Text data distributed processing method based on federal learning

Abstract

The invention discloses a federal learning-based text data distributed processing method, which comprises the following steps of: acquiring information and activity data of specific personnel through a data acquisition and storage module, establishing a unified data standard method, and storing labeled data in a database; unrelated data clearing, missing value and abnormal value processing, data feature statistics and training data grouping are completed; constructing a federated learning training module by using an LSTM neural network model, and training the federated learning training module to obtain an available early warning model; the early warning model is deployed on a local server of a department, track information of a specific person is predicted in real time, whether a dangerous behavior exists or not is judged, and if the possibility of the dangerous behavior exists, measures are taken in time; the problems that in the prior art, data circulation is difficult, calculation cost is low, and global optimization is difficult are solved, meanwhile, the action track of a specific person is predicted, and therefore public safety is guaranteed.

Description

Technical Field

[0001] The present invention relates to fields such as machine learning technology, and specifically, to a text data distributed processing method based on federated learning. Background Art

[0002] In terms of early warning of specific personnel behavior, the existing early warning methods mostly rely on experience for manual warning or use multiple sensors for data detection. These two early warning methods are inefficient, and the warning is not fast enough, the analysis is not comprehensive enough, and it is easy to cause hidden dangers. With the rapid development of information technology and the popularization of the Internet, the application of big data analysis and artificial intelligence technology is becoming more and more extensive. Big data analysis technology can help us process and analyze massive amounts of data, extract useful information from it, and use it for problem diagnosis and prediction. Using machine learning technology can identify patterns and make decisions by learning and training related data.

[0003] In traditional machine learning models, data needs to be centralized on a central server for processing and analysis, which increases the risk of data leakage. In many cases, data from different departments are difficult to communicate with each other, and it is impossible to train deep learning models that require large-scale data with good performance. In this case, federated learning came into being, which allows multiple participants to collaborate to train shared models while keeping their respective data local. In this setting, the original data does not need to leave its original location, and only the parameters or gradient updates of the model are exchanged between participants. In this way, even if the data cannot flow freely, all parties can benefit from collective wisdom and improve the performance of the model while complying with legal requirements for data privacy and security. At the same time, for users of text data, the gating mechanism included in LSTM enables it to avoid the problem of gradient vanishing or gradient exploding when processing long sequences, but this also makes LSTM computationally complex. Especially when processing large-scale data or long sequences, the time cost of training and inference may be very high. In addition, due to the complexity and diversity of data sources and the inconsistent data distribution of different institutions, global optimization may become difficult when processing data with large distribution differences. This is because global optimization usually assumes that all data come from the same distribution, but in this case, this assumption may not hold, which will greatly reduce the convergence efficiency of the global model.

[0004] However, the existing LSTM-based intelligent warning methods have the following disadvantages:

[0005] Difficulty in data circulation: LSTM usually requires a large amount of training data to achieve good performance. For tasks with small data volumes, LSTM may have overfitting problems. As the importance of data security increases, data from different departments is increasingly difficult to circulate, which is not conducive to training intelligent early warning models with good performance.

[0006] High computational cost: The gating mechanism included in LSTM allows it to avoid the problem of gradient vanishing or gradient exploding when processing long sequences, but this also makes LSTM computationally complex. Especially when processing large-scale data or long sequences, the time cost of training and inference may be very high.

[0007] Difficulty in global optimization: Due to the complexity and diversity of data sources and the inconsistency of data distributions among different institutions, global optimization may become difficult when dealing with data with large distribution differences. This is because global optimization usually assumes that all data comes from the same distribution, but in this case, this assumption may not hold, which will greatly reduce the convergence efficiency of the global model.

[0008] The patent published by Shen Zhiheng et al., "A method for medium- and long-term prediction of regional electricity consumption based on federated learning", simultaneously conducts federated training on electricity consumption prediction models for multiple regions. While protecting data privacy, it not only enriches the training samples, but also effectively improves the generalization ability of the model. The local basic model of federated learning adopts an LSTM network. In addition to traditional influencing factors of electricity consumption such as economy and society, the input features of the model also add carbon emissions, and the electricity consumption prediction results are more in line with the future low-carbon development path.

[0009] Authorized by Zhai Yunkai and others, "A New Emerging Major Infectious Disease Early Warning Method Based on LSTM Algorithm", this method establishes a new emerging major infectious disease early warning platform based on LSTM algorithm, including data collection and processing system, feature analysis and early warning indicator screening system and new emerging major infectious disease early warning system, which solves the technical problems of the limited number of cases in the early warning mode of a single hospital and the inability to conduct research and judgment on local areas. It proposes a long-short-term memory artificial neural network early warning model based on the attention mechanism of emerging major infectious diseases. According to the patient's clinical case data, associated imaging data, medical examination data and remote consultation data, etc., multi-time scale pyramid structure time series data are constructed, and long-short-term memory artificial neural network early warning models based on the attention mechanism of emerging major infectious diseases are constructed for time series data of different scales, respectively, to meet the early warning needs of different time sensitivity levels and balance the contradiction between accuracy and response time.

[0010] Wang Yang and others published a method for identifying and warning the behavior of specific people, which includes: collecting basic information of specific people; obtaining real-time trajectory information of specific people based on the basic information; analyzing the real-time behavior of specific people through the real-time trajectory information of specific people, and judging whether the current real-time behavior of specific people is normal travel behavior or abnormal travel behavior; obtaining the historical trajectory of specific people, and further analyzing the historical trajectory and current normal travel behavior or abnormal travel behavior; based on the analysis results, and conducting peer analysis and specific personnel gathering analysis, to warn the behavior of specific people. It binds the mobile phone positioning information to the travel trajectory of specific people to grasp the location information of specific people in real time. Through real-time monitoring, it determines whether there is any abnormality in the behavior of specific people every time they go out, and combines historical trajectory analysis and historical behavior analysis to predict whether illegal behavior will occur, and then give precise strikes.

