Behavior algorithm research and judgment model system

Abstract

The invention discloses a behavior algorithm research and judgment model system, and compared with the prior art, the behavior algorithm research and judgment model system comprises a data collection module, a data processing module, a feature engineering module, a model loading module, an algorithm research and judgment module and a feedback module, the data collection module is used for collecting original data related to behavior analysis and providing a data basis for the whole analysis process, and the feature engineering module is used for extracting features useful for behavior study and judgment from the preprocessed data; the model loading module is used for loading a pre-trained behavior study and judgment model according to a specific application scene, data characteristics and business requirements and / or initializing a new model according to a selected algorithm. The method has good expandability and flexibility, can be customized according to specific requirements and can be easily integrated into various systems, and application scenes are expanded. And subjective errors and uncertainty are reduced by reducing human intervention and automatic analysis.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a behavior algorithm analysis model system. Background Art

[0002] Conventional technical content of the industry at home and abroad at this stage:

[0003] 1. Video processing technology: including video encoding, decoding, transmission, storage, etc., is the basis for processing video data.

[0004] 2. Image processing technology: involves image preprocessing, feature extraction, edge detection, etc., providing basic data for behavior recognition.

[0005] The main disadvantages of current conventional technologies include:

[0006] 1. Slow technological progress: In recent years, the pace of technological progress has slowed down, making it impossible for developers to keep pace with the development of the industry, affecting the innovation atmosphere and development prospects.

[0007] 2. Lack of unified technical standards: Many unified standards and specifications related to technologies have not yet been issued, resulting in differences in some technologies and having an adverse impact on industry and investment activities.

[0008] 3. Inadequate industry supervision: Due to the imperfect industry supervision mechanism and insufficient government supervision, the industry has experienced internal chaos, technological differences, and uncertain prices, which have seriously affected the development of the industry.

[0009] 4. Large influence of external environment: Technology is greatly affected by the external environment. The constant changes in social and economic conditions, interference from external factors and other factors have restricted the development of technology, and some technologies have become obsolete.

[0010] 5. Low security: Some technologies have security risks and may cause serious damage to users’ information, privacy, and data, and may also cause financial losses to users.

[0011] 6. Difficulty in cross-border application: The cross-border application of technology is difficult. Many technologies are not easy to apply across regions and industries. As a result, resource sharing and cooperation between industries are more difficult, which affects the development of the industry. Summary of the invention

[0012] The present application aims to solve one of the technical problems in the related art at least to some extent.

[0013] To this end, one purpose of this application is to propose a behavioral algorithm research and judgment model system to solve many problems in current technology. This system aims to promote technological progress and improve the overall efficiency of the industry through innovative algorithm models and behavioral analysis technologies. First of all, by increasing R&D investment, we will continue to improve and optimize the algorithm model to ensure that it has the ability to respond quickly and adapt to complex scenarios. At the same time, the system supports the gradual unification of industry standards, and can be applied in different fields in accordance with unified operating specifications, thereby reducing technical differences and promoting the coordinated development of various fields. In addition, the system focuses on security in design, integrates multi-level information security protection measures, and ensures high protection of user information and privacy to cope with potential data leakage risks and privacy issues. The system has the ability to cross-border applications in terms of function, so that it is not limited to the application needs of a single industry, but can also adapt to resource sharing and data interoperability in different industries. This cross-border application helps to promote cooperation between different industries, achieve efficient allocation of resources, and further promote the integration and development of industries. For example, the behavioral algorithm research and judgment model system can play an important role in smart homes, public safety, and medical health, and effectively improve the service quality and work efficiency of various industries through customized algorithms and prediction models. Overall, the system takes technological innovation as its core and provides a new solution for the rapid development and healthy standardization of technology through the organic combination of industry standardization, security enhancement and cross-border applications.

[0014] To achieve the above-mentioned purpose, the first embodiment of the present application proposes a behavior algorithm analysis model system, which includes a data collection module, a data processing module, a feature engineering module, a model loading module, an algorithm analysis module and a feedback module, wherein:

[0015] The data collection module is used to collect raw data related to behavior analysis and provide a data basis for the entire analysis process.

[0016] The data processing module is used to pre-process the raw data.

[0017] The feature engineering module is used to extract features useful for behavior analysis from preprocessed data.

[0018] The model loading module is used to load a pre-trained behavior analysis model and / or initialize a new model based on the selected algorithm according to specific application scenarios, data characteristics and business requirements.

[0019] The judgment module is used to load the feature input value model extracted by the feature engineering module into the model selected by the model loading module. The model processes the input features according to the knowledge it has learned and outputs the judgment results.

[0020] The feedback module is used to take corresponding safety measures according to the analysis results, send warning notifications to users, etc. At the same time, it collects user feedback on the results for subsequent model optimization and improvement.

[0021] In one embodiment of the present application, the data collection module collects raw data related to behavior analysis including the following steps:

[0022] S101: Obtaining the video captured by the camera from the monitoring system,

[0023] S102: Obtain text, pictures and videos posted by users from social media platforms.

[0024] S103: Obtain the user's behavior track in the log data from the system log and application log,

[0025] Among them, the user's behavior trajectory refers to the activity records of logging in, browsing, clicking, submitting and / or purchasing generated by the user in the process of using the system or application, and logging in, browsing, clicking, submitting and purchasing are types of log behaviors. The original data includes videos, texts, pictures and behavior trajectories.

[0026] In one embodiment of the present application, the data processing module implements the following steps:

[0027] S201: Data cleaning:

[0028] Video data: Remove blurry frames, dark or noisy frames, delete duplicate clips, and retain high-definition video content.

[0029] Text data: remove redundant characters and marks, delete irrelevant content, clean up duplicate text records, and ensure that text data is clear and concise.

[0030] Image data: Eliminate blurry or repeated images, remove interference noise in images, and ensure image clarity and data quality.

[0031] Behavior trajectory data: Delete redundant or irrelevant records in system logs, remove abnormal operation records, and retain the user's actual behavior path data;

[0032] S202: Data collation:

[0033] Video data: Arrange video frames in chronological order to ensure the continuity of video content, and save them by scene or behavior type.

[0034] Text Data: Sort and categorize text by content type and time, ensuring content is in a consistent format across different data sources.

[0035] Image data: Classify by shooting time, location or content to ensure structured storage of image information.

[0036] Behavior track data: Sort behavior track records by user ID, operation time, and operation content;

[0037] S203: Missing value processing:

[0038] Video data: Lost or missing frames are supplemented by interpolating adjacent frames, or adjacent frames are copied to fill the gaps to ensure video continuity.

[0039] Text data: Fill in missing fields. For example, irrelevant or missing content can be marked with null values ​​to ensure data completeness.

[0040] Image data: Replace missing images with similar images or mark them as empty to ensure that the image data is continuous and complete.

[0041] Behavior trajectory data: Use interpolation or default values ​​to fill missing behavior points to ensure the continuity of behavior trajectory data.

[0042] S204: Data conversion:

[0043] Video data: Convert the video frame by frame into an image sequence to generate a grayscale image or edge map.

