Data processing intelligent analysis method based on AI algorithm

Abstract

The invention discloses a data processing intelligent analysis method based on an AI algorithm, and relates to the technical field of data processing, and the analysis method comprises the specific steps: S100, data collection and preprocessing: obtaining image data through a monitoring camera, an industrial camera and a satellite camera; according to the method, the relation between the feature vectors is represented by building the probability graph model, feature importance evaluation is introduced, the causal relation between variables can be mined, powerful support is provided for decision making, meanwhile, the structural causal model is adopted to evaluate the strength and stability of the causal relation, and the decision making efficiency is improved. By comparing data distribution changes before and after intervention and calculation results on different data subsets, the strength and stability of the causal relationship can be accurately measured, a basis is provided for data analysis, the depth and breadth of data analysis are improved, and a new thought and method are provided for research and application in related fields.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and specifically to a data processing intelligent analysis method based on an AI algorithm. Background Art

[0002] With the rapid development of information technology, data has become an important resource in modern society. In various fields, such as surveillance security, industrial manufacturing, social media analysis, and speech recognition, a large amount of image, text, and voice data is generated every day. These data contain rich information and potential commercial value, but how to efficiently process and analyze these data and mine the useful information has become an urgent problem to be solved. In recent years, the continuous advancement of artificial intelligence technology has provided new ideas and methods for data processing, especially the data processing intelligent analysis method based on AI algorithms, which has attracted much attention due to its powerful data processing and analysis capabilities.

[0003] Although traditional data processing technology can meet certain basic needs to a certain extent, it still has many shortcomings when processing multimodal data. First, traditional methods can usually only process data of a single modality, such as processing images, text or voice separately, and it is difficult to comprehensively analyze and utilize these different modal data, which limits the depth and breadth of data processing and leads to the loss of a large amount of useful information. Secondly, traditional methods have limitations in feature extraction and can often only extract relatively shallow features, but cannot deeply explore the hidden information and correlation relationships in the data. In addition, traditional methods also have shortcomings in causal relationship mining and evaluation, and it is difficult to accurately judge the causal relationship between different variables and its strength and stability.

[0004] Therefore, developing an intelligent analysis method for data processing based on AI algorithm not only improves the accuracy and efficiency of data processing, but also provides strong support for decision-making in various fields and has broad application prospects. Summary of the invention

