Operating personnel detection system for power grid operation safety

Abstract

The invention relates to the technical field of power grid operation safety detection, in particular to an operator detection system for power grid operation safety. According to the operator detection system for the power grid operation safety, the target position, the voltage approach value and the wearing state are jointly calculated by obtaining operator positioning, electrified equipment space distribution, operation voltage and operation image data; multi-dimensional perception of power grid operation risks is achieved, and the defects of an existing system in the aspects of perception dimensionality and accuracy are effectively overcome. Furthermore, the multi-source features are vectorized, a state feature sequence is constructed, the dynamic relation between the personnel and the high-voltage equipment is captured, and the problem that modeling of time sequence changes is difficult to achieve through a traditional rule method is solved. A Transform model containing a multi-head self-attention mechanism is introduced to carry out risk modeling and calculation on a state sequence, a risk type with confidence is output, accurate identification of a high-voltage approaching risk and an illegal wearing risk is realized, an early warning result is output, and the problems of false alarm, missing alarm and response delay in an existing system are solved.

Description

Technical Field

[0001] This invention relates to the field of power grid operation safety detection technology, and more particularly to a personnel detection system for power grid operation safety. Background Technology

[0002] In power system construction, power grid operation and maintenance involve numerous high-risk scenarios such as live-line work and high-altitude work, which place higher demands on the personal safety of workers. Especially during on-site operations such as substation and power line inspection and maintenance, if workers approach high-voltage live parts or fail to wear the required safety protective equipment, serious safety accidents such as electric shock and personal injury can easily occur.

[0003] Most existing power grid operation safety monitoring methods rely on manual inspections or data collection using single sensors, making it difficult to achieve multi-dimensional information fusion and intelligent identification. For example, relying on GPS positioning and electronic fences to determine whether workers have crossed boundaries cannot assess the voltage risk changes during their approach to energized equipment. At the same time, existing risk assessment methods use static rule matching, which cannot effectively model the temporal trend of personnel location changes or the coupling relationship between voltage parameters of energized equipment and spatial location, leading to delayed risk warnings or false alarms and missed warnings. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a personnel detection system for power grid operation safety.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A worker detection system for power grid operation safety includes: The data acquisition module is used to acquire the location data of the workers, the spatial distribution data of the live equipment, the operating voltage of the live equipment, and the image data of the working environment, and to calculate the target location, voltage proximity feature value, and personnel wearing feature data respectively. The feature processing module is used to perform vectorized encoding processing on the target position, voltage proximity feature value and wearing feature data, and combine them in time sequence to obtain the state feature sequence; The risk identification module is used to perform temporal dependency modeling and risk association calculation based on the state feature sequence through a Transformer model containing a multi-head self-attention mechanism and a feedforward network, so as to obtain risk identification results including risk type and confidence level. The early warning output module is used to construct structured early warning data based on the risk identification results and execute early warning output.

[0006] Furthermore, acquiring the operator's location data, electric field strength data, and image data includes the following steps: Positioning data is collected by a positioning terminal worn by the worker, and the positioning data is processed by time synchronization and coordinate transformation to obtain the target location; Based on the target location and the preset spatial distribution data of energized equipment, the spatial distance from the operator to the target energized body is calculated, and the voltage attenuation is calculated in combination with the operating voltage of the energized equipment to obtain the voltage proximity characteristic value. By acquiring image data of the working environment through camera components installed in the power grid working environment, and performing protective equipment detection and category discrimination processing on the image data of the working environment, the characteristic data of personnel wearing protective equipment is obtained.

[0007] Furthermore, performing protective equipment detection and category determination processing on the image data includes the following steps: Based on the collected image data of the work environment, image preprocessing and personnel target extraction processing are performed on the image data to obtain image segments containing the area of ​​the workers; Based on the image fragments, the protective equipment target detection and category recognition processing are performed through an image recognition model to obtain the recognition results of the protective equipment in each part; Based on the identification results and preset protection standards, the wearing status of the workers is determined by rules to obtain personnel wearing characteristic data.

[0008] Furthermore, the image recognition model includes a convolutional neural network model.

