Data compression method and system based on machine learning

Abstract

The invention belongs to the technical field of data compression, and discloses a data compression method and system based on machine learning. The method comprises the following steps: constructing a multi-modal feature extraction model and a multi-modal data compression model by using a machine learning algorithm according to a plurality of historical multi-modal data; collecting real-time multi-modal data, and extracting real-time multi-modal data features of the real-time multi-modal data by using the multi-modal feature extraction model; according to the real-time multi-modal data features, using a continuous learning algorithm to update the multi-modal data compression model to obtain an updated multi-modal data compression model; and according to the real-time multi-modal data characteristics, performing data compression by using the updated multi-modal data compression model to obtain real-time compressed data. According to the method, the problems of insufficient generalization ability, poor new data adaptability, disastrous forgetting and high training data dependency in the prior art are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of data compression, and in particular relates to a data compression method and system based on machine learning. Background Art

[0002] Data compression refers to the technology of reducing the space required for data storage or transmission. With the popularization of the Internet, the Internet of Things, mobile devices and various sensors, the speed and scale of data generation are growing explosively. There is a contradiction between limited storage space and the growing amount of data. Compression is needed to save storage resources, and the limitation of network bandwidth requires reducing the amount of data during data transmission to improve transmission speed and efficiency. Data compression can reduce the purchase and maintenance costs of storage hardware, and compressed data can be transmitted faster, improving the real-time nature of data processing.

[0003] The existing data compression technology has the following defects: 1) Insufficient generalization ability: Existing technologies often perform well on specific types of data, but their generalization ability is insufficient for multimodal data, which means that for different types of data, the compression algorithm may need to be redesigned, resulting in a waste of resources and increased time costs; 2) Poor adaptability to new data: Existing data compression technologies are usually based on static models. Once the training is completed, the model is fixed. When new data appears, these models cannot be adjusted adaptively, resulting in reduced compression efficiency; 3) Catastrophic forgetting: Existing data compression models often suffer from catastrophic forgetting when learning new tasks continuously, that is, they forget the previously learned experience in the process of learning new tasks. This is because traditional training methods are usually performed on fixed data sets and do not take into account the long-term retention of experience; 4) High dependence on training data: Existing data compression models usually require a large amount of training data to achieve good performance. When the amount of data is limited, the performance of the model may be greatly reduced. Summary of the invention

[0004] In order to solve the problems of insufficient generalization ability, poor adaptability to new data, catastrophic forgetting and high dependence on training data in the prior art, the present invention aims to provide a data compression method and system based on machine learning.

[0005] The technical solution adopted by the present invention is: A data compression method based on machine learning comprises the following steps: Based on some historical multimodal data, a multimodal feature extraction model and a multimodal data compression model are constructed using machine learning algorithms; Collect real-time multimodal data, and use a multimodal feature extraction model to extract real-time multimodal data features of the real-time multimodal data; According to the real-time multimodal data characteristics, the multimodal data compression model is updated using a continuous learning algorithm to obtain an updated multimodal data compression model; According to the real-time multimodal data characteristics, an updated multimodal data compression model is used to perform data compression to obtain real-time compressed data.

[0006] Furthermore, based on a number of historical multimodal data, a multimodal feature extraction model and a multimodal data compression model are constructed using a machine learning algorithm, including the following steps: Collecting a number of historical multimodal data, and preprocessing the number of historical multimodal data to obtain a number of preprocessed historical multimodal data; Based on a number of pre-processed historical multimodal data, a multimodal deep learning algorithm is used to construct a multimodal feature extraction model and generate a number of historical multimodal data features; Taking the preset data type as the clustering object, clustering a number of historical multimodal data features to obtain a number of first clustering centers and corresponding historical multimodal data feature clusters; Based on the historical multimodal data feature clusters of several preset data types, an adversarial training deep learning algorithm is used to build an optimized multimodal data compression model and generate several historical data compression experiences; Using the experience replay mechanism of the continuous learning algorithm, an experience replay pool is set up in the optimized multimodal data compression model, and several historical data compression experiences are stored in the experience replay pool; Using the elastic weight connection mechanism of the continuous learning algorithm, the initial loss function of the optimized multimodal data compression model is adjusted to obtain the adjusted loss function and the final multimodal data compression model.

[0007] Furthermore, the multimodal feature extraction model is built based on the CNN-GCN-LSTM algorithm; The multimodal data compression model is built based on the cGAN-MLP-DBN algorithm.

