Non-custodial control data analysis method and control platform

Abstract

The present application relates to the field of data processing, and specifically provides a non-custodial control data analysis method and control platform, which extracts features from each subset of non-custodial monitoring data to obtain the corresponding overall vector representation and detail vector representation, and simplifies according to the similarity of the vector space to obtain an overall vector representation combination and a detail vector representation combination of a set capacity; after fusing the overall search vector representation and detail search vector representation obtained in advance, extract the corresponding internal weight focus vector representation through an internal weight focus strategy; extract the first cross-domain weight focus vector representation and the second cross-domain weight focus vector representation through a cross-domain weight focus strategy; determine the non-custodial performance recognition result based on each first cross-domain weight focus vector representation and each second cross-domain weight focus vector representation. The present application can improve the accuracy and efficiency of the recognition of non-custodial performance recognition results.

Description

Technical Field

[0001] The present application relates to the field of data processing, and specifically, to a non-custodial control data analysis method and control platform. Background Art

[0002] At present, the adoption of non-custodial measures for suspects with minor crimes and little social harm is an important and positive judicial practice. With the development of the Internet of Things, big data and artificial intelligence technologies, intelligent monitoring of non-custodial persons has become possible. At present, it is difficult to extract valuable information from a large amount of monitoring data in the analysis and evaluation of non-custodial monitoring data of non-custodial persons, especially in long-term monitoring scenarios. The processing of monitoring data over a long period of time is limited by the backward data extraction logic and computing power cost considerations, and the confidence of the analysis results is usually not high. Therefore, there is an urgent need for an efficient, accurate and intelligent non-custodial control and analysis method. Summary of the invention

[0003] The purpose of the present invention is to provide a non-custodial control data analysis method and control platform. This application is implemented as follows:

[0004] In a first aspect, the present application provides a method for analyzing non-custodial control data, the method comprising: according to a non-custodial control analysis algorithm, performing feature extraction on each non-custodial monitoring data subset in a target non-custodial monitoring data set to obtain a corresponding overall vector representation and a detailed vector representation, and according to vector space similarity, respectively simplifying the obtained overall vector representation set and the detailed vector representation set to obtain an overall vector representation combination and a detailed vector representation combination of a set capacity; after fusing the pre-learned overall search vector representation and the detailed search vector representation, extracting the corresponding internal weighted focusing vector representation through an internal weighted focusing strategy; the overall search vector representation and the detailed search vector representation are based on non-custodial control analysis algorithm, to obtain a corresponding overall vector representation and a detailed vector representation; The method is obtained by adjusting the corresponding adaptive search vector representation based on the detention monitoring training data set; the internal weighted focusing vector representation integrates the core information of the overall search vector representation and the detailed search vector representation; through the cross-domain weighted focusing strategy, the first cross-domain weighted focusing vector representation between the internal weighted focusing vector representation and each overall vector representation in the overall vector representation combination is extracted respectively, and the second cross-domain weighted focusing vector representation between the internal weighted focusing vector representation and each detail vector representation in the detail vector representation combination is extracted respectively; based on the obtained first cross-domain weighted focusing vector representations and the second cross-domain weighted focusing vector representations, the non-detention performance recognition result of the target non-detention monitoring data set is determined.

[0005] In a second aspect, the present application provides a management and control system, comprising: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method as described above is implemented.

[0006] When extracting the monitoring data vector representation of the target non-detention monitoring data set based on the non-detention control analysis algorithm, the present application does not need to sample and extract the data in the data set. Instead, it uses feature extraction for each non-detention monitoring data subset in the target non-detention monitoring data set to obtain a total vector representation set composed of the total vector representation of each non-detention monitoring data subset, and a detail vector representation set composed of the detail vector representation of each non-detention monitoring data subset. Based on this, according to the vector space similarity, the obtained total vector representation set and detail vector representation set are streamlined respectively, the total vector representation of the local non-detention monitoring data subset is vector-integrated, and the detail vector representation of the local non-detention monitoring data subset is vector-integrated to obtain the total vector representation combination and detail vector representation combination of the set capacity. Based on this, it can be ensured that the total vector representation combination and the detail vector representation combination can represent the main non-detention performance information of the target non-detention monitoring data set. At the same time, the core information in the overall search vector representation and the detail search vector representation is integrated based on the internal weight focusing strategy, and then streamlined based on the cross-domain weight focusing strategy to obtain more representative feature information. By adopting this application, it is possible to effectively simplify the monitoring data vector representation of monitoring data sets of different data capacities, and to simplify features to optimize the speed of non-custodial performance recognition while avoiding information loss, thereby helping to determine the non-custodial performance recognition results in terms of both accuracy and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in describing the embodiments of the present application are briefly introduced below.

[0008] Figure 1 It is a flow chart of a non-custodial control data analysis method provided in an embodiment of the present application.

[0009] Figure 2 It is a schematic diagram of the composition of a management and control system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0010] The following describes the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. The terms used in the implementation method part of the embodiments of the present application are only used to explain the specific embodiments of the present application, and are not intended to limit the present application.

[0011] The execution subject of the non-custodial control data analysis method in the embodiment of the present application is a control system, which may include but is not limited to a server, a personal computer, a laptop, a tablet computer, a smart phone, etc. The control system can be run alone to implement the present application, or it can be connected to the network and implement the present application through interactive operations with other control systems in the network. Among them, the network where the control system is located includes but is not limited to the Internet, a wide area network, a metropolitan area network, a local area network, a VPN network, etc.

[0012] like Figure 1 As shown, the method comprises the following steps:

[0013] Step 100: Based on the non-custodial control analysis algorithm, feature extraction is performed on each non-custodial monitoring data subset in the target non-custodial monitoring data set to obtain the corresponding overall vector representation and detail vector representation, and based on the vector space similarity, the obtained overall vector representation set and detail vector representation set are streamlined to obtain the overall vector representation combination and detail vector representation combination of the set capacity.

[0014] The target non-detained monitoring data set is a collection of real-time monitoring data collected by the control platform for non-detained suspects with minor involvement in the case and little social harm, using electronic terminal devices (electronic bracelets), non-detained codes (mobile phone software) and intelligent detention control platform. For example, the data set may contain multi-dimensional data such as the location information, activity trajectory, and communication records of a non-detained suspect in a certain period of time. Each non-detained monitoring data subset is the monitoring data of the non-detained suspect in a corresponding sub-time interval in the data set. Alternatively, the data set contains multi-dimensional data such as the location information, activity trajectory, and communication records of multiple non-detained suspects in a certain period of time. Each non-detained monitoring data subset is all the monitoring data corresponding to a non-detained suspect in the data set.

[0015] The non-custodial control analysis algorithm is a machine learning model. It can be a neural network model based on deep learning, such as convolutional neural network (CNN), recurrent neural network (RNN) or its variants (LSTM, GRU), etc., or a model based on the Transformer architecture. The control platform uses this algorithm to extract features from each subset of non-custodial monitoring data.

[0016] The purpose of feature extraction is to convert the original non-custodial surveillance data into a more representative and computable vector representation. The overall vector representation is a general description of the overall characteristics of a non-custodial surveillance data subset, which reflects the comprehensive performance characteristics of the suspect during the entire surveillance period. For example, for a non-custodial suspect's location monitoring data subset, the overall vector representation may contain information such as his average activity range and the distribution of main activity locations. The detailed vector representation is a description of more detailed features in the data subset, which can reflect the characteristics of the suspect in a specific time period or specific behavior. Continuing with the location monitoring data as an example, the detailed vector representation may contain information such as his specific activity trajectory within a certain day and the places where he stayed for a long time.

[0017] After obtaining the overall vector representation and detailed vector representation of each non-custodial monitoring data subset, the management and control platform will form an overall vector representation set and a detailed vector representation set respectively. The overall vector representation set is composed of the overall vector representations of all non-custodial monitoring data subsets, and the detailed vector representation set is composed of the detailed vector representations of all non-custodial monitoring data subsets. Next, the management and control platform simplifies these two sets based on the vector space similarity, which can be understood as compression at the feature level. Vector space similarity is an indicator to measure the distance or similarity between two vectors in the vector space. Feasible calculation methods include, for example, cosine similarity, Euclidean distance, etc. The calculation formula for cosine similarity is: ,in and are two vectors, is their dot product, and are their moduli respectively. The calculation formula of Euclidean distance is: , where A i and B i They are vectors and The i-th component of , where n is the dimension of the vector.

[0018] The set capacity is a threshold value pre-set by the control platform to control the number of vectors in the overall vector representation combination and the detailed vector representation combination. Through vector integration processing, the control platform merges the overall vector representation and the detailed vector representation of the local non-custodial surveillance data subset to reduce the number of vectors while retaining the main information in the data set. For example, if there are two non-custodial suspects whose overall vector representations are very similar in the vector space, the control platform can mean-fuse these two vectors to obtain a new overall vector representation, thereby reducing the number of vectors in the overall vector representation set.

[0019] In actual operation, for feature extraction, the control platform can use the convolution layer and pooling layer of the deep learning model to process the non-custodial monitoring data and convert the raw data into vector representation. For example, for image-based non-custodial monitoring data (such as the suspect's facial recognition image), the convolution layer of the convolutional neural network can be used to extract the features of the image, and then the features can be mapped to vector representation through the fully connected layer. For sequence-based non-custodial monitoring data (such as activity trajectory data), a recurrent neural network or Transformer architecture can be used for feature extraction. When performing vector integration processing, after each non-custodial monitoring data subset is embedded and mapped (that is, encoded), the overall vector representation of the non-custodial monitoring data subset is added to the overall vector representation cache, and the detail vector representation is added to the detail vector representation cache. When the number of overall vector representations contained in the overall vector representation cache is not less than the set capacity, the overall vector representations of multiple non-custodial monitoring data subsets in the overall vector representation cache are vector-integrated according to the similarity of the vector space, and the overall vector representation to be added is added to the overall vector representation cache. A similar processing method is also used for the detail vector representation cache. In this way, the control platform can ensure that the overall vector representation combination and the detailed vector representation combination can represent the main non-detention performance information of the target non-detention monitoring data set. At the same time, the streamlined vector representation combination can reduce the complexity of subsequent calculations and improve the speed and efficiency of non-detention performance recognition.

[0020] For example, suppose the control platform collects surveillance data of 100 non-custodial suspects to form a target non-custodial surveillance data set. The surveillance data of each non-custodial suspect constitutes a non-custodial surveillance data subset. The control platform uses the non-custodial control analysis algorithm based on the Transformer architecture to extract features from each data subset and obtain the overall vector representation and detail vector representation of each data subset. Then, these vectors are added to the overall vector representation set and the detail vector representation set respectively. Assuming that the capacity is set to 20, when the number of vectors in the overall vector representation set reaches 20, the control platform begins to integrate these vectors based on the similarity of the vector space. By calculating the cosine similarity, the control platform finds that there are two overall vector representations with very high similarity, so the two vectors are mean-fused to obtain a new overall vector representation, thereby reducing the number of vectors in the overall vector representation set. For the detail vector representation set, a similar processing method is also used, and finally the overall vector representation combination and detail vector representation combination of the set capacity are obtained.

[0021] Step 200: After fusing the pre-learned overall search vector representation and the detailed search vector representation, the corresponding internal weighted focused vector representation is extracted through the internal weighted focused strategy; the overall search vector representation and the detailed search vector representation are obtained by adjusting the corresponding adaptive search vector representations based on the non-custodial monitoring training data set; the internal weighted focused vector representation integrates the core information in the overall search vector representation and the detailed search vector representation.

[0022] As mentioned above, in the non-custodial control scenario, the control platform collects a large amount of non-custodial monitoring data using electronic terminal devices and non-custodial codes. In step 100, the target non-custodial monitoring data set has been feature extracted and the vector set has been streamlined, while step 200 focuses on further mining the core information in the data to prepare for subsequent cross-domain analysis and final non-custodial performance identification.

