Low-voltage operation capability evaluation and optimization method based on fusion operation behavior data

Abstract

The invention provides a low-voltage operation capability evaluation and optimization method based on fusion operation behavior data, and the method comprises the following steps: carrying out the multi-dimensional risk feature labeling of operation behavior data, and recognizing and classifying an electric shock contact behavior and a key operation response delay behavior as a high-risk abnormal behavior; an enhanced sample is generated by adopting an expansion sampling method, and the proportion of high-risk abnormal behaviors in a data set is increased; setting risk sensitivity according to the data source type; adjusting an abnormal identification threshold and a screening priority, and setting the high-risk abnormal behavior as the highest retention level; retaining data screened as abnormal data and marking the data as high-risk key samples; inputting a high-risk key sample and a normal sample into an evaluation model, and adjusting the influence on model training and reasoning; and comparing an original label according to a model output result to obtain a self-adaptive optimization result of the recognition precision and the evaluation effect. According to the invention, the recognition precision and the safety decision reliability of the operation capability evaluation system on high-risk behaviors can be improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent assessment technology for power operations, and in particular to a method for assessing and optimizing low-voltage operation capabilities based on integrated operation behavior data. Background Technology

[0002] In existing machine learning-based low-voltage worker competency assessment processes, outlier removal mechanisms (such as Z-Score or IQR) are typically used to preprocess raw operational behavior data to improve model training effectiveness and data quality. However, in practical applications, certain highly risky behavioral events, such as repeated contact with live equipment or significant delays in critical operation reaction times, occur during training. Although these events are extremely infrequent in the dataset, they possess high warning value and critical assessment significance. Because these high-risk behaviors are statistically far removed from the main data structure, they are easily misidentified as invalid outliers and automatically removed during preprocessing, preventing critical violations from entering the model training and assessment system. This problem causes significant deviations in the model's assessment of personnel risk awareness and operational compliance, leading the system to overestimate operational competency levels and misleading job allocation and retraining recommendations, posing potential safety hazards and management failure risks. Summary of the Invention

[0003] This invention provides a method for low-voltage operation capability assessment and optimization based on integrated operation behavior data. It can improve the accuracy of operation capability assessment system in identifying high-risk behaviors, its response capability, and the reliability of safety decisions, providing strong technical support for building a more scientific, dynamic, and intelligent operation capability management system for high-safety industries such as power.

[0004] The first aspect of this invention provides a method for assessing and optimizing low-pressure operation capabilities based on fused operational behavior data, comprising the following steps: Establish an abnormal behavior classification mechanism, label the operation behavior data with multi-dimensional risk features, and identify and classify electric shock contact behavior and critical operation reaction delay behavior as high-risk abnormal behavior; High-risk anomalous behaviors are augmented by using an augmented sampling method to generate enhanced samples, thereby increasing the proportion of high-risk anomalous behaviors in the dataset, and assigning explicit retention labels during the preprocessing stage. Establish an adaptive anomaly screening model and set risk sensitivity based on data source type; adjust anomaly identification threshold and screening priority, and set high-risk anomalies to the highest retention level; Acquire detection data, perform local feature comparison and trend analysis on the data that are filtered as abnormal, detect high-risk trends based on sliding window, retain the data that are filtered as abnormal and mark them as high-risk key samples; Establish a fusion decision-making mechanism, input high-risk key samples and normal samples into the evaluation model, set sample weights according to risk level, and adjust the impact on model training and inference. A retraining feedback loop is established, and the risk classification parameters and screening strategies are updated based on the model output results compared with the original annotations to obtain adaptive optimization results for recognition accuracy and evaluation effect.

[0005] Furthermore, the establishment of the abnormal behavior classification mechanism also includes: Acquire multidimensional features of work behavior data, including motion trajectory, operation sequence, response time, proximity behavior, voice response, and hand acceleration; Based on the pre-defined high-risk behavior model, feature parameters related to electric shock contact behavior and delayed operation response behavior are extracted; Match the risk behavior labeling system to identify high-risk abnormal events and generate risk indicator fields and behavior label fields; A high-risk behavior classification dataset was established, the sample quality and label consistency were verified, and a subset of high-risk behaviors was extracted for subsequent processing.

[0006] Furthermore, the sample augmentation for the marked high-risk anomalous behavior also includes: Establish a feature template library for high-risk behavior samples, extract key behavioral features, and form template vectors through clustering; Based on the template, augmented sampling is performed using temporal perturbation, spatial interpolation, and generative adversarial networks to generate enhanced samples; To enhance the samples, an enhancement source field and a retention priority field are set so that the samples are retained during the preprocessing stage; Some of the enhanced samples were included in the validation set, and the sample enhancement strategy and sampling frequency were adjusted based on the recognition performance.

[0007] Furthermore, the establishment of the adaptive anomaly screening model also includes: Label the source type of behavioral data, construct a data source trust model, and classify risk sensitivity levels; Set anomaly detection thresholds based on the sensitivity of the data source, and adjust the deviation judgment parameters for each source; Establish an abnormal data retention scoring model and set retention priorities based on data source level, risk label, and time location; The anomaly identification threshold and scoring parameters are dynamically updated based on the model output, which improves the accuracy of retaining high-risk behaviors.

[0008] Furthermore, the acquisition of detection data, and the performance of local feature comparison and trend analysis on the data filtered as abnormal, also includes: Establish a sliding window behavior dataset and extract continuous behavior data within the time range before and after abnormal data; The behavioral characteristics within the sliding window are compared with the standard template, and local deviation indicators such as trajectory offset rate and response delay volatility are calculated. Analyze the trends of characteristic time series to determine whether there is an evolutionary trend of high-risk behaviors; If the trend is significant, the sample is marked as a key sample and retained, and the trend score and retention level fields are recorded.

