Multi-police cooperative combat system and method based on workflow engine

Abstract

The invention provides a multi-police collaborative combat system and method based on a workflow engine, and belongs to the field of information collaboration. The problem of low police resource transfer efficiency is solved; the specific system is as follows: a data fusion module for summarizing service data and constructing a decentralized block chain network; the police strength distribution module is used for predicting the number of cases in each region and distributing police strength; the case analysis module is used for extracting and separating case description information and case judgment information corresponding to each piece of case data, and training a text model; selecting a path for the police car by combining the speed of the police car according to the positions of the police car and the crime area; through efficient sharing and collaborative analysis of multi-police information, data of related departments are rapidly processed by using big data analysis and an artificial intelligence algorithm, a scientific basis is provided for decision making of the related departments, and the case detection efficiency and the social public security prevention and control capability are improved.

Description

TECHNICAL FIELD

[0001] The application discloses a multi-police force cooperative combat system and method based on a workflow engine and relates to the field of information cooperation. BACKGROUND

[0002] The existing police force cooperative combat has the following defects: Inter-police force data is scattered: The data of different police forces in related departments is in a scattered state, different police forces such as criminal investigation, intelligence and technical investigation have their own independent data systems, the data standards and formats are not unified, and there are strict data permission restrictions; for example, when a grassroots police handles a cross-regional personnel track tracking case, the police needs to obtain network monitoring data of the network security department and case data of the criminal investigation department, due to the lack of unified data standards and interface specifications, the police often needs to repeatedly coordinate and communicate among different police forces, which reduces the case investigation efficiency.

[0003] Crime prediction relies on experience: There is a lack of data analysis and model support; the police roughly judges the crime-prone areas and time periods according to the past work experience and the understanding of the public security situation in the jurisdiction, but this method is greatly influenced by subjective factors, and the accuracy is difficult to guarantee.

[0004] Decision efficiency: The data display mode is often abstract, and is mostly presented in the form of a text report and a simple table, and relevant department personnel needs to spend a lot of time to understand and analyze the data, and it is difficult to quickly obtain key information; for example, the command personnel needs to read and analyze a large number of text reports and data tables one by one to understand the crime occurrence situation, police deployment situation and the like in each region, which greatly reduces the decision efficiency. SUMMARY

[0005] In view of the defects in the prior art, the application aims to provide a multi-police force cooperative combat system and method based on a workflow engine, and aims to solve the problem of low police resource mobilization efficiency.

[0006] In order to achieve the above-mentioned purpose, the application is implemented by the following technical scheme: the multi-police force cooperative combat system based on the workflow engine comprises: A data fusion module is used for collecting the business data of each police force in each police station, and a decentralized block chain network is constructed; A police force allocation module is used for separating the case data in the business data, and according to the actual case quantity and case type in each police station responsible for the jurisdiction, the case occurrence quantity of each region is predicted, and the police force is allocated; Case analysis module: used for extracting and separating the case description information and case judgment information corresponding to each case data; the description information and the case judgment information are respectively used as two information sets to train a text model; the text model is used to summarize and classify each case in the police station jurisdiction; Real-time monitoring of whether a case occurs in the jurisdiction; if not, it is not processed; if so, according to the location of the police car and the case area, combined with the speed of the police car, the path of the police car is selected.

[0007] Further, the workflow of the data fusion module is as follows: Take the alliance chain as the framework, deploy at least one verification node and one observation node for each police station; deploy a sorting node for the relevant departments; Generate an initial block and define network rules; each police station authenticates its identity through a CA agency, obtains a digital certificate of the blockchain, and transmits business data into the blockchain; Through federated learning collaboration and encrypted query deployment of privacy protection layer to realize the functions of cross-agency data collaboration and query and data protection; Federated learning collaboration: the relevant departments act as coordinators to uniformly release an initial information query model architecture, clearly define the query object, query range, input features, output style and hyperparameter range of business data, and determine the training target of each police station; Based on the training target, each police station trains the information query model using its own business data, performs local training, and generates gradient parameters; After all the police stations complete local training, the relevant departments aggregate the final training gradient parameters of each police station, update the parameters in the information query model, and broadcast them to all police stations; Encrypted query: ciphertext calculation and result safe acquisition; Data preparation: all data owners encrypt the queryable business data; Request encryption: the query party encrypts the query conditions using the public key of the data owner and the private key of the query party; Local matching: the data owner decrypts the query conditions of the query party, and according to the query conditions, encrypts the content to be queried by the query party with the private key of the query party to obtain query ciphertext; sets the access permission of the query ciphertext to be valid only for the query party, and then feeds back to the query party; Result decryption: the query party decrypts the query ciphertext with the private key to obtain the content to be queried.

[0008] Further, the workflow of the police force allocation module is as follows: Summarize the business data corresponding to each region in the jurisdiction of the police station, and extract case data from the business data; the case data includes the case location, the case time, and the case type; This serves as the area code for all districts within the police station's jurisdiction, and as the code for each case type in the operational data; By summarizing the number of Category I cases occurring in different regions at different times, we obtain Sequence A. （1） ~A （re） ;re represents the total number of regions; By summarizing the number of Category I cases occurring in the same month but in different years and regions, we obtain Sequence B. （1） ~B （12） ; Using the ARIMA model on sequence A （1） Time series analysis was performed on the data. Calculate the autocorrelation coefficient and moving average coefficient, as well as the seasonal autocorrelation coefficient and seasonal moving average coefficient, for Category I cases occurring in Region 1. Let sequence A （1） Chinese data is ~ Where ti represents sequence A （1） The number of time points corresponding to all data in the dataset; calculate ~ The average value aX （0） Calculate in sequence ~ Subtract aX （0） The difference is used as white noise for the first time series prediction; by ~ Using the original data, calculate the lag factor and seasonal lag factor of the original data. Substitute the autocorrelation coefficient and moving average coefficient, along with the seasonal autocorrelation coefficient and seasonal moving average coefficient, into the SARIMA model to calculate the first region in one seasonal cycle S. （s） The number of subsequent Category I cases ; Similarly, repeated prediction Using the same steps, the time is synchronized to the current month, and the number of Category I cases occurring in the 2nd to reth regions after the corresponding seasonal cycle is predicted as the time-series prediction quantity.

