New energy equipment health state evaluation method based on recurrent neural network

Abstract

The application discloses a new energy equipment health state evaluation method based on a recurrent neural network, comprising the following steps: collecting multi-source monitoring data of the new energy equipment and preprocessing to form unified input and working condition information; constructing degradation trend, working condition switching and credibility evidence based on the input data to form an evidence set; generating fast, slow or double update decisions and giving risk budget by using evidence-driven gating; obtaining health characterization by adaptive updating of fast and slow state through a skip double-speed recurrent neural network; outputting health evaluation results and confidence and performing mutual evidence consistency determination; dynamically strengthening updating or backtracking when mutual evidence or credibility is abnormal to form a stable online closed-loop evaluation mechanism. The application adopts the skip double-speed recurrent neural network driven by evidence, realizes online evaluation and adaptive closed-loop control of the health state of the new energy equipment, and has high stability, high sensitivity and strong engineering applicability.

Description

Technical Field

[0001] This invention relates to the field of condition monitoring of new energy equipment, and in particular to a method for assessing the health status of new energy equipment based on recurrent neural networks. Background Technology

[0002] With the large-scale grid connection of new energy equipment such as wind turbines, photovoltaic inverters, and energy storage converters, multi-source monitoring methods, including electrical, thermal, mechanical, and control systems, are generally deployed on the equipment side. Existing technologies typically extract features after unified sampling and preprocessing, and use statistical models or recurrent neural networks and other sequence models to conduct online assessments of health indices, status levels, or degradation trends. They also combine threshold judgments to achieve alarms and operation and maintenance decisions to meet the needs of continuous monitoring and status visualization.

[0003] However, existing technologies mostly use a single time-series backbone for synchronous updates. The handling of missing data, anomalies, and frequent changes in operating conditions often stops at supplementation, elimination, or simple confidence weighting, lacking quality mask and operating condition flag mechanisms that are linked to the evaluation and update behavior. At the same time, it is difficult to take into account both short-term fluctuation response and long-term degradation tracking, and it is easy to make erroneous updates and misjudgments when there is low-quality data or sudden changes in operating conditions. Furthermore, it lacks closed-loop self-correction mechanisms such as forced double updates, update density adjustment, and freeze / rollback based on output mutual verification consistency, resulting in insufficient online robustness.

[0004] Therefore, how to provide a health status assessment method for new energy equipment based on recurrent neural networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a health status assessment method for new energy equipment based on recurrent neural networks. This invention employs an evidence-driven step-by-step dual-speed recurrent neural network to achieve online assessment and adaptive closed-loop control of the health status of new energy equipment, which combines high stability, high sensitivity, and strong engineering applicability.

[0006] The health status assessment method for new energy equipment based on recurrent neural networks according to embodiments of the present invention includes the following steps: Acquire multi-source operation monitoring data of new energy equipment and perform time alignment, missing measurement marking and standardization processing to form a standardized input sequence, missing measurement and quality mask and operating condition indicator sequence; Based on standardized input sequences, missing and quality masks, and operating condition indicator sequences, we construct degradation trend evidence, operating condition switching evidence, and credibility evidence, which are then aggregated to form a three-level evidence input. The standardized input sequence and the three-level evidence input drive the gating system to generate a three-state update decision and simultaneously generate a risk budget. A dual-speed cyclic evaluation backbone containing fast and slow cyclic states is constructed based on a step-by-step recurrent neural network. Based on the three-state update decision and risk budget constraints, fast, slow, or dual updates are performed on the fast and slow cyclic states to obtain the dual-speed cyclic evaluation state. Based on the dual-speed cyclic evaluation state output health status evaluation results and evaluation confidence, a mutual verification consistency judgment result is generated for the health status evaluation results. When the mutual verification consistency judgment result meets the mutual verification triggering condition, double update is enforced and the update density is increased. When the credibility evidence meets the rollback triggering condition, a rollback action set is executed to freeze the fast loop state or slow loop state or rollback health state assessment result. The risk budget and the update threshold configuration of the evidence-driven gating are updated back to form an online closed loop.

[0007] Optionally, the generation of the standardized input sequence, the missing test and quality mask, and the operating condition flag sequence specifically includes: Acquire multi-source operation monitoring data of new energy equipment within the same evaluation cycle to form an original multi-source sampling set; Based on the original multi-source sampling set, a unified time axis and a unified sampling interval are constructed. Resampling and time alignment are performed on each type of multi-source operation monitoring data according to the data source to obtain a time-aligned multi-source sequence set. Perform defect marking on the time-aligned multi-source sequence set, generate defect and quality masks for each time point and each data source, and align and bind the time-aligned multi-source sequence set with the defect and quality masks according to a unified time axis to obtain a mask-bound sequence set. Anomaly removal and amplitude clipping are performed on the masked binding sequence set. The feasible value range is generated by the intersection of the quantile threshold and the physical boundary threshold. Sampling records that exceed the feasible value range are written with an anomaly flag and replaced with boundary values ​​to obtain a cleaned and aligned sequence set. The cleaned and aligned sequence set is normalized to obtain a normalized input sequence; Extract the operating condition indicator sequence from the control-side monitoring data and operation mode records, and merge it with the standardized input sequence and the missing measurement and quality mask according to a unified time axis.

[0008] Optionally, the generation of the third-level evidence input specifically includes: The system receives a standardized input sequence, a defect and quality mask, and a condition flag sequence. Based on a unified time axis, the standardized input sequence is segmented according to a preset evaluation window to form an evaluation window sequence set. Each evaluation window is bound to the defect and quality mask and the condition flag sequence within the corresponding time period to obtain a window-bound sequence set. Long-term statistical and trend extraction processing is performed on standardized input sequences within the same evaluation window on a window-bound sequence set. The long-term statistical results are combined with the trend extraction results to generate a degradation trend evidence sequence. On the window-bound sequence set, state statistics and state transition detection are performed on the working condition flag sequence within the same evaluation window. The state statistics results and state transition detection results are combined to generate a working condition switching evidence sequence. Data quality assessment is performed on standardized input sequences within the same assessment window based on missing tests and quality masks on a window-bound sequence set, and the data quality assessment results are mapped to a credible evidence sequence. Time alignment and scale normalization were performed on the degradation trend evidence sequence, the working condition switching evidence sequence, and the credibility evidence sequence. The aligned and normalized evidence of degradation trends, evidence of operating condition switching, and evidence of credibility are spliced ​​together in the order of the evaluation window to form a three-level evidence input that is consistent with the time of the standardized input sequence.

