An intelligent robot inspection task dynamic optimization and real-time exception identification method

Abstract

The application relates to the technical field of robots, in particular to a dynamic optimization and real-time abnormality identification method for intelligent robot inspection tasks. The technical scheme comprises the following steps: based on a three-layer architecture of terminal robots, edge servers and cloud scheduling centers, a two-level inference mechanism of edge preliminary screening and cloud rechecking is adopted, multi-modal confidence fusion is completed through dynamic weights adapted to abnormal types, after time decay sliding window timing correction, a danger score is calculated in combination with equipment importance, abnormal duration and historical frequency, and hierarchical alarm is completed; meanwhile, a three-level closed-loop rearrangement system is constructed, synchronous edge cloud inference power dynamic scheduling is realized, and the inference task allocation of the edge server and the cloud is dynamically adjusted. The core equipment inspection coverage is greatly improved, the abnormal false alarm rate is obviously reduced, the detection accuracy and operation efficiency are considered, and the application can be widely applied to complex industrial inspection scenes such as power, chemical industry and pipe gallery.

Description

Technical Field

[0001] This invention relates to the field of robotics, specifically to a method for dynamic optimization and real-time anomaly identification of intelligent robot inspection tasks. Background Technology

[0002] With the increasing demand for unmanned operations in industrial settings, quadruped inspection robots, with their strong terrain adaptability and flexible load configuration, have gradually replaced manual labor in high-risk and repetitive daily inspection tasks. Currently, most mainstream inspection scheduling platforms adopt a manual pre-programmed static route model, combined with a single vision detection algorithm for anomaly identification, but this approach still has core shortcomings in practical industrial applications.

[0003] The inspection tasks are static and fixed, lacking adaptability to different scenarios: inspection routes, inspection items, and scheduling rules are all pre-configured manually, making it impossible to dynamically adjust strategies based on the robot's remaining battery power, equipment maintenance level, or real-time anomalies. When an anomaly occurs, only alarm notifications are pushed, relying entirely on manual intervention to adjust the task, resulting in delayed response and a high risk of missing key inspection points in extreme scenarios such as low battery or equipment failure. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention provides a method for dynamic optimization and real-time anomaly identification of intelligent robot inspection tasks. Based on a three-layer architecture of terminal inspection robot – site edge server – cloud scheduling center, it establishes an end-to-end intelligent inspection process system through baseline task construction, multi-source state perception, hierarchical multimodal anomaly identification, three-level closed-loop task re-arrangement, and edge-cloud dynamic computing power scheduling. The specific technical solution is as follows: S1. Import the 2D site grid map and mark the movable area and restricted area. Mark the inspection route including waypoints and navigation parameters. Build the equipment inspection item library and equipment importance weight system. Configure task scheduling rules and safety constraint thresholds. Generate executable baseline inspection tasks.

[0005] S2 periodically collects robot operating status, multimodal sensor data and inspection video stream, and synchronizes inspection progress to the cloud; the cloud aggregates alarm data, equipment operation and maintenance status and historical fault records from the entire site, and dynamically updates the overall operating status.

[0006] S3, the two-level multimodal anomaly identification and graded alarm adopts a layered inference architecture of lightweight initial screening at the edge server and high-precision review in the cloud. It achieves accurate alarm grading through multimodal dynamic fusion, time-series correction and multi-factor hazard score calculation.

[0007] S4. Based on the alarm level, trigger corresponding task adjustments to build a three-level closed-loop system of minute-level real-time emergency rescheduling, day-level mid-term strategy optimization, and week / month-level long-term capability iteration, so as to realize the full-link linkage from immediate response to system self-evolution.

[0008] S5. Based on network bandwidth, edge server load, and alarm level, dynamically adjust the allocation of inference tasks between the edge server and the cloud to balance detection accuracy, real-time performance, and bandwidth usage.

[0009] Preferred navigation parameters include: gait type, travel speed, travel direction, and obstacle stopping rules.

[0010] The preferred method is as follows: S3 specific steps are as follows: 1. Lightweight initial screening at the edge; 2. Single-modal basic confidence quantification; 3. Dynamic weighted fusion driven by anomaly type; 4. Sliding window time series correction with time decay; 5. Multi-factor hazard score calculation and alarm classification.

[0011] Preferred: Lightweight initial screening at the edge includes: deploying a lightweight detection model on the edge server, sampling frames of multimodal data at a fixed frame rate for initial screening, and outputting the single-modal confidence score of each modality; when the confidence score of any single modality is higher than the initial screening threshold, it is judged as a suspected anomaly, and a preset number of frame images before and after the corresponding time period and sensor data at a preset time are extracted and uploaded to the cloud; normal detection results only upload readings and status, saving uplink bandwidth.

[0012] Preferred method: Single-modal basic confidence quantification includes: firstly, the detection results of the three core modalities are normalized in the cloud to obtain the single-modal confidence in the range of 0 to 1; the single-modal confidence includes: visible light visual confidence C. v Confidence level C of thermal imaging anomalies t Acoustic anomaly confidence level C a .

[0013] Preferred: Confidence level of thermal imaging anomalies T max To detect the highest temperature in the area, T t T represents the normal operating temperature threshold for the corresponding equipment. a S' represents the upper limit of the equipment's alarm temperature, and S' is the percentage of the over-temperature area to the target equipment's detection area. The calculation results are forcibly truncated to the 0~1 range to avoid numerical overflow.

[0014] Preferred: Acoustic anomaly confidence level , where D c D is the cosine distance between the current voiceprint and the reference voiceprint. t D is the normal fluctuation threshold. a This is the threshold for abnormal noise alarms; the calculation result is forcibly truncated to the 0~1 range.