[0008] Furthermore, based on the preprocessed historical multimodal data, a multimodal deep learning algorithm is used to construct a multimodal feature extraction model and generate a number of historical multimodal data features, including the following steps: Using the CNN-GCN-LSTM algorithm, an initial multimodal feature extraction model is constructed; the initial multimodal feature extraction model includes a first input layer, an initial image data feature extraction module, an initial graph structure feature extraction module, an initial sequence feature extraction module and a first output layer; Using the first input layer to receive a number of preprocessed historical multimodal data; each preprocessed historical multimodal data includes preprocessed historical image data, preprocessed historical graph structure data and preprocessed historical sequence data; According to a number of pre-processed historical image data, an initial image data feature extraction module is trained to obtain an optimized image data feature extraction module, and a number of historical image data features are generated; According to a number of pre-processed historical graph structure data, an initial graph structure data feature extraction module is trained to obtain an optimized graph structure data feature extraction module, and a number of historical graph structure data features are generated; According to a number of preprocessed historical sequence data, an initial sequence data feature extraction module is trained to obtain an optimized sequence data feature extraction module, and a number of historical sequence data features are generated; Using the first output layer, historical image data features, historical graph structure data features, and historical sequence data features belonging to the same preprocessed historical multimodal data are associated to obtain a number of historical multimodal data features; The first input layer, the optimized image data feature extraction module, the optimized graph structure data feature extraction module, the optimized sequence data feature extraction module and the first output layer are integrated to obtain the final multimodal feature extraction model.

[0009] Furthermore, taking the preset data type as the clustering object, the FCM clustering algorithm is used to cluster a number of historical multimodal data features to obtain a number of first clustering centers and corresponding historical multimodal data feature clusters.

[0010] Furthermore, according to the historical multimodal data feature clusters of several preset data types, an adversarial training deep learning algorithm is used to construct an optimized multimodal data compression model, and generate several historical data compression experiences, including the following steps: Using the cGAN-MLP-DBN algorithm, an initial multimodal data compression model is constructed; the initial multimodal data compression model includes a second input layer, an initial adversarial training module, an initial multimodal data compression module, and a second output layer; Inputting historical multimodal data feature clusters of several preset data types into the second input layer, and pre-training the initial multimodal data compression module according to the historical multimodal data feature clusters of the several preset data types to obtain a pre-trained multimodal data compression module; According to the historical multimodal data feature clusters of several preset data types, the initial adversarial training module is subjected to adversarial training to obtain an optimized adversarial training module, and several expanded historical multimodal data features of each preset data type are generated; According to several historical multimodal data feature clusters of preset data types and several expanded historical multimodal data features, a pre-trained multimodal data compression module is trained to obtain an optimized multimodal data compression module, and several historical data compression experiences are generated; The integrated second input layer, the optimized adversarial training module, the optimized multimodal data compression module and the second output layer obtain an optimized multimodal data compression model.

[0011] Furthermore, real-time multimodal data is collected, and real-time multimodal data features of the real-time multimodal data are extracted using a multimodal feature extraction model, including the following steps: Collecting real-time multimodal data and preprocessing the real-time multimodal data to obtain preprocessed real-time multimodal data; the preprocessed real-time multimodal data includes preprocessed real-time image data, preprocessed real-time graph structure data and preprocessed real-time sequence data; Using a first input layer to receive preprocessed real-time image data, preprocessed real-time graph structure data, and preprocessed real-time sequence data; Using an image data feature extraction module to extract real-time image data features of the pre-processed real-time image data; Use the graph structure data feature extraction module to extract the real-time graph structure data features of the preprocessed real-time graph structure data; Using a sequence data feature extraction module to extract real-time sequence data features of the preprocessed real-time sequence data; The first output layer is used to integrate real-time image data features, real-time graph structure data features, and real-time sequence data features to obtain real-time multimodal data features of real-time multimodal data.

[0012] Furthermore, according to the real-time multimodal data features, a continuous learning algorithm is used to update the multimodal data compression model to obtain an updated multimodal data compression model, including the following steps: Obtain the real-time Euclidean distances between the real-time multimodal data features and several first cluster centers. If all real-time Euclidean distances are greater than the Euclidean distance threshold, proceed to the next step. Otherwise, use the current multimodal data compression model to perform data compression on the multimodal data compression model to obtain real-time compressed data. Setting the real-time multimodal data feature as a new data type other than a preset data type, and inputting the real-time multimodal data feature into a second input layer of the multimodal data compression model; Use the adversarial training module to embed conditional information into real-time multimodal data features to obtain real-time conditional information embedding features; Using the adversarial training module, generate several real-time training data features that belong to the same new data type as the real-time condition information embedding features; According to several real-time training data features, the adversarial training module is subjected to adversarial training to obtain an updated adversarial training module; According to several real-time training data features, a multimodal data compression module is trained to generate several real-time data compression experiences; According to the real-time multimodal data characteristics, a number of historical data compression experiences are randomly extracted from the experience replay pool and mixed with a number of real-time data compression experiences to obtain a number of mixed data compression experiences; Based on several hybrid data compression experiences, a continuous learning algorithm is used to continuously train the multimodal data compression module, and an adjusted loss function is used to generate real-time loss values ​​during the continuous training process; If the real-time loss value is less than the loss value threshold, the updated multimodal data compression module is obtained, otherwise, the continuous training of the multimodal data compression module is continued. An updated adversarial training module and an updated multimodal data compression module are integrated to obtain an updated multimodal data compression model.