[0005] The purpose of the present invention is to make up for the shortcomings of the prior art and to provide a data processing intelligent analysis method based on AI algorithm. The method can efficiently process and analyze multimodal data, including images, text and voice. Through data acquisition and preprocessing, multimodal feature extraction, construction of multivariate causal relationship model, causal relationship strength assessment, model training and optimization steps, it can deeply explore the causal relationship between variables and present the causal connection relationship between variables in a visual form.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: On the one hand, a data processing intelligent analysis method based on AI algorithm, the specific steps of the analysis method are: S100, data acquisition and preprocessing: Acquire image data through surveillance cameras, industrial cameras and satellite cameras; collect text data from enterprise relational databases, government databases, Weibo, news websites and professional forums; collect voice data with the help of handheld microphones, array microphones and telephone microphones, and simultaneously preprocess and standardize the collected data; S200, multimodal feature extraction: extract features using algorithms based on different modal characteristics. Image modal features are extracted using a method based on local binary patterns (LBP). Text modal features are extracted using a method based on the bag-of-words model (BoW) and TF-IDF. Speech modal features are extracted using a method based on linear predictive coding (LPC). Multimodal feature fusion is then performed using a fusion formula. S300, constructing a multivariate causal relationship model: constructing a multivariate causal relationship model based on the probabilistic graphical model and feature importance evaluation. Suppose the feature vector set after feature extraction is , by constructing a probabilistic graphical model to represent the relationship between feature vectors, for each pair of feature vectors and , calculate their joint probability distribution , the calculation formula is: , where N is the total number of data samples, To appear simultaneously and The number of times, according to the joint probability distribution, construct an undirected graph , where the node set V corresponds to the feature vector F, and the edge set E is determined by: , is a pre-set probability threshold, then at the node and Add an edge between them to indicate that there is a strong correlation between them and calculate each eigenvector Information gain Measure the importance of each feature vector in the causal relationship. Set a target variable Y and calculate the entropy of the data set D. , suppose Y has m values, and the probability of each value is , ( ), the calculation formula is: , the data set D is divided into feature vectors The value of is divided into s subsets ( ) Calculate the conditional entropy after partitioning, the formula is: ,in and The subsets are And the number of samples in the data set D, the formula for calculating information gain is: , determine the weight of the feature vector according to the information gain , the formula is: ,Based on the constructed probabilistic graph model, the causal relationship is represented by analyzing the causal direction indicators of the paths and nodes in the graph; S400, Causal strength assessment: Based on the structural causal model, the causal strength is measured by observing the distribution changes of another characteristic variable when one characteristic variable is intervened and not intervened. Simultaneously, the causal strength and direction indicators are calculated on multiple data subsets by random sampling, and the stability of the causal relationship is evaluated based on the consistency of the indicators among different subsets. S500, model training and optimization: The model is adjusted based on probability distribution adaptive adjustment, and the model parameters are set as , assuming that the causal relationship predicted by the previous model for the training data is , the actual causal relationship is y, and the loss function is , for each parameter Set the initial value range , for each parameter , randomly generate two adjacent values ​​within its value range and , respectively calculate the corresponding loss function value ; ,in and Use and The prediction results of the model are used to calculate the parameters The sensitivity of is: , adjust the parameter range and update the parameter value according to the sensitivity , the formula is: ; , k is the coefficient for controlling the shrinkage degree, and a new The above steps of calculating sensitivity, adjusting the value range and updating the parameter value are repeated iteratively until the preset maximum number of iterations is reached. S600, result output and application: Output multimodal data causal relationship map, which presents the causal connection relationship between variables in a visual form. Different types of data variables are represented by nodes of different shapes, colors and sizes, and causal relationships are represented by lines. The thickness and color of the lines correspond to the quantitative indicators of the strength, direction and stability of the causal relationship. At the same time, text annotations are added to the map to display detailed information on the variable name and the specific numerical value of the causal relationship.

[0007] Furthermore, in the multimodal feature extraction in S200, a method based on local binary pattern LBP is used to extract image features. , sample size , divide the image into several small areas, and calculate the LBP value of each small area. Suppose the size of the small area is n×n, and the coordinates of the central pixel are , the calculation formula is: ,in is the gray value of the center pixel, is the gray value of the neighborhood pixel, , and convert these binary values ​​into decimal numbers as the LBP value of the point, count the histogram of the LBP value in each small area, connect these histograms to form the feature vector of the image, suppose the image is divided into k small areas, and the LBP value histogram of each small area is ( ), then the feature vector of the image for: , whose dimension is ,in is the length of the eigenvector.

[0008] Furthermore, in the multimodal feature extraction in S200, a method based on the bag-of-words model BoW and TF-IDF is used to extract text features. , Sample size , construct a vocabulary, count the number of times each word appears in each text, and get the word frequency matrix. Let the vocabulary size be v and the number of texts be , then the word frequency matrix The size is ,in Represents the number of occurrences of the i-th word in the j-th text, and calculates the inverse document frequency IDF. The formula is: ,in Represents the number of texts containing the i-th word and calculates the TF-IDF value. The formula is: , the feature vector of the text It is the TF-IDF vector corresponding to each text, and its dimension is ,in .

[0009] Furthermore, in the multimodal feature extraction in S200, a method based on linear predictive coding (LPC) is used to extract speech features. , sample size , the speech signal is divided into frames, each frame length is N, for each frame of speech signal , ( , solve the linear prediction coefficient by minimizing the sum of squares of the prediction error , ( ,in is the prediction order, and the formula is: ,minimize , by solving the linear equations to obtain the linear prediction coefficients , take the linear prediction coefficient As the feature of speech, let the feature vector extracted from each frame of speech be ,( ), then the feature matrix of speech for: , whose dimension is ,in .

[0010] Furthermore, in S200, a fusion formula is used to perform multimodal feature fusion in multimodal feature extraction, and the image modality feature vector is assumed to be , the dimension is , the text modality feature vector is , the dimension is , the speech modality feature vector is , the dimension is , the fused feature vector is F, the dimension is d, and the feature vector of each modality is normalized: the formula is: ,in yes The norm formula is: Introducing the importance coefficients of the three modes , , and , the fusion formula is: ,in Represents a custom combination operation, which is the element-by-element multiplication of vectors followed by summation.