[0011] Wang Yang et al. published a method for identifying dangerous actions based on deep learning, which includes: Step 1, using a target detection algorithm to detect and identify people and cigarettes respectively; after identifying people, enter Step 2 and Step 4, and after identifying people and cigarettes at the same time, enter Step 5; Step 2, using the UNet image segmentation algorithm to detect whether the dangerous area is included; Step 3, according to the detection results, determine whether the bounding box coordinates corresponding to the person label are included in the dangerous area; Step 4, using the human skeleton point detection algorithm to detect and determine whether the person has fallen; Step 5, further detect the human body parts to identify the face and hands of the person; Step 6, determine whether the bounding box position corresponding to the cigarette is included in the bounding box corresponding to the face or hand. It uses different models for detection, which greatly improves the reliability and accuracy of the corresponding label detection, and at the same time identifies and warns multiple dangers.

[0012] Zhang Long et al. published the "Abnormal Computing Power Federal Detection Method, System and Medium in Edge Computing Network", which includes edge computing network to ensure user service experience through on-demand, flexible and efficient joint scheduling of network resources and computing power resources, monitoring computing power anomalies in edge resource pools, and establishing a computing power anomaly detection operation mechanism; proposing a federated anomaly detection framework for communication between aggregated edge service nodes and edge nodes; facing the problem of abnormal computing power of computing power nodes in edge computing network, proposing an anomaly detection model based on CNN-BILSTM for identifying abnormal computing power; preparing HDFS data sets and PageRank data sets, preparing experimental environments for local training, and using precision, recall and F1 value evaluation indicators to obtain experimental results and analysis. Using this technology, data and privacy security are guaranteed on edge computing nodes, and real-time anomaly detection is performed, effectively improving the performance and reliability of edge computing systems.

[0013] Li Jianbin et al. published "A Federated Learning Security Aggregation Method and System for Distributed Load Forecasting", which uses the LSTM algorithm model as the basic training model. Participants receive initialized model parameters from the central server and use local data to train local models; participants upload the updated local model parameters to the central server; the central server determines the approximate global model parameters based on model parameter similarity, and aggregates the global model parameters using a distance-based weighted method, and then distributes the global model parameters to the participants' local models for the next round of training. The proposed method and system ensure that each load participant does not have to share its local data, protecting the privacy of sensitive information; by eliminating the interference of malicious local models on the global model and expanding the sample data space of each load data participant, the accuracy and effectiveness of the central model are guaranteed, thereby improving the model prediction accuracy. Summary of the invention

[0014] The purpose of the present invention is to provide a distributed processing method for text data based on federated learning. Aiming at the shortcomings of the existing LSTM-based intelligent early warning method, the LSTM model in text prediction is improved by introducing federated learning technology. While solving the problems of data circulation difficulties, computational costs and difficulty in global optimization in the prior art, the action trajectory of specific personnel is predicted, thereby ensuring public safety.

[0015] The present invention is implemented by the following technical solution: a text data distributed processing method based on federated learning, comprising the following steps:

[0016] 1) Complete the collection of information and activity data of specific personnel through the data collection and storage module, including user identity information, social relationships, trajectory information, etc., and establish a unified data standard method to store the labeled data in the database to provide a data basis for subsequent steps;

[0017] 2) Use the data preprocessing module to complete the removal of irrelevant data, the processing of missing values ​​and outliers, the statistics of data features, and the grouping of training data; in the process of using LSTM for data training, each department must first preprocess the data, the main purpose of which is to delete irrelevant data in the original input data, avoid inputting data that does not meet the requirements, including some duplicate data and smooth noise data, and filter out invalid data that is irrelevant to the later data modeling, as well as processing missing values ​​and outliers in the original input data. After the data processing of each department is completed, the distribution information of the training data is counted, including the mean, variance, and covariance matrix of the training data, so as to facilitate the subsequent matching of institutional data with similar distributions for training grouping for federated learning.

[0018] 3) Use the LSTM neural network model to build a federated learning training module, and train it to obtain a usable early warning model; after each department has completed its own data collection, the central server initiates a training request for federated learning, and the department can participate in the federated training at this time. First, submit the data distribution information, and the central server divides different departments into several groups according to the data distribution information, so that departments with similar data distribution are trained together. The neural networks used in the present invention are LSTM trajectory prediction network and LSTM behavior prediction network, which are used for trajectory prediction and behavior prediction. The input of the LSTM trajectory prediction network is trajectory information, combined with the dilated convolution and attention mechanism, it can efficiently perform feature extraction and time series prediction. The trajectory prediction module is based on the historical trajectory information of the specific person whose trajectory needs to be predicted, as the input of LSTM, and the predicted trajectory of the future time domain is predicted by the network. The behavior prediction module is based on the output of the LSTM classification network for secondary inference, and the future trajectory is input according to the user's location and place, and the behavior to be performed by the specific person is classified, and the abnormal behavior of the specific person is discovered in time. After each department completes its own network training, it uploads the model to the central server. After receiving the department model, the central server aggregates the model and then sends the aggregated model to each department for further training until the model accuracy meets the requirements.

[0019] 4) Early warning model deployment: The early warning model is deployed on the department’s local server to make real-time predictions on the trajectory information of specific personnel, and then determine whether there is any dangerous behavior. If there is a possibility of dangerous behavior, timely measures are taken.

[0020] The early warning model is deployed on the department's local server to achieve real-time prediction of key personnel trajectory information. The main task of the early warning model is to monitor and analyze the actions of these key personnel in order to quickly identify potential dangerous behaviors. Once a situation with possible dangerous behavior is predicted, appropriate measures will be taken immediately to ensure timely and effective response. This real-time early warning system helps improve the department's safety and emergency response capabilities.

[0021] At the same time, the early warning model also uses the method of federated learning to train the latest model by aggregating data from different departments. Through federated learning, departments can share experience and information to continuously improve the accuracy and adaptability of the early warning model. This update mechanism ensures that the system can adapt to the changing environment and threats in a timely manner and maintain effective early warning capabilities for new dangerous behaviors. Therefore, the early warning system is not only real-time and accurate, but also can continuously improve overall safety performance as the situation evolves.