[0009] Furthermore, the establishment of the integrated decision-making mechanism also includes: Construct a unified data input fusion channel and organize a structured sample set including behavioral characteristics, label fields, data source fields, and risk level fields; Establish a mapping relationship between sample risk level and corresponding training weight, and introduce a risk impact suppression factor to control the sample weight distribution; The samples and weights are input into the evaluation model for joint training to adjust the degree of influence of the samples on the model training and inference. The risk level coding and training weights of samples are dynamically adjusted based on the model output results to optimize the fusion decision-making mechanism.

[0010] Furthermore, the establishment of the retraining feedback loop also includes: The model output is compared with the original labeled results to evaluate the accuracy of high-risk behavior identification. The risk classification parameters and behavioral characteristic templates are dynamically adjusted based on the differences observed. Optimize the sensitivity weights, threshold rules, and trend scoring factors of the anomaly screening mechanism; Based on the recognition error triggering model retraining, the updated sample set and parameters are input into the model to reconstruct and form a closed-loop optimization.

[0011] The second aspect of the present invention provides a low-voltage operation capability assessment and optimization system based on fused operation behavior data, including a first processing unit for establishing an abnormal behavior classification mechanism, labeling operation behavior data with multi-dimensional risk features, and identifying and classifying electric shock contact behavior and critical operation reaction delay behavior as high-risk abnormal behaviors. The second processing unit is used to augment the high-risk anomalous behaviors marked by the sample. It uses an augmented sampling method to generate augmented samples, increase the proportion of high-risk anomalous behaviors in the dataset, and assign explicit retention labels during the preprocessing stage. The third processing unit is used to establish an adaptive anomaly screening model, set risk sensitivity according to the data source type, adjust the anomaly identification threshold and screening priority, and set high-risk anomalies to the highest retention level. The fourth processing unit is used to acquire detection data, perform local feature comparison and trend analysis on the data that are filtered as abnormal, detect high-risk trends based on sliding window, retain the data that are filtered as abnormal and mark them as high-risk key samples; The fifth processing unit is used to establish a fusion decision-making mechanism, inputting high-risk key samples and normal samples into the evaluation model, setting sample weights according to risk level, and adjusting the impact on model training and inference. The sixth processing unit is used to establish a retraining feedback loop, compare the model output results with the original annotations, update the risk classification parameters and screening strategies, and obtain adaptive optimization results for recognition accuracy and evaluation effect.

[0012] A third aspect of the present invention provides a computer device, comprising: Memory, transceiver, processor, and bus system; The memory is used to store programs; The processor is used to execute the program in the memory, including executing the low-voltage operation capability assessment and optimization method based on fused operation behavior data as described above. The bus system is used to connect the memory and the processor to enable communication between the memory and the processor.

[0013] A fourth aspect of the present invention provides a readable storage medium storing computer-readable instructions, which, when executed by a processor, implement the steps of the low-pressure operation capability assessment and optimization method based on fused job behavior data described above.

[0014] As can be seen from the above technical solutions, the present invention has the following advantages: This invention achieves accurate identification of electric shock contact behaviors and critical operation delay behaviors by constructing a multi-dimensional risk feature annotation mechanism; it expands the training participation of high-risk behavior samples through sample augmentation technology and explicit retention strategies; it constructs an adaptive screening model based on data source sensitivity, dynamically adjusting identification rules to enhance the model's ability to perceive rare high-risk behaviors; it introduces a sliding window trend analysis mechanism to further verify and retain potentially high-risk samples, improving identification accuracy; it integrates a risk weight decision mechanism, making the model more focused on learning and representing safety-critical samples; and it constructs a complete closed loop of identification, correction, and optimization through a retraining feedback loop, driving the model's continuous adaptive evolution. Overall, this invention significantly improves the identification accuracy, response capability, and safety decision reliability of the operational capability assessment system for high-risk behaviors, providing strong technical support for building a more scientific, dynamic, and intelligent operational capability management system for high-safety industries such as power.

[0015] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from an examination of the following, or may be learned from the practice of the invention. Attached Figure Description

[0016] Figure 1 The method flowchart provided by the present invention. Detailed Implementation

[0017] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “corresponding to,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0018] Example 1 The implementation method in this embodiment can be implemented in a system, on a server, or on a terminal; no specific limitation is made. The method in this application will be described from the perspective of system implementation below. As shown in the figure, a method for low-pressure operation capability assessment and optimization based on fused operation behavior data includes the following steps: Establish an abnormal behavior classification mechanism, label the work behavior data with multi-dimensional risk characteristics, identify electric shock contact behavior and behavior with significant delay in reaction time of critical operation as high-risk abnormal events, and classify them as high-risk behavior categories. Establishing an abnormal behavior classification mechanism to address potential high-risk abnormal behaviors during operations is fundamental and crucial for training a highly reliable assessment model. This mechanism analyzes, labels, and classifies raw operational behavior data based on multi-dimensional risk characteristics, ensuring that potential hazardous events are identified, retained, and incorporated into the subsequent modeling process during data preprocessing. Its specific implementation is as follows: First, a multi-dimensional feature extraction module for operational behavior data is constructed. This module is designed for training and actual operation data sources throughout the entire low-voltage live-line work process, including but not limited to the operator's movement trajectory, operation sequence, response time of key links, approach behavior to dangerous areas, voice command response, and changes in hand acceleration. During the data acquisition phase, behavioral data is collected in real time synchronously through AR terminal devices, sensors on work tools, video tracking devices, and edge computing nodes. Subsequently, combined with a pre-defined high-risk behavior feature model, feature parameters highly correlated with "electric shock contact behavior" and "delayed response behavior in key operations" are extracted from each behavioral segment, such as: frequency of spatial trajectory points when contacting live areas, dwell time when entering high-risk areas, reaction signal delay duration, and intervals between consecutive erroneous operations.