[0013] Furthermore, according to the real-time multimodal data characteristics, an updated multimodal data compression model is used to perform data compression to obtain real-time compressed data, including the following steps: Inputting the real-time multimodal data features into a second input layer of the updated multimodal data compression model, and dividing the real-time multimodal data features into a plurality of real-time sub-blocks; Inputting a plurality of real-time sub-blocks into a multimodal data compression module, and encoding each real-time sub-block using the multimodal data compression module to obtain a plurality of real-time compression codewords; Using the second output layer, several real-time compressed codewords are merged to obtain initial real-time compressed data, and the initial real-time compressed data is post-processed to obtain final real-time compressed data.

[0014] A data compression system based on machine learning is used to implement a data compression method. The system includes a model building unit, a feature extraction unit, a model updating unit and a data compression unit which are connected in sequence.

[0015] The beneficial effects of the present invention are: The present invention provides a data compression method and system based on machine learning. The multimodal feature extraction and compression model adopted by the present invention can process multiple types of data, improve the generalization ability of multimodal data, and is suitable for a wider range of data compression scenarios. The multimodal data compression model constructed by the artificial intelligence algorithm can achieve a high compression rate while maintaining a high data reconstruction quality, effectively reducing the distortion of data reconstruction, and can perform continuous learning when facing new data to achieve dynamic updating of the model to adapt to changes in new data, ensure that the compression efficiency is always in the best state, and improve the adaptability to new data. The efficient multimodal data feature extraction model and multimodal data compression model are adopted to reduce unnecessary calculation steps, reduce the overall calculation complexity, and are suitable for deployment in resource-constrained environments. Through generative adversarial training, training data of the same type as new data is generated in real time. In the absence of training data, the continuous learning training effect of the model is guaranteed, and new data can be accurately identified and processed. The application of the continuous learning algorithm transfers the experience learned from historical data to new data, which helps to improve learning efficiency and reduce the demand for training sample volume for new data, thereby reducing the dependence on training data.

[0016] Other beneficial effects of the present invention will be further described in the specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flowchart of the data compression method based on machine learning in the present invention.

[0018] Figure 2 It is a structural block diagram of the data compression system based on machine learning in the present invention. DETAILED DESCRIPTION

[0019] The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments.

[0020] Embodiment 1: like Figure 1 As shown, this embodiment provides a data compression method based on machine learning, comprising the following steps: S1: Based on some historical multimodal data, a multimodal feature extraction model and a multimodal data compression model are constructed using a machine learning algorithm, including the following steps: S1-1: collecting a number of historical multimodal data, and preprocessing the number of historical multimodal data to obtain a number of preprocessed historical multimodal data; S1-2: Based on a number of pre-processed historical multimodal data, a multimodal deep learning algorithm is used to construct a multimodal feature extraction model and generate a number of historical multimodal data features; The multimodal feature extraction model is constructed based on the CNN-GCN-LSTM algorithm, and the multimodal feature extraction model includes a first input layer, an image data feature extraction module constructed based on the CNN algorithm, a graph structure feature extraction module constructed based on the GCN algorithm, a sequence feature extraction module constructed based on the LSTM algorithm, and a first output layer, and the image data feature extraction module, the graph structure feature extraction module, and the sequence feature extraction module are arranged in parallel between the first input layer and the first output layer; According to a number of pre-processed historical multimodal data, a multimodal feature extraction model is constructed using a multimodal deep learning algorithm, and a number of historical multimodal data features are generated, including the following steps: S1-2-1: Use the Convolutional Neural Networks (CNN)-Graph Convolutional Network (GCN)-Long Short-Term Memory (LSTM) algorithm to build an initial multimodal feature extraction model; the initial multimodal feature extraction model includes the first input layer, the initial image data feature extraction module, the initial graph structure feature extraction module, the initial sequence feature extraction module and the first output layer; The CNN network extracts local features in image data, such as edges, textures, and shapes, through convolutional layers and pooling layers to obtain image data features; the GCN network performs feature propagation on graph structure data through operations similar to convolution, extracts node features (position features, image features, etc.) of several segmented images in the graph structure data, and edge features of the positional relationship between nodes, and forms complex, high-dimensional global graph structure features; the LSTM network processes sequence data, such as text, speech, or time series, captures the time dependency in the data, and obtains sequence data features; through the cooperation of the above networks, the feature extraction task of multimodal data is realized to provide preprocessing for subsequent data compression, improve the efficiency and speed of data compression, and reduce the amount of calculation; S1-2-2: using the first input layer to receive a number of preprocessed historical multimodal data; each preprocessed historical multimodal data includes preprocessed historical image data, preprocessed historical graph structure data and preprocessed historical sequence data; S1-2-3: According to a number of pre-processed historical image data, the initial image data feature extraction module is trained to obtain an optimized image data feature extraction module, and a number of historical image data features are generated; S1-2-4: According to a number of pre-processed historical graph structure data, the initial graph structure data feature extraction module is trained to obtain an optimized graph structure data feature extraction module, and a number of historical graph structure data features are generated; S1-2-5: According to a number of preprocessed historical sequence data, the initial sequence data feature extraction module is trained to obtain an optimized sequence data feature extraction module, and a number of historical sequence data features are generated; S1-2-6: using the first output layer, associating historical image data features, historical graph structure data features, and historical sequence data features belonging to the same preprocessed historical multimodal data to obtain a number of historical multimodal data features; S1-2-7: Integrate the first input layer, the optimized image data feature extraction module, the optimized graph structure data feature extraction module, the optimized sequence data feature extraction module and the first output layer to obtain the final multimodal feature extraction model; S1-3: Taking the preset data type as the clustering object, using the Fuzzy C-mean (FCM) clustering algorithm to cluster a number of historical multimodal data features, and obtaining a number of first clustering centers and corresponding historical multimodal data feature clusters, including the following steps: S1-3-1: Use the FCM clustering algorithm to initialize cluster centers in several historical multimodal data features and generate several initial first cluster centers; S1-3-2: Obtain the initial historical Euclidean distance between each historical multimodal data feature and a number of initial first cluster centers, and generate an initial historical fuzzy membership based on the initial historical Euclidean distance; The formula is:

[0021] In the formula, For the i Historical multimodal data characteristics and j The Euclidean distance of the initial first cluster center; For the Historical multimodal data features; For the j The initial first cluster center; is the data characteristic indicator; is the indicator of the first cluster center;

[0022] In the formula, For the Historical multimodal data features; is the data characteristic indicator; They are all initial first cluster center indicators; is the total number of the first cluster centers; For the i Historical multimodal data features to j ,k The distance of the initial first cluster center; For the i Historical multimodal data features belong to The initial historical fuzzy membership of the initial cluster centers; S1-3-3: updating the initial first cluster center according to the initial historical fuzzy membership to obtain several updated first cluster centers; The formula is:

[0023] In the formula, For the j Updated first cluster center; is the fuzzy factor; is the data characteristic indicator; is the total number of data; is the indicator of the first cluster center; For the Historical multimodal data features; For the i Historical multimodal data features belong to The updated historical fuzzy membership of the first cluster center; S1-3-4: Obtain updated historical fuzzy membership of each historical multimodal data feature and a number of updated first cluster centers; S1-3-5: Based on several updated historical fuzzy memberships, the Lagrange multiplier method is used to calculate the merged function to obtain the merged function value and the merged function change value; The formula is:

[0024] In the formula, For the t , t -1 iteration of the combined function value of the Lagrange multiplier method; is the corresponding change value; For the i Characteristic parameters; t is the indicator of the number of iterations; For the i Historical multimodal data features belong to The updated historical fuzzy membership of the first cluster center; is the fuzzy factor; is the data indicator; is the total number of data; is the indicator of the first cluster center; is the total number of cluster centers; For the i Historical multimodal data features to j Updated Euclidean distance of the first cluster center; S1-3-6: If the merge function value is greater than the function threshold, or the merge function change value is greater than the change value threshold, continue to update the first cluster center, otherwise, use the updated first cluster center as the final first cluster center and proceed to the next step; S1-3-7: Obtain the final historical Euclidean distance between each historical multimodal data feature and several final first cluster centers, and divide the historical multimodal data features into the final first cluster centers according to the final historical Euclidean distance to obtain several historical multimodal data feature clusters; the historical multimodal data feature clusters include several historical multimodal data features belonging to the same preset data type; S1-3-8: Setting a corresponding preset data type for each final first cluster center, and spreading the preset data type of the final first cluster center to the corresponding historical multimodal data feature cluster; S1-4: Based on the historical multimodal data feature clusters of several preset data types, an adversarial training deep learning algorithm is used to build an optimized multimodal data compression model and generate several historical data compression experiences; The multimodal data compression model is constructed based on the Conditional Generative Adversarial Network (cGAN)-Multilayer Perceptron (MLP)-Deep Belief Network (DBN) algorithm, and the multimodal data compression model includes a second input layer, an adversarial training module constructed based on the cGAN-MLP algorithm, a multimodal data compression module constructed based on the DBN algorithm, and a second output layer. The adversarial training module includes a conditional information embedder and a conditional information processor both constructed based on the MLP algorithm, and a generator and a discriminator both constructed based on a recurrent neural network (RNN) algorithm. An adversarial training module and a multimodal data compression module are arranged in parallel between the first input layer and the first output layer. The generator is respectively connected to the second input layer, the discriminator, and the multimodal data compression module, and the discriminator is connected to the conditional information processor. The conditional information embedder is used to process the conditional information including multimodal data features, obtain the conditional information embedding features in sequence format, and realize the integration of multimodal conditional information into the generation process; the generator is used to generate a number of real training data features belonging to the same new data type as the conditional information embedding features according to the sequence data and random noise output by the conditional information embedder; the conditional information processor is used to process the additional conditional information of the multimodal data features to help the discriminator more accurately judge the authenticity of the training data features; the discriminator analyzes whether the generated training data features are true and meet the given conditional information; the generator and the discriminator compete with each other through the adversarial training process, the generator tries to generate training data features that can deceive the discriminator, and the discriminator tries to better identify the real and false training data features; the adversarial training module performs well in generating realistic data, especially in the absence of sufficient new data, the generator can help generate data for training, and realize the generation of