A special object risk management and control prediction method and system based on BERT

Abstract

The application discloses a special object risk management and control prediction method and system based on BERT, and the method comprises the following steps: collecting characteristic data of special objects in past traffic accident cases, and establishing a special object sample library; building a risk management and control prediction model, and training the model by using the special object sample library; collecting characteristic data of special objects to be monitored, establishing a special object monitoring database and updating it in time; calling the risk management and control prediction model, predicting the control means of the special objects to be monitored, and sending early warning information. The application solves the problems of the lack of risk early warning and prevention ability of public security organs for special objects and the incompleteness and subjectivity of traditional human intelligence by constructing a special object sample library and training a risk management and control prediction model, and giving reasonable control measures or processing suggestions for the risks possibly caused by special objects through comprehensive analysis of various text information of the special objects.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of data analysis and artificial intelligence, and particularly relates to a special object risk control prediction method and system based on BERT. BACKGROUND

[0002] In the current public security or police work, the risk control of special objects is a key work and a difficult work to ensure social stability. After the special objects cause trouble and disaster, they often cause great difficulties in the work of the public security organs for the relevant post-disaster disposal. Before the trouble and disaster, the public security organs mainly rely on human intelligence to collect relevant action and early warning information. However, according to the traditional human intelligence, it is impossible to be comprehensive and objective, and it is often necessary to consume a large amount of police resources to conduct investigation and visit and individual check on the special object groups with high proportion, which cannot timely and effectively prevent the risk.

[0003] In order to prevent and control the work from the source and fundamentally reduce the trouble and disaster behavior of special objects, the public security or police system needs to break through the data barriers of various related departments, use big data technology to comprehensively analyze the relevant information of special objects, and give reasonable control measures or processing suggestions for the risks that may be caused by special objects. SUMMARY

[0004] In order to solve the problem of lack of risk early warning and prevention ability of public security organs for special objects and the problem of incomprehensive and objective traditional human intelligence, the present application provides a special object risk control prediction method and system based on BERT, which collects information data of special objects by using the information acquisition authority of the public security or police system, and reasonably predicts the risk control of special objects by constructing a large language model.

[0005] The technical scheme provided by the present application is as follows:

[0006] A special object risk control prediction method based on BERT, comprising the following steps:

[0007] S1. Collecting characteristic data of special objects in past trouble cases, and establishing a special object sample library;

[0008] S2. Building a risk control prediction model, and completing model training by using the special object sample library;

[0009] S3. Collecting characteristic data of special objects to be monitored, establishing a special object monitoring database and updating in time;

[0010] S4. Invoking the risk control prediction model, predicting the control means of the special object to be monitored, and sending early warning information.

[0011] Further, in step S1, for special objects involving past traffic accident cases with clear handling results, personal information, family information, community information, medical information and police information of the special objects within a period of time before the traffic accidents are collected and arranged into personal information texts, family information texts, community information texts, medical information texts and police information texts, and the handling results of the traffic accidents are taken as sample labels to form text data samples; a special object sample library is constructed to store a plurality of text data samples of a plurality of special objects.

[0012] Further, step S2 specifically includes the following steps:

[0013] S201. Personal information text classification network, family information text classification network, community information text classification network, medical information text classification network and police information text classification network are respectively built, each information text classification network including a full connection layer, a Softmax layer and a classification output layer;

[0014] S202. The personal information texts, the family information texts, the community information texts, the medical information texts and the police information texts in the special object sample library are respectively input into the BERT model to obtain corresponding text semantic features, and the text semantic features are taken as the input of the full connection layer of the information text classification network to train the personal information text classification network, the family information text classification network, the community information text classification network, the medical information text classification network and the police information text classification network respectively;

[0015] S203. A risk management and control prediction model is built, including a full connection layer, a Softmax layer and a classification output layer; the full connection layer is formed by sequentially splicing the full connection layers of the trained personal information text classification network, the family information text classification network, the community information text classification network, the medical information text classification network and the police information text classification network, and the network parameters thereof are reserved as the initial parameters of the full connection layer of the risk management and control prediction model during training;

[0016] S204. The text semantic features obtained by inputting the personal information texts, the family information texts, the community information texts, the medical information texts and the police information texts in the special object sample library into the BERT model are sequentially spliced and taken as the input of the full connection layer of the risk management and control prediction model to complete the training of the risk management and control prediction model.

