Abnormal hidden danger judgment optimization method and system for production operation defense organization area

Abstract

The invention discloses an abnormal hidden danger judgment optimization method and system for a production operation defense organization area, and the method comprises the steps: obtaining monitoring information of a target area, carrying out the data preprocessing, carrying out the heterogeneous feature extraction and fusion according to the preprocessed monitoring data of the target area, and generating a cross-modal fusion feature; constructing an abnormal vibration analysis model, inputting the abnormal vibration analysis model into the abnormal vibration analysis model for analysis, and judging whether the enclosing wall of the defense organization area is abnormal or not to obtain abnormal vibration analysis information; an intrusion behavior analysis model is constructed, when abnormal vibration is monitored, abnormal hidden danger judgment and analysis are carried out in combination with abnormal vibration analysis information, and early warning prompt is carried out; and when the abnormal invasive organism does not stop the invasion behavior according to the early warning prompt, the target abnormal invasive organism is subjected to invasion trajectory tracking, and a trajectory guidance report is generated and sent to nearby patrol personnel for expelling assistance, so that the accuracy and intelligence of abnormal hidden danger judgment in a complex defense deployment scene are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of hidden danger control in defense areas, and in particular to an optimization method and system for determining abnormal hidden dangers in defense areas for production operations. Background Art

[0002] In the field of industrial production safety, the technology for identifying abnormal hazards in defense zones has long relied on a combination of manual inspections and single-dimensional monitoring equipment. Traditional methods typically use fixed sensors (such as vibration sensors and infrared beamforming devices) to perform static threshold monitoring of perimeter walls or key equipment, triggering alarms when the detection value exceeds a preset safety range. However, this technology has significant drawbacks in practical applications. On the one hand, dynamic interference in complex operating environments (such as heavy machinery vibration and equipment deformation caused by extreme weather) can easily lead to a surge in false alarm rates, and a single sensor cannot distinguish between normal operating vibrations and abnormal signals generated by illegal intrusions. On the other hand, the lack of an effective fusion and analysis mechanism for multi-source heterogeneous monitoring data, and the independent operation of each system, leads to insufficient temporal and spatial correlation in hazard assessment.

[0003] Furthermore, traditional static threshold setting methods struggle to adapt to the dynamic changes of complex deployment scenarios, making it extremely easy to trigger false alarms, requiring frequent manual calibration and resulting in high operational and maintenance costs. In recent years, some improved solutions have attempted to incorporate deep learning technology, but most models rely on large amounts of labeled data for training and lack the ability to achieve collaborative reasoning across sensor data. This results in insufficient early identification of hidden hazards (such as pre-contact vibration before climbing), making it difficult to meet the demand for precise, all-weather prevention and control in high-risk work areas.

[0004] Therefore, a method and system for optimizing abnormal hidden danger determination in production operation deployment areas are provided, thereby improving the accuracy and intelligence of abnormal hidden danger determination in complex deployment scenarios. Summary of the Invention

[0005] The present invention overcomes the defects of the prior art and provides a method and system for optimizing abnormal hidden danger determination in a production operation defense area.

[0006] To achieve the above objectives, the present invention provides, in a first aspect, a method for optimizing abnormal hidden danger determination in a production operation defense area, comprising: Monitor the target area to obtain target area monitoring information and perform data preprocessing. Perform heterogeneous feature extraction and fusion based on the preprocessed target area monitoring data to generate cross-modal fusion features. Constructing an abnormal vibration analysis model, inputting the cross-modal fusion features into the abnormal vibration analysis model for analysis, determining whether there is an abnormality in the defense area wall, and obtaining abnormal vibration analysis information; Construct an intrusion behavior analysis model. When abnormal vibration is detected, the model is combined with the abnormal vibration analysis information to perform abnormal hidden danger analysis, obtain intrusion behavior analysis information and issue early warning prompts. When abnormal invading creatures fail to stop their invading behavior according to the early warning prompts, the invasion trajectory of the target abnormal invading creatures is tracked, and a trajectory guidance report is generated and sent to nearby patrol personnel for expulsion assistance.

[0007] In this solution, the monitoring of the target area to obtain target area monitoring information and performing data preprocessing specifically includes: Based on the preset deployment plan, a heterogeneous sensor network is deployed in the target area, and the target area is monitored around the clock through the deployed heterogeneous sensor network to obtain monitoring information of the target area; Sending the target area monitoring information to the edge computing node for preprocessing according to a preset transmission protocol, wherein the target area monitoring information includes video monitoring data and vibration monitoring data; The cubic spline interpolation algorithm is used to reconstruct the intermediate moment image of the video monitoring data. The time interval between two adjacent frames is divided into several equal points and the pixel motion vector field is obtained. The interpolation weight is calculated based on the pixel motion vector field. The image is reconstructed according to the interpolation weight to eliminate the timing deviation between the video data and the sensor. The vibration monitoring data is decomposed into wavelet packets, and the db4 wavelet basis function is used to divide the original waveform into sub-band frequency domains. The environmental noise is eliminated by the soft threshold method. After the preprocessing operation is completed, the preprocessed target area monitoring information is pushed to the next processing unit through the lightweight message queue built into the edge computing node.

