AI-based unsafe behavior identification method and system

Abstract

The invention discloses an AI-based unsafe behavior identification method and system, and the method comprises the steps: employing a lightweight online modeling method of incremental learning based on the current state information of a worker, and integrating a situation element to dynamically update an individual behavior model; heterogeneous situation data are collected and classified, and key situation features are extracted and dimensionality reduction is carried out; constructing a behavior risk analysis-oriented situational feature system, and performing weight distribution according to the correlation between each situational element and the unsafe behavior to obtain a situational feature weight; in an individual behavior model updating process, according to a change trend of employee behaviors and an identification feedback result, carrying out real-time evaluation on importance of each situation element, dynamically adjusting weights of situation features, and adaptively optimizing a feature fusion strategy; and detecting an abnormal behavior by adopting an abnormal behavior detection method of a self-adaptive threshold to obtain an unsafe behavior identification result. The method can effectively improve the real-time recognition accuracy of unsafe behaviors in the industrial environment, and reduces the false alarm rate.

Description

Technical Field

[0001] The present invention belongs to the field of information technology, and in particular relates to an unsafe behavior identification method and system based on AI. Background Art

[0002] When performing contextualized unsafe behavior identification in industrial scenarios, we face the technical contradiction of how to improve the recognition accuracy while ensuring real-time performance. The complexity and variability of the industrial environment causes employee behavior patterns to be dynamic and uncertain, and traditional rule-based recognition methods are difficult to adapt. Although the introduction of contextual information helps to improve the accuracy of recognition, the acquisition and processing of massive heterogeneous contextual data brings challenges to system latency.

[0003] To resolve this contradiction, the key lies in optimizing the extraction and fusion methods of situational features, and maximizing the situational degree of behavioral risk analysis while meeting real-time requirements. On the one hand, a lightweight online modeling method based on incremental learning is adopted, with the current state of the worker as the core, and the most valuable situational elements for behavioral risk judgment are selectively integrated. The individual behavior model is dynamically updated and improved during the identification process, rather than blindly pursuing the comprehensiveness of situational information. On the other hand, for multi-source heterogeneous situational data, a unified semantic representation and feature dimensionality reduction method is explored to extract the key situational elements that best reflect the behavior pattern, reducing data redundancy and computing overhead. By constructing a situational feature system for behavioral risk analysis, semantic mapping, association calculation and weight allocation are performed on various types of data to form the best match with the behavioral model, and coordinated optimization is performed in the entire architecture to balance the real-time and accuracy of recognition.

[0004] In short, contextual unsafe behavior identification in industrial scenarios requires a balance between modeling granularity and real-time performance according to local conditions, selection of the most critical contextual factors to be included in behavioral risk analysis, and optimization of data organization and calculation methods, so as to timely and accurately identify employees' unsafe behaviors in complex and changing environments, and provide strong support for refined safety management. Summary of the invention

[0005] To achieve the above objectives, the present invention provides an AI-based unsafe behavior identification method and system. Among them, an AI-based unsafe behavior identification method includes:

[0006] Construct an individual behavior model and obtain the current status information of the worker. Based on the current status information of the worker, adopt a lightweight online modeling method of incremental learning, incorporate the most valuable situational elements for behavior risk judgment, and dynamically update the individual behavior model;

[0007] Collect heterogeneous situational data and classify the data, perform semantic mapping and association calculation on various types of data, extract key situational features that best reflect the behavior pattern, and use feature dimensionality reduction technology to reduce the dimension of the key situational features;

[0008] Construct a situational feature system for behavioral risk analysis, assign weights based on the correlation between each situational element and the unsafe behavior, and obtain the situational feature weights; in the process of updating the individual behavior model, perform a real-time evaluation of the importance of each situational element based on the changing trend of employee behavior and the identification feedback results, dynamically adjust the weights of the situational features, and adaptively optimize the feature fusion strategy;

[0009] An abnormal behavior detection method with adaptive threshold is used to detect abnormal behavior, and the normal behavior patterns of employees are continuously learned, and the abnormal behavior judgment threshold is dynamically adjusted to obtain unsafe behavior identification results.

[0010] Preferably, the process of dynamically updating the individual behavior model includes:

[0011] The sensors collect workers’ physiological parameters and behavioral data in real time to obtain workers’ current status information;

[0012] Construct an individual behavior model, extract the most valuable situational elements for risk judgment based on the current status information of the worker, and integrate them into the individual behavior model;

[0013] According to the current status information of the worker, an online incremental learning algorithm is adopted to dynamically update the individual behavior model.