[0044] Text data: Python-based NLP libraries such as NLTK, spaCy, Keras, and scikit-learn can be used to normalize text to the same length by padding, truncating, or using specific vectorization methods.

[0045] Image data: Convert the image into a grayscale image or feature matrix.

[0046] Behavior trajectory data: After numbering the behavior events in different logs, extract the user behavior trajectory according to the preset duration, replace each behavior with the corresponding number, and combine them into time series data in chronological order.

[0047] In one embodiment of the present application, the features include edge information, corner points, texture features of video image data, TF-IDF vectors of text data, and the frequency of occurrence of behaviors of time series data.

[0048] The feature engineering module implements the following steps:

[0049] S301: Feature extraction,

[0050] Video and image data: Use edge detection algorithms such as Canny and Sobel to extract edge information in the image, use Harris corner detection and / or FAST method to identify corners in the image, and calculate the texture features of the image through the gray level co-occurrence matrix. The texture features include contrast, correlation and energy.

[0051] Text data: Split the text into independent words, remove irrelevant prepositions and articles, calculate the TF-IDF weight for each word, and arrange the TF-IDF value of each word into a vector as the TF-IDF vector.

[0052] Time series data: Calculate the frequency of occurrence of each behavior in the time series data;

[0053] S302: Feature selection:

[0054] Relevance: Select features that are highly relevant to the behavior analysis target.

[0055] Validity: Ensure that the selected features can truly reflect the behavioral characteristics and avoid introducing noise.

[0056] Redundancy: Avoid selecting highly correlated features to reduce computational complexity and overfitting risk.

[0057] In one embodiment of the present application, the model loading module implements the following steps:

[0058] 401: Determining model type:

[0059] Choose the appropriate model type based on data characteristics and business requirements:

[0060] For labeled classification or regression tasks, choose supervised learning algorithms such as decision trees, random forests, and / or SVMs.

[0061] For grouping and anomaly identification of unlabeled data, use unsupervised learning algorithms for cluster analysis and / or anomaly detection,

[0062] For processing image, video, text and time series data, select CNN, RNN and / or LSTM deep learning algorithms, combined with spatiotemporal modeling to capture the temporal and spatial variation characteristics of the data through optical flow, trajectory modeling, and spatiotemporal graph network methods;

[0063] S402: Loading pre-trained model:

[0064] If a pre-trained model exists, load the model weights and parameters from storage, check the consistency of the model structure and parameters, and ensure that the pre-trained model meets the current task requirements.

[0065] If the model has been trained in a related scenario, apply transfer learning to migrate the model from the original task to the new task to cope with the situation of insufficient data, thereby improving training efficiency;

[0066] S403: Model initialization:

[0067] In the absence of a pre-trained model, define the model architecture and hyperparameters, such as the number of layers, number of neurons, activation function, learning rate, etc., set the model weights using random initialization and / or transfer learning, and prepare for model training or inference.

[0068] S404: Configure training and evaluation parameters:

[0069] Set the batch size, learning rate, loss function and evaluation index parameters required for training. The evaluation index parameters include accuracy and recall rate.

[0070] S405: Small scale test:

[0071] Perform a small sample test on the loaded model to ensure that the model is loaded correctly and can run normally, ready for the subsequent analysis module.

[0072] In one embodiment of the present application, the algorithm analysis module implements the following steps:

[0073] S501: Data input and preprocessing:

[0074] Input the features extracted by the feature engineering module into the model, ensuring that the format and dimension of the input data are consistent with the input requirements of the model;

[0075] S502: Application of attention mechanism:

[0076] In video and time series data processing, the attention mechanism is applied to focus the model's attention on specific key areas or time periods, automatically identify and highlight key information related to behavior, and filter out noise information to improve the overall accuracy of the data;

[0077] S503: Model Reasoning and Fusion:

[0078] The model selected by the model loading module is used to infer the input feature data to generate preliminary prediction results, wherein the preliminary prediction results include: classification labels for classifying behaviors into normal behaviors, abnormal behaviors, and fraudulent behaviors, and assigning a probability score and / or probability value to the behaviors indicating that the behaviors belong to a certain category,

[0079] Adopt model fusion and integration, through simple weighted average, voting, or more complex stacking generalization, bagging and boosting, to fuse the prediction results of multiple models to obtain more reliable fusion prediction results, where the fusion prediction results include: classification labels for dividing behaviors into normal behaviors, abnormal behaviors, and fraudulent behaviors, and giving behaviors a probability score and / or probability value indicating that the behavior belongs to a certain category;

[0080] S504: Abnormal detection and early warning:

[0081] For the scenario of abnormal behavior identification, statistical methods, distance-based, density-based methods or deep learning methods of anomaly detection technology are applied to identify abnormal events that significantly deviate from normal behavior patterns from the model output.

[0082] Set the threshold range for anomaly detection, and trigger an alert when the score or probability value in the fusion prediction result exceeds a specific threshold;

[0083] S505: Generate analysis results:

[0084] Based on the output of the model, the research and judgment results are converted into easy-to-understand classification labels, scores / probability values ​​or abnormal marks, and output as structured reports to provide a basis for decision-making and subsequent actions;

[0085] S506: Real-time processing and optimization:

[0086] Monitor the model reasoning process and data flow in real time, use caching, parallel processing, and memory optimization to ensure fast response in high-load or real-time scenarios, and adjust model parameters or memory allocation to ensure the efficiency and stability of the analysis process.

[0087] In one embodiment of the present application, the feedback module implements the following steps:

[0088] S601: Execute security measures and warning notifications: If abnormal behavior is detected, the system sends a warning notification to the user or triggers preventive security measures;

[0089] S602: User feedback collection:

[0090] Collect feedback on the research and judgment results from users or relevant personnel, including:

[0091] Questionnaire survey: Design a questionnaire to understand users’ satisfaction and suggestions on the research results.

[0092] User complaint records: record abnormal situations or complaints reported by users to provide reference data for subsequent analysis;

[0093] S603: Model performance evaluation:

[0094] Regularly evaluate the performance of the model and check its accuracy, recall, and F1 score in the current data environment. Accuracy is the proportion of correctly classified results, recall is the proportion of correctly identified anomalies in actual data, and F1 score is a comprehensive score used to balance accuracy and recall.

[0095] S604: Feedback data storage and management:

[0096] All user feedback, model performance evaluation results and judgment data are stored in the feedback database to facilitate subsequent analysis and improvement.

[0097] Categorize and store user feedback, performance indicators, and abnormal records, set validity periods for data in the feedback database, and regularly update and clean up expired data;

[0098] S605: Model iteration and optimization:

[0099] Model parameter tuning: Adjust the model’s learning rate and / or regularization parameters to adapt to the new data distribution,

[0100] Data update: Add new feedback data to the model training data set to improve the model's adaptability to the current data.

[0101] Retraining or fine-tuning: After collecting enough new data, retrain or fine-tune the model to improve the model's prediction accuracy for new scenarios.

