Intelligent integrated panoramic monitoring terminal and data processing method for preventing traffic violations

Abstract

This disclosure provides an intelligent integrated panoramic monitoring terminal and data processing method for combating violations, belonging to the field of power system technology. The method includes: acquiring multi-source heterogeneous violation alarm data and preprocessing it; evaluating the credibility of the preprocessed alarm data based on multi-dimensional feature extraction results; dynamically allocating credibility weights for each violation event based on the credibility evaluation results; and prioritizing and filtering the violation events according to the credibility weights. The solution of this disclosure can dynamically allocate credibility weights for violation events based on multi-source heterogeneous violation alarm data in the power industry to solve the problem of alarm overload caused by high low-confidence false alarm rates.

Description

Technical Field

[0001] This application relates to the field of power system technology, specifically to an intelligent integrated panoramic monitoring terminal for combating violations and a data processing method. Background Technology

[0002] In actual operation, power industry violation monitoring systems typically rely on various sensors, video analytics devices, and terminal acquisition devices to obtain violation alarm information. However, due to limitations in equipment accuracy, environmental interference, and recognition algorithms, alarm data from a single source often has a high false alarm rate. Furthermore, the lack of effective fusion and dynamic reliability assessment mechanisms among multi-source heterogeneous data makes it difficult for the system to accurately identify and prioritize violations when faced with a large amount of repetitive, redundant, or even conflicting alarm information. This leads to problems such as a proliferation of low-confidence alarms and delayed responses to critical events, seriously affecting the efficiency and reliability of on-site safety management. Therefore, there is an urgent need for a method that can combine multi-dimensional features to assess the reliability of multi-source heterogeneous violation alarm data and dynamically allocate weights to improve the accuracy of violation event identification and response capabilities. Summary of the Invention

[0003] In view of this, the present disclosure provides a data processing method for an intelligent integrated panoramic monitoring terminal for combating traffic violations, which at least partially solves the problems existing in the prior art.

[0004] Data processing methods for intelligent integrated panoramic monitoring terminals for combating violations in the power industry include:

[0005] Acquire multi-source heterogeneous traffic violation alarm data and perform preprocessing;

[0006] The credibility of the preprocessed alarm data is evaluated based on the multidimensional feature extraction results.

[0007] The credibility weight of each violation event is dynamically assigned based on the credibility assessment results.

[0008] The violation events are prioritized and false alarms are filtered based on the credibility weight.

[0009] In one specific implementation, the credibility assessment of the preprocessed alarm data based on the multidimensional feature extraction results further includes:

[0010] Obtain the anomaly detection scores S1, S2, ..., Sn for each source data.

[0011] Calculate the average outlier confidence value M_avg = (S1 + S2 + ... + Sn) / n for each source data;

[0012] The following formula is used to determine whether credibility needs to be enhanced: C = α × M_avg + β × max(Si), where α and β are weighting coefficients, and Si is the anomaly score for each source;

[0013] If C > T, the credibility enhancement mechanism is triggered; otherwise, the original value remains unchanged.

[0014] In one specific implementation, the dynamic allocation of credibility weights for each violation event based on the credibility assessment results further includes:

[0015] Obtain the historical false alarm rates R1, R2, ..., Rm for each type of traffic violation within the current time period;

[0016] The dynamic weighting coefficient Wi = e^(τ × t) is calculated based on the time decay factor τ, where t is the time interval between the current time and the last occurrence of the violation.

[0017] The confidence score C and the false alarm rate Rj are combined to generate a comprehensive evaluation value Wij = (C / (Rj + ε)) ×Wi, where ε is a small constant to avoid division by zero;

[0018] Normalize Wij to determine the final weight value.

[0019] In one specific implementation, the step of generating a comprehensive evaluation value Wij = (C / (Rj + ε)) × Wi by combining the confidence score C and the false alarm rate Rj further includes:

[0020] The contribution of using a nonlinear function to improve the false alarm rate is: Wij = (C × log(Rj + 1)) × Wi;

[0021] Set a threshold λ to filter high-risk events. If Wij > λ, it is considered a high priority.

[0022] By introducing spatial dimension information (Sgeo), clustering and scoring similar violations within a region, a spatial correction factor δ is formed.

[0023] The final weight value is calculated using the formula: Wij_new = Wij × (1 + δ).

[0024] In one specific implementation, the prioritization and false alarm filtering of violation events based on the credibility weight further includes:

[0025] Construct a dynamic priority matrix P = {Pij | i=1...n, j=.m}, where Pij represents the ranking value of the i-th event in the j-th dimension;

[0026] The total score for each event is calculated using a weighted summation algorithm: Score_i = Σ(Wi × Pij);

[0027] A false alarm filtering threshold Th is introduced. If Score_i < Th, it is determined to be a false alarm and filtered.

[0028] Sort the unfiltered events by Score_i and output the processing results.

[0029] In one specific implementation, calculating the total score for each event using a weighted summation algorithm further includes:

[0030] The weighted sum is optimized using fuzzy inference logic, and the weight factor γi ∈ [0,1] is defined to represent the importance of different dimensions;

[0031] The fitness of an event to each dimension is calculated using the fuzzy membership function: Fij = tanh(α × Wij + β), where α and β are learning parameters.

[0032] The weighted total score formula is Score_i = Σ(Fij × γi);

[0033] A dynamic adjustment mechanism is set up to update the γi and α, β parameters based on historical feedback.

[0034] In one specific implementation, the credibility assessment of the preprocessed alarm data based on the multidimensional feature extraction results further includes:

[0035] Obtain the device correlation A1, location correlation A2, and time sensitivity A3 for each alarm;

[0036] Construct a triplet model D = (A1, A2, A3) based on historical violation patterns;

[0037] Calculate the overall confidence level C = k1×A1 + k2×A2 + k3×A3, where k1, k2, and k3 are normalization coefficients;

[0038] If C < Th, mark it as low confidence and initiate a duplicate verification mechanism.