[0015] Preferred: Anomaly type-driven dynamic weighted fusion: The cloud first identifies the anomaly type, matches the corresponding modal weights, and then calculates the fusion confidence C. f The formula is: Where i is the mode number, I is the total number of modes, i = 1, 2, ..., I; C i w represents the single-mode confidence score of the i-th mode. i These are the weighting coefficients for the corresponding modes.

[0016] Preferred method: Sliding window timing correction with time decay: A sliding window is constructed using the fusion confidence scores of 5 consecutive frames. The final confidence score is calculated by weighting the results with time decay, filtering out single-frame misjudgments caused by instantaneous light shadows and noise, and achieving the final confidence score. Where α is the weight coefficient of the current frame; L is the sliding window length; t is the time sequence number of the current frame; k is the sequence number of the historical frame within the sliding window; β is the historical frame attenuation coefficient; C f (t) is the multimodal fusion confidence score of the t-th frame; C f (k) Multimodal fusion confidence of the k-th frame; set hard verification rules: output the valid final confidence only when the single-frame fusion confidence of at least 3 frames within the sliding window is higher than the initial screening threshold; otherwise, it is judged as occasional interference and the final confidence is directly set to 0.

[0017] Preferred: Multi-factor hazard score calculation and alarm classification: Hazard score RS=min(100,C×W) d ×K×100), where C: final confidence level; W d : Importance weight of the inspected equipment; K is the historical anomaly coefficient; four levels of alarms are divided according to the risk score range.

[0018] Preferred: Historical anomaly coefficient K=K d ×K h K d : Duration coefficient of abnormality; K h Historical anomaly frequency coefficient.

[0019] Preferably, the historical anomaly coefficient can also be: Where m is the anomaly occurrence number within the preset time period, and M is the total number of anomalies occurring within the preset time period; m = 1, 2, ..., M; l m The duration of the anomaly with ID m; t m t0 is the duration of the anomaly numbered m from the current time point; t0 is the duration of the basic unit.

[0020] Preferred Alarm System: Alarms are categorized into four levels based on hazard score ranges: Critical Alarm: Hazard score ≥ 80, corresponding to high-certainty anomalies in core equipment, triggering the highest priority emergency response; Serious Alarm: 60 ≤ Hazard score < 80, corresponding to clear anomalies in important equipment, triggering high-priority review; General Alarm: 30 ≤ Hazard score < 60, corresponding to weak anomalies in ordinary equipment, included in the routine review list; Anomaly Warning: 10 ≤ Hazard score < 30, corresponding to low-confidence, occasional anomalies, only recorded and archived. Safety fallback rule: If thermal imaging detects a temperature exceeding the equipment's ultimate safety threshold, regardless of the hazard score, it is directly escalated to a critical alarm; detections with a final confidence level below 0.3 are considered invalid, no alarm record is generated, and they are only retained as negative samples.

[0021] Preferred: Minute-level real-time emergency rescheduling includes: executing differentiated rescheduling strategies based on alarm levels, and all rescheduled routes and tasks must pass triple security checks before being sent out for execution: power constraint check, movable area constraint check, and task mutual exclusion check.

[0022] Preferred: Power constraint verification includes: the estimated total energy consumption of all waypoints after rearrangement + return redundancy energy consumption ≤ current remaining power × 80% safety factor. If not met, waypoints are cut off from low to high priority until the constraint is met; Movable area constraint verification: all coordinate points of the replanned path fall within the preset movable area and there are no road sections crossing restricted areas.

[0023] Preferred: Task mutual exclusion verification includes: only one navigation / detection task is being executed on the same robot at any given time, with no instruction conflicts.

[0024] Preferred Logic: Prioritized Reordering: Critical Alarm: Immediately issue a pause command to the robot, interrupting the current inspection task; use the A* algorithm to plan the shortest path from the current location to the anomaly point, and dispatch the robot to the anomaly point first; upon arrival, activate the gimbal preset position system, calling more than 3 different view preset positions to collect data from multiple angles for verification; automatically add temporary detection items for related devices within 5 meters of the anomaly point to expand the investigation scope; after verification, decide whether to continue high-priority waypoint inspections or trigger autonomous return based on remaining battery power. Serious Alarm: After completing all detection actions for the current waypoint, pause the original route execution; recalculate the priority of all remaining waypoints based on the comprehensive priority calculation formula, reorder the route from high to low priority, and prioritize verification of the alarm point; cut the lowest priority 10% of non-core waypoints to ensure the battery budget for key tasks. General / Abnormal Alarm: The current inspection task will not be interrupted; if the subsequent route passes by the abnormal point, an additional inspection will be carried out without changing the main route; if it does not pass by, it will be included in the priority inspection list for the next round of inspections.

[0025] The preferred formula for calculating the overall priority of remaining waypoints is: Where h is the remaining waypoint number, P h The overall priority of the h-th waypoint; W d K represents the importance weight of the corresponding equipment. a E represents the alarm linkage coefficient. h K represents the estimated energy consumption for reaching waypoint h from the current location; e Energy consumption weight.

[0026] Preferred Optimization: Daily Mid-Term Inspection Strategy Optimization: Based on manual review results of potential hazards and alarm statistics, the inspection configuration is adjusted on a daily basis: Inspection Frequency Adjustment: Equipment with more than two confirmed critical / emergency alarms within 30 days automatically has its inspection frequency increased; low-importance equipment with no abnormalities for 90 consecutive days has its inspection frequency appropriately reduced to decrease the resource consumption of ineffective inspections. Detection Item Configuration Optimization: Detection items with frequent false alarms have their confidence thresholds automatically increased to reduce invalid alarms; equipment with frequent missed detections automatically has corresponding detection algorithms added; points with pointer instrument reading errors exceeding the standard have their number of detection frames automatically increased from the default 5 frames to 10 frames (averaged). Inspection Route Optimization: Based on historical alarm distribution statistics of high-risk equipment cluster areas, the inspection route is adjusted to reduce ineffective turnaround sections, lower the total energy consumption of a single inspection trip, and improve the coverage of points inspected per trip.