[0008] In this solution, heterogeneous feature extraction and fusion are performed based on the pre-processed target area monitoring data to generate cross-modal fusion features, specifically including: The next processing unit obtains the pre-processed target area monitoring information transmitted by the edge computing node through a preset data receiving module, and imports the pre-processed target area monitoring information into a preset parallel feature extraction channel for processing. The parallel feature extraction channel includes a video feature extraction channel and a vibration feature extraction channel; The video monitoring data is input into the video feature extraction pipeline built on a lightweight convolutional neural network framework. The input image is processed frame by frame using a depthwise separable convolutional layer. A coordinate attention mechanism is introduced in the bottleneck layer to dynamically focus on abnormal areas and the contours of moving targets, generating a multi-scale visual feature map. The vibration monitoring data is input into a vibration feature extraction pipeline built on a recurrent neural network framework. The forward branch captures the vibration patterns of the historical time window, and the backward branch models future potential correlations. The bidirectional hidden states are fused through a gating mechanism to output a time-series vibration feature vector. The visual feature map is compressed into a region descriptor through spatial pyramid pooling, and tensor splicing and dimension alignment are performed with the temporal vibration feature vector. A multi-head cross-attention mechanism is introduced to perform cross-modal feature fusion; Taking temporal vibration features as query vectors and visual features as key-value pairs, the spatiotemporal correlation matrix of vibration-visual features is calculated through dot product attention, and cross-modal fusion features are generated after weighted aggregation.

[0009] In this solution, the abnormal vibration analysis model is constructed, and the cross-modal fusion features are input into the abnormal vibration analysis model for analysis to determine whether there is an abnormality in the defense area wall, and obtain abnormal vibration analysis information, specifically including: Introducing a big data network, obtaining historical vibration monitoring data of the defense zone walls during normal deployment through the big data network, performing feature extraction on the historical vibration monitoring data, and obtaining historical vibration feature information; An abnormal vibration analysis model is built based on a dynamic Gaussian mixture model. The expectation maximization algorithm is used to optimize the parameters of the abnormal vibration analysis model. The initial mean vector, covariance matrix and mixing coefficient of the Gaussian component are generated by randomly selecting k samples, and the initial probability density is assigned. During the iterative optimization process, the posterior probability of each eigenvector belonging to each Gaussian component is calculated, and the mean, covariance and mixing coefficient of the Gaussian component are re-estimated based on the posterior probability to obtain the model parameter set output after the iteration stops; Optimizing the abnormal vibration analysis model using the model parameter set, constructing a training data set based on the historical vibration feature information, and importing the data set into the optimized abnormal vibration analysis model for training; Obtaining cross-modal fusion features, inputting the cross-modal fusion features into the trained abnormal vibration analysis model, and segmenting the cross-modal fusion features in chronological order using a sliding window mechanism; The Mahalanobis distance between the vibration feature vector and all Gaussian components in each window is calculated, and the calculated Mahalanobis distance is compared with the preset threshold. If it is greater than the preset threshold, it means that there is a hidden danger of abnormal intrusion, and abnormal vibration analysis information is generated and an early warning is issued.

[0010] In this solution, the intrusion behavior analysis model is constructed. When abnormal vibration is detected, the abnormal hidden danger determination analysis is performed in combination with the abnormal vibration analysis information, and the intrusion behavior analysis information is obtained and an early warning prompt is issued. Specifically, the following steps are included: When abnormal vibration is detected, it means that there may be a hidden danger of wall invasion in the abnormal vibration area, and abnormal vibration analysis information is obtained. According to the abnormal vibration analysis information, the abnormal area and abnormal period are traced to obtain regional abnormal tracing information. Obtain cross-modal fusion features, obtain the video surveillance device number in the abnormal area through the regional anomaly tracing information, and extract the video surveillance features of the corresponding numbered video surveillance device in the target abnormal period from the cross-modal fusion features in combination with the abnormal period tracing results to obtain the tracing video surveillance features; The invasion behavior characteristics of different organisms are obtained through big data retrieval, and the obtained invasion behavior characteristics are labeled with behavior categories to form an invasion behavior feature set. The invasion behavior analysis model is constructed using SVDD, and the model is trained based on the invasion behavior feature set. The input intrusion behavior features are mapped to a high-dimensional space using a Gaussian kernel function. The minimum hypersphere boundary containing all intrusion behavior feature samples is obtained in the mapped high-dimensional space and the trained intrusion behavior analysis model is output. Input the traceability video surveillance features into the trained intrusion behavior analysis model to generate a real-time feature vector, calculate the Mahalanobis distance of the real-time feature vector in the kernel space, and compare it with the preset hypersphere radius; If it is smaller than the preset hypersphere radius, it means there is a hidden danger of biological invasion. The cosine similarity between the real-time feature vector and the invasion behavior feature in the kernel space is calculated, and the abnormal behavior category corresponding to the invasion behavior feature with the largest similarity is selected as the belonging category. The invasion behavior analysis information is output and the early warning measures are obtained in the preset strategy library to perform abnormal invasion warning.

[0011] In this solution, when the abnormal intruder fails to stop its intrusion according to the early warning prompt, the target abnormal intruder is tracked and a trajectory guidance report is generated and sent to nearby patrol personnel for expulsion assistance. Specifically, the following steps are performed: Obtaining intrusion behavior analysis information, determining the real-time invasion status of the target intruding organism based on the intrusion behavior analysis information, and if the invasion is not complete, obtaining the real-time location of patrol personnel, and sending the location of the intrusion area to the nearest patrol personnel for guidance; If the invasion is complete, the invasion trajectory of the target intruder is tracked, and the intrusion video monitoring features when the target abnormal intruder is detected in the defense area are extracted through cross-modal fusion features; The intrusion video monitoring features are imported into the extended Kalman filter to estimate the motion state of the target abnormal intruder, and the motion state transfer matrix of the abnormal intruder is calculated by first-order approximation, and the predicted motion state is generated according to the motion state transfer matrix; A Bayesian inference network is introduced to obtain the observed motion state of abnormal invading organisms from the invasion video monitoring features through the motion state estimation algorithm. The observed motion state is used as the input of the Bayesian inference network for network training and to obtain the transient posterior distribution. Inferring the next-moment motion state of the abnormal invading organism in the Bayesian inference network through the transient posterior distribution, generating a plurality of inferred motion states, and selecting the inferred motion state with the highest inference probability value to perform difference calculation with the predicted motion state; Obtain the degree of difference by difference calculation and compare it with the preset difference degree threshold. If the difference is greater than the preset threshold, the predicted motion state is corrected using the corresponding inferred motion state to obtain a corrected predicted operating state. The invasion tracking trajectory of the target abnormal invading organism is generated according to the corrected predicted operating status, and a trajectory guidance report is generated and sent to nearby patrol personnel for expulsion assistance.