[0039] In one specific implementation, the calculation of the overall confidence level C = k1×A1 + k2×A2 + k3×A3 further includes:

[0040] Introducing a dynamic compensation factor Δ, we define Δ = 1 / (1 + exp(β×(t -Tc))), where β is the sensitivity parameter, t is the current time, and Tc is the target time;

[0041] The modified C = C_prev × Δ;

[0042] Δ is corrected using a time-window moving average algorithm to ensure no sudden changes.

[0043] The intelligent integrated panoramic monitoring terminal for combating violations in the power industry includes:

[0044] The data acquisition module acquires multi-source heterogeneous violation alarm data and performs preprocessing.

[0045] The credibility assessment module evaluates the credibility of the preprocessed alarm data based on the multi-dimensional feature extraction results.

[0046] The weight allocation module dynamically allocates the credibility weight of each violation event based on the credibility assessment results;

[0047] The filtering module prioritizes and filters false alarms based on the credibility weights of the violation events.

[0048] This disclosure provides a data processing method for an intelligent integrated panoramic monitoring terminal for combating violations in the power industry. The method includes: acquiring and preprocessing multi-source heterogeneous violation alarm data; evaluating the credibility of the preprocessed alarm data based on multi-dimensional feature extraction results; dynamically allocating credibility weights for each violation event based on the credibility evaluation results; and prioritizing and filtering the violation events according to the credibility weights. This disclosure solves the problem of excessive alarms caused by high low-confidence false alarm rates by dynamically allocating credibility weights for violation events based on multi-source heterogeneous violation alarm data in the power industry. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the exemplary embodiments of this disclosure, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this disclosure and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a flowchart of the data processing method for the intelligent integrated panoramic monitoring terminal for combating traffic violations;

[0051] Figure 2This is a block diagram of an intelligent integrated panoramic monitoring terminal for combating traffic violations. Detailed Implementation

[0052] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of this application. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0053] The data processing method for the intelligent integrated panoramic monitoring terminal for combating violations in the power industry, as described in this invention, aims to achieve accurate identification and efficient response to violation events through the fusion processing and dynamic reliability assessment of multi-source heterogeneous violation alarm data. The core steps of this method are explained in detail below with reference to specific implementation scenarios. Figure 1 As shown, the data processing method of the intelligent integrated panoramic monitoring terminal for combating violations in the power industry according to the present invention includes:

[0054] S101: Acquire multi-source heterogeneous violation alarm data and perform preprocessing.

[0055] In actual power operation sites, the intelligent integrated panoramic monitoring terminal for preventing violations typically consists of a 360° panoramic monitoring terminal, an intelligent early warning receiver (such as a smart bracelet), and an intelligent panoramic monitoring and control system working together. The acquired violation alarm data comes from diverse sources and has heterogeneous formats. In specific implementation, a high-definition binocular camera first collects on-site video streams. The built-in image recognition algorithm based on convolutional neural networks (CNN) (such as YOLOv4 or Faster R-CNN) performs real-time analysis on scenarios such as safety briefings, working at heights, and hot work, generating image-level alarm data for behaviors such as personnel not wearing safety helmets, improperly worn safety belts, and climbing unmanned ladders. Simultaneously, the intelligent early warning receiver acquires the worker's BeiDou positioning information, heart rate monitoring data, and identity verification check-in records through near-field Internet of Things (NFC) communication technology (such as a ZigBee and Bluetooth fusion network), forming personnel status alarm data. In addition, the system also accesses environmental parameter data such as temperature, humidity, and gas concentration output from third-party devices (such as environmental monitoring sensors).

[0056] After acquiring the aforementioned multi-source data, the system executes a unified data preprocessing workflow. For image-based alarm data, format normalization is first performed, converting all video frames to JPEG or H.264 encoding format. Mean or median filtering is then used to remove noise interference introduced during acquisition due to lighting variations or device jitter, ensuring the accuracy of subsequent feature extraction. For location and physiological parameter data, the system employs timestamp alignment and interpolation to address data loss caused by communication delays or signal interruptions, forming a complete spatiotemporal sequence. For structured data (such as device ID, alarm type, and time stamp), the system establishes a unified data dictionary, mapping fields from different sources to a standard data model. For example, "not wearing a helmet" output by device A and "helmet missing" output by device B are uniformly categorized as the standard event type "helmet not worn." After preprocessing, all data is stored in a distributed time-series database in a structured format, providing high-quality input for subsequent multi-dimensional feature extraction.

[0057] S102: The credibility of the preprocessed alarm data is evaluated based on the multidimensional feature extraction results.

[0058] To reduce false alarms caused by errors in data from a single source, this invention introduces a multi-dimensional feature extraction mechanism to evaluate the credibility of each alarm from multiple dimensions, such as device correlation, location correlation, time sensitivity, and historical pattern matching.

[0059] In practical implementation, taking the typical violation of "not wearing a safety belt while working at height" as an example, the system first extracts the camera ID corresponding to the alarm. It then calculates the camera's detection accuracy and failure rate in historical records based on device correlation, forming a device trust coefficient A1. If the camera has recently been calibrated and its recognition accuracy is higher than 95%, A1 is assigned a higher value (e.g., 0.9). Secondly, the system obtains real-time BeiDou positioning data at the time of the alarm and compares it with the preset electronic fence of the work area. If the personnel's location is within the fence boundary and the alarm content is consistent with the safety requirements of the work area, the location correlation A2 is assigned a value of 0.8 or higher; if the positioning data is missing or significantly deviates from the work area, A2 is correspondingly reduced. Thirdly, the time sensitivity A3 is quantified based on the historical patterns of the time period in which the violation occurred. For example, alarms occurring during non-working hours (e.g., late at night) have a lower time sensitivity weight, while those occurring during peak working hours have a higher weight.

