Multi-source heterogeneous sports field data real-time fusion method

Abstract

The invention relates to the field of computers, the field of data fusion and the field of machine learning, and provides a multi-source heterogeneous sports field data real-time fusion method. The method aims at solving the problem of real-time fusion of multi-source heterogeneous sports event data, and the main scheme comprises the following steps: data acquisition: acquiring text and picture information of related sports events on different social media platforms and news websites by utilizing a web crawler technology and calling an API (Application Program Interface); data preprocessing: carrying out standardization processing on the obtained original text, enhancing the contrast of the original image, and increasing the availability; feature extraction and learning: performing feature extraction and learning on the text and picture data processed in the step 2; fusing the extracted text and picture features by using a multi-modal LSTM model; and generating a word cloud picture from the fused result by using the word cloud library, and outputting the word cloud picture to a user for use.

Description

Technical Field

[0001] The invention relates to the fields of computer, data fusion and machine learning, and provides a method for real-time fusion of multi-source heterogeneous sports data. Background Art

[0002] In the field of sports, athlete performance analysis, game strategy formulation, and training effect evaluation all rely on the comprehensive analysis of a large amount of heterogeneous data. Traditional data fusion methods are often based on fixed rules, and have problems such as insufficient flexibility, poor real-time performance, and low accuracy. Therefore, developing a new real-time fusion method for multi-source heterogeneous sports data to improve data processing efficiency and accuracy has become an important direction of current research.

[0003] At present, most of the existing motion data fusion technologies are based on rule engines and traditional statistical methods, which usually rely on predefined rules to integrate data from different sources. The rule engine-based method requires manual definition of rules, resulting in a lack of flexibility and adaptability in the system, making it difficult to cope with complex and changeable motion data characteristics. In addition, traditional statistical methods often rely on prior knowledge and assumptions, and have limited ability to process unknown data relationships, making it difficult to meet the needs of complex motion data analysis. These methods also have shortcomings in mining the correlation between text and images in motion data, resulting in the results being not intuitive enough and the amount of information being limited.

[0004] In view of the above problems, a real-time fusion method of multi-source heterogeneous data based on deep learning is provided to solve the problems existing in the existing technical solutions. By introducing a deep learning model, adaptive learning of multi-source heterogeneous data is achieved, thereby improving the flexibility and adaptability of the system. Secondly, through feature learning of the machine learning model, the dependence on prior knowledge and assumptions is reduced, which can improve the accuracy and universality of data fusion. In addition, a word cloud diagram is constructed by mining the correlation between text and pictures. The fusion effect of multi-source heterogeneous data is improved, and the result display is more intuitive. Summary of the invention

[0005] The purpose of this invention is to solve the problem of real-time fusion of multi-source heterogeneous data in the field of sports. By introducing deep learning and machine learning technologies, the efficiency and accuracy of sports data processing are improved, automatic feature learning and model lightweight are realized, and an end-to-end solution is provided, thereby improving the comprehensiveness and reliability of sports data analysis.

[0006] In order to achieve the above object, the present invention adopts the following technical means:

[0007] A method for real-time fusion of multi-source heterogeneous sports data, comprising the following steps:

[0008] Step 1: Data acquisition, using web crawler technology and calling sports data API interfaces to obtain text and picture information about athletes and games on different social media platforms and sports news websites;

[0009] Step 2: Data preprocessing: standardize the acquired original text, enhance the contrast of the original image, and increase usability;

[0010] Step 3: Feature extraction and learning: extract and learn the features of the text and image data processed in step 2;

[0011] Step 4: Use the multimodal LSTM model to fuse the text and image features extracted in step 3;

[0012] Step 5: Use the word cloud library to generate a word cloud diagram from the result of step 4 and output it to the user for use.

[0013] In the above scheme, step 1 specifically includes the following steps:

[0014] Step 1.1: Data collection, use sports data APIs (such as ESPN API, NBA API, etc.) to obtain text and image data about athletes and games, and use web crawler technology to crawl news article text and image data from major sports news websites.

[0015] In the above scheme, step 2 specifically includes the following steps:

[0016] Step 2.1: Text preprocessing: Get the original data of news articles from sports news websites and forums in step 1, use regular expressions to remove HTML tags, special characters, and punctuation marks from the webpage text, and only keep the plain text content. Then use Python's own encoding function to encode the text to ensure that the encoding format of the text is correct. Use NLTK to perform word segmentation on the questions, remove stop words in the text, split the sentences into a sequence of phrases, and separate them with spaces to form the extracted text.

