A multi-dimensional alignment evaluation benchmark construction method for industrial operation strategy

Abstract

The application discloses a kind of multi-dimension alignment evaluation benchmarks of industrial operation strategy, it is related to industrial intelligent evaluation technical field, including obtaining the historical running data and operating condition record generated in the execution process of multiple to-be-evaluated industrial operation strategy, using K-means clustering algorithm to carry out multi-dimension feature extraction and classification processing to historical running data and operating condition record, obtain scene distribution set, for scene distribution set, carry out anomaly detection and boundary division by isolated forest algorithm, identify the scene distribution bias features contained in scene distribution set, obtain implicit bias identifier;The multi-dimension alignment evaluation benchmarks of the industrial operation strategy can improve the accuracy, reliability and consistency of industrial operation strategy performance evaluation, provide more objective evaluation benchmarks for the selection, optimization and application of industrial operation strategy.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent evaluation technology, specifically to a method for constructing a multi-dimensional alignment evaluation benchmark for industrial operation strategies. Background Technology

[0002] As manufacturing, chemical, and energy industries move towards intelligent and automated processes, industrial operation strategies are increasingly being applied to various aspects such as production scheduling, equipment control, process optimization, and operational status adjustment. Different industrial operation strategies analyze operating parameters, equipment status, and changes in operating conditions during the production process to achieve control and optimization of industrial processes. To evaluate the practical application effects of different industrial operation strategies, it is necessary to establish evaluation benchmarks that reflect the comprehensive performance of the strategies, in order to support technology selection, solution optimization, and application promotion.

[0003] In existing technologies, the performance evaluation of industrial operation strategies typically relies on expert experience to set evaluation rules or on building evaluation models based on existing historical operating data. However, evaluation methods based on expert experience are easily influenced by the evaluator's knowledge background, application scenario, and evaluation preferences, leading to different evaluation results for the same industrial operation strategy from different evaluators. Evaluation methods based on historical operating data are affected by data acquisition environment, equipment operating status, and differences in operating condition distribution, making it difficult to directly use data from different sources for unified evaluation. Furthermore, industrial operation strategies usually need to comprehensively consider multiple evaluation dimensions such as safety, production efficiency, energy consumption, and operational stability. However, different evaluation dimensions are interdependent; for example, increasing safety margins may reduce production efficiency, and reducing energy consumption may affect production capacity. Therefore, existing technologies typically require manually setting weights for different evaluation dimensions when building evaluation models. This weighting process is easily influenced by evaluation needs or application goals, allowing some operation strategies that conform to specific evaluation preferences to receive higher evaluations, reducing the objectivity of the evaluation results. In addition, industrial operation processes typically include various operating conditions such as normal operation scenarios, complex operation scenarios, and abnormal operation scenarios. Existing evaluation methods typically rely on existing operational data for sample selection when constructing evaluation scenarios. However, the distribution of scenarios in historical operational data may be unbalanced, with some common operating conditions having a high proportion of samples while complex or abnormal operating conditions have insufficient samples. This leads to evaluation results that are more biased towards reflecting the performance of strategies in certain scenarios, failing to comprehensively reflect the overall capabilities of industrial operation strategies under different operating conditions. Furthermore, since different industrial operation strategies may originate from different equipment, production environments, or technical paths, their corresponding operational data and evaluation results often exhibit scale differences. Existing technologies lack methods for unified scale comparison and capability quantification of operation strategies from different sources, making it difficult to establish objective and consistent evaluation benchmarks among different industrial operation strategies. Summary of the Invention

[0004] The purpose of this invention is to provide a method for constructing a multi-dimensional alignment evaluation benchmark for industrial operation strategies, thereby solving the problems existing in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for constructing a multi-dimensional alignment evaluation benchmark for industrial operation strategies, comprising: S1, acquiring historical operating data and operating condition records generated during the execution of multiple industrial operation strategies to be evaluated, and using the K-means clustering algorithm to perform multi-dimensional feature extraction and classification processing on the historical operating data and operating condition records to obtain a scene distribution set; S2, for the scene distribution set, performing anomaly detection and boundary division using the isolated forest algorithm to identify the scene distribution bias features contained within the scene distribution set, and obtaining a latent bias identifier; S3, if the latent bias identifier exceeds a pre-established bias judgment threshold, performing feature removal and resampling processing on the scene distribution set to obtain an unbiased scene distribution set; S4, based on the unbiased scene distribution set, using the random forest algorithm to perform feature extraction and classification processing on the safety dimension features, effectiveness, and other parameters. S5. The importance of the rate dimension feature and at least one other evaluation dimension feature is calculated and cross-validated to determine the multidimensional weight values; S6. The operation records corresponding to different industrial operation strategies within the unbiased scenario distribution set are weighted and normalized using the multidimensional weight values ​​to obtain a multidimensional comprehensive score; S7. For the multidimensional comprehensive score, pre-established benchmark verification rules are obtained. If the multidimensional comprehensive score conforms to the benchmark verification rules, the first true capability quantification value corresponding to the operation record is determined; S8. If the multidimensional comprehensive score does not conform to the benchmark verification rules, a penalty calculation is performed on the multidimensional comprehensive score to obtain a second true capability quantification value; S9. Based on the first and second true capability quantification values, the operation records corresponding to different industrial operation strategies are sorted and aligned to obtain the target evaluation benchmark for industrial operation strategy performance evaluation.

[0006] Preferably, step S1 includes acquiring historical operating data and operating condition records generated during the execution of multiple industrial operation strategies to be evaluated, constructing a multi-dimensional feature matrix, performing dimensionality reduction processing on the multi-dimensional feature matrix to obtain an enhanced feature set; determining whether the operating condition fluctuation amplitude of the enhanced feature set is greater than a preset fluctuation threshold; if it is greater than the preset fluctuation threshold, removing abnormal samples to obtain a standard feature set; using the K-means clustering algorithm to determine the cluster center vector on the standard feature set, dividing the steady-state operating interval according to the cluster center vector, and classifying the steady-state operating interval to obtain a scene distribution set.

[0007] Preferably, step S2 includes obtaining the distribution density and feature mapping of the scene distribution set to construct a distribution feature matrix; using the isolated forest algorithm to calculate the path depth of the distribution feature matrix to obtain an anomaly score sequence; setting a segmentation hyperplane based on the anomaly score sequence to obtain a bias feature set; determining whether the bias degree of the bias feature set is greater than a preset bias threshold; if the bias degree is greater than the preset bias threshold, then performing an identifier generation process on the bias feature set to obtain a latent bias identifier.

[0008] Preferably, step S3 includes: if the acquired latent bias identifiers exceed a pre-established bias determination threshold, then extracting the bias feature weights of the scene distribution set based on the latent bias identifiers; dividing the bias feature weights into a subset for removal, performing feature removal on the scene distribution set through the subset for removal, and obtaining a dimensionality-reduced scene distribution set; and fusing the synthetic sample set generated for the dimensionality-reduced scene distribution set with the dimensionality-reduced scene distribution set to perform resampling processing, thereby obtaining an unbiased scene distribution set.

[0009] Preferably, step S4 includes extracting safety dimension features, efficiency dimension features, and comfort dimension features from the unbiased scene distribution set to construct a multi-dimensional feature matrix; using a random forest algorithm to obtain feature splitting gain for the multi-dimensional feature matrix, and calculating the Gini impurity reduction based on the feature splitting gain to determine the initial feature importance; if the initial feature importance meets the preset significance judgment condition, then using out-of-bag data samples to perform cross-validation processing on the initial feature importance to obtain the mean prediction error; and performing normalization mapping processing on the initial feature importance based on the mean prediction error to determine the multi-dimensional weight values.

[0010] Preferably, step S5 includes obtaining operation records corresponding to industrial operation strategies within the unbiased scene distribution set; parsing the operation records to obtain operation action sequences and equipment status responses; calculating the strategy execution deviation based on the operation action sequences and equipment status responses; using multi-dimensional weight values ​​to perform weighted calculations on the strategy execution deviation to obtain a weighted aggregation matrix; determining whether the weighted aggregation matrix meets a preset non-empty condition; if the weighted aggregation matrix meets the preset non-empty condition, then performing extreme value normalization processing on the weighted aggregation matrix to obtain a multi-dimensional comprehensive score, thus completing the weighted calculation and normalization processing of operation records corresponding to different industrial operation strategies within the unbiased scene distribution set to obtain a multi-dimensional comprehensive score.

[0011] Preferably, step S6 includes comparing the multidimensional comprehensive score according to a pre-established benchmark verification rule to obtain a score comparison result; if the score comparison result conforms to the benchmark verification rule, then extracting the interaction frequency features and historical credit rating of the operation record; classifying the interaction frequency features and historical credit rating to construct a capability assessment matrix, and determining the first true capability quantification value corresponding to the operation record based on the capability assessment matrix.

[0012] Preferably, step S7 includes obtaining a multi-dimensional comprehensive score that does not conform to the benchmark verification rules, extracting the violation frequency and deviation range corresponding to the multi-dimensional comprehensive score; classifying the violation frequency and deviation range to obtain a risk level; matching penalty weights and attenuation factors according to the risk level, and using the penalty weights and attenuation factors to calculate credit damage on the multi-dimensional comprehensive score to obtain a corrected score; if the corrected score is lower than a preset quantification base, then penalizing the corrected score to obtain a second true capability quantification value.

[0013] Preferably, step S8 includes parsing the industrial operation records corresponding to the first true capability quantization value and the second true capability quantization value to obtain an initial action sequence; and performing matching processing based on the initial action sequence to obtain an alignment feature sequence.

[0014] Preferably, step S8 further includes extracting the response delay of the alignment feature sequence to construct a multidimensional feature matrix, performing dimensionality reduction processing on the multidimensional feature matrix to obtain the core deviation amplitude; arranging the industrial operation records according to the core deviation amplitude to obtain the target sorting sequence; and extracting performance scores from the target sorting sequence to construct a target evaluation benchmark for evaluating the performance of industrial operation strategies.

[0015] As can be seen from the above technical solution, the present invention has the following beneficial effects:

[0016] This method for constructing a multi-dimensional alignment evaluation benchmark for industrial operation strategies classifies historical operational data and operational condition records generated during the execution of multiple industrial operation strategies to be evaluated into scenarios, identifies bias characteristics in scenario distribution, and performs de-biasing on biased scenario data to reduce the impact of historical data distribution differences on evaluation results. Furthermore, by calculating weights for safety, efficiency, and other evaluation dimensions, a multi-dimensional comprehensive evaluation system is established to avoid evaluation bias caused by single evaluation indicators or manually set weights, thus improving the objectivity and comprehensiveness of evaluation results. Simultaneously, benchmark verification rules are used to verify the comprehensive score and quantify its capabilities, and operation records corresponding to industrial operation strategies from different sources are uniformly sorted and aligned to achieve comparability between different industrial operation strategies. Therefore, this invention can reduce the impact of data distribution differences, scenario bias, and inconsistent evaluation scales during the evaluation of industrial operation strategies, improving the accuracy, reliability, and consistency of industrial operation strategy performance evaluation, and providing a more objective evaluation benchmark for the selection, optimization, and application of industrial operation strategies. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention;

[0018] Figure 2This is a flowchart of the scene bias identification and correction process of the present invention;

[0019] Figure 3 This is a graph showing the change in contour coefficients corresponding to the number of different candidate scene categories in this invention.

[0020] Figure 4 This is a comparison diagram of the implicit bias identification before and after the removal of bias in the four industrial operation strategies of this invention;

[0021] Figure 5 This is a forest plot showing the distribution of importance of each evaluation feature under the five-fold cross-validation of random forest in this invention;

[0022] Figure 6 This is a radar chart comparing the three-dimensional scores of the four industrial operation strategies of this invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Example 1:

[0025] like Figure 1 and Figure 6 As shown, the present invention provides a technical solution: a method for constructing a multi-dimensional alignment evaluation benchmark for industrial operation strategies, comprising:

[0026] S1. Obtain historical operating data and operating condition records generated during the execution of multiple industrial operation strategies to be evaluated. Use the K-means clustering algorithm to perform multi-dimensional feature extraction and classification processing on the historical operating data and operating condition records to obtain a scene distribution set.

[0027] S2. For the scene distribution set, anomaly detection and boundary division are performed using the isolated forest algorithm to identify the scene distribution bias features contained within the scene distribution set and obtain the implicit bias identifier.

[0028] S3. If the latent bias flag exceeds the pre-established bias judgment threshold, feature removal and resampling are performed on the scene distribution set to obtain an unbiased scene distribution set.

[0029] S4. Based on the unbiased scene distribution set, the random forest algorithm is used to calculate the feature importance and cross-validate the security dimension feature, efficiency dimension feature and at least one other evaluation dimension feature to determine the multidimensional weight values.

[0030] S5. By using multi-dimensional weight values, the operation records corresponding to different industrial operation strategies in the unbiased scene distribution set are weighted and normalized to obtain a multi-dimensional comprehensive score.

[0031] S6. For the multidimensional comprehensive score, obtain the pre-established benchmark verification rules. If the multidimensional comprehensive score meets the benchmark verification rules, determine the first true capability quantification value corresponding to the operation record.

[0032] S7. If the multidimensional comprehensive score does not meet the benchmark verification rules, a penalty calculation will be performed on the multidimensional comprehensive score to obtain the second true ability quantification value.

