Parking lot abnormal behavior processing method based on multi-view optimal transmission

Abstract

The present invention relates to the field of computer technology, and in particular to a method for processing abnormal behavior in a parking lot based on multi-view optimal transmission. First, multi-view data sets are established for abnormal behavior data and solution data, respectively, and a corresponding unified representation is obtained by establishing a degradation model. Then, a cost matrix is established through the idle state and feature similarity of the solution. Finally, an optimal match between the abnormal behavior and the solution is obtained by solving an optimal transmission problem. This method solves the defect of a single data source in existing methods, adopts comprehensive representation of multi-view data, overcomes the multi-point optimal matching problem in the abnormal behavior domain and the inspector domain, and effectively alleviates the cold start problem of the model.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method for processing abnormal parking behaviors based on multi-view optimal transmission. Background Art

[0002] The detection and handling of abnormal behavior in parking lots is a key issue concerning people's livelihood. On the one hand, abnormalities in parking lot equipment (such as failure to lift the barrier, barriers hitting cars, payment failures, etc.) will lead to a poor user experience and affect the basic order of the parking lot; on the other hand, abnormal disputes in parking lots (such as scratch accidents, etc.) pose a potential threat to public security management. Assigning the best solution to abnormal behavior (such as calling online customer service, notifying inspectors, etc.) is an effective means to avoid the above-mentioned risk issues. However, due to the complex optimal matching involving multiple devices, multiple types of abnormalities, and multiple solutions, the problem of abnormal behavior handling in large-scale parking lots has always been unresolved. Commonly used manual matching methods are flexible but costly and inefficient; some existing matching methods based on statistical methods or expert rules reduce labor costs, but have poor flexibility and low accuracy; some of the latest deep learning-based methods can effectively model large-scale abnormal behaviors, but are affected by network architecture and optimization strategies, and still have three inherent problems:

[0003] (1) Data from multiple sources cannot be effectively and comprehensively represented. Abnormal behavior monitoring in parking lot abnormal behavior processing systems usually relies on information collected from different terminals, stored in different forms, and has multi-source and multi-form attributes. These characteristics can be collectively referred to as multi-view attributes. Multi-view data includes two categories: multi-modality and multi-feature. For example, dispute anomalies are usually captured by monitoring equipment in parking lots and exist in the form of three-dimensional (3D, i.e. 2D image + 1D time frame) video and one-dimensional (1D) audio; license plate recognition anomalies usually exist in the form of captured 2D license plate images. In addition to different information storage forms (i.e. different modalities), data of the same modality also includes multiple features. For example, device anomaly information is usually recorded in the form of 1D text in the log file of the device terminal, and payment anomaly information comes from the result feedback on the server side and is also recorded in the form of 1D text, but they contain heterogeneous features and cannot be processed with the same strategy.

[0004] (2) Multi-point data in the abnormal behavior domain and the solution domain cannot be effectively matched. Another difficult problem in the parking lot abnormal behavior processing system is the optimal matching problem of multiple points between different domains, that is, how to assign an optimal solution to an abnormal behavior.

[0005] (3) Deep learning models, which require large amounts of training data, are difficult to deploy. Deep learning-based models typically require a large amount of supervisory information, necessitating the collection of historical data before actual use. This poses challenges for deploying the system in new environments. Achieving a cold start of the model and rapid iteration in actual use is also a challenge. Summary of the Invention

[0006] To address the aforementioned issues and address the shortcomings of existing methods, the present invention provides a method for handling abnormal parking behaviors based on multi-view optimal transmission. First, multi-view datasets (libraries) are established for abnormal behavior data and solution data, respectively. Corresponding unified features are obtained by establishing a degradation model. A cost matrix is then established based on the idle state and feature similarity of the solutions. Finally, an optimal transmission problem is solved to obtain the optimal match between abnormal behaviors and solutions. This method can be trained from scratch using expert correction and user feedback, allowing for rapid deployment in new environments.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] The present invention provides a method for handling abnormal parking behavior based on multi-view optimal transmission, comprising the following steps:

[0009] Abnormal behavior multi-view dataset construction: Multi-view data related to abnormal behavior is obtained from hardware terminals or sensors in the parking lot to construct an abnormal behavior multi-view dataset;

[0010] Solution: Multi-view dataset construction: Obtain inspector data to build a multi-view dataset;