[0019] Secondly, a risk behavior labeling system based on expert experience and knowledge graphs was designed. By investigating typical high-risk events during low-voltage live-line work, high-risk behaviors were classified according to risk level, and a standardized behavior labeling system was constructed. Based on this, an expert rule matching mechanism was established to match the aforementioned multi-dimensional features with the labeling rules to determine whether the current behavioral segment meets the identification criteria for a high-risk event. For example, when the work trajectory overlaps with the contact area of ​​live equipment and exceeds a set dwell threshold, and no safety confirmation action is taken, it is marked as "electric shock contact behavior"; when a critical control operation (such as disconnecting a switch) is delayed for more than 3 seconds from trigger prompt to actual action, and no other interfering events occur during this period, it is marked as "delayed operation response behavior." These behaviors are all classified as high-risk abnormal events and identified in the label library.

[0020] Third, a high-risk behavior labeling process is constructed, embedding identified high-risk events into the original data structure as explicit labels. The system adopts a structured data labeling method, adding a "risk indicator field" and a "behavior label field" to the original work data. The "risk indicator field" is used to mark whether an event is high-risk, and the "behavior label field" is used to specify the behavior type and its risk level (such as Level 1 electric shock risk, Level 2 slow reaction, etc.). This step is performed through a combination of automated scripts and manual review. To prevent labeling bias, the system supports a dual-channel review mechanism, where different analysis modules independently identify and compare labeling consistency, automatically triggering manual intervention when discrepancies exist.

[0021] Finally, a classification dataset for high-risk behaviors is established based on the annotation results. The system extracts the labeled high-risk abnormal events from the full dataset, establishes subsets according to behavior type, and stores them together with other normal behavior samples in a data warehouse, providing structured data support for subsequent sample augmentation, anomaly screening, and model training. At this stage, the system also verifies the sample quality, spatiotemporal distribution, and label consistency of various high-risk behaviors, removing unqualified data such as those with label conflicts and sampling interference. Through these steps, accurate identification and classification of key high-risk behaviors can be achieved, providing a high-quality key sample foundation for the entire low-voltage operation capability assessment model.

[0022] High-risk behavior samples are augmented by applying augmentation sampling methods to the labeled high-risk abnormal events to generate enhanced samples, thereby increasing the proportion of high-risk behaviors in the dataset and assigning them explicit retention labels during the data preprocessing stage. To effectively address the problem of scarce sample size and weak model recognition ability for high-risk behaviors in datasets, a high-risk behavior sample augmentation method is further proposed. This method expands the sample size and coverage of the previously labeled high-risk abnormal behaviors, enabling them to acquire sufficient expressive power and risk identification accuracy during subsequent model training. After abnormal behavior classification, this implementation method, based on the original behavior data and label information, constructs a high-risk behavior sample augmentation channel using augmentation sampling technology. Specifically, it includes the following four steps: First, a feature template library for high-risk behavior samples is constructed. For original data samples already labeled as "electric shock contact behavior" or "delayed response behavior in critical operations," core behavioral parameters are extracted, including key feature variables such as the behavior occurrence time window, changes in the operation trajectory, the degree of overlap of dangerous areas, the duration of delayed response, and the frequency of operations. These are then summarized into multiple sets of high-risk behavior template feature vectors using statistical clustering methods. This feature template library serves as a reference standard for generating synthetic samples, ensuring that the data generated during subsequent enhancement processes possesses structural consistency and risk representativeness of the original behaviors. To improve the diversity and representativeness of the templates, the system uses the DBSCAN density clustering method to subclassify the samples, avoiding overly simplistic templates due to feature averaging, thereby covering multiple types of high-risk behavior patterns.

[0023] Secondly, a multi-strategy sample augmentation mechanism is employed to expand the sampling of high-risk behaviors. Specifically, depending on the type of high-risk behavior, three methods—temporal window perturbation, spatial path interpolation, and generative adversarial network (GAN) sample synthesis—are implemented in parallel to augment the samples. Temporal window perturbation generates pseudo-samples by perturbing the time point and duration of the behavior at ±δ levels, thus expanding the temporal diversity of the behavior. Spatial path interpolation introduces small offsets into the original action trajectory, maintaining a consistent trajectory trend while increasing trajectory dispersion. The generative adversarial network method trains a generator model using the features of the original samples to synthesize new samples with similar high-risk characteristics but different expressions, effectively enhancing the behavior's variability. These three augmentation methods can be used independently or in combination. The final number of augmented samples is controlled based on the scarcity of the original samples and the need for class balance, ensuring that the augmentation results do not cause model overfitting or class imbalance.

[0024] Third, high-risk behavior samples after augmentation are given explicit retention labels. After data augmentation, the system embeds an augmentation source field and a retention priority field into each augmented sample. The augmentation source field records the sample generation method (such as "GAN synthesis", "trajectory perturbation", "window perturbation", etc.) so that it can be selectively loaded as needed during subsequent model training. The retention priority field sets different priority parameters based on the risk level. By introducing a "sample preservation strategy label" mechanism, the system incorporates augmented samples and original samples into the model input, and configures a sample weight adjustment strategy during training, so that the augmented samples can effectively drive the model gradient update, rather than just being used as noise data.

[0025] Finally, the effectiveness of the augmented samples is verified and dynamically updated. During the model training phase, a cross-validation strategy is introduced, incorporating some augmented samples into the validation set to evaluate their impact on model convergence speed, risk event identification accuracy, and false positive rate. Loss function fluctuation analysis provides feedback on sample quality. Augmented samples with low identification contribution are eliminated or replaced. Simultaneously, the sample augmentation strategy and sampling density are dynamically adjusted based on changes in the model's identification weights for different types of high-risk behaviors. For example, if the model's ability to identify "operational delay behavior" remains weak, the system will increase the augmentation frequency and retention strength of such samples; conversely, it will reduce their sampling priority and optimize the sample distribution structure. Through this feedback mechanism, the sample augmentation process evolves from static generation to dynamic optimization, ensuring a positive synergy between sample augmentation behavior and model training requirements.