[0033] S8. Based on the first true capability quantification value and the second true capability quantification value, sort and align the operation records corresponding to different industrial operation strategies to obtain the target evaluation benchmark for industrial operation strategy performance evaluation.

[0034] In this embodiment, historical operating data and operating condition records of each industrial operation strategy to be evaluated are first collected during the actual execution of industrial equipment, industrial production line or industrial control system. A one-to-one correspondence is established according to at least one of the following associated identifiers: timestamp, production batch number and task number, so that each operation record can determine its industrial operation strategy, execution time, operating condition and execution result.

[0035] Specifically, historical operating data includes equipment status parameters, control commands, set parameters, actual measured parameters, alarm status, interlock status, operation duration, energy consumption, material consumption, and task completion status; operating condition records include equipment load rate, production task category, process stage, equipment working mode, raw material status, and environmental parameters; through the above correlation processing, historical operating data and operating condition records form original evaluation records with a single operation process as the basic unit.

[0036] In addition, the original evaluation records were checked for data validity.

[0037] Specifically, for continuous sampling parameters, the sampling time is compared with the start and end times of the corresponding operation process, and only data within the start and end time range of the operation process is retained; records with missing timestamps, missing industrial operation strategy identifiers, missing operating condition identifiers, or missing execution results are marked as invalid records and deleted from the scenario construction data; for duplicate records with identical content at the same timestamp, only one is retained; through the above processing, each piece of data entering subsequent steps has a clear data source and a complete operating condition correspondence.

[0038] Valid records are characterized.

[0039] For continuous features, the arithmetic mean and standard deviation of all valid records for the current feature are calculated. The arithmetic mean of the feature is subtracted from the original feature value of the current record, and the difference is divided by the standard deviation of the feature to obtain the standardized feature value. When the standard deviation of a feature is equal to 0, all standardized results of the feature are set to 0. For discrete features such as equipment working mode, production task category, process stage and raw material category, a fixed coding table is pre-established, and a unique integer code is assigned to each category. The same code is used for the same category throughout the entire evaluation process. After processing, each operation record consists of scene features with the same number and the same order.

[0040] In this embodiment, the K-means clustering algorithm is used to classify all scene features.

[0041] First, a candidate range for the number of clustering categories is defined. In this embodiment, the number of candidate categories is set to 2, 3, 4, 5, 6, 7, and 8 categories in sequence. For each number of candidate categories, a complete clustering is performed, and the silhouette coefficient of the corresponding clustering result is calculated. The number of candidate categories with the largest silhouette coefficient value is determined as the final number of scene categories. When the silhouette coefficients corresponding to two candidate categories are the same, the candidate category with the smaller value is selected as the final number of scene categories.

[0042] Furthermore, after determining the number of scene categories, initial cluster centers are selected from the valid operation records, which are the same as the number of scene categories, and the Euclidean distance between each operation record and each cluster center is calculated one by one.

[0043] Specifically, for each operation record, it is assigned to the scene category corresponding to the cluster center with the smallest distance value; when the distance values ​​of an operation record to two cluster centers are the same, it is assigned to the scene category with the smaller number value; after all operation records are assigned, the arithmetic mean of each scene feature of all operation records in each scene category is calculated, and the calculated arithmetic mean of each feature is used to form a new cluster center.

[0044] Then, the distance calculation, scene allocation, and cluster center update are repeated until there are no operation records of scene category changes in two consecutive rounds of clustering results, or the number of iterations reaches 300. When the number of iterations reaches 300, the clustering result obtained in the 300th iteration is used as the final classification result, thereby obtaining the scene distribution set corresponding to the industrial operation scene.

[0045] Furthermore, after completing the scenario classification, the number of valid operation records for each industrial operation strategy in each scenario category is counted. The distribution ratio of the industrial operation strategy in a scenario is obtained by dividing the number of valid operation records of a certain industrial operation strategy in a certain scenario by the total number of valid operation records of the industrial operation strategy. For the same scenario, the distribution ratios of all industrial operation strategies to be evaluated in the scenario are added together, and the sum is divided by the number of industrial operation strategies to be evaluated to obtain the baseline distribution ratio corresponding to the scenario.

[0046] Through the above calculations, each industrial operation strategy generates scenario distribution data that reflects the allocation status of its operation records in various operating scenarios.

[0047] In addition, isolated forest anomaly detection is performed on the scene distribution set.

[0048] This embodiment establishes 300 isolated trees. When establishing each isolated tree, a feature is randomly selected from the scene features, and a split value is randomly determined between the minimum and maximum values ​​of the current sample of that feature. Based on the relationship between the feature value corresponding to the operation record and the split value, the operation record is divided into two child nodes, and the division continues in the same way until a single node retains only one operation record, all operation records within a node have the same feature value, or a pre-set tree depth limit is reached. For each operation record, the number of nodes it traverses from the root node to the isolated node in the 300 isolated trees is counted, and it is converted into an anomaly score between 0 and 1 according to the path length conversion rules of the isolated forest.

[0049] In this embodiment, the abnormal scoring threshold is set to 0.60.

[0050] Specifically, when the anomaly score of an operation record is greater than or equal to 0.60, the operation record is identified as an abnormal scenario record; when the anomaly score is less than 0.60, the operation record is identified as a normal scenario record. For each industrial operation strategy, the number of abnormal scenario records is counted, and then the number of abnormal scenario records is divided by the total number of valid operation records for that industrial operation strategy to obtain the percentage of abnormal scenario records.

[0051] In addition, the distribution of calculation scenarios varies for each industrial operation strategy.

[0052] Specifically, the distribution ratio of the industrial operation strategy in each scenario and the corresponding baseline distribution ratio in each scenario are obtained sequentially. The absolute value of the difference between the two is calculated, and the absolute values ​​corresponding to all scenarios are added together. The sum is then divided by 2 to obtain the scenario distribution difference value of the industrial operation strategy. Subsequently, the scenario distribution difference value is multiplied by 0.5, the abnormal scenario record ratio is multiplied by 0.5, and the two calculation results are added together to obtain the implicit bias indicator corresponding to the industrial operation strategy. The value range of the implicit bias indicator is 0 to 1.

[0053] In this embodiment, the bias determination threshold is set to 0.20.

[0054] Specifically, when the implicit bias indicator of a certain industrial operation strategy is less than or equal to 0.20, the scenario distribution corresponding to the industrial operation strategy is determined to meet the subsequent evaluation requirements; when the implicit bias indicator is greater than 0.20, the current scenario distribution is determined to require feature removal and resampling processing.

[0055] In addition, when performing feature removal, each scene feature is verified individually in turn.

[0056] Specifically, when verifying a scene feature, the feature is temporarily removed from the scene feature set, and scene classification, anomaly detection, and latent bias label calculation are re-executed while keeping other processing parameters unchanged. The latent bias label before feature removal is subtracted from the latent bias label after feature removal to obtain the bias contribution value corresponding to the feature. When the bias contribution value of a feature is greater than or equal to 0.02, and the feature belongs to the data acquisition batch identifier, equipment number, record generation time code, or other data source features that do not participate in the industrial operation strategy capability evaluation, the feature is removed from the scene feature set. Safety evaluation features, efficiency evaluation features, and capability evaluation features directly used by other evaluation dimensions are not subject to the above removal.

[0057] Furthermore, after feature removal is completed, resampling is performed.

[0058] First, all scenario categories are checked one by one. A scenario is identified as a common scenario only when all industrial operation strategies to be evaluated have at least one valid operation record in the same scenario category. For each common scenario, the number of valid operation records for each industrial operation strategy in the common scenario is obtained, and the number of valid operation records with the smallest value is determined as the target number of records for the common scenario. Then, for each industrial operation strategy, operation records with the same number as the target number are extracted from the existing real valid operation records in the common scenario using a no-replacement method, so that the number of operation records for each industrial operation strategy in the same common scenario participating in the subsequent evaluation is consistent.

[0059] Furthermore, after resampling all public scenarios, the implicit bias labels of each industrial operation strategy are calculated again according to the same isolated forest detection rules and bias label calculation rules.

[0060] Specifically, when the implicit bias identifiers of all industrial operation strategies are less than or equal to 0.20, the current processed scenario distribution set is determined as the unbiased scenario distribution set; the unbiased scenario distribution set refers to the data set that meets the bias judgment threshold after being processed according to the unified calculation rules.

[0061] In addition, evaluation dimensions are established based on the unbiased scenario distribution set. These evaluation dimensions include at least a security dimension, an efficiency dimension, and one other evaluation dimension.

[0062] In this embodiment, other evaluation dimensions are defined as stability dimensions. The safety dimension includes the state of safety constraint satisfaction, the number of alarms, the number of interlock triggers, and the duration of critical operating parameters exceeding limits; the efficiency dimension includes task completion time, the amount of task completed per unit time, energy consumption, and material consumption; the stability dimension includes the fluctuation range of critical operating parameters during operation, the deviation between set values ​​and actual values, and the state offset within twenty consecutive sampling periods after the operation ends.

[0063] Furthermore, the importance of each evaluation feature is calculated using the random forest algorithm.

[0064] In this embodiment, 500 decision trees are established and 5-fold cross-validation is used. First, the unbiased scenario distribution set is divided into 5 non-overlapping data subsets according to the number of operation records. In each round, 4 data subsets are selected for model training, and the remaining data subset is used for validation. This process is repeated for 5 rounds, so that each data subset serves as validation data once. After each round of training, the feature importance of each evaluation feature in the security, efficiency, and stability dimensions is recorded. After 5 rounds of validation, the feature importance of the same evaluation feature in the 5 rounds of training is added together, and the sum is divided by 5 to obtain the final feature importance of the evaluation feature.

[0065] Furthermore, the importance of all evaluation features belonging to the safety dimension is summed to obtain the importance of the safety dimension; the importance of all evaluation features belonging to the efficiency dimension is summed to obtain the importance of the efficiency dimension; and the importance of all evaluation features belonging to the stability dimension is summed to obtain the importance of the stability dimension. Then, the importance of the three dimensions is summed to obtain the total importance value. The importance of each dimension is divided by the total importance value to obtain the weights of the safety dimension, efficiency dimension, and stability dimension, so that the sum of the weights of the three dimensions equals 1.

[0066] In addition, a unified numerical scale is applied to all evaluation features.

[0067] For evaluation features such as task completion rate, safety constraint satisfaction rate and control accuracy, which represent improvements in evaluation results, the maximum and minimum values ​​of the current evaluation feature in the unbiased scenario distribution set are first obtained. Then, the minimum value is subtracted from the feature value of the current operation record, and the result is divided by the difference between the maximum and minimum values ​​to obtain a normalized result between 0 and 1.

[0068] For evaluation features where an increase in values ​​such as alarm count, interlock trigger count, task completion time, energy consumption, material consumption, and parameter out-of-bounds duration represents a decrease in evaluation results, first subtract the feature value of the current operation record from the maximum value, and then divide the result by the difference between the maximum and minimum values. When the maximum and minimum values ​​of the same evaluation feature are the same, set all normalized results of the evaluation feature to 0.5.

[0069] As a preferred embodiment, for the same evaluation dimension, each normalized evaluation feature in the dimension is multiplied by its corresponding feature importance, and all products are added together. The sum is then divided by the sum of the importance of all evaluation features in the dimension, thereby obtaining the safety dimension score, efficiency dimension score, and stability dimension score respectively; the values ​​of the three dimension scores are all in the range of 0 to 1.

[0070] Furthermore, the security dimension score is multiplied by the security dimension weight, the efficiency dimension score is multiplied by the efficiency dimension weight, and the stability dimension score is multiplied by the stability dimension weight. The three calculation results are then added together to obtain the multidimensional comprehensive score for each operation record. The value range of the multidimensional comprehensive score is 0 to 1.

[0071] When a percentage score is required, the multidimensional comprehensive score is multiplied by 100, and the result is used as the percentage multidimensional comprehensive score for the corresponding operation record.

[0072] Furthermore, benchmark verification rules are established in advance. This embodiment uses three benchmark verification conditions.

[0073] Specifically, the first condition is that the safety dimension score must not be less than 0.80; the second condition is that the proportion of valid operation records must not be less than 0.95, where the proportion of valid operation records is obtained by dividing the number of valid records after the data validity check by the total number of operation records before the data validity check; the third condition is that the industrial operation strategy has at least one valid operation record participating in the final score in all public scenarios; when all three conditions are met, the multidimensional comprehensive score corresponding to the operation record is determined to meet the benchmark verification rules, and the multidimensional comprehensive score is determined as the first true capability quantification value; when using a percentage output, this value is multiplied by 100 to obtain the first true capability quantification value.

[0074] Specifically, if any one of the three benchmark verification conditions is not met, a penalty calculation is performed on the multidimensional comprehensive score used for benchmark verification.

[0075] The multidimensional comprehensive score used for benchmark verification is the score obtained in step five and verified in step six, which is used as the first multidimensional comprehensive score in the penalty process.

[0076] In addition, the degree of penalty for the security dimension is determined as follows: when the security dimension score is greater than or equal to 0.80, the degree of penalty for the security dimension is set to 0; when the security dimension score is less than 0.80, the actual security dimension score is subtracted from 0.80, and the difference is divided by 0.80 to obtain the degree of penalty for the security dimension.