[0023] The overall search vector representation can be understood as the overall query vector, which reflects the overall characteristics and patterns of non-custodial monitoring data from a macro level, and can guide the control platform to focus on the global information of the data in subsequent analysis. For example, when monitoring the activities of non-custodial suspects, the overall search vector representation may contain information such as the average activity range of the suspect over a long period of time and the distribution of the main activity areas. This information helps the control platform to grasp the suspect's behavior pattern as a whole. The detailed search vector representation is the query vector of the details, which focuses on reflecting the local details and specific features in the data, and enables the control platform to focus on specific behaviors and events during the analysis process. For example, when monitoring the activities of a suspect on a certain day, the detailed search vector representation may contain information such as the specific activity trajectory of the suspect in a specific time period and the location where the suspect stayed for a long time. These detailed information is crucial for judging whether the suspect has abnormal behavior. These two search vector representations are not fixed, but are obtained by adjusting the corresponding adaptive search vector representations based on the non-custodial monitoring training data set. The adaptive search vector representation is learnable. The control platform continuously analyzes and learns the data in the non-custodial monitoring training data set, adjusts the parameters of the adaptive search vector representation, and makes it better adapt to the actual non-custodial monitoring scenario, thereby obtaining a more accurate overall search vector representation and detailed search vector representation. For example, the control platform can use optimization algorithms such as gradient descent to continuously update the parameters of the adaptive search vector representation based on the feedback of the training data, so that the overall search vector representation and detailed search vector representation can more accurately reflect the characteristics of the non-custodial monitoring data.

[0024] After obtaining the overall search vector representation and the detailed search vector representation, the control platform fuses them. The purpose of fusion is to organically combine global information and local information in order to more comprehensively understand the non-custodial monitoring data. One feasible fusion method is to concatenate the two vectors and then map the concatenated vector to a new vector space through a linear transformation layer. Assume that the overall search vector is represented as , the detail search vector is expressed as , the concatenated vector is , the weight matrix of the linear transformation layer is W, the bias vector is b, then the fused vector The formula can be Calculated.

[0025] Next, the control platform extracts the corresponding internal weighted focus vector representation from the fused vector through the internal weighted focus strategy. The internal weighted focus strategy can be implemented as a self-attention mechanism, which is a strategy that can automatically focus on information at different positions in the input sequence, assigning different weights to each element according to the correlation between elements in the input sequence, thereby highlighting important information. In non-custodial control scenarios, the internal weighted focus strategy can help the control platform filter out the core information in the overall search vector representation and the detailed search vector representation from the fused vector.

[0026] Specifically, the implementation of the internal weight focus strategy can be divided into the following steps. First, the control platform will merge the vector The query vector is obtained by three linear transformation layers respectively. , key vector Sum value vector ,Right now ,in are the weight matrices for query, key, and value respectively. Then, the similarity score between the query vector and the key vector is calculated. A feasible calculation method is the dot product operation to obtain the similarity score matrix In order to prevent the similarity score from being too large, the score matrix is ​​usually scaled, that is, ,in is the dimension of the key vector. Next, the scaled score matrix is ​​converted into a probability distribution matrix A=softmax(S') through the softmax function. Each element in the matrix represents the importance of the information at the corresponding position in the final output. Finally, the probability distribution matrix is ​​multiplied by the value vector to obtain the internal weight focus vector representation .

[0027] Through the above process, the internal weighted focus vector representation obtained by the control platform integrates the core information in the overall search vector representation and the detailed search vector representation. It can more effectively represent the key features of non-custodial monitoring data and provide more valuable information for subsequent cross-domain analysis and non-custodial performance identification. For example, when judging whether a non-custodial suspect has violated regulations, the internal weighted focus vector representation can highlight the key information of the suspect's abnormal activities, such as frequent entry and exit of specific areas, contact with suspicious persons, etc., thereby helping the control platform to make more accurate judgments.

[0028] In practical applications, the control platform can use deep learning frameworks (such as TensorFlow, PyTorch, etc.) to implement the vector fusion and self-attention mechanism in step 200. The control platform can use the overall search vector representation and the detailed search vector representation as input to build a corresponding neural network model, and train the parameters of the model to achieve vector fusion and core information extraction. At the same time, the control platform can also optimize and adjust the model according to actual conditions to improve the quality and accuracy of the internal weight focus vector representation.

[0029] Step 300: Through the cross-domain weight focusing strategy, extract the first cross-domain weight focusing vector representation between the internal weight focusing vector representation and each overall vector representation in the overall vector representation combination, and extract the second cross-domain weight focusing vector representation between the internal weight focusing vector representation and each detail vector representation in the detail vector representation combination.

[0030] In step 200, the management and control platform has obtained an internal weighted focusing vector representation that integrates the core information of the overall search vector representation and the detailed search vector representation through an internal weighted focusing strategy, while the overall vector representation combination and the detailed vector representation combination are obtained in step 100 by performing feature extraction and vector representation set simplification on each non-custodial monitoring data subset in the target non-custodial monitoring data set, and represent the overall characteristics and detailed characteristics of the non-custodial monitoring data, respectively.

[0031] The cross-domain weighted focus strategy can be implemented as a cross-attention mechanism, which is a mechanism that can capture the correlation information between different feature vectors. In the non-custodial control scenario, it can help the control platform analyze the relationship between the internal weighted focus vector representation and the overall vector representation and the detail vector representation, so as to discover the hidden patterns and laws in the non-custodial monitoring data. For example, the internal weighted focus vector representation may contain the key feature information of the behavior of non-custodial suspects, while the overall vector representation and the detail vector representation reflect the characteristics of the suspect's monitoring data from different levels. The cross-attention mechanism can find the connection between these features and determine whether the suspect has abnormal behavior.

[0032] In order to implement the cross-domain weight focus strategy, the management and control platform calculates the first cross-domain weight focus vector representation between the internal weight focus vector representation and each overall vector representation in the overall vector representation combination, and the second cross-domain weight focus vector representation between the internal weight focus vector representation and each detail vector representation in the detail vector representation combination. Specifically, the calculation process of the cross-domain weight focus strategy can be divided into the following steps. First, the management and control platform obtains the query vector, key vector and value vector through linear transformation of the internal weight focus vector representation, the overall vector representation combination and the detail vector representation combination. Assume that the internal weight focus vector is represented as , the total vector represents the i-th total vector in the combination is expressed as , the jth detail vector in the detail vector representation combination is represented as , then the query vector can be obtained through linear transformation , the key vector (for the overall vector representation) and (for detail vector representation), the value vector (for the overall vector representation) and (for detail vector representation), where are the corresponding weight matrices respectively.

[0033] Then, we calculate the similarity score between the query vector and the key vector. A feasible calculation method is the dot product operation. For the overall vector representation, the similarity score matrix is ; For detail vector representation, similarity score matrix In order to prevent the similarity score from being too large, the score matrix is ​​usually scaled, that is, and ,in are the key vector dimensions of the overall vector representation and the detail vector representation, respectively.

[0034] Next, the scaled score matrix is ​​converted into a probability distribution matrix through the softmax function. For the overall vector representation, the probability distribution matrix A G =softmax(S G '); For the detail vector representation, the probability distribution matrix A D =softmax(S D '). Each element in these probability distribution matrices represents the importance of the vector at the corresponding position in the final output.

[0035] Finally, the probability distribution matrix is ​​multiplied by the value vector to obtain the first cross-domain weighted focusing vector representation and the second cross-domain weighted focused vector representation These cross-domain weighted focused vector representations contain the association information between the internal weighted focused vector representation and the overall vector representation and the detail vector representation, and can more accurately reflect the characteristics of non-custodial monitoring data.

[0036] In practical applications, the management and control platform can use a deep learning framework to implement the cross-attention mechanism. For example, use TensorFlow or PyTorch to build a neural network model, take the internal weight focus vector representation, the overall vector representation combination, and the detail vector representation combination as input, and complete the above calculation process by defining the corresponding linear transformation layer and softmax layer. The management and control platform can also adjust the parameters of the weight matrix according to actual conditions to optimize the effect of the cross-attention mechanism.

[0037] Step 300 extracts the first cross-domain weight focusing vector representation and the second cross-domain weight focusing vector representation through the cross-domain weight focusing strategy, which can deeply mine the correlation information between different feature vectors in the non-custodial monitoring data, and provide a richer and more accurate basis for the subsequent determination of the non-custodial performance recognition results of the target non-custodial monitoring data set, which helps the management and control platform to manage and monitor non-custodial suspects more effectively.

[0038] Step 400: Determine the non-custodial performance recognition result of the target non-custodial monitoring data set based on the obtained first cross-domain weighted focusing vector representations and the second cross-domain weighted focusing vector representations.

[0039] In the previous steps, the control platform has performed a series of processing on the target non-custodial monitoring data set through the non-custodial control analysis algorithm. In step 100, feature extraction is performed on each non-custodial monitoring data subset in the target non-custodial monitoring data set to obtain an overall vector representation and a detailed vector representation, and the set is streamlined to obtain an overall vector representation combination and a detailed vector representation combination with a set capacity. In step 200, the overall search vector representation and the detailed search vector representation obtained in advance are fused, and the internal weighted focusing vector representation is extracted through the internal weighted focusing strategy. Step 300 uses a cross-domain weighted focusing strategy to extract the first cross-domain weighted focusing vector representation between the internal weighted focusing vector representation and each overall vector representation in the overall vector representation combination, and the second cross-domain weighted focusing vector representation between the internal weighted focusing vector representation and each detail vector representation in the detail vector representation combination.

[0040] The first cross-domain weighted focus vector representation reflects the association information between the internal weighted focus vector representation and the overall vector representation, and reflects the relationship between the global features and core information of non-custodial monitoring data. For example, when monitoring the range of activities of non-custodial suspects, the first cross-domain weighted focus vector representation may highlight the connection between the suspect's main activity area and the overall activity pattern during a specific time period. The second cross-domain weighted focus vector representation reflects the association information between the internal weighted focus vector representation and the detail vector representation, and focuses more on reflecting the association between local detail features and core information. For example, when monitoring the specific behavior of a suspect on a certain day, the second cross-domain weighted focus vector representation may show the correlation between the suspect's stay time at a specific location and the overall behavior pattern.

[0041] There are many ways for the control platform to determine the non-custodial performance recognition results of the target non-custodial monitoring data set. One way is to process based on multiple cascaded reduction mapping layers in the feature integration component. If this method is adopted, the control platform will determine the recognition result based on the first cross-domain weighted focus vector representation and the second cross-domain weighted focus vector representation output by the last reduction mapping layer. The reduction mapping layer plays a role in further processing and integrating feature information in the entire process. Through multiple cascaded reduction mapping layers, more representative and decision-making feature information can be gradually extracted. For example, the first reduction mapping layer may perform preliminary feature fusion and conversion on the input first cross-domain weighted focus vector representation and the second cross-domain weighted focus vector representation, and the subsequent reduction mapping layer performs more in-depth processing on this basis. Finally, the reduction mapping layer at the end outputs feature information that can accurately reflect the non-custodial performance, and the control platform judges the recognition result based on this information.

[0042] Another way is that when the non-custodial control analysis algorithm includes a domain-general algorithm, the control platform first fuses each first cross-domain weighted focus vector representation and each second cross-domain weighted focus vector representation, and maps them into a monitoring data vector representation that matches the non-custodial assessment vector representation dimension of the target non-custodial monitoring data set. The purpose of this step is to enable the processed feature vector to be compatible with the pre-set non-custodial assessment standard to facilitate subsequent accurate evaluation. For example, the fused vector is projected into the same dimensional space as the non-custodial assessment vector representation through linear transformation or non-linear mapping. Assume that the fused vector is , the weight matrix is ​​W, the bias vector is b, then the mapped monitoring data vector is represented by The formula can be Calculated.