[0027] Preferred Features: Weekly / Monthly Long-Term Self-Evolutionary Iteration: Constructing an automated machine learning operation and maintenance closed loop for inspection business and AI training; Automatic Data Collection: Valid anomaly samples marked as "confirmed" by manual review are automatically categorized into the corresponding positive sample dataset; samples marked as "false alarms" are automatically categorized into the negative sample dataset; meter-related samples are automatically associated with manually verified actual readings, updating the meter calibration sample library. Incremental Training Trigger: When the number of new samples of a certain type of anomaly reaches a preset threshold (default 200 valid samples), the system automatically triggers an incremental training task for the model, fine-tuning it based on the original production model; after training, it automatically performs validation set evaluation and outputs an average accuracy report. Gray-Scale Deployment Mechanism: For new versions with evaluation accuracy superior to the current production model, a 7-day gray-scale trial run is first conducted at one site; if the false alarm rate is lower than the original model and there are no missed detections during the trial run, the model is fully deployed to replace the production model; if the effect is unsatisfactory, it automatically rolls back to the original stable version and continues to accumulate samples.

[0028] The technical effects and advantages of this invention are as follows: 1. Significantly improved reliability and execution efficiency of inspection tasks. Through multi-factor waypoint priority scheduling and hierarchical emergency rescheduling mechanism, the inspection strategy can be dynamically adjusted according to the robot's remaining power, equipment importance, and real-time alarms. The inspection coverage of core production equipment is greatly improved, the response time for high-risk anomaly verification is shortened, and the task completion rate under extreme conditions such as low power and minor faults is greatly improved. It effectively avoids the risk of missing key points on static routes and adapts to the dynamic inspection needs of complex sites.

[0029] 2. Significantly improved accuracy and alarm rationality in multimodal anomaly identification. Through dynamic weight fusion adapted to anomaly type, time-decaying sliding window timing correction, and a multi-factor hazard score grading model, single-frame misjudgments caused by light and noise are filtered out, resulting in a significant reduction in the overall false alarm rate in industrial scenarios. Alarm grading is deeply linked to equipment importance and anomaly duration, and the grading results match the actual risk level on site, greatly reducing the maintenance burden caused by invalid alarms.

[0030] 3. The system achieves end-to-end self-evolution and iteration. It constructs an MLOps closed loop of "inspection and detection - hidden danger review - data feedback - model training - gray-scale deployment". The review data generated by daily inspections is automatically converted into training samples without manual collection and organization. The accuracy of the detection model continues to improve with business operation. The mid-term inspection strategy is dynamically optimized according to the alarm distribution, realizing the bidirectional evolution of scheduling strategy and detection capability.

[0031] 4. The utilization rate of edge and cloud resources is comprehensively improved. The dynamic computing power scheduling mechanism can adaptively allocate inference tasks according to network status and edge load. While ensuring the real-time performance and accuracy of detection, it reduces the bandwidth occupation of inspection data, fully releases edge computing power, and can still ensure the continuous operation of inspection work in network fluctuation scenarios, adapting to the complex network and computing power environment of industrial sites. Attached Figure Description

[0032] Figure 1 This is a flowchart illustrating a method for dynamic optimization and real-time anomaly identification of intelligent robot inspection tasks proposed in this invention.

[0033] Figure 2 This is a flowchart illustrating step 2 of the method for dynamic optimization and real-time anomaly identification of intelligent robot inspection tasks proposed in this invention.

[0034] Figure 3 This is a flowchart illustrating step 3 of the method for dynamic optimization and real-time anomaly identification of intelligent robot inspection tasks proposed in this invention. Detailed Implementation

[0035] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the invention, and should not be construed as limiting the invention. Rather, embodiments of the invention include all variations, modifications, and equivalents falling within the spirit and scope of the appended claims.