sufficient training samples in the case of a small amount of new data, ensuring the diversity of training samples, and realizing the continuous training of the multimodal data compression module, which can enhance the generalization ability of the multimodal data compression module, so as to better process real new data; According to the historical multimodal data feature clusters of several preset data types, an adversarial training deep learning algorithm is used to build an optimized multimodal data compression model and generate several historical data compression experiences, including the following steps: S1-4-1: Use the cGAN-MLP-DBN algorithm to build an initial multimodal data compression model; the initial multimodal data compression model includes a second input layer, an initial adversarial training module, an initial multimodal data compression module and a second output layer; S1-4-2: inputting historical multimodal data feature clusters of several preset data types into the second input layer, and pre-training the initial multimodal data compression module according to the historical multimodal data feature clusters of the several preset data types to obtain a pre-trained multimodal data compression module; S1-4-3: According to the historical multimodal data feature clusters of several preset data types, the initial adversarial training module is subjected to adversarial training to obtain an optimized adversarial training module, and several expanded historical multimodal data features of each preset data type are generated; S1-4-4: According to several historical multimodal data feature clusters of preset data types and several expanded historical multimodal data features, a pre-trained multimodal data compression module is trained to obtain an optimized multimodal data compression module, and several historical data compression experiences are generated; S1-4-5: integrating the second input layer, the optimized adversarial training module, the optimized multimodal data compression module and the second output layer to obtain an optimized multimodal data compression model; S1-5: Use the experience replay mechanism of the continuous learning algorithm to set up an experience replay pool in the optimized multimodal data compression model, and store several historical data compression experiences in the experience replay pool; The experience replay pool is used to store the interactive experience of the multimodal data compression model in the environment. These experiences are usually stored in the form of (s, a, r, s'), where s is the current state, that is, the current model parameters of the multimodal data compression model, a is the action taken, including various actions for adjusting the model parameters, r is the reward obtained, which is used to characterize the influence of the action on the state, and s' is the next state, that is, the state after the model parameters are adjusted during the training process; Since the experience in the experience replay pool is randomly drawn, it helps to break the correlation between consecutive experiences, thereby reducing the variance in the training of the multimodal data compression model. By reusing experience, the multimodal data compression model can learn more from limited experience, which is especially useful in the case of new data types or scarce samples. Experience replay helps stabilize the learning process and reduce fluctuations during training. Regularly check the size of the experience replay pool. If it exceeds the preset capacity, remove some experience according to strategies (such as priority sampling and the longest unvisited). S1-6: Using the elastic weight connection mechanism of the continuous learning algorithm, adjusting the initial loss function of the optimized multimodal data compression model, obtaining the adjusted loss function, and obtaining the final multimodal data compression model, including the following steps: S1-6-1: Obtain the initial loss function of the optimized multimodal data compression model, as well as several key weights for adjusting model parameters; By calculating the gradient of the weights with respect to the task loss, we can identify the key weights that are crucial to adjusting the model parameters during the training process. This information can be used to reduce catastrophic forgetting when training new tasks, which is particularly important in continuous learning scenarios. S1-6-2: Obtain the importance measures of several key weights, and generate penalty items for adjusting model parameters based on the importance measures; When training a new task, in order to prevent the multimodal data compression model from forgetting the previously learned experience, an additional penalty term is added to the loss function using the Elastic Weight Consolidation (EWC) mechanism. This penalty term is proportional to the importance measure of the weight and inversely proportional to the amount of change in the weight on the new data. This penalty term ensures that when training on new data, the weights that are important to the old data do not change too much, thereby reducing the risk of catastrophic forgetting. The larger the importance measure, the more restricted the change of the corresponding weight on the new data. In this way, the multimodal data compression model can retain the experience of the old data while learning the new data; S1-6-3: adding the penalty term to the initial loss function of the optimized multimodal data compression model to obtain an adjusted loss function and a final multimodal data compression model; S2: collecting real-time multimodal data, and using a multimodal feature extraction model to extract real-time multimodal data features of the real-time multimodal data, including the following steps: S2-1: collecting real-time multimodal data, and preprocessing the real-time multimodal data to obtain preprocessed real-time multimodal data; the preprocessed real-time