[0043] Then, the control platform takes the non-custodial evaluation vector representation and non-custodial evaluation guidance information of the target non-custodial monitoring data set as input information, and inputs the domain-general algorithm in combination with the monitoring data vector representation to determine the non-custodial performance recognition result of the target non-custodial monitoring data set. The domain-general algorithm is a pre-trained and fine-tuned domain-large model with powerful feature learning and classification capabilities. The non-custodial evaluation vector representation is a quantitative description of non-custodial performance, which contains information on multiple evaluation indicators. The non-custodial evaluation guidance information is additional information used to guide the model to evaluate, such as the weights of different evaluation indicators, the standard range of evaluation, etc. By inputting this information into the domain-general algorithm, the model can classify and judge the non-custodial performance according to the patterns and rules it has learned. For example, the model may output recognition results such as whether the performance of a non-custodial suspect is "good", "average", or "risky".

[0044] In practical applications, the control platform can use a deep learning framework (such as TensorFlow, PyTorch, etc.) to implement the operation in step 400. For the method based on the restoration mapping layer, the control platform can build a corresponding neural network model, define multiple cascaded restoration mapping layers, and adjust the parameters of the model through training so that it can accurately output feature information for identifying non-custodial performance. For the method based on the domain-general algorithm, the control platform can load a pre-trained domain model and input the processed feature vectors and related information into the model for prediction and classification. At the same time, the control platform can also optimize and adjust the model according to the actual situation to improve the accuracy and reliability of the non-custodial performance recognition results.

[0045] As an implementation method, the non-custodial control analysis algorithm includes a feature integration component; step 300, through a cross-domain weighted focusing strategy, respectively extracting a first cross-domain weighted focusing vector representation between an internal weighted focusing vector representation and each overall vector representation in the overall vector representation combination, and respectively extracting a second cross-domain weighted focusing vector representation between an internal weighted focusing vector representation and each detail vector representation in the detail vector representation combination, including:

[0046] Step 310: decomposing the internal weighted focused vector representation into a first internal weighted focused vector representation corresponding to the overall vector representation and a second internal weighted focused vector representation corresponding to the detail vector representation;

[0047] Step 320: Based on the first cross-domain weighted focusing layer in the feature integration component and according to the cross-domain weighted focusing strategy, respectively extract the first internal weighted focusing vector representation and the first cross-domain weighted focusing vector representation between each overall vector representation in the overall vector representation combination;

[0048] Step 330: Based on the second cross-domain weight focusing layer in the feature integration component and according to the cross-domain weight focusing strategy, extract the second internal weight focusing vector representation and the second cross-domain weight focusing vector representation between each detail vector representation in the detail vector representation combination.

[0049] In step 310, the management and control platform decomposes the internal weighted focus vector representation into a first internal weighted focus vector representation corresponding to the overall vector representation, and a second internal weighted focus vector representation corresponding to the detail vector representation. The internal weighted focus vector representation is extracted by fusing the overall search vector representation and the detail search vector representation obtained in advance in step 200 using the internal weighted focus strategy, and it integrates the core information in the overall search vector representation and the detail search vector representation. The decomposition into the first internal weighted focus vector representation and the second internal weighted focus vector representation is to enable subsequent targeted cross-domain weighted focusing operations to be performed in combination with the overall vector representation and the detail vector representation, respectively.

[0050] The control platform can achieve this decomposition through linear transformation. Assume that the internal weight focus vector is expressed as , for the first internal weighted focusing vector representation , can be obtained through the formula Calculated, where is a weight matrix, is the bias vector; similarly, for the second internal weight focusing vector, , can be obtained through the formula Calculated, are the corresponding weight matrices and bias vectors, respectively. These weight matrices and bias vectors are optimized and adjusted during the training process of the non-custodial control analysis algorithm so that the decomposed vectors can better reflect the correlation with the overall vector representation and the detailed vector representation.

[0051] For example, in the monitoring scenario of non-detained persons, the internal weighted focus vector representation may contain comprehensive behavioral characteristics of non-detained persons, such as activity time patterns, activity range, frequency of contact with others, etc. When decomposed, the first internal weighted focus vector representation may focus more on global features related to the overall vector representation, such as the average activity range of non-detained persons over a long period of time, the distribution of major activity areas, etc.; while the second internal weighted focus vector representation pays more attention to local features related to the detail vector representation, such as the specific activity trajectory of non-detained persons in a specific time period of a day, and the places where they stay for a long time.

[0052] In step 320, the control platform extracts the first cross-domain weighted focusing vector representation between the first internal weighted focusing vector representation and each overall vector representation in the overall vector representation combination based on the first cross-domain weighted focusing layer in the feature integration component and the cross-domain weighted focusing strategy. The feature integration component plays an important role in information integration and processing in the entire non-custodial control analysis algorithm, and the first cross-domain weighted focusing layer is specifically used to process the association information between the first internal weighted focusing vector representation and the overall vector representation.

[0053] The cross-domain weighted focus strategy is essentially a cross-attention mechanism that can capture the relationship between different feature vectors. Each overall vector in the combination of and overall vector represents (i represents the index of the overall vector representation) and obtain the query vectors through linear transformation , key vector Sum value vector ,in are the corresponding weight matrices respectively. Then the similarity score between the query vector and the key vector is calculated, and the similarity score matrix can be obtained by the feasible dot product operation To prevent the similarity score from being too large, the score matrix is ​​scaled, that is, ,in is the dimension of the key vector. Then the scaled score matrix is ​​converted into a probability distribution matrix through the softmax function , each element in the matrix represents the importance of the corresponding overall vector representation in the final output. Finally, the probability distribution matrix is ​​multiplied by the value vector to obtain the first cross-domain weighted focus vector representation .

[0054] Taking the monitoring of non-detained persons as an example, the overall vector representation combination may contain the overall activity characteristics of multiple non-detained persons over a period of time, such as the average activity radius of each person, the concentration of activity locations, etc. The first internal weighted focus vector representation contains core information related to the overall situation. Through the calculation of the first cross-domain weighted focus layer, the management and control platform can obtain the degree of correlation between the overall vector representation of each non-detained person and the core information, that is, the first cross-domain weighted focus vector representation. For example, if the overall vector representation of a non-detained person shows that his activity radius has suddenly increased, and the first internal weighted focus vector representation contains information about possible abnormal activities in the near future, then the first cross-domain weighted focus vector representation calculated by the first cross-domain weighted focus layer will highlight the correlation between this abnormal situation and the core information, thereby attracting the attention of the management and control platform.

[0055] In step 330, the management and control platform extracts the second cross-domain weighted focusing vector representation between each detail vector representation in the combination of the second internal weighted focusing vector representation and the detail vector representation based on the second cross-domain weighted focusing layer in the feature integration component and the cross-domain weighted focusing strategy. The second cross-domain weighted focusing layer functions similarly to the first cross-domain weighted focusing layer, except that it processes the association information between the second internal weighted focusing vector representation and the detail vector representation.

[0056] The cross-attention mechanism is also adopted. The control platform focuses the second internal weight vector representation Each detail vector representation in the combination of (j represents the index of the detail vector) and the query vector is obtained by linear transformation. , key vector Sum value vector ,in are the corresponding weight matrices respectively. Calculate the similarity score between the query vector and the key vector to get the similarity score matrix , and then scale it up to get ,in is the dimension of the key vector. The scaled score matrix is ​​converted to a probability distribution matrix through the softmax function Finally, the probability distribution matrix is ​​multiplied by the value vector to obtain the second cross-domain weighted focusing vector representation .

[0057] In practical applications, the control platform can use deep learning frameworks (such as TensorFlow, PyTorch, etc.) to implement steps 310-330. For step 310, the control platform can define the corresponding linear transformation layer in the model and adjust the parameters of the weight matrix and bias vector through training. For steps 320 and 330, the control platform can construct a neural network module containing query, key, value linear transformation layers and softmax layers to realize the calculation of cross-attention mechanism. At the same time, the control platform can also optimize and adjust the model according to the actual non-custodial control needs, such as increasing the number of network layers, adjusting the learning rate, etc., to improve the accuracy and efficiency of extracting the first cross-domain weight focus vector representation and the second cross-domain weight focus vector representation.

[0058] As an embodiment, the feature integration component includes multiple cascaded restoration mapping layers; for the first restoration mapping layer, its execution data includes an overall vector representation combination and a detail vector representation combination; for other restoration mapping layers, the execution data includes each first cross-domain weighted focused vector representation and each second cross-domain weighted focused vector representation output by the previous restoration mapping layer.

[0059] Based on the above situation, step 400 determines the non-custodial performance recognition result of the target non-custodial monitoring data set according to the obtained first cross-domain weighted focused vector representations and the second cross-domain weighted focused vector representations, including:

[0060] Step 410: Determine the non-custodial performance recognition result of the target non-custodial monitoring data set based on the first cross-domain weighted focusing vector representations and the second cross-domain weighted focusing vector representations output by the last restoration mapping layer.

[0061] The restoration mapping layer in the feature integration component, also known as the decoding layer, is an important part of the entire analysis process. The execution data received by the first restoration mapping layer is the overall vector representation combination and the detail vector representation combination, while the subsequent restoration mapping layers use the first cross-domain weighted focused vector representations and the second cross-domain weighted focused vector representations output by the previous restoration mapping layer as execution data. This cascade structure enables feature information to be gradually processed and optimized in different restoration mapping layers, and each layer further mines and integrates information based on the previous layer, so that the final output feature information can more accurately reflect the non-custodial performance of the target non-custodial monitoring data set.

[0062] In the non-custodial control scenario, the first cross-domain weighted focused vector representation reflects the correlation information between the internal weighted focused vector representation and the overall vector representation, reflecting the relationship between the global features and core information of the non-custodial monitoring data; the second cross-domain weighted focused vector representation reflects the correlation information between the internal weighted focused vector representation and the detail vector representation, focusing more on the correlation between local detail features and core information. Through the processing of multiple restoration mapping layers, these correlation information has been continuously strengthened and purified, so that the first cross-domain weighted focused vector representation and the second cross-domain weighted focused vector representation output by the last restoration mapping layer can contain the most representative and decision-making features.

[0063] When determining the non-custodial performance recognition results of the target non-custodial monitoring data set, the control platform will conduct a comprehensive analysis of these vector representations output by the last restoration mapping layer. From a technical point of view, the control platform can use a classification algorithm to complete this task. For example, a feasible support vector machine (SVM) classification algorithm, its basic idea is to find an optimal hyperplane in the feature space to separate samples of different categories. Assume that the first cross-domain weighted focused vector representation and the second cross-domain weighted focused vector representation output by the last restoration mapping layer are combined into a new feature vector , the SVM algorithm will learn a classification function based on the training data ,in is the Lagrange multiplier, is the class label of the sample, is the kernel function, and b is the bias term. By calculating The value of can be used to determine the non-custodial performance category corresponding to the target non-custodial monitoring dataset.

[0064] Another example is the decision tree classification algorithm, which makes classification decisions by building a tree structure. The management and control platform will select appropriate features as nodes for division based on the different feature dimensions represented by the vector output by the last reduction mapping layer, and gradually build a decision tree. For example, the monitoring data of non-detained persons may be divided according to features such as activity range and activity time. If the activity range exceeds the specified area and the activity time is within the non-permitted time period, it can be determined that there is a risk of violation; otherwise, it is judged to be good performance.

[0065] Let's take a specific non-custodial control case as an example. Assume that the control platform monitors multiple non-custodial suspects, and obtains the first cross-domain weighted focus vector representation and the second cross-domain weighted focus vector representation through the previous steps, and then processes them through multiple restoration mapping layers. For one of the suspects, the first cross-domain weighted focus vector representation output by the last restoration mapping layer shows that the overall change trend of his recent activity range is significantly different from the normal situation, while the second cross-domain weighted focus vector representation shows that his stay time in certain specific places is obviously abnormal. The control platform uses a classification algorithm to analyze this information. If the SVM algorithm is used, after calculating the value of the classification function, it is determined that the suspect's non-custodial performance is risky; if the decision tree algorithm is used, the nodes in the decision tree are divided according to the characteristics such as activity range and stay time, and it is also concluded that the suspect is at risk of violation.

