Multi-source data view feature extraction and rapid matching system based on deep learning

Abstract

The invention discloses a multi-source data view feature extraction and quick matching system based on deep learning, and relates to the technical field of computer deep learning, a multi-modal data preprocessing module supports access of various heterogeneous data, technologies such as adaptive normalization are adopted to process data, a feature extraction module constructs a multi-branch attention network architecture, and a multi-source data view feature extraction and quick matching system is established. The trunk uses improved ResNet50 and introduces a channel attention mechanism, output of each branch generates a joint feature through a tensor fusion layer, the matching retrieval module adopts a layered hash coding strategy and a multi-index hash table, the weight learning module dynamically adjusts modal weight through a meta-learning framework and a context awareness mechanism, and the matching retrieval module performs matching retrieval on the modal weight through the meta-learning framework and the context awareness mechanism. The semantic enhancement module constructs a knowledge graph embedding layer and develops a semantic similarity calculation function. The multi-branch attention network improves the accuracy of feature extraction, distributed computing acceleration supports high-concurrency processing, and the method also has the functions of incremental learning, interpretable analysis and security privacy protection, and meets various requirements.

Description

Technical Field

[0001] This invention relates to the field of computer deep learning technology, and in particular to a deep learning-based system for feature extraction and fast matching of multi-source data views. Background Technology

[0002] In today's digital age, the widespread application of multi-source data has become a key characteristic in many fields. From vehicle sensor data and road monitoring image data integrated into intelligent transportation systems, to patient medical record text data and image data combined in the medical field, multi-source data contains a wealth of information. However, these data are heterogeneous, with different data formats, feature dimensions, and semantic meanings, posing significant challenges to the effective processing and utilization of the data. Traditional data processing methods are often limited to processing single-modal data, making it difficult to fully tap the value brought by multi-source data fusion.

[0003] With the rapid development of deep learning technology, it has achieved remarkable results in fields such as image recognition and natural language processing. However, applying deep learning to feature extraction and fast matching of multi-source data still faces many challenges. In feature extraction, the methods for extracting features from different modalities vary significantly. Designing a general and efficient deep feature extraction framework that can take into account the characteristics of multiple modalities and accurately extract valuable features is a pressing problem. Furthermore, during the fusion of multi-modal data, semantic differences and feature misalignment can easily lead to a decline in the quality of the fused features, affecting subsequent matching and analysis.

[0004] In terms of rapid matching and retrieval, traditional matching algorithms struggle to meet the demands in terms of efficiency and accuracy as data volumes continue to grow. This is particularly true in scenarios with high real-time requirements, such as rapid target retrieval in real-time surveillance videos and rapid question matching in intelligent customer service systems, where existing methods often suffer from slow retrieval speeds and low accuracy. Furthermore, the dynamic and diverse nature of multi-source data necessitates that matching algorithms possess greater adaptability and flexibility, capable of optimization and adjustment based on different data characteristics and application scenarios. Simultaneously, the data processing process must consider issues such as security and privacy protection, model interpretability, and efficient resource utilization. These factors place higher demands on deep learning-based multi-source data view feature extraction and rapid matching systems. Summary of the Invention

[0005] The present invention proposes a deep learning-based multi-source data view feature extraction and fast matching system to solve the problems mentioned in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A deep learning-based system for feature extraction and fast matching of multi-source data views includes the following modules: Multimodal data preprocessing module: Image data adopts adaptive histogram equalization technology, which dynamically adjusts the histogram according to the local gray-level distribution of the image. For point cloud data, voxel downsampling method is used to reduce the amount of data while maintaining features. Text data first undergoes lexical and syntactic analysis to remove stop words, and then word vector technology is used to convert the text into numerical vectors. In addition, adaptive normalization method is used to standardize data of different modalities. Deep feature extraction module: The backbone network adopts an improved ResNet50 and introduces a channel attention mechanism to adaptively adjust the channels according to the importance weight of the task. The branch networks include CNN, PointNet++ and BERT, which extract features from the data respectively. Then, the outputs of each branch are fused through a tensor fusion layer to generate a joint feature representation. Fast matching and retrieval module: It adopts a hierarchical hash coding strategy. First, it extracts data features through a deep learning model, maps them into compact binary codes, and uses the triplet loss function to improve the coding discrimination. Based on this, a multi-index hash table is constructed. During the search, Hamming distance is used to quickly filter candidate data similar to the target data. Dynamic weight learning module: A meta-learning framework was designed, and a context-aware mechanism was introduced. By analyzing the context information of the current data, the importance weights of different modal data in the current scenario were calculated. Semantic Enhancement Matching Module: Constructs a knowledge graph embedding layer to map entities and relationships to a low-dimensional vector space. Through the semantic relationships of the knowledge graph, it develops a semantic similarity calculation function and uses an attention mechanism to weight different semantic dimensions during calculation. Distributed computing acceleration module: It adopts a strategy that combines model parallelism and data sharding, and supports each computing node to directly access the data shards stored locally during computing.

[0007] Furthermore, the multimodal data preprocessing module also includes an adaptive noise suppression submodule. This submodule uses variational autoencoder technology to address the noise characteristics in different modal data. For Gaussian noise in image data, the variational autoencoder learns the latent distribution of the image to generate a denoised image. During the learning process, the reparameterization technique is used to separate the randomness of the noise from the latent features of the image.

[0008] Furthermore, the deep feature extraction module also includes a cross-modal alignment submodule. This submodule constructs a contrastive learning framework, which learns cross-modal features by comparing the feature representations of similar and dissimilar samples in different modal data. Specifically, it maximizes the feature similarity of similar sample pairs and minimizes the feature similarity of dissimilar sample pairs, thereby aligning different modal data in the feature space. At the same time, it uses bilinear pooling technology to fuse cross-modal features.

[0009] Furthermore, the fast matching retrieval module also includes a hierarchical index submodule, which constructs a hybrid index structure of KD-tree and hash table. The KD-tree divides the data space into multiple subspaces according to the feature dimensions of the data, so that similar data are clustered in nearby subspaces. The hash table is used to index the subspaces after the KD-tree is divided. The locality-sensitive hashing algorithm is used to map similar data to the same or nearby hash buckets. During retrieval, the subspace containing the target data is first located through the KD-tree, and then the search is performed in the hash bucket corresponding to that subspace.

[0010] Furthermore, the dynamic weight learning module also includes an online meta-learning submodule, which uses the MAML algorithm to adjust the model parameters based on previously learned knowledge when a new task arrives. When facing a new scenario, the model uses the experience learned in other related scenarios to adapt to the new environment. At the same time, a weight decay mechanism is introduced to reduce the impact of historical data on the current model weights over time, so that the model can adapt to changes in new data.