[0011] Unified representation construction based on degradation model: Establish a degradation model and use it to learn unified representations for abnormal behaviors and solutions from corresponding multi-view datasets;

[0012] Cost matrix construction based on cosine similarity: The cost matrix is obtained by calculating the cosine similarity between the unified representation of the current abnormal behavior and the features in the existing solution, combined with the idle state of the current solution;

[0013] Solving the optimal transmission problem and updating parameters: Solve the optimal transmission problem using the Sinkhorn algorithm, and use the stochastic gradient descent algorithm to update the parameter model to obtain the optimal transmission matrix. The assignment result is obtained by calculating the smallest column in the optimal transmission matrix.

[0014] Usage and deployment of the matching system: Deployment is divided into two situations. One is: when it is deployed for the first time or the system processing satisfaction is less than a certain set threshold, manual assignment of the processing method is required (transferring to online customer service or assigning an inspector), and each fault is maintained in the form of multi-view data as a historical fault database. The other situation is: when the data entry is greater than the set threshold or the system processing satisfaction reaches the set threshold, the matching system directly assigns the processing method.

[0015] As a further technical solution of the present invention, the multi-view attributes contained in each abnormal behavior in the abnormal behavior multi-view dataset are: manually filled in basic fault description, parking lot video surveillance key frames, accident monitoring attributes, parking lot payment result feedback, parking lot gate equipment feedback, parking lot monitoring recording clips, vectorized fault processing results and vectorized fault processing feedback.

[0016] As a further technical solution of the present invention, each solution in the solution multi-view dataset includes multi-view attributes: position coordinates of the inspector, the current state vector of the inspector, and the processing capability of the inspector.

[0017] As a further technical solution of the present invention, the degradation model includes a view-specific latent feature learning module and a view-shared unified representation learning module. The view-specific latent feature learning module includes a perceptron structure of a single hidden layer, which is used to accept view data of different sizes and map them into latent feature features of the same dimension; the view-shared unified representation learning module includes two parts of trainable parameters. The first part is a degradation model g(), which is a single-layer fully connected layer including a nonlinear activation function ReLU(); the second part is a set of unified representation parameters Where N is the number of abnormal behaviors, d m is the length of the feature, and the unified representation parameter starts from initialization and is required to reconstruct the potential features of each view. The input data of the degradation learning module, where V represents the number of views, the goal of the degradation model learning is to minimize the following reconstruction loss function:

[0018]

[0019] where ||·||2 represents the l2 norm, f υ (.) denotes the latent feature learning module for the v-th view.

[0020] As a further technical solution of the present invention, the cost matrix Using the existing historical data information, a basic cost evaluation is established between the abnormal behavior characteristics and solutions to be processed, where n and m represent the number of abnormal behaviors and solutions respectively, and C i，jThe larger the value of , the higher the cost of assigning the i-th abnormal behavior to the j-th solution, that is, the less likely the abnormality should be assigned to the solution. The cost is related to the current real-time status of the inspector and whether the inspector is good at handling the abnormal behavior. The real-time status of the inspector is an attribute Ss in the solution data set. Whether the inspector is good at handling the abnormal behavior is determined by the abnormal behavior feature A to be processed. i and the historical feature attributes SA in the solution i The cosine distance D cos (A i , SA i ) means that in the cost matrix C, C i，j The value can be calculated as C i，j =|D cos (A i , SA i )-Ss|.

[0021] As a further technical solution of the present invention, the process of solving the optimal transmission problem is:

[0022] Assuming that the unified representation of abnormal behavior and solution obeys probability distributions p and q, respectively, the task of assigning solutions to abnormal behavior is modeled as the optimal transport (OT) problem with the following discrete distribution: Where <.,.> is the Frobenius point multiplication algorithm, is the cost matrix, is the final transition probability matrix that needs to be optimized and satisfies where a j and s i denote the jth abnormal behavior and the jth i The optimal transmission problem is solved to obtain the transition probability matrix. The column with the largest value in the optimized transition probability matrix for each abnormal behavior corresponds to the optimal solution, thus completing the assignment.

[0023] As a further technical solution of the present invention, the loss function used in the parameter updating process is obtained by combining the reconstruction loss used in constructing the degradation model and the optimal transmission loss used in solving the optimal transmission problem. Where λ is a balancing hyperparameter. The parameters are updated by minimizing the loss function through the random gradient descent algorithm, and the optimal degradation model and probability transfer matrix are obtained through training.