[0014] Preferably, the process of extracting key context features and reducing the dimension of the key context features using feature dimensionality reduction technology includes:

[0015] Acquire a pre-established semantic mapping model and an associated calculation model, input the heterogeneous situational data into the semantic mapping model and the associated calculation model for processing, and extract key situational feature data;

[0016] The key context feature data are subjected to feature dimensionality reduction processing by using a principal component analysis method, by calculating the covariance matrix and eigenvalues ​​of the key context feature data, and determining the feature dimensions to be retained according to a preset dimensionality reduction threshold, thereby obtaining the key context feature data after dimensionality reduction;

[0017] Determine the semantic association relationship between different context data according to the semantic mapping model, and obtain the semantic association features of different context data by constructing a semantic association graph and using a graph embedding algorithm to learn the low-dimensional vector representation of the nodes in the semantic association graph;

[0018] Inputting the reduced-dimensional key context feature data and the semantic association features into a pre-trained behavior pattern recognition model, classifying and predicting the input features through the behavior pattern recognition model, and determining the user behavior pattern type in the current context;

[0019] Obtaining a pre-established behavior-context mapping table, and determining a subset of context data most relevant to the current user behavior by searching the mapping table for a key context corresponding to the current behavior pattern type;

[0020] The context data subset is input into the incremental learning model for online learning to dynamically update the parameters of the incremental learning model. If the computational overhead of the current context data exceeds a preset threshold, the elastic scaling mechanism of system resources is triggered to increase computing nodes.

[0021] Preferably, the process of obtaining the key context feature data after dimension reduction includes:

[0022] Acquire key situational feature data, construct a covariance matrix for the key situational feature data, and perform eigenvalue decomposition on the covariance matrix to obtain an eigenvalue vector;

[0023] According to a preset dimensionality reduction threshold, the eigenvalue vector is subjected to dimensionality reduction processing to determine the feature dimension to be retained;

[0024] For the determined feature dimensions, principal component analysis is performed on the key situational feature data to obtain the key situational feature data after dimensionality reduction;

[0025] Inputting the reduced-dimensional key situational feature data into a classification model for training to obtain a trained classification model;

[0026] The trained classification model is used to classify and predict the target key situational feature data. If the classification result meets the preset conditions, the target key situational feature data is determined as the target classification; according to the target classification result, a corresponding strategy is adopted to process the target object and the processing result is output.

[0027] Preferably, a situational feature system for behavioral risk analysis is constructed, and weights are assigned according to the correlation between each situational element and the unsafe behavior. The process of obtaining the situational feature weights includes:

[0028] According to the pre-established situational feature system, obtain the situational element information of the target behavior;

[0029] Based on the situational factor information, analyzing the correlation between each situational factor and the unsafe behavior through an association rule mining algorithm;

[0030] According to the correlation between the situational factors and unsafe behaviors obtained by mining, the weight coefficient of each situational factor is determined by using the hierarchical analysis method.

[0031] Input the situational factor information and the corresponding weight coefficient into a preset behavior risk assessment model, and calculate the behavior risk value under the situation through the model;

[0032] Determine whether the behavior risk value exceeds a preset risk threshold, and if so, determine the behavior in the situation as a high-risk behavior;

[0033] For situations that are judged to be high-risk, a behavior prediction model is constructed using the random forest algorithm to predict the types of unsafe behaviors that may occur in the situation;

[0034] According to the predicted unsafe behavior type, the corresponding situational feature combination is obtained from the pre-built situation-behavior matching model to form a targeted risk control strategy.

[0035] Preferably, the process of inputting the situational factor information and the corresponding weight coefficient into a preset behavior risk assessment model and obtaining the behavior risk value under the situation through the model calculation includes:

[0036] Obtain at least one situational element information and a weight coefficient corresponding to the situational element information, input the obtained situational element information and the weight coefficient into a preset behavior risk assessment model for model calculation, and obtain a behavior risk value under the situation;

[0037] Determine whether the behavioral risk value exceeds a preset behavioral risk threshold. If the behavioral risk value exceeds the threshold, trigger the preset risk warning information; determine different levels of risk control measures according to the level of the behavioral risk value, and execute corresponding measures.

[0038] Preferably, in the process of updating the individual behavior model, the importance of each situational element is evaluated in real time according to the changing trend of employee behavior and the recognition feedback result, the weight of the situational feature is dynamically adjusted, and the process of adaptively optimizing the feature fusion strategy includes:

[0039] Acquire time series data of employee behavior, analyze the time series data, extract behavior change trend characteristics of the time series data, and determine whether employee behavior has abnormal changes based on the behavior change trend characteristics;

[0040] Obtaining recognition feedback results of the behavior recognition model, and determining the importance of each situational factor to the behavior risk determination based on the recognition feedback results;

[0041] A weighted fusion algorithm is used to take the importance of the situational elements as weights, and each situational feature is weightedly fused to obtain a comprehensive situational feature;

[0042] Inputting the comprehensive situational characteristics into a behavior risk determination model to determine whether the employee's behavior is risky;

[0043] If the behavior risk determination result is inconsistent with the identification feedback result, the weight of the situational feature is dynamically adjusted, the weight parameter is optimized using the gradient descent algorithm, and the feature fusion strategy is updated; then the optimized feature fusion strategy is applied to the behavior risk determination at the next moment to achieve continuous model adaptive optimization.