[0066] In addition to classification algorithms, the management and control platform can also use clustering algorithms to determine the non-custodial performance recognition results. The purpose of the clustering algorithm is to cluster similar samples into one category and different samples into different categories. For example, the K-means clustering algorithm will pre-set the number of clusters k, and then iteratively assign the vector representation output by the last reduction mapping layer to different cluster centers. Assume that k=3, representing the three categories of "good", "general", and "risky" respectively. The management and control platform will calculate the distance from each vector representation to each cluster center and assign it to the category with the nearest cluster center. In this way, the management and control platform can classify the target non-custodial monitoring data set to determine its non-custodial performance recognition results.

[0067] In practical applications, the control platform can use open source machine learning libraries (such as Scikit-learn) to implement the above classification and clustering algorithms. At the same time, in order to improve the accuracy of the recognition results, the control platform can also optimize and adjust the algorithm, such as adjusting the kernel function type and parameters of the SVM algorithm, the initial clustering center selection of the K-means clustering algorithm, etc. In addition, the control platform can also combine other information, such as the historical performance records and social relations of non-detained persons, to make a comprehensive judgment to ensure the reliability and effectiveness of the non-detained performance recognition results.

[0068] As another embodiment, the non-custodial control analysis algorithm includes a domain-general algorithm. Based on this, step 400 determines the non-custodial performance recognition result of the target non-custodial monitoring data set based on the obtained first cross-domain weighted focus vector representations and the second cross-domain weighted focus vector representations, including:

[0069] Step 400A: After fusing each first cross-domain weighted focused vector representation and each second cross-domain weighted focused vector representation, the representation is mapped into a monitoring data vector representation that matches the non-custodial assessment vector representation dimension of the target non-custodial monitoring data set;

[0070] Step 400B: Using the non-custodial assessment vector representation and non-custodial assessment guidance information of the target non-custodial monitoring data set as input information, combined with the general algorithm in the monitoring data vector representation input field, to determine the non-custodial performance recognition result of the target non-custodial monitoring data set.

[0071] In step 400A, the management and control platform fuses each first cross-domain weighted focused vector representation and each second cross-domain weighted focused vector representation, and maps them into a monitoring data vector representation that matches the non-detention assessment vector representation dimension of the target non-detention monitoring data set. Each first cross-domain weighted focused vector representation reflects the association information between the internal weighted focused vector representation and each overall vector representation in the overall vector representation combination, and each second cross-domain weighted focused vector representation reflects the association information between the internal weighted focused vector representation and each detail vector representation in the detail vector representation combination. The non-detention assessment vector representation is a pre-set vector used to evaluate the performance of non-detention personnel, and its dimensions represent the characteristics of various aspects that need to be considered in the evaluation.

[0072] To achieve vector fusion and mapping, the control platform can use linear transformation. Assume that each first cross-domain weighted focus vector represents a matrix , each second cross-domain weighted focusing vector represents a matrix , concatenate them to get the matrix Assume that the mapping matrix is ​​W and the bias vector is b, then the monitoring data vector M can be expressed by the formula The mapping matrix W and the bias vector b here are optimized and adjusted during the training process of the non-custodial control analysis algorithm, in order to make the mapped monitoring data vector representation match the dimension of the non-custodial evaluation vector representation, thereby providing appropriate input for subsequent evaluations.

[0073] In step 400B, the control platform uses the non-custodial evaluation vector representation and non-custodial evaluation guidance information of the target non-custodial monitoring data set as input information, and combines the monitoring data vector representation to input the domain-general algorithm, thereby determining the non-custodial performance recognition result of the target non-custodial monitoring data set. The domain-general algorithm is a pre-trained and fine-tuned domain model. It has powerful feature learning and classification capabilities and can perform complex analysis and judgment based on input information. The non-custodial evaluation vector representation provides the algorithm with the basic framework and dimensional information about the performance evaluation of non-custodial personnel, and the non-custodial evaluation guidance information further guides the algorithm on how to evaluate based on this information, such as the weights of different evaluation indicators, the standard range of evaluation, etc.

[0074] The control platform can concatenate the non-custodial assessment vector representation E, the non-custodial assessment guidance information I, and the monitoring data vector representation M to obtain the input vector X=[E;I;M], and then input it into the model of the domain-general algorithm. Assuming that the model of the domain-general algorithm can be represented as a function f(X), a series of calculations and transformations are performed on the input vector X to finally output the non-custodial performance recognition result. This result can be a classification label, such as "good performance", "some risk", "high risk", etc., or it can be a continuous evaluation score to more accurately measure the performance of non-custodial personnel.

[0075] Continuing with the example of non-detained persons monitoring, the non-detained assessment vector representation may specify several major aspects of the assessment, such as compliance with the scope of activities, normativeness of social behavior, etc. The non-detained assessment guidance information may indicate that the weight of compliance with the scope of activities is 0.6, the weight of normativeness of social behavior is 0.4, and specifies the score range and judgment criteria for each aspect. The control platform inputs this information together with the monitoring data vector representation into the domain-general algorithm, and the algorithm will conduct a comprehensive assessment of the performance of non-detained persons based on the patterns and rules it has learned. If the monitoring data vector representation shows that the scope of non-detained persons' activities is basically in compliance with the regulations, but they have more contact with some people with bad records in their social behavior, the algorithm may give an identification result of "there is a certain risk" based on the weights and standards.

[0076] In practical applications, the control platform can use deep learning frameworks (such as TensorFlow, PyTorch, etc.) to implement steps 400A and 400B. For step 400A, the control platform can define a linear transformation layer in the model and adjust the parameters of the mapping matrix and the bias vector through training. For step 400B, the control platform can load a pre-trained domain-general algorithm model and input the processed input information into the model for prediction. At the same time, in order to improve the accuracy of the non-custodial performance recognition results, the control platform can also optimize and adjust the domain-general algorithm, such as adjusting the model's hyperparameters, increasing the amount of training data, etc. In addition, the control platform can also combine manual review to further verify and correct the results of the algorithm output to ensure the reliability and effectiveness of the recognition results.

[0077] As an implementation mode, each detail vector represents a data slice in the non-custodial monitoring data subset; then, in step 100, the obtained overall vector representation set and the detailed vector representation set are simplified according to the vector space similarity, and the overall vector representation combination and the detailed vector representation combination of the set capacity are obtained, including:

[0078] Step 110: After each embedding mapping is performed on a non-custodial monitoring data subset, the overall vector representation of the non-custodial monitoring data subset is added to the overall vector representation cache, and the detail vector representation of the non-custodial monitoring data subset is added to the detail vector representation cache;

[0079] Step 120: when the number of overall vector representations contained in the overall vector representation cache is not less than the set capacity, based on the vector space similarity, the overall vector representations of the plurality of non-custodial surveillance data subsets in the overall vector representation cache are subjected to vector integration processing, and the overall vector representation to be added is added to the overall vector representation cache;

[0080] And, step 130: for each data slice, when the number of corresponding detail vector representations contained in the detail vector representation cache is not less than the set capacity, according to the vector space similarity, vector integration processing is performed on the corresponding detail vector representations of multiple non-custodial monitoring data subsets in the detail vector representation cache, and the detail vector representation to be added is added to the detail vector representation cache.

[0081] In step 110, after the control platform performs embedding mapping (encoding) on ​​each subset of non-custodial monitoring data, the overall vector representation of the non-custodial monitoring data subset is added to the overall vector representation cache, and the detailed vector representation is added to the detailed vector representation cache. Embedding mapping is the process of converting the original non-custodial monitoring data into a vector representation. It can map the data from the original data space to a low-dimensional vector space, making the data easier to process and analyze. For example, for the activity trajectory data of non-custodial personnel, the control platform can use word embedding technology to convert different activity locations and time information into vector representations.

[0082] The overall vector representation is a general description of the overall characteristics of a non-custodial surveillance data subset, which reflects the comprehensive performance characteristics of the suspect during the entire surveillance period. The detailed vector representation is a description of more detailed characteristics in the data subset, which can reflect the characteristics of the suspect in a specific time period or under specific behavior. The overall vector representation cache and the detailed vector representation cache are used to temporarily store these vector representations for subsequent integration processing.

[0083] The management and control platform can use deep learning models (such as convolutional neural networks, recurrent neural networks, etc.) to implement embedding mapping. Taking convolutional neural networks as an example, it can process non-custodial monitoring data through convolutional layers, pooling layers, and fully connected layers to convert raw data into vector representations. Assuming that the input non-custodial monitoring data is X, after the convolution operation C(X) and pooling operation P(C(X)) of the convolutional layer, and then through the linear transformation F(P(C(f{X}))) of the fully connected layer, the overall vector representation G and the detailed vector representation f{D} are finally obtained. Then, the management and control platform adds f{G} to the overall vector representation cache f{B} G In , f{D} is added to the detail vector representation cache f{B} D middle.

[0084] For example, when monitoring the activities of non-detained persons, a subset of non-detained monitoring data may contain information such as the person's activity trajectory within a week and communication records with others. After embedding and mapping this information, the control platform obtains an overall vector representation, which may contain information such as the person's average activity range and main activity area within the week; at the same time, it obtains a detailed vector representation, which may contain information such as the person's specific activity trajectory within a certain day and the number of communications with specific persons. The control platform adds these vector representations to the corresponding cache respectively.

[0085] In step 120, when the number of overall vector representations contained in the overall vector representation cache is not less than the set capacity, the management and control platform performs vector integration processing on the overall vector representations of multiple non-custodial monitoring data subsets in the overall vector representation cache based on the vector space similarity, and adds the overall vector representation to be added to the overall vector representation cache. The set capacity is a threshold value pre-set by the management and control platform, which is used to control the number of vectors in the overall vector representation cache. Vector space similarity is an indicator to measure the distance or similarity between two vectors in the vector space. Feasible calculation methods include cosine similarity, Euclidean distance, etc. The management and control platform can use a variety of methods to perform vector integration processing. One method is to perform mean fusion of the overall vector representations of adjacent non-custodial monitoring data subsets with the highest similarity in the vector representation cache based on the vector space similarity. For example, suppose there are three overall vector representations f{G}1, f{G}2, and f{G}3 in the overall vector representation cache. By calculating the cosine similarity, it is found that f{G}1 and f{G}2 have the highest similarity. Then the management and control platform can perform mean fusion on them to obtain a new overall vector representation f{G} new =(f{G}1+ f{G}2) / 2, then use f{G} new Replace the positions of f{G}1 and f{G}2 in the cache.

[0086] Another way is to cluster the overall vector representation to be added with the overall vector representation already contained in the vector representation cache based on the similarity of the vector space, and to perform mean fusion on the overall vector representations of each non-custodial monitoring data subset after clustering. For example, the overall vector representation to be added is f{G} add , the control platform will f{G} add Perform cluster analysis together with all the overall vector representations in the cache, divide similar vectors into a cluster, and then perform mean fusion on the vectors in each cluster. Continuing with the monitoring of non-detained persons as an example, when the number of overall vector representations in the overall vector representation cache reaches the set capacity, the control platform finds that the overall vector representations of two non-detained persons are very similar in the vector space, which may mean that the two persons have great similarities in terms of activity range, activity patterns, etc. The control platform performs mean fusion on the two overall vector representations to obtain a new overall vector representation, which not only reduces the number of vectors in the cache, but also retains the main feature information of the two persons.

