A personnel abnormal data early warning method and system

Abstract

The present application relates to the field of pattern recognition, in particular to a personnel abnormal data early warning method and system, personnel behavior data is collected through multiple sensors and preprocessed, wherein the data enhancement step generates pseudo abnormal behavior samples using a reverse generation network; a space-time graph structure is constructed according to the processed data, and a global space-time graph neural network is used to train the space-time graph structure, while combining a hierarchical attention mechanism to learn deep features, obtain fused space-time graph features and generate abnormal classification; an abnormal detection model based on density clustering is used to cluster analyze the behavior data, whether there is abnormal behavior is judged by calculating the abnormal score of each data point, and the parameters are dynamically optimized using a Bayesian model; personnel abnormal data is early warned based on abnormal classification and abnormal score. The present application realizes personnel abnormal data early warning through clustering analysis of personnel behavior data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of pattern recognition, in particular to a personnel abnormal data early warning method and system. BACKGROUND

[0002] It is crucial to perform abnormal early warning on personnel behavior in a monitoring area. Traditional personnel behavior analysis methods usually rely on preset rules or simple statistical models, which are difficult to cope with the complexity and variability of behavior patterns, especially in the face of subtle or newly emerging abnormal behaviors, the accuracy and timeliness of early warning are generally insufficient. These methods often lack comprehensive consideration of the deep spatial correlation and temporal sequence of behavior data, resulting in high false positive rate and false negative rate. In addition, in practical applications, it is extremely difficult to obtain abnormal behavior samples and the number is small, which brings severe challenges to the training of abnormal detection models based on supervised learning. Existing data enhancement techniques have limitations in generating high-quality and representative abnormal samples, resulting in poor model generalization ability and inability to effectively identify unseen abnormal patterns. Therefore, there is an urgent need for a method that can overcome the above technical obstacles, has stronger feature learning ability and data robustness, and realizes accurate and real-time early warning of personnel abnormal behavior, providing a more scientific and reliable basis for safety management in monitoring areas and other scenarios.

[0003] To this end, a personnel abnormal data early warning method and system are proposed. SUMMARY

[0004] The purpose of the present application is to provide a personnel abnormal data early warning method and system, which realizes personnel abnormal data early warning through cluster analysis of personnel behavior data.

[0005] To achieve the above purpose, the present application provides the following technical solutions:

[0006] A personnel abnormal data early warning method, comprising:

[0007] Collecting personnel behavior data through multiple sensors; preprocessing the personnel behavior data, the preprocessing step including denoising, data standardization and data enhancement, generating standard behavior data; the data enhancement generates pseudo abnormal behavior samples through a reverse generation network;

[0008] Constructing a spatio-temporal graph structure according to the standard behavior data, and training the spatial correlation and temporal sequence in the spatio-temporal graph structure using a global spatio-temporal graph neural network to obtain fused spatio-temporal graph features; combining a hierarchical attention mechanism for deep feature learning, the hierarchical attention mechanism including a local spatio-temporal feature extraction layer and a global self-attention aggregation layer, generating an abnormal classification;

[0009] The abnormal detection model based on density clustering is used to perform clustering analysis on the standard behavior data, the abnormal score of each data point is calculated to determine whether there is abnormal behavior, and the parameters are dynamically optimized through the Bayesian model;

[0010] Abnormal classification and abnormal score are used for early warning.

[0011] The personnel behavior data includes personnel identity, timestamp, location data, acceleration data and temperature;

[0012] The personnel behavior data is preprocessed, and the preprocessing step includes denoising, data standardization and data enhancement to generate standard behavior data.

[0013] The specific process of data enhancement is as follows:

[0014] A reverse generation network is built, which receives normal behavior samples as input, and inversely generates the change amount leading to abnormality by learning the internal mode and characteristics of normal behavior;

[0015] Based on the change amount generated by the reverse generation network, a generator network is used to generate pseudo-abnormal samples, and data enhancement is performed in combination with normal behavior samples and abnormal behavior samples.

[0016] The specific process of constructing the spatio-temporal graph structure is as follows:

[0017] The nodes in the spatio-temporal graph structure are the behavior data of a person at a certain time;

[0018] The edges in the spatio-temporal graph structure represent the dependency relationship between nodes, including spatial correlation and temporal dependency; spatial correlation is determined by coordinates, if two nodes are less than a certain threshold in space, a edge is connected in the spatio-temporal graph structure, representing the spatial dependency between nodes; temporal connectivity is that the behavior data of two nodes represents the behavior of a person at adjacent time points, and a time-related edge is established.