[0017] Preferably, the full connection layers of the risk management and control prediction model and each information text classification network all adopt the Relu function as the activation function, and the loss function adopted during training is also the cross-entropy loss function.

[0018] Further, when extracting the text semantic features using the BERT model, a sliding window with a width of N characters is set, and the sliding window is slid from the head of the information text to the tail of the information text by a step D to obtain a plurality of subtexts. The subtext semantic features obtained by inputting each subtext into the BERT model are sequentially spliced to serve as the text semantic features of the entire information text. The width N of the sliding window is greater than the step D.

[0019] Further, in step S3, personal information, family information, community information, medical information and police information of the special object to be monitored are collected, and the personal information, the family information, the community information, the medical information and the police information are respectively arranged into personal information text, family information text, community information text, medical information text and police information text, and are stored in the special object monitoring database in pieces and are updated in time.

[0020] Further, in step S4, for each special object in the monitoring database, the personal information text, the family information text, the community information text, the medical information text and the police information text of the special object are sequentially spliced after being input into the text semantic features obtained by the BERT model, and are input into the trained risk control prediction model to obtain probability values of various handling results of the special object. Then, the handling result with the maximum probability value is taken as the control means, and the control means is sent to the grassroots police or grid administrator of the community where the special object is located through the short message platform.

[0021] A special object risk control prediction system based on the above special object risk control prediction method, comprising:

[0022] A special object sample library for storing text data samples of special objects in past traffic accidents;

[0023] A special object monitoring database for storing various information texts of special objects to be monitored, including personal information text, family information text, community information text, medical information text and police information text;

[0024] A semantic feature extraction module integrated with a BERT model for extracting text semantic features of information texts;

[0025] A risk control prediction module integrated with a trained risk control prediction model for calculating and outputting control means of special objects to be monitored;

[0026] A system control module for retrieving various information texts of special objects to be monitored from the special object monitoring database and sending the information texts to the semantic feature extraction module to obtain corresponding text semantic features, and for receiving various text semantic features from the semantic feature extraction module and sending the text semantic features to the risk control prediction module after splicing the text semantic features;

[0027] The information sending module is configured to obtain the control means of the special object to be monitored from the risk control prediction module, and send the control means to the corresponding grassroots police or grid administrator through a short message platform.

[0028] Further, the system control module further comprises a sliding window processing submodule, configured to divide the information text into a plurality of subtexts in the form of a sliding window, and send each subtext to the semantic feature extraction module.

[0029] Further, the system control module further comprises a sliding window processing submodule, configured to divide the information text into a plurality of subtexts in the form of a sliding window, and send each subtext to the semantic feature extraction module.

[0030] Advantages of the present application:

[0031] By constructing the special object sample library and training the risk control prediction model based on BERT, the various text information of the special object is comprehensively analyzed, and reasonable control measures or processing suggestions are given for the risks possibly caused by the special object, so that the problems of lack of risk early warning and prevention ability of the public security organ and incompleteness and subjectivity of the traditional human intelligence are solved. In addition, the present application improves the effectiveness and accuracy of the risk control prediction model through the two-stage model construction and training mode, and solves the problem that the BERT model cannot be applied to long text in the form of a sliding window. BRIEF DESCRIPTION OF DRAWINGS

[0032] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, illustrate the present application together with the embodiments thereof, and explain the present application, but do not limit the present application.

[0033] Figure 1 is a flow chart of a special object risk control prediction method provided by an embodiment of the present application;

[0034] Figure 2 is a first stage schematic diagram of the risk control prediction model during training, provided by an embodiment of the present application;

[0035] Figure 3 is a second stage schematic diagram of the risk control prediction model during training, provided by an embodiment of the present application;

[0036] Figure 4 is a long text semantic feature extraction method schematic diagram based on a sliding window, provided by an embodiment of the present application;

[0037] Figure 5 is a module composition schematic diagram of a special object risk control prediction system, provided by an embodiment of the present application. DETAILED DESCRIPTION

[0038] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0039] As shown in Figure 1 The special object risk management and control prediction method provided by the present application mainly includes the following steps:

[0040] S1. Collect the characteristic data of special objects in past traffic accident cases, and establish a special object sample library.

