A method and system for collaborative inspection of positive and negative samples based on temporal state evolution analysis

Abstract

This invention belongs to the field of intelligent substation inspection technology, specifically involving a positive and negative sample collaborative inspection method and system based on time-series state evolution analysis. The aim is to solve the problems of high false alarm rates for state transitions, high missed detection rates for gradual faults, and poor environmental adaptability caused by the independent operation of positive and negative sample algorithms and the lack of time-series state evolution analysis in existing intelligent substation inspection systems. The main solution includes: first, the system acquires and preprocesses equipment image streams and historical data to construct a standardized time-series set containing timestamps, state labels, and change vectors; then, it aggregates recent data for each equipment state to form a time-series state chain and embeds operational logic to identify illegal transitions; next, it uses a two-stage collaborative mechanism—the initial screening stage suppresses low-confidence false alarms, and the trend stage fits the gradual slope to trigger early warnings; finally, it dynamically weights and fuses positive and negative sample outputs based on environmental parameters to generate a reliable score and make decisions. This method is suitable for intelligent operation and maintenance scenarios under complex operating conditions.

Description

Technical Field

[0001] This invention belongs to the field of intelligent substation inspection technology, specifically involving a positive and negative sample collaborative inspection method and system based on time-series state evolution analysis. Background Technology

[0002] The core technologies of current intelligent substation inspection systems rely on two types of algorithms:

[0003] 1. Positive Sample Algorithm (Unsupervised Learning): This algorithm constructs a database of images showing the normal state of equipment (such as the standard position of the main transformer oil level and the standard angle of the switch opening) to detect anomalies that deviate from the normal state in real time. Its advantage lies in an anomaly recall rate approaching 100%, but it is susceptible to interference from lighting and fog, leading to a high false alarm rate (>10%).

[0004] 2. Negative Sample Algorithm (Supervised Learning): This algorithm trains a target detection model (such as YOLO or Mask R-CNN) based on a defect sample library (e.g., images of oil leaks or damage), and outputs equipment status classification results (e.g., "switch closed" or "oil temperature too high"). Its advantage lies in high status recognition accuracy (>96%), but due to the scarcity of defect samples and incomplete coverage of operating conditions, it may miss slowly changing faults (e.g., a slow daily drop in oil level).

[0005] The existing technical solutions that are closest to this proposal fall into the following three categories:

[0006] Solution 1: Positive and negative sample cascade fusion (CN117294015A, etc.). This solution first uses a positive sample algorithm to initially screen abnormal images, then inputs the abnormal images into a negative sample algorithm for defect classification, and finally outputs the classification results and generates an alarm. This solution only achieves the cascade of decision results and does not utilize the temporal correlation of state data. For example, when negative samples identify switch state transitions ("open → closed"), misjudgment may occur due to image occlusion. However, the system cannot backtrack to verify historical states, or when positive samples detect continuous small changes in meters, negative samples may miss the detection due to a lack of trend analysis capabilities.

[0007] Option 2: Collaborative approach based on static feature fusion. The process involves extracting pixel difference features (e.g., heatmaps) from positive samples and semantic features (e.g., defect category probabilities) from negative samples, then weighting and fusing them using a feature attention mechanism to output a comprehensive anomaly score. This method only fuses image features at a single moment, neglecting the continuity of state evolution. For example, it fails to consider whether the switch's open / closed state transitions are reasonable (e.g., whether it's possible for the switch to change from "open" to "closed" within 5 seconds); and whether meter readings conform to equipment operating patterns (e.g., whether oil temperature slowly decreases after a sudden rise).

[0008] Option 3: Multimodal signal fusion (CN120522528A). This option integrates acoustic, vibration, and electromagnetic signals to construct a cross-modal correlation model for identifying complex faults. However, this option focuses on physical signal fusion and does not address the timing logic verification of state-related data (switch positions, meter readings), and it relies on the deployment of high-cost sensing equipment.