[0009] Furthermore, the subsequent process of the police force allocation module is as follows: For sequence B （1） Spatial dimension analysis is performed on the data in the dataset; With sequence B （1） Using the earliest occurrence of Category I cases as a benchmark, and with months as the time dimension, sequence B... （1） All data in the system is synchronized with the time. Construct matrices of the number of Category I cases occurring in January for each of the 1st to the reth regions in ascending order of time; use the matrix of the number of Category I cases occurring in the 1st region in January as the baseline matrix, and calculate the correlation coefficient between the baseline matrix and the other matrices. Using the correlation coefficients of the baseline matrix relative to other matrices as weights, a weight matrix for the first region is constructed; using maximum likelihood combined with a spatial lag model, the spatial autoregressive coefficients of the number of Class I cases occurring in the first region in January and the number of Class I cases occurring in other regions in January are calculated. According to sequence B （1） To sequence B （12） Calculate the spatial autoregressive coefficients of the first to the re regions from January to December; Based on the current month, extract the spatial autoregressive coefficients of the first to the re regions in the corresponding month, and substitute the spatial autoregressive coefficients into the spatial lag model to predict the number of Class I cases occurring in the first to the re regions in the current month, which will be used as the spatial prediction quantity. Calculate the product of the number of temporal predictions and the number of spatial predictions for each region, and use it as the inspection coefficient for each region; Police resources are allocated to each area based on the patrol coefficient ratio of each area and the actual number of police officers in the police station.

[0010] Furthermore, the workflow of the case analysis module is as follows: Extract and separate the case description information and case judgment information corresponding to each case data; use the description information as information set A and the case judgment information as information set B, and create an index for the case judgment information corresponding to each case description information; Based on the knowledge distillation mechanism, the BERT model is used as the teacher model and the RNN model is used as the student model. The text model is trained using information set A and information set B. The system monitors in real time whether any incidents occur within its jurisdiction; if no incidents occur, no action is taken; if incidents occur, the system selects a route for the police car based on its location, the location of the incident, and its speed.

[0011] Furthermore, the specific process for training the text model is as follows: The text information corresponding to each case in information set A is used as a subset; the judgment information of the cases corresponding to each subset in information set B is used as the main set. Using the BERT model, all text information in information set A and information set B is converted into text vectors; The first subset in information set A is taken as set A. （te） Set A （te） The corresponding principal set is set B. （st） ; Using the text vectors corresponding to the subset as the question objects and the text vectors corresponding to the main set as the answer objects, the BERT model is trained to fill in the blanks and a teacher model is trained. The RNN model is used as the student model, and the student model is used for fill-in-the-blank learning.

[0012] Furthermore, the route selection process for the police vehicle is as follows: Let the area where the police car is located be region m, and the crime scene be region n; select a route for the police car based on the different routes from region m to region n; Scenario 1: There are no other regions among all the roads from region m to region n; Among all roads from region m to region n, select the route with less traffic and shorter distance for police car A as its target route; Scenario 2: There are other regions among all the roads from region m to region n, but the forward routes to these regions are unique; Repeat the selection criteria from scenario one to choose a target route for police car A; Scenario 3: There are other regions among all the roads from region m to region n, and the routes to these regions are not unique; Count the total number of paths rn from each first-level region to region n, and count the traffic flow tr on the first to rn paths. （1） ~tr （rn） ; Get the movement speed v of police car A （甲） Calculate the latest arrival time t of police car A by selecting the longest path from path 1 to path n. （lat） Calculate the earliest arrival time t of police car A by selecting the shortest path. （ear） ; Get the number of other police cars pw on paths 1 to rn. （1） ,pw （2） ~pw （rn） ; Obtain the position and speed of each police car on paths 1 to rn, calculate the arrival time of each police car on path 1 to rn to region n, and sort them according to the arrival time to region n to obtain tw. （1,1） ~tw （rn,pw（rn）） ; Get the maximum traffic flow of the first path (tr) （max） Calculate the time cost (Ttw) of police car A on the first path. （1） .

[0013] Furthermore, the route selection process for the police vehicle also includes: In TW （1,1） ~tw（1,pw（1）） The statistics belong to the interval [t] （ear） , t （lat） The quantity lp （1） ; Get the total number of vehicles on the current first path (pp) （now） , judge (pp （now） +tr （1） +pp （b） ) <tr （max） Is it valid? If true, then in tw （1,1） tw （1,2） ~tw （1,pw（1）） Extract the (lp) （1） +1) Arrival time of the police car (ty) （ot） , Ttw （1） The value is: Ttw （1） =[(ty) （ot） -t （ear） ) + (ty （ot） -t （lat） )] / 2; If ty （ot） If it does not exist, then Ttw （1） The value is: Ttw （1） =[(tw （1,pw（1）） -t （ear） ) + (tw （1,pw（1）） -t （lat） )] / 2; If not true, then judge (pp) （now） +tr （1） +pp （b） =tr （max） Is it valid? If true, then in tw （1,1） tw （1,2） ~tw （1,pw（1）） Extract the (tp) （a） +1) Arrival time of the police car (ty) （ota） ; tp （a） tp represents the allowed traffic flow for the first route. （a） The formula for calculating tp is: （a） =tr （max） - (pp （now） +tr （1） +pp （b） )-1; Ttw （1） The value is: Ttw （1） =[(ty) （ota） -t （ear） ) + (ty（ota） -t （lat） )] / 2; If ty （ota） If it does not exist, then Ttw （1） The value is: Ttw （1） =[(tw （1,pw（1）） -t （ear） ) + (tw （1,pw（1）） -t （lat） )] / 2.