[0019] The global spatio-temporal graph neural network includes a graph convolution layer, a time convolution layer, a spatio-temporal fusion mechanism and a global pooling layer;

[0020] The graph convolution layer uses a graph convolution network to process the spatial dependency relationship in the spatio-temporal graph structure and extract spatial features; the time convolution layer processes the temporal dependency relationship in the spatio-temporal graph structure and extracts time series features; the spatio-temporal fusion mechanism fuses spatial features and time features to generate spatio-temporal fusion features; the global pooling layer performs global pooling on the fusion features to generate overall feature representation; the spatio-temporal graph neural network is trained based on personnel behavior data to obtain fusion spatio-temporal graph features and generate abnormal classification.

[0021] The hierarchical attention mechanism performs a deep feature learning process as follows:

[0022] The hierarchical attention mechanism includes a local spatio-temporal feature extraction layer and a global self-attention aggregation layer.

[0023] The local spatio-temporal feature extraction layer learns the local features in the spatio-temporal graph structure by the self-attention mechanism, and the global self-attention aggregation layer globally aggregates all spatio-temporal graph features, and captures important dependency relationships at different time and space positions by the self-attention mechanism.

[0024] The hierarchical attention mechanism performs a deep feature learning process as follows:

[0025] The hierarchical attention mechanism includes a local spatio-temporal feature extraction layer and a global self-attention aggregation layer.

[0026] For the clustering cluster with a density less than the threshold value, the data points in the cluster are determined as abnormal behaviors, and the abnormal scores are obtained by calculating the local density deviation of each data point.

[0027] By setting a threshold value, the determination standard of abnormal behaviors is determined; and the parameters are dynamically optimized by a Bayesian model.

[0028] A personnel abnormal data early warning system comprises:

[0029] A data acquisition module acquires personnel behavior data through a plurality of sensors.

[0030] A data processing module pre-processes the personnel behavior data, including denoising, data standardization and enhancement, to generate standard behavior data; wherein the data enhancement generates abnormal behavior samples through a generative adversarial network for generative adversarial training.

[0031] A feature extraction module constructs a spatio-temporal graph structure according to the standard behavior data, and trains the spatial correlation and temporal sequence in the spatio-temporal graph structure by using a global spatio-temporal graph neural network to obtain fused spatio-temporal graph features; a hierarchical attention mechanism is combined for deep feature learning, the hierarchical attention mechanism includes a local spatio-temporal feature extraction layer and a global self-attention aggregation layer, and an abnormal classification is generated.

[0032] An abnormal early warning module uses a density clustering-based anomaly detection model to cluster analyze the standard behavior data, calculates the abnormal scores of each data point, determines whether there is an abnormal behavior, and dynamically optimizes the parameters by a Bayesian model; and performs early warning based on the abnormal classification and the abnormal scores.

[0033] Compared with the prior art, the present application has the following advantages:

[0034] 1、The application extracts rich spatio-temporal features from complex personnel behavior data by combining reverse generation network, global spatio-temporal graph neural network and hierarchical attention mechanism, ensuring accurate identification of abnormal behavior; the generative adversarial training generates abnormal behavior samples in the data augmentation process, so that the model can still perform high precision when processing minority class abnormal behavior; in addition, by learning the internal changes of normal behavior patterns, the model can better capture potential abnormal patterns in different scenarios, thereby effectively reducing false positives and false negatives.

[0035] 2、The application combines the density clustering-based anomaly detection model with the Bayesian dynamic adjustment mechanism, which can perform clustering analysis on behavior data in real time and dynamically adjust the warning standards of abnormal behavior; as the personnel behavior patterns change, the Bayesian dynamic adjustment can optimize the parameters of the detection model, ensuring adaptability and accuracy in various environments; the dynamic adjustment mechanism can continuously optimize the warning threshold according to new data, always maintaining high accuracy, timely discovering abnormal behavior and making warnings, avoiding the limitations of traditional models when facing unknown behaviors.

