximizing output or minimizing energy consumption.

[0070] The specific method for determining the utility function weight set UFWS is as follows: Introducing the "Analytic Hierarchy Process (AHP)" from operations research, the qualitative understanding of operational objectives by experts is transformed into a set of quantitative and logically consistent weight coefficients.

[0071] Define a hierarchical model for decision-making. The top layer is the overall objective: "Achieve optimal self-healing rebalancing". The second layer consists of evaluation criteria, namely the various outputs of the SPSM prediction: "Increase output", "Reduce energy consumption", and "Reduce the system mismatch index (SDI) corresponding to future risks".

[0072] Construct a pairwise comparison judgment matrix. Through a human-computer interaction interface, pose a series of pairwise comparison questions to operations experts or system administrators: "In the current production phase, which is more important, 'increasing output' or 'reducing energy consumption'? How important are they?" Experts must provide judgments on a preset 9-point scale (1 represents equal importance, 9 represents extreme importance). Construct a judgment matrix from all pairwise comparison results. Specifically, this transformation process follows the standard analytic hierarchy process (AHP) 1-9 scale: if the expert considers both criteria equally important, the corresponding matrix element is assigned a value of 1; if the expert considers the former slightly more important, significantly more important, strongly more important, or extremely more important than the latter, the corresponding values ​​are 3, 5, 7, and 9 respectively; 2, 4, 6, and 8 are the intermediate values ​​of the above judgments. Simultaneously, elements symmetrical about the diagonal in the matrix are reciprocals of each other. If, when comparing "increasing production" (row i1) and "reducing energy consumption" (column j1), the expert judges the former to be "significantly important", then the element (i1,j1) in the matrix is ​​assigned the value 5, while the element (j1,i1) is automatically assigned the reciprocal of 5, i.e., 0.2.

[0073] For the judgment matrix constructed in the previous step, the calculation logic is executed to determine the weights: Each column of the judgment matrix is ​​normalized, i.e., each element of that column is divided by the sum of all elements in that column; the normalized matrix is ​​then summed row-wise; each element of the vector obtained by this row-wise summation is divided by the sum of all elements in the vector, i.e., the dimension of the vector. The resulting vector is the weight of each evaluation criterion, thus representing the "Utility Function Weight Set (UFWS)". To ensure the logical consistency of expert judgments, the consistency ratio of the judgment matrix is ​​calculated. If this ratio is lower than a preset threshold of 0.1, the weights are considered valid; otherwise, experts are prompted to re-examine their pairwise comparisons.

[0074] 9) Confidence fusion-related parameters include: The confidence-entropy mapping curve parameter, denoted as CEMP, defines a functional relationship that non-linearly maps confidence scores in the [0,1] interval to information entropy weights. It is determined based on prior knowledge of the relationship between signal quality degradation and increased information uncertainty, ensuring that the weight approaches zero when the score is low and approaches one when the score is high. The method for determining the "confidence-entropy mapping curve parameter (CEMP)" is as follows: The core of this mapping is to construct a function that satisfies the following boundary conditions and monotonicity requirements: when the confidence is 1, the uncertainty is 0, and the weight should be 1; when the confidence is 0, the uncertainty is maximum, and the weight should be 0; as the confidence decreases, the weight should decrease smoothly and non-linearly, and in the lower confidence interval, the weight should decrease faster to reflect the penalty for low-quality information. The specific construction content is as follows: For a given confidence score S (range 0 to 1), the calculation logic for its corresponding information entropy weight W is as follows: Read two core parameters from the configuration file: the curve's "center point parameter Smid" and "steepness parameter k". Subtract the "center point parameter Smid" from the confidence score S to obtain a median difference. Multiply this median difference by the "steepness parameter k" and then take its negative. Calculate the natural constant e raised to the power of the result from the previous step. Add 1 to the exponentiation result from the previous step to obtain a denominator. The final weight W is 1 divided by this denominator.

[0075] "Centerpoint parameter Smid": In this embodiment, it is set to 0.5. It is used to define the point where the confidence score equals 50% as the inflection point where the weight begins to drop sharply.

[0076] "Steepness parameter k": In this embodiment, it is set to 10. It is used to provide a relatively steep transition zone. Through simulation verification, when k=10, as the confidence score decreases from 0.7 to 0.3, the weight will rapidly decay from close to 1 to close to 0, which can effectively distinguish between reliable and unreliable information, while avoiding the problem of drastic weight jumps caused by small fluctuations near the threshold.