[0022] In order to better implement the text data distributed processing method based on federated learning described in the present invention, the following setting is particularly adopted: the step 1) includes the following specific steps:

[0023] 1.1) Collect personal identity information of specific persons (such as ID card, license plate number, occupation, etc.) and collect social relationships between specific persons and other individuals (including relationships with friends, family members, colleagues, etc.);

[0024] 1.2) The trajectory information of a specific person is recorded by a sensing device to obtain historical trajectory data, and the historical trajectory data is divided into a text data sequence of length T; the feature data input by the LSTM network is the motion trajectory of each hour with a fixed sliding window length of T, and the motion trajectory at the next moment, i.e., moment T+1, is used as the target output, so the trajectory data is divided into a text data sequence of length T (in the present invention, T is usually set to 120, i.e., the position of a specific person is collected once every 30 seconds).

[0025] 1.3) Each department establishes a unified data standard method to form a unified coding format and display format for the personal information and activity data of specific personnel, which can achieve better training results. The labeled data is stored in the database to provide a data basis for subsequent steps.

[0026] In order to better implement the text data distributed processing method based on federated learning described in the present invention, the following setting is particularly adopted: the step 2) includes the following steps:

[0027] 2.1) Clear irrelevant data:

[0028] 2.1.1) Deduplication: Delete duplicate data samples to avoid duplication of model training;

[0029] 2.1.2) Smoothing noisy data: Processing the noise in the data through filtering or smoothing techniques ensures that the training data is more stable and reliable;

[0030] 2.1.3) Eliminate invalid data: Filter out data that is irrelevant to model building to ensure that the input features are relevant and meaningful;

[0031] 2.2) Handling missing values ​​and outliers:

[0032] 2.2.1) Missing value processing: Missing values ​​are likely to occur during data collection. Interpolation, filling or deletion methods are used to process missing values ​​in the data to maintain data integrity;

[0033] 2.2.2) Outlier processing: Use appropriate methods such as truncation, replacement or deletion to identify and process outliers to ensure data stability and accuracy;

[0034] 2.3) Statistical data characteristics:

[0035] After preprocessing, each department needs to collect statistics on the distribution information of the training data, which provides a key foundation for subsequent federated learning. The statistical data features include:

[0036] 2.3.1) Mean and variance: Calculate the mean and variance of data features to understand the central tendency and dispersion of the data set;

[0037] 2.3.2) Covariance matrix: Calculate the covariance between different features and explore the correlation between features;

[0038] 2.4) Group training data:

[0039] Based on statistical information, data holding institutions (i.e., federated learning participants) are divided into groups with similar distributions so that each institution can match data with similar distributions for training in federated learning.

[0040] In order to better implement the text data distributed processing method based on federated learning described in the present invention, the following setting is particularly adopted: the federated learning training module constructed by using the LSTM neural network model is specifically:

[0041] An RNN model is selected and improved using a dilated convolution and an attention mechanism; wherein LSTM (Long Short-Term Memory Network) is a special type of RNN (Recurrent Neural Network) that effectively handles the problem of gradient vanishing and gradient exploding through a gating mechanism. However, LSTM has high computational complexity when processing long sequence data, especially when the sequence length increases, which results in a large consumption of computing resources. In order to solve this problem, the present invention uses a dilated convolution and an attention mechanism to improve the model.

[0042] Atrous convolution is a convolution operation that can increase the receptive field without increasing the number of parameters. In 1D sequence processing, atrous convolution can be used to capture long-distance dependencies. Therefore, the present invention uses atrous convolution to solve the problem of complex text data calculation without increasing the amount of additional calculation.

[0043] The dilated convolution introduces a dilation rate d, which allows the elements in the convolution kernel to be spaced d-1 units apart. The expression of the dilated convolution operation is: d is the hole rate, x is the input sequence, w is the convolution kernel, y is the output sequence, and K is the size of the convolution kernel; the attention mechanism allows the model to dynamically focus on different parts of the sequence when processing the sequence, which can more effectively capture long-distance dependencies. In the attention mechanism, the attention score of each sequence element is calculated, and then these scores are used to weight the input sequence to obtain a weighted representation. The attention mechanism adopted by the present invention is: Q(query), K(key), and V(value) are the representations learned by the model: Q is the dimension of the key, which is used to scale the dot product to prevent excessive dot product values ​​from causing the softmax function to enter an area with a small gradient.

[0044] Combining dilated convolution and attention mechanism, a model that can efficiently process long sequence data can be constructed. For example, dilated convolution can be used to increase the receptive field, while the attention mechanism can be used to dynamically select important information in the sequence. Such a model structure can more effectively capture long-distance dependencies while keeping the number of parameters unchanged, and reduce the computational cost caused by the increase in sequence length.

[0045] After using the LSTM neural network model to predict the trajectory of a specific person, the historical trajectory information and other information of the specific person (such as work and social information) are combined and cascaded using the LSTM neural network to obtain the behavior classification results.

[0046] In order to better implement the text data distributed processing method based on federated learning described in the present invention, the following setting is particularly adopted: the training of the federated learning training module includes the following steps:

[0047] 3.1) The central server determines the training task and training model, broadcasts the agreed number of global training rounds E, and randomly initializes and broadcasts the global aggregate model parameters W g ;

[0048] 3.2) Departments choose whether to join the training. The central server determines the subset S that participates in the training according to the agreed ratio of the number of departments participating in each round of local training, and groups and aggregates different departments according to data distribution information (the number of groups is k);

[0049] 3.3) Department s (s∈S) will global aggregation model W g The initial local model of this round of training is trained for K rounds; the formula for local update is as follows:

[0050] 3.4) Department s uploads the gradient of each round The central server then implements global model aggregation based on the average weighted method:

[0051] 3.5) The central server sends the aggregated model parameters W g,k Each department in group k is given a model parameter, and each department in group k continues to train using the model parameters;

[0052] 3.6) Repeat steps 3.3) to 3.5) for multiple rounds until the model achieves the desired effect. Department members of the group that meet the requirements can exit training early.

