Safety management system based on artificial intelligence early warning

Abstract

The invention relates to the technical field of network security, in particular to a security management system based on artificial intelligence early warning. The system comprises a data acquisition unit, and the data acquisition unit is used for collecting on-site safety monitoring data in real time, so that on-site information can be obtained more accurately, and the accuracy is improved; the data analysis unit analyzes abnormal data points in the security monitoring data based on a local abnormal factor algorithm, and introduces time sequence analysis to optimize the local abnormal factor algorithm, so that the accuracy is improved; the risk assessment unit performs risk scoring calculation on abnormal data points in the security monitoring data based on a risk assessment model, and is used for dividing risk levels, so that the early warning information is more specific, and the security is improved; and the early warning notification unit is used for triggering an early warning mechanism according to the divided risk levels.

Description

Technical Field

[0001] The present invention relates to the field of network security technology, and in particular to a security management system based on artificial intelligence early warning. Background Art

[0002] The safety management system based on artificial intelligence early warning is an automated system that uses advanced technical means to conduct real-time monitoring, data collection, analysis and processing, and early warning of monitoring targets. Its main functions include real-time monitoring, data collection, anomaly detection, early warning notification, etc. The purpose of this system is to improve the efficiency of safety management and prevent accidents. The safety management system based on artificial intelligence early warning relies on advanced background technologies, including deep learning and image recognition of artificial intelligence, real-time data collection of the Internet of Things, analysis and processing of big data, elastic resource support of cloud computing, network security protection mechanism, data analysis and visual information communication, and efficient response of edge computing, which together build a powerful monitoring platform. The system can monitor in real time, intelligently analyze and timely warn of potential safety risks, greatly improving the efficiency and accuracy of safety management.

[0003] In the existing safety management system based on artificial intelligence early warning, the problem of false alarms and missed alarms is very serious, and there is a lack of correct management. Some managers rely too much on the system after installing it, and do not formulate corresponding response mechanisms and emergency mechanisms, resulting in a lack of timely response after the system warning. The system may have difficulty in correctly identifying pedestrians in complex backgrounds or low light conditions, leading to accidents. Therefore, a safety management system based on artificial intelligence early warning is designed. Summary of the invention

[0004] The purpose of the present invention is to provide a safety management system based on artificial intelligence early warning to solve the problem raised in the above background technology that in the existing safety management system based on artificial intelligence early warning, the problems of false alarms and missed alarms are very serious and there is a lack of correct management. Some managers rely too much on the system after installing it, and do not formulate corresponding response mechanisms and emergency mechanisms, resulting in a lack of timely response after the system early warning. The system may have difficulty in correctly identifying pedestrians in complex backgrounds or low light conditions, leading to accidents.

[0005] To achieve the above object, the present invention aims to provide a safety management system based on artificial intelligence early warning, including a data acquisition unit, wherein the data acquisition unit is used to collect safety monitoring data on site in real time;

[0006] A data analysis unit, wherein the data analysis unit optimizes the local anomaly factor algorithm by introducing time series analysis based on the abnormal data points in the security monitoring data of the local anomaly factor algorithm;

[0007] Among them, the local anomaly factor algorithm introduced by time series analysis optimization is specifically:

[0008]

[0009] Among them, WeightedLOF t It represents the local anomaly factor value after weighted optimization of time series features at time point t; Represents the time series feature F based on time point t t , the set of k nearest neighbors selected by the adaptive strategy; w(o,t) represents a time weighting factor, the weight of data point o relative to t at time point t; lrd ratio (F t ) represents the time point t, based on the time series feature vector F t The calculated local reachable density ratio; lrd ratio (o) represents the local reachable density ratio of data point o; k represents the parameter that determines the number of neighbors; o represents the neighbor data point;

[0010] A risk assessment unit, wherein the risk assessment unit performs risk scoring calculations on abnormal data points in the security monitoring data based on a risk assessment model and classifies the risk levels;

[0011] An early warning notification unit is used to trigger an early warning mechanism according to the divided risk levels.

[0012] As a further improvement of the technical solution, the data acquisition unit includes a sensor network module and a camera module;

[0013] Wherein, the sensor network module is used to integrate and manage various physical sensors to monitor environmental parameters in real time;

[0014] The camera module is used to capture the video stream of the scene in real time;

[0015] Among them, security monitoring data includes temperature, humidity, smoke, personnel behavior data, and object type data.