[0077] 10.1) The preferred implementation of step S4, based on the root cause chain tracing of mismatch topological clustering, is as follows: In distributed systems, faults are often not isolated, but rather form a "fault propagation chain" or "fault resonance zone" where one node affects another. Single-point localization cannot reveal this systemic, structural risk. The core idea of ​​this invention originates from complex network theory, treating the "mismatch contribution matrix" as a weighted adjacency matrix describing the system's "pathological network." Edges with higher values ​​in the matrix represent stronger mismatch relationships between two nodes. By applying a community detection algorithm, this complex pathological network can be divided into several "mismatch topology clusters" with tightly connected internal connections and sparse external connections. Then, within each cluster, by calculating network centrality, the core nodes on the mismatch propagation path are found, which are the root causes. The specific implementation steps are as follows: Obtain the "Mismatch Contribution Matrix D" generated in the previous steps. Treat each node as a vertex in the network graph. For any non-zero element Di2j2 in the mismatch contribution matrix D, create an edge between vertex i2 and vertex j2, and use the element value Di2j2 as the weight of the edge. This constructs a complete weighted network graph representing the system's mismatch topology. On the network graph constructed in the previous step, execute the community detection algorithm; this embodiment uses the Louvain algorithm. The "Community Detection Resolution Parameter CDR" of this algorithm is read from an external configuration file. By iteratively optimizing the network's modularity, the entire network graph is divided into one or more communities. Each identified community is defined as a "Mismatch Topology Cluster". For each identified "Mismatch Topology Cluster", calculate the network centrality of all nodes within that cluster individually. The centrality calculation method is specified by the "Centrality Metric Type CMT" parameter; in this embodiment, "between-the-nodes centrality" is selected. After calculation, find the node with the highest between-the-nodes centrality within the cluster. The node pair formed by the node and its neighbor node with the largest connecting edge weight within the cluster is identified as the "root cause mismatch node pair" causing the mismatch in that cluster. All identified "mismatch topological clusters" and their corresponding "root cause mismatch node pairs" are summarized and sorted in descending order according to the overall mismatch contribution of the clusters, forming a structured, prioritized mismatch location report. This embodiment uses topological clustering analysis to achieve a leap from "point" diagnosis to "surface" diagnosis, enabling the identification of both single and concurrent faults, thus improving the comprehensiveness and depth of the diagnosis.

[0078] 10.2) The preferred implementation of step S5, based on predictive rebalancing of cost simulation and utility optimization, is as follows: Traditional self-healing strategies are based on static rule-based "open-loop" control, which cannot predict the potential side effects of the strategy itself. If a node's priority is lowered to resolve communication congestion, it may fail to complete time-sensitive critical tasks, leading to more serious production problems. The core idea of ​​this invention is to transform the self-healing decision-making process from a simple "rule matching" problem into a "constrained optimization" problem. Essentially, before making any adjustments, a "virtual future" is constructed. A lightweight "System Performance Proxy Model (SPSM)" is used to simulate the possible consequences of different decisions, and the decision that maximizes the overall "utility"—that is, the benefit minus the cost—is selected. The specific implementation details are as follows: For the identified highest-priority "root cause mismatch node pair," a set of candidate rebalancing instructions covering different adjustment intensities is generated based on their mismatch type. If the type is physically strong and logically weak, a series of instructions such as "increase synchronization frequency by 5%", "increase synchronization frequency by 10%", and "increase synchronization frequency by 20%" are generated. Each instruction in the candidate rebalancing instruction set is traversed. The parameter "increase by 10%" of the instruction is used as input and fed into the pre-trained "System Performance Proxy Model SPSM". SPSM will output a set of predicted values, including estimates of key performance indicators (expected output, overall energy consumption) and the expected value of the system mismatch index SDI within a future time window after executing the instruction.

[0079] Before calculating the utility score, to eliminate the differences in dimensions and numerical ranges among different performance indicators, a normalization operation is performed on each predicted value output by SPSM, mapping it uniformly to the [0,1] interval. This embodiment uses the maximum-minimum normalization method. The calculation logic of this method is as follows: from the preset performance indicator boundary table, query the historical maximum and historical minimum values ​​corresponding to the current predicted value; then, subtract the historical minimum value from the current predicted value to obtain a difference; finally, divide this difference by the result of (historical maximum value minus historical minimum value). All predicted values ​​processed in this step become dimensionless values ​​ranging from 0 to 1, used for subsequent weighted calculations. For each candidate rebalancing instruction, its "rebalancing action utility score" is calculated based on its prediction results. The calculation process is as follows: obtain the "utility function weight set UFWS"; multiply the predicted output increase value by its corresponding weight, add the predicted energy consumption reduction value multiplied by its corresponding weight, and then subtract the predicted future SDI value multiplied by its corresponding weight; sum all weighted terms to obtain the final utility score of the candidate rebalancing instruction. The utility scores of all candidate rebalancing instructions are compared, and the instruction with the highest score is selected as the final rebalancing instruction to be executed. This instruction is then encoded into a format recognizable by the target node controller and distributed through the distributed control network.