[0041] For special objects involved in past traffic accident cases and having clear handling results, collect their various types of characteristic information within a period of time before the accident, such as personal information, family information, community information, medical information and police information. Among them, the personal information includes gender, age, personality, education and work experience, etc., the family information includes family members, economic income, family medical history and major family changes, etc., the community information includes residence time, neighborhood relationship, extreme remarks and abnormal behavior, etc., the medical information includes medical records, frequency of visiting doctors, drug purchase situation (drug type and indication, purchase amount), etc., and the police information includes case history (previous offenses / criminal offenses / contradictions) and case records, etc.

[0042] For each special object, the various types of characteristic information collected about it are sorted into corresponding information texts respectively, obtaining personal information text W1, family information text W2, community information text W3, medical information text W4 and police information text W5. The handling result of the special object in the traffic accident is taken as a sample label Y. In the embodiments of the present application, the handling result of the special object has three kinds of visit and investigation, send to hospital for treatment and strengthen control, that is, the sample label has three categories. Finally, the information text and the sample label jointly constitute a text data sample {W1, W2, W3, W4, W5, Y}. Similarly, the information texts and sample labels of several special objects are collected and sorted to construct a special object sample library containing several text data samples, and they are divided into a training sample library and a test sample library in a certain proportion, such as taking 80% of the special object sample library as the training sample library and the remaining 20% as the test sample library. Of course, in order to ensure the effectiveness of the subsequent prediction model, the special object sample library itself should have a certain scale, such as containing at least hundreds or even thousands of text data samples.

[0043] S2. Build a risk management and control prediction model, and complete model training using the special object sample library.

[0044] The purpose of building and training the risk control prediction model is to extract the text semantic features of the information text through the natural language model, and then accurately classify by using the classification network to obtain the corresponding category, that is, the handling result of the traffic accident. The natural language model used in the text semantic feature extraction in the application is BERT, which is Bidirectional Encoder Representations from Transformer. In simple terms, BERT is an evolved version of Word2Vec, which uses word vectors to represent natural language, but its model depth is extremely large and the number of parameters is particularly large. Taking the BERT-BASE model as an example, it contains 12 hidden layers, each hidden layer has a dimension of 768, each layer contains 12 attention heads, and there are a total of 110M parameters. Only the model parameter file occupies 400MB of space on the hard disk.

[0045] BERT is developed on the basis of the attention mechanism-based Transformer model. By discarding the Decoder module in the Transformer (only keeping the Encoder), BERT has bidirectional encoding capability and strong feature extraction capability. The application adopts the Chinese BERT model BERT-BASE Chinese opened by Google Research, and the specific reference can be made to https: / / github.com / google-research / bert. The Chinese version of the model is trained by Wikipedia, and unlike the word-based model in English version, the Chinese version is a single Chinese character-based model. The application only uses it as a kind of text semantic feature extractor to obtain the text semantic features of the information text of the special object. The specific model configuration parameter information is shown in Table 1.

[0046] Table 1 BERT-BASE Chinese model configuration parameter information

[0047]

[0048] After the text semantic features are extracted, classification can be performed. A conventional method is to splice the semantic feature vectors of information texts of various types extracted by the BERT-BASE model respectively, and then input them into a trained full connection layer and a Softmax layer to obtain classification probabilities of various types. However, we know that various feature information of a special object does not contribute equally to the result of a traffic accident and the result of handling the traffic accident, or in other words, different feature information and the result of a traffic accident and the result of handling the traffic accident do not have the same correlation. For example, medical information and police information of a special object in a short period of time are often more valuable for judging whether the object is likely to cause a traffic accident in a short period of time and are also more referential for handling measures after the traffic accident, while community information, due to its lack of credibility and integrity, although it has some correlation with whether the special object causes a traffic accident, the correlation is often far less than medical information and police information.