[0009] Optionally, the generation of the three-state update decision and risk budget specifically includes: It receives standardized input sequences and Level 3 evidence inputs, drives the gating system with input evidence, aligns and binds Level 3 evidence inputs with standardized input sequences based on a unified timeline, and generates a set of gating window inputs according to the evaluation window. Gated feature fusion is performed on the condition switching evidence and credibility evidence in the gated window input set to generate a gated fast-state evidence score sequence; Gated feature extraction is performed on the degradation trend evidence in the gated window input set to generate a gated slow-state evidence score sequence; A set of three-state update decision quantities is generated based on the gated fast-state evidence score sequence and the gated slow-state evidence score sequence, including fast update decision quantity, slow update decision quantity and double update decision quantity, and bound to the gated window input set to obtain the three-state decision binding set; Risk budget is generated by weighted aggregation and interval mapping based on the three-state determination binding set; A segmented threshold mapping rule is used to map the risk budget sequence to update the threshold configuration and update the density configuration; The update threshold configuration and update density configuration are applied to the three-state decision binding set. Based on the comparison results of the gated fast state evidence score and the fast update threshold, the comparison results of the gated slow state evidence score and the slow update threshold, and the results of the dual update density constraint, a three-state update decision is generated.

[0010] Optionally, the generation of the dual-speed cyclic evaluation state specifically includes: It receives three-state update decisions, risk budgets, update threshold configurations, and update density configurations, and constructs a dual-speed cyclic evaluation backbone, which consists of fast-state cyclic states and slow-state cyclic states, maintaining state synchronization under a unified time axis. The fast-state and slow-state cycles are initialized with initial update frequency and initial state maintenance strategy based on update threshold configuration and update density configuration, respectively. During operation, the three-state update decision controls the update of the corresponding cycle state, and the risk budget dynamically adjusts the update frequency, update trigger threshold and number of consecutive updates. Based on the standardized input sequence and the missing test and quality mask, the standardized input sequence in the current evaluation window is input into the dual-speed cyclic evaluation backbone. The input is filtered for validity based on the missing test and quality mask to form a valid input sequence for this round of cyclic update. When the three-state update decision instruction is fast update, under the condition of keeping the slow-state cyclic state unchanged, only the fast-state cyclic state is updated. This includes performing state recursion on the fast-state cyclic state based on the effective input sequence to generate the updated fast-state cyclic state, which is then combined with the unupdated slow-state cyclic state to form a phased dual-speed cyclic evaluation state. When the three-state update decision instruction is slow update, under the condition of keeping the fast-state cyclic state unchanged, only the slow-state cyclic state is updated. This includes performing state recursion on the slow-state cyclic state based on the effective input sequence and generating the updated slow-state cyclic state, which is then combined with the updated fast-state cyclic state to form a phased dual-speed cyclic evaluation state. When the three-state update decision instruction is double update, the cyclic state update is performed on both the fast-state cyclic state and the slow-state cyclic state simultaneously within the same evaluation window, generating the updated fast-state cyclic state and the updated slow-state cyclic state, forming a dual-speed cyclic evaluation state. During fast update, slow update, or dual update, the frequency and magnitude of cyclic state updates are constrained based on the risk budget, update threshold configuration, and update density configuration. When the update density configuration corresponding to the risk budget indicates an increase in update density, the number of cyclic state updates is increased within the continuous evaluation window. When the update threshold configuration corresponding to the risk budget indicates a tightening of the update threshold, the triggering conditions for cyclic state updates are restricted. Write the updated fast-state and slow-state cyclic states into the dual-speed cyclic evaluation state sequence in chronological order.

[0011] Optionally, the generation of the mutual verification consistency determination result specifically includes: Receive the dual-speed cyclic evaluation state sequence, the three-state update decision sequence, the risk budget sequence, the update threshold configuration and the update density configuration, perform window aggregation on the dual-speed cyclic evaluation state sequence according to the evaluation window, extract the fast-state cyclic state window representation and the slow-state cyclic state window representation corresponding to each evaluation window, and obtain the dual-state window representation set. Based on the dual-state window representation set, a health status assessment output mapping link is constructed. For each assessment window, the fast-state cyclic window representation and the slow-state cyclic window representation are sequentially concatenated and normalized mapping is performed to generate a sequence of health status assessment results. An evaluation confidence generation link is constructed based on a dual-state window representation set and a confidence evidence sequence. For each evaluation window, the short-term fluctuation intensity of the fast-state cyclic state window representation and the trend stability of the slow-state cyclic state window representation are calculated. The evaluation confidence sequence is generated by combining the confidence sampling ratio, outlier ratio and consecutive missing length of the confidence evidence sequence in the corresponding evaluation window. A consistency constraint mapping is performed on the health status assessment result sequence. A health index trend sequence is generated based on the direction and magnitude of change of the health index between adjacent assessment windows. A health level migration sequence is generated based on the migration direction and magnitude of the health level between adjacent assessment windows. The sequence is aligned and bound with the assessment confidence sequence to obtain the output consistency element set. The mutual verification consistency judgment result is generated based on the set of output consistency elements, including the consistency quantity of health index trend, the consistency quantity of health level migration, and the confidence consistency quantity. The mutual verification consistency judgment result is used to perform mutual verification consistency aggregation, generate mutual verification consistency score sequence, and align and bind it with the three-state update decision sequence and risk budget sequence to form mutual verification backflow binding set.

[0012] Optionally, the generation of the online closed loop specifically includes: Receive the mutual verification feedback binding set, extract the mutual verification consistency score sequence, three-state update decision sequence and risk budget sequence, and simultaneously extract the data quality assessment results from the update threshold configuration, update density configuration and credibility evidence sequence, and gather them to form a closed-loop judgment input set; Generate a mutual verification trigger decision quantity on the closed-loop decision input set and determine the mutual verification trigger condition accordingly to obtain the mutual verification trigger flag sequence; When the mutual verification triggering flag sequence indicates that the mutual verification triggering condition is met, the three-state update decision of the corresponding evaluation window is overwritten as a double update, and the update density is increased and the threshold is tightened in the corresponding evaluation window and its subsequent preset continuous evaluation windows to generate a mutual verification enhancement configuration set. Generate a back-off trigger decision quantity on the closed-loop decision input set and determine the back-off trigger condition accordingly to obtain the back-off trigger flag sequence; When the rollback trigger flag sequence indicates that the rollback trigger condition is met, the rollback action set is executed and the rollback execution result set is generated. The mutual verification enhancement configuration set and the rollback execution result set are fed back to update the risk budget sequence and the update threshold configuration and update density configuration of the evidence-driven gating device. They are then aligned and bound to the gating window input set of the next evaluation window. In subsequent evaluation windows, they are used to generate three-state update decisions and update the dual-speed cyclic evaluation state, forming an online closed loop.