[0080] 10.3) Optimal Implementation of Confidence Scoring: Adaptive Weighted Fusion Based on Information Entropy. The traditional "hard masking" mechanism discards data completely when its quality falls below a certain threshold. For example, a signal with a confidence score of 0.4 and a threshold of 0.5, while imperfect, still contains useful information; discarding it entirely is a waste of information. The core idea of ​​this invention is that, in multi-source information fusion, the optimal strategy is to allocate weights based on the information entropy of each source. Sources with higher information entropy have lower confidence scores and thus lower weights. Specific implementation details are as follows: Before step S3, obtain the confidence score ScoreC of the dynamic coupling matrix C and the confidence score ScoreH of the control influence matrix H. Query the "Confidence-Entropy Mapping Curve Parameter CEMP" and convert these two scores into information entropy weights WC and WH, respectively. This mapping relationship is designed as follows: the higher the confidence score, the closer the information entropy weight is to 1; the lower the confidence score, the closer the information entropy weight is to 0. Before calculating topological divergence, a "confidence-weighted fusion matrix M" is first generated. This matrix is ​​generated as follows: the value of any element Mi1j1 in matrix M is calculated by multiplying the corresponding element Ci1j1 of the dynamic coupling matrix C by the weight WC, and then adding the corresponding element Hi1j1 of the control influence matrix H multiplied by the weight WH. Subsequently, this fusion matrix M is used for subsequent topological rank normalization and topological divergence calculation. An adaptive weighted fusion based on information entropy replaces the hard masking mechanism. This allows for dynamic and smooth adjustment of the influence of different information sources in the final decision based on data quality.

[0081] 11) Based on S1 to S3, the overall calculation process of this embodiment forms a complete self-healing closed loop of "perception-diagnosis-decision-execution": Through continuous optimization of S1-S3, a stable, reliable, time-integrated physical-logical mismatch signal (SDI) and a mismatch contribution matrix D are generated. Once the SDI exceeds a preset mismatch threshold, topological clustering and root cause tracing are performed on the mismatch contribution matrix D, outputting a priority mismatch location report. Upon receiving this report, for the highest priority root cause problem, an optimal rebalancing instruction is generated through cost simulation and utility optimization. This instruction is issued to relevant nodes for execution, adjusting their logical interaction modes. The physical and logical states thus change, and this change is perceived in real time by modules S1 and S2, forming a new round of SDI calculation, thus constituting a continuously operating, self-optimizing closed-loop control system.

[0082] It should be noted that the preset mismatch threshold is denoted as SDI. th This parameter defines a specific critical value for the "System Mismatch Index (SDI) characterized by the Physical-Logical Mismatch Signal." A threshold is set if and only if the real-time monitored SDI value exceeds the threshold value. thOnly when the current state deviation exceeds the range of normal fluctuations is it determined that it constitutes a "mismatch event" requiring proactive intervention. Setting a mismatch threshold is key to balancing the core contradiction between "response sensitivity" and "computational resource consumption." Determining the SDI... th The core idea is to transform the problem from an engineering experience-based problem into a decision threshold optimization problem using a binary classification model. Each time-instance state is divided into two categories based on whether a "critical mission failure" will subsequently occur: "normal state" and "pre-failure state". The SDI signal is the classifier score used to distinguish between these two states. Optimal SDI th This is the score value that enables the classifier to achieve optimal overall performance when distinguishing between these two states. This invention uses the "Receiver Operating Characteristic (ROC) curve" analysis method to find this optimal point.