[0016] As a further improvement of the technical solution, the data analysis unit establishes an anomaly detection model based on the local anomaly factor algorithm to detect abnormal data points in the security monitoring data. The construction process of the anomaly detection model is specifically as follows:

[0017] S3.1, import the security monitoring data obtained from the data acquisition unit into the database;

[0018] S3.2. Data preprocessing is performed through data cleaning to remove noise and inconsistency;

[0019] S3.3. Extract key features that are helpful for risk identification from raw security monitoring data;

[0020] Among them, the key features include:

[0021] Extract temperature, humidity, and smoke features from the sensor network module;

[0022] Extracting human behavior detection and object type features from video streams;

[0023] S3.4, calculate the anomaly factor using the local anomaly factor algorithm;

[0024] S3.5, use the preprocessed dataset to train the anomaly detection model;

[0025] S3.6, perform k-fold cross validation on the anomaly detection model;

[0026] S3.7, evaluate the anomaly detection model through evaluation indicators;

[0027] S3.8. Deploy the verified and evaluated anomaly detection model.

[0028] As a further improvement of the technical solution, in S3.4, the local anomaly factor algorithm that does not introduce time series analysis optimization specifically includes:

[0029]

[0030] Among them, P C Represents a temperature data point; lrd ratio (P C ) indicates P C The local reachable density of N k (P C ) indicates P C The set of k nearest neighbors; reachability distance (P C ,o) indicates P C The reachable distance to o; lrd ratio (o) represents the local reachability density of o; LOF(P C ) indicates P C LOF value, LOF(P C ) is closer to 1, C represents a normal temperature data point; if the value of LOF(P) is significantly greater than 1, P C represents abnormal temperature data points;

[0031]

[0032]

[0033] Among them, P S Represents a humidity data point; lrd ratio (P S ) indicates P S The local reachable density of N k (P S ) indicates P S The set of k nearest neighbors; reachability distance (P S ,o) indicates P S The reachable distance to o; lrd ratio (o) represents the local reachability density of o; LOF(P S ) indicates P S LOF value, LOF(P S ) is closer to 1, S Indicates a normal humidity data point; if LOF(P S ) is significantly greater than 1, P S Indicates abnormal humidity data points;

[0034]

[0035] Among them, P Y Represents smoke data points; lrd ratio (P Y ) indicates P Y The local reachable density of N k (P Y ) indicates P Y The set of k nearest neighbors; reachability distance (P Y ,o) indicates P Y The reachable distance to o; lrd ratio (o) represents the local reachability density of o; LOF(P Y ) indicates P Y LOF value, LOF(P Y ) is closer to 1, Y represents a normal smoke data point; if LOF(P Y ) is significantly greater than 1, P Y represents an abnormal smoke data point;

[0036]

[0037] Among them, P R Indicates personnel behavior data points; lrd ratio (P R ) indicates P R The local reachable density of N k (PR ) indicates P R The set of k nearest neighbors; reachability distance (P R ,o) indicates P R The reachable distance to o; lrd ratio (o) represents the local reachability density of o; LOF(P R ) indicates P R LOF value, LOF(P R ) is closer to 1, R represents a normal human behavior data point; if LOF(P R ) is significantly greater than 1, P R Represents abnormal personnel behavior data points;

[0038]

[0039] Among them, P W Indicates the object type data point; lrd ratio (P W ) indicates P W The local reachable density of N k (P W ) indicates P W The set of k nearest neighbors; reachability distance (P W ,o) indicates P W The reachable distance to o; lrd ratio (o) represents the local reachability density of o; LOF(P W ) indicates P W LOF value, LOF(P W ) is closer to 1, W Indicates a normal object type data point; if LOF(P W ) is significantly greater than 1, P W Represents a data point of an abnormal object type.

[0040] As a further improvement of the technical solution, in S3.6, the specific steps involved in the k-fold cross validation are:

[0041] S3.61, randomly divide the entire data set into k subsets of roughly the same size;

[0042] S3.62, select k-1 subsets as training sets and the remaining 1 subset as validation set;

[0043] S3.63, using the selected training set to train an anomaly detection model;

[0044] S3.64. Test the performance of the model on the validation set.