[0017] Step 2.2: Image preprocessing. Perform some basic preprocessing on the original image obtained from sports news websites and posts in step 1. Use a Gaussian filter to perform local mean filtering on the original image. Then calculate the difference image between the original image and the filtered image. Use s=c·log(1+γ) logarithmic transformation, where c is a constant and γ is a positive value used to control the curvature of the logarithmic function. Compress the areas with lower grayscale in the image and stretch the areas with higher grayscale in the image to improve the contrast of the image. Superimpose the enhanced difference image with the original image.

[0018] In the above scheme, step 3 specifically includes the following steps:

[0019] Step 3.1: Extract and transform the text and image features to obtain data vectors representing their features;

[0020] Step 3.2 Feature Learning

[0021] The support vector machine is selected as the machine learning model, and the processed text and image vectors are used as input. During the training process, the model uses Scikit-learn to extract features useful for the task from the TF-IDF features. The F1 score is used as an indicator to evaluate the trained model in the validation set, and the regularization parameters of the model parameters are adjusted according to the evaluation results.

[0022] In the above scheme, step 3.1 specifically includes the following steps:

[0023] Step 3.1.1: Use the bag-of-words model to represent the text as a count vector of words. Each vector contains the number of occurrences of all words in the document. Calculate word frequency, where n t d is the number of times word t appears in text d, n a is the total number of words in the text d, and each element represents the frequency of the corresponding word in the text;

[0024] Step 3.1.2 Use TF-IDF technology to improve the bag-of-words model. Calculate the inverse document frequency, where n t d is the number of times word t appears in text d, n a is the total number of words in the text d, and the importance weight of each word in each document is calculated using ω(t, d)×β(t, D) to obtain the weighted word vector;

[0025] Use the Word2Vec model to obtain the low-dimensional vector representation of the word, multiply the TF-IDF weighted word vector with the corresponding Word2Vec vector to obtain the TF-IDF weighted Word2Vec representation of each word, and perform weighted average of the TF-IDF weighted Word2Vec representations of all words according to the corresponding TF-IDF weighted values. o i is the TF-IDF-weighted Word2Vec representation of each word, W i represents the corresponding TF-IDF weight, N is the total number of words in the document, and the obtained weighted average vector is used as the feature vector of the entire document;

[0026] Step 3.1.3 Use the ORB algorithm to extract features from the image. Each feature point will generate a vector containing a descriptor. Multiple descriptor vectors are extracted from each image. Use the Gaussian kernel function Calculate the similarity of each image descriptor vector, x i and x j represents the descriptor vector in the image feature extraction process, where ||x i -x j || is the distance difference between two descriptor vectors. Randomly initialize the representation of each descriptor vector in the low-dimensional space. Use batch gradient descent to minimize the difference between the similarity between descriptor vectors in the high-dimensional space and the similarity between descriptor vectors in the low-dimensional space. Optimize the low-dimensional embedding, calculate the variance of the low-dimensional embedding vector, perform weighted summation of the variances of all descriptor vectors and their corresponding low-dimensional embedding vectors, and use the obtained weighted sum vector as the feature vector of all images.

[0027] Step 3.1.4: Divide the feature vectors of text and images obtained in steps 3.1.2 and 3.1.3 into training set, validation set and test set.

[0028] In the above scheme, step 4 specifically includes the following steps:

[0029] Step 4.1: Use Keras to build a multimodal LSTM model. The model should include text processing layer, image processing layer, fusion layer and Dense layer. Use layer fusion technology to merge multiple Dense layers into a larger Dense layer. The Dense layer is designed to classify categories according to identity and high quality. Use the softmax activation function to configure the output layer. Input the text and image features extracted in step 3 as input sources into the model. Use sports event classification as labels as training targets to train the model.

[0030] Step 4.2: Use the TextRank algorithm to extract keywords from the data output by the LSTM model, calculate the weight score of the word based on the similarity between the data and the relationship between sports events, create a unique index for each keyword, and output each keyword and the corresponding weight score.

[0031] In the above scheme, step 5 specifically includes the following steps:

[0032] Step 5.1: Use the WordCloud library to generate a word cloud image from the fusion results to support the application requirements of sports event analysis. The generated word cloud image is embedded in the user interface for users to view and analyze, providing users with real-time and accurate information support.