[0049] Therefore, the present application does not directly use a conventional method when building and training a risk control prediction model, but divides the training process into two stages. As shown in Figure 2 In the first stage, a separate classification network including a full connection layer and a Softmax layer is trained by using information texts of various types in a training sample library and corresponding sample labels. Specifically, for personal information texts, a personal information text classification network is built, including a first full connection layer FC1, a Softmax layer and a classification output layer connected in sequence. The classification output layer includes output nodes corresponding to three types of handling results of visit and verification, medical treatment and control enhancement, and the output nodes output classification probabilities of the corresponding handling results. The text semantic feature vector obtained by inputting the personal information text into the BERT-BASE model is input into the first full connection layer FC1 to complete forward propagation and error back propagation of the network, and after a certain period of iterative training, the accuracy of the model is tested by using personal information texts in a test sample library and corresponding sample labels. If the accuracy reaches a predetermined threshold, the training of the personal information text classification network is completed. Similarly, a family information text classification network, a community information text classification network, a medical information text classification network and a police information text classification network are built and trained respectively to obtain network parameters of the trained second full connection layer FC2, the third full connection layer FC3, the fourth full connection layer FC4 and the fifth full connection layer FC5.

[0050] As shown in Figure 3In the second stage, the final risk control prediction model is built and trained, including a full connection layer, a Softmax layer and a classification output layer connected in sequence, wherein the full connection layer is spliced by the first full connection layer FC1, the second full connection layer FC2, the third full connection layer FC3, the fourth full connection layer FC4 and the fifth full connection layer FC5 in the first stage in sequence, and when the network is initialized, the full connection layer parameters use the network parameters of the second full connection layer FC2, the third full connection layer FC3, the fourth full connection layer FC4 and the fifth full connection layer FC5 trained in the first stage, and the other layer parameters are randomly initialized; for each sample in the training sample library, the personal information text, the family information text, the community information text, the medical information text and the police information text are input into the text semantic feature vectors obtained by the BERT-BASE model in sequence, and then input into the full connection layer, the forward propagation and the error back propagation are completed, and after a certain period of iterative training, the accuracy of the risk control prediction model is calculated by using the test sample library, and the training is stopped when the accuracy reaches the preset precision.

[0051] It should be noted that the activation functions of the first full connection layer FC1, the second full connection layer FC2, the third full connection layer FC3, the fourth full connection layer FC4 and the fifth full connection layer FC5 in the embodiment of the application are all selected as Relu functions, and the loss functions used in network training are all cross-entropy loss functions.

[0052] In addition, in order to shorten the network training time, the test accuracy of the family information text classification network, the family information text classification network, the community information text classification network, the medical information text classification network and the police information text classification network in the first stage reaches 80%, and the training of the corresponding network can be stopped. In order to ensure the effectiveness of the risk control prediction model, the test accuracy thereof usually reaches more than 95%.

[0053] Since BERT uses a multi-layer attention mechanism, it can process longer texts, but it still has a certain limit on the length of the input text, i.e. the length of the input text cannot exceed 512 characters. Although the present application has divided the feature information of the special object into personal information text, family information text, community information text, medical information text and police information text, and input them into the BERT-BASE model respectively when extracting the semantic features of the text, there will still be cases exceeding 512 characters in actual use, especially for police information text. The conventional processing method is to segment the input text into subtexts with a length of no more than 512 characters, and then input the subtexts into the BERT model respectively, and then concatenate the obtained semantic features to obtain the semantic features of the entire text. Or directly discard part of the text content, directly cut off the content of no more than 512 characters at the head, tail or middle of the text and input it into the model. However, segmenting the text will destroy the context relationship of the text, and even distort the text semantics, while directly cutting off will lose part of the text content, which is extremely likely to cause the loss of important information, especially for highly condensed texts such as police information text.

