Self-adaptive sensing detection method and system for power industry station monitoring

Abstract

The invention provides a self-adaptive sensing detection method and system for power industry station monitoring, and the method comprises the following steps: S1, collecting the field data of a power industry station, and carrying out the primary processing, control and feature extraction of the data; s2, performing deep processing and feature enhancement on the preprocessed data; s3, performing small target detection and bounding box regression on the multi-scale features to obtain a feature result; s4, inputting the category and position information of the feature result and the historical track sequence into a multi-head attention mechanism to obtain a final event identification feature, and calculating a risk score; s5, the risk scores are graded, a control instruction is output, and an alarm is given out; the self-adaptive sensing detection method and system are compact in structure, low in calculation cost and high in detection precision, and can continuously, dynamically and accurately monitor small targets and abnormal behaviors on the site of the power industry station.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and artificial intelligence, and in particular to an adaptive perception and detection method and system for power industry power plant monitoring. Background Technology

[0002] In the safety monitoring of power industrial plants, small target detection technology plays an extremely important role. Power industrial plant operating areas are densely packed with equipment and have complex structures, with workers often located a considerable distance from the equipment. Key targets captured by cameras often occupy only a tiny fraction of the frame. Accurate identification of these small targets is directly related to assessing personnel safety status and detecting violations, making it a crucial link in ensuring the operational safety of power industrial plants and achieving intelligent monitoring and management. However, due to the limited pixels occupied by small targets in images, their indistinct feature representation, and their susceptibility to changes in ambient lighting, equipment reflections, and occlusion interference, traditional target detection algorithms often struggle to achieve high-precision identification, leading to frequent false positives and false negatives, thus affecting the reliability and practicality of the safety monitoring system. Therefore, improving detection accuracy and model robustness to address the small target detection problem in power industrial plant scenarios has become an urgent need for the development of intelligent monitoring systems.

[0003] Currently, methods for solving the small target detection problem can be mainly divided into several categories. The first category is multi-scale feature fusion methods. These methods enhance the model's ability to recognize targets of different sizes by fusing shallow texture features with high-level semantic features across different layers. This method can improve the detection performance of small targets to some extent, but due to the large differences in semantic and spatial resolution between feature layers, inaccurate feature alignment and semantic information dilution are prone to occur during the fusion process, leading to inaccurate localization and blurred boundaries of small targets. The second category is attention-based feature enhancement methods. By introducing channel attention, spatial attention, or hybrid attention mechanisms, the model can automatically focus on key regions and suppress invalid background interference, thereby improving the saliency of small target features. The drawback of this type of method is that it requires additional computational resources, increasing model complexity and inference time, which is not conducive to real-time deployment in practical monitoring systems. The third category is lightweight network structure optimization methods. These methods use lightweight convolutional designs such as GhostConv and ShuffleNet to reduce redundant computation and improve detection efficiency. This method is suitable for embedded and edge computing environments, but due to the simplified network structure, the feature extraction capability is limited, and accuracy is prone to decline when recognizing small targets in complex backgrounds. The fourth category is upsampling reconstruction and feature recovery methods. These methods enhance the detail recovery capability of feature maps by improving the upsampling module (such as deconvolution, deformable convolution, or dynamic sampling) to better reconstruct the edge features of small targets. Although this method can improve the detail representation of small targets, existing upsampling mechanisms lack adaptability and are difficult to dynamically adjust the sampling strategy according to the target size and position, resulting in unstable feature reconstruction. Summary of the Invention

[0004] To overcome the shortcomings of the prior art, this invention discloses an adaptive sensing and detection method and system for power industry power plant monitoring.

[0005] The technical solution provided by this invention is: an adaptive sensing and detection method for power industry power plant monitoring, comprising the following steps:

[0006] S1: The data acquisition and preprocessing module acquires on-site data from power industry power plants, performs preliminary processing, control, and feature extraction on the data, and transmits the preprocessed data to the processing and feature enhancement module.