[0011] Furthermore, the S300 constructs a probability threshold in a multivariate causal relationship model. It is based on statistical analysis to determine the threshold and calculate the feature vector The joint probability distribution of The mean and standard deviation , assuming that the joint probability distribution approximately obeys the normal distribution, the threshold is set according to the properties of the normal distribution. The formula is: , where x is , is the mean, is the standard deviation, the threshold Set to .

[0012] Furthermore, in the S300, in constructing a multivariate causal relationship model, causal relationships are mined by analyzing the causal direction indicators of the paths and nodes in the graph. and , considering the conditional probability and , define the causal direction index as , the calculation formula is: , , indicating that after considering the importance of features, yes Reasons for , yes Reasons for , the causal relationship is unclear.

[0013] Furthermore, in the causal relationship strength evaluation in S400, for the characteristic variable and The distribution changes under different conditions are used to evaluate the strength of causal relationship. The data are divided into two groups: The data subset in the normal range is denoted as The other group is The data subset whose value is changed by human intervention is denoted as , calculated in these two sets of data The mean and standard deviation ,exist middle, The mean value of , the standard deviation is ,exist middle, The mean value of The standard deviation is , define the causal strength index The calculation formula is: .

[0014] Furthermore, in the causal relationship strength assessment in S400, the causal relationship stability assessment adopts a random sampling method to assess the stability of the causal relationship, and performs N random samplings, each time extracting a 70% data subset from the original data, recorded as , for each sampled data subset, calculate the above causal strength index and causal direction indicators , define the causal stability index and , the formula is: ; ,in and are the average values ​​of the causal strength index and the causal direction index on all sampled data subsets, and is a pre-set threshold, is an indicator function, which equals 1 if the condition is true and 0 otherwise.

[0015] Compared with the prior art, this data processing intelligent analysis method based on AI algorithm has the following beneficial effects: 1. The present invention constructs a probabilistic graphical model to represent the relationship between feature vectors and introduces feature importance evaluation, which can explore the causal relationship between variables and provide strong support for decision-making. At the same time, the present invention also uses a structural causal model to evaluate the strength and stability of the causal relationship. By comparing the changes in data distribution before and after the intervention and the calculation results on different data subsets, the strength and stability of the causal relationship can be accurately measured, which provides a basis for data analysis, not only improves the depth and breadth of data analysis, but also provides new ideas and methods for research and application in related fields.

[0016] 2. The present invention effectively integrates the data features of multiple modes such as images, text and speech through data collection and preprocessing, and multimodal feature extraction steps, providing a more comprehensive and accurate data basis for subsequent analysis, which not only improves the efficiency of data processing, but also makes the analysis results closer to the actual situation. In the feature extraction stage, different algorithms are used for different types of modal data, such as image feature extraction based on local binary patterns LBP, text feature extraction based on bag-of-words model BoW and TF-IDF, and speech feature extraction based on linear predictive coding LPC. The selection and optimization of these algorithms further improve the precision and accuracy of data processing.

[0017] Other advantages, objectives and features of the present invention will be set forth in part in the following description and, in part, will be apparent to those skilled in the art based on an examination of the following or may be taught from the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0019] Figure 1 It is a flow chart of a data processing intelligent analysis method based on AI algorithm; Figure 2 This is a process framework diagram of a data processing intelligent analysis method based on AI algorithm. DETAILED DESCRIPTION

[0020] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation mode, structure, characteristics and effects of the present invention are described in detail below in combination with the accompanying drawings and preferred embodiments.

[0021] Embodiment 1

[0022] Intelligent Transportation System Analysis Data collection and preprocessing: Obtain image data of traffic intersections through surveillance cameras, including vehicles, pedestrians, and traffic signs; collect text data of traffic regulations and road construction information from the database of the traffic management department; use telephone microphones to collect voice communication data between traffic police and drivers; preprocess image data, such as adjusting resolution and removing noise; clean text data to remove irrelevant information; and perform noise reduction on voice data.