[0044] Preferably, the abnormal behavior detection method with adaptive threshold is used to detect abnormal behavior, and the normal behavior pattern of employees is continuously learned. The process of dynamically adjusting the abnormal behavior determination threshold includes:

[0045] Obtain employee behavior data in industrial environments, continuously learn and model the normal behavior patterns of employees, and obtain the feature vectors of normal employee behavior;

[0046] According to the employee's normal behavior feature vector, a clustering algorithm is used to cluster the employee's behavior to obtain cluster centers under different behavior modes;

[0047] For each cluster center, the distance between the employee behavior feature vector in the cluster and the cluster center is calculated, and according to the distance distribution, the abnormal behavior determination threshold of the cluster is adaptively determined;

[0048] When it is detected that the distance between a behavior feature vector of an employee and the center of the cluster to which it belongs exceeds the abnormality determination threshold of the cluster, it is determined that the employee has abnormal behavior;

[0049] According to the complexity and variability of the industrial environment, the normal behavior patterns of employees are updated by incremental learning, and the judgment threshold of abnormal behavior is dynamically adjusted accordingly;

[0050] If the system detects abnormal behavior of employees, it will issue a timely warning and take corresponding safety control measures. At the same time, the abnormal behavior data will be used to update the employee behavior model.

[0051] The present invention also provides an unsafe behavior identification system based on AI, comprising:

[0052] A behavior model dynamic update module is used to build an individual behavior model and obtain the current state information of the worker. Based on the current state information of the worker, a lightweight online modeling method of incremental learning is adopted to incorporate the most valuable situational elements for behavior risk judgment and dynamically update the individual behavior model.

[0053] A situational data processing module is used to collect heterogeneous situational data and classify the data, perform semantic mapping and association calculation on various types of data, extract key situational features that best reflect the behavior pattern, and use feature dimensionality reduction technology to reduce the dimension of the key situational features;

[0054] Situational feature system building module, used to establish a feature system for behavioral risk analysis;

[0055] The feature weight dynamic adjustment module is used to assign weights according to the correlation between each situational element and the unsafe behavior to obtain the situational feature weight; and in the process of updating the individual behavior model, the importance of each situational element is evaluated in real time according to the changing trend of employee behavior and the recognition feedback results, the weight of the situational feature is dynamically adjusted, and the feature fusion strategy is adaptively optimized;

[0056] The abnormal behavior detection module is used to detect abnormal behavior using an abnormal behavior detection method with an adaptive threshold, continuously learn the normal behavior patterns of employees, dynamically adjust the abnormal behavior judgment threshold, and obtain unsafe behavior identification results.

[0057] Compared with the prior art, the present invention has the following advantages and technical effects:

[0058] The present invention discloses a real-time behavioral risk identification method in an industrial environment. Aiming at the dynamic changes of workers' behavioral states, a lightweight online modeling method of incremental learning is adopted to dynamically update individual behavior models. By performing semantic mapping and association calculation on massive heterogeneous situational data, key situational features are extracted and dimension reduction is performed, and a situational feature system for behavioral risk analysis is constructed. In the process of model updating, the weights of situational elements are evaluated and adjusted in real time according to the behavior change trend and recognition feedback, and the feature fusion strategy is optimized. An abnormal behavior detection method with adaptive threshold is adopted to dynamically adjust the judgment threshold and improve the sensitivity of the system. The present invention can effectively improve the real-time recognition accuracy of unsafe behaviors in industrial environments, reduce the false alarm rate, and provide strong support for ensuring the safety of workers. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0060] Figure 1 The figure is a schematic diagram of a method flow of an embodiment of the present invention. DETAILED DESCRIPTION

[0061] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0062] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0063] like Figure 1 As shown, this embodiment provides an unsafe behavior identification method based on AI, including:

[0064] Build an individual behavior model and obtain the current status information of the workers. Based on the current status information of the workers, adopt a lightweight online modeling method of incremental learning, incorporate the most valuable situational elements for behavioral risk judgment, and dynamically update the individual behavior model;

[0065] Collect heterogeneous situational data and classify them, perform semantic mapping and association calculation on various types of data, extract key situational features that best reflect behavior patterns, and use feature dimensionality reduction technology to reduce the dimensionality of key situational features;

[0066] Build a situational feature system for behavioral risk analysis, assign weights based on the correlation between each situational element and unsafe behavior, and obtain situational feature weights; in the process of updating the individual behavior model, evaluate the importance of each situational element in real time based on the changing trend of employee behavior and identification feedback results, dynamically adjust the weight of situational features, and adaptively optimize feature fusion strategies;

[0067] An abnormal behavior detection method with adaptive threshold is used to detect abnormal behavior, and the normal behavior patterns of employees are continuously learned, and the abnormal behavior judgment threshold is dynamically adjusted to obtain unsafe behavior identification results.

[0068] Furthermore, the process of dynamically updating the individual behavior model includes:

[0069] The sensors collect workers’ physiological parameters and behavioral data in real time to obtain workers’ current status information;

[0070] Build an individual behavior model, extract the most valuable situational factors for risk judgment based on the current status information of workers, and integrate them into the individual behavior model;

[0071] According to the current status information of the workers, an online incremental learning algorithm is used to dynamically update the individual behavior model.

[0072] This embodiment improves the real-time and accuracy of identifying worker behavior risks by continuously improving the individual behavior model. Through a lightweight online learning algorithm and an incremental update mechanism, it ensures that the individual behavior model can adapt to changes in worker behavior patterns.