[0007] S2: The processing and feature enhancement module performs deep processing and feature enhancement on the preprocessed data to obtain enhanced multi-scale features, which are then transmitted to the adaptive perception and detection module.

[0008] S3: The adaptive perception detection module performs small target detection and bounding box regression on multi-scale features. It extracts the feature backbone through a lightweight backbone network and dynamic attention mechanism, fuses multi-scale features, and decouples the detection head to input the feature results into the classification branch and regression branch for independent modeling. At the same time, the feature results are sent to the abnormal state recognition and risk module.

[0009] S4: The abnormal state identification and risk assessment module includes an event identification submodule and a risk scoring submodule. The event identification submodule obtains the final event identification features by inputting the category, location information and historical trajectory sequence of the feature results into the multi-head attention mechanism. The risk scoring submodule calculates the risk score based on the final event identification features and then passes the risk score to the system control and alarm response module.

[0010] S5: The system control and alarm response module includes a risk threshold determination submodule, a control decision submodule, and an alarm triggering submodule. Through decision logic and priority scheduling, it classifies risk scores into levels and outputs control commands to issue warnings.

[0011] Preferably, the on-site data in S1 includes data from video surveillance cameras, infrared thermal imagers, and environmental sensors.

[0012] Further preferred, the preliminary processing, control, and feature extraction in S1 specifically include: performing time synchronization and frame-level caching on the collected data; performing preliminary quality control on the original image, including: resolution standardization, frame rate adjustment, and integrity check; and performing dynamic sampling and key frame selection on key areas according to a preset strategy to increase the observation frequency of small targets and high-risk operation areas.

[0013] Further preferably, the depth processing and feature enhancement methods in S2 are as follows:

[0014] The original image is subjected to noise suppression and brightness correction to eliminate the effects of uneven lighting, shadow interference and sensor noise in the monitoring environment;

[0015] Multi-scale feature enhancement of images is performed, including local contrast enhancement and key region saliency enhancement, to make small targets stand out more in complex backgrounds. In view of the characteristics of small target size and sparse distribution, a learnable small target enhancement mechanism is adopted, which adjusts the importance of each channel by feature weighting to highlight the features of potential small targets.

[0016] The processed images are subjected to size standardization and batch normalization to unify the scale of input features and reduce numerical fluctuations during the training and inference phases.

[0017] Further preferred, the specific method of S3 is as follows:

[0018] The adaptive perception detection module includes: a backbone feature extraction submodule, a multi-scale feature fusion submodule, and a decoupled detection head submodule;

[0019] The main feature extraction submodule extracts multi-scale features using a lightweight convolutional structure and a channel-spatial attention mechanism, including channel attention. Spatial attention The calculation is as follows:

[0020] (1)

[0021] (2)

[0022] The weighted features are:

[0023] (3)

[0024] in, and These are global average pooling and global max pooling, respectively. For the Sigmoid function, for Convolution operation;

[0025] In the multi-scale feature fusion submodule, let the shallow features be... Deep features are fusion features Learnable weights adjust:

[0026] (4)

[0027] in, Automatic optimization through gradient descent enhances the response of key regional features across different scales;

[0028] The dynamic attention sampling mechanism adaptively calculates the sampling position through the DySample module. The calculation formula is:

[0029] (5)

[0030] in The offset prediction function learned by the convolutional layer. Used for bilinear interpolation sampling to recover the edge and detail features of small targets; the sampled features are denoted as... ;

[0031] The decoupled detection head will upsample the features The classification and regression branches are modeled independently:

[0032] (6)

[0033] in For class probabilities, This is the bounding box offset. and For learnable parameter matrix and bias.