[0053] In order to better implement the distributed text data processing method based on federated learning described in the present invention, the following setting is particularly adopted: in the step 3.6), during multiple rounds of repetitions of steps 3.3) to 3.5), department members of the group that meet the requirements directly exit training early.

[0054] In order to better realize the distributed text data processing method based on federated learning described in the present invention, after completing the trajectory prediction training, the following setting is particularly adopted: after the trajectory prediction of a specific person is completed using the LSTM neural network model, the historical trajectory information and other information of the specific person (the occupation, social relationship, etc. of the specific person collected in the step 1) are combined to predict dangerous behaviors. These features will be used as inputs to the LSTM classification model, and the activation function usually selects softmax to convert the output into a probability distribution, the sum of the probabilities of each category is 1, and the output layer will give the probability of each behavior category to analyze the behavior of the specific person.

[0055] In order to better realize the distributed text data processing method based on federated learning described in the present invention, the following setting method is particularly adopted: the early warning model also uses the method of federated learning to continuously improve the accuracy and adaptability of the early warning model by aggregating data from different departments.

[0056] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0057] (1) The present invention breaks data silos: Through federated learning, model training can be performed without directly sharing data. Each participant (such as different departments or institutions) can train the LSTM model on local data and only share model parameters or updates instead of raw data. This not only protects data privacy, but also allows more data to be used to train the model, which may reduce the risk of overfitting. This will greatly improve the generalization of the early warning model and help us understand and analyze the on-site situation more accurately.

[0058] (2) The present invention reduces computational costs: In order to reduce the computational cost of LSTM, a model compression technology is proposed to simplify the structure of the LSTM model. This can reduce the number of model parameters and reduce computational complexity, while maintaining model performance as much as possible and reducing communication costs in the federated learning process.

[0059] (3) The present invention improves the model personalization capability: Under the federated learning framework, personalized model training strategies are customized for different participants. The model structure or training process can be adjusted according to the data distribution and characteristics of each participant to improve the performance of the model on specific data.

[0060] (4) Compared with the patent "A method for medium- and long-term prediction of regional electricity consumption based on federated learning" disclosed by Shen Zhiheng et al., the present invention applies dilated convolution and attention mechanism, which can more effectively capture long-distance dependencies while keeping the number of parameters unchanged, and reduce the computational cost caused by the increase in sequence length.

[0061] (5) Compared with the "An Early Warning Method for Emerging Major Infectious Diseases Based on LSTM Algorithm" authorized by Zhai Yunkai et al., this paper invented federated learning and carried out group learning, which effectively ensured data security and accelerated the data flow between different departments.

[0062] (6) Compared with the "Method for Identifying and Predicting the Behavior of Specific Personnel" published by Wang Yang et al., the present invention can grasp the location information of specific persons in real time and conduct retraining. In addition to combining historical trajectory analysis, other information of specific persons is also added to predict the probability of illegal behavior and then stop it. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is the overall flow chart of the present invention.

[0064] Figure 2 It is a data preprocessing module.

[0065] Figure 3 Schematic diagram of federated learning training.

[0066] Figure 4 This is a schematic diagram of the early warning model working. DETAILED DESCRIPTION

[0067] The present invention is further described in detail below in conjunction with examples, but the embodiments of the present invention are not limited thereto.

[0068] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention claimed for protection, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0069] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present invention.

[0070] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0071] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0072] In the present invention, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may include that the first and second features are in direct contact, or may include that the first and second features are not in direct contact but are in contact through another feature between them. Moreover, a first feature being "above", "above" and "above" a second feature includes that the first feature is directly above and obliquely above the second feature, or simply indicates that the first feature is higher in level than the second feature. A first feature being "below", "below" and "below" a second feature includes that the first feature is directly below and obliquely below the second feature, or simply indicates that the first feature is lower in level than the second feature.

[0073] Disclaimer: The collection of information and activity data of specific persons by this invention is permitted by the relevant specific persons, and the laws and regulations of relevant countries and regions are followed during the collection. There is no infringement of personal privacy, so as to ensure the legitimacy of the source of information.

[0074] Embodiment 1:

[0075] The present invention designs a distributed processing method for text data based on federated learning. Aiming at the shortcomings of the existing LSTM-based intelligent warning method, the present invention introduces federated learning technology to improve the LSTM model in text prediction. While solving the problems of data circulation difficulties, computational costs and difficulty in global optimization in the prior art, the action trajectory of specific personnel is predicted, thereby ensuring public safety. Figure 1 to Figure 4 As shown, the following steps are included:

[0076] 1) Complete the collection of information and activity data of specific personnel through the data collection and storage module, including user identity information, social relationships, trajectory information, etc., and establish a unified data standard method to store the labeled data in the database to provide a data basis for subsequent steps;

[0077] 2) Use the data preprocessing module to complete the removal of irrelevant data, the processing of missing values ​​and outliers, the statistics of data features, and the grouping of training data; in the process of using LSTM for data training, each department must first preprocess the data, the main purpose of which is to delete irrelevant data in the original input data, avoid inputting data that does not meet the requirements, including some duplicate data and smooth noise data, and filter out invalid data that is irrelevant to the later data modeling, as well as processing missing values ​​and outliers in the original input data. After the data processing of each department is completed, the distribution information of the training data is counted, including the mean, variance, and covariance matrix of the training data, so as to facilitate the subsequent matching of institutional data with similar distributions for training grouping for federated learning.