[0009] Regarding the closest solution mentioned above, its technical drawbacks and the corresponding solutions in this proposal are as follows:

[0010] The drawbacks of Scheme 1 are: a high false alarm rate for state transitions, the negative sample algorithm misjudging switch states due to transient interference (such as light and shadow occlusion), and the lack of a time-series backtracking mechanism, leading to an increased false alarm rate (measured >10%); and a high missed detection rate for gradual faults, as the negative sample algorithm only outputs a single-point state and cannot correlate with historical data to identify slowly developing defects (such as a 0.5% daily average oil level decrease due to seal aging). This proposal can reduce the false alarm rate for state transitions and capture early characteristics of gradual faults.

[0011] The drawbacks of Scheme 2 are poor environmental adaptability, the static feature fusion does not consider the rationality of state transitions (such as the need to meet electrical operation logic for switch opening / closing), and the false alarm rate is still higher than 8% under complex operating conditions; it also wastes computing resources, requiring real-time calculation of high-dimensional image features, with edge device calculation latency reaching 500ms, making it difficult to support high-frequency state monitoring. This proposal verifies the rationality of states through a lightweight time-series state chain, reducing invalid calculations and reducing edge latency to <200ms.

[0012] The disadvantages of Option 3 are high deployment costs, the need for additional acoustic and vibration sensors, and its unsuitability for status-based scenarios, as it cannot handle the timing logic verification of discrete states such as switch opening / closing and meter readings. This proposal simply reuses the image and status data of existing inspection equipment, without requiring additional hardware, thus achieving low-cost timing status coordination. Summary of the Invention

[0013] The purpose of this invention is to solve the problems of high false alarm rate for state transitions, high missed detection rate for slow-change faults, and poor environmental adaptability caused by the independent operation of positive and negative sample algorithms and the lack of time-series state evolution analysis in existing intelligent substation inspection systems.

[0014] To achieve the above objectives, the present invention employs the following technical means:

[0015] This invention provides a method for collaborative inspection of positive and negative samples based on temporal state evolution analysis, comprising the following steps:

[0016] Step S1, Data Acquisition and Preprocessing: Acquire the device's real-time image stream and historical status database, preprocess the data to obtain data including timestamps. Device image State labels output by the negative sample algorithm and the state change vector output by the positive sample algorithm Standardized time series datasets ;

[0017] Step S2, Timing State Chain Construction: For each device state Aggregate the latest Generate a time-series state chain from the state data at each time point. The time-series state chain Includes a sequence of status labels and state change vector And inject device operation logic rules to mark illegal state transitions;

[0018] Step S3: Execution of the two-stage collaborative mechanism:

[0019] Step S31: Perform initial screening coordination: When the negative sample algorithm outputs a state jump and the confidence level is lower than the first threshold, the positive sample algorithm is triggered to calculate the displacement. If the displacement amount If the displacement is less than the threshold, the alarm is suppressed;

[0020] Step S32, Execute Trend Coordination: When the positive sample algorithm is continuous When a change in the same direction is detected for the first time, the time-series state chain is extracted. The historical state numerical sequence in the data, fitted with the slope of change. If the slope of the change If the slope threshold is exceeded, a slow-change fault warning will be generated;

[0021] Step S4: Dynamic Confidence Weighted Output: Weights are dynamically calculated based on environmental parameters, and the outputs of positive and negative samples are fused to obtain the final state confidence score. And based on the final state confidence score The value is used to output the decision result.

[0022] In the above scheme, the data preprocessing in step S1 includes:

[0023] The Z-score algorithm is used to detect and remove outliers, and linear interpolation is used to fill in the missing data.

[0024] Image data based on timestamps With state data Synchronization is performed, and a sliding window is used for smoothing.

[0025] The state values ​​are normalized using Min-Max.