[0077] The penalty level for valid operation records is determined as follows: when the proportion of valid operation records is greater than or equal to 0.95, the penalty level for valid operation records is set to 0; when the proportion of valid operation records is less than 0.95, the actual proportion of valid operation records is subtracted from 0.95, and the difference is divided by 0.95 to obtain the penalty level for valid operation records.

[0078] In this embodiment, the level of penalty for scene coverage is determined according to the actual number of missing public scenes.

[0079] First, count the total number of public scenarios. Then, count the number of public scenarios with valid scoring records for the industrial operation strategy. Subtract the number of public scenarios with valid scoring records from the total number of public scenarios to get the number of missing public scenarios. Finally, divide the number of missing public scenarios by the total number of public scenarios to get the scenario coverage penalty.

[0080] Furthermore, the penalty for security is multiplied by 0.5, the penalty for valid operation records is multiplied by 0.3, and the penalty for scene coverage is multiplied by 0.2. The three results are then added together to obtain the comprehensive penalty.

[0081] Then, subtract the comprehensive penalty level from 1 and multiply the result by the first multidimensional comprehensive score to obtain the second true ability quantification value; when outputting on a percentage scale, multiply the second true ability quantification value by 100.

[0082] Therefore, operation records that do not meet the benchmark verification rules are corrected by scoring according to the unified numerical rules.

[0083] In addition, the first true capability quantization value is used for operation records that meet the benchmark verification rules, and the second true capability quantization value is used for operation records that do not meet the benchmark verification rules.

[0084] For all operation records of the same industrial operation strategy in the same public scenario, the actual capability quantification values ​​corresponding to each operation record are added together, and the sum is divided by the number of valid operation records participating in the evaluation in the scenario to obtain the scenario capability value of the industrial operation strategy in the public scenario.

[0085] Further determine the unified weight of each public scenario in the final evaluation benchmark.

[0086] Specifically, the total number of valid operation records in each public scenario after resampling is counted, and then the total number of valid operation records in all public scenarios is summed. The number of valid operation records in a certain public scenario is divided by the total number of valid operation records to obtain the scenario weight of that public scenario. Then, for each industrial operation strategy, its scenario capability value in each public scenario is multiplied by the corresponding scenario weight, and all products are summed to obtain the overall capability quantification value of the industrial operation strategy.

[0087] The industrial operation strategies are ranked from the largest to the smallest based on the overall capability quantification value.

[0088] When the overall capability quantification values ​​of the two industrial operation strategies are the same, the average scores of the safety dimension of the two industrial operation strategies are compared, and the industrial operation strategy with the larger value is ranked first; when the average scores of the safety dimension are still the same, the average scores of the efficiency dimension are compared, and the industrial operation strategy with the larger value is ranked first; when the aforementioned scores are still the same, the ranking order is determined by the unique number pre-assigned to the industrial operation strategy from smallest to largest.

[0089] The system establishes a fixed field correspondence relationship and stores the following data: unique industrial operation strategy ID, scenario category, number of scenario records, safety dimension score, efficiency dimension score, stability dimension score, weight of each evaluation dimension, first or second true capability quantification value, scenario capability value of each common scenario, overall capability quantification value, and final sorting order. This forms the target evaluation benchmark for industrial operation strategy performance evaluation.

[0090] To verify the effectiveness of this application, an experiment is designed as follows:

[0091] The experiment used the Tennessee-Eastman process standard chemical process simulation platform as the process prototype to build a semi-physical simulation environment for temperature-pressure coordinated control of a continuous stirred reactor. The upper-level simulation software ran the process dynamics model, while the lower-level control loop was actually executed by a Siemens S7-1500 PLC to collect equipment status parameters. Historical operating data such as equipment status parameters, control commands, alarm status, interlock status, energy consumption, and material consumption were recorded at a 1-minute sampling interval through the OPC UA interface. At the same time, operating conditions such as equipment load rate, production task type, process stage, and raw material status were also recorded.

[0092] Four industrial operation strategies were deployed in the experiment and operated continuously for 30 days under the same process flow and physical equipment constraints.

[0093] Strategy A: PID baseline control strategy (existing engineering tuning parameters);

[0094] Strategy B: Fuzzy PID adaptive control strategy;

[0095] Strategy C: Expert rule base control strategy (based on process engineer experience rules);

[0096] Strategy D: Deep reinforcement learning control strategy (trained offline based on historical data and then deployed online).

[0097] Each strategy was recorded and aggregated in 10-minute intervals. Over 30 days, approximately 2,400 original operation records were generated per strategy, totaling approximately 9,600 original records, which constituted the source of historical operation data and operational status records for this experiment.

[0098] The experimental design comprises five progressive experiments, and the correspondence between each group of experiments and the steps in this application is shown in the table below:

[0099] Table 1. Correspondence between experimental groups and steps in this application

[0100]

[0101] The data validity check is performed according to the rules described in this application: records with missing timestamps, missing policy identifiers, missing operating condition identifiers, or missing execution results are removed; completely duplicate records with the same timestamp are removed; continuous sampling parameters are truncated according to the start and end time range of the operation; and the proportion of valid records for each policy after the check is shown in the table below.

[0102] Table 2 Statistics of Original and Valid Records for Each Strategy

[0103]

[0104] Experiment 1: Verification of Scene Clustering Results

[0105] Feature processing was performed on 9144 valid operation records using four strategies (continuous features were standardized, and discrete features were fixed-encoded). K-means clustering was then performed on candidate categories ranging from 2 to 8, and silhouette coefficients were calculated. The results are shown in the table below.

[0106] Table 3. Silhouette coefficients corresponding to the number of different candidate scene categories

[0107]

[0108] like Figure 3 As shown in Table 3, the silhouette coefficient reaches a maximum value of 0.603 when the number of candidate categories is 5. Therefore, the final number of scenario categories is determined to be 5. Based on the process log verification, the 5 scenario categories correspond to 5 typical operating conditions, namely, low load stable operation with load increase, high load full production, load change and raw material fluctuation disturbance. This is basically consistent with the judgment based on field experience, indicating that K-means clustering can extract a set of scenario distributions with process interpretability from historical operating data and operating condition records.

[0109] Experiment 2: Detection and Verification of Latent Bias Reduction Effect

[0110] Based on the five types of scenario distribution sets, 300 isolated trees were built to detect anomalies in each operation record. Normal / abnormal scenario records were divided according to an anomaly score threshold of 0.60. The implicit bias indicator of each strategy was calculated according to this application (scenario distribution difference value weight 0.5 + abnormal scenario record proportion weight 0.5). The results are shown in the table below:

[0111] Table 4 Comparison of Implicit Bias Labels Before and After Bias Removal for Each Strategy

[0112]

[0113] like Figure 4 As shown in Table 4, the latent bias indicators of strategies A, B, and D before debiasing all exceeded the 0.20 threshold. This is mainly because the distribution ratio of historical operation records for these three strategies in high-load, full-production scenarios was significantly lower than the baseline distribution ratio (i.e., a long-term lack of operational data under high-load conditions). After feature removal (removing data source features irrelevant to capability evaluation, such as equipment numbers and batch identifiers) and resampling without replacement in the same scenario, the latent bias indicators of the three strategies decreased to 0.15, 0.17, and 0.16, respectively, all below the threshold, and the scenario distribution set transformed into an unbiased scenario distribution set. Strategy C, with its original latent bias indicator of 0.18, did not trigger debiasing and its value remained unchanged. These results indicate that isolated forest anomaly detection can effectively identify latent biases caused by uneven scenario coverage in historical data, and debiasing has a clear corrective effect on biased strategies.

[0114] Experiment 3: Verification of Multidimensional Evaluation Weight Calculation Results

[0115] For the unbiased scene distribution set, 12 evaluation features were established across three evaluation dimensions: safety, efficiency, and stability. The importance of each feature was calculated using a random forest algorithm with 500 decision trees and 5-fold cross-validation. The results are shown in the table below:

[0116] Table 5 Summary of the importance of features for each evaluation dimension and their final weights

[0117]

[0118] like Figure 5 As shown in Table 5, in the 5-fold cross-validation, the safety constraint satisfaction rate had the highest importance in the safety dimension (mean 0.114), the task completion time had the highest importance in the efficiency dimension (mean 0.093), and the deviation between the set value and the actual value had the highest importance in the stability dimension (mean 0.076). Furthermore, the overall fluctuation range of the importance of each feature in the 5-fold cross-validation was relatively narrow, indicating that the feature importance ranking remained stable across different validation subsets, and no single subset dominated the weight results. Finally, the weights for the safety dimension (0.36), efficiency dimension (0.37), and stability dimension (0.27) were obtained, with a sum of 1.00. These weights were calculated using data-driven feature importance calculations and did not rely on subjective manual assignment.

[0119] Experiment 4: Comprehensive Scoring and Benchmark Verification / Penalty Mechanism Validation

[0120] The operation records of each strategy within the unbiased scene distribution set are weighted and normalized according to multi-dimensional weights to obtain a three-dimensional score and a multi-dimensional comprehensive score. The scores are then validated according to benchmark verification rules (security dimension score ≥ 0.80, effective operation record ratio ≥ 0.95, full coverage of common scenes). The experimental results are shown in the table below:

[0121] Table 6. Three-dimensional scores, benchmark verification results, and actual capability quantification values ​​for each strategy

[0122]

[0123] like Figure 6As shown, Strategy A's security dimension score of 0.78 is lower than the baseline of 0.80. The calculated security dimension penalty is (0.80-0.78) / 0.80=0.025. After adding the effective record penalty (0) and scene coverage penalty (0), the comprehensive penalty is 0.025×0.5=0.0125. The corrected true capability quantification value is 0.706×(1-0.0125)×100≈69.7, which, after rounding and scene weighting, is finally 66.4. Strategy B triggers an effective record penalty because its effective record ratio of 94.0% is lower than the 95% baseline, and the corrected quantification value drops to 71.8. This result indicates that the benchmark verification and penalty mechanism can objectively reduce the risk of security vulnerabilities or insufficient data integrity in strategies, avoiding the masking of weaknesses in key dimensions by relying solely on a single comprehensive score value.

[0124] Experiment 5: Robustness and Consistency Verification of the Final Evaluation Benchmark Ranking

[0125] The scenario capability values ​​of each strategy in the five common scenarios are weighted and summarized according to the common scenario weight (determined by the proportion of valid operation records in each common scenario after resampling) to obtain the overall capability quantification value and complete the final ranking. At the same time, in order to verify the robustness of the ranking results, 10 rounds of resampling without replacement are independently and repeatedly performed on the unbiased scenario distribution set (each round of resampling seed is different), and the overall capability quantification value and ranking are recalculated. The results are shown in the table below:

[0126] Table 7 Overall Capability Quantification and Final Ranking of Each Strategy

[0127]

[0128] To verify the robustness of the above ranking results, ten rounds of unbiased scene distribution set were independently and repeatedly resampled without replacement (with a different seed in each round). The overall capability quantification and ranking were recalculated, and the number of times each strategy maintained the ranking position shown in Table 7 was counted in the 10 rounds. The results are shown in the table below:

[0129] Table 8. Ranking stability rate of 10 rounds of independent resampling test

[0130]

[0131] In addition, three on-site process engineers were invited to independently rank the four strategies based on their experience, without being informed of the evaluation results of this method. The ranking results given by the three engineers were all that strategy D is better than strategy C, which is better than strategy B, which is better than strategy A, which is completely consistent with the final ranking of this method (consistency rate 100%).

[0132] The ranking stability rate reflects the robustness of the final benchmark ranking results against random fluctuations in the scene samples. Due to the small difference in overall capability quantification values ​​(82.2 and 77.6 respectively), strategies D and C experienced ranking swaps in a few resampling rounds, but this accounted for only 10% of the total 10 rounds, maintaining a consistent overall ranking trend. Strategies A and B showed relatively significant differences in quantification values ​​compared to other strategies, and their rankings remained stable across all 10 rounds. The ranking by human experts was completely consistent with the method ranking, further confirming the good reliability and engineering credibility of the benchmark obtained by this method.

[0133] Example 2:

[0134] S1 includes acquiring historical operating data and operating condition records generated during the execution of multiple industrial operation strategies to be evaluated, constructing a multi-dimensional feature matrix, performing dimensionality reduction on the multi-dimensional feature matrix to obtain an enhanced feature set; determining whether the operating condition fluctuation amplitude of the enhanced feature set is greater than a preset fluctuation threshold; if it is greater than the preset fluctuation threshold, removing abnormal samples to obtain a standard feature set; using the K-means clustering algorithm to determine the cluster center vector on the standard feature set, dividing the steady-state operating interval according to the cluster center vector, and classifying the steady-state operating interval to obtain a scene distribution set.

[0135] In this embodiment, historical operating data and operating condition records are associated with a unified time base.

[0136] Specifically, the data extraction interval is defined as the time from the start to the end of a single execution of an industrial operation strategy. Equipment load, process temperature, process pressure, media flow rate, actuator opening degree, control commands, actual feedback values, energy consumption, and alarm status are matched with the corresponding equipment operating mode, production task category, process stage, and raw material category at that time. Continuous parameters are time-aligned according to a 1-second sampling period. When there are more than two sampled values ​​within 1 second, the arithmetic mean is taken. When the time interval between adjacent valid sampled values ​​does not exceed 5 seconds, linear interpolation is used to fill in the gaps. Records with time intervals exceeding 5 seconds are not included in this feature construction. Discrete operating condition parameters are maintained until the next state change time, thus ensuring that each sampling time corresponds to a complete set of operating features.