[0102] In one embodiment of the present application, the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of a behavioral algorithm analysis model system when executing the program.

[0103] In one embodiment of the present application, the present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that when the program is executed by a processor, the steps of the behavioral algorithm analysis model system are implemented.

[0104] In one embodiment of the present application, the present invention further provides a processor, characterized in that the processor is used to run a program, wherein the program executes the steps of the behavior algorithm analysis model system when running.

[0105] Compared with the prior art, the behavior algorithm analysis model system of the embodiment of the present application has the following advantages:

[0106] 1. Higher accuracy and efficiency:

[0107] Behavioral algorithm analysis model technology, especially models based on advanced technologies such as deep learning, can automatically learn and extract complex features from large amounts of data, thereby achieving more accurate behavior analysis and analysis. This automated and intelligent processing method greatly improves accuracy and efficiency compared to traditional manual analysis or simple algorithms.

[0108] 2. Multimodal data processing capabilities:

[0109] Behavioral algorithm analysis model technology can process multimodal data from different channels, such as videos, images, log data, etc. This multimodal data processing capability enables the model to capture behavioral characteristics more comprehensively and improve the comprehensiveness and accuracy of analysis. In contrast, conventional technologies can often only process a single type of data and have difficulty coping with complex and changing behavioral scenarios.

[0110] 3. Adaptability and robustness:

[0111] With the continuous development of technology, behavioral algorithm analysis model technology has become more adaptive and robust. They can continuously learn and optimize themselves based on new data to adapt to changing behavior patterns and scenarios. At the same time, they are more capable of handling noise and abnormal data and can maintain stable performance in complex environments.

[0112] 4. Real-time and dynamic:

[0113] Behavior algorithm judgment model technology usually has the ability to process and analyze data in real time, and can quickly respond to behavioral changes and make corresponding judgments. This real-time and dynamic performance is particularly important for scenarios that require rapid decision-making and response, such as autonomous driving and intelligent monitoring. In contrast, conventional technologies often find it difficult to achieve such high real-time and dynamic performance.

[0114] 5. Scalability and flexibility:

[0115] Behavioral algorithm judgment model technologies have strong scalability and flexibility, and can be customized and optimized according to different application scenarios and requirements. They can be easily integrated into various systems and work with other technical components to achieve more complex functions and applications. This scalability and flexibility make behavioral algorithm judgment model technology have broad application prospects in multiple fields.

[0116] 6. Reduce human intervention and errors:

[0117] Behavioral algorithm judgment model technology reduces the possibility of human intervention and errors by automatically processing and analyzing data. This can not only improve the accuracy and efficiency of judgment, but also reduce the risks and uncertainties caused by human factors. In contrast, conventional technologies often rely on manual analysis and judgment, which are easily affected by subjective factors and human errors.

[0118] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through the practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0119] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0120] Figure 1 It is a module flow diagram of the behavior algorithm analysis model system of the present invention.

[0121] Figure 2 It is a modular schematic diagram of the data processing module of the present invention.

[0122] Figure 3 It is a modular schematic diagram of the feature engineering module of the present invention.

[0123] Figure 4 It is a modular schematic diagram of the model loading module of the present invention. DETAILED DESCRIPTION

[0124] Embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limitations on the present application. On the contrary, the embodiments of the present application include all changes, modifications and equivalents that fall within the spirit and connotation of the appended claims.

[0125] The behavior algorithm analysis model system of an embodiment of the present application is described below with reference to the accompanying drawings.

[0126] The behavior algorithm analysis model system provided in the embodiment of the present application can be applied to.

[0127] like Figure 1-4 As shown, the behavior algorithm analysis model system of the embodiment of the present application includes a data collection module, a data processing module, a feature engineering module, a model loading module, an algorithm analysis module and a feedback module, wherein:

[0128] The data collection module is used to collect raw data related to behavior analysis and provide a data basis for the entire analysis process.

[0129] The data processing module is used to pre-process the raw data.

[0130] The feature engineering module is used to extract features useful for behavior analysis from preprocessed data.

[0131] The model loading module is used to load a pre-trained behavior analysis model and / or initialize a new model based on the selected algorithm according to specific application scenarios, data characteristics and business requirements.

[0132] The judgment module is used to load the feature input value model extracted by the feature engineering module into the model selected by the model loading module. The model processes the input features according to the knowledge it has learned and outputs the judgment results.

[0133] The feedback module is used to take corresponding safety measures according to the analysis results, send warning notifications to users, etc. At the same time, it collects user feedback on the results for subsequent model optimization and improvement.

[0134] It should be noted that in order to achieve layer-by-layer processing and optimization of data, so that the system can run stably and efficiently and meet various behavior analysis requirements, the relationship between the modules in the present invention is as follows:

[0135] 1. The data collection module obtains the original data and passes it to the data processing module for preprocessing to ensure that the original data has the integrity and availability required for analysis, remove noise and irrelevant content, and lay a good foundation for subsequent steps.

[0136] 2. The data processing module passes the preprocessed clean data to the feature engineering module, so that the feature engineering module can extract key features from structured and standardized data and improve the model's ability to understand and identify behaviors.

[0137] 3. The useful features extracted by the feature engineering module are used as the main content of the model input to ensure that the information received by the model can reflect the important features in the data and help the model learn the behavior pattern more accurately.

[0138] 4. The model loading module loads or initializes the model according to demand, and passes the model structure and parameters to the algorithm analysis module for prediction tasks, ensuring that the model used by the analysis module has been debugged and is suitable for the current analysis task, thereby improving the accuracy and consistency of the prediction results.

[0139] 5. The analysis results generated by the algorithm analysis module are passed to the feedback module for subsequent decision-making and feedback collection. The feedback module takes corresponding measures based on the analysis results and collects user or system feedback information to provide continuous data support for model optimization.

[0140] 6. The feedback module passes user feedback and performance evaluation information back to the model loading module, and if necessary, to the feature engineering module. Through the feedback mechanism, the model is continuously optimized and the feature selection criteria are updated to ensure that the model can adapt to new data and business needs and achieve a virtuous closed loop.

[0141] The functions of the above modules are independent and clear, which is convenient for maintenance and upgrading. A single module can be improved according to different needs to ensure that the system can respond flexibly when the needs change. From data collection to processing, and then to feature extraction, the data quality is gradually improved, making the input data of the model more reliable and clearer, and improving the overall performance of the model. The model loading module supports the fusion of multiple algorithms and models. The algorithm analysis module can select the most appropriate model architecture and further combine the predictions of multiple models to ensure the accuracy and stability of the prediction results. The feedback module provides a real-time feedback mechanism for the model, allowing the system to continuously learn and optimize in practical applications. Through user feedback and performance evaluation, the model can be iteratively updated regularly to maintain accurate identification of behavior patterns. The algorithm analysis module can quickly identify and respond to abnormal behaviors through anomaly detection and early warning mechanisms to ensure system security. The real-time processing module further optimizes the response time, enabling the system to quickly respond to potential threats. Through real-time processing and optimization, the system can maintain stable and efficient operation in scenarios with high load or high real-time requirements, meet large-scale data processing needs, and improve the reliability of the system.