[0026] In the above scheme, the time-series state chain construction step in step S2 includes:

[0027] For each device status Extract the state labels of the k most recent time moments from the database in chronological order. and state change vector Construct a time-series state chain The time-series state chain The expression is:

[0028]

[0029] in, The size of the sliding window. Indicates device status At any moment status labels, Indicates device status At any moment The state change vector;

[0030] For the time-series state chain First-order difference in numerical state calculation And the least squares method is used to fit the linear slope. ;

[0031] Furthermore, by embedding device operation logic, a minimum time interval for state transitions is set, and state transitions that do not meet the minimum time interval are marked as illegal state transitions.

[0032] In the above scheme, the execution steps of the two-stage collaborative mechanism include:

[0033] The initial screening collaboration steps include:

[0034] Jump event detection: Monitor the output state sequence of negative samples and detect changes in state labels between two consecutive frames to generate a jump event queue Ejump;

[0035] Displacement verification: Calculate the pixel displacement at the transition moment using the positive sample algorithm. It also outputs the estimated physical displacement.

[0036] Multi-frame backtracking confirmation: If the displacement amount If the value is greater than or equal to the displacement threshold, backtrack the previous N frames and rerun the negative sample algorithm to confirm the authenticity of the jump.

[0037] Execution trend coordination: When the positive sample algorithm detects changes in the same direction m times consecutively, the time-series state chain is extracted. The historical state numerical sequence in the data, fitted with the slope of change. It is then compared with the slope threshold to generate a slow-change fault warning.

[0038] The steps to implement trend coordination include:

[0039] Trend direction detection, monitoring the change vector of positive samples Count the number of consecutive changes with the same sign to mark potential slowly changing events (Etrend);

[0040] Historical state sequence extraction, from the time-series state chain Extract the state value sequence corresponding to the time period to obtain the numerical state sequence L. The expression of the numerical state sequence L is:

[0041]

[0042] in, Indicates at time The state value;

[0043] And slope fitting and threshold comparison, using linear regression to calculate the slope. The slope The fault probability is compared with the preset fault threshold obtained from the statistical analysis of historical fault data of the equipment, and the gradually changing fault probability is output.

[0044] In the above scheme, step S4, which involves dynamically calculating weights based on environmental parameters and fusing the output results of positive and negative samples, specifically includes:

[0045] Acquire at least one of real-time light intensity, weather code, and equipment age as an environmental parameter;

[0046] The weight parameters are determined based on the environmental parameters using a lookup table or mapping function. and ;

[0047] According to the formula Calculate the final state confidence score ,in This represents the confidence score output by the negative sample algorithm. This represents the natural exponential function.

[0048] In the above scheme, the weight parameter and The adjustment rules include:

[0049] When the light score is lower than the preset value or the weather is rainy or foggy, reduce the weighting parameter. The value;

[0050] When the equipment's service life exceeds the preset limit or its historical failure rate is high, the weighting parameter is increased. The value of .

[0051] In the above scheme, in step S4, the confidence score of the final state is used as a basis for calculation. The value outputs the decision result, including:

[0052] like If so, the negative sample state result will be output directly;

[0053] like If the status is pending, a "Status pending" signal will be output, and a high-definition image will be associated for manual review.

[0054] like This will trigger a drone or robot to conduct a close-range re-inspection.

[0055] The present invention also provides a positive and negative sample collaborative inspection system based on temporal state evolution analysis, wherein when the processor executes the system, it implements the method described therein.

[0056] Because the present invention employs the above-mentioned technical means, it has the following beneficial effects:

[0057] 1. By constructing a time-series state chain, the technical problems of lacking time-dimensional context in state analysis and being unable to identify illegal state transitions are solved, enabling continuous and compliant tracking of device state evolution.

[0058] This invention utilizes a technique of constructing a temporal state chain (specifically, aggregating the historical state sequence of a device using a sliding window). With state change vector This process, which injects equipment operation logic rules to mark illegal transitions, elevates single-point state identification to continuous time-series analysis. This solves the technical problems of existing technologies that only focus on single-moment states, cannot correlate with historical information, and lack judgment criteria on whether state transitions (such as the instantaneous transition from "closing" to "opening" caused by transient interference) conform to equipment operation logic. The direct effect is that it provides a data foundation rich in temporal context and informational logic constraints for subsequent analysis, enabling the initial filtering out of transient abnormal states that clearly violate equipment operation procedures, thus improving the initial reliability of state analysis.