[0087] In step 130, for each data slice, when the number of corresponding detail vector representations contained in the detail vector representation cache is not less than the set capacity, the control platform performs vector integration processing on the corresponding detail vector representations of multiple non-custodial monitoring data subsets in the detail vector representation cache based on vector space similarity, and adds the detail vector representation to be added to the detail vector representation cache. A data slice refers to a specific data segment in a non-custodial monitoring data subset, for example, the activity trajectory of a non-custodial person in a certain day can be regarded as a data slice.

[0088] Similar to the processing method of the overall vector representation, the control platform can integrate the detail vector representation by using mean fusion or cluster fusion. For example, for a certain data slice, there are multiple detail vector representations in the detail vector representation cache. The control platform calculates the vector space similarity and performs mean fusion on the adjacent detail vector representations with the highest similarity, or clusters the detail vector representation to be added with the detail vector representation in the cache, and then performs mean fusion on the vectors in each cluster. For example, when monitoring the activities of non-detained persons, a data slice may be the activity trajectory of the person on a certain morning. When the number of detail vector representations about this data slice in the detail vector representation cache reaches the set capacity, the control platform finds that there are two detail vector representations that are very similar in the vector space, which may mean that the activity trajectories of the two non-detained persons on that morning are very similar. The control platform performs mean fusion on the two detail vector representations to obtain a new detail vector representation, which is then added to the cache.

[0089] In practical applications, the control platform can use deep learning frameworks (such as TensorFlow, PyTorch, etc.) to implement steps 110-130. For embedding mapping, the control platform can build a corresponding neural network model and adjust the parameters of the model through training. For vector integration processing, the control platform can use clustering algorithms (such as K-means clustering) to implement clustering operations and use simple mathematical operations (such as mean calculation) to implement mean fusion operations. At the same time, the control platform can also adjust the size of the set capacity according to actual conditions to balance data accuracy and processing efficiency.

[0090] Steps 110-130 embed, map, vector cache and vector integration processing are performed on the subset of non-custodial monitoring data, so that the management and control platform can effectively manage and process a large amount of non-custodial monitoring data, and obtain the overall vector representation combination and detailed vector representation combination of the set capacity, which provides a more representative and computable data basis for subsequent non-custodial management and control analysis, and helps to improve the accuracy and efficiency of non-custodial performance identification.

[0091] As an implementation manner, in step 300, before respectively extracting the first cross-domain weighted focused vector representation between the internal weighted focused vector representation and each overall vector representation in the overall vector representation combination, and respectively extracting the second cross-domain weighted focused vector representation between the internal weighted focused vector representation and each detail vector representation in the detail vector representation combination, the method further includes:

[0092] Step 301: when the set time is triggered, the overall vector representation combination contained in the overall vector representation cache is obtained, and the detail vector representation combination contained in the detail vector representation cache is obtained.

[0093] In step 301, the set time is a time point or time condition pre-set by the control platform, which can be flexibly set according to the actual non-custodial control needs. For example, the set time can be a fixed time every day (such as 2 a.m.), at which time the control platform conducts a unified analysis of the non-custodial monitoring data collected and processed the previous day; it can also be triggered when the number of vectors in the overall vector representation cache or the detailed vector representation cache reaches a certain threshold, which can ensure that the analysis is carried out in a timely manner after the data accumulates to a certain extent.

[0094] The overall vector representation cache and the detailed vector representation cache are containers used by the management and control platform in step 100 to temporarily store the overall vector representation and the detailed vector representation. In step 100, after the management and control platform performs embedding mapping on each non-custodial monitoring data subset, the overall vector representation of the non-custodial monitoring data subset is added to the overall vector representation cache, and the detailed vector representation is added to the detailed vector representation cache. In addition, when the number of vectors in the cache is not less than the set capacity, the vectors are integrated according to the similarity of the vector space. Therefore, when the set time arrives, the overall vector representation cache and the detailed vector representation cache store the integrated vector representation combinations, which can more effectively represent the main features of the target non-custodial monitoring data set.

[0095] The control platform can use data structures in programming languages ​​(such as lists, arrays, etc.) to implement overall vector representation cache and detailed vector representation cache. For example, in Python, lists can be used to store vector representations, and each vector representation can be a NumPy array. When the set time is triggered, the control platform can obtain all vector representations in the list through simple index operations to form an overall vector representation combination and a detailed vector representation combination.

[0096] In the actual scenario of non-custodial control, assume that the control platform is responsible for monitoring the activities of multiple non-custodial suspects. During the monitoring process of a day, the control platform continuously collects non-custodial monitoring data of each suspect, extracts features for each data subset, obtains overall vector representation and detail vector representation, and adds them to the overall vector representation cache and detail vector representation cache respectively. The time is set to 2 a.m. every day. When the time reaches 2 a.m., the control platform will obtain all overall vector representations in the overall vector representation cache to form an overall vector representation combination. These overall vector representations may contain information such as the average activity range and activity frequency of each suspect in a day. At the same time, the control platform will also obtain all detail vector representations in the detail vector representation cache to form a detail vector representation combination. These detail vector representations may contain information such as the specific activity trajectory of each suspect in a specific time period and contact records with specific personnel.

[0097] The process of obtaining the overall vector representation combination and the detailed vector representation combination can also be implemented in combination with database operations. The management and control platform can store the vector representation in the database, and when the set time is triggered, the corresponding vector representation is extracted from the database through SQL query statements.

[0098] After the control platform obtains the overall vector representation combination and the detail vector representation combination, these vector representation combinations will be used as inputs for the subsequent step 300. In step 300, the control platform will use the cross-domain weight focusing strategy to extract the first cross-domain weight focusing vector representation between the internal weight focusing vector representation and each overall vector representation in the overall vector representation combination, and the second cross-domain weight focusing vector representation between each detail vector representation in the detail vector representation combination. Therefore, the quality and accuracy of the vector representation combination obtained in step 301 directly affect the effect of subsequent analysis.

[0099] In order to ensure the validity of the obtained vector representation combination, the management and control platform can also perform some preprocessing operations after obtaining the vector representation combination. For example, the vector representation is normalized so that the value range of all vector representations is within a unified interval, which can improve the stability and accuracy of subsequent calculations.

[0100] Step 301 plays an important transitional role in the entire non-custodial control data analysis process. The control platform obtains the overall vector representation combination and the detailed vector representation combination at the set time, providing an accurate and effective data basis for subsequent cross-domain analysis, which helps to achieve precise control and supervision of non-custodial suspects.

[0101] As an implementation, vector integration processing is performed on target vector representations of multiple non-custodial surveillance data subsets based on one of the following methods:

[0102] Method 1: Based on the similarity of the vector space, the target vector representations of the adjacent non-custodial surveillance data subsets with the highest similarity in the vector representation cache are mean-fused;

[0103] Method 2: Based on the similarity of vector space, the target vector representation to be added and the target vector representation already contained in the vector representation cache are clustered, and the target vector representations of each non-custodial monitoring data subset after clustering are mean fused; wherein the target vector representation is an overall vector representation or a detailed vector representation, and correspondingly, the vector representation cache is an overall vector representation cache or a detailed vector representation cache.

[0104] In the process of analyzing non-custodial control data, it is an important step for the control platform to perform vector integration processing on the target vector representations of multiple non-custodial monitoring data subsets, which helps to streamline data, reduce redundant information, and retain key features, laying the foundation for subsequent accurate analysis of the performance of non-custodial personnel. Method 1 and Method 2 are two feasible vector integration processing methods for the control platform.

[0105] Method 1 is to perform mean fusion on the target vector representations of the adjacent non-custodial surveillance data subsets with the highest similarity in the vector representation cache based on vector space similarity. Vector space similarity is an indicator that measures the distance or similarity between two vectors in the vector space. The feasible calculation methods include cosine similarity and Euclidean distance. Mean fusion is to add the corresponding elements of the adjacent target vector representations with the highest similarity and take the average value to obtain a new vector representation.

[0106] Taking the activity monitoring of non-detained persons as an example, assume that there are three target vector representations of non-detained monitoring data subsets in the vector representation cache: , respectively representing the activity characteristics of three non-detained persons. By calculating the similarity of the vector space, it is found that and The similarity is the highest, and they are adjacent in the vector representation cache. Then the management and control platform will and Perform mean fusion. , then the new fused vector The control platform will use replace and The location in the vector representation cache is used to reduce the number of vector representations while retaining comprehensive information about the activity characteristics of the two non-detained persons. The management and control platform can use programming languages ​​(such as Python) combined with relevant scientific computing libraries (such as NumPy) to achieve this mean fusion.

[0107] The second method is to cluster the target vector representation to be added and the target vector representation already contained in the vector representation cache based on the similarity of the vector space, and to perform mean fusion on the target vector representations of each non-custodial monitoring data subset after clustering. Clustering is the process of dividing similar vectors into the same cluster. A feasible clustering algorithm is the K-means clustering algorithm. The basic idea of ​​the K-means clustering algorithm is to pre-set the number of clusters k, and then assign vectors to different clusters in an iterative manner so that the sum of the distances from each vector to the center of the cluster to which it belongs is minimized. Assume that there are m target vector representations in the vector representation cache. , the target vector to be added is expressed as The control platform performs K-means clustering on these m + 1 vectors to obtain k clusters For each cluster C i , the control platform will calculate the mean of all vectors in the cluster and obtain a new vector representation as the representative vector of the cluster. Assume that cluster C i There are n i Vector , then the representative vector of the cluster The jth component of .

[0108] Continuing with the example of non-detained personnel activity monitoring, assume that there are four target vector representations of non-detained monitoring data subsets in the vector representation cache: , the target vector to be added is expressed as The control platform performs K-means clustering on these five vectors, setting the number of clusters k=2. After clustering, is divided into cluster C1, is divided into cluster C2. For cluster C1, assume , then the representative vector of cluster C1 ; For cluster C2, assume , then the representative vector of cluster C2 The control platform will use and Replace the positions of the original five vectors in the vector representation cache.

[0109] Method 1 and method 2 each have their own advantages and disadvantages. The computational complexity of method 1 is relatively low, and it only needs to compare the similarities of adjacent vectors and perform mean fusion. It is suitable for situations where the number of vectors in the vector representation cache is small and the similarity of adjacent vectors is high. The computational complexity of method 2 is relatively high, and clustering operations are required, but it can more comprehensively consider the similarities of all vectors, and divide similar vectors into the same cluster for fusion. It is suitable for situations where the number of vectors in the vector representation cache is large and the vector distribution is more complex. The control platform can select the appropriate vector integration processing method according to the actual situation to improve the efficiency and accuracy of non-custodial control data analysis.

[0110] As an implementation method, if there are multiple sets of adjacent non-custodial monitoring data subsets with the greatest similarity, then, in method 1, the target vector representation of the adjacent non-custodial monitoring data subset with the highest similarity in the vector representation cache is mean-fused, including:

[0111] Step 140: Mean fusion is performed on the target vector representations of the adjacent non-custodial monitoring data subsets that are stored the longest and have the highest similarity in the vector representation cache; or, mean fusion is performed on the target vector representations of each group of adjacent non-custodial monitoring data subsets that have the highest similarity in the vector representation cache.

[0112] In the analysis of non-custodial control data, when the target vector representations of multiple non-custodial monitoring data subsets are processed by vector integration, if there are multiple groups of adjacent non-custodial monitoring data subsets with the greatest similarity, step 140 provides a clear processing strategy for the control platform. Step 140 provides two specific operation methods, one is to perform mean fusion on the target vector representations of the adjacent non-custodial monitoring data subsets that are stored the longest and have the highest similarity in the vector representation cache, and the other is to perform mean fusion on the target vector representations of each group of adjacent non-custodial monitoring data subsets with the highest similarity in the vector representation cache.