[0137] Furthermore, after time alignment is completed, a multidimensional feature matrix is ​​constructed according to a pre-fixed feature arrangement order.

[0138] For continuous features, the arithmetic mean and standard deviation of all valid samples of the feature are calculated first. Then, the arithmetic mean is subtracted from the original value of a single sample, and the difference is divided by the standard deviation to obtain the standardized feature value. When the standard deviation is 0, the standardized result of the feature is recorded as 0.

[0139] For discrete characteristics such as production task categories, equipment operating modes, and process stages, numerical processing is performed based on the coding table determined when the evaluation task is established. The same code is used for the same category in all industrial operation strategies to be evaluated, so as to avoid introducing additional differences due to changes in coding rules.

[0140] In addition, principal component analysis was used to reduce the dimensionality of the resulting multidimensional feature matrix.

[0141] First, the covariance between the standardized features is calculated. Then, the direction of each principal component and its corresponding eigenvalue are calculated based on the covariance, and the eigenvalues ​​are arranged in descending order. The contribution rate of each principal component is obtained by dividing the eigenvalue of that principal component by the sum of the eigenvalues ​​of all principal components. The contribution rate is accumulated sequentially starting from the principal component with the largest contribution rate. The accumulation stops when the cumulative contribution rate first reaches 0.90, and all principal components up to that principal component are retained. Then, each operation record is projected onto the retained principal component direction, and the resulting dimensionality-reduced features are recombined according to the original time order to form an enhanced feature set.

[0142] After this processing, the data dimension for subsequent working condition determination is determined by the cumulative contribution rate.

[0143] Furthermore, after the enhanced feature set is formed, each dimensionality-reduced feature is transformed into the 0-to-1 range.

[0144] Specifically, for any dimensionality reduction feature, the maximum and minimum values ​​of that feature in the current evaluation data are obtained. The minimum value is subtracted from the current sample value, and then divided by the difference between the maximum and minimum values ​​to obtain the transformation result. When the maximum and minimum values ​​are the same, the transformation result of the dimensionality reduction feature is uniformly recorded as 0.5. After completing the interval transformation, a fluctuation detection window is formed by 20 consecutive sampling points. The detection window moves by 1 sampling point along the time direction each time. Since the sampling period is 1 second, each detection window corresponds to a 20-second continuous running process.

[0145] The preset fluctuation threshold is calculated from historical steady-state operation records.

[0146] Specifically, operation records with alarm status not triggered, interlock status not triggered, and key process parameters all within the allowable operating range of the equipment are selected from historical operation data as threshold calibration data, and processed according to the same 20-second detection window as the data to be evaluated; for each detection window, the difference between the maximum and minimum values ​​of each enhancement feature within the window is calculated, and then the differences corresponding to all enhancement features are added together and divided by the number of enhancement features, and the result is used as the operating condition fluctuation amplitude of the detection window; the arithmetic mean and standard deviation of the operating condition fluctuation amplitude of all threshold calibration windows are calculated, and the sum of the arithmetic mean and three times the standard deviation is determined as the preset fluctuation threshold; after the threshold is determined, it remains unchanged in the same round of evaluation.

[0147] In addition, for the enhanced feature set of the industrial operation strategy to be evaluated, the operating condition fluctuation amplitude of the 20-second detection window is calculated one by one using the same method.

[0148] When the fluctuation range of the operating conditions in the detection window is not greater than the preset fluctuation threshold, the samples in the window are retained; when the fluctuation range of the operating conditions is greater than the preset fluctuation threshold, the samples in the window are further subjected to anomaly judgment; for each enhanced feature, the arithmetic mean and standard deviation in the window are calculated, and then the absolute value of the difference between the feature value of each sampling point and the corresponding arithmetic mean is calculated.

[0149] When the absolute difference of any enhanced feature is greater than three times the standard deviation of that feature, the corresponding sampling point is identified as an abnormal sample. When a sampling point falls into an adjacent detection window at the same time, as long as the above abnormal conditions are met in any detection window, the sampling point is removed from the enhanced feature set. After all detection windows are processed, the remaining samples are arranged in the original time order to obtain the standard feature set.

[0150] As a preferred implementation, the standard feature set is used to segment the scene using the K-means clustering algorithm.

[0151] Specifically, the number of cluster categories ranges from 2 to 8. Clustering is performed and silhouette coefficients are calculated for each candidate category number. The silhouette value of a single sample is calculated based on its average Euclidean distance to other samples in its category and its average Euclidean distance to the nearest sample in another category. The arithmetic mean of the silhouette values ​​of all samples is used as the silhouette coefficient corresponding to the current candidate category number. The candidate category number with the largest silhouette coefficient is determined as the final number of cluster categories. When there are identical silhouette coefficients, the result with the smaller number of categories is adopted.

[0152] Furthermore, after determining the number of cluster categories, the initial cluster centers are selected using the K-means addition method.

[0153] First, the first initial center is determined from the standard feature set. Subsequent centers are selected based on the minimum Euclidean distance between the sample and the determined center, until the number of initial centers reaches the final number of cluster categories. Then, the Euclidean distance between each standard feature record and each cluster center is calculated, and the record is assigned to the cluster category with the smallest distance value. After one round of allocation, the arithmetic mean of all records in each category on each enhanced feature is calculated, and the arithmetic mean is used to form a new cluster center vector. The record allocation and center update are repeated until the category affiliation of all records is consistent in two consecutive rounds of processing. If the above conditions are not met after 300 iterations, the result of the 300th iteration is used as the clustering result.

[0154] Furthermore, after the cluster center vector is determined, each cluster center is used as the feature center of the corresponding steady-state operating interval.

[0155] Specifically, for a cluster category, the Euclidean distances from all samples in that category to the cluster center are calculated, and then the arithmetic mean and standard deviation of these Euclidean distances are calculated. The sum of the arithmetic mean and twice the standard deviation is determined as the steady-state distance boundary of that category. If the Euclidean distance between a standard feature record and its corresponding cluster center is not greater than the steady-state distance boundary, and the fluctuation range of the detection window to which the record belongs is not greater than a preset fluctuation threshold, the record is included in the steady-state operating range corresponding to that cluster category. Records that do not meet any of the conditions are not included in the steady-state operating range. The steady-state distance boundary is directly calculated from the distance distribution of samples within the current cluster category, and each cluster category is calculated separately.

[0156] When classifying steady-state operating intervals, each steady-state operating interval uses the category identifier of the corresponding cluster center.

[0157] Specifically, records with continuous time in the same category are merged into one scene segment, while records with discontinuous time are treated as independent scene segments, but all belong to the same scene category. For each industrial operation strategy, the number of valid records in each scene category is counted, and the number of valid records in a scene category is divided by the total number of all steady-state valid records of the industrial operation strategy to obtain the scene distribution ratio of the industrial operation strategy in that scene category.

[0158] Finally, a corresponding relationship is established between scene categories, cluster center vectors, steady-state distance boundaries, the number of effective records of each industrial operation strategy in the corresponding scene, and the scene distribution ratio, forming a scene distribution set.

[0159] Example 3:

[0160] S2 includes obtaining the distribution density and feature mapping of the scene distribution set to construct a distribution feature matrix; using the isolated forest algorithm to calculate the path depth of the distribution feature matrix to obtain anomaly score sequences; setting a segmentation hyperplane based on the anomaly score sequences to obtain a bias feature set; determining whether the bias degree of the bias feature set is greater than a preset bias threshold; if the bias degree is greater than the preset bias threshold, then performing an identifier generation process on the bias feature set to obtain a latent bias identifier.

[0161] In this embodiment, the scene distribution set obtained by S1 is organized according to industrial operation strategies and scene categories, and each industrial operation strategy forms a set of scene records under each scene category.

[0162] For any given scenario, the number of valid records for the corresponding industrial operation strategy within that scenario is counted. The number of valid records in that scenario is then divided by the total number of valid records for that industrial operation strategy across all scenarios. The result is used as the distribution density of that industrial operation strategy in the current scenario. The sum of the distribution densities of the same industrial operation strategy across all scenarios is 1, in order to eliminate the impact of differences in the total number of original records between different industrial operation strategies on the comparison of scenario distribution.

[0163] Furthermore, after obtaining the distribution density, feature mapping is performed on the records of each scene.

[0164] Specifically, for each scenario feature retained by S1, the arithmetic mean of that feature in the current scenario is calculated for the current industrial operation strategy. Then, the corresponding feature values ​​of all industrial operation strategies to be evaluated in the same scenario are merged, and the scenario baseline mean and standard deviation of that feature are calculated. The scenario baseline mean is subtracted from the feature mean of the current industrial operation strategy, and the absolute value of the difference is taken and then divided by the scenario baseline standard deviation to obtain the mapping offset value of that feature. When the scenario baseline standard deviation is equal to 0 and the two mean values ​​are equal, the mapping offset value is recorded as 0; when the scenario baseline standard deviation is equal to 0 and the two mean values ​​are not equal, the mapping offset value is recorded as 1.

[0165] Subsequently, the distribution density and mapping offset values ​​of various features corresponding to the same industrial operation strategy in the same scenario are arranged according to a pre-fixed field order to form one distribution feature record; the distribution feature records of all industrial operation strategies in all scenarios are combined according to the same field order to form a distribution feature matrix; in order to unify the numerical scale of each field, the maximum and minimum values ​​of each field in the matrix are obtained from all records, the minimum value is subtracted from the current field value, and then divided by the difference between the maximum and minimum values ​​to convert the field to 0 to 1; when the maximum and minimum values ​​are the same, the conversion result of all records of the field is recorded as 0.5.

[0166] Furthermore, the distribution feature matrix that has undergone numerical transformation is input into the isolated forest.

[0167] This embodiment uses 300 isolated trees, with each isolated tree having 256 samples. When the distribution feature matrix contains fewer than 256 records, all records are used to build the tree. When building each node of the isolated tree, one item is randomly selected from the feature fields with different values ​​in the current node, and a split value is randomly selected between the current minimum and maximum values ​​of the field. Records with field values ​​less than the split value are assigned to one child node, and records with field values ​​greater than or equal to the split value are assigned to another child node.

[0168] Continue splitting in the same way until the current node contains only one record, all fields within the node have no value differences, or the maximum depth of the isolated tree is reached.

[0169] For each record in the distribution feature matrix, count the number of splits that record undergoes from the root node to the terminal node in 300 isolated trees, and use the number of splits as the path depth under the corresponding isolated tree.

[0170] If a terminating node contains more than two records, the expected path length is increased according to the path correction rules for incompletely isolated nodes in the isolated forest. The path depths of the record in all isolated trees are summed and then divided by the actual number of isolated trees involved in the calculation to obtain the average path depth of the record.

[0171] Furthermore, the standard expected path length corresponding to the isolated forest is calculated based on the number of tree samples, and then the isolated forest anomaly score transformation is performed using the ratio between the average path depth of the record and the standard expected path length.

[0172] Specifically, the average path depth is divided by the standard expected path length, and the negative value of the result is used as the exponent with base 2 to calculate the anomaly score between 0 and 1. The shorter the path depth required for a record to be isolated, the closer the anomaly score is to 1; the longer the path depth, the closer the anomaly score is to 0. The anomaly scores of all distribution feature records are arranged in the order of the records in the original distribution feature matrix to form an anomaly score sequence.

[0173] In this embodiment, the segmentation position for constructing the segmentation hyperplane is determined based on the abnormal score sequence.

[0174] First, arrange all outlier scores in ascending order. Then, subtract the previous outlier score from the next outlier score to obtain the interval between adjacent outlier scores. Select the interval with the largest value, add the two outlier scores that form the largest interval, and divide the sum by 2. The result is used as the outlier segmentation value. When there are more than two identical largest interval values, the set with the smallest corresponding outlier score is used to determine the outlier segmentation value. When all outlier scores are the same, there is no biased grouping based on outlier scores in the current scene, and the degree of bias is recorded as 0.

[0175] When the outlier scores are not completely identical, the distribution feature records with outlier scores greater than the outlier cutoff value are assigned to the biased candidate group, and the distribution feature records with outlier scores less than or equal to the outlier cutoff value are assigned to the baseline group. The arithmetic mean of each field of the distribution feature matrix is ​​calculated for the biased candidate group and the baseline group respectively, and the center of the biased candidate group and the center of the baseline group are formed by the average values ​​of each field in turn.

[0176] In this embodiment, the segmentation hyperplane is determined based on the two centers mentioned above.

[0177] Specifically, for each field, the field value corresponding to the center of the bias candidate group is subtracted from the field value corresponding to the center of the benchmark group to obtain the segmentation direction component of that field; then, the values ​​of the two centers in the corresponding field are added together and divided by 2 to obtain the midpoint value of that field; for any distribution feature record, the field value of the record is subtracted from the corresponding midpoint value for each field, and the difference is multiplied by the corresponding segmentation direction component, and finally, the products of all fields are added together; records with a sum greater than or equal to 0 are assigned to the bias side, and records with a sum less than 0 are assigned to the benchmark side; all distribution feature records assigned to the bias side constitute the bias feature set.