[0012] A second aspect of the present invention provides a system for optimizing abnormal hidden danger determination in a production operation defense zone. The system includes: a memory and a processor. The memory contains a program for optimizing abnormal hidden danger determination in a production operation defense zone. When the program is executed by the processor, the following steps are implemented: Monitor the target area to obtain target area monitoring information and perform data preprocessing. Perform heterogeneous feature extraction and fusion based on the preprocessed target area monitoring data to generate cross-modal fusion features. Constructing an abnormal vibration analysis model, inputting the cross-modal fusion features into the abnormal vibration analysis model for analysis, determining whether there is an abnormality in the defense area wall, and obtaining abnormal vibration analysis information; Construct an intrusion behavior analysis model. When abnormal vibration is detected, the model is combined with the abnormal vibration analysis information to perform abnormal hidden danger analysis, obtain intrusion behavior analysis information and issue early warning prompts. When abnormal invading creatures fail to stop their invading behavior according to the early warning prompts, the invasion trajectory of the target abnormal invading creatures is tracked, and a trajectory guidance report is generated and sent to nearby patrol personnel for expulsion assistance.

[0013] The present invention discloses a method and system for optimizing abnormal hidden danger judgment in a production operation defense area, comprising: obtaining monitoring information of a target area and performing data preprocessing, extracting and fusing heterogeneous features based on the preprocessed monitoring data of the target area to generate cross-modal fusion features; constructing an abnormal vibration analysis model, inputting the model into the abnormal vibration analysis model for analysis, judging whether there is an abnormality in the defense area wall, and obtaining abnormal vibration analysis information; constructing an intrusion behavior analysis model, and when abnormal vibration is detected, performing abnormal hidden danger judgment analysis in combination with the abnormal vibration analysis information, and issuing an early warning prompt; when the abnormal intruder fails to stop the intrusion behavior according to the early warning prompt, tracking the intrusion trajectory of the target abnormal intruder, and generating a trajectory guidance report to send to nearby patrol personnel for expulsion assistance, thereby improving the accuracy and intelligence of abnormal hidden danger judgment in complex defense scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the embodiments or exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments or exemplary descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained according to these drawings without paying any creative work.

[0015] Figure 1 A flow chart of an optimization method for determining abnormal hidden dangers in a production operation defense area provided by one embodiment of the present invention; Figure 2 A flowchart of an auxiliary method for eliminating abnormal hidden dangers in a production operation defense area provided by one embodiment of the present invention; Figure 3 A block diagram of an abnormal hidden danger determination optimization system for a production operation defense area provided by one embodiment of the present invention; The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0016] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0017] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0018] Figure 1A flow chart of an optimization method for determining abnormal hidden dangers in a production operation defense area provided by one embodiment of the present invention; like Figure 1 As shown, the present invention provides a flow chart of an optimization method for abnormal hidden danger determination in a production operation defense area, including: S102 monitors the target area to obtain target area monitoring information and performs data preprocessing, extracts and fuses heterogeneous features based on the preprocessed target area monitoring data to generate cross-modal fusion features; S104: constructing an abnormal vibration analysis model, inputting the cross-modal fusion features into the abnormal vibration analysis model for analysis, determining whether there is an abnormality in the defense area wall, and obtaining abnormal vibration analysis information; S106 builds an intrusion behavior analysis model. When abnormal vibration is detected, the model is combined with the abnormal vibration analysis information to perform abnormal hidden danger analysis, obtain intrusion behavior analysis information, and issue an early warning prompt. S108 When the abnormal invading creature fails to stop the invading behavior according to the early warning prompt, the invasion trajectory of the target abnormal invading creature is tracked, and a trajectory guidance report is generated and sent to nearby patrol personnel for expulsion assistance.

[0019] Furthermore, in a preferred embodiment of the present invention, the step of monitoring the target area to obtain target area monitoring information and performing data preprocessing specifically includes: Based on the preset deployment plan, a heterogeneous sensor network is deployed in the target area, and the target area is monitored around the clock through the deployed heterogeneous sensor network to obtain monitoring information of the target area; Sending the target area monitoring information to the edge computing node for preprocessing according to a preset transmission protocol, wherein the target area monitoring information includes video monitoring data and vibration monitoring data; The cubic spline interpolation algorithm is used to reconstruct the intermediate moment image of the video monitoring data. The time interval between two adjacent frames is divided into several equal points and the pixel motion vector field is obtained. The interpolation weight is calculated based on the pixel motion vector field. The image is reconstructed according to the interpolation weight to eliminate the timing deviation between the video data and the sensor. The vibration monitoring data is decomposed into wavelet packets, and the db4 wavelet basis function is used to divide the original waveform into sub-band frequency domains. The environmental noise is eliminated by the soft threshold method. After the preprocessing operation is completed, the preprocessed target area monitoring information is pushed to the next processing unit through the lightweight message queue built into the edge computing node.