[0142] In one embodiment of the present application, the data collection module collects raw data related to behavior analysis including the following steps:

[0143] S101: Obtaining the video captured by the camera from the monitoring system,

[0144] S102: Obtain text, pictures and videos posted by users from social media platforms.

[0145] S103: Obtain the user's behavior track in the log data from the system log and application log,

[0146] Among them, the user's behavior trajectory refers to the activity records of logging in, browsing, clicking, submitting and / or purchasing generated by the user in the process of using the system or application, and logging in, browsing, clicking, submitting and purchasing are types of log behaviors. The original data includes videos, texts, pictures and behavior trajectories.

[0147] In one embodiment of the present application, the data processing module implements the following steps:

[0148] S201: Data cleaning:

[0149] Video data: Remove blurry frames, dark or noisy frames, delete duplicate clips, and retain high-definition video content.

[0150] Text data: remove redundant characters and marks, delete irrelevant content, clean up duplicate text records, and ensure that text data is clear and concise.

[0151] Image data: Eliminate blurry or repeated images, remove interference noise in images, and ensure image clarity and data quality.

[0152] Behavior trajectory data: Delete redundant or irrelevant records in system logs, remove abnormal operation records, and retain the user's actual behavior path data;

[0153] S202: Data collation:

[0154] Video data: Arrange video frames in chronological order to ensure the continuity of video content, and save them by scene or behavior type.

[0155] Text Data: Sort and categorize text by content type and time, ensuring content is in a consistent format across different data sources.

[0156] Image data: Classify by shooting time, location or content to ensure structured storage of image information.

[0157] Behavior track data: Sort behavior track records by user ID, operation time, and operation content;

[0158] S203: Missing value processing:

[0159] Video data: Lost or missing frames are supplemented by interpolating adjacent frames, or adjacent frames are copied to fill the gaps to ensure video continuity.

[0160] Text data: fill in missing fields. For example, irrelevant or missing content can be marked with null values ​​to ensure data integrity.

[0161] Image data: Replace missing images with similar images or mark them as empty to ensure that the image data is continuous and complete.

[0162] Behavior trajectory data: Use interpolation or default values ​​to fill in missing behavior points to ensure the continuity of behavior trajectory data;

[0163] S204: Data conversion:

[0164] Video data: Convert the video frame by frame into an image sequence to generate a grayscale image or edge map.

[0165] Text data: Python-based NLP libraries such as NLTK, spaCy, Keras, and scikit-learn can be used to normalize text to the same length by padding, truncating, or using specific vectorization methods.

[0166] Image data: Convert the image into a grayscale image or feature matrix.

[0167] Behavior trajectory data: After numbering the behavior events in different logs, extract the user behavior trajectory according to the preset duration, replace each behavior with the corresponding number, and combine them into time series data in chronological order;

[0168] It should be noted that, in this embodiment, the specific implementation steps of normalizing the text to the same length by padding, truncating, or using a specific vectorization method are as follows:

[0169] 1. Filling:

[0170] Pad shorter text to a specified length. Usually uses 0 or blank padding to ensure that all text reaches the same length.

[0171] For example, fill text less than 100 words with zeros to 100 words.

[0172] Original text: "I like machine learning"

[0173] Fill the result: "I like machine learning 000..." (to 100 words)

[0174] 2. Truncation:

[0175] For text that is longer than the specified upper limit, the preceding part is retained and the excess part is truncated.

[0176] Example:

[0177] Set a maximum length, for example 100 words, and truncate text longer than 100 words to 100 words.

[0178] Original text: "I like machine learning, especially natural language processing, very interesting. Hope to learn more."

[0179] Truncated result: "I like machine learning, especially natural language processing, it's interesting..."

[0180] 3. Fixed-length vectorization:

[0181] Build a vocabulary, represent the text as a vector with the same length as the vocabulary, calculate features based on term frequency-inverse document frequency TF-IDF, and obtain a vector representation of fixed length.

[0182] Implementation: Regardless of the length of the text, convert each text into a bag-of-words or TF-IDF vector, with the size determined by the vocabulary size (e.g. 5000 dimensions).

[0183] Applicable scenarios: Applicable to tasks such as text classification and topic analysis to ensure that all text vectors are consistent.

[0184] 4. Combination of sequence filling and truncation:

[0185] Description: First fill the text sequence to ensure that the length is not less than the minimum value; for long text, truncate it to control the maximum length. You can set the minimum and maximum length to handle the problem of short text filling and long text truncation.

[0186] Implementation: Assuming the length requirement is between 50 and 100 words, short texts are padded to 50 words and long texts are truncated to 100 words.

[0187] It should be noted that, in this embodiment, the specific data conversion steps of the behavior trajectory data are implemented by:

[0188] 1. Action number:

[0189] Define behavior categories: First, identify the key behavior categories in the system so that each behavior can be assigned a uniform number. Behavior categories usually include the core operations performed by users in the system, such as logging in, browsing, clicking, submitting, purchasing, etc.

[0190] Assign behavior numbers: Assign a unique number to each behavior category to ensure that each behavior has a unique identification in the system, for example:

[0191] Login=1,

[0192] Views = 2,

[0193] Click = 3,

[0194] Submit = 4,

[0195] Purchase = 5,

[0196] Create a behavior coding table: record the behavior name and its corresponding number to facilitate subsequent query and replacement of behavior codes.

[0197] 2. Arrange in chronological order:

[0198] Extract user behavior records: Extract user behavior events from system logs or application logs, including the behavior type and the timestamp of the behavior.

[0199] Sort by time: Sort the extracted behavior records in ascending order by timestamp to ensure that each user's behavior sequence is arranged in the order of occurrence.

[0200] 3. Generate time series:

[0201] Behavior code replacement: Replace each behavior record with its corresponding number to form a set of coding sequences. For example, if a user's behavior sequence is "browse, click, log in, log in, log in", then replace it with the number to form the sequence "32111".

[0202] Combine into a time series: Combine these codes into a complete time series according to the order in which the behaviors occurred. This time series can be regarded as the operation record of the user in a specific time period, which is convenient for subsequent pattern recognition or analysis.

[0203] In one embodiment of the present application, the features include edge information, corner points, texture features of video image data, TF-IDF vectors of text data, and the frequency of occurrence of behaviors of time series data.

[0204] The feature engineering module implements the following steps:

[0205] S301: Feature extraction,

[0206] Video and image data: Use edge detection algorithms such as Canny and Sobel to extract edge information in the image, use Harris corner detection and / or FAST method to identify corners in the image, and calculate the texture features of the image through the gray level co-occurrence matrix. The texture features include contrast, correlation and energy.

[0207] Text data: Split the text into independent words, remove irrelevant prepositions and articles, calculate the TF-IDF weight for each word, and arrange the TF-IDF value of each word into a vector as the TF-IDF vector.