[0045] As a further improvement of the technical solution, in S3.7, precision, recall and F1 score are used as evaluation indicators:

[0046] Among them, precision is used to evaluate the accuracy of the anomaly detection model in marking abnormal data points;

[0047] Recall is used to evaluate the ability of anomaly detection models in actually detecting all anomalous data points;

[0048] The F1 score is used to comprehensively consider both precision and recall to provide a comprehensive performance evaluation.

[0049] The evaluation indicators include:

[0050] Accuracy:

[0051]

[0052] Among them, TP represents the number of true positive examples, that is, the number of samples predicted to be positive and actually positive; FP represents the number of false positive examples, that is, the number of samples predicted to be positive but actually negative.

[0053] Recall:

[0054]

[0055] Among them, FN represents the number of false negative examples, that is, the number of samples that are actually positive but predicted as negative by the model;

[0056] F1 score:

[0057]

[0058] The value range of F1Score is 0 to 1. The closer the F1Score is to 1, the better the model performs in terms of both precision and recall.

[0059] As a further improvement of the present technical solution, the risk assessment unit includes a risk assessment model, and the risk assessment model is constructed based on a convolutional neural network algorithm. The construction process of the risk assessment model is specifically as follows:

[0060] S8.1. Preprocess the features of historical abnormal data points and historical risk scores;

[0061] S8.2. Use a hybrid algorithm combining a convolutional neural network algorithm and a batch normalization algorithm with the dataset to train the risk assessment model;

[0062] S8.3. Use k-fold cross validation to validate the risk assessment model;

[0063] S8.4. Evaluate the risk assessment model using evaluation indicators;

[0064] S8.5. Deploy the verified and evaluated risk assessment model.

[0065] As a further improvement of the technical solution, in S8.2, the hybrid algorithm is specifically:

[0066]

[0067] in, represents an element in the output after batch normalization, located in the i-th row, j-th column and k-th channel; γ represents a learnable scaling factor; represents the variance within the batch; ∈ represents a very small constant; represents the normalization factor; represents the element value in the unnormalized feature map after the convolution operation, located in the i-th row, j-th column and k-th channel; μ B represents the average value of all elements at the same position in the batch; β represents a learnable offset.

[0068] Among them, the convolutional neural network algorithm is specifically as follows:

[0069]

[0070] Among them, I represents the input image; K represents the convolution kernel; B represents the bias term; O represents the output feature map; M represents the height of the convolution kernel; N represents the width of the convolution kernel; i and j represent the position index on the output feature map; m and n represent the index variables inside the convolution kernel.

[0071] The batch normalization algorithm is as follows:

[0072]

[0073] Among them, μ B Represents the mean of all elements at the same feature map position in the mini-batch; represents the variance of all elements in the same feature map position in the mini-batch; N represents the batch size; H represents the height of the feature map; W represents the width of the feature map; γ represents a learnable scaling parameter; β represents a learnable offset parameter, and ∈ represents a very small constant used to prevent division by zero.

[0074] As a further improvement of the technical solution, in the risk assessment unit, the risk score calculation is specifically as follows:

[0075] There are three risk dimensions: severity, impact, and frequency, and each dimension has a scoring range;

[0076] Calculated by weighted sum:

[0077] RS=S×W s +I×W i +F×W i ;

[0078] W s +W i +W i =1;

[0079] Among them, RS represents risk score; S represents severity; W represents s Indicates the weight of severity; I indicates the scope of impact; W i represents the weight of the impact range; F represents the frequency of occurrence; W f The weight representing the frequency of occurrence.

[0080] As a further improvement of the technical solution, the warning notification unit includes a warning module, and the warning module is used to issue a warning according to the level of risk;

[0081] Among them, the classification of risk levels is usually based on the threshold of risk score;

[0082] If RS <T low , then RL=low;

[0083] If T low ≤RS <T high , then RL = medium;

[0084] If RS ≥ T high , then RL = high;

[0085] Among them, T low , T high It indicates the threshold value determined according to the risk score; RL indicates the risk level; RL=low indicates low risk; RL=medium indicates medium risk; RL=high indicates high risk.