[0059] 2. By using a preliminary screening collaboration mechanism, the technical problem of state misjudgment caused by instantaneous interference in the negative sample algorithm is solved, significantly reducing the false alarm rate.

[0060] This invention employs a preliminary screening and collaborative technique (specifically, based on low-confidence (e.g., θ < 85%) state transition events output by the negative sample algorithm, it triggers the positive sample algorithm to determine the displacement at that moment). This process (verification is performed, and the decision to suppress alarms is based on the verification results) changes the traditional cascaded fusion process where only positive samples trigger negative samples. This solves the technical problem that negative sample algorithms are prone to high-confidence misjudgments when faced with transient interference such as light and shadow occlusion and transient foreign objects, leading to false alarms in the system. The achieved effect is the formation of an error correction mechanism based on physical quantity (displacement) verification, which can effectively identify and suppress most false alarms caused by image interference rather than real faults, significantly reducing the relevant false alarm rate from >10% in existing technologies.

[0061] 3. By using a trend collaboration mechanism, the technical problem of the positive sample algorithm being insensitive to slowly developing faults and prone to missed detection is solved, thereby improving the early warning capability for slowly changing faults.

[0062] This invention utilizes a trend-coordination technique (specifically, based on continuous unidirectional change events detected by the positive sample algorithm, triggering the extraction and trend analysis of historical state numerical sequences output by the negative sample algorithm, such as the fitting slope β) to compensate for the lack of historical correlation in the negative sample algorithm by leveraging the sensitivity of positive samples to minute changes. This solves the technical problem of "gradual change faults," such as the continuous slow drop in oil level caused by aging seals, which are difficult for the negative sample algorithm to detect in a timely manner because the amount of a single change does not exceed a threshold, thus leading to missed detections. The achieved effect is that by capturing and confirming minute change trends, early warnings can be issued before the fault becomes obvious (e.g., 5 days in advance), effectively covering the blind spots of traditional monitoring schemes.

[0063] 4. By using a dynamic confidence-weighted output model, the technical problems of poor adaptability and low decision reliability of fixed-weight strategies in complex and ever-changing environments are solved, thereby enhancing the robustness and environmental adaptability of the system.

[0064] This invention employs a dynamic confidence-weighted output technique (specifically, dynamically adjusting the weights of the positive and negative sample algorithm outputs based on real-time environmental parameters). and And calculate the final state confidence score. By implementing a differentiated output strategy based on the Ct value range, the system's decision-making becomes less rigid. This solves the technical problem that fixed weights cannot reflect the algorithm's actual reliability under different conditions, leading to inaccurate overall judgments, in complex operating conditions such as rain, snow, fog, and low light, or when equipment performance degrades due to long service life. The result is that the system can flexibly adjust the level of trust in different pieces of evidence according to the actual situation, just like an expert, making the final decision more intelligent and reliable, and comprehensively improving the system's applicability in real industrial scenarios. Attached Figure Description

[0065] Figure 1 This is a simplified flowchart of the overall process of the present invention;

[0066] Figure 2 This is a simplified flowchart of one embodiment of step 3 of the present invention. Detailed Implementation

[0067] The embodiments of the present invention will be described in detail below. Although the present invention will be described and illustrated in conjunction with some specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, any modifications or equivalent substitutions made to the present invention should be covered within the scope of the claims of the present invention.

[0068] Furthermore, to better illustrate the present invention, numerous specific details are set forth in the following detailed embodiments. Those skilled in the art will understand that the present invention can be practiced without these specific details.

[0069] This invention provides a positive and negative sample collaborative inspection method based on time-series state evolution analysis, aiming to solve the problems of high false alarm rate for state transitions, high missed detection rate for gradual faults, and poor environmental adaptability in existing technologies. The technical solution is described in detail below, step by step. The overall process is as follows: Figure 1 As shown.