[0024] As a further technical solution of the present invention, the use and deployment process of the matching system is as follows:

[0025] The input of multiple abnormal behaviors to be processed is divided into the following situations:

[0026] Case 1: When the matching system is deployed in a new environment (i.e., the number of historical data entries contained in the database is less than the set threshold) or the matching system performance is poor (the average user feedback score of the historical data in the database is lower than the set threshold), the matching system cannot be effectively trained using the existing historical data. At this time, the process turns to an auxiliary judgment method with human participation, namely the manual assignment method.

[0027] Case 2: When both conditions in Case 1 are not met, the matching system can be effectively trained using existing historical data and switch to an automatic judgment approach, which includes the following steps:

[0028] 1) Step 1: Multiple pieces of abnormal behavior information to be processed are uniformly represented through the degradation model;

[0029] 2) Step 2: Obtain the cost matrix by calculating the cosine distance between the unified representation of the abnormal behavior to be processed and the unified representation in the historical data.

[0030] 3) Step 3: By solving the optimal transmission algorithm, the optimal probability transfer matrix is obtained, and the solution corresponding to the column with the maximum probability in the matrix is selected to obtain the final assignment result.

[0031] Once the final assignment result is obtained, the vectorized fault processing result in the abnormal behavior dataset is updated to the number of the assigned solution, and the real-time status value of the inspector of the assigned solution is updated to 1 in the solution dataset; when the abnormal behavior processing is completed, the vectorized fault processing feedback in the abnormal behavior dataset is updated, and the processing capability of the inspector in the solution dataset is updated; when no abnormal behavior is received, the model uses the existing historical data to update the parameters through the parameter update process until the next abnormal behavior information is input, and the above process is repeated.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] Based on the characteristics of parking lot hardware equipment and solutions (hotline transfer and inspectors), the present invention constructs a multi-view database, and realizes the assignment of abnormal behaviors in parking lots through degradation models, optimal transmission and other strategies. Compared with previous methods, the advantages of the present invention are: constructing a degradation model for learning multi-view unified representation and an optimal transmission solution for abnormal behavior processing. The present invention first constructs multi-view data sets for abnormal behavior and inspectors respectively, which solves the defect of single data source of existing methods; generates a unified representation from multi-view data through a degradation model to solve the problem that multi-view data cannot be comprehensively represented; overcomes the multi-point optimal matching problem of the abnormal behavior domain and the inspector domain by establishing and solving the optimal transmission problem; finally, through a two-branch deployment strategy based on a threshold, effectively alleviates the cold start problem of the deep model. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present disclosure and do not constitute improper limitations on the present invention.

[0035] Figure 1 A flow chart of a method for handling abnormal parking behavior based on multi-view optimal transmission provided by the present invention.

[0036] Figure 2 Schematic diagram of the process of learning a unified representation for the degradation model described in the present invention.

[0037] Figure 3 Schematic diagram of the optimal transmission model of the present invention. DETAILED DESCRIPTION

[0038] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0039] Example:

[0040] like Figure 1 As shown, the method and system for handling abnormal parking behaviors based on multi-view optimal transmission in this embodiment include the following steps:

[0041] S1: Construction of Abnormal Behavior Multi-view Dataset A

[0042] Define the multi-view attributes that each exception behavior should contain:

[0043] See Figure 1 : Manually fill in the basic description of the fault (1D text information, optional);

[0044] See Figure 2 : Parking lot video surveillance key frames (3D video information, 10 frames are retained by default);

[0045] See Figure 3 : Accident monitoring attributes (1D text information, including monitoring number, location information, etc.);

[0046] View 4: Parking lot payment result feedback (1D text information);

[0047] View 5: Parking lot barrier device feedback (1D text information);

[0048] Figure 6: Parking lot surveillance recording (1D voice information);

[0049] Figure 7: Vectorized fault handling results (1D vector, where value 0 indicates transfer to customer service, and values 1...K are inspector numbers).

[0050] Figure 8: Vectorized fault handling feedback (1D vector, between 0 and 1, where closer to 0 indicates worse incident handling and closer to 1 indicates better incident handling).