[0073] More specifically, the smart bracelets worn by workers collect physiological parameters such as heart rate and blood pressure in real time, and capture the workers' behavior through cameras. The system uses convolutional neural networks to analyze video frames, extract the posture features of workers, and uses hidden Markov models to identify workers' behaviors in combination with inertial sensor data. According to the changes in the workers' status, an adaptive weight adjustment algorithm is used to dynamically update the weight coefficients of each feature in the individual behavior model. For example, when the worker's fatigue state is detected, the weight of the physiological parameters is increased; when the worker enters a high-risk operation area, the weight of the location information is increased. The system mines the 10 most valuable situational factors for risk judgment from historical data, such as night work and high-altitude work, and selectively integrates them into the individual behavior model. When the risk identification accuracy of the individual behavior model is lower than 90% for 5 consecutive days, the system automatically triggers model optimization and fine-tunes the model using newly collected data through incremental learning. If the risk behavior of the worker is identified to last for more than 10 seconds, an early warning message is immediately sent to the on-site management personnel. The system uses a lightweight online random forest algorithm. Every time 100 new data are collected, incremental learning is performed. The entire process takes no more than 5 seconds, ensuring the real-time update of the individual behavior model. At the same time, the system will dynamically adjust the sampling frequency of the sensor according to the recognition results of the model. When workers are working at high risk, the frequency of data collection is increased to up to 10 times per second. When workers are working normally, the sampling frequency is reduced to once per minute to save equipment power and storage space.

[0074] Furthermore, the process of extracting key context features and reducing the dimension of the key context features using feature dimensionality reduction technology includes:

[0075] Obtaining a pre-established semantic mapping model and an associated computing model, inputting heterogeneous context data into the semantic mapping model and the associated computing model for processing, and extracting key context feature data;

[0076] The principal component analysis method is used to perform feature dimensionality reduction processing on the key situational feature data. By calculating the covariance matrix and eigenvalues ​​of the key situational feature data and determining the feature dimensions to be retained according to the preset dimensionality reduction threshold, the key situational feature data after dimensionality reduction is obtained;

[0077] Determine the semantic association relationship between different context data according to the semantic mapping model, and obtain the semantic association features of different context data by constructing a semantic association graph and using a graph embedding algorithm to learn the low-dimensional vector representation of nodes in the semantic association graph;

[0078] The reduced-dimensional key context feature data and semantic association features are input into the pre-trained behavior pattern recognition model, and the input features are classified and predicted by the behavior pattern recognition model to determine the user behavior pattern type in the current context;

[0079] Obtain a pre-established behavior-context mapping table, and determine a subset of context data most relevant to the current user behavior by searching the mapping table for a key context corresponding to the current behavior pattern type;

[0080] The context data subset is input into the incremental learning model for online learning to dynamically update the parameters of the incremental learning model. If the computing overhead of the current context data exceeds the preset threshold, the elastic scaling mechanism of system resources is triggered to increase computing nodes.

[0081] More specifically, a semantic mapping model containing 1,000 nodes and an association calculation model based on a convolutional neural network are first pre-established. When a large amount of heterogeneous situational data is input, the data is mapped to the semantic space through the semantic mapping model, and the correlation between different data is calculated using the association calculation model, and finally 5,000 key situational features are extracted. Then, the principal component analysis method is used to reduce the dimension of these 5,000 features. The eigenvalues ​​are obtained by calculating the covariance matrix of the feature data. According to the preset dimensionality reduction threshold of 9, the first three principal components are retained, and the key situational feature data is reduced to a 3-dimensional space. Then, a semantic association graph containing 800 nodes is constructed using the semantic mapping model, and the Node2Vec algorithm is used to learn the 64-dimensional vector representation of the nodes to obtain the semantic association features between different situational data. The reduced-dimensional features and semantic association features are input into a pre-trained LSTM-based behavior pattern recognition model to classify the user's behavior pattern with an accuracy rate of 92%. According to the pre-established behavior-context mapping table, the 50 most relevant context data for the current user behavior are found and input into an incremental learning model based on online gradient descent for real-time optimization, which increases the accuracy of behavior pattern recognition to 95%. The current computing overhead of context data is 800 GFLOPS, which exceeds the preset threshold of 500 GFLOPS. Therefore, the elastic scaling mechanism is triggered, and 2 computing nodes are dynamically added, reducing the average latency from 8 seconds to less than 3 seconds.

[0082] Furthermore, the process of obtaining the key context feature data after dimensionality reduction includes:

[0083] Obtain key situational feature data, construct a covariance matrix for the key situational feature data, perform eigenvalue decomposition on the covariance matrix, and obtain an eigenvalue vector;

[0084] According to the preset dimensionality reduction threshold, the eigenvalue vector is reduced in dimension to determine the feature dimension that needs to be retained;

[0085] According to the determined feature dimensions, principal component analysis is performed on the key situational feature data to obtain the key situational feature data after dimensionality reduction;

[0086] Inputting the key situational feature data after dimension reduction into the classification model for training to obtain a trained classification model;

[0087] Using the trained classification model, the target key situational feature data is classified and predicted. If the classification result meets the preset conditions, the target key situational feature data is determined as the target classification;

[0088] According to the target classification results, the corresponding strategy is adopted to process the target object and the processing results are output.