[0086] Among them, warnings based on the risk level include: low risk flashing blue light and slight alarm sound; medium risk flashing yellow light and medium alarm sound; high risk flashing red light and high alarm sound.

[0087] Compared with the prior art, the present invention has the following beneficial effects:

[0088] 1. In this security management system based on artificial intelligence early warning, the sensor network module and the camera module collect the security monitoring data of the scene in real time, which can obtain the on-site information more accurately and improve the accuracy. The abnormal data points in the security monitoring data are analyzed based on the local abnormal factor algorithm, and the local abnormal factor algorithm is optimized by introducing time series analysis to improve the accuracy of the system.

[0089] 2. In this security management system based on artificial intelligence early warning, an anomaly detection model is established to detect abnormal data points in security monitoring data, and the stability and efficiency of detection are improved through the training, verification and evaluation of the anomaly detection model. By dividing the risk level, the early warning information is made more specific and the security is improved. The hybrid algorithm combining the convolutional neural network algorithm and the batch normalization algorithm allows the risk assessment model to process more quickly and improve efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] Figure 1 It is the overall flow chart of the present invention;

[0091] The meaning of each number in the figure is:

[0092] 1. Data collection unit; 2. Data analysis unit; 3. Risk assessment unit; 4. Early warning notification unit. DETAILED DESCRIPTION

[0093] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0094] Example

[0095] See also Figure 1 As shown, a safety management system based on artificial intelligence early warning is provided, including a data acquisition unit 1, which is used to collect safety monitoring data on site in real time;

[0096] In this example, the data acquisition unit 1 includes a sensor network module and a camera module;

[0097] Wherein, the sensor network module is used to integrate and manage various physical sensors to monitor environmental parameters in real time;

[0098] Specifically, a sensor network module is a hardware device that integrates wireless communication and sensor functions. It consists of sensors, microcontrollers, wireless communication modules, and power supplies. The sensor network module can transmit the data collected by the sensor to a remote device or a central controller through wireless communication technology. These modules are usually used to collect environmental data, such as temperature, humidity, light, etc., and send this data to a central processing system for further analysis and decision-making.

[0099] The camera module is used to capture the video stream of the scene in real time;

[0100] Among them, security monitoring data includes temperature, humidity, smoke, personnel behavior data, and object type data.

[0101] The security management system based on artificial intelligence early warning also includes a data analysis unit 2, which analyzes abnormal data points in the security monitoring data based on big data technology and introduces time series analysis to optimize the local abnormal factor algorithm;

[0102] Among them, the local anomaly factor algorithm introduced by time series analysis optimization is specifically:

[0103]

[0104] Among them, WeightedLOF t It represents the local anomaly factor value after weighted optimization of time series features at time point t; Represents the time series feature F based on time point t t , the set of k nearest neighbors selected by the adaptive strategy; w(o,t) represents a time weighting factor, the weight of data point o relative to t at time point t; lrd ratio (F t ) represents the time point t, based on the time series feature vector F t The calculated local reachable density ratio; lrd ratio (o) represents the local reachable density ratio of data point o; k represents the parameter that determines the number of neighbors; o represents the neighbor data point;

[0105] In this example, the data analysis unit 2 establishes an anomaly detection model based on the local anomaly factor algorithm to detect abnormal data points in the security monitoring data. The construction process of the anomaly detection model is specifically as follows:

[0106] S3.1, import the security monitoring data obtained from the data acquisition unit into the database;

[0107] S3.2. Data preprocessing is performed through data cleaning to remove noise and inconsistency;

[0108] S3.3. Extract key features that are helpful for risk identification from raw security monitoring data;

[0109] Among them, the key features include:

[0110] Extract temperature, humidity, and smoke features from the sensor network module;

[0111] Extracting human behavior detection and object type features from video streams;

[0112] S3.4, calculate the anomaly factor using the local anomaly factor algorithm;

[0113] In this example, the local anomaly factor algorithm without time series analysis optimization specifically includes:

[0114]

[0115] Among them, P C Represents a temperature data point; lrd ratio (P C ) indicates P C The local reachable density of N k (P C ) indicates P C The set of k nearest neighbors; reachability distance (P C ,O) indicates P C The reachable distance to o; lrd ratio (o) represents the local reachability density of o; LOF(P C ) indicates P C LOF value, LOF(P C ) is closer to 1, C represents a normal temperature data point; if the value of LOF(P) is significantly greater than 1, P C represents abnormal temperature data points;