[0014] Furthermore, the route selection process for the police vehicle also includes: If (pp) （now） +tr （1） +pp （b） =tr （max） Not valid; Then obtain the total number of traffic lights rg on the first path, and the red light duration rt of the first, second, up to the rg-th traffic lights. （1） ,rt （2） ~rt （rg） ; Obtain the distance X from police car A to the first traffic light. （o-1） The distance X from the first traffic light to the second traffic light （1-2） The distance X from the second traffic light to the third traffic light （2-3） Similarly, the distance X from the (rg-1)th traffic light to the rgth traffic light... （（rg-1）-rg） ; Calculate the distance of police car A through X （1-2） Minimum speed of the corresponding path vs （1） : vs （1） =X （1-2） / rt （1） ; Calculate the latest time (tx) for police car A to arrive at the first traffic light. （甲） : tx （甲） = (2×X) （o-1） ) / (v （甲） +vs （1） ); Determine tx （甲） With rt （1） ,rt （2） ~rt （rg） Is the sum less than t? （lat） ; If it is less than X, then calculate the distance police car A passed through. （2-3） To X （（rg-1）-rg） Minimum speed of the corresponding path vs （2） to vs （rg） ; When Police Car A was driving in X （1-2） To X （（rg-1）-rg） When following the corresponding route, adjust the speed of police car A to vs. （2） to vs （rg） ; Ttw （1） The value is tx （甲） With rt （1） ,rt （2） ~rt （rg） The sum; If not less than, then Ttw （1） The value is t （lat） ; Repeat Ttw （1） The calculation process calculates the time cost Ttw for paths 2 to rn. （2） ~Ttw （rn） ; in Ttw （1） ~Ttw （rn） In the process, the path with the least time cost is selected as the target path for police car A, and the primary area on the target path is designated as the target area for police car A. Repeat the steps of selecting the target path and target area for police car A, and continue to select the path and area for police car A until police car A reaches area n.

[0015] Multi-agency collaborative operations methods based on workflow engines include: Aggregate operational data from various police departments within each police station and build a decentralized blockchain network; Separate case data from business data, and predict the number of cases in each area based on the actual number and type of cases within the jurisdiction of each police station, using time and location as dimensions, and allocate police resources accordingly; Extract and separate the case description information and case judgment information corresponding to each case data; use the description information and case judgment information as two separate information sets to train a text model; use the text model to summarize and classify various cases within the police station's jurisdiction; monitor in real time whether any cases occur within the jurisdiction; if no cases occur, no action is taken; if cases occur, select a route for the police car based on the location of the police car and the area where the incident occurred, combined with the speed of the police car.

[0016] Compared with the prior art, the beneficial effects of the present invention are: Achieving efficient information sharing: This invention establishes unified data standards and interface specifications to comprehensively collect and integrate business data from various police departments within relevant departments, as well as external social data. By utilizing data cleaning, transformation, and loading technologies, data of different formats is converted into a unified format and stored on a big data platform, breaking down the traditional data segmentation between police departments and enabling the free flow and sharing of data from different police departments such as criminal investigation, intelligence, technical investigation, cyber investigation, and video investigation.

[0017] Improving the accuracy of crime prediction: This invention uses machine learning algorithms to conduct in-depth analysis of fused data and construct a crime prediction model. Based on historical case data, time, location and other multi-dimensional data, the model can accurately predict high-crime areas and time periods. By analyzing the types of crimes, time of occurrence and surrounding environmental data of a certain area over a period of time, the model can predict in advance the types of crimes that may occur in the area in a specific future time period, providing a scientific reference for police deployment.

[0018] Providing intelligent decision support: This invention combines actual case data to provide commanders with comprehensive information and analysis for the types of cases discovered in real time, helping them to analyze and characterize cases; at the same time, when handling major public safety incidents, this invention can generate different route plans based on the actual location of police vehicles and the area where the incident occurred, reducing the risk of case spread and improving the efficiency of police resource utilization. Attached Figure Description

[0019] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the system of the present invention; Figure 2 This is a schematic diagram of the method of the present invention; Figure 3 This is a schematic diagram of the model construction of the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] Example 1 Please see Figure 1 The multi-agency collaborative combat system based on a workflow engine includes: Data fusion module: used to aggregate business data from various police departments within each police station (such as criminal investigation case data, cybersecurity network monitoring data, and public security patrol data), and to build a decentralized blockchain network (for data sharing of business data from each police station). The workflow of the data fusion module is as follows: Using the Hytrrledger Fabric consortium blockchain as a framework, deploy at least one verification node and one observation node for each police station; and deploy sorting nodes for relevant departments (such as the Ministry of Justice, the Ministry of State Security, and the Ministry of Emergency Management). The system generates initial blocks for relevant departments and defines network rules (such as node joining / leaving mechanisms and data storage periods); each police station authenticates its identity through a CA authority, obtains a digital certificate for the blockchain, and transmits its business data to the blockchain. By deploying a privacy protection layer through federated learning collaboration and encrypted queries, cross-agency data collaboration and querying capabilities, as well as data protection functions, are achieved. Federated Learning Collaboration: Relevant departments, acting as coordinators, uniformly release the initial information query model architecture (such as XGBoost or neural networks), clarify the query objects, query scope, input features, output styles and hyperparameter ranges of business data, and determine the training objectives of each police station; Each police station trains its information query model using its own business data based on the training objectives, performs local training, and generates gradient parameters (such as neural network weight matrices). After all police stations have completed local training, the relevant departments aggregate the final trained gradient parameters of each police station through the MPC protocol, and update the parameters in the information query model (based on the final trained gradient parameters of each police station) before broadcasting it to all police stations. Encrypted query: Securely retrieve encrypted results from encrypted computation; Data preparation: The data owner (such as Police Station B) encrypts the queryable business data; Request encryption: The querying party (such as Police Station A) uses the public key of the data owner (such as Police Station B) to encrypt the query conditions and the private key of the querying party (such as Police Station A); Local matching: The data owner (e.g., Police Station B) decrypts the query conditions of the querying party (e.g., Police Station A), and encrypts the content to be queried by the querying party (e.g., Police Station A) with the querying party's (e.g., Police Station A) private key according to the query conditions to obtain the query ciphertext; sets the access permission of the query ciphertext to be valid only for the querying party (e.g., Police Station A), and then sends it back to the querying party (e.g., Police Station A). Decryption of results: The querying party (such as Police Station A) uses its private key to decrypt the query ciphertext and obtain the content to be queried.