[0013] The beneficial effects of this invention are: This invention introduces a dual-speed cyclic evaluation backbone within a step-by-step recurrent neural network framework, where fast and slow cyclic states evolve in parallel. This structurally decouples the perception of abnormal fluctuations on short timescales from the characterization of degradation trends on long timescales, while simultaneously ensuring coordinated decision-making. This avoids the performance conflicts arising from a single time-series backbone balancing sensitivity and stability under complex operating conditions. By combining three levels of evidence input—degradation trend evidence, operating condition switching evidence, and credibility evidence—the evaluation update behavior is no longer driven solely by time steps but by evidence-driven three-state update decision control. This enables the equipment to effectively reduce invalid updates and noise accumulation during stable operation and to promptly increase response density during periods of sudden changes in operating conditions or increased risk, achieving a balance between evaluation accuracy and online efficiency.

[0014] Meanwhile, this invention explicitly models the inherent consistency relationship between health status assessment results, assessment confidence, and dual-speed cyclic states through mutual verification consistency judgment and online closed-loop mechanism. Based on this, it triggers forced dual updates, tightens update thresholds, and increases update density, ensuring the reliability of assessments during critical degradation stages. When data quality deteriorates or credibility is insufficient, it can automatically execute freeze and rollback, and simultaneously backflow to correct risk budgets and gating configurations, preventing the accumulation and spread of erroneous information in the cyclic state. This significantly improves the stability, controllability, and engineering applicability of online health status assessment for new energy equipment under complex operating conditions and incomplete data scenarios. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the health status assessment method for new energy equipment based on recurrent neural networks proposed in this invention; Figure 2 This is a schematic diagram illustrating the evidence-driven three-state update decision-making and risk budget generation of the health status assessment method for new energy equipment based on recurrent neural networks proposed in this invention. Figure 3This diagram illustrates the dual-speed step-by-step cyclic evaluation backbone and online closed-loop state evolution of the health status assessment method for new energy equipment based on recurrent neural networks proposed in this invention. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0017] refer to Figures 1-3 A health status assessment method for new energy equipment based on recurrent neural networks includes the following steps: Acquire multi-source operation monitoring data of new energy equipment and perform time alignment, missing measurement marking and standardization processing to form a standardized input sequence, missing measurement and quality mask and operating condition indicator sequence; Based on standardized input sequences, missing and quality masks, and operating condition indicator sequences, we construct degradation trend evidence, operating condition switching evidence, and credibility evidence, which are then aggregated to form a three-level evidence input. The standardized input sequence and the three-level evidence input drive the gating system to generate a three-state update decision and simultaneously generate a risk budget. A dual-speed cyclic evaluation backbone containing fast and slow cyclic states is constructed based on a step-by-step recurrent neural network. Based on the three-state update decision and risk budget constraints, fast, slow, or dual updates are performed on the fast and slow cyclic states to obtain the dual-speed cyclic evaluation state. Based on the dual-speed cyclic evaluation state output health status evaluation results and evaluation confidence, a mutual verification consistency judgment result is generated for the health status evaluation results. When the mutual verification consistency judgment result meets the mutual verification triggering condition, double update is enforced and the update density is increased. When the credibility evidence meets the rollback triggering condition, a rollback action set is executed to freeze the fast loop state or slow loop state or rollback health state assessment result. The risk budget and the update threshold configuration of the evidence-driven gating are updated back to form an online closed loop.

[0018] In this embodiment, the generation of the standardized input sequence, the missing test and quality mask, and the operating condition flag sequence specifically includes: Acquire multi-source operation monitoring data of new energy equipment within the same evaluation cycle to form an original multi-source sampling set; The multi-source operation monitoring data includes electrical side monitoring data, thermal side monitoring data, mechanical side monitoring data, and control side monitoring data, and is written with sampling timestamps and data source identifiers; Based on the original multi-source sampling set, a unified time axis and a unified sampling interval are constructed. Resampling and time alignment are performed on each type of multi-source operation monitoring data according to the data source to obtain a time-aligned multi-source sequence set. Perform defect marking on the time-aligned multi-source sequence set, generate defect and quality masks for each time point and each data source, and align and bind the time-aligned multi-source sequence set with the defect and quality masks according to a unified time axis to obtain a mask-bound sequence set. The missing test and quality mask includes a missing test flag, an outlier flag, and a reliable sampling flag. The missing test flag indicates whether there is original sampling or interpolated sampling at the target time point. If it exists, it is set to valid; if it does not exist, it is set to missing test. Anomaly removal and amplitude clipping are performed on the masked binding sequence set. The feasible value range is generated by the intersection of the quantile threshold and the physical boundary threshold. Sampling records that exceed the feasible value range are written with an anomaly flag and replaced with boundary values ​​to obtain a cleaned and aligned sequence set. The cleaned and aligned sequence set is normalized to obtain a normalized input sequence; The standardization process calculates the baseline mean and baseline standard deviation for each monitoring quantity from each data source. The standardization result is obtained by subtracting the baseline mean of the corresponding monitoring quantity from the monitoring quantity value at each time point and then dividing by the baseline standard deviation of the corresponding monitoring quantity. The baseline mean and baseline standard deviation are retained as the standardization configuration. Extract the operating condition indicator sequence from the control side monitoring data and operation mode records, and merge it with the standardized input sequence and missing and quality masks according to a unified time axis; The operating condition flag sequence includes start / stop flag, grid connection flag, power limit flag, and operating mode flag.