[0113] First, let's look at the case of mean fusion of the target vector representations of adjacent non-custodial surveillance data subsets that are stored the longest and have the highest similarity in the vector representation cache. The vector representation cache is a container used by the management and control platform to temporarily store target vector representations. The target vector representation is a quantitative reflection of the characteristics of the non-custodial surveillance data subset. The longest preservation means that the vector representation has existed in the cache for the longest time and has a relative a priori in the time dimension of the data. The highest similarity is obtained by calculating the similarity of the vector space. The feasible calculation methods are cosine similarity and Euclidean distance. Mean fusion is to add the corresponding elements of the two target vector representations and take the average value.

[0114] The management and control platform can add a timestamp attribute to each vector representation in the vector representation cache to record the time when it enters the cache. When performing vector space similarity calculations, the similarity and storage time of adjacent vectors are compared at the same time. These operations can be implemented using programming languages ​​(such as Python) combined with related scientific computing libraries (such as NumPy).

[0115] For the case where the target vector representations of each group of adjacent non-custodial surveillance data subsets with the highest similarity in the vector representation cache are mean-fused. This method emphasizes equal treatment of all adjacent groups with the highest similarity, regardless of the storage time of the vector representation in the cache. The control platform will perform mean fusion operations on each group of adjacent target vector representations with the highest similarity.

[0116] In practical applications, the management and control platform can use a loop structure to traverse the vector representation cache, find all adjacent groups with the highest similarity, and perform a mean fusion operation on each group. By setting an appropriate similarity threshold, it is ensured that only truly similar adjacent vectors are fused. At the same time, in order to improve computing efficiency, the vector representation cache can be sorted so that adjacent vectors are also adjacent in physical storage, reducing the time complexity of traversal.

[0117] The two implementations of step 140 each have their own applicable scenarios. When the management and control platform is more concerned with the effective use of cache space and the processing of old data, it is more appropriate to select the target vector representation of the adjacent non-custodial monitoring data subsets that have been stored for the longest time and have the highest similarity for mean fusion; when the management and control platform hopes to retain the characteristic information of similar data as comprehensively as possible without missing any group of similar data, it is a better choice to perform mean fusion on the target vector representation of each group of adjacent non-custodial monitoring data subsets with the highest similarity. The management and control platform can flexibly select an appropriate method for vector integration processing based on the actual non-custodial control needs and data characteristics, so as to improve the accuracy and efficiency of non-custodial control data analysis and better realize the management and monitoring of non-custodial personnel.

[0118] As an implementation method, the non-custodial control analysis algorithm is calibrated using the following process:

[0119] Step 10: Based on the non-custodial control analysis algorithm to be adjusted and optimized, feature extraction is performed on each non-custodial monitoring data subset in the non-custodial monitoring training data set to obtain the corresponding overall vector representation and detail vector representation, and based on the vector space similarity, the obtained overall vector representation set and detail vector representation set are simplified to obtain the overall vector representation combination and detail vector representation combination of the set capacity;

[0120] Step 20: After fusing the first adaptive search vector representation to be tuned and optimized with the second adaptive search vector representation, extract the corresponding internal weighted focused vector representation through the internal weighted focused strategy; the internal weighted focused vector representation integrates the core information of the first adaptive search vector representation and the second adaptive search vector representation;

[0121] Step 30: extracting, through a cross-domain weight focusing strategy, a first cross-domain weight focusing vector representation between the internal weight focusing vector representation and each overall vector representation in the overall vector representation combination, and a second cross-domain weight focusing vector representation between the internal weight focusing vector representation and each detail vector representation in the detail vector representation combination;

[0122] Step 40: Based on the errors between the non-custodial performance recognition results and non-custodial assessment information of the non-custodial monitoring training data set obtained by inferring the first cross-domain weighted focusing vector representations and the second cross-domain weighted focusing vector representations, and the corresponding calibration supervision information, the algorithm parameters of the non-custodial control and analysis algorithm are calibrated. The calibrated non-custodial control and analysis algorithm is used to determine the non-custodial performance recognition results of the target non-custodial monitoring data set.

[0123] In step 10, the control platform extracts features from each non-custodial monitoring data subset in the non-custodial monitoring training data set according to the non-custodial control analysis algorithm to be adjusted and optimized, obtains the corresponding overall vector representation and detail vector representation, and simplifies the obtained overall vector representation set and detail vector representation set according to the vector space similarity, to obtain the overall vector representation combination and detail vector representation combination of the set capacity. The non-custodial monitoring training data set is a set of data used by the control platform for training algorithms. It contains multiple non-custodial monitoring data subsets, each of which corresponds to the monitoring data of a non-custodial person in a certain period of time. Feature extraction is the process of converting the original non-custodial monitoring data into a more representative and computable vector representation. The overall vector representation summarizes the overall characteristics of a non-custodial monitoring data subset, and the detail vector representation reflects the detailed characteristics in the data subset. Vector space similarity is used to measure the distance or similarity between two vectors in vector space. The feasible calculation methods include cosine similarity and Euclidean distance.

[0124] Taking the activity monitoring of non-detained persons as an example, the non-detained monitoring training data set may contain the activity trajectories, communication records and other information of 100 non-detained persons within a week. The control platform extracts features from the monitoring data of each person to obtain the overall vector representation and detailed vector representation of each person. Assume that there are 120 vectors in the overall vector representation set, 150 vectors in the detailed vector representation set, and the set capacity is 100. The control platform integrates similar vectors in the overall vector representation set and the detailed vector representation set based on the similarity of the vector space, for example, the adjacent vectors with the highest similarity are averaged and finally the overall vector representation combination and the detailed vector representation combination with a capacity of 100 are obtained. From a technical point of view, the control platform can use deep learning models (such as convolutional neural networks, recurrent neural networks, etc.) for feature extraction and clustering algorithms (such as K-means clustering) for vector integration processing.

[0125] In step 20, after the control platform fuses the first adaptive search vector representation and the second adaptive search vector representation to be adjusted and optimized, it extracts the corresponding internal weighted focused vector representation through the internal weighted focused strategy, which integrates the core information in the first adaptive search vector representation and the second adaptive search vector representation. The first adaptive search vector representation and the second adaptive search vector representation are learnable vectors, which will be continuously adjusted during the algorithm adjustment process to better adapt to the characteristics of non-custodial monitoring data. The internal weighted focused strategy, also known as the self-attention mechanism, can automatically pay attention to the importance of information at different positions in the input sequence, and highlight important information by assigning different weights to each element.

[0126] Specifically, the control platform first represents the first adaptive search vector and the second adaptive search vector represents Perform fusion, for example, obtain the fused vector by splicing operation Then, The query vector is obtained by three linear transformation layers respectively. , key vector Sum value vector ,in are the weight matrices for query, key, and value respectively. Next, the similarity score matrix between the query vector and the key vector is calculated To prevent the similarity score from being too large, the score matrix is ​​scaled to obtain , where d k is the dimension of the key vector. Then the scaled score matrix is ​​converted into a probability distribution matrix A=softmax(S') through the softmax function, and finally the probability distribution matrix is ​​multiplied by the value vector to obtain the internal weight focus vector representation .

[0127] For example, in the monitoring scenario of non-detained persons, the first adaptive search vector indicates that global features such as the range of activities of non-detained persons may be focused on, while the second adaptive search vector indicates that detailed features such as the contact between non-detained persons and specific persons may be focused on. Through the self-attention mechanism, the control platform can extract the internal weighted focus vector representation that contains both global and detailed core information from the fused vector. In terms of technical implementation, the control platform can use deep learning frameworks (such as TensorFlow, PyTorch, etc.) to build a model of the self-attention mechanism and adjust the parameters of the weight matrix through training.

[0128] In step 30, the control platform uses a cross-domain weight focusing strategy to extract a first cross-domain weight focusing vector representation between the internal weight focusing vector representation and each overall vector representation in the overall vector representation combination, and a second cross-domain weight focusing vector representation between the internal weight focusing vector representation and each detail vector representation in the detail vector representation combination. The cross-domain weight focusing strategy, i.e., the cross-attention mechanism, is used to capture the correlation information between different feature vectors.

[0129] The control platform focuses on the internal weight vector representation , each overall vector in the overall vector representation combination represents (i represents the index of the overall vector representation) and each detail vector representation in the combination of the detail vector representation (j represents the index of the detail vector representation) The query vector, key vector and value vector are obtained through linear transformation respectively. For the overall vector representation, the query vector , the key vector , value vector ; For the detail vector representation, the query vector , the key vector , value vector ,in are the corresponding weight matrices respectively. Then the similarity score between the query vector and the key vector is calculated to obtain the similarity score matrix and , and scale to get and ,in and The scaled score matrix is ​​converted into a probability distribution matrix through the softmax function. and Finally, the probability distribution matrix is ​​multiplied by the value vector to obtain the first cross-domain weighted focusing vector representation and the second cross-domain weighted focused vector representation .

[0130] Taking the monitoring of non-detained persons as an example, the overall vector representation combination may contain the overall activity characteristics of multiple non-detained persons over a period of time, and the detailed vector representation combination may contain the detailed activity information of each non-detained person in a specific time period. The internal weighted focus vector representation contains comprehensive core information. Through the cross-attention mechanism, the management and control platform can obtain the degree of correlation between the overall vector representation and the detailed vector representation of each non-detained person and the core information, that is, the first cross-domain weighted focus vector representation and the second cross-domain weighted focus vector representation. In terms of technical implementation, the management and control platform can use the deep learning framework to build a model of the cross-attention mechanism and optimize the parameters of the weight matrix during the training process.

[0131] In step 40, the control platform adjusts the algorithm parameters of the non-custodial control and analysis algorithm based on the error between the non-custodial performance recognition results and non-custodial evaluation information of the non-custodial monitoring training data set inferred according to each first cross-domain weighted focus vector representation and each second cross-domain weighted focus vector representation, and the corresponding calibration supervision information (that is, training labels). The calibrated non-custodial control and analysis algorithm is used to determine the non-custodial performance recognition results of the target non-custodial monitoring data set. The non-custodial performance recognition result is a classification judgment of the performance of non-custodial personnel during the monitoring period, such as "good performance" and "risky"; the non-custodial evaluation information is a quantitative evaluation of the performance of non-custodial personnel, such as an evaluation score. The calibration supervision information is a pre-labeled real result, which is used to measure the accuracy of the algorithm reasoning result.

[0132] The control platform can use the loss function to calculate the error between the inference result and the calibration supervision information. A feasible loss function is the cross entropy loss function. For classification problems, the calculation formula of the cross entropy loss function is: , where y i is the true label, p i is the probability predicted by the algorithm. The control platform updates the parameters of the non-custodial control analysis algorithm based on the gradient of the loss function through the back-propagation algorithm. For example, for a neural network model, the weight matrix and bias vector of each layer are updated.

[0133] In the monitoring scenario of non-detained persons, assuming that the calibration supervision information shows that a non-detained person's performance is "good", and the result obtained by the algorithm reasoning is "risky", the error between the two is calculated through the loss function. The control platform adjusts the parameters of the algorithm based on this error, so that the algorithm can more accurately identify the performance of non-detained persons in subsequent reasoning. In terms of technical implementation, the control platform can use the optimizer provided by the deep learning framework (such as stochastic gradient descent, Adam, etc.) to implement parameter updates.

[0134] As an implementation method, the non-custodial control analysis algorithm includes a domain-general algorithm, and a feature integration component based on an internal weight focus strategy and a cross-domain weight focus strategy architecture, and the feature integration component is used to extract a first cross-domain weight focus vector representation and a second cross-domain weight focus vector representation. Then, in step 40, based on each first cross-domain weight focus vector representation and each second cross-domain weight focus vector representation, the non-custodial performance recognition result and non-custodial evaluation information of the non-custodial monitoring training data set are inferred, including:

[0135] Step 41: In the algorithm pre-training phase, each first cross-domain weighted focused vector representation and each second cross-domain weighted focused vector representation are fused and mapped into a monitoring data vector representation that matches the non-detention evaluation vector representation dimension of the non-detention monitoring training data set; the first non-detention evaluation guidance information is used as input information, combined with the monitoring data vector representation input field general algorithm, to obtain the non-detention evaluation information of the non-detention monitoring training data set;

[0136] Step 42: In the algorithm fine-tuning phase, the first cross-domain weighted focused vector representation and the second cross-domain weighted focused vector representation are fused and mapped into a monitoring data vector representation that matches the non-custodial assessment vector representation dimension of the non-custodial monitoring training data set; the non-custodial assessment vector representation and the second non-custodial assessment guidance information are used as input information, combined with the general algorithm in the monitoring data vector representation input field, to obtain the non-custodial performance recognition result of the non-custodial monitoring training data set.