[0070] Step S1: Data acquisition and preprocessing.

[0071] This step involves connecting the real-time image stream and historical status database of the substation's intelligent inspection system to collect equipment status images. and corresponding time-series state data A standardized time-series dataset can be obtained. ,in For the device image with timestamp t, The status labels output by the negative sample algorithm (such as "switch open", "oil temperature 25℃") This is the state change vector output by the positive sample algorithm (e.g., pointer angle change +3°). The specific operation process is as follows.

[0072] First, the raw data is cleaned by using the Z-score algorithm to detect and remove outliers (such as outlier pixels caused by lens smudges). Then, linear interpolation is used to fill in the missing data, resulting in a continuous time-series data stream with the missing data rate reduced to below 0.1%. For example, if an oil level gauge image is missing 5 frames due to rain or fog, a complete sequence can be generated by interpolating adjacent frames.

[0073] Then, the data is aligned according to different modal types, and the image data is aligned based on timestamps. With state data Synchronization is performed using a sliding window (window size = 10 frames) for smoothing, ensuring that the time deviation between the image and the status data is less than 10ms.

[0074] Finally, normalization is performed on the state values ​​(such as oil level scale and temperature reading) by Min-Max normalization, mapping them to the [0,1] interval to eliminate dimensional differences and improve the stability of subsequent fusion calculations.

[0075] For example, taking the main transformer oil level gauge as an example, infrared images are acquired. (Resolution 1920×1080); Negative sampling algorithm outputs oil level scale values. (Confidence score 85%); Positive sample algorithm calculates the pixel displacement relative to the previous frame. .

[0076] Step S2: Constructing the Temporal State Chain

[0077] For each device status Aggregate its recent Generate a time-series state chain from the state data at each time point. This generates a sequence of equipment state evolution trajectories, which is used to analyze the continuity of state transitions.

[0078]

[0079] in, The size of the sliding window. Indicates device status At any moment status labels, Indicates device status At any moment State change vector; in terms of switch open / closed states For example, (A total of 10 state sequences), associated with the displacement changes of positive sample outputs. ,in Represents pixels.

[0080] The specific operation process is as follows:

[0081] First, state sequence extraction is performed, extracting state labels from a sliding window in the database in chronological order. and change This yields the original state chain. .

[0082] Then calculate the trend characteristics and perform first-order difference calculations on numerical states (such as oil level and temperature). And by fitting a linear slope β (least squares method), an index of state change trend is obtained. (e.g., the average daily decline slope of oil level = 0.05 / day).

[0083] Finally, logical rules are injected. By embedding the device operation logic (such as the minimum time interval between "opening and closing" = 2s), illegal state transitions are marked and invalid transitions are identified (such as "opening and closing" within 0.5s is considered abnormal).

[0084] Step S3: Execution of the two-stage collaborative mechanism

[0085] This stage includes two sub-steps: initial screening coordination and trend coordination. Its purpose is to prevent instantaneous misjudgments and missed detections of slowly changing faults. A simplified flowchart is shown below. Figure 2 As shown.

[0086] The applicable scenario for initial screening collaboration is when the negative sample algorithm outputs a state jump and the confidence level θ < 85%, triggering the positive sample algorithm to calculate the displacement. ,like If the displacement threshold is exceeded, alarms are suppressed. This eliminates false alarms caused by light and shadow occlusion.

[0087] To illustrate with a specific example, negative samples identify the switch's "open → close" state (70% confidence level), while positive samples calculate the displacement. (Threshold = 100 pixels) → Determined as imaging interference, no alarm is issued.

[0088] The initial screening collaboration specifically includes the following steps:

[0089] 1. Jump event detection: Monitor the output state sequence of negative samples and detect changes in state labels between two consecutive frames (e.g., ...). , ), generate the jump event queue Ejump.