[0020] It should be noted that a multi-layered monitoring system, built on a pre-defined deployment plan, deploys a heterogeneous sensor network with complementary sensing capabilities in the target area. This includes dual-spectral cameras and positioning fiber-optic vibration sensors. All sensor nodes transmit data via a hybrid industrial-grade ZigBee and 5G network. The IEEE 1588 precision clock synchronization protocol, based on timestamps, eliminates clock skew between devices and ensures time alignment of multi-source data. After the original monitoring data is encrypted and transmitted to the edge computing node, it is pre-processed according to the physical characteristics of the video stream and vibration signal: the video data adopts the motion-compensated cubic spline interpolation algorithm, and the motion vector of the pixel points between two adjacent frames is calculated by the optical flow method. A motion trajectory model based on affine transformation is constructed, and reconstructed frames are inserted between adjacent frames to compensate for the timing misalignment caused by the difference in the camera sampling rate and the vibration sensor sampling rate; the vibration signal is decomposed into 32 frequency bands through 6-layer wavelet packet decomposition, and the tight support characteristics of the db4 wavelet basis function are used to accurately separate high-frequency noise and effective signals (such as climbing shock waves). The adaptive soft threshold method is used to reduce the noise of each sub-band, in which the threshold parameter is dynamically adjusted according to the sub-band energy ratio to ensure that the signal mutation characteristics are retained while eliminating environmental interference. The pre-processed multi-source data is prioritized and traffic shaped by the lightweight message middleware (such as NanoMQ) built into the edge node. High-priority data (such as vibration alarm signals) is pushed to the cloud analysis platform through a preemptive transmission channel. At the same time, a ring buffer mechanism is used to prevent data loss, providing high-precision input for subsequent multimodal feature fusion and intelligent analysis.

[0021] Furthermore, in a preferred embodiment of the present invention, the heterogeneous feature extraction and fusion based on the pre-processed target area monitoring data to generate cross-modal fusion features specifically includes: The next processing unit obtains the pre-processed target area monitoring information transmitted by the edge computing node through a preset data receiving module, and imports the pre-processed target area monitoring information into a preset parallel feature extraction channel for processing. The parallel feature extraction channel includes a video feature extraction channel and a vibration feature extraction channel; The video monitoring data is input into the video feature extraction pipeline built on a lightweight convolutional neural network framework. The input image is processed frame by frame using a depthwise separable convolutional layer. A coordinate attention mechanism is introduced in the bottleneck layer to dynamically focus on abnormal areas and the contours of moving targets, generating a multi-scale visual feature map. The vibration monitoring data is input into a vibration feature extraction pipeline built on a recurrent neural network framework. The forward branch captures the vibration patterns of the historical time window, and the backward branch models future potential correlations. The bidirectional hidden states are fused through a gating mechanism to output a time-series vibration feature vector. The visual feature map is compressed into a region descriptor through spatial pyramid pooling, and tensor splicing and dimension alignment are performed with the temporal vibration feature vector. A multi-head cross-attention mechanism is introduced to perform cross-modal feature fusion; Taking temporal vibration features as query vectors and visual features as key-value pairs, the spatiotemporal correlation matrix of vibration-visual features is calculated through dot product attention, and cross-modal fusion features are generated after weighted aggregation.

[0022] It should be noted that after the pre-processed monitoring information is pushed to the next processing unit through the message queue of the edge computing node, the protocol parsing and format standardization of the received multi-source data are performed through the preset data receiving module to ensure strict alignment of the timestamps of the video stream and the vibration signal. The processing unit starts the preset parallel feature extraction channel and imports the video monitoring data into the video feature extraction channel built based on a lightweight convolutional neural network: a depth-wise separable convolution layer is used to extract the low-order texture features of the image frame by frame. After compressing the feature dimensions through the bottleneck layer, a coordinate attention mechanism is introduced to jointly assign weights to the spatial position and channel dimensions of the feature map, focusing on enhancing the response strength of abnormal areas. At the same time, a multi-scale feature pyramid structure is used to fuse the target contour information under different receptive fields to generate a visual feature map containing spatial semantics and dynamic motion clues. The synchronous vibration feature extraction pipeline utilizes a recurrent neural network framework, employing a bidirectional gated recurrent unit (Bi-GRU) designed for vibration monitoring data. The forward branch analyzes frequency-domain mutation patterns (such as impact pulses and steady-state resonances) in historical vibration waveforms along the time axis. The backward branch inversely models potential trend correlations within future time windows. A gating mechanism dynamically adjusts the fusion ratio of bidirectional hidden states, outputting a time-series feature vector that characterizes the spatiotemporal evolution of vibration events. The visual feature map is compressed into a fixed-dimensional region descriptor using spatial pyramid pooling. This tensor is then concatenated with the vibration time-series feature vector. Layer-wise normalization eliminates dimensional differences between modalities, and the resulting image is then fed into a multi-headed cross-attention mechanism layer. Using the vibration feature vector as the query vector and the visual features as key-value pairs, a dot product operation is used to calculate the correlation weight matrix between the vibration signal and the visual target on the spatiotemporal grid. Attention-weighted aggregation of the visual features is then performed to generate a cross-modal feature vector that combines vibration energy distribution and visual spatial localization, providing a highly discriminative data foundation for subsequent joint determination of anomalies and potential hazards.

[0023] Furthermore, in a preferred embodiment of the present invention, the abnormal vibration analysis model is constructed, and the cross-modal fusion feature is input into the abnormal vibration analysis model for analysis to determine whether there is an abnormality in the defense area wall, thereby obtaining abnormal vibration analysis information, which specifically includes: Introducing a big data network, obtaining historical vibration monitoring data of the defense zone walls during normal deployment through the big data network, performing feature extraction on the historical vibration monitoring data, and obtaining historical vibration feature information; An abnormal vibration analysis model is built based on a dynamic Gaussian mixture model. The expectation maximization algorithm is used to optimize the parameters of the abnormal vibration analysis model. The initial mean vector, covariance matrix and mixing coefficient of the Gaussian component are generated by randomly selecting k samples, and the initial probability density is assigned. During the iterative optimization process, the posterior probability of each eigenvector belonging to each Gaussian component is calculated, and the mean, covariance and mixing coefficient of the Gaussian component are re-estimated based on the posterior probability to obtain the model parameter set output after the iteration stops; Optimizing the abnormal vibration analysis model using the model parameter set, constructing a training data set based on the historical vibration feature information, and importing the data set into the optimized abnormal vibration analysis model for training; Obtaining cross-modal fusion features, inputting the cross-modal fusion features into the trained abnormal vibration analysis model, and segmenting the cross-modal fusion features in chronological order using a sliding window mechanism; The Mahalanobis distance between the vibration feature vector and all Gaussian components in each window is calculated, and the calculated Mahalanobis distance is compared with the preset threshold. If it is greater than the preset threshold, it means that there is a hidden danger of abnormal intrusion, and abnormal vibration analysis information is generated and an early warning is issued.