[0083] Two sets of time-series data were extracted from the long-term operation logs: one set was a continuous "Physical Logic Mismatch Signal (SDI)" time series; the other set was a synchronized "System Critical Task Success Rate (CTSR)" time series. The SDI data was labeled as follows: for any SDI(t) value at time point t, if the CTSR dropped to a preset failure line (below 95%) within a subsequent preset time window Δt (Δt = 5 minutes), then the SDI(t) data point was labeled as a "positive sample" (representing a precursor to failure); otherwise, it was labeled as a "negative sample" (representing normal operation). The ROC curve is a graph of "True Positive Rate (TPR)" versus "False Positive Rate (FPR)" at all possible thresholds. The logic of its generation steps is as follows: Set a candidate threshold, starting with the minimum SDI in the dataset. Use this candidate threshold to classify and predict all data points: SDI values ​​above the threshold are predicted as positive samples, and values ​​below are predicted as negative samples. Compare the prediction results with the true positive rate (TPR), representing the proportion of correctly predicted positive samples, and calculate the false positive rate (FPR), representing the proportion of incorrectly predicted positive samples. Plot (FPR, TPR) as a point on a two-dimensional coordinate system. Increment the candidate threshold and repeat the corresponding steps after "classifying and predicting all data points" until the candidate threshold reaches the maximum value of SDI. Connecting all the plotted points forms the ROC curve. Optimal SDI th The point on the ROC curve closest to the ideal point at the top left corner (0,1) represents the highest classification accuracy. This embodiment achieves this by maximizing the Youden index. The calculation logic is as follows: for each point on the ROC curve, calculate the difference between the true positive rate (TPR) and the false positive rate (FPR). Find the point where this difference reaches its maximum value. The candidate threshold corresponding to this point is the SDI ultimately adopted in this invention. thIn this embodiment, the final determined SDI th The value is 0.68. This represents the critical point where the optimal mathematical balance can be achieved between "ensuring early warning of 98% of true fault precursors (high TPR)" and "keeping the false alarm rate caused by normal system fluctuations below 5% (low FPR)".

[0084] 12) In step S5 of this invention, a "Rebalancing Action Utility Score" is calculated for each candidate rebalancing instruction using the "System Performance Proxy Model (SPSM)," and denoted as RAUS. This score is a dimensionless scalar, whose theoretical range can be the entire real number domain, but in practice it is distributed within a specific interval. Its numerical trend has a clear technical meaning: When RAUS is significantly positive (RAUS > T1), it indicates that the system predicts that executing the corresponding candidate rebalancing instruction will bring a net positive benefit. Here, T1 is a performance threshold used to define significant effectiveness, and its specific value is uniquely determined by the Objective Threshold Derivation (OTD) method based on the S-curve performance inflection point calculation, as detailed later in this invention. This state means that the instruction can not only effectively reduce the future System Mismatch Index (SDI), but also maintain or even improve key system performance indicators. A higher positive value indicates a higher "cost-effectiveness" of the instruction.

[0085] When RAUS approaches 0 or fluctuates only slightly: This indicates that the system predicts the benefits and costs of executing the corresponding candidate rebalancing instruction are roughly offset, or the overall impact is negligible. This state means that while the instruction can slightly reduce SDI, it will equally damage a certain performance metric, and vice versa. This is an "ineffective" or "inefficient" adjustment, and the decision-making mechanism of this invention will tend to abandon such instructions to avoid unnecessary system disturbances.

[0086] When RAUS is negative: This indicates that the system predicts that executing the corresponding candidate rebalancing instruction will have a net negative impact. This state clearly indicates that although the instruction may superficially address a mismatch problem, the cost of its execution far outweighs its benefits. It allows for the quantification and identification of such "harmful" operations, and their exclusion from the final decision.

[0087] An increase in the element value of the "mismatch contribution matrix" directly leads to an increase in the weight of the corresponding edge when constructing the "mismatch topology graph." This makes the edge more likely to be assigned to a "mismatch topology cluster" by the "community detection algorithm," increasing the overall mismatch contribution of its cluster and thus raising the diagnostic priority of that cluster. Therefore, the element value of the mismatch contribution matrix is ​​positively correlated with the probability of ultimately locating the node pair and generating a rebalancing instruction for it. Computational resources and self-healing actions are prioritized for the node region that contributes the most to the overall system mismatch. By transforming individual "mismatch contribution values" into "topology weights" in the network graph, this invention not only focuses on the "intensity" of the mismatch but also on its "structure," ensuring that diagnosis and response directly address the core of the problem.

[0088] Each weight coefficient in the "Utility Function Weight Set (UFWS)," including the "output improvement weight," has a direct, linear, and positive correlation with its corresponding item in the final "Rebalancing Action Utility Score (RAUS)." When other conditions remain constant, increasing the value of the "output improvement weight" will proportionally increase the RAUS value of candidate rebalancing instructions predicted by the SPSM model to bring higher output gains, making them more likely to be selected as the final execution instructions. This approach transforms abstract, qualitative operational goals into concrete, quantitative algorithm parameters, ensuring that self-healing behavior is no longer merely a technical fix but rather an intelligent decision serving to maximize overall business value.