[0023] Multimodal feature extraction: For image data, the local binary pattern LBP method is used to extract image features. , sample size , divide the image into several small areas, and calculate the LBP value of each small area. Suppose the size of the small area is n×n, and the coordinates of the central pixel are , the calculation formula is: , , and convert these binary values ​​into decimal numbers as the LBP value of the point, count the histogram of the LBP value in each small area, connect these histograms to form the feature vector of the image, suppose the image is divided into k small areas, and the LBP value histogram of each small area is ( ), then the feature vector of the image for: , whose dimension is ,in is the length of the eigenvector;

[0024] For text data, the method based on the bag-of-words model BoW and TF-IDF is used to extract text features. , Sample size , construct a vocabulary, count the number of times each word appears in each text, and get the word frequency matrix. Let the vocabulary size be v and the number of texts be , then the word frequency matrix The size is ,in Represents the number of occurrences of the i-th word in the j-th text, and calculates the inverse document frequency IDF. The formula is: , calculate the TF-IDF value, the formula is: , the feature vector of the text It is the TF-IDF vector corresponding to each text, and its dimension is ,in ; For speech data, the method based on linear predictive coding (LPC) is used to extract speech features. , sample size , the speech signal is divided into frames, each frame length is N, for each frame of speech signal , ( , solve the linear prediction coefficient by minimizing the sum of squares of the prediction error , ( ,in is the prediction order, and the formula is: ,minimize , by solving the linear equations to obtain the linear prediction coefficients , take the linear prediction coefficient As the feature of speech, let the feature vector extracted from each frame of speech be ,( ), then the feature matrix of speech for: , whose dimension is ,in ; The fusion formula is used to fuse multimodal features. Let the image modality feature vector be , the dimension is , the text modality feature vector is , the dimension is , the speech modality feature vector is , the dimension is , the fused feature vector is F, the dimension is d, and the feature vector of each modality is normalized: the formula is: ,in yes The norm formula is: , introducing the importance coefficients of the three modes , , and , the fusion formula is: .

[0025] Constructing a multivariate causal relationship model, let the feature vector set after feature extraction be , construct a probabilistic graphical model to represent the relationship between feature vectors. For each pair of feature vectors and , calculate their joint probability distribution , the calculation formula is: , where N is the total number of data samples, To appear simultaneously and The number of times, analyze the relationship between vehicle speed (image feature) and traffic regulations (text feature), and construct an undirected graph based on the joint probability distribution. , where the node set V corresponds to the feature vector F, and the edge set E is determined by: , is a pre-set probability threshold, then at the node and Add an edge between them to indicate that there is a strong correlation between them and calculate each eigenvector Information gain Measure the importance of each feature vector in the causal relationship. Set a target variable Y and calculate the entropy of the data set D. , suppose Y has m values, and the probability of each value is , ( ), the calculation formula is: , the data set D is divided into feature vectors The value of is divided into s subsets ( ), calculate the conditional entropy after partitioning, the formula is: ,The formula for calculating information gain is: , determine the weight of the feature vector according to the information gain , the formula is: , based on the constructed probabilistic graph model, the causal relationship is represented by analyzing the causal direction indicators of the paths and nodes in the graph; and , considering the conditional probability and , define the causal direction index as , the calculation formula is: , determine whether traffic regulations affect vehicle speed or vehicle speed leads to the adjustment of traffic regulations.

[0026] Causal relationship strength assessment, for characteristic variables and The distribution changes under different conditions are used to evaluate the strength of causal relationship. The data are divided into two groups: The data subset in the normal range is denoted as The other group is The data subset whose value is changed by human intervention is denoted as , calculated in these two sets of data The mean and standard deviation ,exist middle, The mean value of , the standard deviation is ,exist middle, The mean value of The standard deviation is , define the causal strength index The calculation formula is: , a random sampling method is used to evaluate the stability of the causal relationship. N random samplings are performed, and each time 70% of the data subset is extracted from the original data, recorded as , for each sampled data subset, calculate the above causal strength index and causal direction indicators , define the causal stability index and , the formula is: ; .