[0011] Furthermore, the semantic enhancement matching module also includes a knowledge reasoning submodule, which constructs a graph neural network (GNN) and uses the connection relationships between nodes in the knowledge graph for knowledge reasoning. Through the message passing mechanism of the GNN, nodes obtain information from their neighboring nodes and update their own feature representations. On this basis, a multi-hop reasoning mechanism is developed, which allows the model to perform multi-step reasoning in the knowledge graph and mine semantic relationships by passing and aggregating information from neighboring nodes multiple times.

[0012] Furthermore, the distributed computing acceleration module also includes an elastic resource scheduling submodule. This submodule establishes a resource consumption prediction model. By analyzing the complexity of the model and the size of the data, it predicts the amount of resources required for the computing task. Based on the prediction results, it uses a reinforcement learning strategy to dynamically allocate resources. The reinforcement learning agent learns the optimal resource scheduling strategy by interacting with the system environment.

[0013] Furthermore, it also includes an incremental learning module, which uses an online gradient matching algorithm to balance the learning effects of new and old data when introducing new data for learning, retaining the knowledge learned by the model on historical data. At the same time, it develops a sample selection strategy to select new samples for learning by evaluating the uncertainty and redundancy of samples.

[0014] Furthermore, it also includes an interpretability analysis module, which uses the SHAP value method to quantify the contribution of each feature to the model's prediction results. By calculating the SHAP value, we can understand the role and degree of contribution of features in the model's decision-making. In addition, a counterfactual explanation generator is developed to generate counterfactual samples by making small changes to the input data and analyze the changes in the model's prediction results.

[0015] Furthermore, it also includes a security and privacy protection module. This module adopts a federated learning architecture, which allows multiple participants to jointly train the model without sharing the original data. Each participant trains the model locally and then uploads the model parameters to the central server for aggregation. A differential privacy mechanism is deployed to add noise to the data during model training and parameter aggregation.

[0016] Compared with existing technologies, the beneficial effects of this invention are: In terms of feature extraction, the system's multi-branch attention network architecture enables deep feature extraction tailored to the characteristics of different modalities. Through an improved ResNet50 backbone network and a unique channel attention mechanism, the system enhances its ability to capture key features of image data, resulting in more accurate image feature representation. For other modalities such as point clouds and text, the corresponding branch networks can also fully mine their features. The outputs of each branch are fused through a tensor fusion layer, generating a joint feature representation that integrates the advantages of multiple modalities. Compared to traditional single-modal feature extraction methods, this representation more comprehensively reflects the essential characteristics of the data, providing a higher-quality feature foundation for subsequent matching and analysis.

[0017] In terms of fast matching and retrieval, the hierarchical hash encoding strategy and the design of multi-index hash tables greatly improve retrieval efficiency. By learning compact binary encoding and utilizing Hamming distance for fast search, millisecond-level retrieval response is achieved on large-scale datasets, meeting the needs of applications with high real-time requirements. Compared to traditional matching algorithms, the retrieval speed is improved several times, enabling the system to quickly and accurately find target data from massive amounts of data, thus enhancing its practicality and application value.

[0018] Furthermore, the system has been optimized in several key aspects. The dynamic weight learning module enables adaptive adjustment of modality weights under different scenarios, allowing the system to better adapt to complex and ever-changing application environments and improving the model's generalization ability. The semantic enhancement matching module utilizes a knowledge graph embedding layer and semantic similarity calculation functions to significantly improve the accuracy of semantic matching, enabling more precise mining of semantic relationships between data. The distributed computing acceleration module effectively improves the system's computational efficiency and resource utilization through model parallelism, data sharding strategies, and an elastic resource scheduling submodule, supporting the processing of a large number of query requests per second. Simultaneously, the system also features incremental learning, interpretability analysis, and security and privacy protection functions, enhancing the system's stability, interpretability, and security, and meeting the diverse data processing needs of practical applications. Attached Figure Description

[0019] Figure 1 This is a schematic block diagram of the deep learning-based multi-source data view feature extraction and fast matching system proposed in this invention; Figure 2 A bar chart comparing the multimodal data preprocessing performance of traditional methods and our proposed method; Figure 3 Line graph comparing the deep feature extraction accuracy of traditional CNN and the multi-branch network of this system under different training epochs; Figure 4 This is a diagram showing the scalability of the distributed computing acceleration module in this system. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The present invention will now be described in further detail with reference to the accompanying drawings.

[0022] Reference Figures 1 to 4 A deep learning-based system for feature extraction and fast matching of multi-source data views includes the following modules: The multimodal data preprocessing module is capable of handling various heterogeneous data types, including images, point clouds, and text. For image data, it employs adaptive histogram equalization to dynamically adjust the histogram based on the grayscale distribution of local image regions, enhancing image contrast and making details clearer. For example, it effectively improves the visual quality of images taken in low-light environments. For point cloud data, it uses voxel downsampling to reduce data volume while preserving point cloud features, improving subsequent processing efficiency. For instance, in processing point cloud data from large-scale 3D scenes, the data volume can be reduced to one-third of its original size. For text data, it first performs lexical and syntactic analysis to remove stop words, then uses word vector technology to convert the text into numerical vectors for easier unified processing. Furthermore, it uses adaptive normalization to standardize image, point cloud, and text data from different modalities, ensuring a consistent data distribution and preparing for subsequent feature extraction.

[0023] Data access: Supports access to multiple data sources such as images (JPEG / PNG), point clouds (PCD format), and text (TXT / JSON). It parses image data, point cloud data, and text data structures of different modalities through a unified data interface. For example, image data is parsed into an RGB pixel matrix, point cloud data is parsed into a set of three-dimensional coordinates, and text data is parsed into a sequence of word vectors.

[0024] Adaptive normalization: using the formula Data standardization, including batch statistics , Calculations are performed in real time using a sliding window. The data has undergone adaptive normalization. The i-th data point in the original input data. It is a minimum value; it is added to prevent the variance from being too low during calculations. A value of 0 results in a denominator of 0. Empirical parameters γ and β serve as scaling and offset factors, respectively. Their core function is to adjust the distribution range of the normalized data. For example, γ=0.8 and β=1.2 can compress the data distribution to near the mean while maintaining a certain fluctuation range. For image data in low-light scenarios, γ can be increased to 1.0-1.5 and β adjusted to 0.8-1.0 to enhance contrast. For point cloud data after voxel downsampling, γ can be decreased to 0.5-0.7 and β set to 1.5-2.0 to preserve details. For text data, due to the sparsity of word vectors, γ should be kept at 0.8-1.0. β = 1.2-1.5 is used to maintain semantic structure; during dynamic adjustment, the distribution of statistical data can be statistically analyzed in real time through a sliding window. When the image grayscale variance exceeds a threshold such as 200, γ is automatically increased by 0.1 and β is decreased by 0.2. Grid search can also be used to optimize parameters for downstream tasks such as autonomous driving. Experiments show that the default parameters achieve a feature alignment accuracy of 89.3% on cross-modal data, while single-modal parameter optimization can improve performance by 5-8%. For example, in medical imaging scenarios, when γ = 1.2 and β = 0.9, the lesion feature extraction accuracy can be improved to 94.1%.