[0034] Further preferably, in step S4, the final event recognition feature calculation method is as follows:

[0035] Let the detection sequence be subscript Indicates the time step, and the target feature dimension is... Note that the number of heads is ;

[0036] Query, key, and value vector computation: For the first One point to note, at each time step In the middle, through the learnable weight matrix Multiplication generates query ,key Sum vector:

[0037] (7)

[0038] Scaling Dot Product Attention: Computing the 1st Attention output by size:

[0039] (8)

[0040] in, For the mask matrix, the upper triangular elements are set to This allows risk assessment to rely solely on historical and current data; The dimension is the key vector to ensure scaling stability;

[0041] Multi-head output integration: The outputs of each head are concatenated and linearly projected to obtain the final event recognition features:

[0042] (9)

[0043] Further preferably, in step S4, the formula for calculating the risk score is:

[0044] (10)

[0045] in, As a risk score, As a characteristic of the event, Historical data statistical characteristics It is the Sigmoid activation function. This indicates a feature concatenation operation. and For learnable parameter matrix and bias, risk score It can be mapped to three risk levels: low, medium, and high.

[0046] Further optimized, the specific method for classifying risk scores and outputting control commands and issuing warnings in step S5 is as follows:

[0047] Risk score classification: Let the risk score sequence be... ,in Indicates time step The risk value, the risk score With preset threshold , , Comparison of risk level classifications:

[0048] (11)

[0049] in, and The settings are based on historical data statistics and expert experience;

[0050] Control decisions: based on risk level With event type The module executes the decision function. Output control commands :

[0051] (12)

[0052] Control commands include: on-site light warnings, voice prompts, equipment shutdown, and remote notification operations;

[0053] Alarm triggered: The system responds to control commands. Activate the alarm actuator and simultaneously transmit event information. The alarm is recorded in the log database and uploaded to the dispatch center in real time via a network interface. The alarm trigger function can be represented as follows:

[0054] (13)

[0055] in This indicates that an alarm has been triggered. This indicates that it has not been triggered.

[0056] According to another aspect of the present invention, an adaptive sensing and detection system for power industry power plant monitoring is also provided, including a data acquisition and preprocessing module, a processing and feature enhancement module, an adaptive sensing and detection module, an abnormal state identification and risk assessment module, and a system control and alarm response module.

[0057] The data acquisition and preprocessing module is used to acquire and preprocess on-site data;

[0058] The processing and feature enhancement module is used to further process and enhance the features of the collected field data;

[0059] The adaptive perception detection module is used to perform small target detection and bounding box regression on the multi-scale features output by the preprocessing and feature enhancement modules;

[0060] The abnormal state identification and risk assessment module is used to analyze the target detection results output by the adaptive perception and detection module to determine the risk level of power industry station operation behavior, equipment status and safety events;

[0061] The system control and alarm response module is used to process the risk level and event information output by the abnormal state identification and risk assessment module, and to provide real-time intervention and alarm management for the safety of power industry stations.

[0062] The beneficial effects of this invention are:

[0063] This invention addresses the needs of small target detection and real-time risk response in power industry power plant safety monitoring. It proposes an intelligent detection system based on multi-source sensing data fusion and adaptive feature enhancement. While ensuring the accuracy of small target detection, it effectively avoids the missed and false detection problems caused by sparse features, severe occlusion, and complex backgrounds in traditional monitoring methods. By introducing a lightweight backbone network, channel-space attention mechanism, multi-scale feature fusion, and dynamic upsampling technology, it achieves high-precision boundary recovery and feature expression optimization for small targets, enabling comprehensive and accurate capture of minute movement features of personnel, equipment, and tools. Utilizing a multi-head attention mechanism to fuse and analyze historical trajectories and current detection results, and combining risk scoring and event type to generate control commands, it achieves real-time identification and risk quantification of abnormal behavior and equipment status, avoiding the risks of delayed response or incorrect judgment in complex scenarios associated with traditional manual inspections or single-rule systems. The system control and alarm response module adaptively triggers on-site alarms, remote notifications, and safety intervention measures based on risk levels, achieving closed-loop safety management while ensuring real-time performance and response speed.