[0027] Model training and optimization, using adaptive adjustment based on probability distribution to adjust the model, assuming the model parameters are , assuming that the causal relationship predicted by the previous model for the training data is , the actual causal relationship is y, and the loss function is , for each parameter Set the initial value range , for each parameter , randomly generate two adjacent values ​​within its value range and , respectively calculate the corresponding loss function value ; ,in and Use and The prediction results of the model are used to calculate the parameters The sensitivity of is: , adjust the parameter range and update the parameter value according to the sensitivity , the formula is: ; , k is the coefficient for controlling the shrinkage degree, and a new The above steps of calculating sensitivity, adjusting the value range and updating parameter values ​​are repeated continuously until the preset maximum number of iterations is reached.

[0028] Result output and application: Output a multimodal data causal relationship map. In the map, nodes of different shapes, colors and sizes are used to represent different types of data variables such as images, texts and voices. Lines are used to represent causal relationships. The thickness and color of the lines correspond to the quantitative indicators of the strength, direction and stability of the causal relationship, respectively. Text annotations are added to display the variable name and the specific numerical details of the causal relationship. This map can be used by traffic management departments to optimize traffic rules and management strategies.

[0029] In summary, this embodiment applies the invention to intelligent transportation system analysis. By collecting and preprocessing images of traffic intersections, traffic regulations texts, and voice communication data between traffic police and drivers, a multivariate causal relationship model is constructed using a specific multimodal feature extraction method. On the basis of this model, the strength and stability of the causal relationship are evaluated, and more accurate results are obtained through model training and optimization. The final output multimodal data causal relationship map provides an intuitive decision-making basis for traffic management departments, helps to optimize traffic rules and management strategies, and improve the operating efficiency and safety of the transportation system.

[0030] Embodiment 2

[0031] Market Trend Analysis Example Data collection and preprocessing: Image data. The image data of customer traffic in physical stores is obtained through surveillance cameras. For example, in a large shopping mall, surveillance cameras installed at the entrances of multiple stores can record the number of customers entering and leaving the store and their behaviors (such as the length of stay and the area in which they browse goods) in real time. The collected image data is preprocessed and standardized, for example, the resolution and color mode of the image are unified for subsequent processing. Text data: collect text data about industry dynamics from news websites, for example, collect news reports about the fashion industry, including new brand launches, changes in fashion trends, and designer dynamics. Collect consumer evaluation and discussion data on products from professional forums. For example, in fashion forums, consumers will discuss clothing styles, quality, and prices. Clean the text data, remove noise information (such as advertising links and irrelevant emoticons), and perform standardization, such as unified text encoding format; Voice data: Use telephone microphones to collect voice data between customer service and consumers. For example, the customer service center of a fashion brand records voice information about product consultation and complaints when communicating with customers. Perform frame segmentation and noise reduction preprocessing on the voice data to make it meet the requirements for subsequent feature extraction.

[0032] Multimodal feature extraction: For image data, the local binary pattern LBP method is used to extract image features. , sample size , divide the image into several small areas, and calculate the LBP value of each small area. Suppose the size of the small area is n×n, and the coordinates of the central pixel are , the calculation formula is: , , and convert these binary values ​​into decimal numbers as the LBP value of the point, count the histogram of the LBP value in each small area, connect these histograms to form the feature vector of the image, suppose the image is divided into k small areas, and the LBP value histogram of each small area is ( ), then the feature vector of the image for: , whose dimension is ,in is the length of the eigenvector; For text data, the method based on the bag-of-words model BoW and TF-IDF is used to extract text features. , Sample size , construct a vocabulary, count the number of times each word appears in each text, and get the word frequency matrix. Let the vocabulary size be v and the number of texts be , then the word frequency matrix The size is ,in Represents the number of occurrences of the i-th word in the j-th text, and calculates the inverse document frequency IDF. The formula is: , calculate the TF-IDF value, the formula is: , the feature vector of the text It is the TF-IDF vector corresponding to each text, and its dimension is ,in ; For speech data, the method based on linear predictive coding (LPC) is used to extract speech features. , sample size , the speech signal is divided into frames, each frame length is N, for each frame of speech signal , ( , solve the linear prediction coefficient by minimizing the sum of squares of the prediction error , ( ,in is the prediction order, and the formula is: ,minimize , by solving the linear equations to obtain the linear prediction coefficients , take the linear prediction coefficient As the feature of speech, let the feature vector extracted from each frame of speech be ,( ), then the feature matrix of speech for: , whose dimension is ,in ; The fusion formula is used to fuse multimodal features. Let the image modality feature vector be , the dimension is , the text modality feature vector is , the dimension is , the speech modality feature vector is , the dimension is , the fused feature vector is F, the dimension is d, and the feature vector of each modality is normalized: the formula is: ,in yes The norm formula is: , introducing the importance coefficients of the three modes , , and , the fusion formula is: .