[0026] In summary, this implementation method comprehensively improves the proportion and effectiveness of high-risk behavior samples in the dataset by constructing a high-risk behavior template library, introducing a multi-strategy sample enhancement mechanism, setting explicit retention markers, and executing a dynamic update verification process. This provides a solid data foundation for the construction of the capability assessment model and effectively solves the problem of missing high-risk event samples caused by traditional outlier removal methods. It also has good engineering adaptability and innovation.

[0027] Construct an adaptive anomaly screening model, set risk sensitivity parameters according to the source type of operational behavior data, dynamically set anomaly identification thresholds and screening priorities for different data sources, and ensure that high-risk anomalies have the highest retention level. To address the "static rejection with a uniform threshold" problem inherent in conventional outlier screening mechanisms when processing multi-source operational behavior data, an adaptive outlier screening model is proposed. This model constructs a risk sensitivity parameter system based on data source type and dynamically sets identification thresholds and screening priorities by incorporating behavioral context information. This ensures that high-risk outliers are not mistakenly rejected during data preprocessing, thereby enhancing the system's ability to retain hazardous behaviors and improving the risk identification performance of the assessment model. The construction and implementation of this model includes the following four steps: First, a module for identifying the source type of operational behavior data is established. Since low-voltage uninterrupted power supply (UPS) data typically originates from various sensing devices and recording systems, such as AR glasses, motion capture equipment, environmental video streams, handheld tool sensors, and voice input interfaces, it is necessary to label and differentiate these data sources. During the data acquisition phase, the system tags each piece of behavioral data with its source, recording the acquisition hardware, acquisition method, spatiotemporal resolution, and synchronization accuracy. By conducting a multi-dimensional evaluation of data source quality, a "data source trust model" is established, providing a basis for source differentiation in subsequent anomaly identification. This trust model, based on technical indicators such as acquisition latency, error range, and synchronization accuracy, categorizes data sources into high, medium, and low risk sensitivity levels.

[0028] Secondly, a risk sensitivity-based anomaly detection threshold adjustment mechanism is constructed. In the data preprocessing module, traditional static threshold judgment mechanisms often ignore the differences in data source quality, easily classifying the same behavioral event as anomaly in high-confidence data sources while misjudging it as invalid in low-confidence sources. To avoid this problem, this invention introduces a "sensitivity weighting factor" to dynamically adjust the deviation threshold used for anomaly detection (such as Z-Score deviation coefficient, IQR box boundary, etc.) according to the data source type. For example, high-precision time-series trajectory data from AR glasses can use a stricter discrimination threshold (such as Z>1.5) for anomaly detection; while image inference data from external environment cameras, due to their inherent latency and deviation tolerance, have a more relaxed threshold setting (such as Z>2.5). Through this mechanism, different types of data are given differentiated tolerance strategies in anomaly detection, improving the accuracy of screening results and the likelihood of retaining high-risk behaviors.

[0029] Third, a dynamic screening priority control strategy is designed. After the system performs preliminary anomaly identification on multi-source behavioral data, it is necessary to further set retention priorities for data entries judged as "abnormal" to prevent important abnormal information from being indiscriminately removed during subsequent data cleaning or denoising processes. This invention constructs a retention priority scoring model based on the aforementioned three dimensions: data source trust level, behavioral label risk level, and data time location, generating a retention score for each abnormal sample. For example, if an abnormal data point originates from a high-trust data source and its behavioral label is "approaching a charged zone trajectory mutation," then the data retention score is set to the highest level, and it is forcibly retained; conversely, isolated anomalies from low-trust data sources, without behavioral labels, and unrelated to the main trajectory trend can be set to low priority or candidate deletion status.

[0030] Finally, a self-learning and dynamic adjustment mechanism for the screening model is established. To adapt to different operational scenarios and data distribution changes, this implementation introduces an adaptive learning module based on feedback from the evaluation model. During model training, the system continuously monitors training errors and identification result deviations. If it detects a decrease in the recognition rate of a certain type of high-risk behavior or an increase in misidentification, it sends a parameter tuning request to the anomaly screening module, automatically updating the anomaly identification threshold, retention scoring parameters, or data source weight settings. For example, when the model's accuracy in recognizing "delayed operational response" samples decreases, the system lowers the removal threshold for this type of behavior in the corresponding data source, increasing the retention probability, and adjusts the scoring weight to ensure its participation in model retraining. This mechanism realizes a shift from "passive screening" to "feedback-driven screening," enabling the entire anomaly identification process to have dynamic adaptability, significantly improving the system's overall response capability to high-risk abnormal behaviors and the stability of the evaluation system.

[0031] In summary, this invention systematically solves the shortcomings of traditional anomaly identification methods in erroneously rejecting high-risk behaviors by constructing an adaptive anomaly screening model and combining multi-source data identification, sensitivity grading threshold control, retention priority allocation, and feedback self-learning mechanism. It realizes the strategy of "maximizing the retention of risk data" in the behavior assessment system, provides key technical support for the assessment of high reliability operation capabilities, and has strong engineering practicality and innovation.