[0089] More specifically, key context feature data extracted from massive heterogeneous context data are obtained, and a covariance matrix is ​​constructed for these feature data. The dimension of the covariance matrix is ​​the dimension of the feature data. For example, if there are 1000 features, the covariance matrix is ​​a 1000×1000 matrix. Then, the covariance matrix is ​​subjected to eigenvalue decomposition to obtain an eigenvalue vector, and each element in the eigenvalue vector represents the importance of the corresponding feature. According to the preset dimensionality reduction threshold of 95, the eigenvalue vector is subjected to dimensionality reduction processing, and the feature dimensions whose eigenvalue proportion reaches 95% cumulatively are determined as the dimensions to be retained. For example, the dimensions whose cumulative eigenvalue proportion reaches 95% are the first 50 dimensions, and only the first 50 dimensional features are retained. Then, for the determined 50 feature dimensions, principal component analysis is performed on the key context feature data, and the original features are converted into linearly independent principal components through orthogonal transformation to obtain the key context feature data after dimensionality reduction. The feature data after dimensionality reduction is divided into a training set and a test set, of which the training set accounts for 80%, and is input into the SVM classification model for training. The optimal parameters are obtained through grid search to obtain the trained classification model. Using the trained classification model, the key context feature data of the target user is classified and predicted. If the confidence of the classification result exceeds 8, the user is determined as the target classification. According to the target classification result, the corresponding strategy is used to process the target object and the processing result is output. Through the above process, the key context feature data is processed by dimensionality reduction and classification prediction, and the corresponding strategy is adopted according to the prediction result, realizing the automatic identification of target users from massive heterogeneous data.

[0090] Furthermore, a situational feature system for behavioral risk analysis is constructed, and weights are assigned according to the correlation between each situational factor and the unsafe behavior. The process of obtaining the situational feature weights includes:

[0091] According to the pre-established situational feature system, obtain the situational element information of the target behavior;

[0092] Based on the situational factor information, the correlation between each situational factor and unsafe behavior is analyzed through the association rule mining algorithm;

[0093] According to the correlation between the situational factors and unsafe behaviors obtained by mining, the weight coefficient of each situational factor is determined by using the hierarchical analysis method.

[0094] Input the situational factor information and the corresponding weight coefficient into the preset behavior risk assessment model, and calculate the behavior risk value under the situation through the model;

[0095] Determine whether the behavior risk value exceeds the preset risk threshold. If it exceeds the threshold, the behavior in the situation is judged as high-risk behavior;

[0096] For situations that are judged to be high-risk, a behavior prediction model is built using the random forest algorithm to predict the types of unsafe behaviors that may occur in the situation;

[0097] According to the predicted unsafe behavior type, the corresponding situational feature combination is obtained from the pre-built situation-behavior matching model to form a targeted risk control strategy.

[0098] More specifically, first, based on the pre-established multi-dimensional situational feature system including time, location, personnel, equipment, etc., the specific situational element information of the target operator is obtained through sensors, positioning systems, etc. Then, the FP-Growth association rule mining algorithm is used to analyze the correlation between each situational element and unsafe behavior in the massive historical data. For example, through analysis, it is found that in a construction environment with noise greater than 90 decibels, the proportion of behaviors without wearing earmuffs is as high as 75%. Then, the hierarchical analysis method is used to determine the weight coefficients of situational elements such as noise, lighting, and temperature through expert scoring, with a noise weight of 4, a lighting weight of 3, and a temperature weight of 3. After that, the parameters of each element and the weight coefficient of the situation in which the operator is located are input into the behavioral risk assessment model based on BP neural network training, and the behavioral risk value of this situation is calculated to be 85 through the sigmoid activation function. This risk value is compared with the preset 6 risk thresholds. If the threshold is exceeded, the behavior in this situation is judged to be a high-risk behavior. For this high-risk situation, a behavior prediction model was constructed using the random forest algorithm. By integrating 50 decision trees, the top three most likely behaviors were predicted from 10 unsafe behavior types such as not wearing a helmet and not wearing a seat belt based on situational characteristics. Finally, it was found from the pre-trained situation-behavior matching model that in high-altitude work situations with loud noise and dim light, it is easy to fail to wear a safety rope. Based on this, targeted risk control strategies such as wearing shock-absorbing earmuffs and adding lighting equipment were generated and pushed to the operator's terminal in real time.

[0099] Furthermore, the situational factor information and the corresponding weight coefficient are input into a preset behavior risk assessment model, and the process of obtaining the behavior risk value under the situation through the model calculation includes:

[0100] Obtain at least one situational element information and a weight coefficient corresponding to the situational element information, input the obtained situational element information and the weight coefficient into a preset behavior risk assessment model for model calculation, and obtain a behavior risk value under the situation;

[0101] Determine whether the behavioral risk value exceeds the preset behavioral risk threshold. If the behavioral risk value exceeds the threshold, the preset risk warning information is triggered; according to the level of behavioral risk value, determine different levels of risk control measures and implement corresponding measures.