[0019] In this embodiment, the generation of the three-level evidence input specifically includes: The system receives a standardized input sequence, a defect and quality mask, and a condition flag sequence. Based on a unified time axis, the standardized input sequence is segmented according to a preset evaluation window to form an evaluation window sequence set. Each evaluation window is bound to the defect and quality mask and the condition flag sequence within the corresponding time period to obtain a window-bound sequence set. Long-term statistical and trend extraction processing is performed on standardized input sequences within the same evaluation window on a window-bound sequence set. The long-term statistical results are combined with the trend extraction results to generate a degradation trend evidence sequence. The long-term statistics include the window mean, window median, window variance, and window extreme value range; the trend extraction includes calculating the direction and magnitude of change of corresponding statistics between adjacent evaluation windows. On the window-bound sequence set, state statistics and state transition detection are performed on the working condition flag sequence within the same evaluation window. The state statistics results and state transition detection results are combined to generate a working condition switching evidence sequence. The status statistics include the duration and percentage of the operating condition indicator within the window, and the status transition detection includes the number of times the operating condition indicator changes at adjacent time points, the location of the change, and the density of the change. Data quality assessment is performed on standardized input sequences within the same assessment window based on missing tests and quality masks on a window-bound sequence set, and the data quality assessment results are mapped to a credible evidence sequence. The data quality assessment includes the missing data ratio, outlier ratio, consecutive missing data length, and reliable sampling ratio. Time alignment and scale normalization were performed on the degradation trend evidence sequence, the working condition switching evidence sequence, and the credibility evidence sequence. The aligned and normalized evidence of degradation trends, evidence of operating condition switching, and evidence of credibility are spliced ​​together in the order of the evaluation window to form a three-level evidence input that is consistent with the time of the standardized input sequence.

[0020] In this embodiment, the generation of the three-state update decision and risk budget specifically includes: It receives standardized input sequences and Level 3 evidence inputs, drives the gating system with input evidence, aligns and binds Level 3 evidence inputs with standardized input sequences based on a unified timeline, and generates a set of gating window inputs according to the evaluation window. Gated feature fusion is performed on the condition switching evidence and credibility evidence in the gated window input set to generate a gated fast-state evidence score sequence; The gated feature fusion includes weighted aggregation of the state transition detection results of the working condition switching evidence and the data quality assessment results of the credibility evidence, and interval pruning of the weighted aggregation results to obtain the gated fast-state evidence score for fast update determination. The gated fast-state evidence score is determined by the sum of the first gated fusion weight coefficient multiplied by the window value of the working condition switching evidence after scale normalization and the second gated fusion weight coefficient multiplied by the window value of the credibility evidence after scale normalization. Gated feature extraction is performed on the degradation trend evidence in the gated window input set to generate a gated slow-state evidence score sequence; The gated feature extraction includes weighted aggregation of long-term statistical results and trend extraction results of degradation trend evidence, and interval pruning of the weighted aggregation results to obtain a gated slow-state evidence score for slow update determination. The gated slow-state evidence score is determined by the sum of the first slow-state extraction weight coefficient multiplied by the scale-normalized window value of the degradation trend evidence and the second slow-state extraction weight coefficient multiplied by the scale-normalized window value of the change amplitude of degradation trend evidence between adjacent evaluation windows. A set of three-state update decision quantities is generated based on the gated fast-state evidence score sequence and the gated slow-state evidence score sequence, including fast update decision quantity, slow update decision quantity and double update decision quantity, and bound to the gated window input set to obtain the three-state decision binding set; The fast update determination is determined by the difference between the gated fast state evidence score and the gated slow state evidence score within the corresponding evaluation window, as well as the absolute value of the gated fast state evidence score. The slow update determination is determined by the difference between the gated slow state evidence score and the gated fast state evidence score within the corresponding evaluation window, as well as the absolute value of the gated slow state evidence score. The double update determination is determined by the minimum value of the gated fast state evidence score and the gated slow state evidence score, as well as the combined change magnitude of the two. Risk budget is generated by weighted aggregation and interval mapping based on the three-state determination binding set; The input elements for generating the risk budget are limited to evidence of degradation trends, evidence of operating condition switching, and evidence of credibility. Weighted aggregation and interval mapping are performed on the input elements within the same evaluation window to obtain a risk budget sequence corresponding to the evaluation window. The risk budget is determined by the sum of the inverse vectors of the first risk weight coefficient multiplied by the window value of the degradation trend evidence after scaling, the second risk weight coefficient multiplied by the window value of the operating condition switching evidence after scaling, and the third risk weight coefficient multiplied by the window value of the window value of the credibility evidence after scaling. A segmented threshold mapping rule is used to map the risk budget sequence to update the threshold configuration and update the density configuration; The update threshold configuration includes a fast update threshold and a slow update threshold, and the update density configuration includes a dual update density and a single update density. The fast update threshold is determined by subtracting the product of the fast update threshold mapping coefficient and the risk budget from the fast update threshold baseline configuration. The slow update threshold is determined by subtracting the product of the slow update threshold mapping coefficient and the risk budget from the slow update threshold baseline configuration. The dual update density is determined by adding the product of the dual update density mapping coefficient and the risk budget to the dual update density baseline configuration. The update threshold configuration and update density configuration are applied to the three-state decision binding set. Based on the comparison results of the gated fast state evidence score and the fast update threshold, the comparison results of the gated slow state evidence score and the slow update threshold, and the results of the double update density constraint, a three-state update decision is generated. The three-state update decision is limited to three mutually exclusive values: fast update, slow update, and double update.