[0032] Local feature comparison and trend analysis are performed. For data identified as abnormal in the adaptive screening, local feature difference analysis is carried out within a sliding window. If a high-risk behavioral trend is detected, the data is retained and marked as a key sample. To further improve the accuracy of identifying high-risk abnormal behaviors and avoid misjudging data as invalid or abnormal due to local fluctuations or short-term behavioral shifts, an abnormal behavior verification mechanism based on local feature comparison and trend analysis is designed. This mechanism uses a sliding window approach to perform temporal context analysis on data identified as abnormal in the aforementioned adaptive filtering module, leveraging the evolutionary trend of local behavioral patterns to assist in determining whether it possesses potentially high-risk characteristics. Its implementation includes the following four steps: First, a local sliding window behavior dataset is constructed. After initial anomaly screening, the system extracts continuous operation behavior data within a certain time range before and after each piece of data identified as an anomaly as context information, forming a sliding window sequence containing the anomaly point. The window size is dynamically set based on the granularity of the behavior features and the duration of the operation, typically set to a time range of 2-5 seconds before and after, ensuring that the context range effectively covers the complete operation segment. For trajectory-type data, its two-dimensional or three-dimensional spatial path node sequence is extracted; for action response-type data, indicators such as behavior trigger time, action execution time, and execution delay are extracted; for multimodal data, sliding windows are established separately in different modalities. This step provides the data foundation for subsequent local feature analysis.

[0033] Secondly, local feature comparison calculations are performed. Based on continuous data within a sliding window, the system constructs a sequence of local feature vectors and compares them with a standard behavior template or normal behavior curve to calculate a local deviation index. This deviation may include, but is not limited to, trajectory offset rate, response delay fluctuation rate, and the rate of change in the frequency of abnormal actions. For example, if a trajectory point suddenly deviates significantly from the normal work path during operation, the system obtains a spatial deviation value by comparing it with a standard trajectory; if the delay time of a certain operation step significantly exceeds the average value in that segment of the operation, it is marked as a temporal feature anomaly. The system uses a weighted combination of indicators to perform multi-dimensional feature fusion, generating a comprehensive deviation score that reflects the overall degree of deviation between the current local behavior and the standard behavior.

[0034] Third, the system analyzes the evolution trend of local features. It performs time series fitting and trend detection on each feature indicator within the sliding window to determine whether changes in this local feature constitute a continuous precursor to high-risk behavior. By constructing local time series regression models, moving average models, or short-term rate of variation analysis, the system assesses whether the anomaly is in a "rising risk trend phase." For example, in a certain type of operation, if the trajectory deviates continuously, the frequency of erroneous operations increases, and the reaction time continues to lengthen, it can be judged as a "risk behavior evolution trend," meaning that the behavior is not occasional noise but a sign that a more serious violation is about to occur. To enhance the robustness of the model, the system also introduces a multi-feature joint trend judgment mechanism. When at least two or more feature dimensions show drastic changes in the same direction, the behavior is determined to have a significant risk trend.

[0035] Finally, the system performs key sample labeling and retention. If the local trend analysis determines that the anomalous data exhibits a high-risk behavioral evolution trend, the system labels it as a "key sample" and forcibly retains it in the training sample set. Key samples are stored with embedded "trend identifier fields" and "retention level fields," recording their local trend score and retention reasons, serving as crucial input for subsequent model training to identify high-risk evolving behaviors. Conversely, if the anomalous data shows no obvious trend or contextual support within the sliding window, the system downgrades it to a "non-key anomaly," leaving its inclusion in the model input to the next stage's risk fusion mechanism. To improve data utilization efficiency, the system also incorporates an automatic learning mechanism that periodically evaluates the recognition contribution of key samples in model training and updates the sliding window size, feature weights, and trend threshold settings, enabling adaptive evolution of the local feature analysis module.

[0036] In summary, the local feature comparison and trend analysis mechanism proposed in this invention can effectively overcome the static limitations of traditional anomaly identification methods in the time dimension. It re-identifies high-risk behaviors that are easily misidentified but possess significant early warning value from a behavioral evolution perspective. By constructing a sliding window context, calculating local deviation indicators, extracting trend change signals, and implementing dynamic sample retention, this mechanism achieves a shift from "point judgment" to "segment reasoning," providing crucial technical support for the deep identification and modeling of high-risk behavior data. It demonstrates significant practical application and technological innovation value.

[0037] A risk perception and fusion decision-making mechanism is constructed, which inputs both the retained high-risk key samples and normal behavior samples into the evaluation model, and sets risk weights based on the sample risk level to adjust the influence ratio of each sample in the training and inference of the evaluation model. To address the issue of high-risk critical behavior samples being diluted or insufficiently contributing to model training in low-voltage live-line work capability assessment systems, a risk perception fusion decision-making mechanism is proposed. This mechanism integrates the previously retained high-risk critical samples with normal behavior samples into the assessment model and assigns risk weights based on the risk level of the samples. This dynamically adjusts the influence of samples on the learning results during model training and inference, achieving stronger perception and assessment sensitivity for high-risk behaviors. The mechanism specifically includes the following four steps: First, a unified data input fusion channel is constructed. After classifying, enhancing, filtering, and trend-labeling high-risk samples of operational behavior data in the early stages, the system organizes all retained data (including normal and critical samples) into a structured sample set. In the fusion channel, each sample contains a basic behavioral feature vector, a behavioral label field, a data source field, and a risk level field. The risk level field, derived from the aforementioned multi-dimensional assessment (such as expert rule annotation and trend analysis judgment), is used to represent the risk importance of the sample in the assessment system. To support the access of different model structures, this fusion channel adopts a modular feature representation method, including a numerical feature vector matrix, label embedding vectors, and a risk level encoding field, forming a unified data format that can be input into the main model.