[0208] Time series data: Calculate the frequency of occurrence of each behavior in the time series data;

[0209] S302: Feature selection:

[0210] Relevance: Select features that are highly relevant to the behavior analysis target.

[0211] Validity: Ensure that the selected features can truly reflect the behavioral characteristics and avoid introducing noise.

[0212] Redundancy: Avoid selecting highly correlated features to reduce computational complexity and overfitting risk.

[0213] It should be noted that, in this embodiment, in order to maximize the correlation of feature selection, the specific implementation steps of feature selection are as follows:

[0214] 1. Calculate feature correlation:

[0215] Description: Through correlation analysis method, quantify the degree of correlation between each feature and the behavior analysis target.

[0216] Implementation method:

[0217] Correlation coefficient: For numerical features (such as behavior frequency of time series, image texture features, etc.), use methods such as Pearson correlation coefficient and Spearman correlation coefficient to calculate the correlation between the feature and the target variable.

[0218] Information Gain: For categorical or text data features (such as TF-IDF vectors), calculate information gain or mutual information to evaluate the contribution of each feature to classification.

[0219] Purpose: To select highly relevant features to ensure the efficiency and accuracy of the analysis model.

[0220] 2. Evaluate the effectiveness of features:

[0221] Description: Through validity analysis, ensure that the selected features can truly reflect behavioral characteristics.

[0222] Implementation method:

[0223] Statistical tests: Use statistical methods such as t-test and chi-square test to exclude features that have no actual contribution to the target prediction.

[0224] Feature importance ranking: If you use a tree-based model (such as decision trees or random forests), use the feature importance scores provided by the model to remove features with low importance.

[0225] Purpose: Ensure that each feature has a practical effect and avoid noise features interfering with the model.

[0226] 3. Feature redundancy detection:

[0227] Description: Detect highly correlated features and avoid redundancy to reduce computational complexity.

[0228] Implementation method:

[0229] Collinearity analysis: Use the variance inflation factor (VIF) or collinearity matrix to identify highly correlated feature pairs and select one of them to remove.

[0230] Clustering dimensionality reduction: Clustering or principal component analysis (PCA) is performed on highly correlated features to merge multiple related features into a new feature to simplify the feature space.

[0231] Purpose: Eliminate redundant information, reduce computational complexity, and avoid model overfitting.

[0232] 4. Construction of the final feature set:

[0233] Description: Construct the final feature set based on the relevance, effectiveness, and redundancy evaluation results.

[0234] Implementation method:

[0235] Retain features with high relevance, high effectiveness, and low redundancy by priority. Pass the feature set to the model loading module to ensure that the model uses a concise and efficient feature set.

[0236] Purpose: To provide an optimized feature set to support efficient analysis of subsequent models.

[0237] Through the above steps, feature selection is screened and optimized in terms of feature set relevance, effectiveness, and redundancy, ensuring that the analysis model can efficiently and accurately analyze and identify behavioral patterns.

[0238] In one embodiment of the present application, the model loading module implements the following steps:

[0239] 401: Determining model type:

[0240] Choose the appropriate model type based on data characteristics and business requirements:

[0241] For labeled classification or regression tasks, choose supervised learning algorithms such as decision trees, random forests, and / or SVMs.

[0242] For grouping and anomaly identification of unlabeled data, use unsupervised learning algorithms for cluster analysis and / or anomaly detection,

[0243] For processing image, video, text and time series data, select CNN, RNN and / or LSTM deep learning algorithms, combined with spatiotemporal modeling to capture the temporal and spatial variation characteristics of the data through optical flow, trajectory modeling, and spatiotemporal graph network methods;

[0244] S402: Loading pre-trained model:

[0245] If a pre-trained model exists, load the model weights and parameters from storage, check the consistency of the model structure and parameters, and ensure that the pre-trained model meets the current task requirements.

[0246] If the model has been trained in a related scenario, apply transfer learning to migrate the model from the original task to the new task to cope with the situation of insufficient data, thereby improving training efficiency;

[0247] S403: Model initialization:

[0248] In the absence of a pre-trained model, define the model architecture and hyperparameters, such as the number of layers, number of neurons, activation function, learning rate, etc., and use random initialization and / or transfer learning to set the model weights to prepare for model training or inference;

[0249] S404: Configure training and evaluation parameters:

[0250] Set the batch size, learning rate, loss function and evaluation index parameters required for training. The evaluation index parameters include accuracy and recall rate.

[0251] S405: Small scale test:

[0252] Perform a small sample test on the loaded model to ensure that the model is loaded correctly and can run normally, ready for the subsequent analysis module.

[0253] It should be noted that in this embodiment, which is applicable to the case where the amount of data is insufficient, the specific implementation steps of the pre-training model to reduce the training time and improve the effect are as follows:

[0254] 1. Load model file: Load the model's structure file (such as .json, .yaml) and weight file (such as .h5, .pth) from the storage location.

[0255] 2. Check model integrity and consistency: Check the integrity of the model file to ensure that both the structure file and the weight file are loaded successfully and compatible with each other.

[0256] 3. Confirm that the dimensions of the model input and output layers match the feature dimensions of the current data to avoid input mismatch errors.

[0257] 4. Verify the model structure and parameters: Check the number of model layers, the number of neurons in each layer, the activation function, and the hyperparameter configuration to ensure that they meet the current task requirements.

[0258] 5. Check the output layer or the last few layers to confirm that the model meets the goals of the current task, such as the number of classifications or output format.

[0259] 6. Transfer learning application (if applicable): Freeze the bottom-level weights of the pre-trained model and unfreeze only the weights of the top or specific layers to allow these layers to be updated and fine-tuned on the new task.

[0260] 7. Adjust the configuration of the output layer according to task requirements (such as adding new output nodes) to adapt the model to new classification or regression tasks.

[0261] 8. Load the model into memory: Load the model into memory and perform an empty data run or a small sample test to verify whether the model is loaded successfully and can run normally.

[0262] It should be noted that, in this embodiment, in order to lay a foundation for model training so that it can effectively start learning data features, module initialization is achieved through the following specific steps:

[0263] 1. Define the model architecture:

[0264] Define the model architecture based on the data type (e.g., image, time series, text) and task type (e.g., classification, regression). If you choose CNN, specify the number of convolutional layers, convolution kernel size, pooling layers, etc. If you choose RNN / LSTM, define the sequence length, hidden layer size, etc.

[0265] 2. Add a spatiotemporal modeling module such as optical flow, trajectory modeling, or spatiotemporal graph network so that the model can capture the temporal and spatial feature changes in the data.

[0266] 3. Set hyperparameters:

[0267] Configure key training hyperparameters such as learning rate, batch size, regularization coefficient, and optimizer type (such as Adam or SGD).

[0268] 4. Set the activation function (such as ReLU, Sigmoid) and loss function (such as cross entropy loss, mean square error) according to the characteristics of the task.