[0021] In this embodiment, the generation of the dual-speed cycle evaluation state specifically includes: It receives three-state update decisions, risk budgets, update threshold configurations, and update density configurations, and constructs a dual-speed cyclic evaluation backbone, which consists of fast-state cyclic states and slow-state cyclic states, maintaining state synchronization under a unified time axis. The fast-state cycle state consists of a first-state update structure that takes the standardized input sequence within the assessment window as direct input, representing health fluctuations and abnormal changes within a short timescale. The slow-state cycle state consists of a second-state update structure that takes the standardized input sequence after aggregation of multiple assessment windows and evidence of degradation trends as input, representing health evolution and degradation trends within a long timescale. The fast-state and slow-state cycles are initialized with initial update frequency and initial state maintenance strategy based on update threshold configuration and update density configuration, respectively. During operation, the three-state update decision controls the update of the corresponding cycle state, and the risk budget dynamically adjusts the update frequency, update trigger threshold and number of consecutive updates. Based on the standardized input sequence and the missing test and quality mask, the standardized input sequence in the current evaluation window is input into the dual-speed cyclic evaluation backbone. The input is filtered for validity based on the missing test and quality mask to form a valid input sequence for this round of cyclic update. When the three-state update decision instruction is fast update, under the condition of keeping the slow-state cyclic state unchanged, only the fast-state cyclic state is updated. This includes performing state recursion on the fast-state cyclic state based on the effective input sequence to generate the updated fast-state cyclic state, which is then combined with the unupdated slow-state cyclic state to form a phased dual-speed cyclic evaluation state. When the three-state update decision instruction is slow update, under the condition of keeping the fast-state cyclic state unchanged, only the slow-state cyclic state is updated. This includes performing state recursion on the slow-state cyclic state based on the effective input sequence and generating the updated slow-state cyclic state, which is then combined with the updated fast-state cyclic state to form a phased dual-speed cyclic evaluation state. When the three-state update decision instruction is double update, the cyclic state update is performed on both the fast-state cyclic state and the slow-state cyclic state simultaneously within the same evaluation window, generating the updated fast-state cyclic state and the updated slow-state cyclic state, forming a dual-speed cyclic evaluation state. During fast update, slow update, or dual update, the frequency and magnitude of cyclic state updates are constrained based on the risk budget, update threshold configuration, and update density configuration. When the update density configuration corresponding to the risk budget indicates an increase in update density, the number of cyclic state updates is increased within the continuous evaluation window. When the update threshold configuration corresponding to the risk budget indicates a tightening of the update threshold, the triggering conditions for cyclic state updates are restricted. Write the updated fast-state and slow-state cyclic states into the dual-speed cyclic evaluation state sequence in chronological order.

[0022] In this embodiment, the generation of the mutual verification consistency determination result specifically includes: Receive the dual-speed cyclic evaluation state sequence, the three-state update decision sequence, the risk budget sequence, the update threshold configuration and the update density configuration, perform window aggregation on the dual-speed cyclic evaluation state sequence according to the evaluation window, extract the fast-state cyclic state window representation and the slow-state cyclic state window representation corresponding to each evaluation window, and obtain the dual-state window representation set. Based on the dual-state window representation set, a health status assessment output mapping link is constructed. For each assessment window, the fast-state cyclic window representation and the slow-state cyclic window representation are sequentially concatenated and normalized mapping is performed to generate a sequence of health status assessment results. The health status assessment result sequence includes two types of output items: health index and health level. The health index is a continuous value output item, and the health level is a discrete graded output item, and they are bound to each other within the same assessment window. An evaluation confidence generation link is constructed based on a dual-state window representation set and a confidence evidence sequence. For each evaluation window, the short-term fluctuation intensity of the fast-state cyclic state window representation and the trend stability of the slow-state cyclic state window representation are calculated. The evaluation confidence sequence is generated by combining the confidence sampling ratio, outlier ratio and consecutive missing length of the confidence evidence sequence in the corresponding evaluation window. A consistency constraint mapping is performed on the health status assessment result sequence. A health index trend sequence is generated based on the direction and magnitude of change of the health index between adjacent assessment windows. A health level migration sequence is generated based on the migration direction and magnitude of the health level between adjacent assessment windows. The sequence is aligned and bound with the assessment confidence sequence to obtain the output consistency element set. The mutual verification consistency judgment result is generated based on the set of output consistency elements, including the consistency quantity of health index trend, the consistency quantity of health level migration, and the confidence consistency quantity. Among them, the health index trend consistency measure describes the consistency relationship between the health index trend sequence and the trend stability represented by the slow-state cyclic state window; the health level migration consistency measure describes the consistency relationship between the health level migration sequence and the short-term fluctuation intensity represented by the fast-state cyclic state window; and the confidence consistency measure describes the degree of support of the confidence sequence for the health index trend consistency measure and the health level migration consistency measure. Based on the mutual verification consistency judgment results, perform mutual verification consistency aggregation to generate mutual verification consistency score sequence, which is aligned and bound with the three-state update decision sequence and risk budget sequence to form mutual verification backflow binding set; The mutual verification consistency score is obtained by weighting the consistency of the health index trend and the consistency of the health level migration, and modulating it with the confidence consistency.

[0023] In this embodiment, the generation of the online closed loop specifically includes: Receive the mutual verification feedback binding set, extract the mutual verification consistency score sequence, three-state update decision sequence and risk budget sequence, and simultaneously extract the data quality assessment results from the update threshold configuration, update density configuration and credibility evidence sequence, and gather them to form a closed-loop judgment input set; Generate a mutual verification trigger decision quantity on the closed-loop decision input set and determine the mutual verification trigger condition accordingly to obtain the mutual verification trigger flag sequence; The mutual verification trigger determination quantity is obtained by weighted aggregation of the inverse vector of the mutual verification consistency score, the decrease of the mutual verification consistency score between adjacent evaluation windows, and the inverse vector of the evaluation confidence. The mutual verification trigger condition is limited to satisfying one of the following three conditions: the mutual verification trigger determination quantity is not lower than the mutual verification trigger threshold, the mutual verification consistency score is not higher than the mutual verification consistency lower limit threshold, or the decrease of the mutual verification consistency score between adjacent windows is not lower than the decrease trigger threshold. When the mutual verification triggering flag sequence indicates that the mutual verification triggering condition is met, the three-state update decision of the corresponding evaluation window is overwritten as a double update, and the update density is increased and the threshold is tightened in the corresponding evaluation window and its subsequent preset continuous evaluation windows to generate a mutual verification enhancement configuration set. The update density increase is limited to increasing the dual update density in the update density configuration to the upper limit of dual update density and tightening the single update density to the lower limit of single update density. The threshold tightening is limited to tightening the fast update threshold and slow update threshold in the update threshold configuration window by window according to the preset threshold tightening step. Generate a back-off trigger decision quantity on the closed-loop decision input set and determine the back-off trigger condition accordingly to obtain the back-off trigger flag sequence; The rollback trigger determination quantity is obtained by weighting and summing the missing test ratio, outlier ratio, continuous missing test length after window length normalization, and the inverse vector of the reliable sampling ratio. The rollback trigger condition is limited to satisfying one of the following three conditions: the rollback trigger determination quantity is not lower than the rollback trigger threshold, the reliable sampling ratio is not higher than the reliable lower threshold, or the continuous missing test length is not lower than the continuous missing test upper threshold. When the rollback trigger flag sequence indicates that the rollback trigger condition is met, the rollback action set is executed and the rollback execution result set is generated. The rollback action set is limited to at least one of three mutually exclusive actions: freezing fast loop state, freezing slow loop state, and rolling back health status assessment results. Freezing fast loop state means keeping the fast loop state unchanged within a preset continuous assessment window and prohibiting fast update and double update from acting on the fast loop state. Freezing slow loop state means keeping the slow loop state unchanged within a preset continuous assessment window and prohibiting slow update and double update from acting on the slow loop state. Rolling back health status assessment results means rolling back the output of the health status assessment result sequence in the corresponding assessment window to the health status assessment result of the previous assessment window and simultaneously rolling back the assessment confidence. The mutual verification enhancement configuration set and the rollback execution result set are fed back to update the risk budget sequence and the update threshold configuration and update density configuration of the evidence-driven gating device. They are then aligned and bound to the gating window input set of the next evaluation window. In subsequent evaluation windows, they are used to generate three-state update decisions and update the dual-speed cyclic evaluation state, forming an online closed loop.