[0116]

[0117] Among them, P S Represents a humidity data point; lrd ratio (P S ) indicates P S The local reachable density of N k (P S ) indicates P S The set of k nearest neighbors; reachability distance (P S ,o) indicates P S The reachable distance to o; lrd ratio(O) represents the local reachability density of o; LOF(P S ) indicates P S LOF value, LOF(P S ) is closer to 1, S Indicates a normal humidity data point; if LOF(P S ) is significantly greater than 1, P S Indicates abnormal humidity data points;

[0118]

[0119] Among them, P Y Represents smoke data points; lrd ratio (P Y ) indicates P Y The local reachable density of N k (P Y ) indicates P Y The set of k nearest neighbors; reachability distance (P Y ,o) indicates P Y The reachable distance to o; lrd ratio (O) represents the local reachability density of o; LOF(P Y ) indicates P Y LOF value, LOF(P Y ) is closer to 1, Y represents a normal smoke data point; if LOF(P Y ) is significantly greater than 1, P Y represents an abnormal smoke data point;

[0120]

[0121] Among them, P R Indicates personnel behavior data points; lrd ratio (P R ) indicates P R The local reachable density of N k (P R ) indicates P R The set of k nearest neighbors; reachability distance (P R ,o) indicates P R The reachable distance to o; lrd ratio (o) represents the local reachability density of o; LOF(P R ) indicates P R LOF value, LOF(P R ) is closer to 1, Rrepresents a normal human behavior data point; if LOF(P R ) is significantly greater than 1, P R Represents abnormal personnel behavior data points;

[0122]

[0123]

[0124] Among them, P W Indicates the object type data point; lrd ratio (P W ) indicates P W The local reachable density of N k (P W ) indicates P W The set of k nearest neighbors; reachability distance (P W ,o) indicates P W The reachable distance to o; lrd ratio (o) represents the local reachability density of o; LOF(P W ) indicates P W LOF value, LOF(P W ) is closer to 1, W Indicates a normal object type data point; if LOF(P W ) is significantly greater than 1, P W Represents a data point of an abnormal object type.

[0125] Specifically, the Local Outlier Factor (LOF) algorithm is an effective algorithm for detecting outliers in a data set. It is based on an unsupervised learning method and does not require any prior labels or markers. The core idea of ​​the LOF algorithm is that the degree of abnormality of a data point can be judged by comparing its density with other data points in its surrounding neighborhood. Specifically, the algorithm first calculates its k-distance neighborhood for each data point, that is, the set of k data points closest to the point. Then, the local reachable density of each data point is calculated, which reflects the relative density of the data point in its neighborhood. Finally, by comparing the local outlier factor of each data point (that is, the ratio of its density to other data points in the neighborhood), it is determined whether it is an outlier.

[0126] S3.5, use the preprocessed dataset to train the anomaly detection model;

[0127] S3.6, perform k-fold cross validation on the anomaly detection model;

[0128] In this example, the specific steps involved in k-fold cross validation are:

[0129] S3.61, randomly divide the entire data set into k subsets of roughly the same size;

[0130] S3.62, select k-1 subsets as training sets and the remaining 1 subset as validation set;

[0131] S3.63, using the selected training set to train an anomaly detection model;

[0132] S3.64. Test the performance of the model on the validation set.

[0133] S3.7, evaluate the anomaly detection model through evaluation indicators;

[0134] Use precision, recall and F1 score as evaluation metrics:

[0135] Among them, precision is used to evaluate the accuracy of the anomaly detection model in marking abnormal data points;

[0136] Recall is used to evaluate the ability of anomaly detection models in actually detecting all anomalous data points;

[0137] The F1 score is used to comprehensively consider both precision and recall to provide a comprehensive performance evaluation.

[0138] The evaluation indicators include:

[0139] Accuracy:

[0140]

[0141] Among them, TP represents the number of true positive examples, that is, the number of samples predicted to be positive and actually positive; FP represents the number of false positive examples, that is, the number of samples predicted to be positive but actually negative.