[0009] Furthermore, the feature processing module is used to perform the following steps: Based on the target location, the three-dimensional coordinates are normalized by the boundary range of the work area, and mapped to a fixed-dimensional position vector by a position embedding table; Based on the voltage proximity feature value, interval normalization is performed through the safety distance grading rule, and then converted into a proximity risk vector through a linear embedding function; Based on the personnel wearing feature data, category encoding is performed based on a preset category dictionary and mapped to an equipment state vector through an embedding matrix; Time synchronization and concatenation are performed on the position vector, proximity risk vector and equipment state vector to obtain a fused state feature vector; Based on the fused state feature vectors of continuous time steps, sequence segments are constructed according to a preset time window and step size order to obtain the state feature sequence.

[0010] Furthermore, the risk identification module is used to perform the following steps: A learnable location encoding vector is added to each time step based on the state feature sequence to obtain the location-aware feature sequence. The position-aware feature sequence is input into the encoder of a Transformer model containing a multi-head self-attention mechanism and a residual connection structure, and temporal dependency modeling and context feature extraction are performed to obtain a fused representation vector. Based on the fused representation vector, the corresponding risk type and confidence level are calculated through a feedforward network and a Softmax output layer to obtain the risk identification result.

[0011] Furthermore, the Transformer model is trained through the following steps: A training sample set is constructed based on historical operation records. The training sample set includes state feature sequences and target confidence values ​​of corresponding risk labels. Based on the state feature sequence in the training sample set, a position encoding vector is added to the state feature sequence and input into the initialized Transformer model to obtain the risk type and confidence level of the predicted output. Based on the risk confidence level of the predicted output and the target confidence level of the risk label, the loss value is calculated using the mean squared error loss function; Based on the loss value, gradient backpropagation and parameter update are performed on the multi-head attention weights and feedforward network parameters in the Transformer model to obtain a converged risk identification model.

[0012] Furthermore, the mean squared error loss function is as follows: ; in, This is the loss value; This represents the number of training samples; Let be the target confidence value for the i-th training sample; The risk confidence value is the output of the Transformer model for the i-th sample.

[0013] Furthermore, the structured early warning data includes risk type, trigger time, operator identification, and spatial location information.

[0014] Furthermore, the early warning output includes synchronizing structured early warning data to the management personnel's terminal and the operator's mobile terminal, and broadcasting it via audio.

[0015] The beneficial effects of this invention are as follows: By acquiring the location data of workers, the spatial distribution and operating voltage of live equipment, and image data of the work environment, this invention jointly calculates the target location, voltage proximity feature values, and personnel wearing feature data, effectively achieving multi-dimensional perception of risk sources at power operation sites. This overcomes the shortcomings of existing systems that rely on single sensors and cannot simultaneously consider spatial distance, voltage gradient, and protective clothing status. By performing vectorization processing on the above-mentioned multi-source feature data and constructing a state feature sequence in combination with time sequence, the system can express the dynamic approach process and evolution trend between personnel and high-voltage equipment during operation, solving the problems of traditional rule engines being unable to model temporal changes and lacking contextual understanding. Furthermore, by using a Transformer structure that includes a multi-head self-attention mechanism and a feedforward network, deep modeling and risk correlation calculation are performed on the state feature sequence, which can capture the potential coupling relationship between multiple factors such as location, voltage, and clothing, and output identification results including risk type and confidence level, achieving accurate identification of high-voltage proximity risks and illegal operation risks. Finally, based on the identification results, early warning information is output, enabling timely intervention and reminders for high-risk behaviors, improving the efficiency of on-site safety management, and solving the problems of false alarms, missed alarms, and response delays caused by traditional reliance on manual monitoring and static rules. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of a personnel detection system for power grid operation safety according to the present invention.

[0017] Figure 2 This is a flowchart of the training steps of the Transformer model in this invention. Detailed Implementation

[0018] Please see Figures 1-2 As shown, the present invention relates to a worker detection system for power grid operation safety, comprising: The data acquisition module is used to acquire the location data of the workers, the spatial distribution data of the live equipment, the operating voltage of the live equipment, and the image data of the working environment, and to calculate the target location, voltage proximity feature value, and personnel wearing feature data respectively. The feature processing module is used to perform vectorized encoding processing on the target position, voltage proximity feature value and wearing feature data, and combine them in time sequence to obtain the state feature sequence; The risk identification module is used to perform temporal dependency modeling and risk association calculation based on the state feature sequence through a Transformer model containing a multi-head self-attention mechanism and a feedforward network, so as to obtain risk identification results including risk type and confidence level. The early warning output module is used to construct structured early warning data based on the risk identification results and execute early warning output.