[0051] S2: Solution multi-view dataset S construction

[0052] Define the multi-view properties that each solution should contain:

[0053] See Figure 1 : The inspector's location coordinates (1D vector, for the "transfer to customer service" solution, this view attribute is 0);

[0054] See Figure 2 : The inspector's current state vector (1D vector, between 0 and 1, the closer to 0, the more idle the inspector is, and the closer to 1, the busier the inspector is. For the "transfer to customer service" solution, this view attribute is 0)

[0055] See Figure 3 : The inspector’s processing capability (1D vector, weighted average of the unified features of the abnormal behaviors processed by the inspector, where the weight is the view 8 in the abnormal behavior dataset - fault feedback score);

[0056] S3: Unified representation construction based on degradation model

[0057] The overall structure of the degradation model is as follows Figure 2 As shown, the entire model is divided into two parts according to the direction of data flow:

[0058] 1) View-specific latent feature learning module: The specific implementation form of this module is a perceptron structure with a single hidden layer. Its purpose is to accept view data of different sizes and map them into latent features of the same dimension. Since this module is view-specific, it is denoted by f v (.) represents the latent feature learning module of the υ-th view;

[0059] 2) View-sharing unified representation learning module: The unified learning module contains two parts of trainable parameters. The first part is a degradation model g(·), which is implemented as a single fully connected layer containing a nonlinear activation function ReLU(); the second part is a set of unified representation parameters Where N is the number of abnormal behaviors, d m is the length of the feature. The unified representation parameters are initialized and are required to be able to reconstruct the potential features of each view through the degradation model to ensure that it contains the comprehensive information of each view, that is, to meet the view completeness.

[0060] use The input data of the degradation learning module, where V represents the number of views. The goal of the degradation model learning is to minimize the following reconstruction loss function:

[0061]

[0062] where ||·||2 represents the l2 norm.

[0063] S4: Cost matrix construction based on cosine similarity

[0064] Cost Matrix The role of is to use the existing historical data information to establish a basic cost evaluation between the abnormal behavior characteristics and solutions to be processed, where n and m represent the number of abnormal behaviors and solutions respectively; C i，j The larger the value of , the higher the cost of assigning the i-th abnormal behavior to the j-th solution, that is, the less likely the abnormality should be assigned to the solution. Based on this analysis, the cost should be related to the current inspector's real-time status and whether the inspector is good at handling the abnormal behavior. The inspector's real-time status is an attribute Ss in the solution data set, and whether the inspector is good at handling the abnormal behavior can be determined by the abnormal behavior feature A to be processed. i and the historical feature attributes SA in the solution i The cosine distance D cos (A i , SA i ) means that, finally, in the cost matrix C, C i，j The value can be calculated as C i，j =|D cos (A i , SA i )-Ss|.

[0065] S5: Solving the optimal transmission problem and updating parameters

[0066] Through steps S3 and S4, we obtain a unified representation of the abnormal behavior to be processed, a unified representation of each solution, and a one-to-one cost matrix between the two sets. Assuming that the unified representation of the abnormal behavior and the solution obey the probability distribution p and q respectively, the task of assigning solutions to abnormal behaviors can be modeled as the following optimal transport (OT) problem with discrete distribution:

[0067]

[0068] Where <.,.> is the Frobenius point multiplication algorithm, is the cost matrix calculated in S4, is the final transition probability matrix that needs to be optimized and satisfies where aj and s i denote the jth abnormal behavior and the jth i The goal of the optimal transmission problem is to obtain a transition probability matrix by solving the OT problem. The column with the largest value in the optimized probability matrix for each abnormal behavior corresponds to the optimal solution, completing the assignment. The OT problem can be solved using the Sinkhorn algorithm.

[0069] When updating parameters, by combining the reconstruction loss used in constructing the degradation model in S3 and the optimal transmission loss used in solving the optimal transmission problem, the final loss function of the model is Where λ is a balancing hyperparameter. The parameters are updated by minimizing the loss function through random gradient descent, and the optimal degradation model and probability transfer matrix are obtained through training.