[0064] The adaptive sensing and detection system proposed in this invention features a compact structure, low computational overhead, and high detection accuracy. It enables continuous, dynamic, and accurate monitoring of small targets and abnormal behaviors at power industrial sites. Furthermore, it can be deployed on embedded or edge devices to achieve real-time data processing and intelligent control, providing technical support and decision-making basis for the safety management of power industrial sites. Application in power industrial sites and similar high-risk industrial environments can significantly improve the ability to identify small targets and intervene in anomalies, reduce the probability of human error and safety accidents, and ensure the continuous, stable, and safe operation of the system. Attached Figure Description

[0065] Figure 1 A flowchart of an adaptive sensing and detection method for power industry power plant monitoring provided by the present invention;

[0066] Figure 2 The flowchart of risk assessment and alarm in an adaptive sensing and detection method for power industry station monitoring provided by the present invention;

[0067] Figure 3 A heatmap of detection results from the YOLOv10 network model;

[0068] Figure 4 The present invention provides a heat map of the detection results of an adaptive sensing and detection method for power industry power station monitoring. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0070] This invention provides an adaptive sensing and detection method for power industry substation monitoring, with reference to... Figure 1 As shown, it includes the following steps:

[0071] S1: The data acquisition and preprocessing module acquires on-site data from power industry power plants, performs preliminary processing, control, and feature extraction on the data, and transmits the preprocessed data to the processing and feature enhancement module.

[0072] The system acquires on-site data from power industry power plants through multi-source sensing devices, including video surveillance cameras, infrared thermal imagers, and environmental sensors, to cover key information such as personnel operations, equipment status, and environmental changes.

[0073] The acquired images and sensor data are synchronized in time and cached at the frame level to ensure the consistency of multi-channel data in the time dimension and avoid monitoring blind spots caused by data frame loss and delay.

[0074] Preliminary quality control is performed on the original images, including resolution normalization, frame rate adjustment, and integrity checks, to ensure that the input data meets the accuracy requirements of the subsequent feature extraction and detection modules.

[0075] Based on a preset strategy, key areas are dynamically sampled and key frames are selected to increase the observation frequency of small targets and high-risk operation areas, providing a reliable data foundation for subsequent adaptive feature enhancement and detection.

[0076] S2: The processing and feature enhancement module performs deep processing and feature enhancement on the preprocessed data to obtain enhanced multi-scale features, which are then transmitted to the adaptive perception and detection module.

[0077] The raw images provided by the sensing and acquisition module are subjected to noise suppression and brightness correction to eliminate the effects of uneven lighting, shadow interference and sensor noise in the monitoring environment, thereby improving the overall image quality.

[0078] Multi-scale feature enhancement of the image is performed, including local contrast enhancement and key region saliency enhancement, to make small targets stand out more in complex backgrounds, making it easier for subsequent detection modules to capture target edge and texture information. At the same time, in view of the characteristics of small target size and sparse distribution, a learnable small target enhancement mechanism is adopted, which adjusts the importance of each channel through feature weighting to highlight the features of potential small targets.

[0079] The processed images are subjected to size standardization and batch normalization to unify the scale of input features and reduce numerical fluctuations during training and inference, thereby ensuring the stable transmission of features in the network.