[0036] 3、By constructing a spatio-temporal graph structure and combining graph convolution network and time convolution network, the spatial and temporal correlation can be considered simultaneously, and the spatio-temporal features of personnel behavior can be deeply mined; the spatio-temporal graph structure establishes spatial and temporal dependency relationships for each data point, enabling the model to capture the dynamic changes of behavior in all directions when identifying complex behavior patterns; the spatio-temporal feature fusion mechanism organically combines spatial and temporal features, improving the comprehensiveness and accuracy of anomaly detection; this feature fusion method can enhance the system's adaptability to different scenarios and different time periods of abnormal behavior, effectively improving the detection accuracy and flexibility of the system. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 A personnel abnormal data warning method flowchart of the application;

[0038] Figure 2 A spatio-temporal graph structure construction flowchart of the application;

[0039] Figure 3 A personnel abnormal data warning system structure diagram of the application. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0041] Embodiment one:

[0042] The present application provides a personnel abnormal data early warning method, the method flow as shown in Figure 1 The method comprises the following steps:

[0043] Collecting personnel behavior data through multiple sensors; preprocessing the personnel behavior data, the preprocessing steps including denoising, data standardization and data enhancement, generating standard behavior data; the data enhancement generates pseudo abnormal behavior samples through a reverse generation network;

[0044] According to the standard behavior data, a space-time graph structure is constructed, and a global space-time graph neural network is used to train the correlation in space and the sequence in time in the space-time graph structure, to obtain fused space-time graph features; deep feature learning is performed in combination with a hierarchical attention mechanism, the hierarchical attention mechanism including a local space-time feature extraction layer and a global self-attention aggregation layer, to generate an abnormal classification;

[0045] According to the fused space-time graph features, a density clustering-based anomaly detection model is used to perform clustering analysis on the behavior data, to calculate the abnormal score of each data point, to determine whether there is an abnormal behavior, and to dynamically optimize the parameters through a Bayesian model;

[0046] Based on the abnormal classification and the abnormal score, an early warning is performed.

[0047] Specifically, the personnel behavior data includes personnel identity, timestamp, location data, acceleration data and temperature;

[0048] The personnel identity is the identity information of each person entering the monitoring area, including the number and the name, obtained through identity recognition (RFID tag or fingerprint recognition);

[0049] The timestamp records the time when the behavior occurs;

[0050] The location data is the location coordinates of the personnel in the monitoring area, obtained by a positioning sensor;

[0051] The acceleration data is the motion state of the personnel, including X, Y and Z direction acceleration data, measured by an accelerometer;

[0052] The temperature data records the environmental temperature and the personnel temperature;

[0053] The personnel behavior data is preprocessed, the preprocessing steps including denoising, data standardization and data enhancement, to generate standard behavior data;

[0054] The denoising is to eliminate noise in the original data, and the steps include: using a low-pass filter to remove high-frequency noise; smoothing the position data to remove unnecessary jitter; filtering the acceleration data to eliminate short-term vibrations and device errors; and normalizing through Z-score standardization so that data from different sensors can be compared on the same scale.

[0055] The data augmentation generates abnormal behavior samples through a reverse generation network, performs generative adversarial training, generates more abnormal behavior samples, and balances the data distribution.

[0056] By preprocessing the personnel behavior data, including denoising, standardization and data augmentation, the accuracy and robustness of abnormal behavior detection can be significantly improved. The denoising step effectively eliminates high-frequency noise, jitter and errors in the original data, ensuring the clarity and reliability of the data, and providing high-quality input data for subsequent analysis; the standardization process eliminates the dimensional differences between different sensors, so that all behavior data can be compared on a unified scale, further improving the effectiveness of model training.

[0057] The data augmentation specific process is:

[0058] A reverse generation network is built to receive normal behavior samples as input, and to generate changes that lead to abnormalities by learning the internal patterns and features of normal behavior; the normal behavior samples include the behavior of monitored personnel, including the identity, timestamp, position data, acceleration data and temperature of the personnel;

[0059] The reverse generation network learns the internal rules of normal behavior data and establishes a spatiotemporal model of normal behavior. The model understands the characteristics of these behaviors in time and space, such as the smooth change of position when a prisoner walks, the stable fluctuation of acceleration, and the regular fluctuation of temperature data; during the generation process, the reverse generation network not only disturbs single features such as acceleration, but also multi-dimensions the behavior characteristics in time and space, generates potential abnormal behaviors by introducing temporal mutations and spatial position changes, these disturbances can simulate complex abnormal behaviors, generate more abnormal behavior samples that conform to actual application scenarios, enhance the model's ability to recognize complex and unpredictable behavior patterns, such as the behavior of "sudden pause and rapid acceleration", the generator can generate this kind of highly dynamic pseudo-abnormal behavior samples by disturbing the acceleration data and introducing temporal mutations.