[0142] Recall:

[0143]

[0144] Among them, FN represents the number of false negative examples, that is, the number of samples that are actually positive but predicted as negative by the model;

[0145] F1 score:

[0146]

[0147] The value range of F1Score is 0 to 1. The closer the F1Score is to 1, the better the model performs in terms of both precision and recall.

[0148] As a further improvement of the technical solution, in S3.44, the local abnormal factor is specifically:

[0149]

[0150] Among them, lrd ratio (P) represents the local reachability density of P; N k (P) represents the set of k nearest neighbors of P; reachability distance (P,o) represents the reachable distance from P to o; lrd ratio (o) represents the local reachable density of o; LOF(P) represents the LOF value of P. The closer the value of LOF(P) is to 1, the more likely P is a normal data point; if the value of LOF(P) is significantly greater than 1, P may be an outlier.

[0151] S3.8. Deploy the verified and evaluated anomaly detection model.

[0152] The security management system based on artificial intelligence early warning also includes a risk assessment unit 3, which performs risk scoring calculation and risk level classification on abnormal data points in the security monitoring data based on a risk assessment model;

[0153] In this example, the risk score calculation is as follows:

[0154] There are three risk dimensions: severity, impact, and frequency, and each dimension has a scoring range;

[0155] Calculated by weighted sum:

[0156] RS=S×W s +I×W i +F×W i ;

[0157] W s +W i +W i =1;

[0158] Among them, RS represents risk score; S represents severity; W represents s Indicates the weight of severity; I indicates the scope of impact; W i represents the weight of the impact range; F represents the frequency of occurrence; W f The weight representing the frequency of occurrence.

[0159] In this example, the risk assessment unit includes a risk assessment model, and the risk assessment model is constructed based on a convolutional neural network algorithm. The construction process of the risk assessment model is specifically as follows:

[0160] S8.1. Preprocess the features of historical abnormal data points and historical risk scores;

[0161] S8.2. Use a hybrid algorithm combining a convolutional neural network algorithm and a batch normalization algorithm with the dataset to train the risk assessment model;

[0162] In this example, the hybrid algorithm is specifically:

[0163]

[0164] in, represents an element in the output after batch normalization, located in the i-th row, j-th column and k-th channel; γ represents a learnable scaling factor; represents the variance within the batch; ∈ represents a very small constant; represents the normalization factor; represents the element value in the unnormalized feature map after the convolution operation, located in the i-th row, j-th column and k-th channel; μ B represents the average value of all elements at the same position in the batch; β represents a learnable offset.

[0165] Among them, the convolutional neural network algorithm is specifically as follows:

[0166]

[0167] Among them, I represents the input image; K represents the convolution kernel; B represents the bias term; O represents the output feature map; M represents the height of the convolution kernel; N represents the width of the convolution kernel; i and j represent the position index on the output feature map; m and n represent the index variables inside the convolution kernel.

[0168] The batch normalization algorithm is as follows:

[0169]

[0170] Among them, μ B Represents the mean of all elements at the same feature map position in the mini-batch; represents the variance of all elements in the same feature map position in the mini-batch; N represents the batch size; H represents the height of the feature map; W represents the width of the feature map; γ represents a learnable scaling parameter; β represents a learnable offset parameter, and ∈ represents a very small constant used to prevent division by zero.

[0171] S8.3. Use k-fold cross validation to validate the risk assessment model;

[0172] S8.4. Evaluate the risk assessment model using evaluation indicators;

[0173] S8.5. Deploy the verified and evaluated risk assessment model.

[0174] The security management system based on artificial intelligence early warning also includes an early warning notification unit 4, which triggers an early warning mechanism according to the risk level based on the early warning module.

[0175] In this example, the warning notification unit 4 includes a warning module, which is used to issue a warning according to the level of risk.

[0176] Among them, the classification of risk levels is usually based on the threshold of risk score;

[0177] If RS <T low , then RL=low;

[0178] If T low ≤RS <T high , then RL = medium;

[0179] If RS ≥ T high , then RL = high;

[0180] Among them, T low , T high It indicates the threshold value determined according to the risk score; RL indicates the risk level; RL=low indicates low risk; RL=medium indicates medium risk; RL=high indicates high risk.