[0025] Multi-scale feature enhancement: Applying dilated convolution to image data Set the void ratio The three-branch structure captures detailed / intermediate-scale / global features respectively, and outputs enhanced features through concatenation and fusion. Indicates the first position in the output sequence The element at position 1 Represents the input sequence, Represents the input sequence The elements in It is the index of the current calculation position. It is an integer parameter that determines the interval at which elements of the input sequence are selected. It is a loop variable. Weighting coefficients are the coefficients multiplied by the elements of the input sequence. From 0 to The -1 variation corresponds to different weights at different positions, used to adjust the contribution of input sequence elements to the output. It is a positive integer that determines the number of terms in the summation operation, that is, the weights involved in the calculation and the number of elements in the input sequence. The summation starts from... =0 to = -1. The void ratio is... The values ​​of 1 / 3 / 5 are based on a balance between receptive field and feature resolution. =1 corresponds to standard convolution to preserve detailed features. =3 and =5 captures medium-scale and global context by expanding the sampling interval. Validated on the ImageNet dataset, this three-branch combination is superior to... =1 / 2 / 4 or With configurations of 2 / 4 / 6, the feature extraction accuracy in object detection tasks is improved by 4.2%. Furthermore, ablation experiments have confirmed that gradient-weighted fusion of the three hole rates can optimize the complementarity of multi-scale features, effectively avoiding feature blurring or loss of details caused by a single hole rate.

[0026] The deep feature extraction module constructs a unique multi-branch attention network architecture. The backbone network adopts an improved ResNet50, which optimizes the residual blocks based on the traditional ResNet50 and adds a weight adjustment mechanism for cross-layer connections, making information transmission more efficient. Specifically, a learnable weight matrix is ​​introduced into the cross-layer connection path of the residual block. The dimension of this matrix is ​​consistent with the number of feature channels. Through backpropagation, the features of cross-layer connections are dynamically optimized and weighted. For example, after the input feature is processed by residual mapping, it is multiplied by this weight matrix and then added to the original input to form the output feature. At the same time, a lightweight attention gating mechanism is embedded in the bottleneck structure. First, global average pooling is used to extract the overall information of the feature map in the channel dimension. Then, it is processed by two fully connected layers. The first layer compresses the number of channels to 1 / 8 of the original number of channels, and the second layer restores the original number of channels to generate an importance weight vector for each channel. After multiplying this vector by the identity matrix, the features of cross-layer connections are adaptively weighted by channel. Global average pooling is used to compress feature dimensions, fully connected layers are used to calculate channel weights, and RelU and Sigmoid activation functions are used to introduce non-linear and normalized weights, respectively. This mechanism introduces only a few additional parameters and achieves weight updates through end-to-end training, enhancing the flexibility of feature transfer while maintaining computational efficiency. A channel attention mechanism is also introduced, which analyzes each channel of the feature map, calculates the importance weight of each channel for the current task, and then adaptively adjusts the channels according to the weights. For example, in image classification tasks, it can highlight key feature channels related to the category and suppress irrelevant channels, enhancing feature representation capabilities. The branch networks include CNN for images, PointNet++ for point clouds, and BERT for text analysis. Each branch performs deep feature extraction on the data of its corresponding modality, and then the outputs of each branch are fused through a tensor fusion layer. During fusion, the attention mechanism is used to assign weights to the features of image data, point cloud data, and text data from different modalities, generating a joint feature representation and achieving deep fusion of cross-modal data.

[0027] Backbone network optimization: Improve the residual blocks of ResNet50 and introduce a channel attention mechanism in the bottleneck structure. Global average pooling (GAP) compresses the feature dimension, and two fully connected layers (dimension 1024 → 256 → number of channels) calculate the channel weights. It is the ReLU activation function. For the Sigmoid function, This represents the channel attention weight vector, used to measure the importance of each channel in the input feature map F. The weight matrix of the first fully connected layer (FC layer). The weight matrix of the second fully connected layer, F is the input feature map, which is the feature representation after processing the residual blocks in the backbone network. GAP(F) represents the global average pooling operation performed on the input feature map F, which compresses the two-dimensional features of each channel in the feature map F into a scalar.

[0028] Branch network design: Image branch: The CNN uses 5 layers of convolution (3x3 convolution kernel, stride 2) to extract spatial features and output a 256-dimensional feature vector.

[0029] Point cloud branch: PointNet++ outputs 1024-dimensional point cloud features through hierarchical sampling and feature aggregation.

[0030] Text branch: BERT-base extracts text semantics and outputs a 768-dimensional feature vector through pooling layers.

[0031] Tensor Fusion Layer: Using the formula Generate joint features, where attention weights Calculated through fully connected layers The outer product operation maps the features of image data, point cloud data, and text data from different modalities to a unified space. T represents the fusion result obtained through calculation, which is the output after fusing multiple feature maps. N is the number of feature map pairs participating in the fusion, a positive integer representing the upper limit of the number of terms in the summation operation. It is the i-th feature map. This is another corresponding feature map (i and j have a corresponding relationship, specifically determined by the model settings). They generally have the same or compatible dimensions (such as the number of channels, height, width, etc.). The fully connected layer adopts a two-layer structure, with 512 neurons in the first layer and 256 neurons in the second layer, both using ReLU as the activation function. The first layer maps the input cross-modal features to a high-dimensional space, while the second layer reduces the dimensionality to match the attention weight dimension. Through this structural design, the correlation weights between different modal features can be effectively learned. The introduction of the ReLU activation function enhances the non-linear expressive power of the model and makes the calculation of attention weights more flexible. The parameter scale of the two fully connected layers is moderate, which can ensure the learning ability of the model without significantly increasing the computational cost, making it easy to implement in practical engineering.