[0054] Therefore, based on the sliding window principle, the present application further provides a long text semantic feature extraction method. As shown in Figure 4 For a long text W exceeding 512 characters, a text sliding window with a window width of N characters (N < 512) is set, which is slid from the head of the long text W to the tail of the text by a step D, and the subtext in the sliding window is input into the BERT model, and then the obtained subtext semantic features are concatenated in turn to obtain the semantic features of the entire long text. When the step D and the width of the sliding window are the same, it is equivalent to the segmentation processing in the above conventional method. However, in order to overcome the shortcomings of segmentation processing, the step D is smaller than the width of the sliding window, so that there is an overlapping part between the adjacent two subtexts obtained by the sliding window, which not only overcomes the limitation of the length of the input text of the BERT model, but also effectively alleviates the problem of loss of context relationship caused by directly segmenting the long text.

[0055] In some embodiments, the width of the sliding window is 500 characters, and the sliding step D is 400 characters, i.e. there is an overlapping part of 100 characters between the adjacent two subtexts obtained by the sliding window, and the sliding window slides 5 times to cover a long text of 500+400x5=2500 characters.

[0056] S3. Collecting the feature data of the special object to be monitored, establishing a special object monitoring database and updating it in time.

[0057] Similar to step S1, the personal information, family information, community information, medical information and police information of the special object to be monitored are collected and sorted into personal information text, family information text, community information text, medical information text and police information text respectively, and are stored in the special object monitoring database and updated in time.

[0058] S4. Calling the risk control prediction model to predict the control means of the special object to be monitored and sending the early warning information.

[0059] For each special object in the monitoring database, the personal information text, family information text, community information text, medical information text and police information text are input into the text semantic feature vector obtained by the BERT-BASE model in turn after being spliced, and then input into the trained risk control prediction model to obtain the probability value of the three processing results of visit and verification, medical treatment and strengthened control, and finally the processing result corresponding to the maximum probability value is taken as the control means and sent to the grassroots police or grid administrator of the community where the special object is located through the short message platform.

[0060] It is worth noting that in order to further improve the generalization ability and accuracy of the risk control prediction model, we need to continuously expand the size of the special object sample library, use the constantly enriched new samples to perform model iteration on the basis of the parameters of the trained risk control prediction model, and the iteration process is similar to the second phase in step S2, except that the learning rate of the iteration training should be set to a smaller value to prevent model drift.

[0061] Based on the above special object risk control prediction method, the present application also provides a special object risk control prediction system, such as Figure 5As shown, the risk control and prediction system mainly comprises a special object sample library, a special object monitoring database, a semantic feature extraction module, a risk control and prediction module, a system control module and an information sending module. The special object sample library is used to store text data samples of special objects in past traffic accident cases. The special object monitoring database is used to store various information texts of the special object to be monitored, including personal information texts, family information texts, community information texts, medical information texts and police information texts. The semantic feature extraction module is integrated with a BERT model and is used to extract semantic features of the information texts. The risk control and prediction module is integrated with a trained risk control and prediction model and is used to calculate and output control measures of the special object to be monitored. The system control module is used to retrieve various information texts of the special object to be monitored from the special object monitoring database, send the information texts to the semantic feature extraction module to obtain corresponding text semantic features, receive various text semantic features from the semantic feature extraction module, and send the text semantic features to the risk control and prediction module after splicing. The information sending module is used to obtain the control measures of the special object to be monitored from the risk control and prediction module and send the control measures to corresponding grassroots police officers or grid administrators through a short message platform.

[0062] In some embodiments, the system control module further comprises a sliding window processing submodule, which is used to divide the information texts into a plurality of subtexts in the form of a sliding window.

[0063] In some embodiments, the risk control and prediction system further comprises an iterative updating module, which is used to iteratively train the risk control and prediction model according to new samples in the special object sample library and update model parameters in the risk control and prediction module.