[0060] The generator generates pseudo abnormal behavior samples according to the time and space disturbances generated by the reverse generation network; the pseudo abnormal behavior samples simulate the abnormal behavior of the monitoring personnel, such as sudden rapid walking, rapid movement after pausing, and other behavior patterns. The generator generates pseudo abnormal samples by simultaneously disturbing the time series and spatial coordinates, so that the generated pseudo abnormal samples not only have abnormality in behavior content, but also reflect suddenness and irregularity in the time and place of behavior occurrence.

[0061] The pseudo abnormal behavior samples generated by the reverse generation network and the generator can significantly enhance the diversity and complexity of the abnormal behavior data set; by multi-dimensionally disturbing the spatio-temporal features of normal behavior data, including time mutation and spatial position change, the generated abnormal behavior samples are more consistent with the complex behavior patterns in actual application scenarios, and also improve the recognition ability of the model in the face of unknown or uncommon abnormal behavior, especially in monitoring systems. This generated abnormal sample can effectively solve the problem of data scarcity, help the model improve the accuracy, robustness and generalization ability of abnormal detection, and thus provide more accurate abnormal early warning and real-time monitoring.

[0062] Referring to Figure 2 , the specific process of constructing the spatio-temporal graph structure is as follows:

[0063] The nodes in the spatio-temporal graph structure are the behavior features of the monitoring personnel at each time point, including personnel identity, time stamp, position data, acceleration data and temperature data;

[0064] The edges in the spatio-temporal graph structure represent the dependency relationship between nodes, including spatial correlation and temporal dependency;

[0065] The spatial correlation is used to judge the relative position relationship of two nodes in space; if the positions of the personnel represented by two behavior data points are close enough in space, it means that there is a spatial dependency relationship between the two behavior data points. The specific process is to calculate the spatial distance between each two nodes, assuming that the spatial coordinates are and The spatial distance between two nodes is calculated by the Euclidean distance formula: If the calculated spatial distance is less than a set threshold (for example, the threshold is 2 meters), a spatial correlation edge is established between the two nodes, indicating that there is a mutual dependency relationship between the two nodes.

[0066] The time connectivity is used to determine whether two nodes occur at adjacent time points in time, by calculating the time difference of the two nodes, that is, the difference of their timestamps, if the time difference of the two nodes is less than a set threshold (for example, within 5 seconds), a time-related edge is established between the two nodes, indicating that the two behavior data points are adjacent in time and may constitute a continuous behavior sequence; for example, a person changes from a standing state to a walking state in a short time, and there is a time dependence between these behaviors.

[0067] In addition to the spatial and temporal dependence, behavior similarity is introduced as a new edge connecting nodes, which measures the similarity of the behavior features between nodes to determine whether they should be connected by an edge; the cosine similarity is used to measure the similarity of two behavior data points in a multi-dimensional feature space, if the similarity is greater than a threshold, it means that the two nodes show similar behavior patterns and may belong to the same behavior sequence or part of a behavior pattern, therefore a connection edge should be established between the two nodes, by introducing behavior similarity as the connection edge between nodes, the potential similar patterns between behavior data points can be captured more accurately, even if they are not adjacent in space or time.

[0068] All nodes are connected according to spatial distance and time relationship to form a space-time graph structure.

[0069] By constructing the space-time graph structure, the changes and continuity of personnel behavior in space and time are effectively reflected, by introducing behavior similarity, even if two nodes are not adjacent in time or space, the potential similar behavior patterns between them can still be identified, significantly improving the accuracy of the model in complex scenarios, especially when facing irregular, complex or sudden abnormal behaviors, more efficient and accurate abnormal behavior detection and early warning can be provided.