[0036] Example 1 refer to Figure 1 This embodiment proposes a method for dynamic optimization and real-time anomaly identification of intelligent robot inspection tasks. Based on a three-layer architecture of terminal inspection robot - site edge server - cloud scheduling center, it forms an end-to-end intelligent inspection process system through baseline task construction, multi-source state perception, hierarchical multimodal anomaly identification, three-level closed-loop task re-arrangement, and edge-cloud dynamic computing power scheduling. The specific technical solution is as follows: S1. Baseline Inspection Task System Construction: Initial modeling of inspection tasks is completed in the cloud, serving as the foundation for subsequent dynamic optimization. A 2D site occupancy grid map is imported, and movable and restricted areas are marked. Inspection routes containing waypoints and navigation parameters are marked. A device detection item library and device importance weight system are constructed. Task scheduling rules and safety constraint thresholds are configured, generating executable baseline inspection tasks. Map and Route Configuration: A 2D site occupancy grid map is imported, and movable and restricted area constraints are marked. The set of waypoints and directed connecting edges for the inspection route are drawn on the map. Navigation parameters are configured for each connecting edge, including gait type, travel speed, travel direction, and obstacle stopping rules. Waypoints are the point nodes of the inspection route, divided into pure navigation points and detection points. Each waypoint stores the following attributes: world plane coordinates, robot target orientation angle, corresponding device tag number, point dwell time, set of fixed-point detection items, and gimbal preset position parameters. The detection points can be bound to multimodal detection items such as protective equipment wearing detection, meter reading, and thermal imaging temperature measurement, while also being associated with corresponding pan-tilt-zoom (PTZ) preset positions to ensure consistency between the shooting angle and calibration time during inspections. Directed connection edges represent traversable paths between two waypoints. The detection item library and weighting system are constructed as follows: a device detection item library is built, binding each detection object with algorithm type, camera slot position, initial confidence threshold, and meter calibration parameters to form a standardized detection template; simultaneously, importance weights are configured for all equipment, with core production equipment weights set to 2.0, auxiliary production equipment weights to 1.2, and ordinary auxiliary facilities weights to 0.8. Scheduling and safety rule configuration: task scheduling rules are configured, supporting single execution, scheduled execution, and interval execution modes; safety constraint thresholds such as low battery return threshold, task retries, and battery safety redundancy coefficient are set to generate executable baseline inspection tasks. For example, a 2D grid map of the substation is uploaded to the cloud, marking all movable areas and restricted areas for high-voltage equipment; inspection routes are edited on the 2D map workbench, setting 28 waypoints to cover all core equipment; default navigation parameters are configured for each route segment: agile gait on flat ground, low-speed movement, and obstacle avoidance function enabled. An equipment inspection item library is built: main transformers are bound to thermal imaging temperature detection and acoustic abnormal noise detection; high-voltage switchgear is bound to safety helmet detection and oil leakage detection; pointer pressure gauges are bound to meter reading detection, with pre-entered segmented angle-pressure mapping calibration parameters for 5 scale points. Scheduling rules are configured: automatic execution at 8:00 AM daily on weekdays, one task retry, a low battery return threshold of 30%, and the generation of baseline routine inspection tasks.

[0037] S2. Multi-source real-time status acquisition and synchronization: Edge servers and the cloud collaborate to achieve full-dimensional status perception, providing a data foundation for dynamic scheduling and anomaly detection; the edge server periodically collects robot operating status, multimodal sensor data, and inspection video streams, synchronizing inspection progress to the cloud; the cloud aggregates alarm data from the entire site, equipment maintenance status, and historical fault records, dynamically updating the overall operational status. (Reference) Figure 2 Robot Status Acquisition: The edge server collects robot heartbeat data every 10 seconds, including battery level, gait, body temperature, fault status, real-time location, and motion parameters, and synchronizes the inspection progress to the cloud. Multimodal Sensor Data Acquisition: The edge server synchronously accesses real-time data from the robot's visible light camera, thermal imaging camera, and acoustic sensors as input sources for anomaly detection. Global Situational Synchronization: The cloud aggregates alarm data from the entire site, equipment maintenance status, and historical fault records, dynamically updating equipment importance coefficients and the overall operational status. For example, after the task starts on time, the robot performs inspections along the baseline route. The edge server collects robot heartbeat data (battery level, location, gait, body temperature, etc.) every 10 seconds and synchronizes the inspection progress to the cloud; the visible light camera extracts frames at 1 frame / second, and the thermal imaging and acoustic sensors continuously collect data, all of which are connected to the edge initial inspection module.