[0078] 3) Use the LSTM neural network model to build a federated learning training module, and train it to obtain a usable early warning model; after each department has completed its own data collection, the central server initiates a training request for federated learning, and the department can participate in the federated training at this time. First, submit the data distribution information, and the central server divides different departments into several groups according to the data distribution information, so that departments with similar data distribution are trained together. The neural networks used in the present invention are LSTM trajectory prediction network and LSTM behavior prediction network, which are used for trajectory prediction and behavior prediction. The input of the LSTM trajectory prediction network is trajectory information, combined with the dilated convolution and attention mechanism, it can efficiently perform feature extraction and time series prediction. The trajectory prediction module is based on the historical trajectory information of the specific person whose trajectory needs to be predicted, as the input of LSTM, and the predicted trajectory of the future time domain is predicted by the network. The behavior prediction module is based on the output of the LSTM classification network for secondary inference, and the future trajectory is input according to the user's location and place, and the behavior to be performed by the specific person is classified, and the abnormal behavior of the specific person is discovered in time. After each department completes its own network training, it uploads the model to the central server. After receiving the department model, the central server aggregates the model and then sends the aggregated model to each department for further training until the model accuracy meets the requirements.

[0079] 4) Early warning model deployment: The early warning model is deployed on the department’s local server to make real-time predictions on the trajectory information of specific personnel, and then determine whether there is any dangerous behavior. If there is a possibility of dangerous behavior, timely measures are taken.

[0080] The early warning model is deployed on the department's local server to achieve real-time prediction of key personnel trajectory information. The main task of the early warning model is to monitor and analyze the actions of these key personnel in order to quickly identify potential dangerous behaviors. Once a situation with possible dangerous behavior is predicted, appropriate measures will be taken immediately to ensure timely and effective response. This real-time early warning system helps improve the department's safety and emergency response capabilities.

[0081] At the same time, the early warning model also uses the method of federated learning to train the latest model by aggregating data from different departments. Through federated learning, departments can share experience and information to continuously improve the accuracy and adaptability of the early warning model. This update mechanism ensures that the system can adapt to the changing environment and threats in a timely manner and maintain effective early warning capabilities for new dangerous behaviors. Therefore, the early warning system is not only real-time and accurate, but also can continuously improve overall safety performance as the situation evolves.

[0082] Embodiment 2:

[0083] This embodiment is further optimized on the basis of the above embodiment, and the same points as the above technical solution are not repeated here. Figure 1 to Figure 4 As shown, in order to better implement the text data distributed processing method based on federated learning described in the present invention, the following setting method is particularly adopted: the step 1) includes the following specific steps:

[0084] 1.1) Collect personal identity information of specific persons (such as ID card, license plate number, occupation, etc.) and collect social relationships between specific persons and other individuals (including relationships with friends, family members, colleagues, etc.);

[0085] 1.2) The trajectory information of a specific person is recorded by a sensing device to obtain historical trajectory data, and the historical trajectory data is divided into a text data sequence of length T; the feature data input by the LSTM network is the motion trajectory of each hour with a fixed sliding window length of T, and the motion trajectory at the next moment, i.e., moment T+1, is used as the target output, so the trajectory data is divided into a text data sequence of length T (in the present invention, T is usually set to 120, i.e., the position of a specific person is collected once every 30 seconds).

[0086] 1.3) Each department establishes a unified data standard method to form a unified coding format and display format for the personal information and activity data of specific personnel, which can achieve better training results. The labeled data is stored in the database to provide a data basis for subsequent steps.

[0087] Embodiment 3:

[0088] This embodiment is further optimized on the basis of any of the above embodiments, and the same points as the above technical solutions are not repeated here. Figure 1 to Figure 4 As shown, in order to better implement the text data distributed processing method based on federated learning described in the present invention, the following setting method is particularly adopted: the step 2) includes the following steps:

[0089] 2.1) Clear irrelevant data:

[0090] 2.1.1) Deduplication: Delete duplicate data samples to avoid duplication of model training;

[0091] 2.1.2) Smoothing noisy data: Processing the noise in the data through filtering or smoothing techniques ensures that the training data is more stable and reliable;

[0092] 2.1.3) Eliminate invalid data: Filter out data that is irrelevant to model building to ensure that the input features are relevant and meaningful;

[0093] 2.2) Handling missing values ​​and outliers:

[0094] 2.2.1) Missing value processing: Missing values ​​are likely to occur during data collection. Interpolation, filling or deletion methods are used to process missing values ​​in the data to maintain data integrity;

[0095] 2.2.2) Outlier processing: Use appropriate methods such as truncation, replacement or deletion to identify and process outliers to ensure data stability and accuracy;

[0096] 2.3) Statistical data characteristics:

[0097] After preprocessing, each department needs to collect statistics on the distribution information of the training data, which provides a key foundation for subsequent federated learning. The statistical data features include:

[0098] 2.3.1) Mean and variance: Calculate the mean and variance of data features to understand the central tendency and dispersion of the data set;

[0099] 2.3.2) Covariance matrix: Calculate the covariance between different features and explore the correlation between features;

[0100] 2.4) Group training data:

[0101] Based on statistical information, data holding institutions are divided into groups with similar distributions so that each institution can match data with similar distributions for training in federated learning.

[0102] Embodiment 4:

[0103] This embodiment is further optimized on the basis of any of the above embodiments, and the same points as the above technical solutions are not repeated here. Figure 1 to Figure 4 As shown, in order to better implement the text data distributed processing method based on federated learning described in the present invention, the following setting method is particularly adopted: the federated learning training module constructed by using the LSTM neural network model is specifically:

[0104] An RNN model is selected and improved using a dilated convolution and an attention mechanism; wherein LSTM (Long Short-Term Memory Network) is a special type of RNN (Recurrent Neural Network) that effectively handles the problem of gradient vanishing and gradient exploding through a gating mechanism. However, LSTM has high computational complexity when processing long sequence data, especially when the sequence length increases, which results in a large consumption of computing resources. In order to solve this problem, the present invention uses a dilated convolution and an attention mechanism to improve the model.