[0033] The technical solution provided by the present invention is mainly used in the fields of computer, data fusion and machine learning. The method aims to solve the problems of poor flexibility, low real-time performance and low precision in existing data fusion techniques. The following is an analysis of the main technical effects of the method:

[0034] 1. Optimization of data acquisition technology: By combining web crawler technology and API interface calls, real-time acquisition of multi-source heterogeneous data from different sports social media platforms and sports news websites is achieved. This method not only enriches the data sources, but also ensures the timeliness of the data, thereby improving the real-time nature of dynamic monitoring of sports events.

[0035] 2. Innovation in automated feature learning: Using a variety of feature extraction methods (such as bag-of-words model, TF-IDF, Word2Vec, ORB algorithm, etc.), combined with machine learning models (such as support vector machines), automated feature learning of text and image data is achieved. This method not only improves the model's ability to capture the correlation between data, but also improves the model's intelligence and adaptability, and speeds up the process and efficiency of sports data analysis.

[0036] 3. Introduction of multimodal LSTM model: The multimodal LSTM model is used to fuse text and image features, effectively interacting and fusing text and image data, improving the model's ability to capture the correlation between data, enabling the system to understand and analyze data from a broader perspective, and improving the comprehensiveness and accuracy of decision-making.

[0037] 4. Model lightweighting and optimization: Through the design of layer fusion technology, the model's computing resource consumption and storage space occupation are reduced, allowing the system to run efficiently in an environment with limited resources. The optimized model has a faster reasoning speed and can complete complex data processing and analysis tasks in a shorter time, improving the system's response speed and efficiency.

[0038] 5. Provision of end-to-end solutions: From data preprocessing to result output, a comprehensive solution is provided, which reduces manual intervention and errors in the data processing process and improves the reliability and stability of the overall system. The fusion results are displayed using methods such as word cloud diagrams and TextRank algorithms, which intuitively show the frequency and importance of keywords, making it easier for users to quickly grasp the dynamics of sports events.

[0039] 6. Enhanced generalization ability: Through continuous optimization and learning, the model can adapt to changes in different data distributions and features, maintaining robustness and reliability in different scenarios. At the same time, the model has a certain anti-interference ability and can maintain high accuracy and robustness when data is noisy or abnormal.

[0040] In summary, this method realizes the real-time fusion of multi-source heterogeneous data by introducing deep learning and multimodal LSTM models, improves the efficiency and accuracy of data processing, and has obvious advantages over the existing technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a process framework diagram of the present invention. DETAILED DESCRIPTION

[0042] The following is a detailed description of the embodiments of the present invention. Although the present invention will be described and illustrated in conjunction with some specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, modifications or equivalent substitutions made to the present invention should all be included in the scope of the claims of the present invention.

[0043] In addition, in order to better illustrate the present invention, numerous specific details are given in the following specific embodiments. It will be understood by those skilled in the art that the present invention can also be implemented without these specific details.

[0044] The present invention provides a method for real-time fusion of multi-source heterogeneous sports data, comprising the following steps:

[0045] Step 1: Data acquisition, using web crawler technology and calling sports data API interfaces to obtain text and picture information about athletes and games on different social media platforms and sports news websites; when implementing the solution of step 1, how to efficiently collect information from multiple different sources and formats of data is a technical problem that needs to be considered in the present invention. Step 1 uses web crawler technology and API interfaces to obtain relevant text and picture information from different social media platforms and news websites, ensuring the breadth and diversity of data.

[0046] Step 2: Data preprocessing, standardize the original text, enhance the contrast of the original image, and increase usability; how to process and optimize the collected original data in step 2 for subsequent analysis is a technical issue that needs to be considered. By standardizing the text and enhancing the image contrast, the usability of the data and the accuracy of the analysis are improved.

[0047] Step 3: Feature extraction and learning. Extract and learn features of the text and image data processed in step 2. When implementing step 3, how to extract effective features from the processed data and perform effective feature learning is a technical issue that needs to be considered. Use the bag-of-words model, TF-IDF, Word2Vec, and ORB algorithms to perform in-depth feature extraction and conversion of text and images to generate representative data vectors, laying the foundation for subsequent fusion and analysis.

[0048] Step 4: Use the multimodal LSTM model to fuse the text and image features extracted in step 3; by building a multimodal LSTM model, the text and image features are deeply fused, which enhances the model's ability to understand and predict complex relationships.

[0049] Step 5: Use the word cloud library to generate a word cloud graph from the results of step 4 and output it to the user for use. Use the WordCloud library to generate a word cloud graph and present complex data results in a visual form, which is convenient for users to analyze and make decisions on sports events.