[0080] The module outputs the enhanced multi-scale features to the adaptive perception detection module, providing high-quality, distinguishable input features for the accurate identification of small targets in the future;

[0081] S3: The adaptive perception and detection module is used to perform small target detection and bounding box regression on the multi-scale features output by the preprocessing and feature enhancement module, thereby achieving high-precision identification of small targets such as personnel, safety helmets, and work tools in power industry stations. The module achieves adaptive perception of targets at different scales and in complex backgrounds through a lightweight backbone network and dynamic attention mechanism. The core of the module includes a backbone feature extraction submodule, a multi-scale feature fusion submodule, and a decoupled detection head submodule. Let the input features be... The steps are as follows:

[0082] Feature backbone extraction: Input features are enhanced using a lightweight convolutional structure and a channel-spatial attention mechanism, including channel attention. Spatial attention The calculation is as follows:

[0083] (1)

[0084] (2)

[0085] The weighted features are:

[0086] (3)

[0087] in, and These are global average pooling and global max pooling, respectively. For the Sigmoid function, for Convolution operation;

[0088] Multi-scale feature fusion: Let the shallow features be... Deep features are fusion features Learnable weights adjust:

[0089] (4)

[0090] Automatic optimization through gradient descent is used to enhance the response of key regional features across different scales;

[0091] Dynamic upsampling: The sampling position is adaptively calculated through the DySample module. :

[0092] (5)

[0093] in The offset prediction function learned by the convolutional layer. Used for bilinear interpolation sampling to recover edge and detail features of small targets, the upsampled features are denoted as ;

[0094] Decoupled detection head: Upsampled features The classification and regression branches are modeled independently:

[0095] (6)

[0096] in For class probabilities, This is the bounding box offset. and For learnable parameter matrix and bias;

[0097] S4: The anomaly state identification and risk assessment module analyzes the target detection results output by the adaptive perception and detection module to determine the risk level of power industry station operations, equipment status, and safety events. This module achieves precise identification and risk quantification of small target behaviors through multi-dimensional feature fusion and rule constraints.

[0098] The core of this module includes an event recognition submodule and a risk scoring submodule. The event recognition submodule captures the time-series dependency of work behavior and equipment status by inputting target category, location information and historical trajectory sequence into the multi-head attention mechanism. The multi-head structure effectively improves the ability to identify abnormal behavior of small targets by paying attention to features of different subspaces in parallel.

[0099] Let the detection sequence be subscript Indicates the time step, and the target feature dimension is... Note that the number of heads is The calculation of the masked multi-head attention mechanism is as follows:

[0100] Query, key, and value vector computation: For the first One point to note, at each time step In the middle, through the learnable weight matrix Multiplication generates query ,key Sum vector:

[0101] (7)

[0102] Scaling Dot Product Attention: Computing the 1st Attention output by size:

[0103] (8)

[0104] in, For the mask matrix, the upper triangular elements are set to This is used to shield information from future time steps, ensuring that risk assessments rely solely on historical and current data. The dimension is the key vector to ensure scaling stability;

[0105] Multi-head output integration: The outputs of each head are concatenated and linearly projected to obtain the final event recognition features:

[0106] (9)

[0107] Subsequently, the risk scoring submodule is based on event characteristics. Historical data statistical characteristics Calculate risk score :

[0108] (10)

[0109] in, It is the Sigmoid activation function. This indicates a feature concatenation operation. and For learnable parameter matrix and bias, risk score It can be mapped to three risk levels: low, medium, and high, and transmitted to the system control and alarm response module to achieve real-time early warning and safety management;

[0110] S5: The System Control and Alarm Response module processes the risk level and event information output by the Abnormal State Identification and Risk Assessment module, enabling real-time intervention and alarm management for power industry substation safety. This module controls field equipment and triggers alarm systems through decision logic and priority scheduling to ensure timely responses to high-risk events. The flowchart is shown below. Figure 2 As shown;

[0111] The core of the module includes a risk threshold determination submodule, a control decision submodule, and an alarm triggering submodule. Let the risk scoring sequence be... ,in Indicates time step The risk value of the module is calculated as follows:

[0112] Risk threshold determination: First, assign risk scores... With preset threshold , , Comparison to achieve risk level classification:

[0113] (11)

[0114] in, and Settings can be configured based on historical data statistics and expert experience;

[0115] S502 Control Decision: Based on Risk Level With event type The module executes the decision function. Output control commands :

[0116] (12)

[0117] Control commands include on-site light warnings, voice prompts, equipment shutdown, and remote notification operations to ensure timely response to potential safety hazards;

[0118] Alarm triggered: The system responds to control commands. Activate the alarm actuator and simultaneously transmit event information. Records are stored in a log database and can be uploaded to the dispatch center in real time via a network interface, enabling remote monitoring and historical tracking. The alarm trigger function can be represented as:

[0119] (13)

[0120] in This indicates that an alarm has been triggered. This indicates that it has not been triggered.