[0024] It should be noted that historical vibration monitoring data of the defense zone walls under normal deployment conditions is first accessed through a big data network. This data contains vibration waveform and spectral characteristics under different time periods and environmental conditions, such as steady-state vibration during equipment operation and random fluctuations caused by wind and rain interference. Multidimensional feature extraction is performed on the raw vibration data. Using a combination of wavelet packet decomposition and energy entropy calculation, the time-domain waveform is converted into a set of feature vectors representing vibration intensity, frequency distribution, and mutation characteristics, forming a historical vibration feature information database. Based on this feature database, a dynamic Gaussian mixture model (GMM) is constructed as the core framework for anomaly analysis. The model parameters are iteratively optimized using the expectation-maximization (EM) algorithm. The K-means++ algorithm is used in the initialization phase to replace pure random selection. K cluster centers are selected from the historical features as the initial mean vectors of the Gaussian components. The covariance matrix is ​​initialized to a diagonal matrix to improve computational stability. The mixing coefficients are non-uniformly distributed based on the sample distribution density, completing the preliminary modeling of the probability density function.

[0025] During the EM algorithm's iterative process, the E-step first calculates the posterior probability that each eigenvector belongs to each Gaussian component. Subsequently, in the M-step, the mean, covariance, and mixing coefficient of each component are reestimated based on the probability weights. To prevent degradation of the covariance matrix, a regularization term is introduced to constrain its condition number. An early stopping mechanism terminates training when the change in the log-likelihood between consecutive iterations is less than a threshold or when the maximum number of iterations is reached. The converged Gaussian component parameter set is then output. This parameter set is used to solidify the GMM structure, partitioning historical feature data into training and validation sets to verify the model's ability to fit known normal events and ensure its generalization performance to unseen data. Furthermore, during the real-time analysis phase, the model receives joint feature vectors from the cross-modal fusion pipeline (e.g., multi-dimensional features fused from visual localization and vibration energy). A sliding window mechanism is used to partition the continuous time series data into fixed-length analysis units. The feature vectors within each window are fed into the trained GMM to calculate the Mahalanobis distance between them and the Gaussian components. When the minimum Mahalanobis distance of the feature vectors in a certain window exceeds the threshold, it is determined to be an abnormal intrusion event, and an early warning signal containing the abnormality type, confidence score and timestamp is generated. The multi-level response chain (such as sound and light alarms, drone inspection scheduling) is triggered simultaneously, and the verified abnormal samples are injected into the incremental learning module through the feedback loop to continuously optimize the model's adaptability to new intrusion patterns. Furthermore, in a preferred embodiment of the present invention, the intrusion behavior analysis model is constructed. When abnormal vibration is detected, abnormal hidden danger determination analysis is performed in combination with the abnormal vibration analysis information to obtain intrusion behavior analysis information and issue an early warning prompt, specifically including: When abnormal vibration is detected, it means that there may be a hidden danger of wall invasion in the abnormal vibration area, and abnormal vibration analysis information is obtained. According to the abnormal vibration analysis information, the abnormal area and abnormal period are traced to obtain regional abnormal tracing information. Obtain cross-modal fusion features, obtain the video surveillance device number in the abnormal area through the regional anomaly tracing information, and extract the video surveillance features of the corresponding numbered video surveillance device in the target abnormal period from the cross-modal fusion features in combination with the abnormal period tracing results to obtain the tracing video surveillance features; The invasion behavior characteristics of different organisms are obtained through big data retrieval, and the obtained invasion behavior characteristics are labeled with behavior categories to form an invasion behavior feature set. The invasion behavior analysis model is constructed using SVDD, and the model is trained based on the invasion behavior feature set. The input intrusion behavior features are mapped to a high-dimensional space using a Gaussian kernel function. The minimum hypersphere boundary containing all intrusion behavior feature samples is obtained in the mapped high-dimensional space and the trained intrusion behavior analysis model is output. Input the traceability video surveillance features into the trained intrusion behavior analysis model to generate a real-time feature vector, calculate the Mahalanobis distance of the real-time feature vector in the kernel space, and compare it with the preset hypersphere radius; If it is smaller than the preset hypersphere radius, it means there is a hidden danger of biological invasion. The cosine similarity between the real-time feature vector and the invasion behavior feature in the kernel space is calculated, and the abnormal behavior category corresponding to the invasion behavior feature with the largest similarity is selected as the belonging category. The invasion behavior analysis information is output and the early warning measures are obtained in the preset strategy library to perform abnormal invasion warning.