[0022] Police force allocation module: Used to separate case data from business data, and predict the number of cases in each area based on the actual number and type of cases in each police station's jurisdiction, and allocate police force according to time and location. The workflow of the police force allocation module is as follows: The system compiles operational data from various areas within the police station's jurisdiction and extracts case data from this data. Case data includes the location, time, and type of the crime. This serves as the area code for all districts within the police station's jurisdiction, and as the code for each case type in the operational data; Set the three-dimensional variable X （i，t，k） Represents: (Within the police station's jurisdiction) the number of cases k occurring in area i within time period t; (i.e., i represents the location of the incident, t represents the time of the incident, and k represents the type of incident number; X) （i，t，k） The overall representation indicates the type ID of a case. Considering only the time dimension, the number of cases of type I occurring in different regions at different times is arranged in chronological order to obtain sequence A. （1） ~A （re） ;re indicates the total number of areas within the police station's jurisdiction; Considering only the spatial dimension, the number of cases of type I occurring in the same month but in different years and different regions is grouped into sequence B according to the region number and the chronological order of the year. （1） ~B （12） (That is, the number of Category I cases occurring in the entire region from January to December); Using the ARIMA model on sequence A （1） Time series analysis was performed on the data. Calculate the difference order d, autoregressive parameter p, moving average parameter q, and seasonal difference order D for type I cases occurring in region 1 using (grid search, cross-validation, or) automated order determination algorithms. （s） Seasonal autoregressive parameter P （s） Seasonal moving average parameter Q （s） and seasonal cycles S （s） ; The autocorrelation coefficient of Class I cases is calculated based on the difference order d, the autoregressive parameter p, and the moving average parameter q. ~ and moving average coefficient θ （1） ~θ （q） ; Based on the seasonal difference order D （s） Seasonal autoregressive parameter P （s） Seasonal moving average parameter Q （s） Calculate the seasonal autocorrelation coefficient of Category I cases. ~ and seasonal moving average coefficient ~ ; Let sequence A （1） Chinese data is ~ Where ti represents sequence A （1） The number of time points corresponding to all data in the dataset; calculate ~ The average value aX （0） Calculate in sequence ~ Subtract aX （0） The difference is used as the white noise et in the first time series prediction. （0，1） ~et （0，ti） ; by ~ The original data is based on the difference order d and the seasonal difference order D. （s） Calculate the lag factor and seasonal lag factor of the original data, and then calculate the autocorrelation coefficient ψ. （1） ~ψ （p） and moving average coefficient θ （1） ~θ （q） And substitute in the seasonal autocorrelation coefficient ~ and seasonal moving average coefficient ~ Substituting into the SARIMA model, the calculation of the first region in one seasonal cycle S （s） The number of subsequent Category I cases : ; by ~ , Given the original data, calculate ~ , The average value aX （1） Calculate in sequence ~ , Subtract aX （1） The quotient, as the white noise et of the second time series prediction. （1，1） ~et （1，ti） et （1，（ti+1）） ; repeat The calculation process calculates the first region in two seasonal cycles S. （s） The number of subsequent Category I cases : Similarly, calculate the S of the first region in z seasonal cycles. （s） The number of subsequent Category I cases was ; (Retrieve the current month corresponding to the current time) Repeated prediction Using the same steps, the time is synchronized to the current month, and the number of Category I cases in the 2nd to reth regions after the corresponding seasonal cycle is predicted as the time series prediction quantity. For sequence B （1） Spatial dimension analysis is performed on the data in the dataset; With sequence B （1） Using the earliest occurrence of Category I cases as a benchmark, and with months as the time dimension, sequence B... （1） All data in the system is synchronized with the time. Construct a matrix sequentially in ascending order of time, showing the number of Category I cases occurring in January for each of the 1st to the *re*th regions; the matrix has 1 column and the number of rows depends on "Sequence B". （1） From the earliest occurrence of Category I cases in China to "Sequence B" （1） The number of months spanned within the period of "the latest occurrence of a Category I case"; Using the matrix of the number of Category I cases occurring in the first region in January as the baseline matrix, the correlation coefficient of the baseline matrix relative to other matrices is calculated; Using the correlation coefficients of the baseline matrix relative to other matrices as weights, a weight matrix for the first region is constructed; using maximum likelihood (MLE) or generalized moment estimation (GMM) combined with spatial lag model (SAR), the spatial autoregressive coefficients of the number of Category I cases in the first region in January and the number of Category I cases in other regions in January are calculated. According to sequence B （2） To sequence B （12） Calculate the spatial autoregression coefficients of the number of Category I cases occurring in the first region from February to December and the number of Category I cases occurring in the same months in other regions; Calculate the spatial autoregressive coefficients of the 2nd to the reth regions from January to December; Based on the current month, extract the spatial autoregressive coefficients of the first to the re regions in the corresponding month, and substitute the spatial autoregressive coefficients into the spatial lag model (SAR) to predict the number of Class I cases occurring in the first to the re regions in the current month, which is used as the spatial prediction quantity. Calculate the product of the number of temporal predictions and the number of spatial predictions for each region, and use it as the inspection coefficient for each region; Police resources are allocated to each area based on the patrol coefficient ratio of each area and the actual number of police officers in the police station.