[0024] Example 1: To verify the feasibility of this invention in practice, it was applied to a health assessment scenario of a cluster of grid-connected new energy equipment operating in a coastal region. In this scenario, the new energy equipment operates in a high-humidity, high-salt-spray, and frequently fluctuating power environment. Its operating status is affected by multiple factors, including weather changes, grid connection dispatch commands, and equipment aging. The health status exhibits characteristics of frequent short-term fluctuations, slow long-term degradation, and overlapping effects. Existing assessment methods typically analyze based on fixed time scales or single models, making it difficult to distinguish between operational disturbances and actual degradation in a timely manner. This often leads to problems such as large fluctuations in assessment results, frequent false alarms, or insensitivity to early degradation, introducing significant uncertainty into operation and maintenance decisions.

[0025] In this scenario, multi-source operational monitoring data, including electrical, thermal, mechanical, and operational condition data of the new energy equipment, are first collected. Time alignment, missing data marking, and standardization are then performed to form a standardized input sequence that can be directly used as model input. Simultaneously, missing data and quality masks, as well as operational condition indicator sequences reflecting data integrity and reliability, are generated. Based on this, degradation trend evidence, operational condition switching evidence, and credibility evidence are constructed according to the time-dimensional changes in operational data. This enables the model to simultaneously perceive the long-term direction of equipment operation changes, short-term operational condition disturbances, and data quality status. Subsequently, the above information is input into the evidence-driven gating system to dynamically generate a three-state update decision (fast update, slow update, or dual update), and a risk budget is simultaneously formed to reflect the comprehensive requirements for update frequency and stability in the current evaluation phase.

[0026] During the health status assessment process, a dual-speed cyclic assessment backbone based on a jump-step recurrent neural network simultaneously operates in fast-state and slow-state cyclic states. The fast-state cyclic state focuses on capturing short-term health changes caused by power fluctuations and operating condition switching, while the slow-state cyclic state continuously depicts the degradation trend of the equipment during long-term operation. When the operating environment is stable and the data is reliable, the assessment backbone automatically reduces the update frequency to maintain a stable state. When frequent changes in operating conditions or deviations in the health trend are detected, a corresponding update strategy is triggered through a three-state update decision to achieve rapid response to critical change stages. Based on the dual-speed cyclic assessment states, health status assessment results and assessment confidence levels are continuously output, and the consistency between short-term assessment results and long-term trends is verified through mutual verification consistency judgment.

[0027] In actual operation, when the mutual verification consistency judgment results confirm short-term anomalies and long-term trends, the assessment process automatically increases the update density, strengthening the ability to track potential risk stages. When credible evidence reflects a decline in monitoring data quality, the assessment process performs a rollback, freezing or reverting assessment results for some cyclical states to avoid misjudgments caused by noise or missing data, and feeding relevant information back for subsequent risk budgeting and update threshold adjustments. Through this online closed-loop mechanism, the assessment results can dynamically and adaptively adjust according to the operational status and data quality.

[0028] In practical operation, this method maintains the continuity and smoothness of assessment results during stable equipment operation phases, promptly reflects health status trends during load fluctuations or environmental changes, and significantly reduces assessment instability caused by operating condition disturbances or data anomalies. Compared with traditional single-scale assessment methods, this invention demonstrates significant improvements in the timeliness of health status response, the consistency of assessment results, and the ability to identify long-term degradation, effectively supporting continuous monitoring and risk warning of new energy equipment operation status.

[0029] Table 1. Overall Performance Comparison Results of Health Status Assessment Methods for New Energy Equipment

[0030] The accuracy of health status assessment shows that traditional statistical assessment methods based on fixed time windows achieve an accuracy of 86.2% under complex operating conditions. This is mainly due to their inability to distinguish between short-term operational disturbances and long-term degradation trends, making the assessment results susceptible to fluctuations. Introducing a single-scale recurrent neural network improves the accuracy to 89.5%, indicating that time-series modeling capabilities improve the assessment effect to some extent. However, due to the single time scale, it still struggles to simultaneously account for both rapid fluctuations and slow degradation. The method of this invention achieves an assessment accuracy of 94.8%, significantly higher than the comparative methods. This demonstrates that the dual-speed recurrent assessment backbone constructed based on a step-by-step recurrent neural network can simultaneously characterize fast-state changes and slow-state evolution, resulting in a more comprehensive and stable health status representation.

[0031] Regarding the number of evaluation periods in advance for early degradation identification, statistical evaluation methods can only identify degradation in advance by 2 evaluation periods, single-scale recurrent neural network methods improve this to 3 evaluation periods, while the method of this invention can identify degradation in advance by 6 evaluation periods. This difference stems from the continuous accumulation of degradation trend evidence in the slow-state cyclic state, ensuring that long-term subtle changes are not masked by short-term fluctuations. Furthermore, under the mutual evidence consistency triggering mechanism, relevant degradation signals can be strengthened and updated in the early stages, thus significantly improving the early identification capability.