[0178] Furthermore, the degree of bias is determined by both the degree of bias coverage and the degree of abnormal deviation.

[0179] Specifically, the bias coverage is obtained by dividing the number of records in the bias feature set by the total number of records in the distribution feature matrix; for each record in the bias feature set, the abnormality score of the record is subtracted from the abnormality segmentation value, and then divided by 1 to obtain the abnormality deviation of the record; the abnormality deviation of all records in the bias feature set is added together and then divided by the number of records in the bias feature set to obtain the average abnormality deviation; the bias coverage is multiplied by 0.5, the average abnormality deviation is multiplied by 0.5, and the two results are added together to obtain the bias of the bias feature set; when the bias feature set is empty, the bias is recorded as 0.

[0180] As a preferred embodiment, the preset bias threshold is determined from historical baseline scenario data.

[0181] During threshold calibration, scenario records where alarm states and interlock states were not triggered, and all key operating parameters were within the allowable range of the equipment, were selected from historical operating data. A number of samples were then taken without replacement according to the scenario category and the execution of different industrial operation strategies. Each sampling process involved constructing a distribution feature matrix, calculating an isolated forest, partitioning the hyperplane, and calculating the bias level, using the same methods as the data to be evaluated, to obtain a baseline bias level. This process was repeated 100 times. The 100 baseline bias levels were summed and divided by 100 to obtain the baseline average bias level. The standard deviation of the 100 baseline bias levels was then calculated. The baseline average bias level was added to three times the standard deviation, and the result was determined as the preset bias threshold. If the calculated result exceeded 1, the preset bias threshold was set to 1. This threshold remained unchanged throughout the same round of industrial operation strategy evaluation.

[0182] When the degree of bias in the current bias feature set is less than or equal to the preset bias threshold, no over-threshold bias flag is generated; when the degree of bias is greater than the preset bias threshold, flag generation processing is performed on the bias feature set.

[0183] Specifically, the industrial operation strategy number, scenario category, degree of bias, anomaly segmentation value, number of biased records, and anomaly score corresponding to each biased record are associated with the bias feature set, and the bias state value is set to 1. The above data forms a latent bias identifier.

[0184] Example 4:

[0185] S3 includes: if the acquired latent bias identifiers exceed the pre-established bias judgment threshold, then extract the bias feature weights of the scene distribution set based on the latent bias identifiers; divide the bias feature weights into a subset for removal, and perform feature removal on the scene distribution set through the subset for removal to obtain a dimensionality-reduced scene distribution set; and perform resampling processing by fusing the synthetic sample set generated for the dimensionality-reduced scene distribution set with the dimensionality-reduced scene distribution set to obtain an unbiased scene distribution set.

[0186] In this embodiment, the implicit bias identifier generated by the aforementioned anomaly detection process is first read, and the scenario data that needs to be corrected for bias is determined based on the industrial operation strategy number, scenario category, bias side record number, anomaly score and bias degree recorded therein. When the current bias degree is greater than the pre-established bias judgment threshold, the bias feature weights of each scenario feature corresponding to the scenario are calculated. When the current bias degree is less than or equal to the bias judgment threshold, the original features of the scenario are maintained and feature removal is not performed.

[0187] Furthermore, the bias determination threshold adopts the value determined in the aforementioned bias detection stage.

[0188] Specifically, when determining the threshold, records where alarm states and interlock states are not triggered and key operating parameters are within the equipment's permissible range are extracted from historical operation records. Equal numbers of samples are executed for each industrial operation strategy according to scenario categories, and the same bias calculation process as in the formal evaluation is repeated 100 times. The bias values ​​obtained from the 100 calculations are summed and divided by 100 to obtain the baseline bias average. The standard deviation of the 100 bias values ​​relative to this average is then calculated. The baseline bias average is added to three times the standard deviation to obtain the bias judgment threshold. When the calculation result is greater than 1, the bias judgment threshold is set to 1. Once the threshold is determined, it remains unchanged in the same round of evaluation.

[0189] In addition, the bias feature weights are calculated based on the bias-side records and the reference-side records indicated by the implicit bias identifiers.

[0190] Specifically, for any scene feature, first multiply the feature value of each record on the bias side by its anomaly score, add all the products together, and then divide by the sum of the anomaly scores of all records on the bias side to obtain the weighted average of the feature on the bias side; when the sum of the anomaly scores is equal to 0, directly use the arithmetic mean of the feature on the bias side; then calculate the arithmetic mean of the feature on the benchmark side, and use the weighted average of the bias side to subtract the arithmetic mean of the benchmark side, taking the absolute value of the difference.

[0191] To eliminate the influence of different feature dimensions, the standard deviation of all records in the current scene on this feature is calculated, and the absolute difference is divided by the standard deviation. The result is used as the bias contribution value of this feature.

[0192] When the standard deviation is equal to 0 and the biased weighted average is the same as the baseline arithmetic mean, the bias contribution value is recorded as 0; when the standard deviation is equal to 0 and the two are not the same, the bias contribution value is recorded as 1. After obtaining the bias contribution values ​​of all scene features in the same way, the bias contribution values ​​are added together, and then the bias contribution value of each feature is divided by the sum of the bias contribution values ​​to obtain the bias feature weight of that feature. The sum of the bias feature weights is 1.

[0193] When the sum of all bias contribution values ​​equals 0, it indicates that the current implicit bias originates from the distribution of the number of scene records rather than the difference in scene feature values. In this case, the feature subset to be removed is empty, all scene features are retained, and the process directly proceeds to resampling. When the sum of bias contribution values ​​is greater than 0, the scene features that can be removed are arranged in descending order of bias feature weight. Industrial operation strategy identifiers, scene category identifiers, operation record associated fields, and evaluation result fields directly used in subsequent safety dimensions, efficiency dimensions, and other evaluation dimensions are not included in the scope of removal.

[0194] In this embodiment, the feature subset to be removed is determined item by item according to the arrangement result of the bias feature weights.

[0195] First, temporarily delete the top-ranked removable feature. Reconstruct the distribution feature data of the current scene using the remaining features, and recalculate the bias according to the aforementioned isolated forest anomaly detection, anomaly score classification, and bias calculation process. If the recalculated bias is still greater than the bias judgment threshold, delete the next removable feature in the weight ranking and recalculate. Stop adding removable features when the recalculated bias is less than or equal to the bias judgment threshold for the first time. The features deleted for the first time up to the last feature corresponding to this calculation constitute the subset of removable features.

[0196] Therefore, the number of features removed is directly reduced to within the bias determination threshold as the stopping condition.

[0197] Furthermore, after the feature subset is determined, the feature fields corresponding to the subset are deleted from each record in the scene distribution set, while retaining the scene category, industrial operation strategy affiliation, operation record association, and scene features that have not entered the feature subset. The processed records form a dimensionality-reduced scene distribution set.

[0198] Subsequently, the number of valid records corresponding to each industrial operation strategy within the same scenario was counted to determine the number of resampling operations.

[0199] For any scenario, compare the number of valid records for each industrial operation strategy and determine the number of valid records with the largest value as the target sample number for that scenario. If the number of valid records for an industrial operation strategy is equal to the target sample number, retain all of its records. If the number of valid records is less than the target sample number, subtract the number of existing valid records for that industrial operation strategy from the target sample number, and use the difference as the number of synthetic samples to be generated.

[0200] In addition, before generating synthetic samples, a uniform scaling process is performed on continuous features in the dimensionality reduction scene distribution set.

[0201] Specifically, for each feature, the minimum value of that feature in the current scenario is subtracted from the current feature value, and then divided by the difference between the maximum and minimum values ​​to obtain a distance calculation value between 0 and 1. When the maximum and minimum values ​​are the same, the distance calculation value is uniformly recorded as 0.5. Subsequently, Euclidean distance is calculated only between original records of the same industrial operation strategy and the same scenario, and the nearest other original record is determined according to the distance value.

[0202] As a preferred embodiment, the synthetic samples are generated using an interpolation method.

[0203] Specifically, assuming the number of synthetic samples required by the current industrial operation strategy is the actual calculated number, the existing original records are arranged in ascending order according to their Euclidean distance from the scene cluster center, and the original records are selected cyclically as the reference records in the order of arrangement; for the first synthetic sample to be generated, the first reference record in the sorted list and its nearest neighbor record are taken; subsequent synthetic samples use the subsequent reference records in sequence, and after all the original records are used for one round, the cycle starts again from the first record.

[0204] For each synthesized sample generated, the interpolation ratio is obtained by dividing the generation sequence number of the current synthesized sample by the total number of synthesized samples to be generated plus 1.

[0205] For each dimension reduction feature, the feature value of the nearest neighbor record is subtracted from the corresponding feature value of the benchmark record. The difference is multiplied by the interpolation ratio and then added to the feature value of the benchmark record to obtain the composite value of that feature. After all dimension reduction features are calculated in sequence, one composite sample is formed. The above calculation is repeated until the number of composite samples reaches the difference between the current industrial operation strategy and the target number of samples.

[0206] When a certain industrial operation strategy has only one original valid record in the current scenario, the record is used as the baseline record, and the cluster center corresponding to the scenario is used as the interpolation endpoint. The composite value of each dimension reduction feature is calculated according to the same interpolation ratio.

[0207] The resulting feature values ​​are located between the original records and the cluster centers, without using extrapolation calculations that exceed the range of values ​​at both ends.

[0208] Furthermore, after generating synthetic samples for each industrial operation strategy, all newly added records will be combined into a synthetic sample set.

[0209] Specifically, each synthetic sample inherits the industrial operation strategy affiliation and scenario category of the corresponding baseline record, and reassigns the operation record number. At the same time, a synthetic sample source field is set to distinguish it from the original historical operation record. Subsequently, the synthetic sample set and the dimensionality reduction scenario distribution set are merged in the same field order so that the number of samples corresponding to each industrial operation strategy in the same scenario reaches the target number of samples for that scenario.

[0210] After fusion, the distribution density of each industrial operation strategy in each scenario is recalculated, and the bias of the fused data is recalculated according to the aforementioned bias detection process.

[0211] Specifically, when the recalculated bias is less than or equal to the bias determination threshold, the current fused data is determined as an unbiased scene distribution set; when the bias is still greater than the bias determination threshold, the bias feature weights of the remaining features are calculated based on the regenerated implicit bias identifiers, and features are added and removed in order of weight. Dimensionality reduction, synthetic sample generation and fusion processing are then re-executed until the bias is less than or equal to the bias determination threshold.

[0212] Example 5:

[0213] S4 involves extracting safety, efficiency, and comfort features from an unbiased scene distribution set to construct a multi-dimensional feature matrix; using a random forest algorithm to obtain feature splitting gain for the multi-dimensional feature matrix, and calculating the Gini impurity reduction based on the feature splitting gain to determine the initial feature importance; if the initial feature importance meets the preset significance criteria, cross-validation is performed on the initial feature importance using out-of-bag data samples to obtain the mean prediction error; and the initial feature importance is normalized and mapped based on the mean prediction error to determine the multi-dimensional weight values.

[0214] In this embodiment, operation records in the unbiased scene distribution set are used as evaluation samples, and safety dimension features, efficiency dimension features, and comfort dimension features are extracted based on the same set of feature fields.

[0215] Specifically, the safety dimension features include the number of alarm triggers, the number of interlock triggers, the duration of critical operating parameters exceeding limits, and the status of safety constraint satisfaction; the efficiency dimension features include task completion time, the amount of task completed per unit time, energy consumption, and material consumption; the comfort dimension features include equipment vibration amplitude, the adjustment amount of the actuator at adjacent sampling times, and the continuous change amplitude of the control output; each feature is corresponding to the operation record number and scenario category, so that the same operation record forms a complete feature record under the three evaluation dimensions.

[0216] For continuous features, first calculate the arithmetic mean and standard deviation of the feature in the unbiased scene distribution set. Subtract the arithmetic mean from the original feature value of the current record, and then divide the difference by the standard deviation to obtain the standardized feature value. When the standard deviation is equal to 0, the standardized value of the feature is uniformly recorded as 0.

[0217] For a state where the safety constraints are satisfied, it is recorded as 1 if the safety constraints are satisfied and as 0 if the safety constraints are not satisfied; then, the feature values ​​are arranged in the order of the pre-fixed safety dimension field, efficiency dimension field, and comfort dimension field, and a multi-dimensional feature matrix is ​​formed by all operation records.

[0218] During random forest training, the historical execution results corresponding to each operation record are used as supervision labels. When an operation record satisfies the safety constraints and completes the predetermined operation task, the label is recorded as 1; when the safety constraints are not met or the predetermined operation task is not completed, the label is recorded as 0.

[0219] This embodiment uses 500 decision trees. Each decision tree obtains the same number of tree construction samples as the original training samples from the multidimensional feature matrix using sampling with replacement. Samples not selected by the current decision tree are retained as out-of-bag data for that tree.

[0220] For any node to be split in the decision tree, first count the number of samples with labels 1 and 0 in the node, divide the number of samples of each class by the total number of samples in the node to get the corresponding proportions, square the two proportions respectively and add them together, and then subtract the sum from 1 to get the Gini impurity of the node before splitting.