[0050] In the above technical solution, step 1 specifically includes the following steps:

[0051] Step 1.1: Data collection, use the sports data API to obtain text and image data about athletes and games, and use web crawler technology to crawl news article text and image data from major news websites.

[0052] In the above technical solution, step 2 specifically includes the following steps:

[0053] Step 2.1: Text preprocessing: Get the original data of news articles from news websites and forums in step 1, use regular expressions to remove HTML tags, special characters, and punctuation marks from the webpage text, and only keep the plain text content. Then use Python's own encoding function to encode the text to ensure that the encoding format of the text is correct. Use NLTK to perform word segmentation on the text, remove stop words in the text, split the sentences into a sequence of phrases, and separate them with spaces to form the extracted text.

[0054] Step 2.2: Image preprocessing. Perform some basic preprocessing on the original image obtained from the news website or post in step 1. Use a Gaussian filter to perform local mean filtering on the original image. Then calculate the difference image between the original image and the filtered image. Use s=c·log(1+γ) logarithmic transformation, where c is a constant and γ is a positive value used to control the curvature of the logarithmic function. Compress the areas with lower grayscale in the image and stretch the areas with higher grayscale in the image to improve the contrast of the image. Superimpose the enhanced difference image with the original image.

[0055] In the above technical solution, step 3 specifically includes the following steps:

[0056] Step 3.1: Extract and transform the text and image features to obtain data vectors representing their features;

[0057] Step 3.2 Feature Learning

[0058] The support vector machine is selected as the machine learning model, and the processed text and image vectors are used as input. During the training process, the model uses Scikit-learn to extract features useful for the task from the TF-IDF features. The F1 score is used as an indicator to evaluate the trained model in the validation set, and the regularization parameters of the model parameters are adjusted according to the evaluation results.

[0059] In the above technical solution, step 3.1 specifically includes the following steps:

[0060] Step 3.1.1: Use the bag-of-words model to represent the text as a count vector of words. Each vector contains the number of occurrences of all words in the document. Calculate word frequency, where n t d is the number of times word t appears in text d, n a is the total number of words in the text d, and each element represents the frequency of the corresponding word in the text;

[0061] Step 3.1.2 Use TF-IDF technology to improve the bag-of-words model. Calculate the inverse document frequency, where n t d is the number of times word t appears in text d, n a is the total number of words in the text d, and the importance weight of each word in each document is calculated using ω(t, d)×β(t, D) to obtain the weighted word vector;

[0062] Use the Word2Vec model to obtain the low-dimensional vector representation of the word, multiply the TF-IDF weighted word vector with the corresponding Word2Vec vector to obtain the TF-IDF weighted Word2Vec representation of each word, and perform weighted average of the TF-IDF weighted Word2Vec representations of all words according to the corresponding TF-IDF weighted values. o i is the TF-IDF-weighted Word2Vec representation of each word, W i represents the corresponding TF-IDF weight, N is the total number of words in the document, and the obtained weighted average vector is used as the feature vector of the entire document;

[0063] Step 3.1.3 Use the ORB algorithm to extract features from the image. Each feature point will generate a vector containing a descriptor. Multiple descriptor vectors are extracted from each image. Use the Gaussian kernel function Calculate the similarity of each image descriptor vector, x i and x j represents the descriptor vector in the image feature extraction process, where ||x i -x j|| is the distance difference between two descriptor vectors. Randomly initialize the representation of each descriptor vector in the low-dimensional space. Use batch gradient descent to minimize the difference between the similarity between descriptor vectors in the high-dimensional space and the similarity between descriptor vectors in the low-dimensional space. Optimize the low-dimensional embedding, calculate the variance of the low-dimensional embedding vector, perform weighted summation of the variances of all descriptor vectors and their corresponding low-dimensional embedding vectors, and use the obtained weighted sum vector as the feature vector of all images.

[0064] Step 3.1.4: Divide the feature vectors of text and images obtained in steps 3.1.2 and 3.1.3 into training set, validation set and test set.

[0065] In the above technical solution, step 4 specifically includes the following steps:

[0066] Step 4.1: Use Keras to build a multimodal LSTM model. The model should include a text processing layer, an image processing layer, a fusion layer, and a Dense layer. Use layer fusion technology to merge multiple Dense layers into a larger Dense layer. The Dense layer is designed to classify categories based on identity and high quality. Use the softmax activation function to configure the entire output layer. Input the text and image features extracted in step 3 as input sources into the model. Use event classification as a label as a training target, and mainly include social events and natural disasters to train the model.