[0019] In some embodiments, the system is deployed in a substation live-line maintenance scenario. Before the operation begins, personnel wearing positioning-enabled terminal devices enter the work area. The system acquires the personnel's location data in real time and spatially correlates it with pre-modeled spatial distribution data of live-line equipment. Simultaneously, the system accesses the operating voltage parameters of live-line equipment recorded in the substation equipment ledger for subsequent risk estimation. During the operation, on-site cameras continuously collect images of the work environment to identify the personnel's protective equipment wearing status. Unlike existing solutions that rely solely on single positioning or video monitoring, this system does not directly rely on sensor thresholds for judgment. Instead, it integrates positioning, voltage, and visual information to calculate the voltage proximity of the personnel's current spatial location, forming a voltage proximity feature value. Simultaneously, an image recognition model analyzes whether the personnel are properly wearing safety helmets, insulating gloves, and other protective equipment, obtaining personnel wearing feature data. Furthermore, instead of using traditional rule engines or static threshold judgment methods, the system performs unified vectorized encoding processing on the target location, voltage proximity feature value, and wearing feature data, constructing a state feature sequence in chronological order. This state feature sequence not only reflects the risk status of workers at a single moment but also fully preserves the process information of personnel movement paths, voltage risk changes, and the evolution of protective status over time. The state feature sequence is input into a Transformer model containing a multi-head self-attention mechanism and a feedforward network structure to model the feature correlations between different time steps. Through the multi-head self-attention mechanism, location information, voltage proximity trends, and protective gear status at different time periods can be adaptively assigned different weights, thereby uncovering potential coupled risk patterns between multiple factors, such as high-risk behaviors caused by continuous proximity to high-voltage equipment without adequate protection. Compared to existing methods relying on human experience or recurrent neural networks, this modeling approach has significant advantages in long-term time-series dependency representation and multi-feature fusion. Finally, based on the risk type and corresponding confidence level output by the model, the identification results are structurally expressed, and early warning data containing risk level, trigger time, worker identification, and spatial location information is generated, thereby enabling timely reminders and interventions for high-risk work behaviors.

[0020] Furthermore, acquiring the operator's location data, electric field strength data, and image data includes the following steps: Positioning data is collected by a positioning terminal worn by the worker, and the positioning data is processed by time synchronization and coordinate transformation to obtain the target location; Based on the target location and the preset spatial distribution data of energized equipment, the spatial distance from the operator to the target energized body is calculated, and the voltage attenuation is calculated in combination with the operating voltage of the energized equipment to obtain the voltage proximity characteristic value. By acquiring image data of the working environment through camera components installed in the power grid working environment, and performing protective equipment detection and category discrimination processing on the image data of the working environment, the characteristic data of personnel wearing protective equipment is obtained.

[0021] In some embodiments, before entering the power grid work area, operators wear terminal devices capable of location and timestamp acquisition. These terminals continuously acquire the operators' raw location data at a preset sampling period. To ensure consistency of multi-source data in subsequent processing, the location data is first synchronized using a unified time reference to align location points from different sampling times. Then, based on the spatial coordinate system used in the work area, the location data undergoes coordinate transformation, for example, from a geographic coordinate system to a three-dimensional Cartesian coordinate system referenced to the work area, thus obtaining the target location for subsequent spatial calculations. This processing method avoids spatial distance calculation errors caused by inconsistent coordinate systems. After obtaining the target location, the spatial distance between the operators and each target energized body is estimated using pre-constructed spatial distribution data of energized equipment. This estimation process calculates the Euclidean distance or shortest spatial distance based on the target location and the geometric center or key feature points of the equipment in the energized equipment spatial distribution model to obtain the relative spatial relationship between the operators and the energized bodies. Building upon this, the operating voltage parameters of the energized equipment are further introduced. Based on a pre-defined voltage attenuation model, the voltage attenuation is calculated for the spatial distance, thus mapping the purely geometric distance to a voltage proximity characteristic value with electrical safety significance. Simultaneously, image data of the working environment is continuously collected by camera components deployed in the power grid operation environment. Preprocessing operations are performed on the image data, including image scale normalization, illumination correction, and noise suppression, to improve the stability of subsequent recognition. After preprocessing, protective equipment detection and category discrimination are performed on the image data. Specifically, an image recognition model analyzes the personnel area and performs target detection and category recognition for key protective equipment such as safety helmets and insulating gloves, outputting corresponding wearing status labels.