[0033] Construct a multivariate causal relationship model and a probability graph model. Suppose the feature vector set after feature extraction is , constructing a probabilistic graphical model to represent the relationship between feature vectors, such as calculating the joint probability distribution of the store customer flow image feature vector and the news report text feature vector , the calculation formula is: , where N is the total number of data samples, To appear simultaneously and The number of times, according to the joint probability distribution, construct an undirected graph , where the node set V corresponds to the feature vector F, and the edge set E is determined by: , is a pre-set probability threshold, then at the node and Add an edge between them to indicate that there is a strong correlation between them and calculate each eigenvector Information gain Measure the importance of each feature in the causal relationship. Set a target variable Y and calculate the entropy of the data set D. , suppose Y has m values, and the probability of each value is , ( ), the calculation formula is: , the data set D is divided into feature vectors The value of is divided into s subsets ( ) Calculate the conditional entropy after partitioning, the formula is: ,The formula for calculating information gain is: , determine the weight of the feature vector according to the information gain , the formula is: , based on the constructed probabilistic graph model, the causal relationship is represented by analyzing the causal direction indicators of the paths and nodes in the graph; and , considering the conditional probability and , define the causal direction index as , the calculation formula is: ,If CSI is greater than 1, it indicates that after considering the feature importance, the news report text feature vector is the cause of the store customer flow image feature vector; ,if CSI is less than 1, the opposite is true; ,if CSI is equal to 1, the causal relationship is unclear.

[0034] Causal relationship strength assessment, considering the store customer flow image feature vector and the news report text feature vector, the data is divided into two groups, one of which is a data subset where the store customer flow image feature vector is in the normal range , denoted as The other group is a subset of data that has been artificially intervened (for example, by changing store promotion strategies to change customer traffic). , denoted as , respectively calculate the mean and standard deviation of the news report text feature vectors in these two sets of data, and then define the causal relationship strength index CX to evaluate the causal relationship strength between the two. middle, The mean value of , the standard deviation is ;exist middle, The mean value of The standard deviation is , define the causal strength index The calculation formula is: , a random sampling method is used to evaluate the stability of the causal relationship. N random samplings are performed, and each time 70% of the data subset is extracted from the original data, recorded as , for each sampled data subset, calculate the above causal strength index and causal direction indicators , define the causal stability index and , the formula is: ; .

[0035] The model is adjusted based on the adaptive adjustment of probability distribution, and the model parameters are set as , assuming that the causal relationship predicted by the previous model for the training data is , the actual causal relationship is y, and the loss function is , for each parameter Set the initial value range , for each parameter , randomly generate two adjacent values ​​within its value range and , respectively calculate the corresponding loss function value ; ,in and Use and The prediction results of the model are used to calculate the parameters The sensitivity of is: , adjust the parameter range and update the parameter value according to the sensitivity , the formula is: ; , k is the coefficient for controlling the shrinkage degree, and a new The above steps of calculating sensitivity, adjusting the value range and updating parameter values ​​are repeated continuously until the preset maximum number of iterations is reached.

[0036] Result output and application: Output multimodal data causal relationship map, which presents the causal connection relationship between variables in a visual form. For example, circular nodes represent store customer flow image feature vectors, square nodes represent news report text feature vectors, and triangular nodes represent voice feature vectors. Lines represent causal relationships, and the thickness and color of the lines correspond to the quantitative indicators of the strength, direction, and stability of the causal relationship. At the same time, text annotations are added to the map to display the variable name, specific values ​​of the causal relationship, and detailed information. According to the causal relationship map, market trends are analyzed. For example, if the news report text feature vector has a strong causal influence on the store customer flow image feature vector, and the direction is positive, then it can be inferred that the relevant news reports have a promoting effect on the store's customer flow. Enterprises can adjust their marketing strategies based on these analysis results, such as strengthening cooperation with the media and publishing more attractive news content to increase store traffic.