[0269] The model loading module of the present invention plays a core role in the behavior analysis model system, and its advantages are reflected in flexibility, efficiency and reliability. First, the module can automatically select the appropriate model type according to business needs and data characteristics, and support a variety of model architectures including supervised learning, unsupervised learning and deep learning, so as to adapt to various data types such as images, texts, and time series. For models that have been trained on similar tasks, the module supports transfer learning, which greatly improves the model's adaptability to new tasks by freezing the underlying feature layer and fine-tuning the high layer, and is particularly suitable for application scenarios with limited data volume. The module also integrates spatiotemporal modeling technologies, such as optical flow, trajectory modeling, and spatiotemporal graph networks, which can capture the temporal and spatial characteristics of behavioral data and provide more accurate support for complex behavioral pattern recognition. In addition, the module designs an automated initialization and weight setting process, and through a variety of weight initialization methods (such as Xavier and He initialization), ensures that the model starts training from the appropriate state and improves the convergence speed.

[0270] In one embodiment of the present application, the algorithm analysis module implements the following steps:

[0271] S501: Data input and preprocessing:

[0272] Input the features extracted by the feature engineering module into the model, ensuring that the format and dimension of the input data are consistent with the input requirements of the model.

[0273] S502: Application of attention mechanism:

[0274] In video and time series data processing, the attention mechanism is applied to focus the model's attention on specific key areas or time periods, automatically identify and highlight key information related to the behavior, such as specific frames in the video or significant events in the time series, and filter out noise information to improve the overall accuracy of the data.

[0275] S503: Model Reasoning and Fusion:

[0276] The model selected by the model loading module is used to infer the input feature data to generate preliminary prediction results, wherein the preliminary prediction results include: classification labels for classifying behaviors into normal behaviors, abnormal behaviors, and fraudulent behaviors, and assigning a probability score and / or probability value to the behaviors indicating that the behaviors belong to a certain category,

[0277] Model fusion and integration are used to fuse the prediction results of multiple models through simple weighted averaging, voting, or more complex stacking generalization, bagging and boosting to obtain more reliable fusion prediction results, where the fusion prediction results include: classification labels for behaviors into normal behaviors, abnormal behaviors, and fraudulent behaviors, and assigning a probability score and / or probability value to the behavior indicating that the behavior belongs to a certain category.

[0278] S504: Abnormal detection and early warning:

[0279] For the scenario of abnormal behavior identification, statistical methods, distance-based, density-based methods or deep learning methods of anomaly detection technology are applied to identify abnormal events that significantly deviate from normal behavior patterns from the model output.

[0280] Set the threshold range for anomaly detection, and trigger an alert when the score or probability value in the fusion prediction result exceeds a specific threshold.

[0281] S505: Generate analysis results:

[0282] Based on the output of the model, the analysis results are converted into easy-to-understand classification labels, scores / probability values ​​or anomaly marks, and output as structured reports to provide a basis for decision-making and subsequent actions.

[0283] S506: Real-time processing and optimization:

[0284] Monitor the model reasoning process and data flow in real time, use caching, parallel processing, and memory optimization to ensure fast response in high-load or real-time scenarios, and adjust model parameters or memory allocation to ensure the efficiency and stability of the analysis process.

[0285] It should be noted that, in this embodiment, in order to effectively identify abnormal behavior, the statistical method of anomaly detection technology, the distance-based method, the density-based method or the deep learning method are applied to identify abnormal events that significantly deviate from the normal behavior pattern from the model output. The specific implementation steps are as follows:

[0286] 1. Data preparation:

[0287] Extract scores, probability values ​​or other key features from the fusion prediction results as input data for anomaly detection.

[0288] Normalize the data to ensure that the scales of different features are consistent and avoid differences between features affecting the detection effect.

[0289] 2. Choose anomaly detection method:

[0290] Choose the appropriate anomaly detection method based on the type of data and the complexity of the abnormal behavior:

[0291] Statistical method: Applicable to scenarios with clear feature distribution, and the abnormal range is set through parameters such as mean and standard deviation.

[0292] Distance-based methods: Use algorithms such as k-nearest neighbor (k-NN) or isolation forest to calculate the distance or similarity between samples and identify outliers that are significantly different from other samples.

[0293] Density-based methods: Applicable to nonlinearly distributed data, using algorithms such as local outlier factor (LOF) or DBSCAN to detect sample points where local density is significantly lower than other areas.

[0294] Deep learning methods: such as autoencoders or GAN models, are suitable for complex data and can identify abnormal data by reconstructing errors or generating adversarial network scores.

[0295] 3. Set anomaly detection threshold

[0296] Depending on the method selected, set the detection threshold:

[0297] Statistical method: Set the value of the score or probability value exceeding a certain standard deviation as the abnormal threshold, such as 3 times the standard deviation from the mean.

[0298] Distance or density based methods: Set a distance threshold, or use the minimum density point in the density distribution as the anomaly threshold.

[0299] Deep learning method: Set a reconstruction error threshold, and samples exceeding this value are marked as abnormal samples.

[0300] 4. Anomaly score calculation:

[0301] Calculate an anomaly score or outlier metric for each sample based on the chosen method, for example:

[0302] Statistical methods calculate the degree of deviation from the mean;

[0303] Calculate the average distance between the sample and its k nearest neighbors based on the distance method;

[0304] The density method calculates the local density of the sample;

[0305] Deep learning methods use the reconstruction error or the discriminator score as the anomaly score.

[0306] 5. Threshold determination and abnormal marking:

[0307] Compare the anomaly score of each sample with the set anomaly threshold:

[0308] If the score exceeds the threshold, the sample is marked as abnormal behavior.

[0309] The labeling results are passed to the early warning module to trigger an early warning when an anomaly is detected.

[0310] 6. Abnormal feedback and adjustment:

[0311] Regularly evaluate anomaly detection results, analyze false positives or missed negatives, optimize detection parameters or method selection, and ensure the accuracy and stability of anomaly detection.

[0312] Through the above steps, anomaly detection technology can accurately identify anomalies in fusion prediction results based on statistics, distance, density or deep learning methods, and identify abnormal events that significantly deviate from normal behavior patterns from the model output.

[0313] In one embodiment of the present application, the feedback module implements the following steps:

[0314] S601: Execute security measures and warning notifications: If abnormal behavior is detected, the system sends a warning notification to the user or triggers preventive security measures.

[0315] S602: User feedback collection:

[0316] Collect feedback on the research and judgment results from users or relevant personnel, including:

[0317] Questionnaire survey: Design a questionnaire to understand users’ satisfaction and suggestions on the research results.

[0318] User complaint records: record abnormal situations or complaints reported by users to provide reference data for subsequent analysis.

[0319] S603: Model performance evaluation:

[0320] Regularly evaluate the performance of the model and check its accuracy, recall, and F1 score in the current data environment. Accuracy is the proportion of correctly classified results, recall is the proportion of correctly identified anomalies in actual data, and F1 score is a comprehensive score used to balance accuracy and recall.

[0321] S604: Feedback data storage and management:

[0322] All user feedback, model performance evaluation results and judgment data are stored in the feedback database to facilitate subsequent analysis and improvement.