[0181] Among them, warnings based on the risk level include: low risk flashing blue light and slight alarm sound; medium risk flashing yellow light and medium alarm sound; high risk flashing red light and high alarm sound.

[0182] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A safety management system based on artificial intelligence early warning, characterized in that: include: A data collection unit (1), the data collection unit (1) is used to collect on-site safety monitoring data in real time; A data analysis unit (2), wherein the data analysis unit (2) analyzes abnormal data points in the security monitoring data based on a local abnormal factor algorithm, and introduces time series analysis to optimize the local abnormal factor algorithm; Among them, the local anomaly factor algorithm introduced by time series analysis optimization is specifically: Among them, WeightedLOF t It represents the local anomaly factor value after weighted optimization of time series features at time point t; Represents the time series feature F based on time point t t , the set of k nearest neighbors selected by the adaptive strategy; w(o,t) represents a time weighting factor, the weight of data point o relative to t at time point t; lrd ratio (F t ) represents the time point t, based on the time series feature vector F t The calculated local reachable density ratio; lrd ratio (o) represents the local reachable density ratio of data point o; k represents the parameter that determines the number of neighbors; o represents the neighbor data point; A risk assessment unit (3), wherein the risk assessment unit (3) calculates risk scores for abnormal data points in the security monitoring data based on a risk assessment model; An early warning notification unit (4), the early warning notification unit (4) is used to trigger an early warning mechanism according to the divided risk levels.

2. The safety management system based on artificial intelligence early warning according to claim 1 is characterized by: The data acquisition unit (1) comprises a sensor network module and a camera module; Wherein, the sensor network module is used to integrate and manage real-time monitoring data of various physical sensors; The camera module is used to capture the video stream of the scene in real time; Among them, security monitoring data includes temperature, humidity, smoke, personnel behavior data, and object type data.

3. The safety management system based on artificial intelligence early warning according to claim 1 is characterized by: The data analysis unit (2) establishes an anomaly detection model based on a local anomaly factor algorithm to detect abnormal data points in the security monitoring data. The construction process of the anomaly detection model is specifically as follows: S3.1, import the security monitoring data obtained from the data acquisition unit into the database; S3.

2. Data preprocessing is performed through data cleaning to remove noise and inconsistency; S3.3, extract key features that are helpful for risk identification from raw security monitoring data through sensor network modules and camera modules; Among them, key features include: Extract temperature, humidity, and smoke features from the sensor network module; Extracting human behavior detection and object type features from video streams; S3.4, calculate the anomaly factor using the local anomaly factor algorithm optimized by introducing time series analysis; S3.5, use the preprocessed dataset to train the anomaly detection model; S3.6, perform k-fold cross validation on the anomaly detection model; S3.7, evaluate the anomaly detection model through evaluation indicators; S3.

8. Deploy the verified and evaluated anomaly detection model.

4. The safety management system based on artificial intelligence early warning according to claim 3 is characterized by: In S3.4, the local anomaly factor algorithm that does not introduce time series analysis optimization specifically includes: Among them, P C Represents a temperature data point; lrd ratio (P C ) indicates P C The local reachable density of N k (P C ) indicates P C The set of k nearest neighbors; reachability distance (P C ,o) indicates P C The reachable distance to o; lrd ratio (o) represents the local reachability density of o; LOF(P C ) indicates P C LOF value, LOF(P C ) is closer to 1, C represents a normal temperature data point; if the value of LOF(P) is significantly greater than 1, P C represents abnormal temperature data points; Among them, P S Represents a humidity data point; lrd ratio (P S ) indicates P S The local reachable density of N k (P S ) indicates P S The set of k nearest neighbors; reachability distance (P S ,o) indicates P S The reachable distance to o; lrd ratio (o) represents the local reachability density of o; LOF(P S ) indicates P S LOF value, LOF(P S ) is closer to 1, S Indicates a normal humidity data point; if LOF(P S ) is significantly greater than 1, P S represents an abnormal humidity data point; Among them, P Y Represents smoke data points; lrd ratio (P Y ) indicates P Y The local reachable density of N k (P Y ) indicates P Y The set of k nearest neighbors; reachability distance (P Y ,o) indicates P Y The reachable distance to o; lrd ratio (o) represents the local reachability density of o; LOF(P Y ) indicates P Y LOF value, LOF(P Y ) is closer to 1, Y represents a normal smoke data point; if LOF(P Y ) is significantly greater than 1, P Y represents an abnormal smoke data point; Among them, P R Indicates personnel behavior data points; lrd ratio (P R ) indicates P R The local reachable density of N k (P R ) indicates P R The set of k nearest neighbors; reachability distance (P R ,o) indicates P R The reachable distance to o; lrd ratio (o) represents the local reachability density of o; LOF(P R ) indicates P R LOF value, LOF(P R ) is closer to 1, R represents a normal human behavior data point; if LOF(P R ) is significantly greater than 1, P R Represents abnormal personnel behavior data points; Among them, P W Indicates the object type data point; lrd ratio (P W ) indicates P W The local reachable density of N k (P W ) indicates P W The set of k nearest neighbors; reachability distance (P W ,o) indicates P W The reachable distance to o; lrd ratio (o) represents the local reachability density of o; LOF(P W ) indicates P W LOF value, LOF(P W ) is closer to 1, W Indicates a normal object type data point; if LOF(P W ) is significantly greater than 1, P W Represents a data point of an abnormal object type.