[0032] The fast matching and retrieval module employs a hierarchical hashing encoding strategy. First, a deep learning model learns the feature representation of the data, then maps these features into compact binary codes. To learn more discriminative codes, a triplet loss function is used, which ensures similar data are closer together in the encoding space, while dissimilar data are further apart. Based on this, a multi-index hash table is constructed with a hierarchical structure, where each layer is divided according to different data features. During the search, Hamming distance is used to quickly filter out candidate data similar to the target data, significantly improving retrieval speed by more than 5 times compared to traditional methods, meeting the needs of rapid retrieval of large-scale data.

[0033] Hierarchical hashing encoding: using a triplet loss function Learn binary encoding, where 'a' is the anchor sample, 'p' is a positive sample of the same class as 'a', and 'n' is a negative sample of a different class than 'a'. It is a feature extraction function. This represents the square of the Euclidean distance. It is to set the boundary parameters with the smallest positive and negative intervals. =0.5). This parameter is determined based on ablation experiments on standard datasets such as CIFAR-10 and ImageNet. At a hash value of 0.5, the discriminative power and retrieval accuracy of hash coding achieve an optimal balance, ensuring both tight clustering of similar samples in the coding space and maintaining sufficient distance between samples of different classes. For different modalities of data, when processing high-dimensional features such as images, [the following can be done]: Fine-tuning to 0.6-0.8 enhances feature discrimination, while handling semantically sparse modalities such as text. It is advisable to maintain a value of 0.4-0.5 to avoid excessively widening the distance between semantically similar samples. Specific adjustments can be made dynamically by optimizing the retrieval accuracy of the validation set.

[0034] Multi-index hash tables: Eight independent hash tables are constructed, each using a different random projection vector. The number of these eight independent hash tables is determined based on a synergistic optimization of data scale and modal features. When processing a 1TB dataset, the eight tables can map to the eight storage nodes of the data shards (each node handles approximately 128GB of data), enabling dynamic matching between hash retrieval and data storage location, reducing cross-node query overhead. For image, point cloud, and text modal data, the eight tables can map cross-modal features to differentiated hash spaces through different random projection vectors, reducing the probability of feature collisions. Experiments show that on a million-level multimodal dataset, the eight-table configuration improves the accuracy of Hamming distance retrieval by 11.7% and reduces retrieval latency by 23% compared to the four-table or 16-table schemes. This number of tables ensures the ability to distinguish high-dimensional features while avoiding storage redundancy caused by an excessive number of tables, achieving a balance between retrieval efficiency and resource consumption. (The last sentence about Hamming distance is unrelated and appears to be a separate, incomplete thought.) Perform a fast search, where b i and b j These represent two binary encoded vectors, used to represent information such as the characteristics of the data object. L represents the length of the binary encoded vector, that is, the number of elements in the encoding. and These represent binary encoded vectors b respectively. i and b j The kth element. It is an indicator function. Represents the binary encoded vector b i and b j The Hamming distance between them.

[0035] The dynamic weight learning module employs a meta-learning framework that rapidly and adaptively adjusts model parameters based on different task scenarios. Specifically, by learning from multiple different tasks, the meta-learner captures the commonalities and differences between tasks, enabling it to quickly adjust the model's initial parameters when facing new tasks, allowing it to converge to the optimal solution more quickly. Simultaneously, a context-aware mechanism is introduced. By analyzing the contextual information of the current data, such as its source and application scenario, the module calculates the importance weights of different modalities of image data, point cloud data, and text data in the current scenario. For example, in autonomous driving scenarios, the weights of image and point cloud data are dynamically adjusted according to different environments (e.g., day or night, city roads or highways), achieving optimal modality fusion in different scenarios.

[0036] Meta-learning framework: Employing the MAML algorithm The learning rate is updated in an inner loop across 5 tasks. =0.01, where the selection of 5 tasks follows the principle of balancing relevance and dissimilarity, prioritizing tasks with similar data distribution to the target task but different semantic scenarios (e.g., selecting urban roads, highways, and nighttime driving in autonomous driving scenarios). K-means clustering analysis is used to analyze the differences in feature distribution among tasks, ensuring that the selected tasks provide sufficient generalization information while avoiding redundancy. Experiments show that this strategy can improve the convergence speed of the meta-learning framework on new tasks by more than 3 times. The outer loop uses the Adam optimizer to update the meta-parameters. Where θ is the model's parameter vector, and θ′ is the model's parameter vector after one update. This represents the gradient operator with respect to the parameter θ. In the mission The loss function is used to measure the difference between the model's predicted value and the true value on the task. It is a model determined by the parameter θ, which processes the input data and produces the output.

[0037] Context-aware mechanism: through formulas Calculate the modal weights, where The similarity score is calculated by using cosine similarity between the current input and historical samples. Indicates the first The similarity score of each modality, where M represents the number of modalities involved in the calculation. It is an exponential function that performs exponential operations on the input value. This is a loop variable used to iterate through all modalities from 1 to M, performing a weighted summation operation. Similarity score. Before computation, L2 normalization is uniformly performed on feature vectors of different modalities to eliminate the influence of differences in feature magnitude. For features with inconsistent dimensions (such as 256-dimensional images and 1024-dimensional point clouds), a shared projection matrix W∈R is used. (1024×256) By mapping low-dimensional features to a unified high-dimensional space and then calculating cosine similarity, this method improves the accuracy of similarity calculation by 27.3% in cross-modal data testing. It is the first The weights of each modality are used to dynamically adjust the weight ratio of different modalities (image, point cloud, text) in the fusion process, so that the fusion is more in line with the characteristics of the current input.

[0038] The semantic enhancement matching module constructs a knowledge graph embedding layer, mapping entities and relationships to a low-dimensional vector space. Leveraging the rich semantic relationships within the knowledge graph, it provides richer semantic information for multi-source data. A semantic similarity calculation function is developed, comprehensively considering the textual semantics, structural semantics, and semantic associations within the knowledge graph. During calculation, an attention mechanism is used to assign weights to different semantic dimensions, thereby more accurately calculating the semantic similarity between data and improving the accuracy of semantic matching. For example, in intelligent question-answering systems, it can more accurately match questions and answers.