[0022] Furthermore, performing protective equipment detection and category determination processing on the image data includes the following steps: Based on the collected image data of the work environment, image preprocessing and personnel target extraction processing are performed on the image data to obtain image segments containing the area of ​​the workers; Based on the image fragments, the protective equipment target detection and category recognition processing are performed through an image recognition model to obtain the recognition results of the protective equipment in each part; Based on the identification results and preset protection standards, the wearing status of the workers is determined by rules to obtain personnel wearing characteristic data.

[0023] In some embodiments, camera components in the work environment continuously acquire image data of the work environment, including personnel, at a fixed frame rate. Considering the characteristics of power grid work sites, such as large variations in lighting, complex backgrounds, and diverse personnel postures, image preprocessing is first performed on the acquired raw image data. This includes scale normalization to unify the input resolution, brightness and contrast correction to reduce interference from strong light or shadows, and noise suppression through filtering, thereby improving the stability and robustness of subsequent recognition algorithms. After preprocessing, personnel target extraction is performed on the image data. A target detection algorithm locates the image region corresponding to the personnel from the complex background, obtaining image segments containing the personnel region. This allows subsequent analysis to focus on areas directly related to safety assessment. After obtaining the image segments, they are input into a pre-trained image recognition model to perform protective equipment target detection and category recognition processing on key parts of the personnel. The image recognition model is built on a deep learning structure and can simultaneously complete the spatial localization and category discrimination of protective equipment in a single forward inference process. For example, it can distinguish different equipment types such as safety helmets, insulating gloves, and protective clothing, and output the corresponding detection confidence and category labels. This method transforms raw visual information into structured protective equipment identification results, enabling the wearing status to be expressed in an algorithm-processable data format. Compared to traditional methods based on manual inspection or simple color and shape matching, this identification method has stronger adaptability in complex work scenarios. After obtaining the identification results of protective equipment for each part, the wearing status of the worker is further determined by rules based on preset protection standards. The protection standards are based on work specifications and define the required combination of protective equipment for different types of work. For example, in live-line work scenarios, both safety helmets and insulating gloves must be worn simultaneously. By matching the identification results with the protection standards, it is determined whether the current worker's protective equipment is complete, missing, or improperly worn, and corresponding personnel wearing feature data is generated accordingly. This wearing feature data is expressed in a structured form and can be directly used in subsequent multi-feature fusion and risk identification calculations.

[0024] Furthermore, the image recognition model includes a convolutional neural network model.

[0025] In one specific embodiment, the image recognition model for protective equipment detection and category discrimination is constructed using a convolutional neural network (CNN). The CNN takes a pre-processed image fragment of the worker as input and extracts spatial feature information from the image through multi-layer convolution operations. The convolutional layers utilize local receptive fields and weight sharing mechanisms to encode the texture features and edge structures of protective equipment such as safety helmets and insulating gloves at different scales and perspectives, thereby reducing the interference of complex backgrounds and changes in personnel posture on the recognition results. During the convolutional feature extraction process, the multi-channel feature maps output by the intermediate layers of the network are used to characterize the differences in spatial location and shape of different protective equipment. By introducing downsampling and feature aggregation operations into the CNN, efficient representation of key areas of the protective equipment is achieved, enabling the model to maintain recognition accuracy while possessing good computational efficiency. Furthermore, the high-level features of the convolutional neural network are mapped to specific protective equipment category labels and corresponding confidence values ​​through a fully connected or detector head structure. This is used to characterize the identification results of protective equipment in various parts of the body. In practical applications, the convolutional neural network model is trained using sample data containing different work scenarios, different lighting conditions, and different protective equipment wearing states, enabling the model to distinguish between various states such as complete wearing, incomplete wearing, and improper wearing. By matching the identification results output by the convolutional neural network with preset protection standards, personnel wearing feature data reflecting the compliance level of workers' protection can be further generated.