[0323] Categorize and store user feedback, performance indicators, and abnormal records, set validity periods for data in the feedback database, and regularly update and clean up expired data.

[0324] S605: Model iteration and optimization:

[0325] Model parameter tuning: Adjust the model’s learning rate and / or regularization parameters to adapt to the new data distribution,

[0326] Data update: Add new feedback data to the model training data set to improve the model's adaptability to the current data.

[0327] Retraining or fine-tuning: After collecting enough new data, retrain or fine-tune the model to improve the model's prediction accuracy for new scenarios.

[0328] It should be noted that in this embodiment, in order to improve the accuracy of the behavior algorithm of the invention, the specific implementation steps of retraining or fine-tuning are as follows:

[0329] 1. Data preparation:

[0330] Organize the newly collected feedback data and user behavior data into the training data set to ensure that the data includes the latest abnormal behavior feedback, user feedback, performance evaluation results, etc.

[0331] Standardize or normalize new data to ensure data consistency and avoid interference of data scale differences on the model.

[0332] 2. Data segmentation:

[0333] The updated dataset is divided into training, validation, and test sets to ensure that the model can be validated during the training process and the performance can be evaluated in the test set.

[0334] Make sure your dataset contains enough normal and abnormal samples so that the model can better learn the classification boundaries and features.

[0335] 3. Hyperparameter adjustment:

[0336] According to the characteristics of the new data, the learning rate and regularization parameters of the model are appropriately adjusted to control the update speed of the model and avoid overfitting.

[0337] For deep learning models, you can adjust parameters such as batch size or gradient clipping to improve training stability.

[0338] 4. Fine-tune the model:

[0339] Freeze the underlying feature extraction layer of the model, maintain the original feature extraction capability, and only update the weights of the high-level or classification layer to reduce training time and improve the model's adaptability to new tasks.

[0340] Fine-tune the new data for a specified number of epochs and monitor the performance of the validation set (such as accuracy, recall, and F1 score) to ensure that the model gradually improves during the fine-tuning process.

[0341] 5. Retraining (if applicable):

[0342] If the amount of new data is large or the task scenario changes significantly, unfreeze all layers and retrain the model to fully update the weights.

[0343] Use appropriate early stopping mechanisms or model checkpoints to avoid overtraining and improve the generalization ability of the model.

[0344] 6. Evaluation and testing:

[0345] Use the test set to fully evaluate the fine-tuned or retrained model, calculate the accuracy, recall, and F1 score, and verify whether the model is adapted to the new data characteristics.

[0346] If the evaluation results meet expectations, deploy the new model to the actual environment; if the results are not good, adjust the hyperparameters or model architecture and retrain.

[0347] 7. Model deployment and monitoring:

[0348] Deploy the updated model to the system, replacing the old version, to ensure its stability and accuracy in the production environment.

[0349] Regularly monitor the real-time performance of the model and record its prediction results in actual scenarios for further optimization and improvement.

[0350] Through the above steps, the model can be fine-tuned or retrained with new data to improve its accuracy in identifying abnormal behaviors and its ability to adapt to new scenarios, thereby improving the overall effectiveness and reliability of behavioral algorithm analysis.

[0351] The behavioral algorithm judgment model of the present invention automatically learns complex features from massive data through advanced technologies such as deep learning in terms of accuracy and efficiency, thereby improving the accuracy of behavioral analysis. Its multimodal data processing capabilities support the integration of multiple data sources such as video, image, and log data to achieve comprehensive behavioral feature capture. In addition, the model has a high degree of adaptability and robustness, can be continuously optimized according to new data, adapt to complex and changeable behavioral patterns, and effectively process noise data, thereby maintaining stable performance in different environments. Its real-time and dynamic nature enables the model to respond quickly to behavioral changes and meet the needs of scenarios such as autonomous driving and intelligent monitoring that have high requirements for real-time response. The behavioral algorithm judgment model technology also has good scalability and flexibility, can be customized according to specific needs and easily integrated into a variety of systems to expand application scenarios. By reducing human intervention, automated analysis reduces subjective errors and uncertainties, and further improves the reliability of judgment. Overall, the present invention is significantly superior to traditional methods in terms of accuracy, efficiency, adaptability, scalability, and real-time performance, laying a solid foundation for the accuracy and reliability of early warning monitoring.

[0352] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and deform the above embodiments within the scope of the present application.

Claims

1. A behavioral algorithm analysis model system, characterized in that: The behavioral algorithm analysis model system includes data collection module, data processing module, feature engineering module, model loading module, algorithm analysis module and feedback module, among which: The data collection module is used to collect raw data related to behavior analysis and provide a data basis for the entire analysis process. The data processing module is used to pre-process the raw data. The feature engineering module is used to extract features useful for behavior analysis from preprocessed data. The model loading module is used to load a pre-trained behavior analysis model and / or initialize a new model based on the selected algorithm according to specific application scenarios, data characteristics and business requirements. The judgment module is used to load the feature input value model extracted by the feature engineering module into the model selected by the model loading module. The model processes the input features according to the knowledge it has learned and outputs the judgment results. The feedback module is used to take corresponding safety measures according to the analysis results, send warning notifications to users, etc. At the same time, it collects user feedback on the results for subsequent model optimization and improvement.

2. The behavior algorithm analysis model system as claimed in claim 1, characterized in that: The data collection module collects raw data related to behavior analysis including the following steps: S101: Obtaining the video captured by the camera from the monitoring system, S102: Obtain text, pictures and videos posted by users from social media platforms. S103: Obtain the user's behavior track in the log data from the system log and application log, Among them, the user's behavior trajectory refers to the activity records of logging in, browsing, clicking, submitting and / or purchasing generated by the user in the process of using the system or application, and logging in, browsing, clicking, submitting and purchasing are types of log behaviors. The original data includes videos, texts, pictures and behavior trajectories.

3. The behavior algorithm analysis model system as claimed in claim 2, wherein the data processing module implements the following steps: S201: Data cleaning: Video data: Remove blurry frames, dark or noisy frames, delete duplicate clips, and retain high-definition video content. Text data: remove redundant characters and marks, delete irrelevant content, clean up duplicate text records, and ensure that text data is clear and concise. Image data: Eliminate blurry or repeated images, remove interference noise in images, and ensure image clarity and data quality. Behavior trajectory data: Delete redundant or irrelevant records in system logs, remove abnormal operation records, and retain the user's actual behavior path data; S202: Data collation: Video data: Arrange video frames in chronological order to ensure the continuity of video content, and save them by scene or behavior type. Text data: Sort and categorize text by content type and time, ensuring that content is in the same format across different data sources. Image data: Classify by shooting time, location or content to ensure structured storage of image information. Behavior track data: Sort behavior track records by user ID, operation time, and operation content; S203: Missing value processing: Video data: Lost or missing frames are supplemented by interpolating adjacent frames, or adjacent frames are copied to fill the gaps to ensure video continuity. Text data: fill in missing fields. For example, irrelevant or missing content can be marked with null values ​​to ensure data integrity. Image data: Replace missing images with similar images or mark them as empty to ensure that the image data is continuous and complete. Behavior trajectory data: Use interpolation or default values ​​to fill in missing behavior points to ensure the continuity of behavior trajectory data; S204: Data conversion: Video data: Convert the video frame by frame into an image sequence to generate a grayscale image or edge map. Text data: Python-based NLP libraries such as NLTK, spaCy, Keras, and scikit-learn can be used to normalize text to the same length by padding, truncating, or using specific vectorization methods. Image data: Convert the image into a grayscale image or feature matrix. Behavior trajectory data: After numbering the behavior events in different logs, extract the user behavior trajectory according to the preset duration, replace each behavior with the corresponding number, and combine them into time series data in chronological order.