[0089] 13) To quantitatively verify the superiority of the closed-loop self-healing scheme based on topological clustering and utility prediction proposed in this invention compared with existing technologies, a computing environment based on the MATLAB platform was built in this embodiment. A mathematical model of a distributed robot collaborative operation system containing 20 nodes was constructed within this environment. The experimental design aims to demonstrate the beneficial effects of this invention by comparing the performance of the method of this invention (experimental group) with the benchmark method (control group) in handling typical fault scenarios. The control group adopts a simplified strategy representing existing technologies: a single node pair is located by finding the maximum value in the "mismatch contribution matrix," and a preset, static binary adjustment rule is executed. Furthermore, to verify the robustness of the confidence-weighted fusion mechanism of this invention, scenario four is added. In this scenario, the signal quality of the simulated physical coupling sensing module deteriorates, resulting in a confidence score of 0.4 for its output "dynamic coupling matrix," which is lower than the preset shielding threshold of 0.5. See Table 1 below for details. Table 1: Performance Comparison Experimental Data of the Closed-Loop Self-Healing Method of the Present Invention and the Control Group Method To comprehensively evaluate the effectiveness of the two methods, this analysis introduces a system health index, denoted as SHI. The calculation logic for this parameter is as follows: normalize the "system throughput after loop closure" (divide by the theoretical maximum throughput in this scenario, which is 100 items / minute in this example); subtract the "mismatch index after loop closure" from 1; the mismatch index after loop closure is the representation of the system mismatch index SDI in the closed-loop self-healing scheme; multiply the two results. That is, the steps for calculating the SHI value are: obtain the throughput value and divide it by 100 to obtain the normalized throughput; obtain the system mismatch index value and subtract this value from 1 to obtain the stability factor; multiply the normalized throughput by the stability factor to obtain the final SHI value. The system health index is used to comprehensively quantify the overall operating status after executing the self-healing closed loop. Its value range is [0,1]. The higher the value, the better the production performance is taken into account while restoring stability, that is, a healthier self-healing is achieved.

[0090] Comparative analysis of scenario two (multi-node vibration resonance): In this scenario, the node pair (7,8) with the most severe physical vibration is the "symptom" of the fault, while the node pair (3,4) with rhythm misalignment caused by the upstream controller is the "cause" of the fault. The peak method used in the control group can only identify the symptom, so its diagnosis result is (7,8). The method of this invention, through topological clustering, identifies this as a "mismatched topological cluster" of size 4 composed of nodes {3,4,7,8}, as shown in the "Cluster Information" column of the table. Then, by performing network centrality calculation within this cluster, the "root cause node pair" as the source of the fault is further explored and located as (3,4). This difference in the diagnostic process is the fundamental reason for the differences in all subsequent effects.

[0091] Based on an incorrect diagnosis, the control group performed a standard-intensity synchronization frequency boost (+10%) on nodes (7,8), which provided some relief but failed to eradicate the problem. Therefore, the "mismatch index after loop closure" was 0.45, and the system throughput only recovered to 85. This invention, based on precise root cause diagnosis, directly performed a more targeted and stronger synchronization frequency boost (+15%) on nodes (3,4), correcting the timing deviation at its source. This operation reduced the "mismatch index after loop closure" to 0.18, fundamentally improving system stability and leading to a recovery of the "system throughput after loop closure" to a high level of 96. By calculating a comprehensive "system health index," the invention achieved a SHI value of 0.787 in this scenario, a significant improvement of 68.2% compared to the control group's 0.468. This data demonstrates the disruptive advantage of the invention's "root cause localization based on topological clustering" design. It elevates fault diagnosis from "symptom recognition" to "mechanism tracing," ensuring that self-healing measures can directly address the core issues. This enables it to achieve repair results far exceeding those of existing technologies in complex and interconnected fault scenarios, thereby improving system availability and production efficiency.