[0026] Furthermore, the feature processing module is used to perform the following steps: Based on the target location, the three-dimensional coordinates are normalized by the boundary range of the work area, and mapped to a fixed-dimensional position vector by a position embedding table; Based on the voltage proximity feature value, interval normalization is performed through the safety distance grading rule, and then converted into a proximity risk vector through a linear embedding function; Based on the personnel wearing feature data, category encoding is performed based on a preset category dictionary and mapped to an equipment state vector through an embedding matrix; Time synchronization and concatenation are performed on the position vector, proximity risk vector and equipment state vector to obtain a fused state feature vector; Based on the fused state feature vectors of continuous time steps, sequence segments are constructed according to a preset time window and step size order to obtain the state feature sequence.

[0027] It should be noted that after acquiring the target location, voltage proximity feature value, and personnel wearing characteristic data, this data is not directly used for risk assessment. Instead, it undergoes unified feature processing and representation transformation to eliminate differences in units, value ranges, and semantic levels among different data sources. For the target location data, the three-dimensional spatial coordinates of the workers are normalized based on the boundary parameters of the work area, mapping them to a unified numerical range to avoid interference from scale differences in different work areas on subsequent model training. Based on this, a location embedding table is introduced to look up and map the normalized three-dimensional coordinates, transforming spatial location information into fixed-dimensional location vectors. This allows spatial location to participate in subsequent feature fusion and temporal modeling in high-dimensional vector form, rather than being limited to geometric distance calculations. For the voltage proximity feature value, considering the nonlinear differences in safety risks of equipment at different voltage levels, continuous values ​​are not directly used as model input. Instead, based on preset safety distance grading rules, the voltage proximity feature value is normalized to several risk ranges with clear safety semantics. After interval normalization, the interval normalization result is converted into a proximity risk vector using a linear embedding function, enabling electrical risk information to be expressed in a learnable vector form, thereby enhancing the model's ability to distinguish different voltage risk levels. For personnel wearing feature data, since it is essentially discrete category information, the wearing status can be categorized based on a pre-defined protective equipment category dictionary, and the discrete encoding can be mapped to an equipment status vector through an embedding matrix. Through the learning process of the embedding matrix, the similarity and differences between different protective equipment states can be reflected in the vector space, thus avoiding the problems of high dimensionality and weak semantics in traditional one-hot encoding, allowing the wearing status to participate in multi-feature fusion as a continuous feature. After constructing the location vector, proximity risk vector, and equipment status vector, time synchronization processing is performed on the three types of vectors, using a unified timestamp as the alignment benchmark to ensure accurate correspondence of multi-source features within the same time step. Subsequently, the aligned multi-class feature vectors are spliced ​​in a pre-defined order to form a fused status feature vector, enabling a unified expression of the spatial location, voltage risk, and protective status information of the worker within a single time step. Finally, based on the fused state feature vectors of continuous time steps, sequence segments are constructed according to the preset time window length and sliding step size to form a state feature sequence.

[0028] Furthermore, the risk identification module is used to perform the following steps: A learnable location encoding vector is added to each time step based on the state feature sequence to obtain the location-aware feature sequence. The position-aware feature sequence is input into the encoder of a Transformer model containing a multi-head self-attention mechanism and a residual connection structure, and temporal dependency modeling and context feature extraction are performed to obtain a fused representation vector. Based on the fused representation vector, the corresponding risk type and confidence level are calculated through a feedforward network and a Softmax output layer to obtain the risk identification result.