[0067] Step 4.2: Use the TextRank algorithm to extract keywords from the data output by the LSTM model, calculate the weight score of the word based on the similarity between the data and the connection relationship between sports events, create a unique index for each keyword, and output each keyword and the corresponding weight score.

[0068] In the above technical solution, step 5 specifically includes the following steps:

[0069] Step 5.1: Use the WordCloud library to generate a word cloud image from the fusion results to support the application requirements of sports event analysis. The generated word cloud image is embedded in the user interface for users to view and analyze, providing users with real-time and accurate information support.

[0070] Because the present invention adopts the above scheme, it has the following characteristics:

[0071] Real-time multi-source data fusion: Using a multimodal LSTM model to fuse text and image features can effectively interact and fuse text and image data, improving the model's ability to capture the correlation between data, enabling the system to understand and analyze data from a broader perspective, and improving the comprehensiveness and accuracy of decision-making.

[0072] Automated feature learning: It adopts a variety of custom feature extraction methods to increase the diversity of model learning. It uses machine learning models for automated feature learning without human intervention, which improves the intelligence level and adaptability of the model and speeds up the speed and efficiency of data analysis.

[0073] Model lightweighting and optimization: The layer fusion technology design reduces the model's computing resource consumption and storage space occupation, allowing the system to run efficiently in an environment with limited resources. The optimized model has a faster reasoning speed and can complete complex data processing and analysis tasks in a shorter time, improving the system's response speed and efficiency.

[0074] End-to-end solution: It provides an end-to-end solution from data preprocessing to result output, which reduces manual intervention and errors in the data processing process, improves the reliability and stability of the overall system, and reduces the complexity and cost of data processing. The fusion results are displayed using methods such as word cloud diagrams and TextRank algorithms, which intuitively show the frequency and importance of keywords, making it easier for users to quickly understand the content of sports events.

[0075] Compared with the existing technology, what are the technical advantages of this application proposal?

[0076] Real-time multi-source data fusion: Using a multimodal LSTM model to fuse text and image features can effectively interact and fuse text and image data, improving the model's ability to capture the correlation between data, enabling the system to understand and analyze data from a broader perspective, and improving the comprehensiveness and accuracy of decision-making.

[0077] Automated feature learning: Automated feature learning is performed through machine learning models without manual intervention, which improves the intelligence level and adaptability of the model. Using the support vector machine model, high-level abstract features can be extracted from the features extracted by custom extraction, which improves the efficiency and accuracy of data analysis.

[0078] Model lightweighting and optimization: The layer fusion technology design reduces the model's computing resource consumption and storage space occupation, allowing the system to run efficiently in an environment with limited resources. The optimized model has a faster reasoning speed and can complete complex data processing and analysis tasks in a shorter time, improving the system's response speed and efficiency.

[0079] End-to-end solution: It provides an end-to-end solution from data preprocessing to result output, which reduces manual intervention and errors in the data processing process and improves the reliability and stability of the overall system. The fusion results are displayed using methods such as word cloud diagrams and TextRank algorithms, which intuitively show the frequency and importance of keywords, making it easier for users to quickly understand the content of sports events.

Claims

1. A method for real-time fusion of multi-source heterogeneous sports data, characterized in that: The following steps are involved: Step 1: Data acquisition, using web crawler technology and calling sports data API interfaces to obtain text and picture information about athletes and games on different social media platforms and sports news websites; Step 2: Data preprocessing: standardize the acquired original text, enhance the contrast of the original image, and increase usability; Step 3: Feature extraction and learning: extract and learn the features of the text and image data processed in step 2; Step 4: Use the multimodal LSTM model to fuse the text and image features extracted in step 3; Step 5: Use the word cloud library to generate a word cloud diagram from the result of step 4 and output it to the user for use.

2. The method for real-time fusion of multi-source heterogeneous sports data according to claim 1, characterized in that: Step 1 specifically includes the following steps: Step 1.1: Data collection, use the sports data API to obtain text and image data about athletes and games, and use web crawler technology to crawl news article text and image data from major sports news websites.