[0039] Knowledge graph embedding: using the TransE algorithm (in The embedding vector of the head entity represents the vector representation of the entity that serves as the starting point of a relation in the knowledge graph, with a dimension of 200. The embedding vector represents the relationship between the head entity and the tail entity, and its dimension is also 200. The embedding vector of the tail entity represents the vector representation of the entity that serves as the endpoint of a relation in the knowledge graph, with a dimension of 200. The embedding vectors of entities and relations are learned, with an embedding dimension of 200, a training batch size of 128, and a learning rate of 0.001, trained for 500 epochs on a Freebase subset. This set of parameters is optimized based on the graph structure characteristics of the Freebase subset. An embedding dimension of 200 performs optimally in experiments balancing semantic representation capability and computational cost, capturing complex relationships between entities while avoiding the curse of dimensionality. A training batch size of 128 is suitable for the node size of most knowledge graphs, achieving a good trade-off between memory usage and gradient stability. A learning rate of 0.001 has been verified as the optimal combination of convergence speed and accuracy in link prediction tasks through grid search. When dealing with large-scale graphs with more than 100,000 nodes, the embedding dimension can be increased to 300-500 to enhance representation capability. When there are more than 50 relation types, it is advisable to reduce the learning rate to 0.0005 and increase the batch size to 256, dynamically adjusting the parameters to match the graph complexity.

[0040] Semantic similarity calculation: via formula Calculation, where attention weights Generated by a multi-head attention mechanism, used to measure the first [head] in a query. The semantic vector and the document in the document The importance of the associations between semantic vectors and Let q represent the query content and d represent the document, respectively. Sim(q,d) represents the semantic similarity between query q and document d, and N and M are the number of semantic vectors in the query and document, respectively.

[0041] The distributed computing acceleration module employs a strategy combining model parallelism and data sharding. For model parallelism, the deep neural network is divided into layers and distributed across different computing nodes for parallel computation. Each node is responsible for the computational tasks of a specific layer. A parameter server architecture ensures gradient synchronization between nodes, guaranteeing consistency in model training. For data sharding, large-scale data is divided into blocks based on data characteristics or identifiers using a hash function and stored on different storage nodes. During computation, each computing node can directly access its locally stored data shards, reducing data transfer overhead and supporting over 100,000 query requests per second, effectively improving the system's processing capacity.

[0042] Model parallelism: dividing the network into layers Where N is the set of network layers, representing the set of all network layers obtained after dividing the entire network into layers. Gradients are allocated to 4 GPU nodes and synchronized via a parameter server architecture. Where P represents the number of GPU nodes, and here P=4. It is the gradient calculated by the i-th GPU node at step t; The total gradient is obtained after synchronization via the parameter server architecture at step t, and the parameters are updated using asynchronous SGD. Layer partitioning follows the principles of balanced computation and memory optimization. By analyzing the floating-point operations (FLOPs) and feature map memory usage of each layer, the network is divided into four modules proportionally. For example, convolutional layers and fully connected layers are dynamically allocated according to computational density to ensure that the computational load deviation of each GPU node does not exceed 15%. At the same time, when synchronizing gradients using the parameter server architecture, parameter updates of higher-dimensional layers are transmitted first to reduce cross-node communication overhead.

[0043] Data sharding: using hash functions The 1TB dataset is divided into blocks and stored across 8 nodes, m=8, with each node configured with 128GB of memory and NVMe SSD. Here, k is the identifier of the data element; m is the number of nodes; and h(k) is the result calculated using a hash function.

[0044] In this invention, the multimodal data preprocessing module further includes an adaptive noise suppression submodule. This submodule employs variational autoencoder (VAE) technology to address the noise characteristics of different modalities. For Gaussian noise in image data, the VAE learns the latent distribution of the image to generate a denoised image. During the learning process, reparameterization techniques are used to separate the randomness of the noise from the latent features of the image, thereby accurately removing the noise. After processing, the accuracy of subsequent feature extraction is ensured even in a -5dB signal-to-noise ratio environment.

[0045] Variational autoencoder: using the formula Learning data distribution, where, It is the input data; It is a latent variable; It is a multivariate normal distribution (Gaussian distribution), used for latent variables. Model the distribution. These are the parameters of the model; It is determined by parameters Definitely, based on input The obtained hidden variables The mean vector; Based on input The obtained hidden variables A vector composed of the variances of each dimension. This means converting the variance vector into a diagonal covariance matrix. Indicates that given input and parameters Below, latent variables Variational posterior distribution; through reparameterization techniques Generate clean samples, where μ is the mean of the latent variables and σ is the standard deviation vector. It is element-wise multiplication; where I is the identity matrix, with a latent variable dimension of 100.

[0046] Noise processing: In a -5dB signal-to-noise ratio environment, the weights of the reconstruction loss and KL divergence are adjusted ( =0.2), balancing noise reduction and information preservation.

[0047] In this invention, the deep feature extraction module further includes a cross-modal alignment submodule. This submodule constructs a contrastive learning framework that learns more discriminative cross-modal features by comparing the feature representations of similar and dissimilar samples in different modalities. Specifically, by maximizing the feature similarity of similar sample pairs and minimizing the feature similarity of dissimilar sample pairs, different modalities can be better aligned in the feature space. Simultaneously, bilinear pooling is used to fuse cross-modal features, generating more representative alignment features, thereby improving the semantic alignment accuracy of images and text and providing a more reliable feature foundation for subsequent cross-modal matching.

[0048] Contrastive learning framework: using formulas The batch size is 256, and the positive-to-negative sample ratio is 1:10. It is a contrastive learning loss function. and These are the feature vectors of positive sample pairs. It is the feature vector of other samples (including negative samples). It is a function that calculates the similarity between feature vectors. It is a temperature parameter.

[0049] Bilinear pooling: via the formula Where Bi-Vector is the generated aligned feature vector. BN is the batch normalization operation, ReLU is the modified linear unit activation function, and W1 and W2 are projection matrices of dimension 2048→512, used to adjust the input feature vector. and Perform a projection transformation.

[0050] In this invention, the fast matching and retrieval module further includes a hierarchical indexing submodule. This submodule constructs a hybrid index structure of KD-tree and hash table. The KD-tree is used to spatially partition the data, dividing the data space into multiple subspaces based on the data's feature dimensions, so that similar data are clustered in nearby subspaces. The hash table is used to quickly index the subspaces partitioned by the KD-tree, employing a locality-sensitive hashing algorithm to map similar data to the same or nearby hash buckets. During retrieval, the subspace that may contain the target data is first quickly located using the KD-tree, and then a precise search is performed in the corresponding hash bucket of that subspace. This achieves millisecond-level retrieval response on datasets of millions, greatly improving retrieval efficiency.

[0051] Hybrid index structure: Construct a 3-level KD tree, with each level containing 1000 data points per node. Each leaf node is associated with a hash table, employing the Locality Sensitive Hash (LSH) algorithm. ,in It is the input data vector. It is a random vector. It's the offset. =0.5 is a fixed parameter that affects the quantization interval of the hash function. It is the calculated hash value.

[0052] Search optimization: Use a priority queue to traverse the KD tree, and set the pruning threshold to 1.5 times the current minimum distance to reduce unnecessary hash table lookups.