[0060] After extracting the features of the three dimensions mentioned above, the system constructs a triplet model D=(A1, A2, A3) and calculates the comprehensive confidence level C = k1×A1 + k2×A2 + k3×A3 using a weighted summation method, where k1, k2, and k3 are normalization coefficients that can be dynamically adjusted according to different work scenarios (for example, in the scenario of working at heights, the weight of equipment correlation can be set to 0.5, the weight of location correlation to 0.3, and the weight of time sensitivity to 0.2). To further improve the evaluation accuracy, the system introduces a dynamic compensation factor Δ, which is calculated as Δ = 1 / (1 + exp(β×(t-Tc))), where t is the current time, Tc is the center point of the typical occurrence time window of this type of violation, and β is the sensitivity parameter. If C is lower than the preset threshold Th (e.g., 0.6), the system marks the alarm as "low confidence" and triggers a re-verification mechanism, such as calling the historical video of adjacent cameras for secondary comparison, or sending a confirmation prompt to the operator through the intelligent warning receiver to avoid misjudgment caused by the deviation of a single data source.

[0061] In another implementation, the credibility assessment of the preprocessed alarm data based on the multidimensional feature extraction results specifically employs a method based on deep feature fusion and spatiotemporal consistency verification. Specifically, the system first inputs the preprocessed image-based alarm data into a pre-trained lightweight convolutional neural network (such as an improved MobileNetV3) for deep feature extraction, obtaining a deep visual feature vector F_vis, including target contours, textures, color distribution, and key point heatmaps. Simultaneously, the BeiDou positioning trajectory, heart rate change sequence, and equipment vibration data reported by the intelligent early warning receiver are converted into a behavioral feature vector F_iot via a time-series encoder (such as an LSTM network). Subsequently, the system performs cross-modal feature concatenation of F_vis and F_iot, and calculates the contribution weight of each modality feature to the current alarm event through an attention mechanism, forming a multidimensional fusion feature vector F_fusion. Building upon this foundation, the system introduces a spatiotemporal consistency verification module: On one hand, it performs kernel density estimation by comparing the timestamp corresponding to the current alarm with the occurrence time periods of historical violations within the same work area. If the current time falls within a historically high-incidence period, a higher time consistency score T_score is assigned. On the other hand, it uses BeiDou positioning data to calculate the spatial distance between the current alarm location and the location of the most recent confirmed violation of the same type, mapping it to a spatial consistency score S_score using a Gaussian kernel function. Finally, the system inputs the fused feature vector F_fusion into a credibility regression model trained based on gradient boosting tree (XGBoost), and uses T_score and S_score as external correction factors to calculate the comprehensive credibility C = C_model × (T_score^γ1) × (S_score^γ2) through a weighted product, where C_model is the basic confidence probability output by the model, and γ1 and γ2 are balancing coefficients. This achieves a multi-dimensional credibility comprehensive evaluation of each alarm data from visual semantics, sensor behavior, to spatiotemporal patterns.

[0062] S103: Dynamically assign credibility weights to each violation event based on the credibility assessment results.

[0063] After completing the credibility assessment of each alarm, the present invention further dynamically allocates the credibility weight of each violation event based on the historical false alarm rate and time decay characteristics to achieve differentiated processing.

[0064] In practice, the system maintains a database of false alarm rates for different types of violations, recording the cumulative number of alarms and confirmed false alarms for each type of violation (such as "not wearing a safety helmet," "oxygen and acetylene cylinders turned upside down," "crossing a fence," etc.) over the past 30 days. It then calculates the historical false alarm rates R1, R2, ..., Rm for each type of violation in the current time period. Taking "crossing a fence" as an example, if this type of behavior is frequently falsely reported in the past week (e.g., due to visual algorithm misjudgment caused by changes in lighting), its false alarm rate Rj is high, and its weight will be automatically reduced.

[0065] Simultaneously, the system introduces a time decay factor τ to calculate the dynamic weighting coefficient Wi = e^(τ×t), where t is the time interval between the current time and the most recent actual occurrence (confirmed manually or by the system) of the violation event. If a certain type of violation has not been actually confirmed for a long period, its Wi value will decrease exponentially with increasing t, indicating that the reference value of this type at the current stage is reduced. Combining the comprehensive confidence level C and the false alarm rate Rj obtained in the second step, the system generates a comprehensive evaluation value Wij = (C / (Rj+ε)) ×Wi, where ε is a small constant to avoid division by zero. For events with high confidence and low false alarm rate, the Wij value increases significantly; conversely, if the confidence level is low and the false alarm rate is high, Wij approaches zero.

[0066] Furthermore, to enhance the ability to identify high-risk scenarios, this invention employs a nonlinear function to optimize the comprehensive evaluation value: Wij' = C × log(Rj+1) × Wi. This form amplifies the weight contribution when Rj is small (i.e., low false alarm rate type), enhancing the algorithm's sensitivity to stable event types. In addition, the system introduces spatial dimension information Sgeo to cluster and score similar violations occurring within a short period in the same area (such as the same tower or the same transmission line), forming a spatial correction factor δ. If multiple cameras simultaneously report "not wearing a helmet" behavior in the same area, δ takes a positive value (e.g., 0.2), further increasing the weight of that event. The final weight value is calculated as Wij_final = Wij' × (1+δ), and after normalization, the sum of the weights of all violations is 1, facilitating subsequent priority ranking.