[0105] Atrous convolution is a convolution operation that can increase the receptive field without increasing the number of parameters. In 1D sequence processing, atrous convolution can be used to capture long-distance dependencies. Therefore, the present invention uses atrous convolution to solve the problem of complex text data calculation without increasing the amount of additional calculation.

[0106] The dilated convolution introduces a dilation rate d, which allows the elements in the convolution kernel to be spaced d-1 units apart. The expression of the dilated convolution operation is: d is the hole rate, x is the input sequence, w is the convolution kernel, y is the output sequence, and K is the size of the convolution kernel; the attention mechanism allows the model to dynamically focus on different parts of the sequence when processing the sequence, which can more effectively capture long-distance dependencies. In the attention mechanism, the attention score of each sequence element is calculated, and then these scores are used to weight the input sequence to obtain a weighted representation. The attention mechanism adopted by the present invention is: Q(query), K(key), and V(value) are the representations learned by the model: Q is the dimension of the key, which is used to scale the dot product to prevent excessive dot product values ​​from causing the softmax function to enter an area with a small gradient.

[0107] Combining dilated convolution and attention mechanism, a model that can efficiently process long sequence data can be constructed. For example, dilated convolution can be used to increase the receptive field, while the attention mechanism can be used to dynamically select important information in the sequence. Such a model structure can more effectively capture long-distance dependencies while keeping the number of parameters unchanged, and reduce the computational cost caused by the increase in sequence length.

[0108] After completing trajectory prediction for a specific person, the historical trajectory information and other information about the specific person (such as work and social information) are combined and cascaded using the LSTM neural network to obtain the behavior classification results.

[0109] Embodiment 5:

[0110] This embodiment is further optimized on the basis of any of the above embodiments, and the same points as the above technical solutions are not repeated here. Figure 1 to Figure 4 As shown, in order to better implement the text data distributed processing method based on federated learning described in the present invention, the following setting is particularly adopted: the training of the federated learning training module includes the following steps:

[0111] 3.1) The central server determines the training task and training model, broadcasts the agreed number of global training rounds E, and randomly initializes and broadcasts the global aggregate model parameters W g ;

[0112] 3.2) Departments choose whether to join the training. The central server determines the subset S that participates in the training according to the agreed ratio of the number of departments participating in each round of local training, and groups and aggregates different departments according to data distribution information (the number of groups is k);

[0113] 3.3) Department s (s∈S) will global aggregation model W g The initial local model of this round of training is trained for K rounds; the formula for local update is as follows:

[0114] 3.4) Department s uploads the gradient of each round The central server then implements global model aggregation based on the average weighted method:

[0115] 3.5) The central server sends the aggregated model parameters W g, k is given to each department of group k, and each department of group k continues to train using the model parameters;

[0116] 3.6) Repeat steps 3.3) to 3.5) for multiple rounds until the model achieves the desired effect. Department members of the group that meet the requirements can exit training early.

[0117] Embodiment 6:

[0118] This embodiment is further optimized on the basis of any of the above embodiments, and the same points as the above technical solutions are not repeated here. Figure 1 to Figure 4 As shown, in order to better implement the distributed text data processing method based on federated learning described in the present invention, the following setting is particularly adopted: in the step 3.6), during multiple rounds of repetitions of steps 3.3) to 3.5), department members of the group that meet the requirements directly exit the training early.

[0119] Embodiment 7:

[0120] This embodiment is further optimized on the basis of embodiment 4, and the same points as the above technical solution are not repeated here. Figure 1 to Figure 4 As shown, in order to better implement the distributed text data processing method based on federated learning described in the present invention, the following setting is particularly adopted: after the trajectory prediction of a specific person is completed, the dangerous behavior prediction is performed by combining the historical trajectory information and other information of the specific person (the occupation, social relationship, etc. of the specific person collected in step 1 of Example 1). These features will be used as the input of the LSTM classification model, and the activation function usually selects softmax to convert the output into a probability distribution, the sum of the probabilities of each category is 1, and the output layer will give the probability of each behavior category to analyze the behavior of the specific person.

[0121] Embodiment 8:

[0122] This embodiment is further optimized on the basis of any of the above embodiments, and the same points as the above technical solutions are not repeated here. Figure 1 to Figure 4 As shown, in order to better realize the distributed text data processing method based on federated learning described in the present invention, the following setting method is particularly adopted: the early warning model also uses the method of federated learning to continuously improve the accuracy and adaptability of the early warning model by aggregating data from different departments.

[0123] Embodiment 9:

[0124] A distributed processing method for text data based on federated learning, such as Figure 1 to Figure 4 As shown, including:

[0125] Part 1: Data acquisition and storage module:

[0126] Collect information and activity data of specific people, including user identity information, social relationships, trajectory information, etc., and establish a unified data standard method to store the labeled data in the database to provide a data basis for subsequent steps. Specifically:

[0127] Data collection:

[0128] The first step is to collect personal identity information of specific persons (such as ID card, license plate number, occupation, etc.) and collect social relationships between specific persons and other individuals (including relationships with friends, family members, colleagues, etc.);

[0129] The second step is to record the trajectory information of specific persons through sensing equipment to obtain historical trajectory data, and divide the historical trajectory data into a text data sequence of length T; the feature data input by the LSTM network is the motion trajectory of each hour with a fixed sliding window length of T, and the motion trajectory at the next moment, i.e., moment T+1, is used as the target output, so the trajectory data is divided into a text data sequence of length T (in the present invention, T is usually set to 120, i.e., the position of a specific person is collected once every 30 seconds).

[0130] Data Storage:

[0131] The third step is for each department to establish a unified data standard method, and form a unified coding format and display format for the personal information and activity data of specific personnel, which can achieve better training results. The labeled data is stored in the database to provide a data basis for the subsequent steps.