[0092] Comparative Analysis Scenario 3 (High-Cost Repair Scenario): In this scenario, the mismatch problem at node (15,16) has two repair paths: one is to significantly reduce the communication priority, which can eliminate the mismatch the fastest, but will severely sacrifice system performance; the other is to slightly increase the synchronization frequency, which eliminates the mismatch more slowly, but has little impact on performance. The static rule of the control group chose the former, resulting in poor performance consequences. The "System Performance Proxy Model SPSM" of this invention predicts these two paths before making a decision. The prediction results show that the "predicted utility score" of the former is a significant negative value, while that of the latter is a positive value of 8.9. Therefore, the decision system intelligently avoids the high-cost repair option. The direct result of this decision is that the solution of this invention increases the "mismatch index after loop closure" from 0.11 to 0.22 in exchange for increasing the "system throughput after loop closure" from 50 items / minute to 95 items / minute. Although the absolute value of the system mismatch index increased, the SHI value of this invention (0.741) was improved by 66.5% compared to the control group (0.445) by calculating the comprehensive indicator of the "system health index". This data strongly demonstrates the practicality of the "decision-making mechanism based on utility prediction" of this invention. It reveals that optimal self-healing does not always pursue the lowest system mismatch index, but rather achieves the best balance between system stability and core performance. By quantifying and optimizing this balance point, this invention successfully avoids over-repair caused by traditional methods, ensuring that self-healing behavior always serves the ultimate goal of maximizing the overall value of the system.

[0093] Comparative Analysis Scenario 4 (Low Confidence Signal): This scenario aims to verify the processing capability of this invention when faced with imperfect input information. The control group's "hard shielding" mechanism completely discarded information in the physical dimension because the confidence level of the physical sensing signal (0.4) was lower than the shielding threshold (0.5), causing the diagnostic system to "go blind" and fail to identify the fault. The "adaptive weighted fusion" mechanism of this invention demonstrates superior robustness. It does not discard the information but assigns it a lower weight based on its confidence level and fuses it with high-quality logical information. This "soft fusion" approach retains sufficient effective information, enabling the diagnostic system to still successfully identify fault clusters and locate the problem node. Ultimately, the "system health index" of this invention reached 0.598 under this harsh condition, achieving a 144.1% improvement compared to the low index (0.245) of the control group due to diagnostic failure.

[0094] 14) By implementing the "Objective Threshold Derivation (OTD)" methodology, the internal decision-making indicator "Rebalancing Action Utility Score (RAUS)" is transformed into a practical application framework with clear technical boundaries and optimized operational strategies. To objectively evaluate the final effect of self-healing decisions, this invention selects "System Critical Task Success Rate (CTSR)" as a verifiable effectiveness indicator (VEI). This indicator directly quantifies the ability of a distributed system to complete its core business functions after performing self-healing operations. Through regression analysis of experimental data, "Rebalancing Action Utility Score (RAUS)" and VEI "System Critical Task Success Rate (CTSR)" show a significant positive correlation with a S-shaped curve relationship. Specifically, when RAUS is negative or in the low positive range, CTSR remains at a low level, indicating that the system is still in a faulty or sub-healthy state; when RAUS increases to the median range, CTSR increases exponentially; in the high value range of RAUS, CTSR approaches 100% and reaches saturation.

[0095] The first threshold point (T1, the efficacy inflection point): This efficacy threshold is defined as the first derivative of the S-curve, i.e., the point where the rate of increase in CTSR reaches its maximum value. This point marks the point where the increase in RAUS has the most dramatic effect on health improvement. In this embodiment, the preferred value for T1 is 10.2.

[0096] The second threshold point (T2, saturation inflection point): This performance threshold is defined as the point where the function value of the S-curve first reaches 95% of its saturation value (100%). This point signifies that a highly optimized state has been reached, and the marginal benefit of further improving RAUS begins to diminish. In this embodiment, the preferred value for T2 is 25.0.

[0097] A threshold T1 of 10.2 marks the objective boundary at which a self-healing instruction transitions from "potentially effective" to "significantly effective." Instructions below this value have uncertain effectiveness.

[0098] A threshold T2 of 25.0 marks the objective boundary at which the self-healing instruction transitions from "effective repair" to "optimal repair," meaning that the instruction not only solves the problem but may also push the system to a better operating point than before the failure.

[0099] Define the cost and utility of the following intervention levels: Zero-level intervention (suppression and alarm): The cost is increased cognitive load for maintenance personnel, and the benefit is to avoid erroneous automatic operations.

[0100] Level 1 intervention (standard automatic execution): The cost is the usual computing and communication overhead, and the benefit is restoring the normal operation of the system.

[0101] Secondary intervention (executing and recording optimization examples): The cost adds the overhead of data recording and model updating on the basis of the primary intervention, and the utility is to drive long-term learning and evolution of the system.

[0102] The intervention level is accurately mapped to the RAUS interval divided by thresholds T1 and T2, so as to achieve an optimized balance between decision risk and benefit. This mapping demonstrates a significant progress of the present invention at the decision execution level, that is, matching the "level" of action according to the "confidence" of decision-making.