[0121] Another aspect of the present invention provides an adaptive sensing and detection system for power industry substation monitoring, comprising:

[0122] The system includes a data acquisition and preprocessing module, a processing and feature enhancement module, an adaptive perception and detection module, an abnormal state identification and risk assessment module, and a system control and alarm response module.

[0123] The data acquisition and preprocessing module is used to acquire and preprocess on-site data;

[0124] The processing and feature enhancement module is used to further process and enhance the features of the collected field data;

[0125] The adaptive perception detection module is used to perform small target detection and bounding box regression on the multi-scale features output by the preprocessing and feature enhancement modules;

[0126] The abnormal state identification and risk assessment module is used to analyze the target detection results output by the adaptive perception and detection module to determine the risk level of power industry station operation behavior, equipment status and safety events;

[0127] The system control and alarm response module is used to process the risk level and event information output by the abnormal state identification and risk assessment module, and to provide real-time intervention and alarm management for the safety of power industry stations.

[0128] The principle of this method is:

[0129] This invention first acquires multi-source field data from power industry power plants through a sensing and acquisition module, including video, infrared, and environmental sensor information. The acquired data undergoes time synchronization, keyframe selection, and quality control to provide a reliable foundation for subsequent processing. Secondly, a preprocessing and feature enhancement module performs noise suppression, brightness correction, multi-scale feature enhancement, and normalization on the acquired data, thereby highlighting small target features and improving feature representation capabilities. Then, a lightweight backbone network, channel-space attention mechanism, multi-scale feature fusion, and dynamic upsampling are introduced into the adaptive sensing and detection module to achieve high-precision detection and bounding box regression of small targets. Next, an anomaly identification and risk assessment module fuses historical trajectories with current detection information through a multi-head attention mechanism to quantitatively assess abnormal behavior of small targets and equipment risks, achieving continuous dynamic risk scoring. Finally, a system control and alarm response module generates control commands based on risk level and event type, triggering on-site alarms and remote management, realizing closed-loop control from small target detection and anomaly identification to safety intervention.

[0130] Example

[0131] This embodiment uses the outdoor monitoring area of ​​a 220kV substation as the test scenario to deploy the adaptive perception and detection system of the present invention to detect and identify the safety helmet wearing status and operation behavior of the workers.

[0132] The system acquires on-site video data through fixed high-definition surveillance cameras, with a sampling frame rate of 25fps and a resolution of 1920×1080. The system automatically performs frame buffering and time synchronization processing, and filters out key frames containing personnel activities as input.

[0133] Brightness equalization and noise suppression are performed on keyframe images, and then a multi-scale enhancement algorithm is used to improve local contrast, making small targets (helmets, tools) more prominent in the feature map;

[0134] After extracting multi-layer features through a lightweight feature extraction network, a channel-space joint attention mechanism is used to enhance the response of key regions. The weights of shallow and deep features are self-adjusted through an adaptive multi-scale feature fusion module. Then, the dynamic upsampling module (DySample) is used to recover the detailed features of small targets. Finally, the detection results are output through an efficient decoupled detection head, including personnel category, helmet wearing status, operation behavior, and tool position.

[0135] The detected target information and historical trajectory features are input into a multi-head attention model to automatically analyze personnel behavior patterns and output a risk score. If the score exceeds a set threshold, the system automatically identifies it as an abnormal event (not wearing a safety helmet or entering a restricted area).