[0038] Secondly, the system designs a mapping relationship between sample risk levels and weights. Based on the risk level of each sample, the system assigns training weights to each sample. Risk levels are divided into multiple tiers (e.g., 0 represents normal behavior, 1 is low risk, 2 is medium risk, and 3 is high risk), each corresponding to a different sample weight value (e.g., w0=1.0, w1=1.2, w2=1.5, w3=2.0). The higher the risk level, the greater the sample's contribution to the loss function during training, thus guiding the model to focus more on learning and representing risky behaviors. This weight mapping relationship can be set by experts or dynamically adjusted based on model feedback. To avoid overfitting to high-risk behaviors, the system also introduces a "risk influence suppression factor," which automatically normalizes the total sample influence based on the weight distribution in each training round, ensuring stable convergence during training.

[0039] Third, samples and weights are jointly trained in the capability assessment model. During model training, the system inputs the behavioral features and weight values ​​of samples from the fusion channel into the model structure. Taking a typical deep neural network as an example, the loss function (such as cross-entropy loss or mean squared error loss) during training is weighted and corrected, specifically by multiplying the error of each sample by its corresponding risk weight and accumulating the results. This makes the model parameter updates more focused on the ability to identify high-risk samples. In tree models (such as XGBoost and LightGBM) or graph neural network structures, risk weights are also used to weight the sample splitting gain or node contribution. This process ensures that high-risk samples can play a significant role in training guidance regardless of the model structure used. To improve the model's ability to identify high-risk behaviors during the inference stage, the system also introduces a risk weight factor during the model evaluation stage to post-process the output confidence of the inference results, enhancing the system's sensitivity to anomalies.

[0040] Finally, a risk feedback-based fusion decision optimization mechanism is established. During the model deployment phase, the system continuously collects data on the deviation between the model output and actual behavioral data, and analyzes the accuracy and false positive rate of risk sample identification. If the identification rate of a certain type of high-risk behavior sample is found to be insufficient, the system will increase the risk level encoding and training weight of that type of sample in the training set to strengthen the focus on that behavior in the next round of training. Simultaneously, the system also supports feedback adjustments to sample risk levels; that is, if some samples originally set as high-risk are frequently misclassified, the system will appropriately lower their weight values ​​based on the validation set performance.

[0041] In summary, this invention constructs a risk perception fusion decision-making mechanism, introduces a risk level and weight system into the evaluation model, achieves unified modeling of critical and normal samples, and enhances the model's sensitivity and reliability in identifying high-risk behaviors through sample weight adjustment and feedback optimization mechanisms. This mechanism effectively avoids the problem of data dilution of critical samples, improves the model's ability to perceive and respond to operational risks in real-world working environments, and possesses high engineering practical value and technological innovation.

[0042] Establish an abnormal behavior retraining feedback loop based on model output results, compare the high-risk behavior identification results of the evaluation model output with the original annotation results, and dynamically update the risk classification parameters and screening mechanism configuration to achieve adaptive optimization of high-risk behavior identification accuracy and evaluation model performance. To further improve the accuracy of high-risk behavior identification and overall model performance in the low-voltage uninterrupted power supply (UPS) capability assessment system, an abnormal behavior retraining feedback loop based on model output results is constructed. This aims to achieve a closed-loop information system and dynamic collaborative update mechanism between model evaluation results and the anomaly identification mechanism. This feedback loop not only enables rapid response and correction of model identification errors but also dynamically optimizes risk classification parameters and anomaly screening mechanism configurations, thereby ensuring continuous evolution and high reliability of the system during long-term operation. This implementation method specifically includes the following four steps: First, a module is constructed to compare the high-risk identification results with the original labeled results. After the model completes one round of training and outputs the evaluation results, the system automatically extracts the model's identification output of high-risk behaviors in the operational behavior data and compares it one by one with the original manual or rule-based labeled results previously generated by the abnormal behavior classification module. This comparison includes the calculation of indicators such as recognition accuracy (True Positive), false positive, and false negative, while recording the specific predicted probability value of each key sample and its difference from its corresponding risk level. Through this module, the system can systematically evaluate the current model's ability to identify various high-risk behaviors and identify which categories or patterns of behavior have biased, missed, or misjudged issues in the model.

[0043] Secondly, the risk classification parameters are dynamically adjusted based on the difference analysis results. After acquiring the identification deviation data, the system inputs it into the risk classification parameter tuning module to recalculate the risk level weight distribution and feature importance ranking of abnormal behavior types. For example, if the system finds that behaviors such as "repeatedly touching live equipment" frequently generate low-confidence predictions in the model identification, it will increase the risk level parameter of this type of behavior in the sample labeling, thereby enhancing its corresponding sample weight in the next round of training. At the same time, the system adaptively adjusts the feature templates of high-risk behaviors, that is, it identifies the behavioral feature substructures that the current model has not effectively learned through sample clustering analysis, thereby expanding or refining the original risk behavior template library to ensure that subsequent enhanced samples better reflect the evolution trend of real high-risk behaviors.

[0044] Third, the anomaly screening mechanism configuration is optimized based on the identification feedback information. The system uses the identification deviation results and model confidence feedback to drive the parameter adjustment of the adaptive anomaly screening model. Specifically, this includes updating the sensitivity weights of various data sources, anomaly threshold adjustment rules, and local trend scoring factors. For example, when the system identifies that a certain type of abnormal data source is frequently misjudged as non-risk behavior, the anomaly screening threshold corresponding to that data source is appropriately lowered to increase the retention probability of that type of sample; if a certain risk category cannot be identified by the screening mechanism due to fluctuations in sample feature distribution, a backup feature channel is automatically called for double screening. In addition, to avoid over-adjustment leading to anomaly generalization, the system introduces a threshold adjustment smoothing factor, dynamically controlling the update magnitude according to the change in the proportion of misjudged samples, thus achieving robustness and convergence of parameter updates.