[0070] The global space-time graph neural network includes a graph convolution layer, a time convolution layer, a space-time fusion mechanism and a global pooling layer;

[0071] The graph convolution layer uses a graph convolution network to process the spatial dependence relationship in the space-time graph structure and extract spatial features; the graph convolution network performs weighted summation on the features of adjacent nodes to identify the spatial dependence relationship between nodes, for example, if the behavior of a person has spatial similarity (such as behavior consistency at adjacent positions), the graph convolution layer will extract this spatial pattern and integrate it into the feature representation of the node;

[0072] The time convolution layer focuses on processing the temporal dependence in the spatio-temporal graph, extracting the time series features of each node. Through the time convolution operation, the time nodes and time persistence of behavior changes are identified, for example, if the state transition of personnel from "standing" to "walking" occurs in a short time, the time convolution layer can capture the dynamic changes on these time series, thereby helping the system to identify and understand the time dimension of the behavior pattern;

[0073] The spatio-temporal fusion mechanism fuses the spatial features and the temporal features to generate spatio-temporal fusion features, for example, the rapid acceleration of personnel walking, combined with the features of the time and space dimensions, helps to accurately identify abnormal behaviors;

[0074] The global pooling layer pools the spatio-temporal fusion features, usually using global max pooling or global average pooling to extract the detailed information of each node in the graph into global features, reducing redundant information, and gathering all the node features in the spatio-temporal graph into a fixed-length feature vector to generate an overall feature representation;

[0075] The spatio-temporal graph neural network is trained based on personnel behavior data to obtain the fusion spatio-temporal graph features.

[0076] Specifically, the global spatio-temporal graph neural network includes 2 layers of graph convolution layers (64 hidden units per layer), 3 layers of time convolution layers (using 1D convolution with a kernel size of 3), using the Adam optimizer, the learning rate is set to 0.001, the batch size is 32, and a total of 200 cycles are trained; After the global pooling layer, the fusion spatio-temporal graph features are input to a fully connected classifier, which is trained by a binary cross-entropy loss function, and finally outputs the classification result of the behavior belonging to "normal" or "abnormal".

[0077] The spatial dependence of personnel behavior is effectively modeled through the graph convolution layer, which can identify the spatial correlation between individuals and individuals, and individuals and the environment. At the same time, the time dynamics of behavior are accurately captured through the time convolution layer, which can deeply understand the evolution process of the behavior state. The spatio-temporal fusion mechanism deeply fuses these two key dimension information, realizes the comprehensive description of complex behavior patterns, and effectively solves the problem of low recognition accuracy caused by ignoring the spatio-temporal coupling of traditional methods.

[0078] The specific process of deep feature learning of the hierarchical attention mechanism is as follows:

[0079] The hierarchical attention mechanism includes a local spatio-temporal feature extraction layer and a global self-attention aggregation layer.

[0080] The local spatio-temporal feature extraction layer learns the local features in the spatio-temporal graph structure by a self-attention mechanism; specifically, the query vector, the key vector and the value vector of each node are calculated by the self-attention mechanism, the dependence relationship weight between each node and other nodes is obtained by calculating the correlation (usually using dot product to calculate the similarity) between the query vector and the key vector. Then the value vector is weighted and summed to obtain the local spatio-temporal feature representation; the features of each node are weighted and updated according to the dependence relationship between them, so as to obtain the deep learning representation of the local spatio-temporal feature;

[0081] After the local spatio-temporal feature extraction layer, the global self-attention aggregation layer is entered to globally weight and aggregate the features of all spatio-temporal graphs. The key of this process is to capture important dependence relationships at different time and space positions. First, global self-attention calculation is performed on all nodes in the entire spatio-temporal graph. Similar to the local feature extraction layer, each node generates a query, key and value vector; then the global attention weight, i.e. the relationship between each node and other nodes in the global, is calculated. Through these relationships, the model can capture which nodes (time points, space positions) have an important influence on other nodes in the entire spatio-temporal graph; finally, the features of all nodes are weighted and aggregated by the global self-attention mechanism to obtain a global feature representation, which contains the mutual dependence relationship between nodes in the entire spatio-temporal graph.

[0082] After the local spatio-temporal feature extraction layer and the global self-attention aggregation layer, the model will obtain a deep feature representation that integrates local details and global relationships;

[0083] Deep feature learning through hierarchical attention mechanism can significantly improve the feature representation ability of spatio-temporal graph data. The local spatio-temporal feature extraction layer effectively captures the relationship between each node and its neighborhood through the self-attention mechanism, and mines local detail information; while the global self-attention aggregation layer helps the model identify and integrate important dependence relationships in the entire spatio-temporal graph through global weighted aggregation.