[0137] Step 41: For the algorithm pre-training phase, the control platform fuses each first cross-domain weighted focused vector representation and each second cross-domain weighted focused vector representation, and maps them into a monitoring data vector representation that matches the non-detention evaluation vector representation dimension of the non-detention monitoring training data set. Then, the first non-detention evaluation guidance information is used as input information, combined with the monitoring data vector representation input field general algorithm, to obtain the non-detention evaluation information of the non-detention monitoring training data set. Each first cross-domain weighted focused vector representation reflects the association information between the internal weighted focused vector representation and each overall vector representation in the overall vector representation combination, and each second cross-domain weighted focused vector representation reflects the association information between the internal weighted focused vector representation and each detail vector representation in the detail vector representation combination. The non-detention evaluation vector representation is a pre-set vector used to evaluate the performance of non-detention personnel, and its dimension represents the characteristics of various aspects that need to be considered in the evaluation.

[0138] To achieve vector fusion and mapping, the control platform can use linear transformation. Assume that each first cross-domain weighted focus vector represents the composition matrix f{Z} G , each second cross-domain weighted focusing vector represents the composition matrix f{Z} D , concatenate them to get the matrix f{Z}=[f{Z}G ; f{Z} D ]. Assume that the mapping matrix is ​​f{W} and the bias vector is f{b}, then the monitoring data vector representation f{M} can be calculated by the formula f{M}=f{W}f{Z}+f{b}. The mapping matrix f{W} and the bias vector f{b} here are optimized and adjusted during the algorithm pre-training process. The purpose is to make the mapped monitoring data vector representation match the dimension of the non-custodial evaluation vector representation, so as to provide appropriate input for subsequent evaluation. The first non-custodial evaluation guidance information is additional information used to guide the domain-general algorithm to perform non-custodial evaluation, such as the weights of different evaluation indicators, the standard range of evaluation, etc. The domain-general algorithm is a pre-trained and fine-tuned large domain model. It has powerful feature learning and classification capabilities, and can perform complex analysis and judgment based on the input information.

[0139] Taking the monitoring scenario of non-detained persons as an example, the first cross-domain weighted focus vector representation may include the association between the overall characteristics of non-detained persons such as the scope of activities and frequency of activities over a long period of time and the core information, and the second cross-domain weighted focus vector representation may include the association between the specific activity trajectory of non-detained persons in a specific period of time, the detailed characteristics such as contact with specific persons and the core information. The dimensions of the non-detained evaluation vector representation may cover multiple aspects such as the compliance of non-detained persons with regulations and changes in social harm. The first non-detained evaluation guidance information may stipulate that the weight of compliance with the scope of activities is 0.6, and the weight of social behavior norms is 0.4, and the score range and judgment criteria for each aspect are given. The control platform fuses and maps the first cross-domain weighted focus vector representation and the second cross-domain weighted focus vector representation to obtain the monitoring data vector representation, and then inputs it into the domain general algorithm together with the first non-detained evaluation guidance information. The algorithm evaluates the performance of non-detained persons based on this information, for example, giving a comprehensive evaluation score, which is the non-detained evaluation information. The control platform can use deep learning frameworks (such as TensorFlow, PyTorch, etc.) to implement vector fusion, mapping, and input and calculation of domain-wide algorithms. For vector fusion and mapping, the control platform can build a linear transformation layer and adjust the parameters of the mapping matrix and bias vector through training. For domain-wide algorithms, the control platform can load pre-trained models and input the processed input information into the model for prediction.

[0140] In step 42, for the algorithm fine-tuning phase, the control platform fuses the first cross-domain weighted focused vector representation and the second cross-domain weighted focused vector representation, and maps them into a monitoring data vector representation that matches the non-detention evaluation vector representation dimension of the non-detention monitoring training data set. Then, the non-detention evaluation vector representation and the second non-detention evaluation guidance information are used as input information, and combined with the monitoring data vector representation input field general algorithm, the non-detention performance recognition result of the non-detention monitoring training data set is obtained. Different from step 41, in the algorithm fine-tuning phase, the non-detention evaluation vector representation needs to be included in the input information, and the second non-detention evaluation guidance information needs to be used. The second non-detention evaluation guidance information may be guidance information that is further refined or adjusted on the basis of pre-training, which is more in line with the actual non-detention control scenario.

[0141] The first cross-domain weighted focus vector representation and the second cross-domain weighted focus vector representation are also fused and mapped into a monitoring data vector representation by linear transformation. Assume that the first cross-domain weighted focus vector representation is , the second cross-domain weighted focusing vector is expressed as , concatenate them and perform linear transformation to obtain the monitoring data vector representation Then, the control platform represents the non-custodial evaluation vector 2. Second Non-custodial Assessment Guidance Information and monitoring data vector representation Concatenate and get the input vector , and input it into the model of the domain-general algorithm.

[0142] Continuing with the example of non-detained personnel monitoring, in the algorithm fine-tuning phase, the non-detained evaluation vector indicates the specific values ​​of each evaluation indicator, and the second non-detained evaluation guidance information may adjust the evaluation criteria according to the actual situation, such as increasing the attention paid to contact with suspicious persons. The control platform inputs this information together with the monitoring data vector representation into the domain-general algorithm. The algorithm classifies and judges the performance of non-detained personnel based on the patterns and rules it has learned, and outputs the non-detained performance recognition results, such as "good performance", "certain risks", "high risk", etc.

[0143] In practical applications, the control platform can use a smaller learning rate to adjust the parameters of the domain-wide algorithm based on pre-training during algorithm fine-tuning to avoid overfitting of the model during fine-tuning. At the same time, the control platform can add more annotated data for fine-tuning to improve the model's adaptability to actual non-custodial control scenarios.

[0144] As an implementation mode, in step 40, based on the errors between the non-custodial performance recognition results and non-custodial assessment information of the non-custodial monitoring training data set inferred according to the obtained first cross-domain weighted focus vector representations and the obtained second cross-domain weighted focus vector representations, and the corresponding adjustment supervision information, the algorithm parameters of the non-custodial control analysis algorithm are adjusted, including:

[0145] Step 43: In the algorithm pre-training phase, based on the error between the non-custodial evaluation information obtained by inference and the non-custodial evaluation information in the calibration supervision information, the algorithm parameters of the feature integration component are calibrated;

[0146] Step 44: In the algorithm fine-tuning phase, based on the error between the non-custodial performance recognition result obtained by reasoning and the non-custodial performance recognition result in the calibrated supervision information, the algorithm parameters of the feature integration component and the domain-general algorithm are calibrated.

[0147] Steps 43-44 in the implementation of step 40 are key steps in the calibration process of the non-custodial control analysis algorithm, respectively targeting the algorithm pre-training phase and the algorithm fine-tuning phase, and based on the error between the results obtained by reasoning and the calibration supervision information, the algorithm parameters of the non-custodial control analysis algorithm are calibrated to improve the accuracy and adaptability of the algorithm so that it can be better applied to the control scenarios of non-custodial personnel.

[0148] Step 43 is that in the algorithm pre-training phase, the control platform adjusts the algorithm parameters of the feature integration component based on the error between the non-detention assessment information obtained by reasoning and the non-detention assessment information in the adjustment supervision information. In the algorithm pre-training phase, the control platform, according to step 41, fuses each first cross-domain weighted focus vector representation and each second cross-domain weighted focus vector representation, and maps them into a monitoring data vector representation that matches the non-detention assessment vector representation dimension of the non-detention monitoring training data set, and then uses the first non-detention assessment guidance information as input information, combined with the monitoring data vector representation input field general algorithm, to obtain the non-detention assessment information of the non-detention monitoring training data set. The non-detention assessment information is a quantitative assessment of the performance of non-detained personnel, for example, it can be a comprehensive assessment score that reflects the performance of non-detained personnel in terms of compliance with regulations, social harm, etc.

[0149] The calibration supervision information is the real result that has been pre-labeled, and the non-custodial evaluation information is an accurate evaluation value that can be used as a reference. The control platform measures the accuracy of the algorithm in the pre-training phase by calculating the error between the non-custodial evaluation information obtained by reasoning and the non-custodial evaluation information in the calibration supervision information. A feasible error calculation method is to use a loss function, such as the mean square error (MSE) loss function, whose formula is: , where y iis the non-custodial assessment information (true value) in the adjustment supervision information, is the non-custodial assessment information (predicted value) obtained by inference, and n is the number of samples.

[0150] The feature integration component plays an important role in information integration and processing in the entire non-detention control analysis algorithm. It is used to extract the first cross-domain weighted focus vector representation and the second cross-domain weighted focus vector representation. The control platform uses the back propagation algorithm to update the algorithm parameters of the feature integration component based on the error calculated by the loss function. The back propagation algorithm calculates the gradient of the error relative to each parameter in the feature integration component, and then updates the parameters in the opposite direction of the gradient, so that the error gradually decreases. Taking the activity monitoring of non-detained persons as an example, assume that the non-detained assessment information of a non-detained person in the calibration supervision information is 80 points, and the assessment information obtained by algorithm reasoning is 70 points. The control platform uses the mean square error loss function to calculate the error, and then updates the parameters in the feature integration component based on this error. If there is a weight matrix W in the feature integration component, the control platform will calculate the gradient of the error relative to W, and then adjust the value of W according to the above update formula so that the assessment information obtained by the next reasoning is closer to the true value.

[0151] The management and control platform can use deep learning frameworks (such as TensorFlow, PyTorch, etc.) to implement error calculation and parameter update. These frameworks provide automatic differentiation functions, which can easily calculate the gradient of the error relative to the parameter, and provide various optimizers (such as stochastic gradient descent, Adam, etc.) to implement parameter updates.

[0152] Step 44 is the algorithm fine-tuning stage. The control platform adjusts the algorithm parameters of the feature integration component and the domain-general algorithm based on the error between the non-detention performance recognition result obtained by reasoning and the non-detention performance recognition result in the adjustment supervision information. In the algorithm fine-tuning stage, the control platform, according to step 42, fuses the first cross-domain weighted focus vector representation and the second cross-domain weighted focus vector representation, maps them into a monitoring data vector representation that matches the non-detention evaluation vector representation dimension of the non-detention monitoring training data set, takes the non-detention evaluation vector representation and the second non-detention evaluation guidance information as input information, and inputs the domain-general algorithm in combination with the monitoring data vector representation to obtain the non-detention performance recognition result of the non-detention monitoring training data set. The non-detention performance recognition result is a classification judgment of the performance of non-detention personnel during the monitoring period, such as "good performance", "certain risks", "high risk", etc.

[0153] Similarly, the non-custodial performance recognition results in the calibration supervision information are pre-labeled true classification labels. The control platform evaluates the accuracy of the algorithm in the fine-tuning phase by calculating the error between the non-custodial performance recognition results obtained by inference and the non-custodial performance recognition results in the calibration supervision information. For classification problems, a feasible loss function is the cross entropy loss function.

[0154] In the algorithm fine-tuning phase, the parameters of both the feature integration component and the domain-general algorithm need to be adjusted. Because in the fine-tuning process, the entire algorithm needs to be more in line with the actual non-custodial control scenario, and both the feature integration component and the domain-general algorithm have an impact on the final recognition result. The control platform also uses the back propagation algorithm to update the parameters in the feature integration component and the domain-general algorithm based on the error calculated by the cross entropy loss function. The update method is similar to step 43, by calculating the gradient of the error relative to each parameter, and then updating the parameters in the opposite direction of the gradient.