[0103] The condition for setting the suppression and alarm interval is RAUS<T1: its upper boundary 10.2 is determined by the "efficiency inflection point" T1 of the system decision. Below this value, the positive effect of the decision has not been fully verified. The action is to suppress the automatic execution of the rebalancing instruction. At the same time, the mismatch event and the candidate rebalancing instruction with the lowest utility score are reported to the operation and maintenance monitoring platform, as a high-priority alarm of "suggesting manual intervention".

[0104] The condition for setting the standard self-healing interval is RAUS∈[T1,T2]: its boundary is jointly defined by the efficiency inflection point T1 marking "significantly effective" and the saturation inflection point T2 marking "optimal repair". The execution action is to automatically execute the rebalancing instruction with the highest utility score.

[0105] The optimization and learning interval is set as RAUS>T2: its lower boundary 25.0 is determined by the "saturation inflection point" T2 reached by system repair, representing a high-value optimization opportunity. The instruction is automatically executed, and the complete data pair of "problem-decision-effect" this time is marked as a "golden sample" and stored in a dedicated database, which is used for incremental retraining of the "System Performance Proxy Model SPSM" in the future, so as to realize the continuous evolution of the system.

[0106] For the construction process of the calculation logic involved in this application, algorithms such as regression analysis in machine learning can be adopted to establish a mathematical model by analyzing the internal trends and mutual relationships of the collected parameters. This process can be implemented with the help of professional computing tools (such as the Scikit-learn library of Python or the R language environment). In all calculation processes, in order to eliminate the influence caused by different physical dimensions and ensure that the data are compared and analyzed on the same scale, the input parameters in each formula are subjected to dimensionless processing. The adopted dimensionless technology includes but is not limited to maximum-minimum normalization or Z-Score standardization.

[0107] The algorithm of this invention is implemented as a Python script. Before executing the core logic, the program first executes a data loading module (e.g., using the widely used pandas library in Python) configured to read the aforementioned spreadsheet file and load its contents into the program's working memory (e.g., a DataFrame data structure). Subsequent algorithm steps will directly query and retrieve the required configuration parameters from this in-memory data structure.

[0108] It should be emphasized that the foregoing embodiments are merely illustrative of preferred implementations of the present invention and are not intended to limit the scope of protection of the present invention. This application also provides a computer-readable storage medium having computer program instructions stored thereon.

Claims

1. A self-healing closed-loop fault handling method for industrial equipment based on monitoring and control, characterized in that, The specific steps include: S1: Obtain the dynamic coupling matrix, which is configured to quantify the physical interaction strength between pairs of nodes in a distributed system. The physical interaction strength is calculated based on the cross-correlation of the vibration signals of the pairs of nodes. S2: Obtain the control influence matrix, which is configured to quantify the degree of logical dependency between each pair of nodes in a plurality of nodes, the degree of logical dependency being calculated based on the information entropy of the control command flow between the pairs of nodes; S3: Perform topological rank sorting and normalization on the dynamic coupling matrix and the control influence matrix respectively to eliminate the differences in data scale and units between the two matrices; Topological divergence is calculated on the dynamic coupling matrix and the control influence matrix to generate a physical-logical mismatch signal; wherein the physical-logical mismatch signal is configured to characterize the degree of mismatch between the physical interaction topology and the logical control topology of the distributed system. The topological divergence calculation specifically involves using Jensen-Shannon divergence to calculate the difference between two normalized matrices. S4: Based on the physical logic mismatch signal, locate the mismatched node pair in the distributed system, wherein the mismatched node pair is the primary contributor causing the physical logic mismatch signal to exceed a preset mismatch threshold, including: The mismatch contribution matrix, generated during the topological divergence calculation and representing the contribution of each node to the mismatch, is used as part of the physical-logical mismatch signal; and... By identifying the peak element in the mismatch contribution matrix, and taking the node pair corresponding to the peak element as the mismatch node pair; Identifying peak elements to locate mismatched node pairs specifically includes: processing the mismatch contribution matrix into a weighted network graph; performing a community detection algorithm on the weighted network graph to identify one or more mismatched topology clusters composed of strong mismatched node pairs; calculating the network centrality of each node within each mismatched topology cluster to identify the root cause node; and identifying the root cause node and the node pair formed by it and its neighboring node with the largest mismatch contribution value within the mismatched topology cluster as the root cause mismatched node pair that caused the cluster mismatch. S5: In response to locating the mismatched node pair, generate a rebalancing instruction for adjusting the logical control interaction between the mismatched node pair.