[0132] Part 2: Data preprocessing module:

[0133] Complete the removal of irrelevant data, processing of missing values ​​and outliers, statistics of data features, and grouping of training data; in the process of using LSTM for data training, each department must first pre-process the data, the main purpose of which is to delete irrelevant data in the original input data to avoid inputting data that does not meet the requirements, including some duplicate data and smooth noise data, and at the same time filter out invalid data that is irrelevant to the later data modeling, as well as process missing values ​​and outliers in the original input data. After completing the data processing of each department, the distribution information of the training data is counted, including the mean, variance, and covariance matrix of the training data, so as to facilitate the subsequent matching of institutional data with similar distributions for training grouping for federated learning.

[0134] Specifically:

[0135] Clear irrelevant data:

[0136] Deduplication: Delete duplicate data samples to avoid duplication in model training;

[0137] Smoothing noisy data: Processing the noise in the data through filtering or smoothing techniques ensures that the training data is more stable and reliable;

[0138] Eliminate invalid data: Filter out data that is irrelevant to model building to ensure that the input features are relevant and meaningful;

[0139] Handling missing values ​​and outliers:

[0140] Missing value processing: Missing values ​​are likely to occur during data collection. Use interpolation, filling or deletion to handle missing values ​​in the data to maintain data integrity;

[0141] Outlier processing: Use appropriate methods such as truncation, replacement or deletion to identify and process outliers to ensure data stability and accuracy;

[0142] Statistical data features:

[0143] After preprocessing, each department needs to collect statistics on the distribution information of the training data, which provides a key foundation for subsequent federated learning. The statistical data features include:

[0144] Mean and variance: Calculate the mean and variance of data features to understand the central tendency and dispersion of the data set;

[0145] Covariance matrix: calculate the covariance between different features and explore the correlation between features;

[0146] Group training data:

[0147] Based on statistical information, data holding institutions are divided into groups with similar distributions so that each institution can match data with similar distributions for training in federated learning.

[0148] Part III: Federated Learning Training Module:

[0149] Joint modeling module: Figure 1 As shown in the federated learning training module, after each department has completed its own data collection, the central server initiates a federated learning training request (the central server publishes a training task), and the department can now participate in the federated training (the department performs LSTM model training).

[0150] First, submit the data distribution information. The central server (department module) divides different departments into several groups according to the data distribution information, so that departments with similar data distribution are trained together. The neural networks used in the present invention are LSTM trajectory prediction network and LSTM behavior prediction network, which are used for trajectory prediction and behavior prediction. The input of LSTM trajectory prediction network is trajectory information, combined with dilated convolution and attention mechanism, it can efficiently perform feature extraction and time series prediction. The trajectory prediction module is based on the information of the historical trajectory of the specific person whose trajectory needs to be predicted, as the input of LSTM, and the predicted trajectory of the future time domain is obtained by network prediction. The behavior prediction module is based on the output of LSTM classification network for secondary inference, according to the user's location and place, input future trajectory, classify the behavior that specific personnel will perform, and promptly discover abnormal behavior of specific personnel. After the department completes its own network training, it uploads the model to the central server. After receiving the department model, the central server performs model aggregation, and then sends the aggregated model to each department for continued training until the model accuracy meets the requirements.

[0151] LSTM (Long Short-Term Memory Network) is a special type of RNN (Recurrent Neural Network) that effectively handles the problem of gradient vanishing and gradient exploding through a gating mechanism. However, LSTM has high computational complexity when processing long sequence data, especially when the sequence length increases, which leads to a large consumption of computing resources. To solve this problem, the present invention uses dilated convolution and attention mechanism to improve the model.

[0152] Atrous convolution is a convolution operation that can increase the receptive field without increasing the number of parameters. In 1D sequence processing, atrous convolution can be used to capture long-distance dependencies. Therefore, the present invention uses atrous convolution to solve the problem of complex text data calculation without increasing the amount of additional calculation.

[0153] The dilated convolution introduces a dilation rate d, which allows the elements in the convolution kernel to be spaced d-1 units apart. The expression of the dilated convolution operation is: d is the hole rate, x is the input sequence, w is the convolution kernel, y is the output sequence, and K is the size of the convolution kernel; the attention mechanism allows the model to dynamically focus on different parts of the sequence when processing the sequence, which can more effectively capture long-distance dependencies. In the attention mechanism, the attention score of each sequence element is calculated, and then these scores are used to weight the input sequence to obtain a weighted representation. The attention mechanism adopted by the present invention is: Q(query), K(key), and V(value) are the representations learned by the model: Q is the dimension of the key, which is used to scale the dot product to prevent excessive dot product values ​​from causing the softmax function to enter an area with a small gradient.

[0154] Combining dilated convolution and attention mechanism, a model that can efficiently process long sequence data can be constructed. For example, dilated convolution can be used to increase the receptive field, while the attention mechanism can be used to dynamically select important information in the sequence. Such a model structure can more effectively capture long-distance dependencies while keeping the number of parameters unchanged, and reduce the computational cost caused by the increase in sequence length.

[0155] After completing trajectory prediction for a specific person, the historical trajectory information and other information about the specific person (such as work and social information) are combined and cascaded using the LSTM network to obtain the behavior classification results.

[0156] The training of the federated learning training module includes the following steps:

[0157] Step 1: The central server determines the training task and training model, broadcasts the agreed number of global training rounds E, and randomly initializes and broadcasts the global aggregate model parameters W g ;

[0158] Step 2: The department chooses whether to join the training. The central server determines the subset S to participate in the training according to the agreed proportion of the number of departments participating in each round of local training, and groups and aggregates different departments according to data distribution information (the number of groups is k);

[0159] Step 3: Department s (s∈S) aggregates the global model W g The initial local model of this round of training is trained for K rounds; the formula for local update is as follows:

[0160] Step 4: Department s uploads the gradient of each round The central server then implements global model aggregation based on the average weighted method:

[0161] Step 5: The central server sends the aggregated model parameters W g,kEach department in group k is given a model parameter, and each department in group k continues to train using the model parameters;

[0162] Step 6: Repeat steps 3) to 5) for multiple rounds until the model achieves the desired effect. Department members of the group that meets the requirements can exit the training early.