multimodal data includes preprocessed real-time image data, preprocessed real-time graph structure data, and preprocessed real-time sequence data; S2-2: using the first input layer to receive the preprocessed real-time image data, the preprocessed real-time graph structure data, and the preprocessed real-time sequence data; S2-3: using an image data feature extraction module to extract real-time image data features of the preprocessed real-time image data; S2-4: using a graph structure data feature extraction module to extract real-time graph structure data features of the preprocessed real-time graph structure data; S2-5: using a sequence data feature extraction module to extract real-time sequence data features of the preprocessed real-time sequence data; S2-6: using the first output layer, integrating the real-time image data features, the real-time graph structure data features, and the real-time sequence data features to obtain the real-time multimodal data features of the real-time multimodal data; S3: According to the real-time multimodal data features, the multimodal data compression model is updated using a continuous learning algorithm to obtain an updated multimodal data compression model, including the following steps: S3-1: Obtain the real-time Euclidean distance between the real-time multimodal data features and several first cluster centers. If all the real-time Euclidean distances are greater than the Euclidean distance threshold, proceed to the next step. Otherwise, use the current multimodal data compression model to perform data compression on the multimodal data compression model to obtain real-time compressed data. It is used to automatically identify the data features of new data types and prepare in advance the continuous training environment of the multimodal data compression model, including the preparation of training samples, so that the multimodal data compression model can accurately process the new data features of the data type; S3-2: setting the real-time multimodal data feature to a new data type other than the preset data type, and inputting the real-time multimodal data feature into a second input layer of the multimodal data compression model; S3-3: Use the conditional information embedder of the adversarial training module to embed conditional information into the real-time multimodal data features to obtain real-time conditional information embedding features; S3-4: using the generator of the adversarial training module, generating a number of real-time training data features that belong to the same new data type as the real-time conditional information embedding feature; S3-5: performing adversarial training on the adversarial training module according to a number of real-time training data features to obtain an updated adversarial training module; S3-6: training the multimodal data compression module according to the characteristics of a number of real-time training data, and generating a number of real-time data compression experiences; S3-7: According to the real-time multimodal data characteristics, a number of historical data compression experiences are randomly extracted from the experience replay pool, and mixed with a number of real-time data compression experiences to obtain a number of mixed data compression experiences; By combining historical data compression experience, the multimodal data compression model can transfer the experience learned from previous data to new data, which helps improve learning efficiency and reduce the number of training samples required for new data. Combining old experience can help prevent the multimodal data compression model from forgetting the experience learned on old data when training on new data. This is a common continuous learning challenge called "catastrophic forgetting". Old experience can provide a broader view of data distribution, which helps the multimodal data compression model generalize better on new data. S3-8: Based on several hybrid data compression experiences, use a continuous learning algorithm to continuously train the multimodal data compression module, and use an adjusted loss function to generate real-time loss values ​​during the continuous training process; S3-9: If the real-time loss value is less than the loss value threshold, an updated multimodal data compression module is obtained, otherwise, continuous training of the multimodal data compression module is continued; S3-10: Integrate the updated adversarial training module and the updated multimodal data compression module to obtain an updated multimodal data compression model; S4: According to the real-time multimodal data characteristics, using the updated multimodal data compression model, data compression is performed to obtain real-time compressed data, including the following steps: S4-1: inputting the real-time multimodal data features into the second input layer of the updated multimodal data compression model, and dividing the real-time multimodal data features into a plurality of real-time sub-blocks; S4-2: inputting a plurality of real-time sub-blocks into a multimodal data compression module, and using the multimodal data compression module to encode each real-time sub-block to obtain a plurality of real-time compression codewords; The multimodal data compression module is equipped with multiple compression channels, which can process data features of different modes at the same time, thus improving the adaptability to multimodal data scenarios; S4-3: using the second output layer, merging a plurality of real-time compressed codewords to obtain initial real-time compressed data, and post-processing the initial real-time compressed data to obtain final real-time compressed data; Through pre-training, the DBN network has learned the effective representation of multimodal data features and can compress data based on multimodal features, so that the compressed data takes up less storage space or bandwidth while retaining information.