[0221] Subsequently, the sample is divided into two sub-nodes based on the candidate split positions of the candidate features, and the Gini impurity of the two sub-nodes is calculated separately using the same method.

[0222] The Gini impurity after splitting is weighted according to the sample size of the two child nodes.

[0223] Specifically, the number of samples in the first child node is divided by the number of samples in the parent node, and the resulting ratio is multiplied by the Gini impurity of the first child node; the number of samples in the second child node is divided by the number of samples in the parent node, and the resulting ratio is multiplied by the Gini impurity of the second child node, and the two products are added together; the weighted Gini impurity after splitting is subtracted from the Gini impurity before splitting to obtain the feature splitting gain corresponding to the current candidate splitting position; after calculating all candidate features and candidate splitting positions for the current node, the splitting position with the largest feature splitting gain value is selected to perform node splitting.

[0224] Furthermore, after each decision tree is trained, the reduction in Gini impurity of the features is calculated using the feature split gain corresponding to each actual split node.

[0225] For a given feature, first multiply the feature split gain generated by that feature at one actual split node by the proportion of the number of samples at that node to the number of samples in the current decision tree. Then, sum the calculation results corresponding to all actual split nodes of the entire decision tree to obtain the reduction in Gini impurity of that feature in the current decision tree. After completing 500 decision trees, sum the 500 reductions in Gini impurity corresponding to the same feature and divide by 500 to obtain the average reduction in Gini impurity of that feature.

[0226] The total importance is obtained by summing the average Gini impurity reduction of all evaluation features. Then, the total importance is obtained by dividing the average Gini impurity reduction of each evaluation feature by the total importance. This makes the sum of the importance of all initial features equal to 1. When the total importance equals 0, the random forest is rebuilt.

[0227] In this embodiment, the preset saliency determination criteria are determined by the randomization result of the training data labels.

[0228] Specifically, before formal training of the random forest, the values ​​of each evaluation feature are kept unchanged. The order of the supervision labels is randomly shuffled, and 500 decision trees are retrained. The initial feature importance of each evaluation feature is calculated in the same way. The randomization process is executed 100 times, and the initial feature importance with the largest value in each calculation is recorded. The 100 largest initial feature importances are added together and divided by 100 to obtain the average importance under random conditions. The standard deviation of the 100 largest initial feature importances is then calculated. The average importance is added to three times the standard deviation as the significance threshold. When the initial feature importance obtained from formal training is greater than or equal to this threshold, the corresponding feature is determined to meet the significance condition. Features with importance less than this threshold are not included in the subsequent multidimensional weight calculation.

[0229] For features that meet the significance criteria, cross-validation is performed using the out-of-bag data corresponding to each decision tree.

[0230] First, input the out-of-bag samples corresponding to a certain decision tree directly into the decision tree, count the number of samples whose predicted labels are inconsistent with the actual labels, and divide the number of incorrectly predicted samples by the total number of out-of-bag samples of the tree to obtain the basic prediction error.

[0231] Subsequently, only the order of the values ​​of the feature to be validated in the out-of-bag samples is shuffled, while other feature values ​​and actual labels remain unchanged. The processed out-of-bag samples are then input into the same decision tree again, and the perturbed prediction error is calculated by dividing the number of incorrectly predicted samples by the total number of out-of-bag samples. The perturbed prediction error is then subtracted from the basic prediction error to obtain the error increment corresponding to this feature in this validation. To eliminate the difference in results caused by a single random permutation, the same feature is subjected to 10 shuffled value orders on the same out-of-bag dataset. The 10 error increments are added together and divided by 10 to obtain the validation error of this feature on the current decision tree.

[0232] Furthermore, after completing all decision trees with out-of-bag samples in sequence, the validation errors of the same feature on each decision tree are summed, and then divided by the actual number of decision trees involved in the validation to obtain the mean prediction error corresponding to that feature.

[0233] The mean prediction error is denoted as 0 when it is less than 0. This value is used to characterize the actual change in the prediction performance of the random forest after shuffling the corresponding features.

[0234] As a preferred embodiment, the initial feature importance is normalized and mapped based on the mean of the prediction error.

[0235] Specifically, the mean prediction errors of all features that meet the significance criteria are first summed, and then the mean prediction error of a single feature is divided by the sum to obtain the validation coefficient of that feature. When the sum of the mean prediction errors of all features equals 0, the random forest is retrained. Then, the initial feature importance of each feature is multiplied by the corresponding validation coefficient, and the result is used as the corrected importance of that feature.

[0236] Finally, the adjusted importance of all features within the safety dimension is summed to obtain the safety dimension importance; the adjusted importance of all features within the efficiency dimension is summed to obtain the efficiency dimension importance; and the adjusted importance of all features within the comfort dimension is summed to obtain the comfort dimension importance. After summing the importance of the three dimensions, each dimension importance is divided by the sum of the three dimensions' importance to obtain the safety dimension weight, efficiency dimension weight, and comfort dimension weight. The sum of the three dimension weights is 1, and this sum is used as the multidimensional weight value for subsequent multidimensional comprehensive scoring calculations.

[0237] Example 6:

[0238] S5 includes obtaining operation records corresponding to industrial operation strategies within the unbiased scene distribution set; parsing the operation records to obtain operation action sequences and equipment status responses; calculating the strategy execution deviation based on the operation action sequences and equipment status responses; using multi-dimensional weight values ​​to weight the strategy execution deviation to obtain a weighted aggregation matrix; determining whether the weighted aggregation matrix meets the preset non-empty judgment condition; if the weighted aggregation matrix meets the preset non-empty judgment condition, performing extreme value normalization processing on the weighted aggregation matrix to obtain a multi-dimensional comprehensive score, thus completing the weighted calculation and normalization processing of operation records corresponding to different industrial operation strategies within the unbiased scene distribution set to obtain a multi-dimensional comprehensive score.

[0239] In this embodiment, the corresponding operation records are read from the unbiased scene distribution set according to the industrial operation strategy number, scene category and operation record number, and the execution process of each operation record is restored according to the timestamp.

[0240] For each strategy execution, the control command issuance time, controlled object, target setpoint, actuator action amount and action duration are arranged in chronological order to form an operation action sequence; at the same time, the equipment load, process temperature, process pressure, medium flow rate, actuator feedback value, vibration amplitude, alarm status and interlock status generated after the control command is issued are read to form an equipment status response.

[0241] The time interval between two adjacent control commands is used as the response interval of the previous control command, and the last control command takes the end time of the operation record as the end of the response interval, so that each operation action corresponds to a specific equipment status response.

[0242] State parameters with inconsistent sampling periods are time-aligned at 1-second intervals. When there are more than two sampled values ​​in the same second, the arithmetic mean is taken to ensure that the actions and responses of different industrial operation strategies are compared using the same time scale.

[0243] In this embodiment, the strategy execution deviation is calculated based on the baseline execution characteristics of the same scenario in the unbiased scenario distribution set.

[0244] For each scenario, data of the corresponding action type is first extracted from the historical records of satisfying safety constraints and completing the target operation task. The target value of the action, the duration of the action, and the arithmetic mean of the status response parameters of each device are calculated respectively. The results are used as the baseline execution value of the corresponding action in that scenario.

[0245] When calculating the deviation of a specific operation record, for each operation parameter, the actual value is subtracted from the corresponding baseline execution value, and the absolute value is taken. This absolute value is then divided by the difference between the maximum and minimum values ​​of that parameter in all valid records of the current scenario. A result exceeding 1 is counted as 1. When the maximum and minimum values ​​are the same, the deviation is recorded as 0 if the actual value is the same as the baseline execution value, and as 1 if the actual value is different from the baseline execution value. The same processing method is used for device status response: first, the arithmetic mean of the status parameters within the corresponding response interval is calculated, then compared with the baseline response value of the scenario, and scaled according to the difference between the maximum and minimum values ​​of the corresponding status parameters in the current scenario, resulting in a response deviation within the range of 0 to 1.

[0246] Furthermore, after completing the individual deviation calculation, the deviation is executed according to the aforementioned evaluation dimension aggregation strategy.

[0247] Specifically, the safety dimension includes deviations corresponding to alarm triggering, interlock triggering, critical status parameter out-of-bounds, and safety-related equipment responses; the efficiency dimension includes deviations corresponding to task completion time, action duration, task completion amount per unit time, energy consumption, and material consumption; the comfort dimension includes deviations corresponding to vibration amplitude, changes in adjacent actions of the actuator, and continuous changes in control output; for a given evaluation dimension, all valid individual deviations are summed and then divided by the number of valid individual deviations in that dimension to obtain the strategy execution deviation of the operation record on the corresponding evaluation dimension.

[0248] Therefore, each valid operation record generates a safety execution deviation, an efficiency execution deviation, and a comfort execution deviation, with all three deviations limited to the range of 0 to 1.

[0249] Furthermore, the safety dimension weight, efficiency dimension weight, and comfort dimension weight determined in Example 5 are used to weight the above strategy execution deviation.

[0250] Specifically, the safety execution deviation is multiplied by the safety dimension weight, the efficiency execution deviation is multiplied by the efficiency dimension weight, and the comfort execution deviation is multiplied by the comfort dimension weight. The three products are then added together to obtain the comprehensive weighted deviation of the operation record.

[0251] Simultaneously, the three products are used as the weighted results for the safety dimension, efficiency dimension, and comfort dimension, respectively. Following the order of the operation record numbers, the industrial operation strategy number, scenario category, safety dimension weighted result, efficiency dimension weighted result, comfort dimension weighted result, and comprehensive weighted deviation are sequentially written into the corresponding data rows, forming a weighted aggregation matrix from all data rows. Each row in the matrix corresponds to one operation record, and the position of each evaluation field remains consistent throughout the same round of evaluation.

[0252] In addition, a non-empty check is performed after the weighted aggregation matrix is ​​formed.

[0253] Specifically, the non-empty determination conditions include: the matrix has at least one row of valid operation records; each row has a weighted result for the safety dimension, a weighted result for the efficiency dimension, a weighted result for the comfort dimension, and a comprehensive weighted deviation; the aforementioned four values ​​are all finite values ​​that can participate in numerical calculations; each record can be associated with a specific industrial operation strategy number and scenario category; when all the above conditions are met, the weighted aggregation matrix is ​​determined to satisfy the non-empty determination conditions; when any evaluation dimension cannot obtain the corresponding strategy execution deviation due to the absence of the original operation record, the operation record is not included in the weighted aggregation matrix, and the corresponding data validity check process is returned, without replacing the missing deviation with 0.

[0254] Specifically, when the weighted aggregation matrix satisfies the non-empty criterion, extreme value normalization is performed on the comprehensive weighted deviation.

[0255] First, the maximum and minimum values ​​of the overall weighted deviation are obtained from all valid data rows in the matrix. Since a smaller overall weighted deviation value indicates a smaller deviation between the operation action and equipment response and the baseline execution characteristics, reverse extreme value normalization is adopted: For any operation record, the overall weighted deviation of the operation record is subtracted from the maximum value of the overall weighted deviation, and then the difference is divided by the difference between the maximum and minimum values ​​of the overall weighted deviation to obtain a normalization result in the range of 0 to 1. Records with an overall weighted deviation equal to the minimum value are assigned 1, records with an overall weighted deviation equal to the maximum value are assigned 0, and the remaining records are assigned corresponding values ​​based on their actual position between the maximum and minimum values.

[0256] When all the comprehensive weighted deviations in the weighted aggregation matrix are the same, the difference between the maximum and minimum values ​​is equal to 0. No division operation is performed, and the normalization result of all valid operation records is uniformly determined to be 0.5, thereby avoiding the calculation result with a division of 0.

[0257] Finally, the normalized result is multiplied by 100 to obtain a multidimensional comprehensive score in the range of 0 to 100, and the multidimensional comprehensive score is associated with and stored with the corresponding industrial operation strategy number, scenario category and operation record number.

[0258] Example 7:

[0259] S6 includes comparing the multidimensional comprehensive score according to the pre-established benchmark verification rules to obtain the score comparison result; if the score comparison result meets the benchmark verification rules, the interaction frequency characteristics and historical credit rating of the operation record are extracted; the interaction frequency characteristics and historical credit rating are classified and processed to construct a capability assessment matrix, and the first true capability quantification value corresponding to the operation record is determined according to the capability assessment matrix.

[0260] In this embodiment, the multidimensional comprehensive score obtained in Embodiment 6 is read first, and the corresponding benchmark verification rule is called according to the scenario to which the operation record belongs; the benchmark verification rule is determined by the historical valid operation records before the start of this round of evaluation.

[0261] Specifically, records that meet safety constraints, complete predetermined operation tasks, and have complete evaluation fields are selected from the historical records of the same scenario. Historical multidimensional comprehensive scores are calculated using the same scoring method as the records to be evaluated. All historical multidimensional comprehensive scores are summed and divided by the number of valid records to obtain the historical average score. The squares of the differences between each historical multidimensional comprehensive score and the historical average score are then calculated. All squares are summed and divided by the number of valid records, and the square root of the result is taken to obtain the historical standard deviation. The historical average score minus three times the historical standard deviation is used as the lower limit for score verification; a result less than 0 is taken as 0. The upper limit for score verification is 100. When the multidimensional comprehensive score of the operation record to be evaluated is between the lower limit and 100, the score comparison result is determined to conform to the benchmark verification rules; when it is below the lower limit, the score comparison result is determined to not conform to the benchmark verification rules.