[0067] In another implementation, the dynamic allocation of credibility weights for each violation event based on the credibility assessment results employs a dynamic weight allocation method that links fuzzy Bayesian inference with scene risk levels. Specifically, the system first constructs a scene risk level matrix R_scene based on the current work scene type (such as safety briefing, working at height, and hot work) and on-site environmental parameters (such as wind speed, temperature and humidity, and work height), dividing each scene into three risk levels: high, medium, and low. An initial risk coefficient ρ is preset for each level (0.8 for high risk, 0.5 for medium risk, and 0.2 for low risk). Subsequently, for each violation event, the system uses the comprehensive confidence level C calculated in the second step as the prior probability, and constructs a likelihood function P(false alarm|C) by combining it with the historical false alarm rate Rj of that violation type. The posterior probability P(true violation|C) is then updated using the Bayesian formula: P(true violation|C) = (C × P(true violation prior)) / (C × P(true violation prior) + (1-C) × P(false alarm prior)), where P(true violation prior) and P(false alarm prior) are dynamically updated based on historical statistical data. On this basis, the system introduces a dynamic adjustment factor λ = ρ × (1 + η×log(1+N)), where η is the scene sensitivity coefficient and N is the cumulative number of reports of the same type of event in the same area within the current time period, used to suppress excessive weight expansion in alarm storm scenarios. Finally, the system calculates the dynamic credibility weight W = P(real violation|C) × λ for each violation event, and normalizes all events within the same time window to give events with high confidence, high scenario risk, and low false alarm history a significantly higher weight allocation. At the same time, the system uses a Bayesian framework to achieve adaptive fusion of confidence and historical false alarm rate, avoiding the weight allocation bias caused by a single threshold truncation.

[0068] S104: Prioritize and filter false alarms for violation events based on the credibility weight.

[0069] After completing the dynamic weight allocation, the system constructs a dynamic priority matrix P = {Pij | i=1..n, j=1..m}, where Pij represents the ranking value of the i-th violation event in the j-th dimension (such as confidence level, false alarm rate, spatial density, time urgency, etc.). The total score for each event is calculated using a weighted summation algorithm: Score_i = Σ(Wij_final ×Pij). To prevent low-confidence events from interfering with scheduling resources, the system sets a false alarm filtering threshold Th (e.g., 0.15). If Score_i < Th, it is directly determined as a false alarm, filtered from the alarm list, and not pushed to the system.

[0070] For the remaining events, the system sorts them from highest to lowest Score_i and outputs the final priority alarm sequence. In practical applications, high-priority events (e.g., Score_i > 0.8) are simultaneously pushed via pop-up windows, vibration, and voice alerts to the handheld terminals of on-site management personnel and the intelligent early warning receivers (wristbands) through the intelligent panoramic monitoring and control system, and the event information is automatically written into the supervision work order module, forming a closed-loop processing. Medium-priority events (e.g., Score_i between 0.5 and 0.8) are only pushed to the on-duty monitoring center, where the on-duty personnel confirm them again by referring to the video footage. Low-priority but unfiltered events (Score_i between 0.15 and 0.5) are stored in the historical database for subsequent model training and false alarm rate updates.

[0071] In another implementation, the prioritization and false alarm filtering of violation events based on the credibility weights employs a collaborative filtering method based on dynamic time windows and event graph association. Specifically, the system first constructs a sliding time window (e.g., a window length of 30 seconds), initially sorts all violation events generated within the window according to the credibility weights calculated in the third step, and forms a queue of events to be processed. Subsequently, the system calls a pre-constructed event association graph, which uses historically confirmed violation events as core nodes and encodes the spatiotemporal association relationships between different event types through a graph neural network (e.g., "not wearing a helmet" and "crossing a fence" are highly correlated in time, and "oxygen acetylene cylinder upside down" and "smoking on site" are strongly correlated in space). For the set of events within the current window, the system calculates the centrality score of each event in the association graph and the co-occurrence strength with other events in the same window. If a low-weight event has a strong correlation with multiple high-weight events, its weight is compensated and increased through the graph convolution propagation mechanism. Conversely, if a high-weight event is isolated in the association graph and its confidence source is singular, a temporary deweighting mechanism is triggered. Building upon this foundation, the system introduces a dynamic priority queue management strategy: for events with a weight exceeding the first threshold (e.g., 0.7), they are immediately pushed to the on-site intelligent early warning receiver and monitoring screen for real-time alarms; for events with a weight between the second and first thresholds (e.g., 0.4 to 0.7), the system places them in a delayed confirmation queue and continuously receives new events for correlation verification during window sliding. If a new event with a high correlation to the original event appears in a subsequent window, the original event is automatically promoted to a high-priority queue; for events with a weight below the second threshold (e.g., 0.4), the system uses an event correlation graph for false alarm filtering. If such an event has no effective supporting nodes in the correlation graph and its weight is below the threshold for three consecutive time windows, it is directly discarded and recorded as a suspected false alarm for subsequent model optimization. Through this dynamic time window and event graph collaborative filtering method, the system effectively suppresses the interference of isolated low-confidence alarms on monitoring personnel while ensuring rapid response to high-confidence events, achieving refined sorting and intelligent filtering of violation events.

[0072] To further enhance the robustness of the ranking, this invention introduces fuzzy inference logic into the weighted summation algorithm. A weight factor γi ∈ [0,1] is defined to represent the importance of different dimensions, and a fuzzy membership function Fij = tanh(α×Wij+β) is used to calculate the fitness of an event for each dimension, where α and β are parameters that can be updated through self-learning from historical feedback data. The final total score formula is adjusted to Score_i = Σ(Fij × γi), enabling the system to continuously optimize the importance allocation between dimensions based on on-site operational data, avoiding the limitations of manually setting weights. Through these steps, this invention effectively solves the problem of rampant false alarms from multi-source heterogeneous alarm data in power scenarios, achieving dynamic allocation of credibility and efficient response to violation events.