[0163] Part 4: Early warning model deployment: Deploy the early warning model on the department’s local server to make real-time predictions of specific personnel’s trajectory information. Figure 4 As shown in the figure, first, the current and historical trajectory information of the target person is input into the LSTM trajectory prediction network to predict the future trajectory of the specific person. After obtaining the future trajectory prediction result, the historical trajectory information and other data of the specific person are sent to the LSTM behavior prediction network to determine whether there is any dangerous behavior. If there is a possibility of dangerous behavior, timely measures are taken.

[0164] The early warning model is deployed on the department's local server to achieve real-time prediction of key personnel trajectory information. The main task of the early warning model is to monitor and analyze the actions of these key personnel in order to quickly identify potential dangerous behaviors. Once a situation with possible dangerous behavior is predicted, appropriate measures will be taken immediately to ensure timely and effective response. This real-time early warning system helps improve the department's safety and emergency response capabilities.

[0165] At the same time, the early warning model also uses the method of federated learning to train the latest model by aggregating data from different departments. Through federated learning, departments can share experience and information to continuously improve the accuracy and adaptability of the early warning model. This update mechanism ensures that the system can adapt to the changing environment and threats in a timely manner and maintain effective early warning capabilities for new dangerous behaviors. Therefore, the early warning system is not only real-time and accurate, but also can continuously improve overall safety performance as the situation evolves.

[0166] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification or equivalent change made to the above embodiment based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A distributed text data processing method based on federated learning, characterized in that: The steps include: 1) Complete the collection of information and activity data of specific personnel through the data collection and storage module, establish a unified data standard method, and store the labeled data in the database; 2) Use the data preprocessing module to complete the removal of irrelevant data, the processing of missing values ​​and outliers, the statistics of data features and the grouping of training data; 3) Use the LSTM neural network model to build a federated learning training module, and train it to obtain a usable early warning model; 4) Early warning model deployment: The early warning model is deployed on the department’s local server to make real-time predictions on the trajectory information of specific personnel, and then determine whether there is any dangerous behavior. If there is a possibility of dangerous behavior, timely measures are taken.

2. A distributed text data processing method based on federated learning according to claim 1, characterized in that: The step 1) comprises the following specific steps: 1.1) Collect personal identity information of specific persons and collect social relationships between specific persons and other individuals; 1.2) Record the trajectory information of a specific person through a sensor device to obtain historical trajectory data, and divide the historical trajectory data into a text data sequence of length T; 1.3) Each department shall establish a unified data standard method to form a unified coding format and display format for the personal information and activity data of specific personnel, and store the labeled data in the database.

3. The text data distributed processing method based on federated learning according to claim 1, characterized in that: The step 2) comprises the following steps: 2.1) Clear irrelevant data: 2.1.1) Deduplication: Delete duplicate data samples; 2.1.2) Smoothing noisy data: Processing the noise in the data through filtering or smoothing techniques; 2.1.3) Eliminate invalid data: filter out data that is not relevant to model building; 2.2) Handling missing values ​​and outliers: 2.2.1) Missing value processing: Use interpolation, filling or deletion methods to handle missing values ​​in the data; 2.2.2) Outlier processing: Use truncation, replacement or deletion methods to identify and process outliers; 2.3) Statistical data characteristics: 2.3.1) Mean and variance: Calculate the mean and variance of data features; 2.3.2) Covariance matrix: calculate the covariance between different features; 2.4) Group training data: Based on statistical information, data holding institutions are divided into groups with similar distributions so that each institution can match data with similar distributions for training in federated learning.

4. A distributed text data processing method based on federated learning according to claim 1, 2 or 3, characterized in that: The specific steps of constructing the federated learning training module using the LSTM neural network model are as follows: Select an RNN model and use dilated convolution and attention mechanism to improve the RNN model; the expression of dilated convolution operation is: d is the hole rate, x is the input sequence, w is the convolution kernel, y is the output sequence, and K is the size of the convolution kernel; the attention mechanism is: Q(query), K(key), and V(value) are the representations learned by the model: Q is the dimension of the key, which is used to scale the dot product to prevent excessive dot product values ​​from causing the softmax function to enter an area with a small gradient.

5. A distributed text data processing method based on federated learning according to claim 1, 2 or 3, characterized in that: The training of the federated learning training module includes the following steps: 3.1) The central server determines the training task and training model, broadcasts the agreed number of global training rounds E, and randomly initializes and broadcasts the global aggregate model parameters W g ; 3.2) Departments choose whether to join the training. The central server determines the subset S that participates in the training according to the agreed ratio of the number of departments participating in local training in each round, and groups and aggregates different departments according to data distribution information; 3.3) Department s will global aggregation model W g As the initial local model of this round of training, K rounds of training are performed; the formula for local update is as follows: θ s ←θ s -υ l (p s (θ s )+γ‖θ s -W g,s ‖); 3.4) Department s uploads the gradient θ of each round s -W g,s , and then the central server implements global model aggregation based on the average weighted method: 3.5) The central server sends the aggregated model parameters W g,k Each department in group k is given a model parameter, and each department in group k continues to train using the model parameters; 3.6) Repeat steps 3.3) to 3.5) for multiple rounds until the model achieves the desired effect.

6. The method for distributed processing of text data based on federated learning according to claim 5, characterized in that: In the step 3.6), during multiple rounds of repetitions of steps 3.3) to 3.5), department members of the group that meet the requirements directly exit the training early.

7. A text data distributed processing method based on federated learning according to claim 1, 2, 3 or 6, characterized in that: After using the LSTM neural network model to predict the trajectory of a specific person, the historical trajectory information and other information of the specific person are combined, and the LSTM neural network is used for cascading to obtain the behavior classification results.

8. A distributed text data processing method based on federated learning according to claim 1, 2, 3 or 6, characterized in that: The early warning model also uses the method of federated learning to continuously improve the accuracy and adaptability of the early warning model by aggregating data from different departments.

Images