[0045] Finally, a retraining trigger mechanism is constructed and model reconstruction is completed. Based on the aforementioned parameter optimization results, the system determines whether the current model needs to trigger a retraining operation. If the recognition error exceeds a set threshold, or the recognition accuracy of key samples continuously declines, the system automatically initiates the retraining process, using the updated sample set, weight configuration, and selection parameters as input to reconstruct the capability assessment model. This retraining process can be performed using incremental learning or full retraining, provided that data consistency and parameter synchronization are ensured. After training is completed, the system will re-evaluate the performance indicators of the new model and reintegrate the new round of recognition results into the feedback loop, forming a closed-loop optimization mechanism of recognition, comparison, adjustment, and retraining.

[0046] Example 2 A low-voltage operation capability assessment and optimization system based on fused operation behavior data includes a first processing unit, which is used to establish an abnormal behavior classification mechanism, perform multi-dimensional risk feature labeling on operation behavior data, and identify and classify electric shock contact behavior and critical operation reaction delay behavior as high-risk abnormal behavior. The second processing unit is used to augment the high-risk anomalous behaviors marked by the sample. It uses an augmented sampling method to generate augmented samples, increase the proportion of high-risk anomalous behaviors in the dataset, and assign explicit retention labels during the preprocessing stage. The third processing unit is used to establish an adaptive anomaly screening model, set risk sensitivity according to the data source type, adjust the anomaly identification threshold and screening priority, and set high-risk anomalies to the highest retention level. The fourth processing unit is used to acquire detection data, perform local feature comparison and trend analysis on the data that are filtered as abnormal, detect high-risk trends based on sliding window, retain the data that are filtered as abnormal and mark them as high-risk key samples; The fifth processing unit is used to establish a fusion decision-making mechanism, inputting high-risk key samples and normal samples into the evaluation model, setting sample weights according to risk level, and adjusting the impact on model training and inference. The sixth processing unit is used to establish a retraining feedback loop, compare the model output results with the original annotations, update the risk classification parameters and screening strategies, and obtain adaptive optimization results for recognition accuracy and evaluation effect.

[0047] Example 3 A computer device, comprising: Memory, transceiver, processor, and bus system; The memory is used to store programs; The processor is used to execute the program in the memory, including executing the low-voltage operation capability assessment and optimization method based on fused operation behavior data as described above. The bus system is used to connect the memory and the processor to enable communication between the memory and the processor.

[0048] Example 4 A readable storage medium storing computer-readable instructions, which, when executed by a processor, implement the steps of the low-pressure operation capability assessment and optimization method based on fused job behavior data described above.

[0049] In summary, this invention, by constructing an abnormal behavior retraining feedback loop driven by model recognition output, comprehensively realizes dynamic closed-loop control of high-risk behaviors, from identification deviation detection, risk parameter optimization, anomaly screening and adjustment to model retraining. This mechanism possesses continuous learning, automatic adjustment, and model evolution capabilities, significantly improving the adaptability and risk identification accuracy of the low-voltage uninterrupted power supply capability assessment system, and has high engineering value and significant technological innovation significance.

[0050] The low-voltage operation capability assessment and optimization method proposed in this invention, which integrates operational behavior data, effectively solves problems in existing technologies such as the erroneous rejection of high-risk behavior samples during preprocessing, insufficient model recognition of key abnormal behaviors, and the inability to self-correct assessment result biases. This method achieves accurate identification of electric shock contact behaviors and key operational delay behaviors by constructing a multi-dimensional risk feature annotation mechanism; expands the training participation of high-risk behavior samples through sample augmentation techniques and explicit retention strategies; constructs an adaptive screening model based on data source sensitivity, dynamically adjusting recognition rules to enhance the model's perception of rare high-risk behaviors; introduces a sliding window trend analysis mechanism to further verify and retain potential high-risk samples, improving recognition accuracy; integrates a risk weight decision mechanism, making the model more focused on learning and representing safety-critical samples; and constructs a complete closed loop of recognition, correction, and optimization through a retraining feedback loop, driving continuous adaptive evolution of the model. Overall, this invention significantly improves the accuracy, responsiveness, and reliability of safety decisions in operation capability assessment systems for high-risk behaviors, providing strong technical support for building a more scientific, dynamic, and intelligent operation capability management system for high-safety industries such as power.

[0051] It is understood that those skilled in the art can combine various implementation methods in the above embodiments under the guidance of the above examples to obtain technical solutions with multiple implementation methods.

[0052] Those skilled in the art will recognize that the units of the various examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of the invention.

[0053] In the embodiments provided by the present invention, it should be understood that the division of units is only a logical functional division. In actual implementation, there may be other division methods, such as multiple units can be combined into one unit, one unit can be split into multiple units, or some features can be ignored.

[0054] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0055] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0056] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for assessing and optimizing low-voltage operation capabilities based on fused operational behavior data, characterized in that, Includes the following steps: Establish an abnormal behavior classification mechanism, label the operation behavior data with multi-dimensional risk features, and identify and classify electric shock contact behavior and critical operation reaction delay behavior as high-risk abnormal behavior; High-risk anomalous behaviors are augmented by using an augmented sampling method to generate enhanced samples, thereby increasing the proportion of high-risk anomalous behaviors in the dataset, and assigning explicit retention labels during the preprocessing stage. Establish an adaptive anomaly screening model and set risk sensitivity based on the data source type; Adjust the anomaly identification threshold and screening priority, and set high-risk abnormal behaviors to the highest retention level; Acquire detection data, perform local feature comparison and trend analysis on the data that is filtered as abnormal, detect high-risk trends based on sliding window, retain the data that is filtered as abnormal and mark it as high-risk key samples; Establish a fusion decision-making mechanism, input high-risk key samples and normal samples into the evaluation model, set sample weights according to risk level, and adjust the impact on model training and inference. A retraining feedback loop is established, and the risk classification parameters and screening strategies are updated based on the model output results compared with the original annotations to obtain adaptive optimization results for recognition accuracy and evaluation effect.