[0026] It should be noted that after the monitoring system determines that abnormal vibrations are occurring in a certain area, it first activates the abnormality tracing mechanism. Combining the time-frequency characteristics of the vibration signal (such as the time point of energy mutation and frequency distribution offset) with the coordinates of the geofence, it determines the physical location and corresponding time period of the abnormal vibrations. This generates tracing information, including the fence section number and start / end timestamps. Based on this tracing information, the system retrieves the video stream data from the video surveillance device with the corresponding area number (such as the dual-spectrum dome camera deployed in fence section 3) during the abnormal period from the cross-modal fusion feature library. The system extracts the spatiotemporal features of all video frames within this period (including moving target trajectories, heat source distribution, and contour change rate) to construct a tracing video surveillance feature vector set. At the same time, a big data platform was used to collect typical intrusion behavior samples from various organisms (such as humans and wild animals). Keyframes were captured and features were annotated (joint angle sequences for climbing movements, boundary-crossing speed curves, and limb swing spectra) from the sample videos. This generated an intrusion behavior feature set containing behavioral category labels. The Support Vector Data Description (SVDD) algorithm was then used to construct an intrusion behavior analysis model. The original features were mapped to a high-dimensional reproducing kernel Hilbert space using a Gaussian kernel function. The boundary of the smallest hypersphere that enclosed all normal behavior samples was determined. During training, the Sequential Minimization (SMO) algorithm was used to accelerate the calculation of the hypersphere radius and center position. The feature vectors of the traced video were mapped to the kernel space, and the normalized Mahalanobis distance to the center of the hypersphere was calculated. If the distance was less than the hypersphere radius, the behavior was identified as a biological intrusion. Cosine similarity matching was then used to determine the specific category (e.g., human climbing, animal collision). The model then linked to the hierarchical response rules in the pre-set policy library (human intrusion triggers an audible and visual alarm and sends the coordinates to the security terminal, while animal activity triggers a repellent device). The behavioral characteristics and response results were simultaneously recorded in a self-learning knowledge base, enabling the model to continuously adapt to new intrusion patterns.

[0027] Figure 2 A flowchart of an auxiliary method for eliminating abnormal hidden dangers in a production operation defense area provided by one embodiment of the present invention; like Figure 2 As shown, the present invention provides a flow chart of an auxiliary method for eliminating abnormal hidden dangers in a production operation defense area, comprising: S202, obtaining intrusion behavior analysis information, determining the real-time invasion status of the target intruding organism based on the intrusion behavior analysis information, and if the invasion is not complete, obtaining the real-time location of patrol personnel, and sending the location of the intrusion area to the nearest patrol personnel for guidance; S204: If the target intruder is in a complete invasion state, the invasion trajectory of the target intruder is tracked, and the intrusion video monitoring features when the target abnormal intruder is detected invading the defense area are extracted through cross-modal feature fusion; S206, importing the intrusion video monitoring features into the extended Kalman filter to estimate the motion state of the target abnormal intruder, calculating the motion state transfer matrix of the abnormal intruder through first-order approximation, and generating a predicted motion state according to the motion state transfer matrix; S208, introducing a Bayesian inference network, obtaining the observed motion state of abnormal invading organisms from the invasion video monitoring features through a motion state estimation algorithm, and using the observed motion state of the abnormal invading organisms as input to the Bayesian inference network for network training and obtaining a transient posterior distribution; S210, inferring the next-moment motion state of the abnormal invading organism in a Bayesian inference network using the transient posterior distribution to generate a plurality of inferred motion states, and performing difference calculation between the inferred motion state with the highest inference probability value and the predicted motion state; S212, obtaining a difference degree through difference calculation and comparing it with a preset difference degree threshold, if the difference degree is greater than the preset threshold, correcting the predicted motion state using the corresponding inferred motion state to obtain a corrected predicted motion state; S214: Generate an invasion tracking trajectory of the target abnormal invading organism based on the corrected predicted operating state, and generate a trajectory guidance report to send to nearby patrol personnel for expulsion assistance.

[0028] It should be noted that the system first constructs a set of dynamic motion parameters (including position, velocity, and acceleration) for the intruder based on intrusion behavior analysis information and the coordinates of the target detection box in the real-time video stream. If the intrusion status is incomplete (e.g., wandering outside the fenced area), Beidou / GPS dual-mode positioning data deployed on patrol terminals is used to calculate the real-time distance between the intrusion area and each patrol officer using the Dijkstra algorithm. Alert information containing geofence coordinates is prioritized and pushed to the patrol terminal with the shortest path. An AR navigation guidance layer is overlaid to facilitate rapid response. For targets in a fully intruded state, cross-modal fusion features (such as thermal imaging trajectory and voiceprint azimuth) are extracted during the abnormal period and fed into an extended Kalman filter for nonlinear motion modeling. The target's motion equation is linearized using the Jacobian matrix to construct a state vector containing position, velocity, and heading angle. The Kalman gain is calculated based on the observation data output by the video analysis module. The state covariance matrix is ​​iteratively updated to reduce the coupling effect of process and observation noise, and the predicted motion trajectory of the target in the two-dimensional plane is output. A Bayesian inference network is also introduced to encode video monitoring features (motion optical flow field and pose estimation parameters) into observation likelihood probabilities. Monte Carlo sampling is performed on the target motion prior distribution (e.g., maximum walking speed for people and turning angle limits for animals) to generate multiple candidate motion state hypotheses and their posterior probability distributions. The KL divergence is used to measure the distributional difference between the predicted state and the Bayesian inference result. When the difference exceeds a dynamic threshold, the Kalman filter state vector is weightedly corrected using the maximum a posteriori probability hypothesis to eliminate trajectory errors caused by target occlusion or sensor drift. The corrected motion state is then interpolated using cubic spline interpolation to generate a smooth intrusion trajectory curve. High-risk nodes (e.g., path turning points near oil tank areas) are automatically marked. A tactical map with time-space guidance is generated and transmitted to the patrol personnel's smart terminal, forming a closed-loop security response system from trajectory prediction to on-site disposal, preventing intruders from causing damage to the protected area.

[0029] Figure 3 An embodiment of the present invention provides a system 3 for optimizing abnormal hidden danger determination in a production operation defense zone. The system includes a memory 31 and a processor 32. The memory 31 contains a program for optimizing abnormal hidden danger determination in a production operation defense zone. When the program is executed by the processor 32, the following steps are implemented: Monitor the target area to obtain target area monitoring information and perform data preprocessing. Perform heterogeneous feature extraction and fusion based on the preprocessed target area monitoring data to generate cross-modal fusion features. Constructing an abnormal vibration analysis model, inputting the cross-modal fusion features into the abnormal vibration analysis model for analysis, determining whether there is an abnormality in the defense area wall, and obtaining abnormal vibration analysis information; Construct an intrusion behavior analysis model. When abnormal vibration is detected, the model is combined with the abnormal vibration analysis information to perform abnormal hidden danger analysis, obtain intrusion behavior analysis information and issue early warning prompts. When abnormal invading creatures fail to stop their invading behavior according to the early warning prompts, the invasion trajectory of the target abnormal invading creatures is tracked, and a trajectory guidance report is generated and sent to nearby patrol personnel for expulsion assistance.