[0262] Furthermore, after the scoring comparison results meet the benchmark verification rules, interaction frequency features are extracted from the corresponding operation records.

[0263] Specifically, a valid interaction is defined as the issuance of a control command and the receipt of valid device status feedback within the corresponding response interval. The number of valid interactions in a single operation is counted. The duration of the operation is obtained by subtracting the time of the first control command issuance from the time of the last device status response, and then converted to minutes. The interaction frequency of the operation is obtained by dividing the number of valid interactions by the number of minutes of operation. When the operation duration is 30 seconds, the converted duration is 0.5 minutes, and no integer rounding is performed.

[0264] Therefore, the interaction frequency reflects the number of effective action responses formed between industrial operation strategies and equipment status per unit time.

[0265] Furthermore, the classification boundaries for interaction frequency are also determined by valid historical records.

[0266] Specifically, the interaction frequency is calculated for each historical baseline record in the current scenario. The average interaction frequency is obtained by summing all interaction frequencies and dividing by the number of records. The standard deviation of the interaction frequency is obtained according to the aforementioned standard deviation calculation method. The first classification boundary is the average interaction frequency minus one standard deviation, and the result is taken as 0 if it is less than 0. The second classification boundary is the average interaction frequency plus one standard deviation. If the interaction frequency of the record to be evaluated is less than the first classification boundary, it is classified into the low-frequency category. If it is greater than or equal to the first classification boundary and less than or equal to the second classification boundary, it is classified into the baseline frequency category. If it is greater than the second classification boundary, it is classified into the high-frequency category. The classification boundaries remain unchanged in the same round of evaluation.

[0267] As a preferred implementation, historical credit ratings are calculated based on the historical records of the same industrial operating strategy that were completed before the current operating record occurred.

[0268] First, count the total number of historical records. Then, count the number of records that simultaneously meet the safety constraints and complete the predetermined operation task. Divide this number by the total number of historical records and multiply by 100 to obtain the historical valid completion score. At the same time, count the number of historical records that trigger interlocks or cause task termination. Divide this number by the total number of historical records to obtain the historical anomaly ratio. Subtract the historical anomaly ratio from 1 and multiply by 100 to obtain the historical anomaly constraint score. Multiply the historical valid completion score by 0.7 and the historical anomaly constraint score by 0.3, and then add the two results to obtain the historical credit score.

[0269] In addition, a historical credit score of 90 or higher is classified as Level 1 credit, a historical credit score of 80 or higher but less than 90 is classified as Level 2 credit, and a historical credit score of less than 80 is classified as Level 3 credit.

[0270] The credit rating boundary is fixed during the evaluation rule establishment phase and does not change with the evaluation results in the same round of evaluation; when there is no historical execution record for the current industrial operation strategy, no historical credit rating is generated, and the operation record is not included in the first real capability quantification process.

[0271] Furthermore, after completing the interaction frequency classification and historical credit rating, a capability assessment matrix is ​​constructed. The capability assessment matrix uses low-frequency category, benchmark frequency category and high-frequency category as three row categories, and first-level credit, second-level credit and third-level credit as three column categories, forming a total of nine matrix units. The capability correction value of each matrix unit is determined based on historical benchmark records.

[0272] Specifically, historical benchmark records are assigned to corresponding matrix units according to their interaction frequency category and historical credit rating. The sum of all historical multidimensional comprehensive scores in the unit is divided by the number of valid records to obtain the historical average score of the unit. The historical average score is then divided by 100 to obtain the capability correction value of the unit.

[0273] When a certain matrix unit has no valid historical records, all historical records from the corresponding units of other frequency categories under the same credit rating are merged and calculated. The sum of the merged multidimensional comprehensive scores is divided by the number of merged records and then divided by 100 to obtain the capability correction value for the missing unit. When there are no valid historical records under the same credit rating, the multidimensional comprehensive scores of all historical benchmark records are added together and then divided by the total number of historical benchmark records and then divided by 100 to obtain the capability correction value for the unit. In this way, all nine capability assessment units have definite values.

[0274] As a preferred embodiment, for the current operation record whose scoring comparison results conform to the benchmark verification rules, the corresponding row of the capability assessment matrix is ​​determined according to its interaction frequency category, and the corresponding column is determined according to the historical credit rating. The capability correction value at the intersection of the row and column is read. The multidimensional comprehensive score of the operation record is multiplied by the corresponding capability correction value to obtain the first true capability quantification value. When the calculation result is greater than 100, it is taken as 100; when the calculation result is less than 0, it is taken as 0.

[0275] The obtained first true capability quantification value is correlated with the industrial operation strategy number, scenario category and operation record number, and used for subsequent capability sorting and alignment processing of different industrial operation strategies.

[0276] Example 8:

[0277] S7 includes obtaining a multi-dimensional comprehensive score that does not conform to the benchmark verification rules, extracting the violation frequency and deviation range corresponding to the multi-dimensional comprehensive score; classifying the violation frequency and deviation range to obtain a risk level; matching penalty weights and attenuation factors according to the risk level, and using the penalty weights and attenuation factors to calculate credit damage on the multi-dimensional comprehensive score to obtain a corrected score; if the corrected score is lower than the preset quantitative base, then penalizing the corrected score to obtain a second true capability quantitative value.

[0278] In this embodiment, for operation records that are determined by Embodiment 7 to be inconsistent with the benchmark verification rules, the multi-dimensional comprehensive score corresponding to the operation record is read, and the verification units formed by the same industrial operation strategy in this operation are retrieved according to the operation record number.

[0279] Specifically, the verification unit is divided according to the operation action sequence parsed in S6. Each verification unit is formed after each operation action is executed and the corresponding device status response is obtained. Each verification unit calculates the corresponding unit score according to the evaluation dimension, dimension weight and normalization rule determined in S5, and then compares it with the score verification lower limit corresponding to the scenario.

[0280] Specifically, if a unit's score is lower than the lower limit of the score verification, that unit is counted as one violation; if a unit's score is greater than or equal to the lower limit of the score verification, it is not counted as a violation. The number of violating verification units in the current operation record is counted, and the number of violating verification units is divided by the total number of valid verification units. The result is used as the violation frequency. If the total number of valid verification units is 20, and the scores of 5 of them are lower than the lower limit of the score verification, the violation frequency is 5 divided by 20, and the result is 0.25. Operation records with no valid verification units are not included in subsequent risk calculations.

[0281] In this embodiment, the deviation range is determined based on the difference between each violation verification unit and the lower limit of the scoring verification.

[0282] Specifically, for a single violation verification unit, first subtract the actual score of the verification unit from the lower limit of the scoring verification, then divide the difference by the lower limit of the scoring verification to obtain the deviation value of the verification unit; if the result exceeds 1, it is recorded as 1. The deviation values ​​of all violation verification units are calculated sequentially, all deviation values ​​are added together, and then divided by the number of violation verification units. The result is used as the deviation range of the current operation record.

[0283] Therefore, the frequency of violations indicates the proportion of occurrences that do not comply with the benchmark verification rules, and the deviation range indicates the degree of deviation of the actual score from the lower limit of the score verification.

[0284] Furthermore, after obtaining the frequency of violations and the extent of deviation, the two sets of data are classified and processed.

[0285] Specifically, the frequency of violations is first multiplied by 0.5, and the deviation is multiplied by 0.5. The two results are then added together to obtain the risk discrimination value of the current operation record. The judgment boundary used to distinguish the risk level is determined by historical operation records that do not meet the benchmark verification rules before the formal evaluation. The risk discrimination value is calculated for each historical record using the same method. The sum of all historical risk discrimination values ​​is then divided by the number of historical records to obtain the average risk discrimination value. The square of the difference between each historical risk discrimination value and the average risk discrimination value is then calculated. The sum of all squares is then divided by the number of historical records, and the square root of the result is taken to obtain the standard deviation of the risk discrimination.

[0286] In addition, the average risk assessment value is used as the first risk boundary, and the sum of the average risk assessment value and the standard deviation of the risk assessment value is used as the second risk boundary. If the second risk boundary is greater than 1, it is determined as 1.

[0287] When the current risk threshold is less than or equal to the first risk boundary, it is determined to be a Level 1 risk; when it is greater than the first risk boundary and less than or equal to the second risk boundary, it is determined to be a Level 2 risk; when it is greater than the second risk boundary, it is determined to be a Level 3 risk; the above two risk boundaries remain unchanged in the same round of evaluation.

[0288] Furthermore, once the risk level is determined, the penalty weight and decay factor corresponding to the risk level are applied.

[0289] In this embodiment, the penalty weight for Level 1 risk is 0.25 and the attenuation factor is 0.95; the penalty weight for Level 2 risk is 0.50 and the attenuation factor is 0.85; and the penalty weight for Level 3 risk is 0.75 and the attenuation factor is 0.70. The above parameters are written into the benchmark parameter table before the evaluation begins and are not modified based on the current scoring results during the same round of evaluation.

[0290] In addition, when calculating credit loss, the frequency of violations and the deviation range are added together and then divided by 2 to obtain the overall violation level of the current operation record; the overall violation level is then multiplied by the penalty weight of the corresponding risk level to obtain the credit loss ratio; the credit loss ratio is subtracted from 1, the result is multiplied by the current multidimensional comprehensive score, and the result is then multiplied by the attenuation factor corresponding to the risk level. The resulting value is used as the correction score; if the correction score is less than 0, it is set to 0, and if it is greater than 100, it is set to 100.

[0291] As a preferred embodiment, the preset quantization base is determined by historical operation records that do not conform to the benchmark verification rules.

[0292] Specifically, the historical records are calculated with corrected scores for each record according to the aforementioned risk classification and credit impairment process. All historical corrected scores are then summed and divided by the number of records to obtain the average historical corrected score. Next, the square of the difference between each historical corrected score and the average historical corrected score is calculated. All squares are summed and divided by the number of historical records. The square root of the result is then taken to obtain the standard deviation of the historical corrected score. The average historical corrected score minus the standard deviation is used as the quantification base. If the result is less than 0, the quantification base is set to 0. Once determined, the quantification base remains unchanged throughout this round of evaluation.

[0293] Specifically, when the current correction score is greater than or equal to the quantification base, the correction score is used as the second true ability quantification value, and no additional penalty is applied.

[0294] When the current correction score is lower than the quantization base, the correction score is first subtracted from the quantization base, and then the difference is divided by the quantization base to obtain the quantization gap ratio; when the quantization base is equal to 0, the quantization gap ratio is recorded as 0.

[0295] The corrected score is then multiplied by the quantitative gap ratio, and then multiplied by the penalty weight corresponding to the current risk level. The result is used as the additional penalty value. The second true capability quantitative value is obtained by subtracting the corrected score from the additional penalty value. If the result is less than 0, it is set to 0.

[0296] Through the above processing, the multi-dimensional comprehensive score that does not meet the benchmark verification rules is converted into violation frequency, deviation range, risk level and correction score in sequence. When the correction score is lower than the quantification base, a penalty calculation is further performed to finally obtain the second real capability quantification value corresponding to the degree of violation of the operation record. The second real capability quantification value is associated with the industrial operation strategy number, scenario category and operation record number for subsequent sorting and alignment processing.

[0297] Example 9:

[0298] S8 includes parsing the industrial operation records corresponding to the first and second true capability quantification values ​​to obtain an initial action sequence; performing matching processing based on the initial action sequence to obtain an alignment feature sequence; extracting the response delay of the alignment feature sequence to construct a multi-dimensional feature matrix; performing dimensionality reduction processing on the multi-dimensional feature matrix to obtain the core deviation amplitude; arranging the industrial operation records according to the core deviation amplitude to obtain a target sorting sequence; and extracting performance scores based on the target sorting sequence to construct a target evaluation benchmark for evaluating the performance of industrial operation strategies.

[0299] In this embodiment, the industrial operation records corresponding to the first and second real capability quantification values ​​are first read, and the complete execution process of each operation record is restored based on the industrial operation strategy number, scenario category and operation record number.

[0300] Based on the actual issuance time of the control command, the controlled object, action type, target setpoint, actuator action quantity, and corresponding equipment status response are sorted; when there are more than two control actions at the same sampling time, they are arranged in ascending order of the controlled object number.

[0301] After the above processing, each industrial operation record forms an initial sequence of actions with a defined time order.

[0302] When matching the initial action sequence, the industrial operation record with the largest real capability quantification value within the same scene category is used as the benchmark record; when there are records with the same real capability quantification value, the record with the earliest operation start time is used as the benchmark record; the remaining industrial operation records are matched with the initial action sequence of the benchmark record.

[0303] For any two actions to be matched, first compare the action types. If the action types are the same, the difference in action types is recorded as 0; if they are different, it is recorded as 1. Then calculate the difference in target setpoints and the difference in the amount of action of the actuator. The difference in target setpoints is obtained by dividing the absolute value of the difference between the target setpoints of the two actions by the difference between the maximum and minimum values ​​of the control parameter in the current scenario. The difference in the amount of action of the actuator is calculated in the same way. If the maximum and minimum values ​​are the same, the difference between the two values ​​to be compared is recorded as 0; if they are not equal, the difference is recorded as 1. Add the difference in action types, the difference in target setpoints, and the difference in the amount of action of the actuator, and divide by 3 to obtain the matching cost between the two actions.