[0073] In one specific implementation, the credibility assessment of the preprocessed alarm data based on the multi-dimensional feature extraction results further employs a multi-source anomaly detection score fusion and threshold trigger enhancement mechanism. The system first obtains anomaly detection scores for the same target event from different data sources. For example, it obtains the visual detection confidence S_vis of the "not wearing a safety helmet" behavior detected by the YOLOv4 algorithm from a 360° panoramic monitoring terminal, the heart rate mutation index S_hr (anomaly score based on heart rate variability) of the worker from an intelligent early warning receiver, and the deviation degree S_loc (normalized offset) of the worker's position from the electronic fence from the BeiDou positioning system, thus forming an anomaly score sequence S1, S2, …, Sn. Subsequently, the system calculates the average anomaly confidence value M_avg = (S1+S2+…+Sn) / n for each source data, and records the maximum value max(Si). Based on this, the system introduces a weighted fusion formula C = α×M_avg + β×max(Si), where α and β are dynamically adjusted weighting coefficients. Initially, α can be set to 0.6 and β to 0.4, ensuring the fusion result balances overall average performance while highlighting the contribution of a single high-confidence source. When the calculated fusion confidence C exceeds a preset threshold T (e.g., T=0.75), the system automatically triggers a confidence enhancement mechanism. Specifically, the system uses adjacent cameras or supplementary lighting devices to perform secondary re-captures and re-identification of the same area, obtaining additional verification data. Simultaneously, the system sends an active confirmation command to the intelligent early warning receiver of on-site personnel, requiring them to confirm their safety status via button feedback or biometric identification, and feeding the confirmation result back to the confidence assessment module as an enhancement factor. If C does not exceed the threshold T, the original confidence level remains unchanged, and the system directly proceeds to the subsequent weight allocation stage. Through the aforementioned enhanced mechanism of multi-source fusion and threshold triggering, misjudgments caused by accidental errors from a single data source are effectively avoided. At the same time, the introduction of human-computer interaction verification in high uncertainty scenarios significantly improves the reliability of alarm credibility assessment.

[0074] In one specific implementation, the dynamic allocation of credibility weights for each violation event based on the credibility assessment results employs a combination of historical false alarm rate statistics and time decay weighting. The system first maintains a rolling time window (e.g., the last 30 days) in the database for each violation type (e.g., "not wearing a helmet," "crossing a hard fence," "oxygen acetylene cylinder upside down," etc.), and calculates the historical false alarm rates R1, R2, ..., Rm for each type in real time within the current time period. The false alarm rate is defined as the proportion of the number of times a violation of this type is manually or systematically verified as a false alarm to the total number of alarms for that type. For behaviors like "crossing a hard fence," which are easily affected by light and shadow interference leading to false alarms, if the false alarm rate is as high as 0.4 in the past week, then Rj is set to 0.4; while for behaviors like "not wearing a seatbelt," which are more stable in identification, the false alarm rate may only be 0.05. Based on this, the system introduces a time decay factor τ (τ is generally between 0.1 and 0.5, and the specific value can be adjusted according to the frequency of on-site data updates) and calculates the dynamic weighting coefficient Wi = e^(τ×t), where t is the time interval (in hours) between the current time and the last time that the violation type was confirmed as a real violation. For example, if a certain type of violation was just confirmed within the last 24 hours, then t is small and Wi approaches 1; if the violation has not been confirmed for more than 7 days, then t is large and the Wi index decays to below 0.1, indicating that the reference value of this type of violation is reduced at the current stage. Subsequently, the system combines the comprehensive confidence level C, the historical false alarm rate Rj, and the dynamic weighting coefficient Wi obtained in the second step to generate a comprehensive evaluation value Wij = (C / (Rj+ε)) ×Wi, where ε is a small constant (such as 0.01) to prevent division by zero errors. Taking "not wearing a helmet" as an example, if C=0.85, Rj=0.05, and Wi=0.95, then Wij = (0.85 / (0.05+0.01))×0.95 ≈ 13.46; while for "crossing a fence," if C=0.60, Rj=0.40, and Wi=0.30, then Wij = (0.60 / (0.40+0.01))×0.30 ≈ 0.44. Finally, the system normalizes the Wij of all events, so that the sum of the confidence weights of all violations within the same time window is 1, thereby achieving dynamic weight allocation. This ensures that events with low false alarm rates, high confidence, and recent confirmation history receive significantly higher weights, while the weights of events with high false alarm rates or long-term unconfirmed events are effectively suppressed.

[0075] In one specific embodiment, the calculation of the comprehensive evaluation value Wij as described in claim 3 further employs a nonlinear function to enhance the contribution of low false alarm rate events, and is optimized in conjunction with a spatial clustering correction factor. Specifically, the system first replaces the original linear form Wij = (C / (Rj+ε)) × Wi with the nonlinear form Wij = C × log(Rj+1) × Wi, where the logarithmic function log(Rj+1) has a large gradient change when Rj is low (i.e., event types with low false alarm rates), which can significantly amplify the weight contribution of low false alarm rate events. For example, for an event with a false alarm rate Rj = 0.05, log(1.05) ≈ 0.021, while for an event with a false alarm rate Rj = 0.40, log(1.40) ≈ 0.146, and the ratio of the two is approximately 6.95. In fact, low false alarm rate events achieve a higher nonlinear amplification effect. Subsequently, the system sets a threshold λ (e.g., λ=0.5) to filter high-risk events. If the calculated Wij > λ, the event is marked as a high-priority candidate event and directly enters the subsequent sorting process without waiting for further verification. Based on this, the system introduces spatial dimension information (Sgeo) to cluster and score similar violations within the region: using BeiDou positioning data, the system delineates a spatial radius (e.g., 50 meters) centered on the current alarm location, counts the number N_spatial of similar violations occurring within this radius and in the current time window (e.g., the previous 10 minutes), and calculates the spatial density factor δ = 1 / (1 + e^(-k×(N_spatial-μ))), where k is the sensitivity coefficient (e.g., 0.5) and μ is the clustering initiation threshold (e.g., 2). When N_spatial reaches or exceeds μ, δ approaches a positive value (up to 0.5), indicating the existence of spatially clustered violations within the region, and the weight of such events should be increased; conversely, if N_spatial is 0 or 1, δ approaches 0, and no further correction is made. Finally, the system generates a spatially corrected comprehensive evaluation value Wij_new = Wij × (1+δ), achieving dynamic enhancement of weights in the spatial dimension. By combining the aforementioned nonlinear function with spatial clustering correction, the system achieves more significant weight differentiation in low false alarm rate events, while leveraging spatial clustering characteristics to strengthen attention to regional cluster violations, effectively improving the accuracy of weight allocation and alarm response sensitivity in high-risk scenarios.