[0023] Case Analysis Module: Used to extract and separate case description information and case judgment information corresponding to each case data; the description information and case judgment information are used as two information sets to train a text model; the text model is used to summarize and classify various cases within the police station's jurisdiction; Extract and separate the case description information and case judgment information corresponding to each case data; use the description information as information set A and the case judgment information as information set B, and create an index for the case judgment information corresponding to each case description information (so that each data in information set A corresponds to a certain data in information set B); Please see Figure 3 Based on the knowledge distillation mechanism, the text model is trained using information set A and information set B; The text information corresponding to each case in information set A is taken as a subset, denoted as the subset; the judgment information of each case corresponding to the subset (information set A) in information set B is taken as a subset, denoted as the master set; each subset corresponds to a master set, and each master set corresponds to one or more subsets; Using the BERT model, all text information in information set A and information set B is converted into text vectors; Using the text vectors corresponding to the subset as the question objects and the text vectors corresponding to the main set as the answer objects, the BERT model is trained to fill in the blanks and a teacher model is trained. The first subset in information set A is taken as set A. （te） (In set B) set A （te） The corresponding principal set is set B. （st） ; The BERT model undergoes fill-in-the-blank learning, and the steps are as follows: Step 1: Create fill-in-the-blank questions: Extract set A （te） For each phrase in the set A, calculate the word vector. （te） Each phrase in the set corresponds to a word vector and (set A) （te） The cosine similarity of the text vectors of the entire text; (In descending order) the phrases whose word vectors correspond to the top 30% of the cosine similarity are designated as key phrases, and the remaining 70% of the phrases whose word vectors correspond as non-key phrases. From set A （te） Among the key phrases and non-key phrases, those with a higher proportion of key phrases than non-key phrases are grouped into type A multiple-choice questions. Step 2: Do the fill-in-the-blank questions: Set B （st） The corresponding text vector serves as the standard answer; select from set B (excluding set B). （st） Other main sets of (of) are combined into options; Based on the weights and biases of the BERT model's classification layer, the softmax function is used to guess the correct probability of each option; the option with the highest correct probability is selected as the guessed answer. Step 3: Compare with the reference answer: Calculate the cosine similarity and cross-entropy loss of the text vectors of the corresponding phrases of the reference answer and the guessed answer; The gradient descent algorithm was used to optimize the weights and biases of the BERT model's classification layer until the cosine similarity of the text vectors of the corresponding phrases of the reference answer and the guessed answer was maximized and the cross-entropy loss was minimized. Step 4: From set A （te） Among the key phrases and non-key phrases, those with a lower proportion of key phrases than non-key phrases are grouped into type B multiple-choice questions. Based on the type b multiple choice question format, repeat the training process from the first to the third step, further optimize the weights and biases of the BERT model's classification layer, and complete the construction of the teacher model; A student model is constructed based on the RNN model; The student model is used for fill-in-the-blank learning. The steps for fill-in-the-blank learning are as follows: Data input and parallel processing: Set A （te） Set B and other main sets are simultaneously input into the teacher model and the student RNN model to be trained; In set A （te） In the middle, the teacher model generates multiple-choice questions and standard answers (c) based on the format of type a multiple-choice questions. （st） (c) （st） (The text vector representing the standard answer). The teacher model is selected from set B (not set B). （st） Other principal sets, with the standard answer c （st） Combine into multiple options; The student model then outputs its own prediction result, the corresponding text vector being v. （st） ; Backpropagation: Calculate the text vector c （st） With v （st） Cosine similarity and cross-entropy loss; Adam or SGD with momentum is selected as the optimizer. The model parameters in the teacher model are kept unchanged. Only the weights and biases of the embedded and hidden layers of the RNN model are optimized by gradient backpropagation until the cosine similarity of the text vectors of the corresponding phrases of the reference answer and the guessed answer is maximized and the cross-entropy loss is minimized or the preset maximum number of iterations is reached. The lightweight student model is then output. The trained student model was used to summarize and classify various cases within the police station's jurisdiction; The system monitors in real time whether any incidents occur within its jurisdiction; if no incidents occur, no action is taken; if incidents occur, the system selects a route for the police car based on its location and the location of the incident, combined with the speed of the police car. The police vehicle's route is selected based on its location and the location of the crime scene, combined with its speed. Let the area where the police car is located be region m, and the crime scene be region n; select a route for the police car based on the different routes from region m to region n; Scenario 1: There are no other regions