[0038] S3, a two-level multimodal anomaly identification and graded alarm system, adopts a layered inference architecture of lightweight initial screening at the edge server and high-precision review in the cloud. It achieves accurate alarm grading through multimodal dynamic fusion, time-series correction, and multi-factor hazard score calculation. (Reference) Figure 3 The specific steps of S3 are as follows: 1. Lightweight initial screening at the edge: A lightweight detection model is deployed on the edge server to perform initial screening of multimodal data by sampling frames at a fixed frame rate, and outputs the single-modal confidence score of each modality; when the confidence score of any single modality is higher than the initial screening threshold (default 0.3), it is judged as a suspected anomaly, and the images of the 5 frames before and after the corresponding time period and 10 seconds of sensor data are captured and uploaded to the cloud; normal detection results only upload the readings and status, saving uplink bandwidth. 2. Single-modal basic confidence quantification: The cloud first normalizes the detection results of the three core modalities to obtain the single-modal confidence score in the range of 0~1: Visible light visual confidence score C v The high-precision, lightweight, single-stage target detection algorithm, optimized for industrial scenarios and optimized in the cloud, directly outputs the matching probability of abnormal features, with a value ranging from 0 to 1. Specific details of this existing technology will not be elaborated here. Thermal imaging anomaly confidence C. t The calculation is based on the highest temperature in the detection area, the temperature threshold, and the percentage of the area exceeding the temperature limit. The formula is as follows: T max To detect the highest temperature in the area, T t T represents the normal operating temperature threshold for the corresponding equipment. a S' represents the upper limit of the equipment's alarm temperature, and S' is the percentage of the target equipment's detection area relative to the over-temperature zone. The calculation result is forcibly truncated to the 0-1 range to avoid numerical overflow. For example, when inspecting the No. 1 main transformer route, the thermal imaging detected a maximum temperature of 85℃ in the area. The normal temperature threshold for this equipment is 60℃, the alarm temperature limit is 100℃, and the over-temperature zone accounts for 15%. Substituting these values ​​into the thermal imaging confidence formula, we can calculate: Ct =min{1,max[0,(85-60) / (100-60)]×0.15}=0.625×0.15=0.09375. Acoustic anomaly confidence level C a The calculation is based on the spectral difference between the real-time voiceprint and the device's reference voiceprint, using the following formula: , where D c D represents the cosine distance (difference) between the current voiceprint and the reference voiceprint. t D is the normal fluctuation threshold. a This is the threshold for abnormal noise alarms; the calculation result is forcibly truncated to the 0-1 range. For example, if the cosine distance between the acoustically detected voiceprint and the reference voiceprint is 0.5, the normal fluctuation threshold is 0.3, and the abnormal noise alarm threshold is 0.8, then substituting these values ​​into the acoustic confidence formula will calculate: C a =min{1,max[0,(0.5-0.3) / (0.8-0.3)]}=0.4, the acoustic confidence score of 0.4 is higher than the initial screening threshold of 0.3, and is judged as a suspected anomaly. The edge captures 5 frames of visible light and thermal imaging data and 10 seconds of acoustic data and uploads them to the cloud. 3. Anomaly type-driven dynamic weighted fusion: The cloud first identifies the anomaly type, matches the corresponding modal weights and then calculates the fusion confidence score C. f The formula is: Where i is the mode number, I is the total number of modes, i = 1, 2, ..., I; C i w represents the single-mode confidence score of the i-th mode. i These are the weighting coefficients for the corresponding modes. Weights are dynamically adapted according to the anomaly type: For appearance-related anomalies (safety helmet not worn, oil leaks, meter appearance anomalies): visible light is the core mode, w1=0.7, w2=0.2, w3=0.1; for overheating-related anomalies (equipment temperature rise, connector overheating): thermal imaging is the core mode, w1=0.25, w2=0.6, w3=0.15; for mechanical fault-related anomalies (bearing noise, discharge sound): acoustics is the core mode, w1=0.15, w2=0.25, w3=0.6. If a sensor for a certain mode malfunctions or the data is invalid, the corresponding weight is reset to 0, and the weights of the other modes are proportionally amplified to ensure that the fusion result is always normalized to the 0~1 range. For example, if the cloud identifies an overheating anomaly, the matching weights are: thermal imaging 0.6, acoustics 0.15, visible light 0.25. The visible light detection in the current frame shows no abnormalities in the device's appearance, with C1=0.625, C1=0.4, and C3=0.2. Substituting these values ​​into the fusion formula, we calculate: C=(0.6×0.625+0.15×0.4+0.25×0.2) / (0.6+0.15+0.25)==0.485. 4. Timing correction with a sliding window and time decay: A sliding window is constructed using the fusion confidence scores of 5 consecutive frames. The final confidence score is calculated using time decay weighting to filter out misjudgments in a single frame caused by instantaneous light shadows and noise. The final confidence score C is calculated using the following formula: , where α is the weight coefficient of the current frame, ranging from 0 to 1, with a default value of 0.4; L is the sliding window length, with a default value of 5. t is the time sequence number of the current frame, representing the latest frame for which the final confidence score is being calculated. It is the end point of the entire sliding window and the frame closest to the current state. For example, when processing the detection result of frame 100, t=100. k is the sequence number of the historical frames within the sliding window, a traversal variable in the summation formula, representing all historical frames within the sliding window except the current frame, with a value range of [t-L+1, t-1]. For example, when the window length L=5 and the current t=100, k sequentially takes four historical frames: 96, 97, 98, and 99. β is the historical frame attenuation coefficient, ranging from 0 to 1, with a default value of 0.8. Frames older have lower weights; this implements the design logic that older frames have less impact on the final result. C f (t) is the multimodal fusion confidence score of frame t, which is the fusion confidence score of the latest frame obtained by dynamic weighted fusion calculation based on anomaly type. Its value ranges from 0 to 1, representing the anomaly probability of the current single frame. C f (k) The multimodal fusion confidence of the k-th frame, representing the fusion confidence of each historical frame within the window. Each k corresponds to the single-frame fusion result of a historical frame, with values ​​ranging from 0 to 1. A hard verification rule is set: only when the single-frame fusion confidence of at least 3 frames within the sliding window is higher than the initial screening threshold is a valid final confidence output; otherwise, it is judged as occasional interference, and the final confidence is directly set to 0. This method differs from simple multi-frame averaging by giving higher weight to the latest frame and decreasing weight to historical frames. It filters out occasional misjudgments caused by single-frame lighting and noise through multi-frame processing, while ensuring that the final confidence reflects the real-time abnormal state of the site in a timely manner. It does not allow the lag of historical frames to affect the abnormal response speed, thus adapting to the real-time requirements of industrial inspection. For example, the sliding window has 5 frames, the current frame is the 5th frame, and the fusion confidence of the previous 4 frames are 0.32, 0.41, 0.45, and 0.47, respectively. The current frame weight α = 0.4, and the attenuation coefficient β = 0.8. The weighted average of historical frames is calculated as: Sum of historical frame weights: 0.8 + 0.8 2 +0.8 3 +0.8 4 =0.8 + 0.64 + 0.512 + 0.4096 = 2.3616; Weighted sum of historical frames: 0.8 4 ×0.32+0.8 3 ×0.41+0.8 2×0.45+0.8×0.47≈0.4096×0.32+0.512×0.41+0.64×0.45+0.8×0.47≈1.015; Weighted average of historical frames: 1.015÷2.3616≈0.43; Final confidence level: C=0.4×0.485+0.6×0.43≈0.452; Verification rule: If 4 out of 5 frames have a fused confidence level higher than the initial screening threshold of 0.3, they are judged as valid anomalies. 5. Multi-factor hazard score calculation and alarm classification: Based on the final confidence level, combined with the equipment importance, anomaly duration, and historical anomaly frequency, the hazard score RS in the range of 0~100 is calculated. The formula is: RS=min(100,C×W d ×K×100), where the parameters are defined and their values ​​are determined as follows: C: final confidence level, ranging from 0 to 1, representing the probability that the anomaly actually exists; W d The importance weight of the inspected equipment ranges from 0.8 to 2.0, with core production equipment having the highest weight; K is the historical anomaly coefficient; it can also be calculated, K = K d ×K h ;K d : Abnormal duration coefficient, 1.0 for duration < 10 seconds, 1.2 for duration ≤ 10 seconds < 60 seconds, and 1.5 for duration ≥ 60 seconds; K h Historical anomaly frequency coefficient: 1.0 for the first occurrence within 30 days, 1.2 for 2-3 occurrences within 30 days, and 1.5 for ≥4 occurrences within 30 days. For example, equipment importance W... d =2.0, the anomaly lasting 40 seconds corresponds to K d =1.2, the first occurrence of K within 30 days h =1.0. Substituting into the hazard score formula: RS=min(100,0.452×2.0×1.2×1.0×100)=min(100,108.48)=100; triggering the temperature anomaly fallback rule, it is judged as a critical alarm, and a hazard record is generated and pushed to the manual review process. The historical anomaly coefficient can also be... Where m is the anomaly number within the preset time period, and M is the total number of anomalies within the preset time period; m = 1, 2, ..., M; the preset time period can be 30 days, but other value settings are not excluded, which will not be elaborated here. m The duration of the anomaly with the number m can be in seconds, with durations less than 1 second counted as 1 second; the duration evaluation unit for l0 can be 1 second or t. mThis refers to the duration of the anomaly numbered m from the current time point, in days. t0 is the base unit duration, which can be 10 days, but other numerical settings are not excluded. This method allows for the consideration of the frequency and duration of anomalies, avoiding data distortion and masking. It also ensures that the current state is dominant by decreasing the weight of historical frames, and that historical state evaluation guarantees objective and consistent assessments, leading to more accurate calculations. 6. Four alarm levels are defined based on hazard score ranges: Critical Alarm: Hazard score ≥ 80, corresponding to high-deterministic anomalies in core equipment, triggering the highest priority emergency response; Serious Alarm: 60 ≤ Hazard score < 80, corresponding to clear anomalies in important equipment, triggering high-priority review; General Alarm: 30 ≤ Hazard score < 60, corresponding to weak anomalies in ordinary equipment, included in the routine review list; Anomaly Notice: 10 ≤ Hazard score < 30, corresponding to low-confidence, occasional anomalies, only recorded and archived. Set a safety fallback rule: If the thermal imaging detects that the temperature exceeds the device's limit safety threshold, it will be directly upgraded to an emergency alarm regardless of the danger score; if the final confidence score is lower than 0.3, it will be judged as an invalid detection, no alarm record will be generated, and it will only be kept as a negative sample.