[0090] 2. Displacement verification: Calculate the pixel displacement at the transition moment using the positive sample algorithm. (Through optical flow or feature point matching) Output physical displacement estimate.

[0091] 3. Multi-frame backtracking confirmation, if If the threshold is reached, the negative sample algorithm is rerun for the previous 5 frames to confirm the authenticity of the jump (if 3 out of 5 frames are "closed", then it is confirmed as a real jump).

[0092] Trend coordination is applicable when the positive sample algorithm detects changes in the same direction for m=5 consecutive times (such as a continuous drop in oil level), and extracts the historical state sequence of the negative samples. The slope β of the fitted change is compared with a threshold to provide an early warning 5 days in advance for oil level slow-drop faults caused by seal aging. For example, if the oil level drops for 5 consecutive days (average daily drop of 0.5%), the fitted slope β = −0.05 / day (threshold = -0.03 / day) → generates a "slow-drop defect warning".

[0093] Trend coordination includes the following steps:

[0094] Trend direction detection monitors the change vector Δt of positive samples and counts the number of consecutive changes with the same sign (e.g., ...). ), mark potential slowly changing events as Etrend.

[0095] Historical state sequence extraction, from temporal state chain Extract the state value sequence for the corresponding time period to obtain the numerical state sequence. (e.g., oil level markings).

[0096] The slope fitting and threshold comparison uses linear regression to calculate the slope β, which is then compared with the preset fault threshold (based on historical fault data statistics of the equipment) to output the gradually changing fault probability (if β < βthreshold, an early warning is issued).

[0097] Step S4: Dynamic confidence weighted output

[0098] This step dynamically calculates weights based on environmental parameters (light intensity score, historical equipment failure rate). and Combine positive and negative sample outputs:

[0099]

[0100] Output final state confidence score To achieve adaptive decision-making.

[0101] For example, in rainy or foggy weather (sunshine score = 30 / 100): adjust downwards. Reduce the weight of negative samples to 0.3; for example, increase the weight of historically high-failure equipment. Up to 1.2, the weighting of positive sample changes is enhanced.

[0102] Specifically, the following steps are included:

[0103] First, environmental parameters are collected. Real-time light intensity, weather code, equipment service life and other parameters are obtained from the inspection system to generate an environmental vector Venv=[light, weather, service life].

[0104] Then, the weights are dynamically calculated, and the environment vector is mapped to the weight parameters through a lookup table. and (Table 1):

[0105] Table 1:

[0106]

[0107] Finally, a decision is output based on... Value selection output strategy:

[0108] Directly output the state results of the negative samples;

[0109] Output "Status pending," linking to high-resolution images for manual verification;

[0110] : Triggered close-range re-inspection of the drone.

Claims

1. A method for collaborative inspection of positive and negative samples based on temporal state evolution analysis, characterized in that, Includes the following steps: Step S1, Data Acquisition and Preprocessing: Acquire the device's real-time image stream and historical status database, preprocess the data to obtain data including timestamps. Device image State labels output by the negative sample algorithm and the state change vector output by the positive sample algorithm Standardized time series datasets ; Step S2, Timing State Chain Construction: For each device state Aggregate the latest Generate a time-series state chain from the state data at each time point. The time-series state chain Includes a sequence of status labels and state change vector And inject device operation logic rules to mark illegal state transitions; Step S3: Execution of the two-stage collaborative mechanism: Step S31: Perform initial screening coordination: When the negative sample algorithm outputs a state jump and the confidence level is lower than the first threshold, the positive sample algorithm is triggered to calculate the displacement. If the displacement amount If the displacement is less than the threshold, the alarm is suppressed; Step S32, Execute Trend Coordination: When the positive sample algorithm is continuous When a change in the same direction is detected for the first time, the time-series state chain is extracted. The historical state numerical sequence in the data, fitted with the slope of change. If the slope of the change If the slope threshold is exceeded, a slow-change fault warning will be generated; Step S4: Dynamic Confidence Weighted Output: Weights are dynamically calculated based on environmental parameters, and the outputs of positive and negative samples are fused to obtain the final state confidence score. And based on the final state confidence score The value is used to output the decision result.