[0070] S6: Use and deployment of matching systems

[0071] In order to solve the problem that the deep learning model cannot be cold-started, this embodiment designs a human-machine combined training method. The complete flow chart is as follows: Figure 1 As shown, the dotted line represents the path of manual participation in auxiliary judgment, and the solid line represents the path of the proposed matching system performing judgment;

[0072] The input of multiple abnormal behaviors to be processed is divided into the following situations:

[0073] Case 1: When the matching system is deployed in a new environment (i.e., the number of historical data entries in the database is less than the set threshold) or the matching system performance is poor (the average user feedback score of the historical data in the database is lower than the set threshold), the matching system cannot be effectively trained using the existing historical data. In this case, the process turns to manual participation and auxiliary judgment;

[0074] Case 2: When both conditions in Case 1 are not met, the matching system can be effectively trained using existing historical data. At this point, the process shifts to the system's automatic judgment path, which includes the following steps:

[0075] Step 1: Multiple pieces of abnormal behavior information to be processed are uniformly represented through the degradation learning module;

[0076] Step 2: Obtain the cost matrix by calculating the cosine distance between the unified representation of the abnormal behavior to be processed and the unified representation in the historical data.

[0077] Step 3: By solving the optimal transmission algorithm, the optimal probability transfer matrix is obtained, and the solution corresponding to the column with the largest probability in the matrix is selected to obtain the final assignment result; after obtaining the final assignment result, the view 7 is updated in the abnormal behavior dataset as the number of the assigned solution, and the view of the assigned solution is updated in the solution dataset. Figure 2 The value of (real-time status) is 1. Once the abnormal behavior processing is completed, the view 8 (feedback score) in the abnormal behavior dataset is updated, and the view in the solution dataset is updated. Figure 3 (Processing capacity) When no abnormal behavior is received, the model is updated by using the existing historical data through the process described in S6 until the next abnormal behavior information is input, and the process described in S6 is repeated.

[0078] The network structure, algorithm and modeling process not described in detail in this article are common technologies in this field.

[0079] It should be noted that the purpose of disclosing the embodiments is to facilitate a further understanding of the present invention. However, those skilled in the art will appreciate that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the contents disclosed in the embodiments, and the scope of protection claimed by the present invention shall be determined by the scope defined in the claims.

Claims

1. A method for handling abnormal parking behaviors based on multi-view optimal transmission, characterized in that: The following steps are involved: Abnormal behavior multi-view dataset construction: Multi-view data related to abnormal behavior is obtained from the hardware terminals in the parking lot to construct an abnormal behavior multi-view dataset; Solution: Multi-view dataset construction: Obtain inspector data to build a multi-view dataset; Unified representation construction based on degradation model: A degradation model is established and used to learn unified representations for abnormal behavior and solutions from corresponding multi-view datasets. The degradation model includes a view-specific latent feature learning module and a view-shared unified representation learning module. The view-specific latent feature learning module contains a single hidden layer perceptron structure that accepts view data of different sizes and maps them into latent features of the same dimension. The unified representation learning module for view sharing contains two parts of trainable parameters. The first part is a degradation model g(·), which is a single-layer fully connected layer containing a nonlinear activation function ReLU(); the second part is a set of unified representation parameters Where N is the number of abnormal behaviors, d m is the length of the feature, and the unified representation parameter starts from initialization and is required to reconstruct the potential features of each view. Represents the input data of the degradation learning module, where V represents the number of views, and the goal of degradation model learning is to minimize the following reconstruction loss function: where ||·||2 represents the l2 norm, f v (·) represents the latent feature learning module of the v-th view; Cost matrix construction based on cosine similarity: The cost matrix is obtained by calculating the cosine similarity between the unified representation of the current abnormal behavior and the features in the existing solution, combined with the idle state of the current solution; Solving the optimal transmission problem and updating parameters: Solve the optimal transmission problem using the Sinkhorn algorithm, and use the stochastic gradient descent algorithm to update the parameter model to obtain the optimal transmission matrix. The assignment result is obtained by calculating the smallest column in the optimal transmission matrix. Usage and deployment of the matching system: Deployment is divided into two situations. When it is deployed for the first time or the system processing satisfaction is less than a certain threshold, manual assignment of the handling method is required, and each fault is maintained in the form of multi-view data as a historical fault database. When the data entry is greater than the set threshold or the system processing satisfaction reaches the set threshold, the matching system directly assigns the handling method.