[0039] S4. Based on the alarm level, trigger corresponding level task adjustments, execute real-time emergency rescheduling at the minute level, complete mid-term inspection strategy optimization at the daily level, and achieve long-term self-evolution iteration of the detection model at the weekly / monthly level; construct a three-level closed-loop system of minute-level real-time emergency rescheduling - daily mid-term strategy optimization - weekly / monthly long-term capability iteration, realizing full-link linkage from immediate response to system self-evolution. 1. Minute-level real-time emergency rescheduling: Execute differentiated rescheduling strategies based on alarm levels. All rescheduled routes and tasks must pass three safety checks before execution: Power constraint check: The estimated total energy consumption of all waypoints after rescheduling + return redundancy energy consumption ≤ current remaining power × 80% safety factor. If not met, waypoints are cut from low to high priority until the constraint is met; Movable area constraint check: All coordinates of the replanned path fall within the preset movable area, without crossing prohibited sections; Task mutual exclusion check: Only one navigation / detection task exists on the same robot at any given time, with no instruction conflicts. Priority Reordering Logic: Critical Alarm: Immediately issue a pause command to the robot, interrupting the current inspection task; use the A* algorithm to plan the shortest path from the current location to the anomaly point, and dispatch the robot to the anomaly point first; upon arrival, activate the gimbal preset position system, calling more than 3 different view preset positions to collect data from multiple angles for verification; automatically add temporary detection items for related devices within 5 meters of the anomaly point to expand the investigation scope; after verification, decide whether to continue high-priority waypoint inspections or trigger autonomous return based on remaining battery power. Serious Alarm: After completing all detection actions for the current waypoint, pause the original route execution; recalculate the priority of all remaining waypoints based on the comprehensive priority calculation formula, reorder the route from high to low priority, and prioritize verification of the alarm point; cut the lowest priority 10% of non-core waypoints to ensure the battery budget for key tasks. General / Anomaly Alarm: Do not interrupt the current inspection task; if the subsequent route passes by the anomaly point, add supplementary inspections without changing the main route; if not, include it in the priority detection list for the next round of inspections. The formula for calculating the overall priority of remaining waypoints is: Where h is the remaining waypoint number, P h The overall priority of the h-th waypoint is determined by the numerical value; higher values ​​indicate higher priority. d K represents the importance weight of the corresponding equipment. a This is the alarm linkage coefficient. It takes a value of 1 when there is no alarm, and increases to 1.25~2.0 depending on the level of the associated alarm (increasing by 0.25 for each level increase); E h K represents the estimated energy consumption for reaching waypoint h from the current location; e1. Energy Consumption Weighting: The lower the robot's remaining battery power, the higher the energy consumption weight. 2. Daily Mid-Term Inspection Strategy Optimization: Based on manual review results of potential hazards and alarm statistics, the inspection configuration is adjusted daily: Inspection Frequency Adjustment: Equipment with more than two confirmed critical / emergency alarms within 30 days automatically has its inspection frequency increased; low-importance equipment with no abnormalities for 90 consecutive days has its inspection frequency appropriately reduced to decrease resource consumption from ineffective inspections. Detection Item Configuration Optimization: Detection items with frequent false alarms have their confidence thresholds automatically increased to reduce invalid alarms; equipment with frequent missed detections automatically has corresponding detection algorithms added; points with pointer instrument reading errors exceeding the standard automatically have their detection frame count increased from the default 5 frames to 10 frames (average). Inspection Route Optimization: Based on historical alarm distribution statistics of high-risk equipment cluster areas, the inspection route is adjusted to reduce ineffective turnaround sections, lower the total energy consumption of a single inspection trip, and improve the coverage of points inspected per trip. 3. Weekly / Monthly Long-Term Self-Evolutionary Iteration: Building an automated machine learning operation and maintenance closed loop for inspection business and AI training: Automatic data collection: Valid anomaly samples marked as "confirmed" by manual review are automatically categorized into the corresponding positive sample dataset; samples marked as "false alarms" are automatically categorized into the negative sample dataset; meter-related samples are automatically associated with manually verified actual readings, updating the meter calibration sample library. Incremental training trigger: When the number of new samples of a certain type of anomaly reaches a preset threshold (default 200 valid samples), the system automatically triggers an incremental training task for the model, fine-tuning it based on the original production model; after training, it automatically performs validation set evaluation and outputs an average accuracy report. Gray-scale deployment mechanism: For new versions with evaluation accuracy superior to the current production model, a 7-day gray-scale trial is first conducted at one site; if the false alarm rate is lower than the original model and there are no missed detections during the trial, the model is fully deployed to replace the production model; if the effect is not satisfactory, it automatically rolls back to the original stable version and continues to accumulate samples. For example, power verification: the robot currently has 62% remaining power. The total estimated energy consumption for heading to the anomaly point, verification, and return is 45% of the remaining power, which is below the 80% safety factor and meets the constraints. Movable area verification: the planned path falls entirely within the movable area, without crossing restricted areas. Task mutual exclusion verification: there are no other tasks running after terminating the current task, and there are no conflicts. After the verification is passed, the A* algorithm is used to plan the shortest path from the current location to the No. 1 main transformer, and the robot is scheduled to go to the anomaly point first. Upon arrival, it automatically activates three preset gimbal positions (front panoramic view, close-up of the temperature rise area, and view of adjacent equipment) to collect images and acoustic data from multiple angles for secondary verification. Simultaneously, temperature detection items for two surrounding switchgear cabinets are added to expand the investigation scope. After the verification is completed, the remaining power meets the inspection requirements of the core equipment, and the remaining inspections continue to be completed according to the recalculated waypoint priority. Mid-term strategy optimization: After the maintenance personnel review and confirm that the anomaly is a valid potential hazard, the system automatically adjusts the inspection frequency of the No. 1 main transformer from once a day to three times a day; at the same time, the weight of thermal imaging detection in this area is increased, and the number of detection frames is increased from 5 frames to 10 frames.Long-term model iteration: The verification images are automatically included in the positive sample dataset for thermal imaging overheat detection. When the cumulative number of new samples of the same type reaches 200, the system automatically triggers incremental training of the model. After training is completed, the validation set evaluation is automatically performed. After the average accuracy is better than the current production model, the site is selected for a 7-day gray-scale trial run. During the trial run, the false alarm rate is lower than the original model and there are no missed detections. Finally, the model is fully deployed to replace the production model.