[0025] Embodiment 2: like Figure 2 As shown, this embodiment provides a data compression system based on machine learning, which is used to implement a data compression method. The system includes a model building unit, a feature extraction unit, a model updating unit and a data compression unit connected in sequence; A model building unit, used to build a multimodal feature extraction model and a multimodal data compression model based on a number of historical multimodal data using a machine learning algorithm; A feature extraction unit, used for collecting real-time multimodal data, and extracting real-time multimodal data features of the real-time multimodal data using a multimodal feature extraction model; A model updating unit, used to update the multimodal data compression model according to the real-time multimodal data features using a continuous learning algorithm to obtain an updated multimodal data compression model; The data compression unit is used to compress the data according to the real-time multimodal data characteristics using an updated multimodal data compression model to obtain real-time compressed data.

[0026] The present invention provides a data compression method and system based on machine learning. The multimodal feature extraction and compression model adopted by the present invention can process multiple types of data, improve the generalization ability of multimodal data, and is suitable for a wider range of data compression scenarios. The multimodal data compression model constructed by the artificial intelligence algorithm can achieve a high compression rate while maintaining a high data reconstruction quality, effectively reducing the distortion of data reconstruction, and can perform continuous learning when facing new data to achieve dynamic updating of the model to adapt to changes in new data, ensure that the compression efficiency is always in the best state, and improve the adaptability to new data. The efficient multimodal data feature extraction model and multimodal data compression model are adopted to reduce unnecessary calculation steps, reduce the overall calculation complexity, and are suitable for deployment in resource-constrained environments. Through generative adversarial training, training data of the same type as new data is generated in real time. In the absence of training data, the continuous learning training effect of the model is guaranteed, and new data can be accurately identified and processed. The application of the continuous learning algorithm transfers the experience learned from historical data to new data, which helps to improve learning efficiency and reduce the demand for training sample volume for new data, thereby reducing the dependence on training data.

[0027] The present invention is not limited to the above optional implementations, and anyone can derive other various forms of products under the enlightenment of the present invention. The above specific implementations should not be understood as limiting the scope of protection of the present invention. The scope of protection of the present invention should be based on the definition in the claims, and the description can be used to interpret the claims.

Claims

1. A data compression method based on machine learning, characterized in that: The steps include: Based on some historical multimodal data, a multimodal feature extraction model and a multimodal data compression model are constructed using machine learning algorithms; Collect real-time multimodal data, and use a multimodal feature extraction model to extract real-time multimodal data features of the real-time multimodal data; According to the real-time multimodal data characteristics, the multimodal data compression model is updated using a continuous learning algorithm to obtain an updated multimodal data compression model; According to the real-time multimodal data characteristics, an updated multimodal data compression model is used to perform data compression to obtain real-time compressed data.

2. The data compression method based on machine learning according to claim 1, characterized in that: Based on some historical multimodal data, a multimodal feature extraction model and a multimodal data compression model are constructed using a machine learning algorithm, including the following steps: Collecting a number of historical multimodal data, and preprocessing the number of historical multimodal data to obtain a number of preprocessed historical multimodal data; Based on a number of pre-processed historical multimodal data, a multimodal deep learning algorithm is used to construct a multimodal feature extraction model and generate a number of historical multimodal data features; Taking the preset data type as the clustering object, clustering a number of historical multimodal data features to obtain a number of first clustering centers and corresponding historical multimodal data feature clusters; Based on the historical multimodal data feature clusters of several preset data types, an adversarial training deep learning algorithm is used to build an optimized multimodal data compression model and generate several historical data compression experiences; Using the experience replay mechanism of the continuous learning algorithm, an experience replay pool is set in the optimized multimodal data compression model, and several historical data compression experiences are stored in the experience replay pool; Using the elastic weight connection mechanism of the continuous learning algorithm, the initial loss function of the optimized multimodal data compression model is adjusted to obtain the adjusted loss function and the final multimodal data compression model.

3. The data compression method based on machine learning according to claim 2, characterized in that: The multimodal feature extraction model is constructed based on the CNN-GCN-LSTM algorithm; The multimodal data compression model is built based on the cGAN-MLP-DBN algorithm.

4. The data compression method based on machine learning according to claim 3, characterized in that: According to a number of pre-processed historical multimodal data, a multimodal feature extraction model is constructed using a multimodal deep learning algorithm, and a number of historical multimodal data features are generated, including the following steps: Using the CNN-GCN-LSTM algorithm, an initial multimodal feature extraction model is constructed; the initial multimodal feature extraction model includes a first input layer, an initial image data feature extraction module, an initial graph structure feature extraction module, an initial sequence feature extraction module and a first output layer; Using the first input layer to receive a plurality of preprocessed historical multimodal data; each of the preprocessed historical multimodal data includes preprocessed historical image data, preprocessed historical graph structure data and preprocessed historical sequence data; According to a number of pre-processed historical image data, an initial image data feature extraction module is trained to obtain an optimized image data feature extraction module, and a number of historical image data features are generated; According to a number of pre-processed historical graph structure data, an initial graph structure data feature extraction module is trained to obtain an optimized graph structure data feature extraction module, and a number of historical graph structure data features are generated; According to a number of preprocessed historical sequence data, an initial sequence data feature extraction module is trained to obtain an optimized sequence data feature extraction module, and a number of historical sequence data features are generated; Using the first output layer, historical image data features, historical graph structure data features, and historical sequence data features belonging to the same preprocessed historical multimodal data are associated to obtain a number of historical multimodal data features; The first input layer, the optimized image data feature extraction module, the optimized graph structure data feature extraction module, the optimized sequence data feature extraction module and the first output layer are integrated to obtain the final multimodal feature extraction model.

5. The data compression method based on machine learning according to claim 4, characterized in that: Taking the preset data type as the clustering object, the FCM clustering algorithm is used to cluster several historical multimodal data features to obtain several first clustering centers and corresponding historical multimodal data feature clusters.