2. The self-healing closed-loop fault handling method for industrial equipment based on monitoring and control according to claim 1, characterized in that: In S1, the physical interaction intensity is calculated based on the cross-correlation of the vibration signals of each pair of nodes, including: calculating the normalized cross-correlation spectrum of the vibration signals of each pair of nodes within a preset characteristic frequency band, and taking the peak value of the normalized cross-correlation spectrum as the physical interaction intensity.

3. The self-healing closed-loop fault handling method for industrial equipment based on monitoring and control according to claim 2, characterized in that: The calculation of the normalized cross-correlation spectrum and the acquisition of the peak value specifically includes: performing time-frequency analysis on the vibration signal using continuous wavelet transform to generate time-frequency coefficients; based on the time-frequency coefficients, synchronously calculating the cross-correlation coefficient characterizing the energy correlation of the signal and the phase lock value characterizing the phase synchronization of the signal; and combining the cross-correlation coefficient and the phase lock value to generate a multidimensional dynamic coupling descriptor characterizing the physical interaction strength.

4. The self-healing closed-loop fault handling method for industrial equipment based on monitoring and control according to claim 3, characterized in that: In S2, the logical dependency degree is calculated based on the information entropy of the control command flow between each pair of nodes, including: calculating the conditional information entropy of the control command flow from the first node to the second node relative to all control command flows issued by the first node, and using the reciprocal of the conditional information entropy as the logical dependency degree. The calculation of conditional information entropy specifically includes: identifying predefined timing instruction modules composed of multiple instructions from the control instruction stream; calculating the conditional information entropy of instruction type transitions within the timing instruction modules and the timing information entropy of the time intervals between the timing instruction modules; and weighting and combining the conditional information entropy of the instruction type transitions and the timing information entropy to generate the logical dependency degree.

5. The self-healing closed-loop fault handling method for industrial equipment based on monitoring and control according to claim 4, characterized in that: The Jensen-Shannon divergence method is used to calculate the difference, specifically including: continuously calculating a series of instantaneous divergence values ​​within a preset time sliding window; obtaining a context state signal characterizing the current global operating condition of the distributed system; adaptively adjusting the weight coefficients of the exponentially weighted moving average filter based on the context state signal; and smoothing the series of instantaneous divergence values ​​using an exponentially weighted moving average filter with adaptive weight coefficients to generate a time-integrated physical-logical mismatch signal characterizing the long-term mismatch trend.

6. The self-healing closed-loop fault handling method for industrial equipment based on monitoring and control according to claim 5, characterized in that: In S5, the rebalancing instruction is configured to adaptively select at least one of the following adjustment strategies based on the mismatch type of the mismatched node pair: If the mismatch type is physically strong coupling but logically weak dependency, then increase the synchronization frequency of control commands between the mismatched node pairs; If the mismatch type is physically weakly coupled but logically strongly dependent, then the communication priority of control commands between the mismatched node pairs is reduced.

7. The self-healing closed-loop fault handling method for industrial equipment based on monitoring and control according to claim 6, characterized in that: The adaptive selection of adjustment strategies specifically includes: generating a set of candidate rebalancing instructions covering different adjustment granularities for the root cause mismatch node pairs; using a pre-trained system performance proxy model to predict the comprehensive impact of executing each candidate rebalancing instruction on preset key system performance indicators and future physical logic mismatch signals; calculating a utility score for each candidate rebalancing instruction based on the comprehensive impact; and selecting and generating the candidate rebalancing instruction with the highest utility score as the final rebalancing instruction.

8. The self-healing closed-loop fault handling method for industrial equipment based on monitoring and control according to claim 7, characterized in that: Before S3, the process also includes: obtaining confidence scores for the dynamic coupling matrix and the control influence matrix; and, In S3, if the confidence score of any matrix is ​​lower than a preset confidence threshold, then the contribution of the low confidence score matrix is ​​masked when performing the topological divergence calculation. The process of masking the contribution of the low-confidence scoring matrix specifically includes: converting the confidence scores of the dynamic coupling matrix and the control influence matrix into their respective information entropy weights; using the information entropy weights to weight and fuse the dynamic coupling matrix and the control influence matrix during the topological divergence calculation to generate a confidence-weighted fusion matrix; and performing subsequent processing on the confidence-weighted fusion matrix to dynamically reduce the contribution of the low-confidence scoring matrix to the final mismatch signal.

Images