[0029] In some embodiments, a learnable position encoding mechanism is first introduced for the state feature sequence. Specifically, a set of trainable position encoding vectors is superimposed on the fused state feature vector corresponding to each time step in the sequence, enabling the model to explicitly perceive temporal sequence information while maintaining the original feature semantics, thereby distinguishing the relative positions of different time steps in the sequence. This position encoding is not generated by a fixed function, but is optimized along with the model parameters during training, enabling it to adaptively match the temporal characteristics such as personnel movement speed and voltage change rhythm in power grid operation scenarios, avoiding the problem of insufficient adaptability of traditional fixed position encoding to different operation rhythms. After the position encoding is superimposed, a position-aware feature sequence is formed, providing a temporally semantic input representation for subsequent temporal dependency modeling. Subsequently, the position-aware feature sequence is input into a Transformer encoder containing a multi-head self-attention mechanism and a residual connection structure to perform feature modeling. During the multi-head self-attention calculation process, each attention head performs weighted calculation on the correlation between different time steps in the sequence, enabling the model to simultaneously pay attention to the behavioral changes of the operator over multiple time spans, such as continuous approach to high-voltage live conductors, rapid displacement in a short period of time, or sudden changes in the state of protective equipment. Through multi-head parallel computation, the impact of different risk factors at different time scales can be modeled separately, thereby enhancing the model's ability to express complex risk patterns. Simultaneously, a residual connection structure is introduced, allowing the output of each layer to superimpose contextual modeling results while retaining the original feature information. This effectively alleviates the gradient decay problem during deep network training, ensuring that temporal features are not excessively weakened during multi-layer propagation. After multi-layer encoding, a fusion representation vector comprehensively reflects personnel location evolution, voltage risk accumulation, and changes in protection status is obtained. After obtaining the fusion representation vector, a feedforward network performs nonlinear mapping and dimensional transformation on the vector, compressing the high-dimensional temporal representation into the feature space required for risk determination. At the output of the feedforward network, a Softmax output layer is further introduced to perform normalized probability calculations on the scores corresponding to various preset risk types, obtaining the prediction results of the risk types and their corresponding confidence values. The confidence value not only reflects the relative probability that the current operational state belongs to a certain risk type but can also be used for subsequent early warning threshold setting and risk level classification, thereby achieving a quantitative expression of risk identification results.

[0030] Furthermore, the Transformer model is trained through the following steps: A training sample set is constructed based on historical operation records. The training sample set includes state feature sequences and target confidence values ​​of corresponding risk labels. Based on the state feature sequence in the training sample set, a position encoding vector is added to the state feature sequence and input into the initialized Transformer model to obtain the risk type and confidence level of the predicted output. Based on the risk confidence level of the predicted output and the target confidence level of the risk label, the loss value is calculated using the mean squared error loss function; Based on the loss value, gradient backpropagation and parameter update are performed on the multi-head attention weights and feedforward network parameters in the Transformer model to obtain a converged risk identification model.

[0031] In some embodiments, a training sample set for risk learning is first constructed based on historical power grid operation records. These historical operation records originate from actual substation maintenance, line inspections, and other operational processes. For each continuous operational activity, its corresponding state feature sequence is extracted according to a preset time window. Combined with safety accident records, expert evaluation results, or post-event risk review conclusions, a continuous numerical target confidence value is assigned to each state feature sequence to characterize the objective degree of danger of the operational activity in the overall risk level. By introducing a target confidence value instead of a single category label, the training samples can simultaneously reflect subtle differences between different risk levels, providing more continuous and fine-grained supervisory information for the model to learn changes in risk strength. During the model training phase, the state feature sequences in the training sample set are used as input data. Learnable positional encoding vectors are superimposed on the feature vectors at each time step at the input, allowing the model to simultaneously learn temporal sequence information and risk evolution patterns during training. The positionally encoded state feature sequences are input into the initialized Transformer model, sequentially undergoing multi-head self-attention calculation and feedforward network mapping, outputting the risk type prediction result and risk confidence prediction value corresponding to the sequence. In this model, risk type is used to constrain the model to distinguish different risk forms at the semantic level, while risk confidence is used as the core regression objective to characterize the model's ability to quantitatively judge the overall risk level. Further, based on the predicted risk confidence and the target confidence values ​​labeled in the training samples, a loss value is calculated using the mean squared error loss function. This loss function uses the numerical deviation between the predicted and target confidence as the optimization objective, prompting the model to continuously reduce the error in estimating the risk level during training. This allows the model to focus more on the continuous changing trend of risk strength, rather than just whether the risk category is hit. This training method significantly differs from the binary or multi-class classification logic centered on "whether it crosses the boundary" or "whether it violates regulations" in existing technologies, and is more suitable for the actual characteristics of gradually accumulating and dynamically changing risks in power grid operation scenarios. Based on the loss value, gradient backpropagation and parameter updates are performed on the multi-head attention weight parameters in the Transformer model and the connection weights in the feedforward network, enabling the model to gradually learn the contribution relationship between different time steps and different risk factors in the state feature sequence to the final risk confidence. As the number of training rounds increases, when the loss value tends to stabilize and converge, a risk identification model that accurately reflects the evolution of risks in power grid operations is obtained. Through the above training process, the model acquires the ability to assess the risk intensity of complex operational behaviors.