[0030] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0031] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed across multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the scheme of this embodiment.

[0032] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0033] Those skilled in the art will appreciate that all or part of the steps of the above-mentioned method embodiments may be implemented by hardware associated with program instructions, and the aforementioned program may be stored in a computer-readable storage medium. When the program is executed, the program executes the steps of the above-mentioned method embodiments. The aforementioned storage medium includes various media that can store program codes, such as mobile storage devices, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0034] Alternatively, if the integrated units described above are implemented as software modules and sold or used as standalone products, they can also be stored on a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of the present invention, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product, stored on a storage medium, includes instructions for enabling a computer device (such as a personal computer, server, or network device) to execute all or part of the methods described in various embodiments of the present invention. The aforementioned storage media include various media capable of storing program code, such as removable storage devices, ROM, RAM, magnetic disks, or optical disks.

[0035] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for optimizing abnormal hidden danger determination in a production operation defense area, characterized in that: include: Monitor the target area to obtain target area monitoring information and perform data preprocessing. Perform heterogeneous feature extraction and fusion based on the preprocessed target area monitoring data to generate cross-modal fusion features. Constructing an abnormal vibration analysis model, inputting the cross-modal fusion features into the abnormal vibration analysis model for analysis, determining whether there is an abnormality in the defense area wall, and obtaining abnormal vibration analysis information; Construct an intrusion behavior analysis model. When abnormal vibration is detected, the model is combined with the abnormal vibration analysis information to perform abnormal hidden danger analysis, obtain intrusion behavior analysis information and issue early warning prompts. When abnormal invading creatures fail to stop their invading behavior according to the early warning prompts, the invasion trajectory of the target abnormal invading creatures is tracked, and a trajectory guidance report is generated and sent to nearby patrol personnel for expulsion assistance.

2. The method for optimizing abnormal hidden danger determination in a production operation defense area according to claim 1, characterized in that: The monitoring of the target area to obtain target area monitoring information and performing data preprocessing specifically includes: Based on the preset deployment plan, a heterogeneous sensor network is deployed in the target area, and the target area is monitored around the clock through the deployed heterogeneous sensor network to obtain monitoring information of the target area; Sending the target area monitoring information to the edge computing node for preprocessing according to a preset transmission protocol, wherein the target area monitoring information includes video monitoring data and vibration monitoring data; The cubic spline interpolation algorithm is used to reconstruct the intermediate moment image of the video monitoring data. The time interval between two adjacent frames is divided into several equal points and the pixel motion vector field is obtained. The interpolation weight is calculated based on the pixel motion vector field. The image is reconstructed according to the interpolation weight to eliminate the timing deviation between the video data and the sensor. The vibration monitoring data is decomposed into wavelet packets, and the db4 wavelet basis function is used to divide the original waveform into sub-band frequency domains. The environmental noise is eliminated by the soft threshold method. After the preprocessing operation is completed, the preprocessed target area monitoring information is pushed to the next processing unit through the lightweight message queue built into the edge computing node.

3. The method for optimizing abnormal hidden danger determination in a production operation defense area according to claim 1, characterized in that: The heterogeneous feature extraction and fusion are performed based on the pre-processed target area monitoring data to generate cross-modal fusion features, specifically including: The next processing unit obtains the pre-processed target area monitoring information transmitted by the edge computing node through a preset data receiving module, and imports the pre-processed target area monitoring information into a preset parallel feature extraction channel for processing. The parallel feature extraction channel includes a video feature extraction channel and a vibration feature extraction channel; The video monitoring data is input into the video feature extraction pipeline built on a lightweight convolutional neural network framework. The input image is processed frame by frame using a depthwise separable convolutional layer. A coordinate attention mechanism is introduced in the bottleneck layer to dynamically focus on abnormal areas and the contours of moving targets, generating a multi-scale visual feature map. The vibration monitoring data is input into a vibration feature extraction pipeline built on a recurrent neural network framework. The forward branch captures the vibration patterns of the historical time window, and the backward branch models future potential correlations. The bidirectional hidden states are fused through a gating mechanism to output a time-series vibration feature vector. The visual feature map is compressed into a region descriptor through spatial pyramid pooling, and tensor splicing and dimension alignment are performed with the temporal vibration feature vector. A multi-head cross-attention mechanism is introduced to perform cross-modal feature fusion; Taking temporal vibration features as query vectors and visual features as key-value pairs, the spatiotemporal correlation matrix of vibration-visual features is calculated through dot product attention, and cross-modal fusion features are generated after weighted aggregation.