4. The behavior as claimed in claim 3, wherein the features include edge information, corner points, texture features of video image data, TF-IDF vectors of text data, and the frequency of occurrence of the behavior of time series data, The feature engineering module implements the following steps: S301: Feature extraction, Video and image data: Use edge detection algorithms such as Canny and Sobel to extract edge information in the image, use Harris corner detection and / or FAST method to identify corners in the image, and calculate the texture features of the image through the gray level co-occurrence matrix. The texture features include contrast, correlation and energy. Text data: Split the text into independent words, remove irrelevant prepositions and articles, calculate the TF-IDF weight for each word, and arrange the TF-IDF value of each word into a vector as the TF-IDF vector. Time series data: Calculate the frequency of occurrence of each behavior in the time series data; S302: Feature selection: Relevance: Select features that are highly relevant to the behavior analysis target. Validity: Ensure that the selected features can truly reflect the behavioral characteristics and avoid introducing noise. Redundancy: Avoid selecting highly correlated features to reduce computational complexity and overfitting risk.

5. The behavior algorithm analysis model system as claimed in claim 4, characterized in that: The model loading module implements the following steps: 401: Determining model type: Choose the appropriate model type based on data characteristics and business requirements: For labeled classification or regression tasks, choose supervised learning algorithms such as decision trees, random forests, and / or SVMs. For grouping and anomaly identification of unlabeled data, use unsupervised learning algorithms for cluster analysis and / or anomaly detection, For processing image, video, text and time series data, select CNN, RNN and / or LSTM deep learning algorithms, combined with spatiotemporal modeling to capture the temporal and spatial variation characteristics of the data through optical flow, trajectory modeling, and spatiotemporal graph network methods; S402: Loading pre-trained model: If a pre-trained model exists, load the model weights and parameters from storage, check the consistency of the model structure and parameters, and ensure that the pre-trained model meets the current task requirements. If the model has been trained in a related scenario, apply transfer learning to migrate the model from the original task to the new task to cope with the situation of insufficient data, thereby improving training efficiency; S403: Model initialization: In the absence of a pre-trained model, define the model architecture and hyperparameters, such as the number of layers, number of neurons, activation function, learning rate, etc., set the model weights using random initialization and / or transfer learning, and prepare for model training or inference. S404: Configure training and evaluation parameters: Set the batch size, learning rate, loss function and evaluation index parameters required for training. The evaluation index parameters include accuracy and recall rate. S405: Small scale test: Perform a small sample test on the loaded model to ensure that the model is loaded correctly and can run normally, ready for the subsequent analysis module.

6. The behavior algorithm analysis model system as claimed in claim 5, characterized in that: The algorithm analysis module implements the following steps: S501: Data input and preprocessing: Input the features extracted by the feature engineering module into the model, ensuring that the format and dimension of the input data are consistent with the input requirements of the model; S502: Application of attention mechanism: In video and time series data processing, the attention mechanism is applied to focus the model's attention on specific key areas or time periods, automatically identify and highlight key information related to behavior, and filter out noise information to improve the overall accuracy of the data; S503: Model Reasoning and Fusion: The model selected by the model loading module is used to infer the input feature data to generate preliminary prediction results, wherein the preliminary prediction results include: classification labels for classifying behaviors into normal behaviors, abnormal behaviors, and fraudulent behaviors, and assigning a probability score and / or probability value to the behaviors indicating that the behaviors belong to a certain category, Adopt model fusion and integration, through simple weighted average, voting, or more complex stacking generalization, bagging and boosting, to fuse the prediction results of multiple models to obtain more reliable fusion prediction results, where the fusion prediction results include: classification labels for classifying behaviors into normal behaviors, abnormal behaviors, and fraudulent behaviors, and assigning a possibility score and / or probability value for the behavior to indicate that the behavior belongs to a certain category; S504: Abnormal detection and early warning: For the scenario of abnormal behavior identification, statistical methods, distance-based, density-based methods or deep learning methods of anomaly detection technology are applied to identify abnormal events that significantly deviate from normal behavior patterns from the model output. Set the threshold range for anomaly detection, and trigger an alert when the score or probability value in the fusion prediction result exceeds a specific threshold; S505: Generate analysis results: Based on the output of the model, the research and judgment results are converted into easy-to-understand classification labels, scores / probability values ​​or abnormal marks, and output as structured reports to provide a basis for decision-making and subsequent actions; S506: Real-time processing and optimization: Monitor the model reasoning process and data flow in real time, use caching, parallel processing, and memory optimization to ensure fast response in high-load or real-time scenarios, and adjust model parameters or memory allocation to ensure the efficiency and stability of the analysis process.

7. The behavior algorithm analysis model system as claimed in claim 6, characterized in that: The feedback module implements the following steps: S601: Execute security measures and warning notifications: If abnormal behavior is detected, the system sends a warning notification to the user or triggers preventive security measures; S602: User feedback collection: Collect feedback on the research and judgment results from users or relevant personnel, including: Questionnaire survey: Design a questionnaire to understand users’ satisfaction and suggestions on the research results. User complaint records: record abnormal situations or complaints reported by users to provide reference data for subsequent analysis; S603: Model performance evaluation: Regularly evaluate the performance of the model and check its accuracy, recall, and F1 score in the current data environment. Accuracy is the proportion of correctly classified results, recall is the proportion of correctly identified anomalies in actual data, and F1 score is a comprehensive score used to balance accuracy and recall. S604: Feedback data storage and management: All user feedback, model performance evaluation results and judgment data are stored in the feedback database to facilitate subsequent analysis and improvement. Categorize and store user feedback, performance indicators, and abnormal records, set validity periods for data in the feedback database, and regularly update and clean up expired data; S605: Model iteration and optimization: Model parameter tuning: Adjust the model’s learning rate and / or regularization parameters to adapt to the new data distribution, Data update: Add new feedback data to the model training data set to improve the model's adaptability to the current data. Retraining or fine-tuning: After collecting enough new data, retrain or fine-tune the model to improve the model's prediction accuracy for new scenarios.

8. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the module according to any one of claims 1 to 7 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the module described in any one of claims 1 to 7 are implemented.

10. A processor, characterized in that: The processor is used to run a program, wherein the program executes the method of any module of claims 1 to 7 when running.

Images