[0076] In one specific implementation, the prioritization and false alarm filtering of violation events based on the credibility weight employs a dynamic sorting method based on a multi-dimensional priority matrix and a weighted total score. The system first constructs a dynamic priority matrix P = {Pij | i=1..n, j=1..m} for each violation event, where n is the total number of events to be processed within the current time window, and m is the preset number of evaluation dimensions. Evaluation dimensions include, but are not limited to: the comprehensive credibility weight Wij of the event (calculated from the results of claim 3 or 4), the historical hazard level H of the event type (e.g., violations related to working at height are assigned a value of 0.9, and improper attire is assigned a value of 0.3), the safe distance D between the event location and critical equipment (the closer the distance, the higher the value), the number of times the event recurs within the time window F (the more recurrences, the higher the value), and the risk level R_scene of the current work scenario. The system calculates ranking values ​​for each dimension. For example, for the comprehensive credibility weight dimension, the Wij values ​​of all events within the window are sorted from high to low, with the first-ranked event assigned a value of 1, and the k-th ranked event assigned a value of (1 / k) or mapped to the [0,1] interval using linear normalization. Subsequently, the system calculates the total score for each event using a weighted summation algorithm: Score_i = Σ(W_dim_j × Pij), where W_dim_j is the preset weight coefficient for the j-th dimension (e.g., 0.5 for credibility weight, 0.2 for historical hazard level, 0.15 for safe distance, 0.1 for repetition count, and 0.05 for scenario risk). This weight coefficient can be dynamically adjusted according to different operational scenarios. Based on this, the system introduces a false alarm filtering threshold Th (e.g., Th=0.3). If Score_i < Th, the event is determined to be a false alarm, directly filtered, and not entered into the alarm queue; if Score_i ≥ Th, the event is retained. Finally, the system sorts all retained events in descending order of Score_i to form a priority alarm list. The sorting results are pushed to the alarm dashboard of the intelligent panoramic monitoring and control system in real time. High-priority events (such as Score_i > 0.8) simultaneously trigger vibration and voice alerts at the on-site intelligent early warning receiver.Taking a real-world scenario as an example, at a high-altitude work site, both the "safety belt not worn" event (Wij=0.85, H=0.9, D=0.8, F=1, R_scene=0.9) and the "person falls to the ground" event (Wij=0.45, H=0.8, D=0.5, F=1, R_scene=0.9) occur simultaneously. Calculations show that Score_safety belt ≈ 0.85×0.5+0.9×0.2+0.8×0.15+1×0.1+0.9×0.05=0.785 and Score_falls to the ground ≈ 0.45×0.5+0.8×0.2+0.5×0.15+1×0.1+0.9×0.05=0.505. Both are greater than Th=0.3, but the safety belt event has a significantly higher priority than the falling event. The system prioritizes the former, achieving accurate identification and priority response to critical violations.

[0077] In one specific implementation, the weighted summation algorithm of claim 5 is further optimized using fuzzy inference logic to enhance its adaptability to weight allocation in complex scenarios. The system first defines a weight factor γi ∈ [0,1] to represent the importance of different dimensions. This factor no longer uses a fixed preset value but is dynamically generated by the fuzzy inference system. Specifically, the system constructs a fuzzy rule base based on current scenario features (such as job type, personnel density, and historical false alarm rate). For example, if "job type is climbing" and "historical false alarm rate is low," then the "credibility dimension weight γ1" is high (takes 0.7); if "job type is general" and "personnel density is high," then the "repetition count dimension weight γ4" is medium (takes 0.5). The output of the rule base is defuzzified to obtain the real-time weight factors for each dimension. Based on this, the system uses a fuzzy membership function to calculate the fitness of each event to each dimension: Fij = tanh(α×Wij + β), where α and β are learnable parameters. Initially, they can be set to α=2.0 and β=-1.0, so that Fij approaches 0 when Wij is low and approaches 1 when Wij is high, forming a smooth membership mapping. For example, for the "not wearing a helmet" event, if its confidence weight Wij=0.85, then Fij = tanh(2.0×0.85-1.0)=tanh(0.7)≈0.604; for the "crossing the fence" event, if Wij=0.35, then Fij=tanh(2.0×0.35-1.0)=tanh(-0.3)≈-0.291. In this case, the system takes the absolute value or adjusts it through the activation function for subsequent calculations. Subsequently, the system calculates the weighted total score using the formula Score_i = Σ(Fij × γi), achieving a fuzzy multi-dimensional score. To further enhance adaptability, the system introduces a dynamic adjustment mechanism: after each round of alarm processing, the system collects feedback results from manual review (e.g., whether an alarm is confirmed as a genuine violation or marked as a false alarm). Using this feedback as a monitoring signal, the system incrementally updates the α and β parameters and the output weights of the fuzzy rule base using gradient descent, allowing the fuzzy membership function and weight factors to gradually fit the actual distribution characteristics of the current site. Through this fuzzy inference and parameter self-learning mechanism, the system can dynamically optimize the contribution of each dimension according to different operational scenarios and real-time feedback, avoiding the problem of fixed weights performing poorly in complex and changing environments, and significantly improving the robustness and adaptability of priority ranking.