[0155] Continuing with the example of non-detained personnel monitoring, suppose that the non-detained performance recognition result of a non-detained person in the calibration supervision information is "good performance", while the result of algorithm reasoning is "there is a certain risk". The control platform uses the cross entropy loss function to calculate the error, and then updates the parameters in the feature integration component and the domain-general algorithm based on this error. For example, for a hidden layer weight matrix W in the domain-general algorithm hidden , the control platform will calculate the error relative to W hidden The gradient of , and then adjust W according to the update formula hidden , which enables the algorithm to more accurately identify the performance of non-detained persons in subsequent reasoning.

[0156] The control platform usually uses a smaller learning rate in the algorithm fine-tuning phase than in the pre-training phase to avoid overfitting of the model during the fine-tuning process. At the same time, the control platform can add more labeled data for fine-tuning to improve the adaptability of the model to actual non-custodial control scenarios. Steps 43 and 44 are respectively in the algorithm pre-training phase and the algorithm fine-tuning phase. By calculating the error between the inference result and the adjustment supervision information, the algorithm parameters of the feature integration component and the domain general algorithm are adjusted, and the performance of the non-custodial control analysis algorithm is gradually optimized, so that it can more accurately identify the performance of non-custodial personnel and provide a more reliable decision-making basis for the management and monitoring of non-custodial personnel.

[0157] The present application embodiment provides a management and control system, such as Figure 2As shown, the management and control system 100 includes: a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, such as through a bus 102. Optionally, the management and control system 100 may also include a transceiver 104. It should be noted that in actual applications, the transceiver 104 is not limited to one, and the structure of the management and control system 100 does not constitute a limitation on the embodiments of the present application.

[0158] An embodiment of the present application provides a management and control system. The management and control system in the embodiment of the present application includes: one or more processors; a memory; one or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by one or more processors, and when the one or more programs are executed by the processor, the non-custodial management and control data analysis method provided by the present application is implemented.

[0159] The above description is only a partial implementation method of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A non-custodial control data analysis method, characterized in that: The method comprises: According to the non-custodial control analysis algorithm, feature extraction is performed on each non-custodial monitoring data subset in the target non-custodial monitoring data set to obtain the corresponding overall vector representation and detail vector representation, and according to the vector space similarity, the obtained overall vector representation set and detail vector representation set are simplified to obtain the overall vector representation combination and detail vector representation combination of the set capacity; After fusing the pre-learned overall search vector representation and the detailed search vector representation, a corresponding internal weighted focused vector representation is extracted through an internal weighted focused strategy; the overall search vector representation and the detailed search vector representation are obtained by adjusting the corresponding adaptive search vector representations based on a non-custodial monitoring training data set; the internal weighted focused vector representation integrates the core information in the overall search vector representation and the detailed search vector representation; By using a cross-domain weight focusing strategy, respectively extracting a first cross-domain weight focusing vector representation between the internal weight focusing vector representation and each overall vector representation in the overall vector representation combination, and respectively extracting a second cross-domain weight focusing vector representation between the internal weight focusing vector representation and each detail vector representation in the detail vector representation combination; After fusing each of the first cross-domain weighted focused vector representations and each of the second cross-domain weighted focused vector representations, mapping them into a monitoring data vector representation that matches the non-custodial assessment vector representation dimension of the target non-custodial monitoring data set; The non-custodial control analysis algorithm includes a domain-general algorithm, which takes the non-custodial assessment vector representation and non-custodial assessment guidance information of the target non-custodial monitoring data set as input information, combines the monitoring data vector representation with the input of the domain-general algorithm, and determines the non-custodial performance recognition result of the target non-custodial monitoring data set.

2. The method according to claim 1, characterized in that The non-custodial control analysis algorithm includes a feature integration component; the cross-domain weight focusing strategy is used to extract the first cross-domain weight focusing vector representation between the internal weight focusing vector representation and each overall vector representation in the overall vector representation combination, and the second cross-domain weight focusing vector representation is extracted between the internal weight focusing vector representation and each detail vector representation in the detail vector representation combination, including: Decomposing the internal weighted focused vector representation into a first internal weighted focused vector representation corresponding to the overall vector representation and a second internal weighted focused vector representation corresponding to the detail vector representation; Based on the first cross-domain weight focusing layer in the feature integration component, according to the cross-domain weight focusing strategy, respectively extract the first internal weight focusing vector representation, and the first cross-domain weight focusing vector representation between each overall vector representation in the overall vector representation combination; Based on the second cross-domain weighted focusing layer in the feature integration component, according to the cross-domain weighted focusing strategy, respectively extract the second internal weighted focusing vector representation, and the second cross-domain weighted focusing vector representation between each detail vector representation in the detail vector representation combination; The feature integration component includes a plurality of cascaded restoration mapping layers; for a first restoration mapping layer, the execution data includes the overall vector representation combination and the detail vector representation combination; for other restoration mapping layers, the execution data includes each first cross-domain weighted focused vector representation and each second cross-domain weighted focused vector representation output by the previous restoration mapping layer; The non-custodial performance recognition result of the target non-custodial monitoring data set is determined based on the first cross-domain weighted focusing vector representations and the second cross-domain weighted focusing vector representations output by the last restoration mapping layer.

3. The method according to claim 1 or 2, characterized in that Each detail vector represents a data slice in the corresponding non-custodial monitoring data subset; the overall vector representation set and the detail vector representation set are simplified according to the vector space similarity to obtain the overall vector representation combination and the detail vector representation combination of the set capacity, including: After each embedding mapping is performed on a non-custodial monitoring data subset, the overall vector representation of the non-custodial monitoring data subset is added to the overall vector representation cache, and the detail vector representation of the non-custodial monitoring data subset is added to the detail vector representation cache; When the number of overall vector representations contained in the overall vector representation cache is not less than the set capacity, based on vector space similarity, performing vector integration processing on the overall vector representations of multiple non-custodial monitoring data subsets in the overall vector representation cache, and adding the overall vector representation to be added to the overall vector representation cache; And, for each data slice, when the number of corresponding detail vector representations contained in the detail vector representation cache is not less than the set capacity, based on the vector space similarity, vector integration processing is performed on the corresponding detail vector representations of multiple non-custodial monitoring data subsets in the detail vector representation cache, and the detail vector representation to be added is added to the detail vector representation cache.

4. The method according to claim 3, characterized in that Before respectively extracting the first cross-domain weighted focusing vector representation between the internal weighted focusing vector representation and each overall vector representation in the overall vector representation combination by using the cross-domain weighted focusing strategy, and respectively extracting the second cross-domain weighted focusing vector representation between the internal weighted focusing vector representation and each detail vector representation in the detail vector representation combination, the method further includes: When the set time is triggered, obtaining the overall vector representation combination contained in the overall vector representation buffer, and obtaining the detail vector representation combination contained in the detail vector representation buffer; Perform vector integration processing on target vector representations of multiple non-custodial surveillance data subsets based on one of the following methods: Method 1: Based on the similarity of the vector space, the target vector representations of the adjacent non-custodial surveillance data subsets with the highest similarity in the vector representation cache are mean-fused; Method 2: Based on the similarity of the vector space, the target vector representation to be added and the target vector representation already contained in the vector representation cache are clustered, and the target vector representations of each non-custodial monitoring data subset after clustering are mean-fused; The target vector representation is an overall vector representation or a detailed vector representation, and correspondingly, the vector representation cache is an overall vector representation cache or a detailed vector representation cache.

5. The method according to claim 4, characterized in that If there are multiple sets of adjacent non-custodial monitoring data subsets with the greatest similarity, then the step of performing mean fusion on the target vector representations of the adjacent non-custodial monitoring data subsets with the greatest similarity in the vector representation cache includes: The target vector representations of the adjacent non-custodial monitoring data subsets that are stored the longest and have the highest similarity in the vector representation cache are mean-fused; or, the target vector representations of each group of adjacent non-custodial monitoring data subsets that have the highest similarity in the vector representation cache are respectively mean-fused.

6. The method according to claim 1, characterized in that The non-custodial control analysis algorithm is calibrated using the following process: According to the non-custodial control analysis algorithm to be adjusted and optimized, feature extraction is performed on each non-custodial monitoring data subset in the non-custodial monitoring training data set to obtain the corresponding overall vector representation and detail vector representation, and according to the vector space similarity, the obtained overall vector representation set and detail vector representation set are simplified to obtain the overall vector representation combination and detail vector representation combination of the set capacity; After fusing the first adaptive search vector representation to be tuned and optimized with the second adaptive search vector representation, extracting a corresponding internal weighted focused vector representation through an internal weighted focused strategy; the internal weighted focused vector representation integrates the core information in the first adaptive search vector representation and the second adaptive search vector representation; By using a cross-domain weight focusing strategy, respectively extracting a first cross-domain weight focusing vector representation between the internal weight focusing vector representation and each overall vector representation in the overall vector representation combination, and a second cross-domain weight focusing vector representation between the internal weight focusing vector representation and each detail vector representation in the detail vector representation combination; Based on the errors between the non-custodial performance recognition results and non-custodial assessment information of the non-custodial monitoring training data set obtained by inferring the first cross-domain weighted focusing vector representations and the second cross-domain weighted focusing vector representations, and the corresponding calibration supervision information, the algorithm parameters of the non-custodial control and analysis algorithm are calibrated, and the calibrated non-custodial control and analysis algorithm is used to determine the non-custodial performance recognition results of the target non-custodial monitoring data set.

7. The method according to claim 6, characterized in that The non-custodial control analysis algorithm includes a feature integration component based on an internal weight focus strategy and a cross-domain weight focus strategy architecture, wherein the feature integration component is used to extract the first cross-domain weight focus vector representation and the second cross-domain weight focus vector representation; then, based on each obtained first cross-domain weight focus vector representation and each second cross-domain weight focus vector representation, the non-custodial performance recognition result and non-custodial evaluation information of the non-custodial monitoring training data set are inferred, including: In the algorithm pre-training phase, the first cross-domain weighted focused vector representations and the second cross-domain weighted focused vector representations are fused and mapped into monitoring data vector representations that match the non-detention evaluation vector representation dimension of the non-detention monitoring training data set; the first non-detention evaluation guidance information is used as input information, combined with the monitoring data vector representation, and input into the field general algorithm to obtain the non-detention evaluation information of the non-detention monitoring training data set; In the algorithm fine-tuning stage, the first cross-domain weighted focused vector representation and the second cross-domain weighted focused vector representation are fused and mapped into a monitoring data vector representation that matches the non-custodial assessment vector representation dimension of the non-custodial monitoring training data set; the non-custodial assessment vector representation and the second non-custodial assessment guidance information are used as input information and combined with the monitoring data vector representation to be input into the domain-general algorithm to obtain the non-custodial performance recognition result of the non-custodial monitoring training data set.

8. The method according to claim 7, characterized in that The algorithm parameters of the non-custodial control analysis algorithm are adjusted based on the errors between the non-custodial performance recognition results and non-custodial assessment information of the non-custodial monitoring training data set obtained by inference and the corresponding adjustment supervision information, including: In the algorithm pre-training phase, based on the error between the non-custodial assessment information obtained by inference and the non-custodial assessment information in the calibration supervision information, the algorithm parameters of the feature integration component are calibrated; In the algorithm fine-tuning stage, algorithm parameters of the feature integration component and the domain-general algorithm are calibrated based on the error between the non-custody performance recognition result obtained by reasoning and the non-custody performance recognition result in the calibration supervision information.

9. A management and control system, characterized in that: include: one or more processors; Memory; one or more computer programs; The one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method according to any one of claims 1 to 8 is implemented.

Images