[0037] In summary, this embodiment applies the invention in the market trend analysis scenario, collects and preprocesses data from multiple sources, extracts multimodal features, builds models, evaluates causal relationship strength and stability, trains and optimizes the model, and finally outputs a causal relationship map. This process can analyze market trends and provide a basis for companies to adjust their marketing strategies, demonstrating the effectiveness of the invention in actual scenarios.

[0038] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technical personnel in this field can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A data processing intelligent analysis method based on AI algorithm, characterized in that: The specific steps of this analysis method are: S100, data acquisition and preprocessing: Acquire image data through surveillance cameras, industrial cameras and satellite cameras; collect text data from enterprise relational databases, government databases, Weibo, news websites and professional forums; collect voice data with the help of handheld microphones, array microphones and telephone microphones, and simultaneously preprocess and standardize the collected data; S200, multimodal feature extraction: extract features using algorithms based on different modal characteristics. Image modal features are extracted using a method based on local binary patterns (LBP). Text modal features are extracted using a method based on the bag-of-words model (BoW) and TF-IDF. Speech modal features are extracted using a method based on linear predictive coding (LPC). Multimodal feature fusion is then performed using a fusion formula. S300, constructing a multivariate causal relationship model: constructing a multivariate causal relationship model based on the probabilistic graphical model and feature importance evaluation. Suppose the feature vector set after feature extraction is , by constructing a probabilistic graphical model to represent the relationship between feature vectors, for each pair of feature vectors and , calculate their joint probability distribution , the calculation formula is: , where N is the total number of data samples, To appear simultaneously and The number of times, according to the joint probability distribution, construct an undirected graph , where the node set V corresponds to the feature vector F, and the edge set E is determined by: , is a pre-set probability threshold, then at the node and Add an edge between them to indicate that there is a strong correlation between them and calculate each eigenvector Information gain Measure the importance of each feature vector in the causal relationship. Set a target variable Y and calculate the entropy of the data set D. , suppose Y has m values, and the probability of each value is , ( ), the calculation formula is: , the data set D is divided into feature vectors The value of is divided into s subsets ( ), calculate the conditional entropy after partitioning, the formula is: ,in and The subsets are And the number of samples in the data set D, the formula for calculating information gain is: , determine the weight of the feature vector according to the information gain , the formula is: ,Based on the constructed probabilistic graph model, the causal relationship is represented by analyzing the causal direction indicators of the paths and nodes in the graph; S400, Causal strength assessment: Based on the structural causal model, the causal strength is measured by observing the distribution changes of another characteristic variable when one characteristic variable is intervened and not intervened. Simultaneously, the causal strength and direction indicators are calculated on multiple data subsets by random sampling, and the stability of the causal relationship is evaluated based on the consistency of the indicators among different subsets. S500, model training and optimization: The model is adjusted based on probability distribution adaptive adjustment, and the model parameters are set as , assuming that the causal relationship predicted by the previous model for the training data is , the actual causal relationship is y, and the loss function is , for each parameter Set the initial value range , for each parameter , randomly generate two adjacent values ​​within its value range and , respectively calculate the corresponding loss function value ; ,in and Use and The prediction results of the model are used to calculate the parameters The sensitivity of is: , adjust the parameter range and update the parameter value according to the sensitivity , the formula is: ; , k is the coefficient for controlling the shrinkage degree, and a new The above steps of calculating sensitivity, adjusting the value range and updating the parameter value are repeated iteratively until the preset maximum number of iterations is reached. S600, result output and application: Output multimodal data causal relationship map, which presents the causal connection relationship between variables in a visual form. Different types of data variables are represented by nodes of different shapes, colors and sizes, and causal relationships are represented by lines. The thickness and color of the lines correspond to the quantitative indicators of the strength, direction and stability of the causal relationship. At the same time, text annotations are added to the map to display detailed information on the variable name and the specific numerical value of the causal relationship.

2. According to claim 1, a data processing intelligent analysis method based on AI algorithm is characterized in that: In the multimodal feature extraction, a method based on local binary pattern LBP is used to extract image features. , sample size , divide the image into several small areas, and calculate the LBP value of each small area. Suppose the size of the small area is n×n, and the coordinates of the central pixel are , the calculation formula is: ,in is the gray value of the center pixel, is the gray value of the neighborhood pixel, , and convert these binary values ​​into decimal numbers as the LBP value of the point, count the histogram of the LBP value in each small area, connect these histograms to form the feature vector of the image, suppose the image is divided into k small areas, and the LBP value histogram of each small area is ( ), then the feature vector of the image for: , whose dimension is ,in is the length of the eigenvector.