2. The method for handling abnormal parking behaviors based on multi-view optimal transmission according to claim 1 is characterized in that: The multi-view attributes of each abnormal behavior in the abnormal behavior multi-view dataset are: manually filled-in basic fault description, parking lot video surveillance key frames, accident monitoring attributes, parking lot payment result feedback, parking lot gate equipment feedback, parking lot monitoring recording clips, vectorized fault handling results and vectorized fault handling feedback.

3. The method for handling abnormal parking behaviors based on multi-view optimal transmission according to claim 2 is characterized in that: Each solution in the solution multi-view dataset includes multi-view attributes: the position coordinates of the inspector, the current state vector of the inspector, and the processing capability of the inspector.

4. The method for handling abnormal parking behaviors based on multi-view optimal transmission according to claim 3 is characterized in that: The cost matrix Using the existing historical data information, a basic cost evaluation is established between the abnormal behavior characteristics and solutions to be processed, where n and m represent the number of abnormal behaviors and solutions respectively, and C i,j The larger the value of , the higher the cost of assigning the i-th abnormal behavior to the j-th solution, that is, the less likely the abnormality should be assigned to the solution. The cost is related to the current real-time status of the inspector and whether the inspector is good at handling the abnormal behavior. The real-time status of the inspector is an attribute Ss in the solution data set. Whether the inspector is good at handling the abnormal behavior is determined by the abnormal behavior feature A to be processed. i and the historical feature attributes SA in the solution i The cosine distance D cos (A i , SA i ) means that in the cost matrix C, C i,j The value is calculated as C i，j =|D cos (A i , SA i )-Ss|.

5. The method for handling abnormal parking behaviors based on multi-view optimal transmission according to claim 4 is characterized in that: The process of solving the optimal transmission problem is: Assuming that the unified representation of abnormal behavior and solution obeys probability distributions p and q, respectively, the task of assigning solutions to abnormal behavior is modeled as the optimal transport OT problem with the following discrete distribution: where <·,·> is the Frobenius point multiplication algorithm, is the cost matrix, is the final transition probability matrix that needs to be optimized and satisfies where a j and s i They represent the j-th abnormal behavior and the i-th solution feature respectively. The transition probability matrix is obtained by solving the optimal transmission problem. The column with the largest value in the optimized transition probability matrix for each abnormal behavior corresponds to the optimal solution, completing the assignment.

6. The method for handling abnormal parking behaviors based on multi-view optimal transmission according to claim 5 is characterized in that: The loss function used in the parameter update process is obtained by combining the reconstruction loss used in constructing the degradation model and the optimal transmission loss used in solving the optimal transmission problem. Where λ is a balancing hyperparameter. The parameters are updated by minimizing the loss function through the random gradient descent algorithm, and the optimal degradation model and probability transfer matrix are obtained through training.

7. The method for handling abnormal parking behaviors based on multi-view optimal transmission according to claim 6 is characterized in that: The usage and deployment process of the matching system is as follows: The input of multiple abnormal behaviors to be processed is divided into the following situations: Case 1: When the matching system is deployed in a new environment (i.e., the number of historical data entries in the database is less than the set threshold), or when the matching system performance is poor (i.e., the average user feedback score of the historical data in the database is lower than the set threshold), the matching system cannot be effectively trained using the existing historical data. In this case, the process switches to a manual auxiliary judgment method (i.e., manual assignment). Case 2: When both conditions in Case 1 are not met, the matching system can be effectively trained using existing historical data and switch to an automatic judgment approach, which includes the following steps: Step 1: Multiple pieces of abnormal behavior information to be processed are uniformly represented through the degradation model; Step 2: Obtain the cost matrix by calculating the cosine distance between the unified representation of the abnormal behavior to be processed and the unified representation in the historical data; By solving the optimal transmission algorithm, the optimal probability transfer matrix is obtained, and the solution corresponding to the column with the maximum probability in the matrix is selected to obtain the final assignment result; After obtaining the final assignment result, the vectorized fault processing result in the abnormal behavior dataset is updated with the number of the assigned solution, and the real-time status value of the inspector of the assigned solution is updated to 1 in the solution dataset. When the abnormal behavior processing is completed, the vectorized fault processing feedback in the abnormal behavior dataset is updated, and the processing capability of the inspector in the solution dataset is updated. When no abnormal behavior is received, the model is updated through the parameter update process using the existing historical data until the next abnormal behavior information is input, and the above process is repeated.

Images