3. The method for real-time fusion of multi-source heterogeneous sports data according to claim 1, characterized in that: Step 2 specifically includes the following steps: Step 2.1: Text preprocessing: Get the original data of news articles from sports news websites and forums in step 1, use regular expressions to remove HTML tags, special characters, and punctuation marks from the webpage text, and only keep the plain text content. Then use Python's own encoding function to encode the text to ensure that the encoding format of the text is correct. Use NLTK to perform word segmentation on the questions, remove stop words in the text, split the sentences into a sequence of phrases, and separate them with spaces to form the extracted text. Step 2.2: Image preprocessing. Perform some basic preprocessing on the original image obtained from the news website or post in step 1. Use a Gaussian filter to perform local mean filtering on the original image. Then calculate the difference image between the original image and the filtered image. Use s=c·log(1+γ) logarithmic transformation, where c is a constant and γ is a positive value used to control the curvature of the logarithmic function. Compress the areas with lower grayscale in the image and stretch the areas with higher grayscale in the image to improve the contrast of the image. Superimpose the enhanced difference image with the original image.

4. The method for real-time fusion of multi-source heterogeneous sports data according to claim 1, characterized in that: Step 3 specifically includes the following steps: Step 3.1: Extract and transform the text and image features to obtain data vectors representing their features; Step 3.2 Feature Learning The support vector machine is selected as the machine learning model, and the processed text and image vectors are used as input. During the training process, the model uses Scikit-learn to extract features useful for the task from the TF-IDF features. The F1 score is used as an indicator to evaluate the trained model in the validation set, and the regularization parameters of the model parameters are adjusted according to the evaluation results.

5. The method for real-time fusion of multi-source heterogeneous sports data according to claim 3, characterized in that: Step 3.1 specifically includes the following steps: Step 3.1.1: Use the bag-of-words model to represent the text as a count vector of words, each vector containing the number of occurrences of all words in the document. Calculate word frequency, where n t d is the number of times word t appears in text d, n a is the total number of words in the text d, and each element represents the frequency of the corresponding word in the text; Step 3.1.2 Use TF-IDF technology to improve the bag-of-words model. Calculate the inverse document frequency, where n t d is the number of times word t appears in text d, n a is the total number of words in the text d, and the importance weight of each word in each document is calculated using ω(t, d)×β(t, D) to obtain the weighted word vector; Use the Word2Vec model to obtain the low-dimensional vector representation of the word, multiply the TF-IDF weighted word vector with the corresponding Word2Vec vector to obtain the TF-IDF weighted Word2Vec representation of each word, and perform weighted average of the TF-IDF weighted Word2Vec representations of all words according to the corresponding TF-IDF weighted values. o i is the TF-IDF-weighted Word2Vec representation of each word, W i represents the corresponding TF-IDF weight, N is the total number of words in the document, and the obtained weighted average vector is used as the feature vector of the entire document; Step 3.1.3 Use the ORB algorithm to extract features from the image. Each feature point will generate a vector containing a descriptor. Multiple descriptor vectors are extracted from each image. Use the Gaussian kernel function Calculate the similarity of each image descriptor vector, x i and x j represents the descriptor vector in the image feature extraction process, where ||x i -x j || is the distance difference between two descriptor vectors. Randomly initialize the representation of each descriptor vector in the low-dimensional space. Use batch gradient descent to minimize the difference between the similarity between descriptor vectors in the high-dimensional space and the similarity between descriptor vectors in the low-dimensional space. Optimize the low-dimensional embedding, calculate the variance of the low-dimensional embedding vector, perform weighted summation of the variances of all descriptor vectors and their corresponding low-dimensional embedding vectors, and use the obtained weighted sum vector as the feature vector of all images. Step 3.1.4: Divide the feature vectors of text and images obtained in steps 3.1.2 and 3.1.3 into training set, validation set and test set.

6. The method for real-time fusion of multi-source heterogeneous sports data according to claim 3, characterized in that: Step 4 specifically includes the following steps: Step 4.1: Use Keras to build a multimodal LSTM model. The model should include text processing layer, image processing layer, fusion layer and Dense layer. Use layer fusion technology to merge multiple Dense layers into a larger Dense layer. The Dense layer is designed to classify categories according to identity and high quality. Use the softmax activation function to configure the output layer. Input the text and image features extracted in step 3 as input sources into the model. Use sports event classification as labels as training targets to train the model. Step 4.2: Use the TextRank algorithm to extract keywords from the data output by the LSTM model, calculate the weight score of the word based on the similarity between the data and the relationship between sports events, create a unique index for each keyword, and output each keyword and the corresponding weight score.

7. The method for real-time fusion of multi-source heterogeneous sports data according to claim 3, characterized in that: Step 5 specifically includes the following steps: Step 5.1: Use the WordCloud library to generate a word cloud image from the fusion results to support the application requirements of sports event analysis. The generated word cloud image is embedded in the user interface for users to view and analyze, providing users with real-time and accurate information support.

Images