[0032] Furthermore, the mean squared error loss function is as follows: ; in, This is the loss value; This represents the number of training samples; Let be the target confidence value for the i-th training sample; The risk confidence value is the output of the Transformer model for the i-th sample.

[0033] Specifically, in the calculation process, for each training sample, the difference between the predicted risk confidence value and the target confidence value is first calculated, and this difference is squared to amplify the impact of larger prediction biases on the overall loss, thereby prompting the model to pay more attention to high-risk misclassified samples during training. Subsequently, the squared errors corresponding to all training samples are summed and averaged to obtain the loss value used for parameter optimization. Through this method, the gradient update direction of the model during the backpropagation stage is directly directed to reduce the risk confidence prediction error, which helps the model gradually approach the true risk distribution. By introducing this loss function, the model can learn the differences in risk levels corresponding to different state feature sequences, thus outputting smoother, more stable, and more discriminative risk confidence results in subsequent online identification processes.

[0034] Furthermore, the structured early warning data includes risk type, trigger time, operator identification, and spatial location information.

[0035] Specifically, the risk type is determined by the risk category result output by the Transformer model, used to characterize the risk attribute corresponding to the current work state, such as high-voltage approach risk, lack of protective equipment risk, or combined risk. The trigger time is determined by the timestamp when the risk confidence first exceeds a preset threshold. This timestamp comes from the time step index corresponding to the state feature sequence, thus accurately reflecting the starting moment of the risk occurrence or evolution. The operator identification is provided by the unique personnel identification information pre-bound in the positioning terminal or work management system, enabling the early warning result to be accurately associated with the specific operator, avoiding misassociation problems in multi-person work scenarios. The spatial location information is obtained by direct mapping from the target location data. The target location is the actual spatial coordinates of the operator after coordinate transformation and normalization, used to characterize the specific spatial position of the operator relative to the live equipment when the risk is triggered. Through the above structured approach, the risk identification result, which was originally only used for internal model judgment, is transformed into an early warning data unit with clear semantics and engineering usability, enabling subsequent operations such as alarm push, trajectory backtracking, responsibility analysis, and safety assessment to be processed based on a unified data format.

[0036] Furthermore, the early warning output includes synchronizing structured early warning data to the management personnel's terminal and the operator's mobile terminal, and broadcasting it via audio.

[0037] In some embodiments, structured early warning data is first mapped to data channels corresponding to the management terminal and the operator's mobile terminal based on the worker's identifier and the binding relationship between the worker's identifier and the work task. The early warning data synchronized to the management terminal is presented in a complete structured form, including risk type, trigger time, spatial location information, and risk confidence level, to support dispatching and command, risk assessment, and post-event traceability analysis. The early warning data synchronized to the operator's mobile terminal is lightweighted according to on-site response needs, focusing on retaining the risk type and immediate location information to reduce information load and improve response speed. After the early warning is pushed out, an audio broadcast process is automatically triggered based on the risk type and confidence level, converting the early warning information into corresponding voice prompts and broadcasting them in real time on the operator's mobile terminal. The triggering condition for the audio broadcast is jointly controlled by the risk confidence level and a preset threshold, ensuring that even in scenarios where the operator's line of sight is limited, there is high noise, or their attention is focused on the task, they can still perceive the risk promptly through hearing. Through this method, the risk results identified by the algorithm are directly converted into safety prompts that can be perceived by the operator in real time.

[0038] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A personnel detection system for power grid operation safety, characterized in that, include: The data acquisition module is used to acquire the location data of the workers, the spatial distribution data of the live equipment, the operating voltage of the live equipment, and the image data of the working environment, and to calculate the target location, voltage proximity feature value, and personnel wearing feature data respectively. The feature processing module is used to perform vectorized encoding processing on the target position, voltage proximity feature value and wearing feature data, and combine them in time sequence to obtain the state feature sequence; The risk identification module is used to perform temporal dependency modeling and risk association calculation based on the state feature sequence through a Transformer model containing a multi-head self-attention mechanism and a feedforward network, so as to obtain risk identification results including risk type and confidence level. The early warning output module is used to construct structured early warning data based on the risk identification results and execute early warning output.