[0053] In this invention, the dynamic weight learning module further includes an online meta-learning submodule. This submodule employs the MAML algorithm, enabling it to quickly adjust model parameters based on previously learned knowledge when a new task arrives. When facing a new scenario, the model can leverage experience learned from other relevant scenarios to quickly adapt to the new environment. Simultaneously, a weight decay mechanism is introduced to gradually reduce the influence of historical data on the current model weights over time, allowing the model to better adapt to changes in new data. Within five time steps after a scenario switch, the model converges to the optimal weight configuration, ensuring the system's performance stability in dynamic environments.

[0054] MAML optimization: using formulas The inner loop learning rate =0.01, outer loop learning rate =0.001, θ is the initial parameter of the model; These are the parameters optimized using MAML. It is the gradient operator with respect to the parameter θ. In the mission loss function on In the mission The loss function on. The model is determined by the parameter θ. This formula reflects the meta-learning parameter update process based on gradient descent when the model learns in parallel on multiple tasks (supporting 5 tasks).

[0055] Weight decay mechanism: through the formula Dynamically adjust the weights of historical data, among which These are the initial values ​​for the weights of historical data; In time step Weighting of historical data =0.05 is the attenuation coefficient. The time step is updated every 100 samples.

[0056] In this invention, the semantic enhancement matching module further includes a knowledge reasoning submodule. This submodule constructs a graph neural network (GNN) and utilizes the connections between nodes in the knowledge graph for knowledge reasoning. Through the message passing mechanism of the GNN, nodes can obtain information from their neighboring nodes and update their own feature representations. Based on this, a multi-hop reasoning mechanism is developed, allowing the model to perform multi-step reasoning in the knowledge graph. By repeatedly passing and aggregating information from neighboring nodes, deeper semantic relationships are uncovered. Supporting knowledge reasoning with more than 5 hops, it can handle complex semantic queries and provide more powerful reasoning capabilities for semantic enhancement matching.

[0057] Graph Neural Networks: Using Formulas To perform knowledge reasoning, among which It is a node In the Hidden state after layer update; It is a node In the The hidden state of the layer For the first Layer weight matrix. For nodes The neighborhood group, The normalization constant is It's an activation function that introduces non-linearity; here, the number of layers... =3, with 200 hidden dimensions per layer.

[0058] Multi-hop reasoning: through formulas Calculate the multi-hop reasoning score, where For query With the answer Multi-hop reasoning score, This refers to the search query content; The answer is yes, K=5 represents the number of hops and the hop weight. (Based on the characteristic that the transmission strength of semantic relationships in knowledge graphs decreases as the reasoning path lengthens, near-distance reasoning captures more direct semantic connections and therefore has a higher weight; while far-distance reasoning often involves potential indirect relationships and has a smaller impact on the result, thus its weight is significantly reduced. This allocation method has been experimentally verified to improve question-answering matching accuracy, reflecting a differentiated consideration of the semantic importance of different hop counts.) It is a query With the passing of the first Relevant representations of the answer after skipping reasoning The similarity.

[0059] In this invention, the distributed computing acceleration module further includes an elastic resource scheduling submodule. This submodule establishes a resource consumption prediction model, predicting the amount of resources required for computing tasks by analyzing the model's complexity (such as the number of network layers and parameters) and the amount of data. Based on the prediction results, a reinforcement learning strategy is used for dynamic resource allocation. This reinforcement learning algorithm employs the PPO (Proximal Policy Optimization) algorithm. The state space is defined as a multi-dimensional vector containing node CPU utilization, memory usage, task queue length, and the current resource allocation ratio. Node load data is obtained through periodic sampling, and task queue states are encoded according to priority levels. The action space is designed as a set of discrete operations for resource allocation adjustment, including proportional adjustments to the number of CPU cores, memory capacity, and storage bandwidth (e.g., ±10%, ±20% granularity). The policy function π(a|s) is parameterized through a neural network, with the cumulative reward value (task completion rate - resource waste cost) as the optimization objective. Through multiple rounds of environmental interaction training, the dynamic resource allocation strategy is optimized. Reinforcement learning agents learn optimal resource scheduling strategies by continuously interacting with the system environment. Based on the current system load and task requirements, they rationally allocate computing resources, thereby improving resource utilization and enhancing the overall operating efficiency of the system.

[0060] Resource prediction model: through formula Predict resource consumption, of which The model complexity is the predicted value of resource consumption. The amount of data is measured by the number of parameters and computational complexity. In MB, parameters =0.2, =0.05, =0.01, used to determine the weight of each factor on resource consumption.

[0061] Reinforcement learning scheduling: employing strategies ,in It is the optimal strategy. It is the state space and the action space. For resource allocation decisions, From strategy The generated trajectory, In time step In a state And take action The reward received at that time Indicating in strategy Generated trajectory Expectation value above, discount factor =0.9, The strategy for finding the maximum value .

[0062] This invention also includes an incremental learning module. This module employs an online gradient matching algorithm, which balances the learning effects of new and old data when introducing new data for learning, preserving the knowledge learned by the model from historical data. Simultaneously, a sample selection strategy is developed to select the most representative new samples for learning by evaluating sample uncertainty and redundancy. This effectively solves the catastrophic forgetting problem that easily occurs during incremental learning, ensuring that the model maintains good performance even with continuously updated data.

[0063] Gradient matching: using the formula Preserving historical knowledge, among which These are the optimal model parameters obtained after optimization using an online gradient matching algorithm. They are the parameter solutions that satisfy the condition of "minimizing the sum of squared weighted differences between the old and new models' gradients". The model parameters θ that minimize the objective function are selected and used to determine the optimal parameters during the optimization process. For old model parameters, Based on the current model parameters θ, the sample The corresponding loss function value; Based on old model parameters ,sample The corresponding loss function value. It is the input sample. These are the true labels corresponding to the samples. Batch size B=64. This represents the gradient with respect to the model parameter θ. The old model parameter θ is synchronized periodically. After every 500 training batches, the current parameter θ is copied to θ to ensure the stable storage of old knowledge. The weight coefficient is adjusted by a dynamic forgetting factor β∈[0,1]. The initial value is set to 0.8, and it decays exponentially with training iterations according to β=β×0.995, so that the gradient matching weight of the new task gradually increases and the influence of the old model gradient gradually weakens.