[0084] The density clustering-based anomaly detection model calculates the density distribution of each data point by clustering analysis on standard behavior data; specifically, two key parameters, neighborhood radius and minimum point number, are used to calculate the number of points contained in the neighborhood within the radius for each data point. If the number of points in the neighborhood is greater than or equal to the threshold, the point is considered as a core point belonging to a certain cluster. If the number of points in the neighborhood of a point is less than the threshold, the point is considered as a noise point or an abnormal point, and an abnormal score is obtained by calculating the local density deviation of each data point;

[0085] The local density deviation calculation method is to assume that the data point There are multiple points in the neighborhood of each data point, and the density of each point is calculated The density of the data point is compared with the density of other points in its neighborhood If the density of the data point is significantly lower than the density of other points in its neighborhood, it means that it deviates greatly from the normal behavior pattern and represents abnormal behavior

[0086] By calculating the local density deviation, the model can identify points that deviate greatly from surrounding data points, i.e., these points may represent abnormal behavior, and the greater the deviation, the stronger the abnormality of the point.

[0087] According to the performance of historical data and current data, the Bayesian adjustment mechanism can dynamically optimize parameters such as neighborhood radius, minimum point number, and threshold

[0088] For each data point, an anomaly score is calculated based on its local density deviation; first, a local density index is calculated for each data point, which considers the distance between it and its nearest neighbors; then, the local density of the data point is compared with the average local density of all other points in its neighborhood; if a point's local density is significantly lower than the average density of surrounding points, it means that it is in a relatively sparse area and deviates greatly from the normal behavior pattern; the degree of this deviation, i.e., the local density deviation, is directly used as the anomaly score of the point; therefore, the higher the score, the more likely the data point is abnormal behavior.

[0089] The density clustering-based anomaly detection model can effectively identify abnormal behavior that deviates significantly from the normal behavior pattern, especially when dealing with high-dimensional and complex data. Through the calculation of local density deviation, the model can accurately identify potential abnormal points, and the Bayesian adjustment mechanism dynamically optimizes key parameters to ensure that the detection accuracy continuously improves as the data characteristics change. This not only improves the accuracy of anomaly detection, but also adapts to various environments and data distribution changes, greatly enhancing the adaptability and reliability of the system in practical applications

[0090] Based on anomaly classification and anomaly scoring, if a data point is determined to be abnormal and the anomaly score exceeds the set threshold, the point is considered to be a serious anomaly, and the system will trigger a high-priority warning; for abnormal behavior with a lower anomaly score, the system will judge it as a minor anomaly and trigger a low-priority warning for subsequent observation.

[0091] By multi-sensor data collection and preprocessing, combined with spatio-temporal graph neural network and hierarchical attention mechanism, the spatial correlation and time sequence characteristics of personnel behavior can be accurately captured, and the precision of anomaly detection is improved; data augmentation technology generates pseudo-anomaly samples to enhance the robustness of the model, while the anomaly detection model based on density clustering and the Bayesian dynamic optimization mechanism can adapt to different environmental changes and dynamically adjust key parameters, thereby effectively identifying potential abnormal behavior; the early warning mechanism based on anomaly classification and scoring can quickly respond to various abnormal situations, improving safety and management efficiency.

[0092] Embodiment two:

[0093] The present application provides a personnel anomaly data early warning system, referring to Figure 3 , comprising:

[0094] Data acquisition module: collecting personnel behavior data through multiple sensors;

[0095] Data processing module: preprocessing personnel behavior data, including denoising, data standardization and enhancement, to generate standard behavior data; wherein data enhancement generates abnormal behavior samples through a reverse generation network for generative adversarial training;

[0096] Feature extraction module: constructing a spatio-temporal graph structure according to the standard behavior data, and using a global spatio-temporal graph neural network to train the spatial correlation and temporal sequence in the spatio-temporal graph structure to obtain fused spatio-temporal graph features; combining a hierarchical attention mechanism for deep feature learning, the hierarchical attention mechanism includes a local spatio-temporal feature extraction layer and a global self-attention aggregation layer to generate abnormal classification;

[0097] Anomaly early warning module: based on the fused spatio-temporal graph features, using an anomaly detection model based on density clustering to perform clustering analysis on the behavior data, calculating the anomaly score of each data point to determine whether there is abnormal behavior, and dynamically optimizing the parameters through a Bayesian model; early warning based on abnormal classification and abnormal scoring.