[0102] More specifically, when obtaining the situational element information of the target behavior, the video surveillance system can be used to collect on-site images, and image recognition technology can be used to extract key elements such as personnel location and action. For example, if the operator is identified as being in a dangerous area, the corresponding regional risk weight coefficient is 8. The extracted situational element information and weight coefficient are input into the behavioral risk assessment model built based on the Bayesian network. The model calculates and finds that the behavioral risk value in this situation is 75, which exceeds the preset risk threshold of 6. The system automatically triggers a high-risk prompt and announces through voice, "Attention, there are major safety hazards in the current area, please evacuate as soon as possible." At the same time, according to the high-risk level interval of 7-9 where the risk value is located, the system automatically executes risk control measures including cutting off the power supply of dangerous equipment and blocking dangerous areas, effectively avoiding accidents.

[0103] Furthermore, in the process of updating the individual behavior model, the importance of each situational factor is evaluated in real time according to the changing trend of employee behavior and the recognition feedback results, and the weight of the situational features is dynamically adjusted. The process of adaptively optimizing the feature fusion strategy includes:

[0104] Obtain the time series data of employee behavior, analyze the time series data, extract the behavior change trend characteristics of the time series data, and determine whether the employee behavior has changed abnormally based on the behavior change trend characteristics;

[0105] Obtain the recognition feedback results of the behavior recognition model, and determine the importance of each situational factor to the behavior risk assessment based on the recognition feedback results;

[0106] The weighted fusion algorithm is used to take the importance of situational factors as weights, and each situational feature is weightedly fused to obtain a comprehensive situational feature;

[0107] Input comprehensive situational characteristics into the behavioral risk determination model to determine whether the employee's behavior is risky;

[0108] If the behavioral risk determination result is inconsistent with the recognition feedback result, the weight of the situational feature is dynamically adjusted, the weight parameter is optimized using the gradient descent algorithm, and the feature fusion strategy is updated; then the optimized feature fusion strategy is applied to the behavioral risk determination at the next moment to achieve continuous model adaptive optimization.

[0109] Furthermore, the abnormal behavior detection method with adaptive threshold is used to detect abnormal behavior, and the normal behavior pattern of employees is continuously learned. The process of dynamically adjusting the abnormal behavior determination threshold includes:

[0110] Obtain employee behavior data in industrial environments, continuously learn and model the normal behavior patterns of employees, and obtain the feature vectors of normal employee behavior;

[0111] According to the employee's normal behavior feature vector, a clustering algorithm is used to cluster the employee's behavior and obtain the cluster center under different behavior modes;

[0112] For each cluster center, calculate the distance between the employee behavior feature vector in the cluster and the cluster center, and adaptively determine the abnormal behavior judgment threshold of the cluster based on the distance distribution;

[0113] When it is detected that the distance between an employee's behavior feature vector and the center of the cluster to which it belongs exceeds the abnormality judgment threshold of the cluster, it is judged that the employee has abnormal behavior;

[0114] According to the complexity and variability of the industrial environment, the normal behavior patterns of employees are updated by incremental learning, and the judgment threshold of abnormal behavior is dynamically adjusted accordingly; by dynamically adjusting the abnormal behavior judgment threshold, the system's detection sensitivity to unsafe behaviors of employees is improved, and the probability of normal behavior being misjudged as abnormal is reduced, thereby reducing the system's false alarm rate.

[0115] If the system detects abnormal behavior among employees, it will issue a timely warning and take corresponding safety control measures. At the same time, the abnormal behavior data will be used to update the employee behavior model to continuously improve the accuracy of anomaly detection.

[0116] More specifically, in an industrial environment, sensors are used to collect employee behavior data, including information such as the employee's location, action, and operation. The long short-term memory neural network (LSTM) is used to model the employee's behavior sequence, and the feature vector of the employee's normal behavior is obtained through training. The dimension of the feature vector is 128. The employee behavior feature vector is clustered using the K-means clustering algorithm. The number of clusters is set to 5, and 5 cluster centers are obtained, representing 5 typical employee behavior patterns. For each cluster, the Euclidean distance between the employee behavior feature vector and the cluster center in the cluster is calculated, and the distribution of the distance is statistically analyzed. According to the mean and standard deviation of the distance distribution, the 3-sigma principle is used to determine the abnormal behavior judgment threshold, that is, the distance exceeding the mean plus 3 times the standard deviation is considered abnormal. When the distance between an employee's behavior feature vector and the cluster center to which it belongs exceeds the abnormal judgment threshold, the employee is judged to have abnormal behavior. Considering the changes in the industrial environment, the cluster center is updated every hour using the newly collected employee behavior data, and the abnormal judgment threshold is adjusted accordingly. Through dynamic threshold adjustment, the abnormal judgment threshold is increased by 10% to improve the sensitivity of abnormal detection, while the misjudgment rate of normal behavior is controlled within 5%. When abnormal behavior of employees is detected, the system automatically triggers an early warning, adds abnormal behavior data to the training set, and uses incremental learning to update the employee behavior model in real time and optimize model parameters. After 30 consecutive days of system operation and iterative optimization, the accuracy of abnormal behavior detection has increased from the initial 85% to 95%, effectively improving the intelligent level of industrial safety management and control.

[0117] This embodiment also includes an AI-based unsafe behavior identification system, including:

[0118] The behavior model dynamic update module is used to build individual behavior models and obtain the current status information of workers. Based on the current status information of workers, it adopts a lightweight online modeling method of incremental learning, incorporates the most valuable situational elements for behavior risk judgment, and dynamically updates the individual behavior model.