[0040] S5. Based on network bandwidth, edge server load, and alarm level, dynamically adjust the allocation of inference tasks between the edge server and the cloud to balance detection accuracy, real-time performance, and bandwidth usage. The dynamic scheduling of edge-cloud inference computing power adjusts the allocation of inference tasks between the edge and cloud based on the on-site network status, edge server load, and alarm level, taking into account detection accuracy, real-time performance, and bandwidth efficiency. The scheduling rules are as follows: Normal operating conditions: When the edge server GPU load is below 70% and the network bandwidth is stable, routine detection is performed entirely on the edge server, with only alarm and normal reading data uploaded to reduce bandwidth usage. Constrained operating conditions: When the network bandwidth is below 2Mbps or the edge server GPU load is above 80%, the edge server is only responsible for frame extraction and data preprocessing, while high-precision inference tasks are uploaded to the cloud for execution, ensuring uninterrupted inspection operations. Alarm operating conditions: When a severe or higher-level alarm is triggered, the edge server suspends unnecessary routine inference tasks, prioritizing the uploading of alarm data. The cloud centralizes computing power to perform multimodal fusion verification, improving anomaly detection accuracy. For example, during routine inspections, the edge server's GPU load is 45%, the network bandwidth is stable, all routine tests are completed at the edge, only readings and alarm data are uploaded, and the bandwidth usage is kept below 1Mbps; when the on-site network fluctuates and the bandwidth drops to 1.5Mbps, the system automatically shuts down unnecessary cloud-based verifications and pushes all inferences down to the edge to ensure that the inspection operation is not interrupted; when a critical alarm is triggered, the edge suspends non-core routine detection inferences and prioritizes the uploading of alarm data, and the cloud centralizes computing power to complete multimodal verifications.

[0041] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0042] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for dynamic optimization and real-time anomaly identification of intelligent robot inspection tasks, characterized in that, The method for dynamic optimization and real-time anomaly identification of intelligent robot inspection tasks includes: S1. Import the 2D site grid map and mark the movable area and restricted area. Mark the inspection route including waypoints and navigation parameters. Build the equipment inspection item library and equipment importance weight system. Configure task scheduling rules and safety constraint thresholds. Generate executable baseline inspection tasks. S2: Periodically collect robot operating status, multimodal sensor data and inspection video stream, and synchronize inspection progress to the cloud; the cloud aggregates alarm data, equipment operation and maintenance status and historical fault records from the entire site, and dynamically updates the overall operating status; S3 adopts a layered inference architecture that uses lightweight initial screening at the edge server and high-precision review in the cloud. It achieves accurate alarm classification through multimodal dynamic fusion, time-series correction, and multi-factor hazard score calculation. S4. Trigger corresponding task adjustments based on alarm levels to build a three-level closed-loop system of minute-level real-time emergency rescheduling, day-level mid-term strategy optimization, and week / month-level long-term capability iteration, realizing full-link linkage from immediate response to system self-evolution. S5. Based on network bandwidth, edge server load, and alarm level, dynamically adjust the allocation of inference tasks between the edge server and the cloud to balance detection accuracy, real-time performance, and bandwidth usage.