[0098] The data acquisition module collects personnel behavior data through multiple sensors, including personnel identity, timestamp, location data, acceleration data and temperature data; these sensors are deployed in the A monitoring area, and the personnel identity is obtained through RFID tags, the location coordinates are obtained through GPS sensors, the motion state is measured through three-axis accelerometers, and the environment temperature is recorded through temperature sensors; for example, at a certain time, the system collects a group of original data, including personnel identity "A-101", timestamp "15:47:00", location data "(2.5m, 1.8m, 0.2m)", acceleration data "(0.12, 0.08, 0.10)" and temperature data "36.5°C".

[0099] The data processing module pre-processes the collected personnel behavior data, including denoising, data standardization and data augmentation, to generate standard behavior data; the denoising process uses a low-pass filter and smoothing processing to eliminate noise and jitter in the original data; data standardization uses Z-score standardization so that data from different sensors can be compared on the same scale;

[0100] Data augmentation generates pseudo-abnormal behavior samples through a generative adversarial network; the generative adversarial network receives normal behavior samples as input, learns their internal patterns and reversely generates change amounts that lead to abnormalities; then, the generator network generates pseudo-abnormal samples simulating behaviors such as "sudden stop and rapid acceleration" or "rapid sprint" based on these change amounts, which significantly enhances the diversity and complexity of the abnormal behavior data set, effectively solving the problem of scarce abnormal data.

[0101] The feature extraction module constructs a spatio-temporal graph structure based on the standard behavior data, where nodes are the behavior data of personnel at a certain time, and edges represent the dependency between nodes, including spatial correlation and temporal dependence; for example, if the distance between the positions of two nodes is less than 2 meters or the time difference is less than 5 seconds, a connection edge is established between them; then, a global spatio-temporal graph neural network is used to train the spatial correlation and temporal sequence in the spatio-temporal graph structure to obtain fused spatio-temporal graph features; the network includes a graph convolution layer (for extracting spatial features), a temporal convolution layer (for extracting temporal sequence features), and a spatio-temporal fusion mechanism (for fusing the two); at the same time, a hierarchical attention mechanism is used for deep feature learning, which includes a local spatio-temporal feature extraction layer and a global self-attention aggregation layer, which captures important dependency relationships at different times and spatial locations through a self-attention mechanism to generate abnormal classification.

[0102] The anomaly warning module uses a density clustering-based anomaly detection model to perform clustering analysis on the behavior data based on the fused spatio-temporal graph features; the model calculates the density distribution of each data point and judges that data points in a cluster with a density less than a threshold as abnormal behavior; the anomaly score is obtained by calculating the local density deviation of each data point, and the higher the score, the stronger the abnormality of the point; in addition, the Bayesian model dynamically optimizes parameters such as neighborhood radius, minimum number of points and threshold to ensure that detection accuracy improves as data changes; based on anomaly classification and anomaly score, a warning is issued; if a data point is determined to be abnormal and the anomaly score exceeds a set threshold, the system will trigger a high-priority warning; for abnormal behavior with a lower anomaly score, a low-priority warning is triggered for subsequent observation.

[0103] The application can extract rich spatio-temporal features from complex personnel behavior data by combining the reverse generation network, the global spatio-temporal graph neural network and the hierarchical attention mechanism, and ensure accurate identification of abnormal behavior; the abnormal detection model based on density clustering is combined with the Bayesian dynamic adjustment mechanism, which can analyze the clustering of behavior data in real time, and dynamically adjust the early warning standard to adapt to the change of personnel behavior mode, and ensure the adaptability and accuracy in various environments; by constructing the spatio-temporal graph structure and combining the graph convolution network and the time convolution network, the application can simultaneously consider the correlation of space and time, and improve the comprehensiveness and accuracy of abnormal detection.

[0104] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A personnel abnormal data early warning method, characterized in that: include: Collect personnel behavior data through multiple sensors; Preprocessing the personnel behavior data, wherein the preprocessing steps include denoising, data standardization, and data enhancement to generate standard behavior data; The data enhancement generates pseudo abnormal behavior samples through an inverse generation network; A spatiotemporal graph structure is constructed based on the standard behavioral data, and a global spatiotemporal graph neural network is used to train the spatial correlation and temporal dependency in the spatiotemporal graph structure to obtain fused spatiotemporal graph features. Deep feature learning is performed in conjunction with a hierarchical attention mechanism, which includes a local spatiotemporal feature extraction layer and a global self-attention aggregation layer to generate anomaly classifications. Use a density-based anomaly detection model to perform cluster analysis on standard behavior data, calculate the anomaly score for each data point, determine whether there is abnormal behavior, and dynamically optimize parameters through a Bayesian model; Provide early warning based on anomaly classification and anomaly scoring.