3. The data processing intelligent analysis method based on AI algorithm according to claim 1 is characterized in that: In S200, the multimodal feature extraction uses a method based on the bag-of-words model BoW and TF-IDF to extract text features. , Sample size , construct a vocabulary, count the number of times each word appears in each text, and get the word frequency matrix. Let the vocabulary size be v and the number of texts be , then the word frequency matrix The size is ,in Represents the number of occurrences of the i-th word in the j-th text, and calculates the inverse document frequency IDF. The formula is: ,in Represents the number of texts containing the i-th word and calculates the TF-IDF value. The formula is: , the feature vector of the text It is the TF-IDF vector corresponding to each text, and its dimension is ,in .

4. The data processing intelligent analysis method based on AI algorithm according to claim 1 is characterized in that: In the multimodal feature extraction, a method based on linear predictive coding (LPC) is used to extract speech features. , sample size , the speech signal is divided into frames, each frame length is N, for each frame of speech signal , ( , solve the linear prediction coefficient by minimizing the sum of squares of the prediction error , ( ,in is the prediction order, and the formula is: ,minimize , by solving the linear equations to obtain the linear prediction coefficients , take the linear prediction coefficient As the feature of speech, let the feature vector extracted from each frame of speech be ,( ), then the feature matrix of speech for: , whose dimension is ,in .

5. The data processing intelligent analysis method based on AI algorithm according to claim 1 is characterized in that: In the S200, a fusion formula is used to perform multimodal feature fusion in multimodal feature extraction. Assume that the image modal feature vector is , the dimension is , the text modality feature vector is , the dimension is , the speech modality feature vector is , the dimension is , the fused feature vector is F, the dimension is d, and the feature vector of each modality is normalized: the formula is: ,in yes The norm formula is: Introducing the "importance coefficients" of the three modes , , and , the fusion formula is: ,in Represents a custom combination operation, which is the element-by-element multiplication of vectors followed by summation.

6. The data processing intelligent analysis method based on AI algorithm according to claim 1 is characterized in that: S300, constructing a probability threshold in a multivariate causal relationship model It is based on statistical analysis to determine the threshold and calculate the feature vector The joint probability distribution of The mean and standard deviation , assuming that the joint probability distribution approximately obeys the normal distribution, the threshold is set according to the properties of the normal distribution. The formula is: , where x is , is the mean, is the standard deviation, the threshold Set to .

7. The data processing intelligent analysis method based on AI algorithm according to claim 1 is characterized in that: In the S300, in constructing a multivariate causal relationship model, causal relationships are mined by analyzing the causal direction indicators of the paths and nodes in the graph. and , considering the conditional probability and , define the causal direction index as , the calculation formula is: , , indicating that after considering the importance of features, yes Reasons for , yes Reasons for , the causal relationship is unclear.

8. The data processing intelligent analysis method based on AI algorithm according to claim 1 is characterized in that: In the causal relationship strength assessment in S400, causal relationship strength assessment, for the characteristic variable and The distribution changes under different conditions are used to evaluate the strength of causal relationship. The data are divided into two groups: The data subset in the normal range is denoted as The other group is The data subset whose value is changed by human intervention is denoted as , calculated in these two sets of data The mean and standard deviation ,exist middle, The mean value of , the standard deviation is ,exist middle, The mean value of The standard deviation is , define the causal strength index The calculation formula is: .

9. The data processing intelligent analysis method based on AI algorithm according to claim 1 is characterized in that: In S400, the causal relationship stability assessment in the causal relationship strength assessment uses a random sampling method to assess the stability of the causal relationship. N random samplings are performed, and each time a 70% data subset is extracted from the original data, which is recorded as , for each sampled data subset, calculate the above causal strength index and causal direction indicators , define the causal stability index and , the formula is: ; ,in and are the average values ​​of the causal strength index and the causal direction index on all sampled data subsets, and is a pre-set threshold, is an indicator function, which equals 1 if the condition is true and 0 otherwise.

Images