among all the roads from region m to region n; Among all roads from region m to region n, select the route with less traffic and shorter distance for police car A as its target route; Scenario 2: There are other regions among all the roads from region m to region n, but the forward routes to these regions are unique; (For example: There are four areas 1, 2, 3 and 4 between area m and area n, but the road segments between these areas are: area m → area 1 → area 2 → area 3 → area 4 → area n; That is, from region m, you can only go to region 1; from region 1, you can only go to region 2; from region 2, you can only go to region 3, and so on; that is, the route from one region to the next must be adjacent to the previous region, and the route from one region to another is unique. Repeat the selection criteria from scenario one to choose a target route for police car A; Scenario 3: There are other regions among all the roads from region m to region n, and the routes to these regions are not unique; (For example: There are four areas 1, 2, 3 and 4 between area m and area n. The road segments in these areas are: area m → area 1 → area 2 → area 3 → area 4 → area n; m region → 1 region → 4 region → n region; m region → 3 region → n region; From region m, one can go to region 1 or region 2; From area 1, you can go to area 2 or area 4; From region 3, one can go to region 4, or region n, etc. That is, in scenario three, the route from one area to another is not unique. (Along the direction from region m to region n) Let the region directly adjacent to region m be a first-level region (similar to regions 1 and 3 that region m can go to in the example above, which are also first-level regions). Count the total number of paths from each primary region to region n (without duplicates) rn, and count the traffic flow tr on the 1st, 2nd, up to the rnth path. （1） tr （2） ~tr （rn） ; Get the current time and the movement speed v of police car A. （甲）Calculate the latest arrival time t of police car A by selecting the longest path from path 1 to path n. （lat） Calculate the earliest arrival time t of police car A by selecting the shortest path. （ear） ; Get the number of other police cars pw on paths 1 to rn. （1） ,pw （2） ~pw （rn） ; Obtain the position and speed of each police car on paths 1 to rn, calculate the arrival time of each police car on path 1 to rn to region n, and sort them according to the arrival time to region n to obtain tw. （1,1） tw （1,2） ~tw （1,pw（1）） ~tw （rn,pw（rn）） ; Among them, tw （1,1） This indicates the time when the first police car on the first path arrives at region n. tw （1,2） This indicates the time it takes for the second police car on the first path to arrive at region n. And so on, tw （1,pw（1）） This indicates the pw-th node on the first path. （1） The time it takes for the police cars to arrive at area n; tw （rn,pw） This represents the pw-th path on the n-th path. （rn） The time it takes for the police cars to arrive at area n; Get the maximum traffic flow of the first path (tr) （max） Calculate the time cost (Ttw) of police car A on the first path. （1） ; In TW （1,1） tw （1,2） ~tw （1,pw（1）） The statistics belong to the interval [t] （ear） , t （lat） The quantity lp （1） ; Get the total number of vehicles on the current first path (pp) （now） , judge (pp （now） +tr （1） +pp （b） ) <tr （max） Is it valid? If true, then in tw （1,1） tw （1,2） ~tw （1,pw（1）） Extract the (lp) （1） +1) Arrival time of the police car (ty) （ot） , Ttw （1）The value is: Ttw （1） =[(ty) （ot） -t （ear） ) + (ty （ot） -t （lat） )] / 2; If ty （ot） If it does not exist, then Ttw （1） The value is: Ttw （1） =[(tw （1,pw（1）） -t （ear） ) + (tw （1,pw（1）） -t （lat） )] / 2; If not true, then judge (pp) （now） +tr （1） +pp （b） =tr （max） Is it valid? If true, then in tw （1,1） tw （1,2） ~tw （1,pw（1）） Extract the (tp) （a） +1) Arrival time of the police car (ty) （ota） ; tp （a） tp represents the allowed traffic flow for the first route. （a） The formula for calculating tp is: （a） =tr （max） - (pp （now） +tr （1） +pp （b） )-1; Ttw （1） The value is: Ttw （1） =[(ty) （ota） -t （ear） ) + (ty （ota） -t （lat） )] / 2; If ty （ota） If it does not exist, then Ttw （1） The value is: Ttw （1） =[(tw （1,pw（1）） -t （ear） ) + (tw （1,pw（1）） -t （lat） )] / 2; If (pp) （now） +tr （1） +pp （b） =tr （max） Not valid; Then obtain the total number of traffic lights rg on the first path, and the red light duration rt of the first, second, up to the rg-th traffic lights. （1） ,rt（2） ~rt （rg） ; Get the distance X from police car A (on the first path) to the first traffic light. （o-1） The distance X from the first traffic light to the second traffic light （1-2） The distance X from the second traffic light to the third traffic light （2-3） And so on, the distance X from the (rg-1)th traffic light to the rgth traffic light （（rg-1）-rg） ; Calculate the distance of police car A through X （1-2） Minimum speed of the corresponding path vs （1） : vs （1） =X （1-2） / rt （1） ; Calculate the latest time (tx) for police car A to arrive at the first traffic light. （甲） : tx （甲） = (2×X) （o-1） ) / (v （甲） +vs （1） ); Determine tx （甲） With rt （1） ,rt （2） ~rt （rg） Is the sum less than t? （lat） ; If it is less than X, then calculate the distance police car A passed through. （2-3） To X （（rg-1）-rg） Minimum speed of the corresponding path vs （2） to vs （rg） ; When Police Car A was driving in X （1-2） To X （（rg-1）-rg） When following the corresponding route, adjust the speed of police car A to vs. （2） to vs （rg） ; Ttw （1） The value is tx （甲） With rt （1） ,rt （2） ~rt （rg） The sum; If not less than, then Ttw （1） The value is t （lat） ; Repeat Ttw （1） The calculation process calculates the time cost Ttw for paths 2 to rn. （2） ~Ttw （rn） ; in Ttw （1） ~Ttw （rn）In the process, the path with the least time cost is selected as the target path for police car A, and the primary area on the target path is designated as the target area for police car A. Repeat the steps of selecting the target path and target area for police car A, and continue to select the path and area for police car A until police car A reaches area n.