[0032] The stability index of the evaluation results reflects the smoothness and resilience of the health assessment output under continuous operating conditions. The stability index of traditional statistical methods is 0.71, showing significant fluctuations during frequent changes in operating conditions; the single-scale recurrent neural network method improves this to 0.76, but instability still exists during periods of declining data quality. The method of this invention achieves a stability index of 0.89, primarily due to the dynamic constraints of risk budget on update frequency and update threshold, enabling the model to remain sensitive in high-risk phases and actively converge in low-risk phases, avoiding ineffective oscillations.

[0033] In terms of the number of false rollbacks, the statistical evaluation method experienced 18 rollbacks during the running cycle, the single-scale recurrent neural network method experienced 14 rollbacks, while the method of this invention only experienced 6 rollbacks. This indicates that the credibility evidence and mutual verification consistency judgment mechanism effectively reduces unnecessary rollbacks caused by missing tests and abnormal data. By freezing or rolling back the precise triggering conditions of the loop state, the rollback action is made more targeted, reducing interference with the normal evaluation process.

[0034] In terms of continuous assessment consistency score, the method of this invention achieved 0.86, significantly higher than the comparative methods. This indicates a stronger consistency relationship between the trend of health index changes, health level migration behavior, and assessment confidence, demonstrating the role of the mutual verification consistency aggregation mechanism in suppressing assessment drift and enhancing long-term credibility.

[0035] As can be seen from the data in Table 1, this invention achieves simultaneous improvements in multiple key dimensions such as accuracy, advance capability, stability, and consistency through the collaborative design of evidence-driven gating mechanism, three-state update decision, risk budget constraint, and dual-speed step-by-step cyclic evaluation backbone. This fully demonstrates the comprehensive advantages of this method in assessing the health status of new energy equipment under complex operating environments.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for assessing the health status of new energy equipment based on recurrent neural networks, characterized in that, The steps include the following: Acquire multi-source operation monitoring data of new energy equipment and perform time alignment, missing measurement marking and standardization processing to form a standardized input sequence, missing measurement and quality mask and operating condition indicator sequence; Based on standardized input sequences, missing and quality masks, and operating condition indicator sequences, we construct degradation trend evidence, operating condition switching evidence, and credibility evidence, which are then aggregated to form a three-level evidence input. The standardized input sequence and the three-level evidence input drive the gating system to generate a three-state update decision and simultaneously generate a risk budget. A dual-speed cyclic evaluation backbone containing fast and slow cyclic states is constructed based on a step-by-step recurrent neural network. Based on the three-state update decision and risk budget constraints, fast, slow, or dual updates are performed on the fast and slow cyclic states to obtain the dual-speed cyclic evaluation state. Based on the dual-speed cyclic evaluation state output health status assessment results and assessment confidence, a mutual verification consistency judgment result is generated for the health status assessment results. When the mutual verification consistency judgment result meets the mutual verification triggering condition, double update is enforced and the update density is increased. When the credibility evidence meets the rollback triggering condition, a rollback action set is executed to freeze the fast loop state or slow loop state or rollback health state assessment result. The risk budget and the update threshold configuration of the evidence-driven gating are updated back to form an online closed loop.

2. The method for assessing the health status of new energy equipment based on recurrent neural networks according to claim 1, characterized in that, The generation of the standardized input sequence, the missing and quality mask, and the operating condition flag sequence specifically includes: Acquire multi-source operation monitoring data of new energy equipment within the same evaluation cycle to form an original multi-source sampling set; Based on the original multi-source sampling set, a unified time axis and a unified sampling interval are constructed. Resampling and time alignment are performed on each type of multi-source operation monitoring data according to the data source to obtain a time-aligned multi-source sequence set. Perform defect marking on the time-aligned multi-source sequence set, generate defect and quality masks for each time point and each data source, and align and bind the time-aligned multi-source sequence set with the defect and quality masks according to a unified time axis to obtain a mask-bound sequence set. Anomaly removal and amplitude clipping are performed on the masked binding sequence set. The feasible value range is generated by the intersection of the quantile threshold and the physical boundary threshold. Sampling records that exceed the feasible value range are written with an anomaly flag and replaced with boundary values ​​to obtain a cleaned and aligned sequence set. The cleaned and aligned sequence set is normalized to obtain a normalized input sequence; Extract the operating condition indicator sequence from the control-side monitoring data and operation mode records, and merge it with the standardized input sequence and the missing measurement and quality mask according to a unified time axis.

3. The method for assessing the health status of new energy equipment based on recurrent neural networks according to claim 1, characterized in that, The generation of the three-level evidence input specifically includes: The system receives a standardized input sequence, a defect and quality mask, and a condition flag sequence. Based on a unified time axis, the standardized input sequence is segmented according to a preset evaluation window to form an evaluation window sequence set. Each evaluation window is bound to the defect and quality mask and the condition flag sequence within the corresponding time period to obtain a window-bound sequence set. Long-term statistical and trend extraction processing is performed on standardized input sequences within the same evaluation window on a window-bound sequence set. The long-term statistical results are combined with the trend extraction results to generate a degradation trend evidence sequence. On the window-bound sequence set, state statistics and state transition detection are performed on the working condition flag sequence within the same evaluation window. The state statistics results and state transition detection results are combined to generate a working condition switching evidence sequence. Data quality assessment is performed on standardized input sequences within the same assessment window based on missing tests and quality masks on a window-bound sequence set, and the data quality assessment results are mapped to a credible evidence sequence. Time alignment and scale normalization were performed on the degradation trend evidence sequence, the working condition switching evidence sequence, and the credibility evidence sequence. The aligned and normalized evidence of degradation trends, evidence of operating condition switching, and evidence of credibility are spliced ​​together in the order of the evaluation window to form a three-level evidence input that is consistent with the time of the standardized input sequence.

4. The method for assessing the health status of new energy equipment based on recurrent neural networks according to claim 1, characterized in that, The generation of the three-state update decision and risk budget specifically includes: It receives standardized input sequences and Level 3 evidence inputs, drives the gating system with input evidence, aligns and binds Level 3 evidence inputs with standardized input sequences based on a unified timeline, and generates a set of gating window inputs according to the evaluation window. Gated feature fusion is performed on the condition switching evidence and credibility evidence in the gated window input set to generate a gated fast-state evidence score sequence; Gated feature extraction is performed on the degradation trend evidence in the gated window input set to generate a gated slow-state evidence score sequence; A set of three-state update decision quantities is generated based on the gated fast-state evidence score sequence and the gated slow-state evidence score sequence, including fast update decision quantity, slow update decision quantity and double update decision quantity, and bound to the gated window input set to obtain the three-state decision binding set; Risk budget is generated by weighted aggregation and interval mapping based on the three-state determination binding set; A segmented threshold mapping rule is used to map the risk budget sequence to update the threshold configuration and update the density configuration; The update threshold configuration and update density configuration are applied to the three-state decision binding set. Based on the comparison results of the gated fast state evidence score and the fast update threshold, the comparison results of the gated slow state evidence score and the slow update threshold, and the results of the dual update density constraint, a three-state update decision is generated.