[0064] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; under the idea of the present application, the technical features of the above embodiments or different embodiments can also be combined, and there are many other changes of different aspects of the present application as described above. In order to be simple, they are not provided in details; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A special object risk management and prediction method based on BERT, characterized in that, Includes the following steps: S1. Collect characteristic data of special objects in past hit-and-run cases and establish a sample database of special objects: For special individuals involved in past incidents with clear outcomes, their personal, family, community, medical, and police information from a period prior to the incident is collected and compiled into corresponding personal, family, community, medical, and police information texts. The incident handling results are used as sample tags to form text data samples. A special individual sample library is constructed to store several text data samples of several special individuals. S2. Build a risk management and prediction model, and train the model using a special object sample library: S201. Construct personal information text classification networks, family information text classification networks, community information text classification networks, medical information text classification networks, and police information text classification networks respectively. Each information text classification network includes a fully connected layer, a Softmax layer, and a classification output layer. S202. Input the personal information text, family information text, community information text, medical information text and police information text from the special object sample library into the BERT model to obtain the corresponding text semantic features. Then, use the text semantic features as the input to the fully connected layer of the information text classification network to train the corresponding personal information text classification network, family information text classification network, community information text classification network, medical information text classification network and police information text classification network respectively. S203. Construct a risk management and control prediction model, including a fully connected layer, a softmax layer, and a classification output layer; the fully connected layer is formed by sequentially splicing the fully connected layers of the trained personal information text classification network, family information text classification network, community information text classification network, medical information text classification network, and police information text classification network, and retaining their network parameters as the initial parameters of the fully connected layer when training the risk management and control prediction model; S204. Input the personal information text, family information text, community information text, medical information text and police information text from the special object sample library into the BERT model respectively. The resulting text semantic features are then concatenated and used as the input to the fully connected layer of the risk management prediction model to complete the training of the risk management prediction model. S3. Collect characteristic data of special objects to be monitored, establish a special object monitoring database and update it in a timely manner; S4. Call the risk management prediction model to predict the management measures for the special objects to be monitored and send early warning information.

2. The special object risk management and prediction method as described in claim 1, characterized in that, The fully connected layers of the risk management prediction model and each information text classification network use the ReLU function as the activation function, and the loss function used during training is also the cross-entropy loss function.

3. The special object risk management and prediction method as described in claim 1, characterized in that, When using the BERT model to extract semantic features of text, a sliding window with a width of N characters is set. The window is slid from the beginning to the end of the text with a step size D to obtain several sub-texts. The semantic features of each sub-text are then input into the BERT model and concatenated to form the semantic features of the entire text. The width N of the sliding window should be greater than the step size D.

4. The special object risk management and prediction method as described in claim 1, characterized in that, In step S3, personal information, family information, community information, medical information, and police information of the special individuals to be monitored are collected, and these are organized into corresponding personal information text, family information text, community information text, medical information text, and police information text, and stored one by one in the special individual monitoring database and updated in a timely manner.

5. The special object risk management and prediction method as described in claim 4, characterized in that, In step S4, for each special object in the monitoring database, its personal information text, family information text, community information text, medical information text, and police information text are respectively input into the semantic features obtained by the BERT model and then concatenated. These features are then input into the trained risk management and prediction model to obtain the probability values ​​of various accident handling results. The accident handling result with the highest probability value is then used as the management and control measure and sent to the grassroots police officer or grid administrator in the community where the special object is located via SMS platform.

6. A special object risk management and prediction system based on the special object risk management and prediction method of claim 1, characterized in that, include: A special object sample library is used to store text data samples of special objects from past accident cases; The special object monitoring database is used to store various types of information texts of special objects to be monitored, including personal information texts, family information texts, community information texts, medical information texts, and police information texts; The semantic feature extraction module integrates the BERT model to extract the semantic features of information text. The risk management and prediction module integrates a trained risk management and prediction model, which is used to calculate and output the management and control measures for the special objects to be monitored. The system control module is used to retrieve various information texts of the special objects to be monitored from the special object monitoring database and send them to the semantic feature extraction module to obtain the corresponding text semantic features. It is also used to receive various text semantic features from the semantic feature extraction module, and then send them to the risk control prediction module after splicing them together. The information sending module is used to obtain the control measures for the special objects to be monitored from the risk control and prediction module, and send them to the relevant grassroots police officers or grid administrators through the SMS platform.

7. The special object risk management and prediction system as described in claim 6, characterized in that, The system control module also includes a sliding window processing submodule, which is used to divide the information text into several sub-texts in the form of a sliding window, and send each sub-text to the semantic feature extraction module.

8. The special object risk management and prediction system as described in claim 6, characterized in that, It also includes an iterative update module, which is used to iteratively train the risk management prediction model based on newly added samples in the special object sample library and update the model parameters in the risk management prediction module.

Images