[0064] Sample selection: via formula Select samples, where S is the subset of samples to be selected, and D is the entire set of samples. This means finding, among all sets S that satisfy the condition S is a subset of D (S⊂D), the subset S that maximizes the subsequent function expression. Calculated from the predicted entropy, Based on kernel density estimation, This is the weighting coefficient (value 0.1). The prediction entropy calculation is based on the softmax output probability distribution of the classification task, i.e., calculating the information entropy for each class prediction probability vector of each sample to measure the uncertainty of the model's prediction for that sample. Kernel density estimation uses a Gaussian kernel function, controlling the smoothness of the sample space by adjusting the bandwidth parameter. This kernel function is computationally efficient and possesses good local smoothing properties, effectively capturing dense regions of data distribution in sample redundancy assessment. In practical applications, the optimal bandwidth parameter can be determined through cross-validation, making the redundancy assessment more closely reflect the data characteristics.

[0065] This invention also includes an interpretability analysis module. This module uses the SHAP (Shape-Up Point) method to quantify the contribution of each feature to the model's prediction results. By calculating the SHAP value, it is clear which features play a key role in the model's decision-making and the degree of their contribution. Furthermore, a counterfactual explanation generator is developed. By making minor changes to the input data, counterfactual samples are generated, and the changes in the model's prediction results are analyzed. This provides a more intuitive and understandable explanation for the model's decision-making, improving explanation accuracy and helping users understand the model's behavior and decision-making process.

[0066] SHAP value calculation: using the formula Quantify feature contribution, among which Features The SHAP value, where N is the set of all features. Let |S| be the contribution value of the feature subset S, and |N| be the number of elements in the feature subset S and the total number of elements in all feature sets N, respectively. It contains features The contribution value of the feature subset is approximated by the KernelSHAP method with 200 samplings. This sampling number is set based on a balance between feature dimension and computational efficiency. In scenarios where the feature dimension is ≤50, 200 samplings can ensure the accuracy of SHAP value calculation through Monte Carlo approximation. However, when the feature dimension is >50, a hierarchical adaptive sampling strategy is automatically activated—sampling is done in groups after sorting by feature importance, with each group sampling twice the square root of the dimension (e.g., 100-dimensional features are divided into 5 groups, with each group sampling 20 times). By aggregating the results through importance weighting, in ImageNet high-dimensional feature testing, this strategy reduces computation time by 40% compared to a fixed 200 samplings, while maintaining an interpretation accuracy error of <3%.

[0067] Counterfactual explanation: through formula Generate counterfactual samples, where It is the original sample. These are generated counterfactual samples, with a step size. =0.01, It is to calculate the original sample After model Output and Counterfactual Samples After model Output similarity. Indicates about The formula iterates (100 times) based on the gradient to update the original sample, generating counterfactual samples. The iteration terminates using a dual-judgment mechanism: when the number of iterations reaches 100 or the predicted similarity between the generated sample and the original sample reaches a certain threshold. The process terminates when the similarity exceeds 0.9. The similarity is calculated based on the softmax probability vector of the model's output layer, and gradient norm monitoring is introduced. If the gradient change is less than 1e-4 for 5 consecutive iterations, the process terminates early. This invention also includes a security and privacy protection module. This module employs a federated learning architecture, allowing multiple participants to collaboratively train the model without sharing the original data. Each participant trains the model locally and then uploads the model parameters to a central server for aggregation. To further protect data privacy, a differential privacy mechanism is deployed, adding appropriate noise to the data during model training and parameter aggregation, making it impossible to infer specific information from the original data from the model parameters. When the value is 0.5, the risk of data leakage is reduced, ensuring the security and privacy of data during sharing and use.

[0068] Federated learning: using formulas Aggregate model parameters, where Aggregated global model parameters For the client Sample size It is the first The client in the first The parameters of the local model after training are as follows. It is the sum of the sample sizes of all clients. This refers to the number of clients, with 10 local updates per training round. During actual aggregation, the quality of client data is first evaluated by comparing the distribution differences between client data and global data (e.g., analyzing the similarity of data features) and detecting label noise (e.g., identifying mislabels through cross-validation) to calculate a quality adjustment factor. Based on this, the weights of each client are adjusted—clients with large data distribution differences or a lot of noise will have their weights reduced proportionally; clients with excessively low data quality will have their aggregation weights reduced or even be temporarily excluded.

[0069] Differential privacy: through formula Noise is added, where Pr represents the probability, M(x) is the output after applying the privacy protection mechanism M to the input data x, and x′ is the data adjacent to x. S is a subset of the output space. It is a privacy budget parameter that controls the strength of privacy protection. It is a very small error parameter. A Gaussian mechanism is used. To add noise, among which It has a mean of 0 and a covariance matrix of A Gaussian distribution, where I is the identity matrix and the noise standard deviation is... ,in, It is a function The sensitivity of the feature extraction module. This sensitivity calculation employs a differentiated strategy based on module characteristics. The L2 norm is used to measure the maximum impact of a single sample input change on the output feature vector, while the matching retrieval module uses the number of changes in the result set caused by a single sample query as a sensitivity indicator. In practical applications, the standard deviation of Gaussian noise (e.g., ...) is dynamically adjusted according to the gradient norm in the feature extraction module. When σ=10, the retrieval module sets the noise parameters according to the sensitivity level (σ=5). When ≤5, σ=2. When σ > 5, σ = 3).

[0070] Reference Figure 1This document clearly presents the core module structure and interrelationship logic of a deep learning-based multi-source data view feature extraction and fast matching system, covering six core modules: multimodal data preprocessing, deep feature extraction, fast matching and retrieval, dynamic weight learning, semantic enhancement matching, and distributed computing acceleration. Each module clearly labels its core technologies and key functions, intuitively demonstrating the entire process architecture from data access and preprocessing, feature extraction and fusion, to matching and retrieval, and performance optimization. This provides a clear visual reference for understanding the overall system working mechanism and highlights the design logic of each module collaboratively supporting efficient multi-source data processing.

[0071] Reference Figure 2 The performance of traditional methods and the adaptive method of this invention in the preprocessing of three types of modal data—image, point cloud, and text—was compared (with latency as the evaluation metric). The charts clearly show that the adaptive preprocessing method of this invention has significantly lower latency than traditional methods across all three types of data, verifying the effectiveness of techniques such as adaptive normalization, multi-scale feature enhancement, and adaptive noise suppression. Quantitative analysis demonstrates that the preprocessing module of this system can significantly reduce processing latency while maintaining data quality.

[0072] Reference Figure 3 This study demonstrates the changing trends in deep feature extraction accuracy between traditional CNNs and our proposed multi-branch network across different training epochs. As the number of training epochs increases, the accuracy of our multi-branch network consistently surpasses that of traditional CNNs, with faster convergence speed and a significant final accuracy advantage. This also validates the superiority of our multi-branch attention network architecture (improved ResNet50 + channel attention mechanism + cross-modal alignment), proving its ability to extract the essential features of multimodal data more accurately.