2. A personnel abnormal data early warning method according to claim 1, characterized in that: The personnel behavior data includes personnel identity, timestamp, location data, acceleration data and temperature; The preprocessing steps include denoising, data normalization and data enhancement to generate standard behavioral data.

3. A personnel abnormal data early warning method according to claim 2, characterized in that: The specific process of data enhancement is as follows: Build a reverse generation network, receive normal behavior samples as input, and reversely generate the changes that lead to abnormalities by learning the inherent patterns and characteristics of normal behavior; Based on the variation generated by the reverse generation network, pseudo abnormal samples are generated through the generator network, and data enhancement is performed by combining normal behavior samples and abnormal behavior samples.

4. A personnel abnormal data early warning method according to claim 1, characterized in that: The specific process of constructing the spatiotemporal graph structure is as follows: The nodes in the spatiotemporal graph structure are the behavioral data of a person at a certain moment; The edges in the spatiotemporal graph structure represent the dependency relationship between nodes, including spatial correlation and temporal dependence. Spatial correlation is determined by coordinates. If the distance between two nodes is less than a set threshold in space, an edge is connected in the spatiotemporal graph structure to represent the spatial correlation relationship between the nodes. Temporal dependence is established when the human behavior data represented by two nodes occur at adjacent moments in time.

5. The method for early warning of abnormal personnel data according to claim 3, characterized in that: The global spatiotemporal graph neural network includes a graph convolution layer, a temporal convolution layer, a spatiotemporal fusion mechanism and a global pooling layer; The graph convolution layer uses a graph convolution network to process the spatial correlation relationship in the spatiotemporal graph structure and extract spatial features; the temporal convolution layer processes the temporal dependency relationship in the spatiotemporal graph structure and extracts temporal features; The spatiotemporal fusion mechanism fuses spatial features with temporal features to generate spatiotemporal fusion features; the global pooling layer globally pools the fused features to generate an overall feature representation; the global spatiotemporal graph neural network is trained based on personnel behavior data to obtain fused spatiotemporal graph features and generate anomaly classification.

6. The method for early warning of abnormal personnel data according to claim 1, characterized in that: The specific process of deep feature learning by the hierarchical attention mechanism is as follows: The hierarchical attention mechanism includes a local spatiotemporal feature extraction layer and a global self-attention aggregation layer; The local spatiotemporal feature extraction layer performs weighted learning on the local features in the spatiotemporal graph structure through the self-attention mechanism, and the global self-attention aggregation layer performs global weighted aggregation on all spatiotemporal graph features, capturing important dependencies at different time and spatial positions through the self-attention mechanism.

7. The method for early warning of abnormal personnel data according to claim 1, characterized in that: The specific process of clustering analysis of standard behavior data using the density-based anomaly detection model is as follows: The density clustering-based anomaly detection model performs cluster analysis on standard behavior data and calculates the density distribution of each data point; For clusters with density less than the threshold, the data points in the cluster are judged to be abnormal behaviors, and the abnormality score is obtained by calculating the local density deviation of each data point; By setting thresholds, the criteria for determining abnormal behavior are determined; parameters are dynamically optimized through the Bayesian model.

8. A personnel abnormal data early warning system, characterized in that: include: Data collection module: collects personnel behavior data through multiple sensors; Data processing module: pre-processes personnel behavior data, including denoising, data standardization and enhancement, to generate standard behavior data; Data enhancement uses an inverse generative network to generate pseudo-abnormal behavior samples for generative adversarial training. Feature extraction module: This module constructs a spatiotemporal graph structure based on standard behavioral data and uses a global spatiotemporal graph neural network to train the spatial correlation and temporal dependency within the spatiotemporal graph structure to obtain fused spatiotemporal graph features. Deep feature learning is performed using a hierarchical attention mechanism consisting of a local spatiotemporal feature extraction layer and a global self-attention aggregation layer to generate anomaly classifications. Abnormal warning module: Uses a density-based anomaly detection model to perform cluster analysis on standard behavior data, calculates the anomaly score for each data point, determines whether there is abnormal behavior, and dynamically optimizes parameters using a Bayesian model; Provide early warning based on anomaly classification and anomaly scoring.

Images