[0136] The control module triggers corresponding measures based on the risk level: for medium-risk events, the system issues a voice prompt through the on-site loudspeaker; for high-risk events, it simultaneously triggers warning lights and uploads alarm information to the monitoring center to achieve real-time safety intervention.

[0137] Figure 3 The detection results are based on the baseline YOLOv10 model. Figure 4 The detection results after introducing a feature enhancement network, a decoupled detection head, and an adaptive perception module are shown in the figure. As can be seen from the figure, the present invention can more accurately identify small targets such as substation workers, safety helmets, and tools. The target response area in the heat map of the detection results is more concentrated and the boundary is clearer, indicating that both the detection accuracy and robustness are improved.

[0138] It should be understood that, inspired by the inventive concept, those skilled in the art can make various improvements and modifications based on the above description without departing from the scope of the invention, and these modifications still fall within the protection scope of the invention.

[0139] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims.

[0140] It should be understood that the present invention is not limited to the precise structure shown in the above description, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. An adaptive sensing and detection method for power industry substation monitoring, characterized in that, Includes the following steps: S1: The data acquisition and preprocessing module acquires on-site data from power industry power plants, performs preliminary processing, control, and feature extraction on the data, and transmits the preprocessed data to the processing and feature enhancement module. S2: The processing and feature enhancement module performs deep processing and feature enhancement on the preprocessed data to obtain enhanced multi-scale features, which are then transmitted to the adaptive perception and detection module. S3: The adaptive perception detection module performs small target detection and bounding box regression on multi-scale features. It extracts the feature backbone through a lightweight backbone network and dynamic attention mechanism, fuses multi-scale features, and decouples the detection head to input the feature results into the classification branch and regression branch for independent modeling. At the same time, the feature results are sent to the abnormal state recognition and risk module. S4: The abnormal state identification and risk assessment module includes an event identification submodule and a risk scoring submodule. The event identification submodule obtains the final event identification features by inputting the category, location information and historical trajectory sequence of the feature results into the multi-head attention mechanism. The risk scoring submodule calculates the risk score based on the final event identification features and then passes the risk score to the system control and alarm response module. S5: The system control and alarm response module includes a risk threshold determination submodule, a control decision submodule, and an alarm triggering submodule. Through decision logic and priority scheduling, it classifies risk scores into levels and outputs control commands to issue warnings.

2. The adaptive sensing and detection method for power industry substation monitoring according to claim 1, characterized in that, The field data in S1 includes data from video surveillance cameras, infrared thermal imagers, and environmental sensors.

3. The adaptive sensing and detection method for power industry substation monitoring according to claim 1, characterized in that, The preliminary processing, control, and feature extraction in S1 specifically include: performing time synchronization and frame-level caching on the collected data; performing preliminary quality control on the original image, including resolution standardization, frame rate adjustment, and integrity checks; and dynamically sampling and selecting key frames for key areas according to a preset strategy to increase the observation frequency of small targets and high-risk operation areas.

4. The adaptive sensing and detection method for power industry substation monitoring according to claim 1, characterized in that, The depth processing and feature enhancement methods in S2 are specifically as follows: The original image is subjected to noise suppression and brightness correction to eliminate the effects of uneven lighting, shadow interference and sensor noise in the monitoring environment; Multi-scale feature enhancement of images is performed, including local contrast enhancement and key region saliency enhancement, to make small targets stand out more in complex backgrounds. In view of the characteristics of small target size and sparse distribution, a learnable small target enhancement mechanism is adopted, which adjusts the importance of each channel by feature weighting to highlight the features of potential small targets. The processed images are subjected to size standardization and batch normalization to unify the scale of input features and reduce numerical fluctuations during the training and inference phases.