[0078] In one specific implementation, the credibility assessment of the preprocessed alarm data based on the multi-dimensional feature extraction results further employs a triplet model based on device correlation, location correlation, and time sensitivity, combined with a repeated verification mechanism to achieve secondary confirmation of low-confidence alarms. The system first extracts three core features for each alarm event: device correlation A1, location correlation A2, and time sensitivity A3. Device correlation A1 is quantified based on the historical performance data of the terminal device that generated the alarm. The system maintains a statistical table of the recognition accuracy of each camera and the data reporting success rate of each smart warning receiver. Taking the "not wearing a safety helmet" alarm as an example, if the camera that generated the alarm has an accuracy rate of 96% in the past 100 detections, then A1 is set to 0.96; if a smart bracelet experiences frequent abnormal fluctuations in heart rate data over the past week but none correspond to actual violations, its data confidence is low, and A1 is correspondingly reduced. Location relevance A2 is calculated based on the matching degree between BeiDou positioning data and electronic fences. The system divides the work area into core work area, edge area, and restricted area. Alarms generated in the core work area are assigned A2=1.0, those in the edge area are assigned A2=0.7, and those in the restricted area are assigned A2=0.3. If positioning data is missing, A2 is taken as 0.5. Time sensitivity A3 is calculated based on the historical occurrence time distribution of violation types. The system generates a probability density function for each violation type within 24 hours using the kernel density estimation method. For violations of the "working at height" category, if the current time is during the peak period of 9:00-17:00 on weekdays, A3 is taken as above 0.9; if it is during the early morning hours, A3 is taken as below 0.2. After extracting the aforementioned three-dimensional features, the system constructs a triplet model D=(A1, A2, A3) and calculates the overall confidence level C = k1×A1 + k2×A2 + k3×A3 using a linear weighted formula, where k1, k2, and k3 are normalization coefficients (e.g., k1=0.4, k2=0.35, k3=0.25), ensuring that the value of C is within the range of [0,1]. If the calculated C is lower than the preset threshold Th (e.g., Th=0.5), the system marks the alarm as "low confidence level" and automatically initiates a duplicate verification mechanism. The repeated verification mechanism specifically includes: on the one hand, the system calls up 5-second video clips before and after the camera, and uses a more refined target tracking algorithm (such as DeepSORT) to trace the target's trajectory back to verify whether the target has indeed committed a violation; on the other hand, the system sends a collaborative confirmation command to other nearby intelligent early warning receivers. If the adjacent workers confirm the alarm is true through their wristbands, the C value is updated as enhanced evidence; if no confirmation is received within the preset waiting time (such as 10 seconds) and the video backtracking fails to provide effective support, the system marks the alarm as "pending manual review" and pushes it to the confirmation queue in the monitoring center for manual judgment by the on-duty personnel.By combining the triplet model with the repeated verification mechanism, the system effectively reduces the interference of low-confidence alarms on real-time monitoring, while multi-level verification ensures that the false negative rate of key violations is kept within an acceptable range.

[0079] In one specific implementation, the calculation of the comprehensive confidence level C = k1×A1 + k2×A2 + k3×A3 as described in claim 7 further introduces a dynamic compensation factor Δ to achieve fine-tuning of the impact on the time dimension. The system first defines the dynamic compensation factor Δ = 1 / (1 + exp(β×(t - Tc))), where β is a sensitivity parameter (the value range is usually set to 0.2 to 0.8, which can be dynamically adjusted according to the fluctuation of the field data), t is the current alarm occurrence time (a continuous time variable in hours), and Tc is the center point of the historical high-incidence time window for this type of violation (for example, for the behavior of "not wearing a safety belt while working at height", according to historical data, its high-incidence period is from 9:00 am to 11:00 am, then Tc is taken as 10.0 hours). The sigmoid function makes Δ approach 1 when t is much less than Tc; Δ is approximately 0.5 when t is close to Tc; and Δ approaches 0 when t is much greater than Tc, thus achieving dynamic decay of the confidence level of violations at different time periods. Subsequently, the system multiplies the initial comprehensive confidence level C_prev calculated in claim 7 with the dynamic compensation factor to obtain the corrected confidence level C = C_prev × Δ. For example, if an alarm for "not wearing a safety belt while working at height" occurs at 2:00 AM (t=2.0), and β=0.5, Tc=10.0, then Δ = 1 / (1+exp(0.5×(2-10))) = 1 / (1+exp(-4)) ≈ 1 / (1+0.0183) ≈ 0.982, the confidence level only decreases slightly. If the alarm occurs at 4:00 PM (t=16.0), then Δ = 1 / (1+exp(0.5×(16-10))) = 1 / (1+exp(3)) ≈ 1 / (1+20.09) ≈ 0.047, the confidence level is significantly reduced, reflecting the historical pattern that the probability of this type of violation is extremely low during this period. Based on this, the system further introduces a time window moving average algorithm to correct Δ, ensuring that the confidence level will not change drastically due to a single time point shift. Specifically, the system maintains a sliding time window of length L (e.g., 24 hours) for each type of violation, records the historical Δ value sequence at each moment within the window, and calculates the smoothing correction value Δ_smooth = (Δ_current + ΣΔ_history) / (L+1) at the current moment, where ΣΔ_history is the sum of all historical Δ values ​​within the window.By replacing the original Δ with Δ_smooth in the calculation of C = C_prev × Δ_smooth, the system can effectively suppress confidence jumps caused by single-point data fluctuations or Tc setting deviations, making the overall confidence level exhibit a smooth and gradual change characteristic in the time dimension. This not only follows the historical time distribution pattern of violations but also avoids excessive punishment for occasional abnormal periods, thereby improving the stability and reliability of the confidence assessment in the time dimension.

[0080] In addition, such as Figure 2 As shown, the present invention also provides an intelligent integrated panoramic monitoring terminal 200 for combating violations in the power industry, which includes:

[0081] Data acquisition module 201 acquires multi-source heterogeneous violation alarm data and performs preprocessing;

[0082] The credibility assessment module 202 assesses the credibility of the preprocessed alarm data based on the multi-dimensional feature extraction results.

[0083] The weight allocation module 203 dynamically allocates the credibility weight of each violation event based on the credibility assessment results.

[0084] The filtering module 204 prioritizes and filters false alarms for violation events based on the credibility weight.

[0085] The functions of each module of the intelligent integrated panoramic monitoring terminal 200 for combating violations in the power industry of the present invention have been described above with reference to the method embodiments, and will not be repeated here.