2. The method according to claim 1, characterized in that, The data preprocessing in step S1 includes: The Z-score algorithm is used to detect and remove outliers, and linear interpolation is used to fill in the missing data. Image data based on timestamps With state data Synchronization is performed, and a sliding window is used for smoothing. The state values ​​are normalized using the Min-Max method.

3. The method according to claim 1, characterized in that, The time-series state chain construction step described in step S2 includes: For each device status Extract the state labels of the k most recent time moments from the database in chronological order. and state change vector Construct a time-series state chain The time-series state chain The expression is: in, The size of the sliding window. Indicates device status At any moment status labels, Indicates device status At any moment The state change vector; For the time-series state chain First-order difference in numerical state calculation And the least squares method is used to fit the linear slope. ; Furthermore, by embedding device operation logic, a minimum time interval for state transitions is set, and state transitions that do not meet the minimum time interval are marked as illegal state transitions.

4. The method according to claim 1, characterized in that, In step S2, the injected device operation logic rules include: setting a minimum time interval for state transitions according to the device operation logic, and marking state transitions that do not meet the minimum time interval as illegal state transitions.

5. The method according to claim 1, characterized in that, The two-stage collaborative mechanism execution steps include: The initial screening collaboration steps include: Jump event detection: Monitor the output state sequence of negative samples and detect changes in state labels between two consecutive frames to generate a jump event queue Ejump; Displacement verification: Calculate the pixel displacement at the transition moment using the positive sample algorithm. It also outputs the estimated physical displacement. Multi-frame backtracking confirmation: If the displacement amount If the value is greater than or equal to the displacement threshold, backtrack the previous N frames and rerun the negative sample algorithm to confirm the authenticity of the jump. Execution trend coordination: When the positive sample algorithm detects changes in the same direction m times consecutively, the time-series state chain is extracted. The historical state numerical sequence in the data, fitted with the slope of change. It is then compared with the slope threshold to generate a slow-change fault warning; The steps to implement trend coordination include: Trend direction detection, monitoring the change vector of positive samples Count the number of consecutive changes with the same sign to mark potential slowly changing events (Etrend); Historical state sequence extraction, from the time-series state chain Extract the state value sequence corresponding to the time period to obtain the numerical state sequence L. The expression of the numerical state sequence L is: in, Indicates at time The state value; And slope fitting and threshold comparison, using linear regression to calculate the slope. The slope The fault probability is compared with the preset fault threshold obtained from the statistical analysis of historical fault data of the equipment, and the gradually changing fault probability is output.

6. The method according to claim 1, characterized in that, In step S4, the weights are dynamically calculated based on environmental parameters, and the output results of positive and negative samples are fused. Specifically, this includes: Acquire at least one of real-time light intensity, weather code, and equipment age as an environmental parameter; The weight parameters are determined based on the environmental parameters using a lookup table or mapping function. and ; According to the formula Calculate the final state confidence score ,in This represents the confidence score output by the negative sample algorithm. This represents the natural exponential function.

7. The method according to claim 6, characterized in that, The weight parameters and The adjustment rules include: When the light score is lower than the preset value or the weather is rainy or foggy, reduce the weighting parameter. The value; When the equipment's service life exceeds the preset limit or its historical failure rate is high, the weighting parameter is increased. The value of .

8. The method according to claim 7, characterized in that, In step S4, the final state credibility score is used as a basis. The value outputs the decision result, including: like If so, the negative sample state result will be output directly; like If the status is pending, a "Status pending" signal will be output, and a high-definition image will be associated for manual review. like This will trigger a drone or robot to conduct a close-range re-inspection.

9. A positive and negative sample collaborative inspection system based on temporal state evolution analysis, characterized in that, The system includes a processor configured to perform the method as described in any one of claims 1-8.

Images