2. The personnel detection system for power grid operation safety according to claim 1, characterized in that, The acquisition of the operator's location data, electric field strength data, and image data includes the following steps: Positioning data is collected by a positioning terminal worn by the worker, and the positioning data is processed by time synchronization and coordinate transformation to obtain the target location; Based on the target location and the preset spatial distribution data of energized equipment, the spatial distance from the operator to the target energized body is calculated, and the voltage attenuation is calculated in combination with the operating voltage of the energized equipment to obtain the voltage proximity characteristic value. By acquiring image data of the working environment through camera components installed in the power grid working environment, and performing protective equipment detection and category discrimination processing on the image data of the working environment, the characteristic data of personnel wearing protective equipment is obtained.

3. The personnel detection system for power grid operation safety according to claim 2, characterized in that, Performing protective equipment detection and category determination processing on the image data includes the following steps: Based on the collected image data of the work environment, image preprocessing and personnel target extraction processing are performed on the image data to obtain image segments containing the area of ​​the workers; Based on the image fragments, the protective equipment target detection and category recognition processing are performed through an image recognition model to obtain the recognition results of the protective equipment in each part; Based on the identification results and preset protection standards, the wearing status of the workers is determined by rules to obtain personnel wearing characteristic data.

4. The personnel detection system for power grid operation safety according to claim 3, characterized in that, The image recognition model includes a convolutional neural network model.

5. A personnel detection system for power grid operation safety according to claim 3, characterized in that, The feature processing module is used to perform the following steps: Based on the target location, the three-dimensional coordinates are normalized by the boundary range of the work area, and mapped to a fixed-dimensional position vector by a position embedding table; Based on the voltage proximity feature value, interval normalization is performed through the safety distance grading rule, and then converted into a proximity risk vector through a linear embedding function; Based on the personnel wearing feature data, category encoding is performed based on a preset category dictionary and mapped to an equipment state vector through an embedding matrix; Time synchronization and concatenation are performed on the position vector, proximity risk vector and equipment state vector to obtain a fused state feature vector; Based on the fused state feature vectors of continuous time steps, sequence segments are constructed according to a preset time window and step size order to obtain the state feature sequence.

6. The personnel detection system for power grid operation safety according to claim 1, characterized in that, The risk identification module is used to perform the following steps: A learnable location encoding vector is added to each time step based on the state feature sequence to obtain the location-aware feature sequence. The position-aware feature sequence is input into the encoder of a Transformer model containing a multi-head self-attention mechanism and a residual connection structure, and temporal dependency modeling and context feature extraction are performed to obtain a fused representation vector. Based on the fused representation vector, the corresponding risk type and confidence level are calculated through a feedforward network and a Softmax output layer to obtain the risk identification result.

7. A personnel detection system for power grid operation safety according to claim 6, characterized in that, The Transformer model is trained through the following steps: A training sample set is constructed based on historical operation records. The training sample set includes state feature sequences and target confidence values ​​of corresponding risk labels. Based on the state feature sequence in the training sample set, a position encoding vector is added to the state feature sequence and input into the initialized Transformer model to obtain the risk type and confidence level of the predicted output. Based on the risk confidence level of the predicted output and the target confidence level of the risk label, the loss value is calculated using the mean squared error loss function; Based on the loss value, gradient backpropagation and parameter update are performed on the multi-head attention weights and feedforward network parameters in the Transformer model to obtain a converged risk identification model.

8. A personnel detection system for power grid operation safety according to claim 7, characterized in that, The mean squared error loss function is as follows: ； in, This is the loss value; This represents the number of training samples; Let be the target confidence value for the i-th training sample; The risk confidence value is the output of the Transformer model for the i-th sample.

9. A personnel detection system for power grid operation safety according to claim 1, characterized in that, The structured early warning data includes risk type, trigger time, operator identification, and spatial location information.

10. A personnel detection system for power grid operation safety according to claim 1, characterized in that, The early warning output includes synchronizing structured early warning data to the management personnel's terminal and the operator's mobile terminal, and broadcasting it via audio.

Images