2. The method for evaluating and optimizing low-voltage operation capabilities based on fused operational behavior data according to claim 1, characterized in that, The establishment of the abnormal behavior classification mechanism also includes: Acquire multidimensional features of work behavior data, including motion trajectory, operation sequence, response time, proximity behavior, voice response, and hand acceleration; Based on the pre-defined high-risk behavior model, feature parameters related to electric shock contact behavior and delayed operation response behavior are extracted; Match the risk behavior labeling system to identify high-risk abnormal events and generate risk indicator fields and behavior label fields; A high-risk behavior classification dataset was established, the sample quality and label consistency were verified, and a subset of high-risk behaviors was extracted for subsequent processing.

3. The method for evaluating and optimizing low-voltage operation capabilities based on fused operational behavior data according to claim 1, characterized in that, The sample augmentation for marked high-risk anomalous behaviors also includes: Establish a feature template library for high-risk behavior samples, extract key behavioral features, and form template vectors through clustering; Based on the template, augmented sampling is performed using temporal perturbation, spatial interpolation, and generative adversarial networks to generate enhanced samples; To enhance the samples, an enhancement source field and a retention priority field are set so that the samples are retained during the preprocessing stage; Some of the enhanced samples were included in the validation set, and the sample enhancement strategy and sampling frequency were adjusted based on the recognition performance.

4. The method for evaluating and optimizing low-voltage operation capabilities based on fused operation behavior data according to claim 1, characterized in that, The establishment of the adaptive anomaly screening model also includes: Label the source type of behavioral data, construct a data source trust model, and classify risk sensitivity levels; Set anomaly detection thresholds based on the sensitivity of the data source, and adjust the deviation judgment parameters for each source; Establish an abnormal data retention scoring model and set retention priorities based on data source level, risk label, and time location; The anomaly identification threshold and scoring parameters are dynamically updated based on the model output, which improves the accuracy of retaining high-risk behaviors.

5. The method for evaluating and optimizing low-voltage operation capabilities based on fused operational behavior data according to claim 1, characterized in that, The process of acquiring detection data and performing local feature comparison and trend analysis on the data filtered as abnormal also includes: Establish a sliding window behavior dataset and extract continuous behavior data within the time range before and after abnormal data; The behavioral characteristics within the sliding window are compared with the standard template, and local deviation indicators such as trajectory offset rate and response delay volatility are calculated. Analyze the trends of characteristic time series to determine whether there is an evolutionary trend of high-risk behaviors; If the trend is significant, the sample is marked as a key sample and retained, and the trend score and retention level fields are recorded.

6. The method for evaluating and optimizing low-voltage operation capabilities based on fused operation behavior data according to claim 1, characterized in that, The establishment of the integrated decision-making mechanism also includes: Construct a unified data input fusion channel and organize a structured sample set including behavioral characteristics, label fields, data source fields, and risk level fields; Establish a mapping relationship between sample risk level and corresponding training weight, and introduce a risk impact suppression factor to control the sample weight distribution; The samples and weights are input into the evaluation model for joint training to adjust the degree of influence of the samples on the model training and inference. The risk level coding and training weights of samples are dynamically adjusted based on the model output results to optimize the fusion decision-making mechanism.

7. The method for evaluating and optimizing low-voltage operation capabilities based on fused operational behavior data according to claim 1, characterized in that, The establishment of the retraining feedback loop also includes: The model output is compared with the original labeled results to evaluate the accuracy of high-risk behavior identification. The risk classification parameters and behavioral characteristic templates are dynamically adjusted based on the differences observed. Optimize the sensitivity weights, threshold rules, and trend scoring factors of the anomaly screening mechanism; Based on the recognition error triggering model retraining, the updated sample set and parameters are input into the model to reconstruct and form a closed-loop optimization.

8. A low-voltage operation capability assessment and optimization system based on fused operational behavior data, characterized in that, It includes a first processing unit, which is used to establish an abnormal behavior classification mechanism, perform multi-dimensional risk feature labeling on work behavior data, and identify and classify electric shock contact behavior and critical operation reaction delay behavior as high-risk abnormal behavior. The second processing unit is used to augment the high-risk anomalous behaviors marked by the sample. It uses an augmented sampling method to generate augmented samples, increase the proportion of high-risk anomalous behaviors in the dataset, and assign explicit retention labels during the preprocessing stage. The third processing unit is used to establish an adaptive anomaly screening model and set risk sensitivity according to the data source type. Adjust the anomaly identification threshold and screening priority, and set high-risk abnormal behaviors to the highest retention level; The fourth processing unit is used to acquire detection data, perform local feature comparison and trend analysis on the data that are filtered as abnormal, detect high-risk trends based on sliding window, retain the data that are filtered as abnormal and mark them as high-risk key samples; The fifth processing unit is used to establish a fusion decision-making mechanism, inputting high-risk key samples and normal samples into the evaluation model, setting sample weights according to risk level, and adjusting the impact on model training and inference. The sixth processing unit is used to establish a retraining feedback loop, compare the model output results with the original annotations, update the risk classification parameters and screening strategies, and obtain adaptive optimization results for recognition accuracy and evaluation effect.

9. A computer device, characterized in that, include: Memory, transceiver, processor, and bus system; The memory is used to store programs; The processor is used to execute the program in the memory, including executing the low-voltage operation capability assessment and optimization method based on fused operation behavior data as described in any one of claims 1 to 7; The bus system is used to connect the memory and the processor to enable communication between the memory and the processor.

10. A readable storage medium storing computer-readable instructions, characterized in that, When the computer-readable instructions are executed by the processor, they implement the steps of the low-voltage operation capability assessment and optimization method based on fused operation behavior data as described in any one of claims 1 to 7.

Images