[0119] The context data processing module is used to collect heterogeneous context data and classify the data, perform semantic mapping and association calculation on various types of data, extract key context features that best reflect the behavior pattern, and use feature dimensionality reduction technology to reduce the dimension of key context features;

[0120] Situational feature system building module, used to establish a feature system for behavioral risk analysis;

[0121] The feature weight dynamic adjustment module is used to assign weights according to the correlation between each situational factor and the unsafe behavior to obtain the situational feature weights. In the process of updating the individual behavior model, the importance of each situational factor is evaluated in real time according to the changing trend of employee behavior and the recognition feedback results, and the weights of situational features are dynamically adjusted to adaptively optimize the feature fusion strategy.

[0122] The abnormal behavior detection module is used to detect abnormal behavior using an abnormal behavior detection method with an adaptive threshold, continuously learn the normal behavior patterns of employees, dynamically adjust the abnormal behavior judgment threshold, and obtain unsafe behavior identification results.

[0123] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. An unsafe behavior identification method based on AI, characterized in that: include: Construct an individual behavior model and obtain the current status information of the worker. Based on the current status information of the worker, adopt a lightweight online modeling method of incremental learning, incorporate the most valuable situational elements for behavior risk judgment, and dynamically update the individual behavior model; Collect heterogeneous situational data and classify the data, perform semantic mapping and association calculation on various types of data, extract key situational features that best reflect the behavior pattern, and use feature dimensionality reduction technology to reduce the dimension of the key situational features; Construct a situational feature system for behavioral risk analysis, assign weights based on the correlation between each situational element and unsafe behavior, and obtain situational feature weights; In the process of updating the individual behavior model, the importance of each situational factor is evaluated in real time according to the changing trend of employee behavior and the recognition feedback results, the weight of the situational feature is dynamically adjusted, and the feature fusion strategy is adaptively optimized; An abnormal behavior detection method with adaptive threshold is used to detect abnormal behavior, and the normal behavior patterns of employees are continuously learned, and the abnormal behavior judgment threshold is dynamically adjusted to obtain unsafe behavior identification results.

2. The AI-based unsafe behavior identification method according to claim 1 is characterized in that: The process of dynamically updating the individual behavior model includes: The sensors collect workers’ physiological parameters and behavioral data in real time to obtain workers’ current status information; Construct an individual behavior model, extract the most valuable situational elements for risk judgment based on the current status information of the worker, and integrate them into the individual behavior model; According to the current status information of the worker, an online incremental learning algorithm is adopted to dynamically update the individual behavior model.

3. The AI-based unsafe behavior identification method according to claim 1, characterized in that: The process of extracting key context features and reducing the dimension of the key context features using feature dimensionality reduction technology includes: Acquire a pre-established semantic mapping model and an associated calculation model, input the heterogeneous situational data into the semantic mapping model and the associated calculation model for processing, and extract key situational feature data; The key context feature data are subjected to feature dimensionality reduction processing by using a principal component analysis method, by calculating the covariance matrix and eigenvalues ​​of the key context feature data, and determining the feature dimensions to be retained according to a preset dimensionality reduction threshold, thereby obtaining the key context feature data after dimensionality reduction; Determine the semantic association relationship between different context data according to the semantic mapping model, and obtain the semantic association features of different context data by constructing a semantic association graph and using a graph embedding algorithm to learn the low-dimensional vector representation of the nodes in the semantic association graph; Inputting the reduced-dimensional key context feature data and the semantic association features into a pre-trained behavior pattern recognition model, classifying and predicting the input features through the behavior pattern recognition model, and determining the user behavior pattern type in the current context; Obtaining a pre-established behavior-context mapping table, and determining a subset of context data most relevant to the current user behavior by searching the mapping table for a key context corresponding to the current behavior pattern type; The context data subset is input into the incremental learning model for online learning to dynamically update the parameters of the incremental learning model. If the computational overhead of the current context data exceeds a preset threshold, the elastic scaling mechanism of system resources is triggered to increase computing nodes.

4. The AI-based unsafe behavior identification method according to claim 3 is characterized in that: The process of obtaining the key situational feature data after dimension reduction includes: Acquire key situational feature data, construct a covariance matrix for the key situational feature data, and perform eigenvalue decomposition on the covariance matrix to obtain an eigenvalue vector; According to a preset dimensionality reduction threshold, the eigenvalue vector is subjected to dimensionality reduction processing to determine the feature dimension to be retained; For the determined feature dimensions, principal component analysis is performed on the key situational feature data to obtain the key situational feature data after dimensionality reduction; Inputting the reduced-dimensional key situational feature data into a classification model for training to obtain a trained classification model; The trained classification model is used to classify and predict the target key situational feature data. If the classification result meets the preset conditions, the target key situational feature data is determined as the target classification; according to the target classification result, a corresponding strategy is adopted to process the target object and the processing result is output.