[0086] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this invention, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A data processing method for anti-violation intelligent integrated panoramic monitoring terminal, characterized in that, include: Acquire multi-source heterogeneous traffic violation alarm data and perform preprocessing; The credibility of the preprocessed alarm data is evaluated based on the multidimensional feature extraction results. The credibility weight of each violation event is dynamically assigned based on the credibility assessment results. Based on the aforementioned credibility weights, violation events are prioritized and false alarms are filtered out. in The credibility assessment of the preprocessed alarm data based on the multidimensional feature extraction results further includes: Obtain the anomaly detection scores S1, S2, ..., Sn for each source data. Calculate the average outlier confidence value M_avg = (S1 + S2 + ... + Sn) / n for each source data; The following formula is used to determine whether credibility needs to be enhanced: C = α × M_avg + β × max(Si), where α and β are weighting coefficients, and Si is the anomaly score for each source; If C > T, the credibility enhancement mechanism is triggered; otherwise, the original value remains unchanged. The dynamic allocation of credibility weights for each violation event based on the credibility assessment results further includes: Obtain the historical false alarm rates R1, R2, ..., Rm for each type of traffic violation within the current time period; The dynamic weighting coefficient Wi = e^(τ × t) is calculated based on the time decay factor τ, where t is the time interval between the current time and the last occurrence of the violation. The confidence score C and the false alarm rate Rj are combined to generate a comprehensive evaluation value Wij = (C / (Rj + ε)) × Wi, where ε is a small constant to avoid division by zero; Normalize Wij to determine the final weight value.

2. The anti-vandalism intelligent integrated panoramic monitoring terminal data processing method according to claim 1, characterized in that, The method of generating a comprehensive evaluation value Wij = (C / (Rj + ε)) × Wi by combining the confidence score C and the false alarm rate Rj further includes: The contribution of using a nonlinear function to improve the false alarm rate is: Wij = (C × log(Rj + 1)) × Wi; Set a threshold λ to filter high-risk events. If Wij > λ, it is considered a high priority. By introducing spatial dimension information (Sgeo), clustering and scoring similar violations within a region, a spatial correction factor δ is formed. The final weight value is calculated using the formula: Wij_new = Wij × (1 + δ).

3. The data processing method for the intelligent integrated panoramic monitoring terminal for anti-violation behavior according to claim 1, characterized in that, The prioritization and false alarm filtering of violation events based on the credibility weight further includes: Construct a dynamic priority matrix P = {Pij | i=1...n, j=1..m}, where Pij represents the ranking value of the i-th event in the j-th dimension; The total score for each event is calculated using a weighted summation algorithm: Score_i = Σ(Wi × Pij); A false alarm filtering threshold Th is introduced. If Score_i < Th, it is determined to be a false alarm and filtered. Sort the unfiltered events by Score_i and output the processing results.

4. The data processing method for the intelligent integrated panoramic monitoring terminal for anti-violation behavior according to claim 3, characterized in that, The calculation of the total score for each event using a weighted summation algorithm further includes: The weighted sum is optimized using fuzzy inference logic, and the weight factor γi ∈ [0,1] is defined to represent the importance of different dimensions; The fitness of an event to each dimension is calculated using the fuzzy membership function: Fij = tanh(α × Wij + β), where α and β are learning parameters. The weighted total score formula is Score_i = Σ(Fij × γi); A dynamic adjustment mechanism is set up to update the γi and α, β parameters based on historical feedback.

5. The data processing method for the intelligent integrated panoramic monitoring terminal for anti-violation behavior according to claim 1, characterized in that, The credibility assessment of the preprocessed alarm data based on the multidimensional feature extraction results further includes: Obtain the device correlation A1, location correlation A2, and time sensitivity A3 for each alarm; Construct a triplet model D = (A1, A2, A3) based on historical violation patterns; Calculate the overall confidence level C = k1×A1 + k2×A2 + k3×A3, where k1, k2, and k3 are normalization coefficients; If C < Th, mark it as low confidence and initiate a duplicate verification mechanism.

6. The data processing method for the intelligent integrated panoramic monitoring terminal for anti-violation behavior according to claim 5, characterized in that, The calculation of the overall confidence level C = k1×A1 + k2×A2 + k3×A3 further includes: Introducing a dynamic compensation factor Δ, we define Δ = 1 / (1 + exp(β×(t-Tc))), where β is the sensitivity parameter, t is the current time, and Tc is the target time; The modified C = C_prev × Δ; Δ is corrected using a time-window moving average algorithm to ensure no sudden changes.

7. A smart integrated panoramic monitoring terminal for preventing traffic violations, characterized in that, include: The data acquisition module acquires multi-source heterogeneous violation alarm data and performs preprocessing. The credibility assessment module evaluates the credibility of the preprocessed alarm data based on the multi-dimensional feature extraction results. The weight allocation module dynamically allocates the credibility weight of each violation event based on the credibility assessment results; The filtering module prioritizes and filters false alarms for violation events based on the credibility weights. in The credibility assessment of the preprocessed alarm data based on the multidimensional feature extraction results further includes: Obtain the anomaly detection scores S1, S2, ..., Sn for each source data. Calculate the average outlier confidence value M_avg = (S1 + S2 + ... + Sn) / n for each source data; The following formula is used to determine whether credibility needs to be enhanced: C = α × M_avg + β × max(Si), where α and β are weighting coefficients, and Si is the anomaly score for each source; If C > T, the credibility enhancement mechanism is triggered; otherwise, the original value remains unchanged. The dynamic allocation of credibility weights for each violation event based on the credibility assessment results further includes: Obtain the historical false alarm rates R1, R2, ..., Rm for each type of traffic violation within the current time period; The dynamic weighting coefficient Wi = e^(τ × t) is calculated based on the time decay factor τ, where t is the time interval between the current time and the last occurrence of the violation. The confidence score C and the false alarm rate Rj are combined to generate a comprehensive evaluation value Wij = (C / (Rj + ε)) × Wi, where ε is a small constant to avoid division by zero; Normalize Wij to determine the final weight value.

Images