5. The method for assessing the health status of new energy equipment based on recurrent neural networks according to claim 1, characterized in that, The generation of the dual-speed cyclic evaluation state specifically includes: It receives three-state update decisions, risk budgets, update threshold configurations, and update density configurations, and constructs a dual-speed cyclic evaluation backbone, which consists of fast-state cyclic states and slow-state cyclic states, maintaining state synchronization under a unified time axis. The fast-state and slow-state cycles are initialized with initial update frequency and initial state maintenance strategy based on update threshold configuration and update density configuration, respectively. During operation, the three-state update decision controls the update of the corresponding cycle state, and the risk budget dynamically adjusts the update frequency, update trigger threshold and number of consecutive updates. Based on the standardized input sequence and the missing test and quality mask, the standardized input sequence in the current evaluation window is input into the dual-speed cyclic evaluation backbone. The input is filtered for validity based on the missing test and quality mask to form a valid input sequence for this round of cyclic update. When the three-state update decision instruction is fast update, under the condition of keeping the slow-state cyclic state unchanged, only the fast-state cyclic state is updated. This includes performing state recursion on the fast-state cyclic state based on the effective input sequence to generate the updated fast-state cyclic state, which is then combined with the unupdated slow-state cyclic state to form a phased dual-speed cyclic evaluation state. When the three-state update decision instruction is slow update, under the condition of keeping the fast-state cyclic state unchanged, only the slow-state cyclic state is updated. This includes performing state recursion on the slow-state cyclic state based on the effective input sequence and generating the updated slow-state cyclic state, which is then combined with the updated fast-state cyclic state to form a phased dual-speed cyclic evaluation state. When the three-state update decision instruction is double update, the cyclic state update is performed on both the fast-state cyclic state and the slow-state cyclic state simultaneously within the same evaluation window, generating the updated fast-state cyclic state and the updated slow-state cyclic state, forming a dual-speed cyclic evaluation state. During fast update, slow update, or dual update, the frequency and magnitude of cyclic state updates are constrained based on the risk budget, update threshold configuration, and update density configuration. When the update density configuration corresponding to the risk budget indicates an increase in update density, the number of cyclic state updates is increased within the continuous evaluation window. When the update threshold configuration corresponding to the risk budget indicates a tightening of the update threshold, the triggering conditions for cyclic state updates are restricted. Write the updated fast-state and slow-state cyclic states into the dual-speed cyclic evaluation state sequence in chronological order.

6. The method for assessing the health status of new energy equipment based on recurrent neural networks according to claim 1, characterized in that, The generation of the mutual verification consistency determination result specifically includes: Receive the dual-speed cyclic evaluation state sequence, the three-state update decision sequence, the risk budget sequence, the update threshold configuration and the update density configuration, perform window aggregation on the dual-speed cyclic evaluation state sequence according to the evaluation window, extract the fast-state cyclic state window representation and the slow-state cyclic state window representation corresponding to each evaluation window, and obtain the dual-state window representation set. Based on the dual-state window representation set, a health status assessment output mapping link is constructed. For each assessment window, the fast-state cyclic window representation and the slow-state cyclic window representation are sequentially concatenated and normalized mapping is performed to generate a sequence of health status assessment results. An evaluation confidence generation link is constructed based on a dual-state window representation set and a confidence evidence sequence. For each evaluation window, the short-term fluctuation intensity of the fast-state cyclic state window representation and the trend stability of the slow-state cyclic state window representation are calculated. The evaluation confidence sequence is generated by combining the confidence sampling ratio, outlier ratio and consecutive missing length of the confidence evidence sequence in the corresponding evaluation window. A consistency constraint mapping is performed on the health status assessment result sequence. A health index trend sequence is generated based on the direction and magnitude of change of the health index between adjacent assessment windows. A health level migration sequence is generated based on the migration direction and magnitude of the health level between adjacent assessment windows. The sequence is aligned and bound with the assessment confidence sequence to obtain the output consistency element set. The mutual verification consistency judgment result is generated based on the set of output consistency elements, including the consistency quantity of health index trend, the consistency quantity of health level migration, and the confidence consistency quantity. The mutual verification consistency judgment result is used to perform mutual verification consistency aggregation, generate mutual verification consistency score sequence, and align and bind it with the three-state update decision sequence and risk budget sequence to form mutual verification backflow binding set.

7. The method for assessing the health status of new energy equipment based on recurrent neural networks according to claim 1, characterized in that, The generation of the online closed loop specifically includes: Receive the mutual verification feedback binding set, extract the mutual verification consistency score sequence, three-state update decision sequence and risk budget sequence, and simultaneously extract the data quality assessment results from the update threshold configuration, update density configuration and credibility evidence sequence, and gather them to form a closed-loop judgment input set; Generate a mutual verification trigger decision quantity on the closed-loop decision input set and determine the mutual verification trigger condition accordingly to obtain the mutual verification trigger flag sequence; When the mutual verification triggering flag sequence indicates that the mutual verification triggering condition is met, the three-state update decision of the corresponding evaluation window is overwritten as a double update, and the update density is increased and the threshold is tightened in the corresponding evaluation window and its subsequent preset continuous evaluation windows to generate a mutual verification enhancement configuration set. Generate a back-off trigger decision quantity on the closed-loop decision input set and determine the back-off trigger condition accordingly to obtain the back-off trigger flag sequence; When the rollback trigger flag sequence indicates that the rollback trigger condition is met, the rollback action set is executed and the rollback execution result set is generated. The mutual verification enhancement configuration set and the rollback execution result set are fed back to update the risk budget sequence and the update threshold configuration and update density configuration of the evidence-driven gating device. They are then aligned and bound to the gating window input set of the next evaluation window. In subsequent evaluation windows, they are used to generate three-state update decisions and update the dual-speed cyclic evaluation state, forming an online closed loop.

Images