5. The AI-based unsafe behavior identification method according to claim 1, characterized in that: Construct a situational feature system for behavioral risk analysis, and assign weights based on the correlation between each situational factor and the unsafe behavior. The process of obtaining the situational feature weights includes: According to the pre-established situational feature system, obtain the situational element information of the target behavior; Based on the situational factor information, analyzing the correlation between each situational factor and the unsafe behavior through an association rule mining algorithm; According to the correlation between the situational factors and unsafe behaviors obtained by mining, the weight coefficient of each situational factor is determined by using the hierarchical analysis method. Input the situational factor information and the corresponding weight coefficient into a preset behavior risk assessment model, and calculate the behavior risk value under the situation through the model; Determine whether the behavior risk value exceeds a preset risk threshold, and if so, determine the behavior in the situation as a high-risk behavior; For situations that are judged to be high-risk, a behavior prediction model is constructed using the random forest algorithm to predict the types of unsafe behaviors that may occur in the situation; According to the predicted unsafe behavior type, the corresponding situational feature combination is obtained from the pre-built situation-behavior matching model to form a targeted risk control strategy.

6. The AI-based unsafe behavior identification method according to claim 5, characterized in that: The process of inputting the situational factor information and the corresponding weight coefficient into the preset behavior risk assessment model and calculating the behavior risk value under the situation through the model includes: Obtain at least one situational element information and a weight coefficient corresponding to the situational element information, input the obtained situational element information and the weight coefficient into a preset behavior risk assessment model for model calculation, and obtain a behavior risk value under the situation; Determine whether the behavioral risk value exceeds a preset behavioral risk threshold. If the behavioral risk value exceeds the threshold, trigger the preset risk warning information; determine different levels of risk control measures according to the level of the behavioral risk value, and execute corresponding measures.

7. The AI-based unsafe behavior identification method according to claim 1, characterized in that: In the process of updating the individual behavior model, the importance of each situational factor is evaluated in real time according to the changing trend of employee behavior and the recognition feedback results, the weight of the situational feature is dynamically adjusted, and the process of adaptively optimizing the feature fusion strategy includes: Acquire time series data of employee behavior, analyze the time series data, extract behavior change trend characteristics of the time series data, and determine whether employee behavior has abnormal changes based on the behavior change trend characteristics; Obtaining recognition feedback results of the behavior recognition model, and determining the importance of each situational factor to the behavior risk determination based on the recognition feedback results; A weighted fusion algorithm is used to take the importance of the situational elements as weights, and each situational feature is weightedly fused to obtain a comprehensive situational feature; Inputting the comprehensive situational characteristics into a behavior risk determination model to determine whether the employee's behavior is risky; If the behavior risk determination result is inconsistent with the identification feedback result, the weight of the situational feature is dynamically adjusted, the weight parameter is optimized using the gradient descent algorithm, and the feature fusion strategy is updated; then the optimized feature fusion strategy is applied to the behavior risk determination at the next moment to achieve continuous model adaptive optimization.

8. The AI-based unsafe behavior identification method according to claim 1, characterized in that: The abnormal behavior detection method with adaptive threshold is used to detect abnormal behavior, and the normal behavior patterns of employees are continuously learned. The process of dynamically adjusting the abnormal behavior determination threshold includes: Obtain employee behavior data in industrial environments, continuously learn and model the normal behavior patterns of employees, and obtain the feature vectors of normal employee behavior; According to the employee's normal behavior feature vector, a clustering algorithm is used to cluster the employee's behavior to obtain cluster centers under different behavior modes; For each cluster center, the distance between the employee behavior feature vector in the cluster and the cluster center is calculated, and according to the distance distribution, the abnormal behavior determination threshold of the cluster is adaptively determined; When it is detected that the distance between a behavior feature vector of an employee and the center of the cluster to which it belongs exceeds the abnormality determination threshold of the cluster, it is determined that the employee has abnormal behavior; According to the complexity and variability of the industrial environment, the normal behavior patterns of employees are updated by incremental learning, and the judgment threshold of abnormal behavior is dynamically adjusted accordingly; If the system detects abnormal behavior of employees, it will issue a timely warning and take corresponding safety control measures. At the same time, the abnormal behavior data will be used to update the employee behavior model.

9. An AI-based unsafe behavior identification system, characterized in that: include: A behavior model dynamic update module is used to build an individual behavior model and obtain the current state information of the worker. Based on the current state information of the worker, a lightweight online modeling method of incremental learning is adopted to incorporate the most valuable situational elements for behavior risk judgment and dynamically update the individual behavior model. A situational data processing module is used to collect heterogeneous situational data and classify the data, perform semantic mapping and association calculation on various types of data, extract key situational features that best reflect the behavior pattern, and use feature dimensionality reduction technology to reduce the dimension of the key situational features; Situational feature system building module, used to establish a feature system for behavioral risk analysis; A feature weight dynamic adjustment module is used to assign weights according to the correlation between each situational factor and the unsafe behavior to obtain the situational feature weight; In the process of updating the individual behavior model, the importance of each situational factor is evaluated in real time according to the changing trend of employee behavior and the recognition feedback results, the weight of the situational feature is dynamically adjusted, and the feature fusion strategy is adaptively optimized; The abnormal behavior detection module is used to detect abnormal behavior using an abnormal behavior detection method with an adaptive threshold, continuously learn the normal behavior patterns of employees, dynamically adjust the abnormal behavior judgment threshold, and obtain unsafe behavior identification results.

Images