[0073] Reference Figure 4 The study also presented the trend of increasing system training speed and changes in resource utilization as the number of nodes increased. As the number of nodes expanded from 1 to 16, the training speed continued to increase, while resource utilization remained at a high level without significant decline. This intuitively demonstrates the effectiveness of the model parallelism, data sharding, and elastic resource scheduling strategies in the distributed computing acceleration module, proving that the system has good scalability and can achieve a linear increase in processing efficiency by adding computing nodes, meeting the needs of large-scale, high-concurrency data processing.

[0074] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based multi-source data view feature extraction and fast matching system, characterized in that, Includes the following modules: Multimodal data preprocessing module: Image data adopts adaptive histogram equalization technology, which dynamically adjusts the histogram according to the local gray-level distribution of the image. For point cloud data, voxel downsampling method is used. Text data first undergoes lexical and syntactic analysis to remove stop words, and then word vector technology is used to convert the text into numerical vectors. Adaptive normalization method is used to standardize data of different modalities. Deep feature extraction module: The backbone network adopts an improved ResNet50 and introduces a channel attention mechanism to adaptively adjust the channels according to the importance weight of the task. The branch networks include CNN, PointNet++ and BERT, which extract features from the data respectively. Then, the outputs of each branch are fused through a tensor fusion layer to generate a joint feature representation. Fast matching and retrieval module: It adopts a hierarchical hash coding strategy. First, it extracts data features through a deep learning model, maps them into compact binary codes, and uses the triplet loss function to improve the coding discrimination. Based on this, a multi-index hash table is constructed. During the search, Hamming distance is used to quickly filter candidate data similar to the target data. Dynamic weight learning module: Through a meta-learning framework, a context-aware mechanism is introduced to calculate the importance weights of different modalities in the current scenario by analyzing the context information of the current data. Semantic Enhancement Matching Module: Constructs a knowledge graph embedding layer to map entities and relationships to a low-dimensional vector space. Through the semantic relationships of the knowledge graph, it develops a semantic similarity calculation function and uses an attention mechanism to weight different semantic dimensions during calculation. Distributed computing acceleration module: It adopts a combination of model parallelism and data sharding, and supports each computing node to directly access the data shards stored locally during computing.

2. The deep learning-based multi-source data view feature extraction and fast matching system according to claim 1, characterized in that, The multimodal data preprocessing module also includes an adaptive noise suppression submodule. This submodule uses variational autoencoder technology to address the noise characteristics in different modal data. For Gaussian noise in image data, the variational autoencoder learns the latent distribution of the image to generate a denoised image. During the learning process, the reparameterization technique is used to separate the randomness of noise from the latent features of the image.

3. The deep learning-based multi-source data view feature extraction and fast matching system according to claim 1, characterized in that, The deep feature extraction module also includes a cross-modal alignment submodule, which constructs a contrastive learning framework. By comparing the feature representations of similar and dissimilar samples in different modal data, cross-modal features are learned. Specifically, by maximizing the feature similarity of similar sample pairs and minimizing the feature similarity of dissimilar sample pairs, the data from different modalities are aligned in the feature space. At the same time, bilinear pooling is used to fuse the cross-modal features.

4. The deep learning-based multi-source data view feature extraction and fast matching system according to claim 1, characterized in that, The fast matching and retrieval module also includes a hierarchical indexing submodule, which constructs a hybrid index structure of KD-tree and hash table. The KD-tree divides the data space into multiple subspaces according to the feature dimensions of the data, so that similar data are clustered in nearby subspaces. The hash table is used to index the subspaces after the KD-tree is divided. The locality-sensitive hashing algorithm is used to map similar data to the same or nearby hash buckets. During retrieval, the subspace containing the target data is first located through the KD-tree, and then the search is performed in the hash bucket corresponding to that subspace.

5. The deep learning-based multi-source data view feature extraction and fast matching system according to claim 1, characterized in that, The dynamic weight learning module also includes an online meta-learning submodule, which uses the MAML algorithm to adjust the model parameters based on previously learned knowledge when a new task arrives. When facing a new scenario, the model uses the experience learned in other relevant scenarios to adapt to the new environment. At the same time, a weight decay mechanism is introduced to reduce the impact of historical data on the current model weights over time, so that the model can adapt to changes in new data.

6. The deep learning-based multi-source data view feature extraction and fast matching system according to claim 1, characterized in that, The semantic enhancement matching module also includes a knowledge reasoning submodule, which constructs a graph neural network (GNN) to perform knowledge reasoning by utilizing the connection relationships between nodes in the knowledge graph. Through the message passing mechanism of the GNN, nodes obtain information from their neighboring nodes and update their own feature representations. Based on this, a multi-hop reasoning mechanism is developed, which allows the model to perform multi-step reasoning in the knowledge graph. By passing and aggregating information from neighboring nodes multiple times, semantic relationships are mined.

7. The deep learning-based multi-source data view feature extraction and fast matching system according to claim 1, characterized in that, The distributed computing acceleration module also includes an elastic resource scheduling submodule. This submodule establishes a resource consumption prediction model. By analyzing the complexity of the model and the size of the data, it predicts the amount of resources required for the computing task. Based on the prediction results, it uses a reinforcement learning strategy to dynamically allocate resources. The reinforcement learning agent learns the optimal resource scheduling strategy by interacting with the system environment.

8. The deep learning-based multi-source data view feature extraction and fast matching system according to claim 1, characterized in that, It also includes an incremental learning module, which uses an online gradient matching algorithm to balance the learning effects of new and old data when introducing new data for learning, retaining the knowledge learned by the model on historical data. At the same time, it develops a sample selection strategy to select new samples for learning by evaluating the uncertainty and redundancy of samples.

9. The deep learning-based multi-source data view feature extraction and fast matching system according to claim 1, characterized in that, It also includes an interpretability analysis module, which uses the SHAP value method to quantify the contribution of each feature to the model's prediction results. By calculating the SHAP value, we can understand the role and degree of contribution of features in the model's decision-making. In addition, a counterfactual explanation generator is developed to generate counterfactual samples by making small changes to the input data and analyze the changes in the model's prediction results.

10. The deep learning-based multi-source data view feature extraction and fast matching system according to claim 1, characterized in that, It also includes a security and privacy protection module, which adopts a federated learning architecture that allows multiple participants to jointly train the model without sharing the original data. Each participant trains the model locally and then uploads the model parameters to a central server for aggregation. A differential privacy mechanism is deployed to add noise to the data during model training and parameter aggregation.

Images