2. The method for dynamic optimization and real-time anomaly identification of intelligent robot inspection tasks according to claim 1, characterized in that, Navigation parameters include: gait type, speed, direction of travel, and obstacle stopping rules.

3. The method for dynamic optimization and real-time anomaly identification of intelligent robot inspection tasks according to claim 1, characterized in that, The specific steps of S3 are as follows:

1. Lightweight initial screening at the edge; 2. Single-modal basic confidence quantification; 3. Dynamic weighted fusion driven by anomaly type; 4. Sliding window time series correction with time decay; 5. Multi-factor hazard score calculation and alarm classification.

4. The method for dynamic optimization and real-time anomaly identification of intelligent robot inspection tasks according to claim 3, characterized in that, The lightweight initial screening at the edge includes: deploying a lightweight detection model on the edge server, sampling frames of multimodal data at a fixed frame rate for initial screening, and outputting the single-modal confidence score for each modality; when the confidence score of any single modality is higher than the initial screening threshold, it is judged as a suspected anomaly, and a preset number of frames before and after the corresponding time period and sensor data at a preset time are extracted and uploaded to the cloud; normal detection results only upload the readings and status, saving uplink bandwidth.

5. The method for dynamic optimization and real-time anomaly identification of intelligent robot inspection tasks according to claim 3, characterized in that, The unimodal basic confidence quantification includes: firstly, the detection results of the three core modalities are normalized in the cloud to obtain the unimodal confidence in the range of 0 to 1; the unimodal confidence includes: visible light visual confidence C. v Confidence level C of thermal imaging anomalies t Acoustic anomaly confidence level C a .

6. The method for dynamic optimization and real-time anomaly identification of intelligent robot inspection tasks according to claim 1, characterized in that, Minute-level real-time emergency rescheduling includes: implementing differentiated rescheduling strategies based on alarm levels; and requiring all rescheduled routes and tasks to pass triple security checks before execution: power constraint check, movable area constraint check, and task mutual exclusion check.

7. The method for dynamic optimization and real-time anomaly identification of intelligent robot inspection tasks according to claim 6, characterized in that, The power constraint verification includes: the estimated total energy consumption of all waypoints after rearrangement + return redundancy energy consumption ≤ current remaining power × 80% safety factor. If it is not met, waypoints are cut off from low to high priority until the constraint is met; movable area constraint verification: all coordinate points of the replanned path fall within the preset movable area and there are no road sections crossing restricted areas.

8. The method for dynamic optimization and real-time anomaly identification of intelligent robot inspection tasks according to claim 6, characterized in that, Task mutual exclusion verification includes ensuring that only one navigation / detection task is being executed on the same robot at any given time, with no instruction conflicts.

9. The method for dynamic optimization and real-time anomaly identification of intelligent robot inspection tasks according to claim 1, characterized in that, Daily Mid-Term Inspection Strategy Optimization: Based on manual review results of potential hazards and alarm statistics, the inspection configuration is adjusted on a daily basis: Inspection Frequency Adjustment: Equipment with more than two confirmed critical / emergency alarms within 30 days will have its inspection frequency automatically increased; low-importance equipment with no abnormalities for 90 consecutive days will have its inspection frequency appropriately reduced to decrease the resource consumption of ineffective inspections; Detection Item Configuration Optimization: Detection items that frequently generate false alarms will have their confidence threshold automatically increased to reduce ineffective alarms; Equipment that frequently misses detections will have corresponding detection algorithms automatically added. For points where the reading error of the pointer instrument exceeds the standard, the number of detection frames will be automatically increased from the default 5 frames to 10 frames to take the average value. Inspection route optimization: Based on historical alarm distribution statistics, high-risk equipment cluster areas are identified, and the inspection route is adjusted to reduce invalid turnaround sections, thereby reducing the total energy consumption of a single inspection trip and improving the coverage of inspection points per trip.

10. The method for dynamic optimization and real-time anomaly identification of intelligent robot inspection tasks according to claim 1, characterized in that, Weekly / monthly long-term self-evolutionary iteration: Building an automated machine learning operation and maintenance closed loop for inspection business and AI training: Automatic data collection: Valid abnormal samples marked as "confirmed" by manual review are automatically classified into the corresponding positive sample dataset; samples marked as "false alarm" are automatically classified into the negative sample dataset; meter samples are automatically associated with the actual reading values ​​verified by manual review and the meter calibration sample library is updated. Incremental training trigger: When the number of new samples of a certain type of anomaly reaches a preset threshold, the system automatically triggers the incremental training task of the model to fine-tune it based on the original production model. After training is completed, the validation set evaluation is automatically performed, and the average accuracy report is output. Gray-scale deployment mechanism: If the evaluation accuracy of the new version is better than the current production model, a 7-day gray-scale trial run is first selected on one site. If the false alarm rate is lower than the original model and there are no missed detections during the trial run, the full version is deployed to replace the production model. If the effect does not meet the standard, it is automatically rolled back to the original stable version and continues to accumulate samples.

Images