5. The safety management system based on artificial intelligence early warning according to claim 3 is characterized by: In S3.6, the specific steps involved in the k-fold cross validation are: S3.61, randomly divide the entire data set into k subsets of roughly the same size; S3.62, select k-1 subsets as training sets and the remaining 1 subset as validation set; S3.63, using the selected training set to train an anomaly detection model; S3.

64. Test the performance of the model on the validation set.

6. The safety management system based on artificial intelligence early warning according to claim 3 is characterized by: In S3.7, precision, recall and F1 score are used as evaluation indicators: Among them, precision is used to evaluate the accuracy of the anomaly detection model in marking abnormal data points; Recall is used to evaluate the ability of anomaly detection models in actually detecting all anomalous data points; The F1 score is used to comprehensively consider both precision and recall.

7. The safety management system based on artificial intelligence early warning according to claim 1 is characterized by: The risk assessment unit (3) includes a risk assessment model, and the risk assessment model is constructed based on a convolutional neural network algorithm. The construction process of the risk assessment model is specifically as follows: S8.

1. Preprocess the features of historical abnormal data points and historical risk scores; S8.

2. Use a hybrid algorithm combining a convolutional neural network algorithm and a batch normalization algorithm with the dataset to train the risk assessment model; S8.

3. Use k-fold cross validation to validate the risk assessment model; S8.

4. Evaluate the risk assessment model using evaluation indicators; S8.

5. Deploy the verified and evaluated risk assessment model.

8. The safety management system based on artificial intelligence early warning according to claim 7 is characterized by: In S8.2, the hybrid algorithm is specifically: in, represents an element in the output after batch normalization, located in the i-th row, j-th column and k-th channel; γ represents a learnable scaling factor; represents the variance within the batch; ∈ represents a very small constant; represents the normalization factor; represents the element value in the unnormalized feature map after the convolution operation, located in the i-th row, j-th column and k-th channel; μ B represents the average value of all elements at the same position in the batch; β represents a learnable offset.

9. The safety management system based on artificial intelligence early warning according to claim 1 is characterized by: In the risk assessment unit (3), the risk score calculation is specifically as follows: There are three risk dimensions: severity, impact, and frequency, and each dimension has a scoring range; Calculated by weighted sum: RS=S×W s +I×W i +F×W i ; IN s +W i +W i =1; Among them, RS represents risk score; S represents severity; W represents s Indicates the weight of severity; I indicates the scope of impact; W i represents the weight of the impact range; F represents the frequency of occurrence; W f The weight representing the frequency of occurrence.

10. The safety management system based on artificial intelligence early warning according to claim 9 is characterized in that: The early warning notification unit (4) comprises an early warning module, and the early warning module is used to issue an early warning according to the risk level; Among them, the classification of risk levels is based on the threshold of risk score; If RS <T low , then RL=low; If T low ≤RS <T high , then RL = medium; If RS ≥ T high , then RL = high; Among them, T low , T high It indicates the threshold value determined according to the risk score; RL indicates the risk level; RL=low indicates low risk; RL=medium indicates medium risk; RL=high indicates high risk.

Images