6. The data compression method based on machine learning according to claim 5, characterized in that: According to the historical multimodal data feature clusters of several preset data types, an adversarial training deep learning algorithm is used to build an optimized multimodal data compression model and generate several historical data compression experiences, including the following steps: Using the cGAN-MLP-DBN algorithm, an initial multimodal data compression model is constructed; the initial multimodal data compression model includes a second input layer, an initial adversarial training module, an initial multimodal data compression module and a second output layer; Inputting historical multimodal data feature clusters of several preset data types into the second input layer, and pre-training the initial multimodal data compression module according to the historical multimodal data feature clusters of the several preset data types to obtain a pre-trained multimodal data compression module; According to the historical multimodal data feature clusters of several preset data types, the initial adversarial training module is adversarially trained to obtain an optimized adversarial training module, and several expanded historical multimodal data features of each preset data type are generated; According to several historical multimodal data feature clusters of preset data types and several expanded historical multimodal data features, a pre-trained multimodal data compression module is trained to obtain an optimized multimodal data compression module, and several historical data compression experiences are generated; The integrated second input layer, the optimized adversarial training module, the optimized multimodal data compression module and the second output layer obtain an optimized multimodal data compression model.

7. The data compression method based on machine learning according to claim 6, characterized in that: Collecting real-time multimodal data and using a multimodal feature extraction model to extract real-time multimodal data features of the real-time multimodal data includes the following steps: Collecting real-time multimodal data and preprocessing the real-time multimodal data to obtain preprocessed real-time multimodal data; the preprocessed real-time multimodal data includes preprocessed real-time image data, preprocessed real-time graph structure data and preprocessed real-time sequence data; Using a first input layer to receive preprocessed real-time image data, preprocessed real-time graph structure data, and preprocessed real-time sequence data; Using an image data feature extraction module to extract real-time image data features of the pre-processed real-time image data; Use the graph structure data feature extraction module to extract the real-time graph structure data features of the preprocessed real-time graph structure data; Using a sequence data feature extraction module to extract real-time sequence data features of the preprocessed real-time sequence data; The first output layer is used to integrate real-time image data features, real-time graph structure data features, and real-time sequence data features to obtain real-time multimodal data features of real-time multimodal data.

8. The data compression method based on machine learning according to claim 7, characterized in that: According to the real-time multimodal data characteristics, the multimodal data compression model is updated using a continuous learning algorithm to obtain an updated multimodal data compression model, including the following steps: Obtain the real-time Euclidean distances between the real-time multimodal data features and several first cluster centers. If all real-time Euclidean distances are greater than the Euclidean distance threshold, proceed to the next step. Otherwise, use the current multimodal data compression model to perform data compression on the multimodal data compression model to obtain real-time compressed data. Setting the real-time multimodal data feature as a new data type other than a preset data type, and inputting the real-time multimodal data feature into a second input layer of the multimodal data compression model; Use the adversarial training module to embed conditional information into real-time multimodal data features to obtain real-time conditional information embedding features; Using the adversarial training module, generate several real-time training data features that belong to the same new data type as the real-time condition information embedding features; According to several real-time training data features, the adversarial training module is subjected to adversarial training to obtain an updated adversarial training module; According to several real-time training data features, a multimodal data compression module is trained to generate several real-time data compression experiences; According to the real-time multimodal data characteristics, a number of historical data compression experiences are randomly extracted from the experience replay pool and mixed with a number of real-time data compression experiences to obtain a number of mixed data compression experiences; Based on several hybrid data compression experiences, a continuous learning algorithm is used to continuously train the multimodal data compression module, and an adjusted loss function is used to generate real-time loss values ​​during the continuous training process; If the real-time loss value is less than the loss value threshold, an updated multimodal data compression module is obtained, otherwise, continuous training of the multimodal data compression module is continued; An updated adversarial training module and an updated multimodal data compression module are integrated to obtain an updated multimodal data compression model.

9. The data compression method based on machine learning according to claim 8, characterized in that: According to the real-time multimodal data characteristics, an updated multimodal data compression model is used to compress the data to obtain real-time compressed data, including the following steps: Inputting the real-time multimodal data features into a second input layer of the updated multimodal data compression model, and dividing the real-time multimodal data features into a plurality of real-time sub-blocks; Inputting a plurality of real-time sub-blocks into a multimodal data compression module, and encoding each real-time sub-block using the multimodal data compression module to obtain a plurality of real-time compression codewords; Using the second output layer, several real-time compression codewords are merged to obtain initial real-time compression data, and the initial real-time compression data is post-processed to obtain final real-time compression data.

10. A data compression system based on machine learning, used to implement the data compression method according to any one of claims 1 to 9, characterized in that: The system comprises a model building unit, a feature extraction unit, a model updating unit and a data compression unit which are connected in sequence.

Images