5. The adaptive sensing and detection method for power industry substation monitoring according to claim 1, characterized in that, The specific method of S3 is as follows: The adaptive perception detection module includes: a backbone feature extraction submodule, a multi-scale feature fusion submodule, and a decoupled detection head submodule; The main feature extraction submodule extracts multi-scale features using a lightweight convolutional structure and a channel-spatial attention mechanism, including channel attention. Spatial attention The calculation is as follows: （1） （2） The weighted features are: （3） in, and These are global average pooling and global max pooling, respectively. For the Sigmoid function, for Convolution operation; In the multi-scale feature fusion submodule, let the shallow features be... Deep features are fusion features Learnable weights adjust: （4） in, Automatic optimization through gradient descent enhances the response of key regional features across different scales; The dynamic attention sampling mechanism adaptively calculates the sampling position through the DySample module. The calculation formula is: （5） in The offset prediction function learned by the convolutional layer. Used for bilinear interpolation sampling to recover the edge and detail features of small targets; the sampled features are denoted as... ; The decoupled detection head will upsample the features The classification and regression branches are modeled independently: （6） in For class probabilities, This is the bounding box offset. and For learnable parameter matrix and bias.

6. The adaptive sensing and detection method for power industry substation monitoring according to claim 1, characterized in that, In step S4, the final event recognition feature calculation method is as follows: Let the detection sequence be subscript Indicates the time step, and the target feature dimension is... Note that the number of heads is ; Query, key, and value vector computation: For the first One point to note, at each time step In the middle, through the learnable weight matrix Multiplication generates query ,key Sum vector: （7） Scaling Dot Product Attention: Computing the 1st Attention output by size: （8） in, For the mask matrix, the upper triangular elements are set to This allows risk assessment to rely solely on historical and current data; The dimension is the key vector to ensure scaling stability; Multi-head output integration: The outputs of each head are concatenated and linearly projected to obtain the final event recognition features: （9）。 7. The adaptive sensing and detection method for power industry substation monitoring according to claim 1, characterized in that, In S4, the formula for calculating the risk score is: （10） in, As a risk score, As a characteristic of the event, Historical data statistical characteristics It is the Sigmoid activation function. This indicates a feature concatenation operation. and For learnable parameter matrix and bias, risk score It can be mapped to three risk levels: low, medium, and high.

8. The adaptive sensing and detection method for power industry substation monitoring according to claim 1, characterized in that, In step S5, the risk score is graded and control commands are output, and warnings are issued using the following method: Risk score classification: Let the risk score sequence be... ,in Indicates time step The risk value, the risk score With preset threshold , , Comparison of risk level classifications: （11） in, and The settings are based on historical data statistics and expert experience; Control decisions: based on risk level With event type The module executes the decision function. Output control commands : （12） Control commands include: on-site light warnings, voice prompts, equipment shutdown, and remote notification operations; Alarm triggered: The system responds to control commands. Activate the alarm actuator and simultaneously transmit event information. The alarm is recorded in the log database and uploaded to the dispatch center in real time via a network interface. The alarm trigger function can be represented as follows: （13） in This indicates that an alarm has been triggered. This indicates that it has not been triggered.

9. An adaptive sensing and detection system for monitoring power plant substations, characterized in that, include: The system includes a data acquisition and preprocessing module, a processing and feature enhancement module, an adaptive perception and detection module, an abnormal state identification and risk assessment module, and a system control and alarm response module. The data acquisition and preprocessing module is used to acquire and preprocess on-site data; The processing and feature enhancement module is used to further process and enhance the features of the collected field data; The adaptive perception detection module is used to perform small target detection and bounding box regression on the multi-scale features output by the preprocessing and feature enhancement modules; The abnormal state identification and risk assessment module is used to analyze the target detection results output by the adaptive perception and detection module to determine the risk level of power industry station operation behavior, equipment status and safety events; The system control and alarm response module is used to process the risk level and event information output by the abnormal state identification and risk assessment module, and to provide real-time intervention and alarm management for the safety of power industry stations.

Images