[0024] Example 2 Please see Figure 2 Multi-agency collaborative combat methods based on workflow engines include: Step S1: Aggregate the business data of each police department within each police station (such as criminal investigation case data, cyber security network monitoring data, and public security patrol data), and build a decentralized blockchain network (to share the corresponding business data of each police station). Step S2: Separate the case data from the business data, and predict the number of cases in each area based on the actual number and type of cases within the jurisdiction of each police station, using time and location as dimensions, and allocate police resources accordingly; Step S3: Extract and separate the case description information and case judgment information corresponding to each case data; use the description information and case judgment information as two information sets to train the text model; use the text model to summarize and classify the various cases within the police station's jurisdiction; Extract and separate the case description information and case judgment information corresponding to each case data; treat the description information as information set A and the case judgment information as information set B, and create an index for the case judgment information corresponding to each case description information (so that each data in information set A corresponds to a certain data in information set B); monitor in real time whether any cases occur in the jurisdiction; if no cases occur, no action is taken; if cases occur, select a route for the police car based on the location of the police car and the area where the case occurred, combined with the speed of the police car.

[0025] The above formulas are all dimensionless calculations. The formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation. For example, there are weighting coefficients and proportional coefficients. The values ​​set are to quantify each parameter to obtain a specific value, which is convenient for subsequent comparison. The values ​​of the weighting coefficients and proportional coefficients are only required to not affect the proportional relationship between the parameters and the quantified values.

[0026] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-agency collaborative combat system based on a workflow engine, characterized in that: The system includes: Data fusion module: used to aggregate business data and build a decentralized blockchain network; Police force allocation module: Used to separate case data from business data, and predict the number of cases in each area based on the actual number and type of cases, using time and location as dimensions, and allocate police force accordingly; Case Analysis Module: Used to extract and separate case description information and case judgment information corresponding to each case data; the description information and case judgment information are used as two information sets to train a text model; the text model is used to summarize and classify various cases within the police station's jurisdiction; The system monitors in real time whether any incidents occur within its jurisdiction; if no incidents occur, no action is taken; if incidents occur, the system selects a route for the police car based on its location, the location of the incident, and its speed.

2. The multi-agency collaborative combat system based on a workflow engine according to claim 1, characterized in that, The workflow of the data fusion module is as follows: Using a consortium blockchain as a framework, at least one verification node and one observation node are deployed for each police station; sorting nodes are deployed for relevant departments. The initial block is generated, and network rules are defined; each police station authenticates its identity through a CA authority, obtains a digital certificate for the blockchain, and transmits its business data to the blockchain. By deploying a privacy protection layer through federated learning collaboration and encrypted queries, cross-agency data collaboration and querying capabilities, as well as data protection functions, are achieved. The initial information query model architecture was released in a unified manner, clarifying the query objects, query scope, input characteristics, output styles and hyperparameter ranges of business data, and determining the training objectives of each police station; Based on the training objective, each business data is used to train the information query model, which is then trained locally to generate gradient parameters. After completing local training, aggregate the gradient parameters of the final training and update the parameters in the information query model; Encrypted query: Securely retrieve encrypted results from encrypted computation; Data preparation: The data owner encrypts the queryable business data; Request encryption: The querying party uses the public key of the data owner to encrypt the query conditions and the querying party's private key; Local matching: The data owner decrypts the query conditions of the queryer, and encrypts the content to be queried by the queryer using the queryer's private key according to the query conditions to obtain the query ciphertext; the access permission of the query ciphertext is set to be valid only for the queryer, and then fed back to the queryer. Decryption of results: The querying party uses its private key to decrypt the ciphertext and obtain the content to be queried.

3. The multi-agency collaborative combat system based on a workflow engine according to claim 1, characterized in that, The workflow of the police force allocation module is as follows: Summarize the business data corresponding to each region and extract case data from the business data; the case data includes the location of the incident, the time of the incident, and the case type. This serves as the area code for all districts within the police station's jurisdiction, and as the code for each case type in the operational data; By summarizing the number of Category I cases occurring in different regions at different times, we obtain Sequence A. （1） ~A （re） ; By summarizing the number of Category I cases occurring in the same month but in different years and regions, we obtain Sequence B. （1） ~B （12） ; Using the ARIMA model on sequence A （1） Time series analysis was performed on the data. Calculate the autocorrelation coefficient and moving average coefficient, as well as the seasonal autocorrelation coefficient and seasonal moving average coefficient, for Category I cases occurring in Region 1. Let sequence A （1） Chinese data is ~ Where ti represents sequence A （1） The number of time points corresponding to all data in the dataset; calculate ~ The average value aX （0） Calculate in sequence ~ Subtract aX （0） The difference is used as white noise for the first time series prediction; by ~ Using the original data, calculate the lag factor and seasonal lag factor of the original data. Substitute the autocorrelation coefficient, moving average coefficient, seasonal autocorrelation coefficient, and seasonal moving average coefficient into the SARIMA model to calculate the lag factor for the first region in one seasonal cycle S. （s） The number of subsequent Category I cases ; Similarly, repeated prediction Using the same steps, the time is synchronized to the current month, and the number of Category I cases occurring in the 2nd to reth regions after the corresponding seasonal cycle is predicted as the time-series prediction quantity.

4. The multi-agency collaborative combat system based on a workflow engine according to claim 3, characterized in that, The subsequent process of the police force allocation module is as follows: For sequence B （1） Spatial dimension analysis is performed on the data in the dataset; With sequence B （1） Using the earliest occurrence of Category I cases as a benchmark, and with months as the time dimension, sequence B... （1） All data in the system is synchronized with the time. Construct a matrix of the number of Category I cases in ascending order of time; use the matrix of the number of Category I cases in the first region in January as the baseline matrix, and calculate the correlation coefficient between the baseline matrix and the other matrices. The weight matrix for the first region is constructed using the correlation coefficients of the baseline matrix relative to other matrices as weights. Using the maximum likelihood combined with the spatial lag model, the spatial autoregressive coefficients of the number of Category I cases in the first region in January and the number of Category I cases in other regions in January were calculated. According to sequence B （1） To sequence B （12） Calculate the spatial autoregressive coefficients of the first to the re regions from January to December; Extract the spatial autoregressive coefficients of the first to the re regions in the corresponding month, and substitute the spatial autoregressive coefficients into the spatial lag model to predict the number of Class I cases in the first to the re regions in the current month, which will be used as the spatial prediction quantity. Calculate the product of the number of temporal predictions and the number of spatial predictions for each region, and use it as the inspection coefficient for each region; Police resources are allocated to each area based on the patrol coefficient ratio and the actual number of police officers.