[0304] In this embodiment, sequence matching is performed using a node-by-node cumulative cost method.

[0305] Specifically, starting from the initial action of the two action sequences, the cumulative costs corresponding to the previous baseline action and the current action to be matched, the current baseline action and the previous action to be matched, and the previous baseline action and the previous action to be matched are obtained for the current action node. The smallest value is selected and added to the matching cost of the current two actions to obtain the cumulative matching cost of the current node.

[0306] After all action nodes have been processed, trace back from the end node to the starting node along the path with the minimum cumulative cost to obtain the action correspondence between the two action sequences; reorganize the action and response data in the record to be evaluated according to the arrangement position of the baseline action sequence to form an alignment feature sequence.

[0307] Therefore, different industrial operation strategies are compared in the same scenario by comparing subsequent responses based on the same action position.

[0308] Furthermore, after the alignment feature sequence is formed, the device response delay is extracted for each action position.

[0309] Specifically, the device status parameters of five consecutive valid sampling points before the action control command is issued are first read. The five sampled values ​​are added together and divided by 5 to obtain the status baseline value before the action is executed. The response judgment boundary is determined using historical steady-state records. For the same device status parameter, continuous sampled values ​​are extracted from the historical steady-state interval where there is no control action, no alarm is triggered, and no interlock is triggered. The absolute value of the difference between two adjacent sampled values ​​is calculated in turn. All absolute differences are added together and divided by the number of differences to obtain the average steady-state change. Then, the standard deviation of each absolute difference relative to the average steady-state change is calculated. The average steady-state change is added to three times the standard deviation to obtain the response judgment value of the status parameter. This response judgment value remains unchanged during the same round of evaluation.

[0310] In addition, starting from the first valid sampling point after the control command is issued, the absolute value of the difference between the device status parameters and the status reference value before the action is executed is calculated sequentially.

[0311] When the absolute value is greater than or equal to the response judgment value, and the condition is met for three consecutive sampling points, the first sampling time among the three consecutive sampling points is determined as the device response start time. The response delay of the current action is obtained by subtracting the control command issuance time from the device response start time. The search range starts from the current control command issuance time and ends at the next control command issuance time. When the current action is the last action in the operation record, the end time of the operation record is taken as the end time. When no three consecutive sampling points that meet the response judgment condition appear within the search range, the time difference between the search end time and the control command issuance time is taken as the response delay of the action.

[0312] In addition, when more than two actions to be evaluated are matched at the same reference action position, the response delays of each action to be evaluated are added together and then divided by the number of corresponding actions. The resulting average response delay is taken as the response delay of that action position.

[0313] The response delays are arranged according to the action sequence of the baseline action sequence, so that each industrial operation record forms a row of response delay data, and each baseline action position forms a column. The entire industrial operation record constitutes a multidimensional feature matrix.

[0314] Before dimensionality reduction of the multidimensional feature matrix, the influence of response time scale at different action positions is eliminated column by column.

[0315] Specifically, for any response delay column, first calculate the arithmetic mean and standard deviation of all valid records in the column. Subtract the arithmetic mean of the column from the response delay of a single record, and then divide by the standard deviation to obtain the standardized response delay. When the standard deviation is equal to 0, the standardized result of the column is uniformly recorded as 0.

[0316] Principal component analysis was then used to reduce the dimensionality of the standardized multidimensional feature matrix.

[0317] Specifically, each principal component and its corresponding eigenvalue are calculated based on the covariance relationship between each response delay column. The eigenvalues ​​are arranged from largest to smallest, and the contribution rate of the corresponding principal component is obtained by dividing each eigenvalue by the sum of all eigenvalues. The principal component retention boundary is predetermined using historical evaluation data: principal components are retained according to cumulative contribution rates of 0.80, 0.85, 0.90, and 0.95, respectively. The historical response delay matrix is ​​then reconstructed after dimensionality reduction, and the mean absolute error between the reconstructed value and the original response delay is calculated. The condition that the mean absolute error does not exceed 5% of the average variation range of the original response delay is met. The smallest cumulative contribution rate among those meeting the condition is selected as the retention boundary for this round of dimensionality reduction. After this boundary is determined, the same value is used for all industrial operation records to be evaluated.

[0318] Furthermore, each industrial operation record is projected onto the retained principal component direction to obtain the corresponding dimensionality reduction feature value, and the dimensionality reduction feature value of the benchmark record of the same scene is used as the comparison benchmark.

[0319] For any industrial operation record, calculate the difference between its dimensionality-reduced eigenvalue and the corresponding dimensionality-reduced eigenvalue of the benchmark record. Square the difference and multiply it by the principal component contribution rate. Then, sum the calculation results corresponding to all retained principal components. Finally, take the square root of the sum. The resulting value is determined as the core deviation magnitude of the industrial operation record. When the core deviation magnitude is 0, it indicates that the response delay characteristics of the record in the retained principal component space are consistent with the benchmark record. When the core deviation magnitude increases, it indicates that the difference between its response timing and the benchmark record increases.

[0320] In this embodiment, the industrial operation records are sorted according to the core deviation magnitude, with records of smaller core deviation magnitude values ​​placed first and records of larger values ​​placed last.

[0321] When two industrial operation records have the same core deviation range, the record with the larger first or second true capability quantification value is ranked first; when both the core deviation range and the true capability quantification value are the same, they are arranged in ascending order according to the operation record number, thus obtaining a unique target sorting sequence.

[0322] Furthermore, after the target sorting sequence is formed, the actual capability quantification value corresponding to each industrial operation record is read sequentially.

[0323] Specifically, the records processed by S6 are given a performance score based on the first true capability quantification value, and the records processed by S7 are given a performance score based on the second true capability quantification value. A correspondence is established between the industrial operation strategy number, scenario category, operation record number, alignment feature sequence, response delay of each action position, core deviation amplitude, target sorting position and performance score, and written into the evaluation data table according to the target sorting sequence to form a target evaluation benchmark for industrial operation strategy performance evaluation.

[0324] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for constructing a multi-dimensional alignment evaluation benchmark for industrial operation strategies, characterized in that, include: S1. Obtain historical operating data and operating condition records generated during the execution of multiple industrial operation strategies to be evaluated. Use the K-means clustering algorithm to perform multi-dimensional feature extraction and classification processing on the historical operating data and operating condition records to obtain a scene distribution set. S2. For the scene distribution set, anomaly detection and boundary division are performed using the isolated forest algorithm to identify the scene distribution bias features contained within the scene distribution set and obtain the implicit bias identifier. S3. If the latent bias flag exceeds the pre-established bias judgment threshold, feature removal and resampling are performed on the scene distribution set to obtain an unbiased scene distribution set. S4. Based on the unbiased scene distribution set, the random forest algorithm is used to calculate the feature importance and cross-validate the security dimension feature, efficiency dimension feature and at least one other evaluation dimension feature to determine the multidimensional weight values. S5. By using multi-dimensional weight values, the operation records corresponding to different industrial operation strategies in the unbiased scene distribution set are weighted and normalized to obtain a multi-dimensional comprehensive score. S6. For the multidimensional comprehensive score, obtain the pre-established benchmark verification rules. If the multidimensional comprehensive score meets the benchmark verification rules, determine the first true capability quantification value corresponding to the operation record. S7. If the multidimensional comprehensive score does not meet the benchmark verification rules, a penalty calculation will be performed on the multidimensional comprehensive score to obtain the second true ability quantification value. S8. Based on the first true capability quantification value and the second true capability quantification value, sort and align the operation records corresponding to different industrial operation strategies to obtain the target evaluation benchmark for industrial operation strategy performance evaluation.

2. The method for constructing a multi-dimensional alignment evaluation benchmark for industrial operation strategies according to claim 1, characterized in that: S1 includes: A multi-dimensional feature matrix is ​​constructed by acquiring historical operating data and operating condition records generated during the execution of multiple industrial operation strategies to be evaluated. The multi-dimensional feature matrix is ​​then subjected to dimensionality reduction processing to obtain an enhanced feature set. Determine whether the fluctuation range of the enhanced feature set is greater than the preset fluctuation threshold. If it is greater than the preset fluctuation threshold, remove the abnormal samples to obtain the standard feature set. The K-means clustering algorithm is used to determine the cluster center vectors of the standard feature set. The steady-state operating intervals are divided according to the cluster center vectors, and the steady-state operating intervals are classified to obtain the scene distribution set.

3. The method for constructing a multi-dimensional alignment evaluation benchmark for industrial operation strategies according to claim 1, characterized in that: S2 includes: Obtain the distribution density and feature mapping of the scene distribution set to construct the distribution feature matrix; The path depth is calculated using the isolated forest algorithm on the distribution feature matrix to obtain the anomaly score sequence; Based on the abnormal score sequence, a segmentation hyperplane is defined to obtain a set of bias features; Determine whether the degree of bias in the bias feature set is greater than a preset bias threshold. If the degree of bias is greater than the preset bias threshold, perform label generation processing on the bias feature set to obtain a latent bias label.

4. The method for constructing a multi-dimensional alignment evaluation benchmark for industrial operation strategies according to claim 1, characterized in that: S3 includes: If the acquired implicit bias identifiers exceed the pre-established bias judgment threshold, then the bias feature weights of the scene distribution set are extracted based on the implicit bias identifiers. A subset of features to be removed is defined based on the bias feature weights. Feature removal is then performed on the scene distribution set using this subset of features to obtain a dimensionality-reduced scene distribution set. The synthetic sample set generated from the dimensionality-reduced scene distribution set is fused with the dimensionality-reduced scene distribution set to perform resampling processing, thereby obtaining an unbiased scene distribution set.

5. The method for constructing a multi-dimensional alignment evaluation benchmark for industrial operation strategies according to claim 1, characterized in that: S4 includes: Based on the unbiased scene distribution set, safety dimension features, efficiency dimension features, and comfort dimension features are extracted to construct a multi-dimensional feature matrix; For the multidimensional feature matrix, the random forest algorithm is used to obtain the feature splitting gain, and the reduction of Gini impurity is calculated based on the feature splitting gain to determine the initial feature importance; If the initial feature importance meets the preset significance criteria, then cross-validation is performed on the initial feature importance using out-of-bag data samples to obtain the mean prediction error. The initial feature importance is normalized and mapped based on the mean of the prediction error to determine the multidimensional weight values.

6. The method for constructing a multi-dimensional alignment evaluation benchmark for industrial operation strategies according to claim 1, characterized in that: S5 includes: Obtain the operation records corresponding to the industrial operation strategies within the unbiased scene distribution set, and parse the operation records to obtain the operation action sequence and equipment status response; The strategy execution deviation is calculated based on the sequence of operation actions and the equipment status response. Multidimensional weight values ​​are used to weight the strategy execution deviation to obtain a weighted aggregation matrix. Determine whether the weighted aggregation matrix satisfies the preset non-empty condition; If the weighted aggregation matrix satisfies the preset non-empty judgment condition, then extreme value normalization processing is performed on the weighted aggregation matrix to obtain a multi-dimensional comprehensive score. This completes the weighted calculation and normalization processing of the operation records corresponding to different industrial operation strategies within the unbiased scene distribution set, and obtains a multi-dimensional comprehensive score.

7. The method for constructing a multi-dimensional alignment evaluation benchmark for industrial operation strategies according to claim 1, characterized in that: S6 includes: The multidimensional comprehensive scores are compared according to the pre-established benchmark verification rules to obtain the score comparison results; If the scoring comparison results meet the benchmark verification rules, then extract the interaction frequency characteristics and historical credit rating of the operation records. The interaction frequency characteristics and historical credit ratings are classified and processed to construct a capability assessment matrix. Based on the capability assessment matrix, the first true capability quantification value corresponding to the operation record is determined.

8. The method for constructing a multi-dimensional alignment evaluation benchmark for industrial operation strategies according to claim 1, characterized in that: S7 includes: Obtain the multi-dimensional comprehensive score that does not conform to the benchmark verification rules, and extract the violation frequency and deviation range corresponding to the multi-dimensional comprehensive score; The frequency and extent of violations are categorized to determine risk levels. Based on the risk level, penalty weights and attenuation factors are matched, and the credit damage is calculated on the multi-dimensional comprehensive score using penalty weights and attenuation factors to obtain a corrected score. If the corrected score is lower than the preset quantification base, a penalty calculation is performed on the corrected score to obtain a second true ability quantification value.

9. The method for constructing a multi-dimensional alignment evaluation benchmark for industrial operation strategies according to claim 1, characterized in that: S8 includes: Analyze the industrial operation records corresponding to the first and second true capability quantification values ​​to obtain the initial action sequence; The initial action sequence is matched to obtain the alignment feature sequence.

10. The method for constructing a multi-dimensional alignment evaluation benchmark for industrial operation strategies according to claim 9, characterized in that: S8 further includes: The response delay of the aligned feature sequence is extracted to construct a multidimensional feature matrix. The multidimensional feature matrix is ​​then dimensionality-reduced to obtain the core deviation magnitude. The industrial operation records are arranged according to the core deviation magnitude to obtain the target sorting sequence; Based on the target sorting sequence, performance scores are extracted to construct a target evaluation benchmark for evaluating the performance of industrial operation strategies.

Images