4. The method for optimizing abnormal hidden danger determination in a production operation defense area according to claim 1, characterized in that: The abnormal vibration analysis model is constructed, and the cross-modal fusion feature is input into the abnormal vibration analysis model for analysis to determine whether there is an abnormality in the defense area wall, thereby obtaining abnormal vibration analysis information, specifically including: Introducing a big data network, obtaining historical vibration monitoring data of the defense zone walls during normal deployment through the big data network, performing feature extraction on the historical vibration monitoring data, and obtaining historical vibration feature information; An abnormal vibration analysis model is built based on a dynamic Gaussian mixture model. The expectation maximization algorithm is used to optimize the parameters of the abnormal vibration analysis model. The initial mean vector, covariance matrix and mixing coefficient of the Gaussian component are generated by randomly selecting k samples, and the initial probability density is assigned. During the iterative optimization process, the posterior probability of each eigenvector belonging to each Gaussian component is calculated, and the mean, covariance and mixing coefficient of the Gaussian component are re-estimated based on the posterior probability to obtain the model parameter set output after the iteration stops; Optimizing the abnormal vibration analysis model using the model parameter set, constructing a training data set based on the historical vibration feature information, and importing the data set into the optimized abnormal vibration analysis model for training; Obtaining cross-modal fusion features, inputting the cross-modal fusion features into the trained abnormal vibration analysis model, and segmenting the cross-modal fusion features in chronological order using a sliding window mechanism; The Mahalanobis distance between the vibration feature vector and all Gaussian components in each window is calculated, and the calculated Mahalanobis distance is compared with the preset threshold. If it is greater than the preset threshold, it means that there is a hidden danger of abnormal intrusion, and abnormal vibration analysis information is generated and an early warning is issued.

5. The method for optimizing abnormal hidden danger determination in a production operation defense area according to claim 1, characterized in that: The intrusion behavior analysis model is constructed, and when abnormal vibration is detected, abnormal hidden danger determination analysis is performed in combination with the abnormal vibration analysis information, intrusion behavior analysis information is obtained, and early warning prompts are issued, specifically including: When abnormal vibration is detected, it means that there may be a hidden danger of wall invasion in the abnormal vibration area, and abnormal vibration analysis information is obtained. According to the abnormal vibration analysis information, the abnormal area and abnormal period are traced to obtain regional abnormal tracing information. Obtain cross-modal fusion features, obtain the video surveillance device number in the abnormal area through the regional anomaly tracing information, and extract the video surveillance features of the corresponding numbered video surveillance device in the target abnormal period from the cross-modal fusion features in combination with the abnormal period tracing results to obtain the tracing video surveillance features; The invasion behavior characteristics of different organisms are obtained through big data retrieval, and the obtained invasion behavior characteristics are labeled with behavior categories to form an invasion behavior feature set. The invasion behavior analysis model is constructed using SVDD, and the model is trained based on the invasion behavior feature set. The input intrusion behavior features are mapped to a high-dimensional space using a Gaussian kernel function. The minimum hypersphere boundary containing all intrusion behavior feature samples is obtained in the mapped high-dimensional space and the trained intrusion behavior analysis model is output. Input the traceability video surveillance features into the trained intrusion behavior analysis model to generate a real-time feature vector, calculate the Mahalanobis distance of the real-time feature vector in the kernel space, and compare it with the preset hypersphere radius; If it is smaller than the preset hypersphere radius, it means there is a hidden danger of biological invasion. The cosine similarity between the real-time feature vector and the invasion behavior feature in the kernel space is calculated, and the abnormal behavior category corresponding to the invasion behavior feature with the largest similarity is selected as the belonging category. The invasion behavior analysis information is output and the early warning measures are obtained in the preset strategy library to perform abnormal invasion warning.

6. The method for optimizing abnormal hidden danger determination in a production operation defense area according to claim 1, characterized in that: When the abnormal intruder fails to stop its intrusion according to the warning prompt, the target abnormal intruder is tracked and a trajectory guidance report is generated and sent to nearby patrol personnel for expulsion assistance, specifically including: Obtaining intrusion behavior analysis information, determining the real-time invasion status of the target intruding organism based on the intrusion behavior analysis information, and if the invasion is not complete, obtaining the real-time location of patrol personnel, and sending the location of the intrusion area to the nearest patrol personnel for guidance; If the invasion is complete, the invasion trajectory of the target intruder is tracked, and the intrusion video monitoring features when the target abnormal intruder is detected in the defense area are extracted through cross-modal fusion features; The intrusion video monitoring features are imported into the extended Kalman filter to estimate the motion state of the target abnormal intruder, and the motion state transfer matrix of the abnormal intruder is calculated by first-order approximation, and the predicted motion state is generated according to the motion state transfer matrix; A Bayesian inference network is introduced to obtain the observed motion state of abnormal invading organisms from the invasion video monitoring features through the motion state estimation algorithm. The observed motion state is used as the input of the Bayesian inference network for network training and to obtain the transient posterior distribution. Inferring the next-moment motion state of the abnormal invading organism in the Bayesian inference network through the transient posterior distribution, generating a plurality of inferred motion states, and selecting the inferred motion state with the highest inference probability value to perform difference calculation with the predicted motion state; Obtain the degree of difference by difference calculation and compare it with the preset difference degree threshold. If the difference is greater than the preset threshold, the predicted motion state is corrected using the corresponding inferred motion state to obtain a corrected predicted operating state. The invasion tracking trajectory of the target abnormal invading organism is generated according to the corrected predicted operating status, and a trajectory guidance report is generated and sent to nearby patrol personnel for expulsion assistance.

7. A system for optimizing abnormal hidden danger judgment in production operation defense areas, characterized in that: The system includes: a memory and a processor. The memory contains a program for optimizing the abnormal hidden danger determination method for a production operation defense area. When the program is executed by the processor, the following steps are implemented: Monitor the target area to obtain target area monitoring information and perform data preprocessing. Perform heterogeneous feature extraction and fusion based on the preprocessed target area monitoring data to generate cross-modal fusion features. Constructing an abnormal vibration analysis model, inputting the cross-modal fusion features into the abnormal vibration analysis model for analysis, determining whether there is an abnormality in the defense area wall, and obtaining abnormal vibration analysis information; Construct an intrusion behavior analysis model. When abnormal vibration is detected, the model is combined with the abnormal vibration analysis information to perform abnormal hidden danger analysis, obtain intrusion behavior analysis information and issue early warning prompts. When abnormal invading creatures fail to stop their invading behavior according to the early warning prompts, the invasion trajectory of the target abnormal invading creatures is tracked, and a trajectory guidance report is generated and sent to nearby patrol personnel for expulsion assistance.

Images