5. The multi-agency collaborative combat system based on a workflow engine according to claim 1, characterized in that, The workflow of the case analysis module is as follows: Extract and separate the case description information and case judgment information corresponding to each case data; use the description information as information set A and the case judgment information as information set B, and create an index for the case judgment information corresponding to each case description information; Based on the knowledge distillation mechanism, the BERT model is used as the teacher model and the RNN model is used as the student model. The text model is trained using information set A and information set B. The system monitors in real time whether any incidents occur within its jurisdiction; if no incidents occur, no action is taken; if incidents occur, the system selects a route for the police car based on its location, the location of the incident, and its speed.

6. The multi-agency collaborative combat system based on a workflow engine according to claim 5, characterized in that, The specific process for training the text model is as follows: The text information corresponding to each case in information set A is used as a subset; the judgment information of the cases corresponding to each subset in information set B is used as the main set. Using the BERT model, all text information in information set A and information set B is converted into text vectors; The first subset in information set A is taken as set A. （te） Set A （te） The corresponding principal set is set B. （st） ; Using the text vectors corresponding to the subset as the question objects and the text vectors corresponding to the main set as the answer objects, the BERT model is trained to fill in the blanks and a teacher model is trained. The RNN model is used as the student model, and the student model is used for fill-in-the-blank learning.

7. The multi-agency collaborative combat system based on a workflow engine according to claim 6, characterized in that, The route selection process for the police car is as follows: Let the area where the police car is located be region m, and the crime scene be region n; select a route for the police car based on the different routes from region m to region n; Scenario 1: There are no other regions among all the roads from region m to region n; Among all roads from region m to region n, select the route with less traffic and shorter distance for police car A as its target route; Scenario 2: There are other regions among all the roads from region m to region n, but the forward routes to these regions are unique; Repeat the selection criteria from scenario one to choose a target route for police car A; Scenario 3: There are other regions among all the roads from region m to region n, and the routes to these regions are not unique; Count the total number of paths rn from each first-level region to region n, and count the traffic flow tr on the first to rn paths. （1） ~tr （rn） ; Get the police car's moving speed v （甲） Calculate the latest arrival time t of the police car by selecting the longest path from the 1st to the nth path. （lat） Calculate the earliest arrival time t of the police car by selecting the shortest path. （ear） ; Get the number of other police cars pw on paths 1 to rn. （1） ,pw （2） ~pw （rn） ; Obtain the position and speed of each police car on paths 1 to rn, calculate the arrival time of each police car on path 1 to rn to region n, and sort them according to the arrival time to region n to obtain tw. （1,1） ~tw （rn,pw（rn）） ; Get the maximum traffic flow of the first path (tr) （max） Calculate the time cost (Ttw) of the police car on the first path. （1） .

8. The multi-agency collaborative combat system based on a workflow engine according to claim 7, characterized in that, The route selection process for the police car also includes: In TW （1,1） ~tw （1,pw（1）） The statistics belong to the interval [t] （ear） , t （lat） The quantity lp （1） ; Get the total number of vehicles on the current first path (pp) （now） , judge (pp （now） +tr （1） +pp （b） ) <tr （max） Is it valid? If true, then in tw （1,1） tw （1,2） ~tw （1,pw（1）） Extract the (lp) （1） +1) Arrival time of the police car (ty) （ot） , Ttw （1） The value is: Ttw （1） =[(ty) （ot） -t （ear） ) + (ty （ot） -t （lat） )] / 2; If ty （ot） If it does not exist, then Ttw （1） The value is: Ttw （1） =[(tw （1,pw（1）） -t （ear） ) + (tw （1,pw（1）） -t （lat） )] / 2; If not true, then judge (pp) （now） +tr （1） +pp （b） =tr （max） Is it valid? If true, then in tw （1,1） tw （1,2） ~tw （1,pw（1）） Extract the (tp) （a） +1) Arrival time of the police car (ty) （ota） ; tp （a） tp represents the allowed traffic flow for the first route. （a） =tr （max） - (pp （now） +tr （1） +pp （b） )-1; Ttw （1） The value is: Ttw （1） =[(ty) （ota） -t （ear） ) + (ty （ota） -t （lat） )] / 2; If ty （ota） If it does not exist, then Ttw （1） =[(tw （1,pw（1）） -t （ear） ) + (tw （1,pw（1）） -t （lat） )] / 2; Calculate the time cost Ttw for paths 2 to rn. （2） ~Ttw （rn） ; in Ttw （1） ~Ttw （rn） In this process, the path with the least time cost is selected as the target path for the police car, and the primary area on the target path is designated as the target area for the police car. Repeat the steps of selecting a target route and target area for the police car, and continue to select routes and areas for the police car until the police car reaches area n.

9. A multi-agency collaborative combat method based on a workflow engine, applicable to the multi-agency collaborative combat system based on a workflow engine as described in any one of claims 1-8, characterized in that, The method includes: Aggregate business data and build a decentralized blockchain network; Separate case data from business data, and predict the number of cases in each area based on the actual number and type of cases, using time and location as dimensions, and allocate police resources accordingly; Extract and separate the case description information and case judgment information corresponding to each case data; use the description information and case judgment information as two separate information sets to train a text model; use the text model to summarize and classify various cases within the police station's jurisdiction; monitor in real time whether any cases occur within the jurisdiction; if no cases occur, no action is taken; if cases occur, select a route for the police car based on the location of the police car and the area where the incident occurred, combined with the speed of the police car.

Images