Audio data search method and system based on deep neural network of big data

Abstract

The present application relates to a method and system for searching audio data using a deep neural network based on big data, which is applied to the field of audio search technology, and includes preprocessing the original audio data to obtain key features; extracting the key features based on a preset feature extraction algorithm to obtain audio features; constructing a graph structure based on the audio features, and establishing an approximate neighbor search algorithm based on the graph structure; searching the audio data to be searched based on the approximate neighbor search algorithm, and displaying the audio data to be searched and the search results through visualization technology. The present application combines graph embedding technology to improve the efficiency of large-scale audio data search, achieve efficient indexing and rapid retrieval of audio data, and increase the accuracy of the search.

Description

Technical Field

[0001] The present application relates to the field of audio search technology, and in particular to a method and system for constructing an audio data search method based on a deep neural network of big data. Background Art

[0002] With the rapid development of the Internet and multimedia technology, audio data has exploded. Traditional audio data search methods mainly rely on matching algorithms based on keywords or specific features, but these methods have obvious limitations when dealing with large-scale and diverse audio data. The development of big data technology and deep neural networks has brought new opportunities for audio data search, especially showing great potential in improving search efficiency and accuracy.

[0003] Traditional audio search methods perform poorly in handling the diversity and complexity of audio data, resulting in low search accuracy and relevance.

[0004] Based on this, there is a need to provide an audio data search method that increases the accuracy and relevance of search results. Summary of the invention

[0005] In order to improve the problem that traditional audio search methods perform poorly in processing the diversity and complexity of audio data, resulting in low accuracy and relevance of search results, the present application provides an audio data search method and system based on a deep neural network based on big data.

[0006] In the first aspect, the present application provides a method for searching audio data using a deep neural network based on big data, which adopts the following technical solutions: including:

[0007] Preprocess the raw audio data to obtain key features;

[0008] Extract the key features based on a preset feature extraction algorithm to obtain audio features;

[0009] Constructing a graph structure based on the audio features, and establishing an approximate nearest neighbor search algorithm based on the graph structure;

[0010] The audio data to be searched is searched based on the approximate nearest neighbor search algorithm, and the audio data to be searched and the search results are displayed through visualization technology.

[0011] Optionally, extracting the key features based on a preset feature extraction algorithm to obtain audio features further includes:

[0012] Train the preset self-supervised learning model according to the self-supervised learning loss function;

[0013] Generate a training data set according to the trained self-supervised learning model and the audio features;

[0014] Building a graph structure based on the audio features and the training data set;

[0015] The self-supervised learning loss function formula is as follows:

[0016] ;

[0017] in, refers to the self-supervised learning loss function, represents the number of samples, represents the number of categories or features, Indicates The sample in The actual labels on the categories; Indicates The sample in Predicted probabilities for the categories.

[0018] Optionally, constructing a graph structure based on the audio features and establishing an approximate neighbor search algorithm based on the graph structure includes:

[0019] Mapping the audio features into nodes in the graph structure;

[0020] Calculating the similarity between the audio features according to a preset similarity algorithm, and establishing the edges of the graph structure according to the similarity between the audio features;

[0021] Based on the graph embedding technology, the nodes and edges in the graph are converted into vector representations in a low-dimensional space to obtain embedded representations of the nodes in the graph structure;

[0022] The approximate nearest neighbor search algorithm is established based on the embedded representation.

[0023] Optionally, the similarity algorithm formula is as follows:

[0024] ;

[0025] in, Representing audio features and audio features The similarity between A feature vector representing audio features.

[0026] Optionally, after constructing a graph structure based on the audio features and establishing an approximate nearest neighbor search algorithm based on the graph structure, the method further includes:

[0027] The training data set is transmitted to the graph structure for indexing, and training features and labels are output, wherein the training features include: the distance between the search point and the query point, the distance of the k search points closest to the query point in the queue, and the distance from the nearest search point in the result queue to the query point after the current search ends, and the labels include: continue search and terminate search;

[0028] Transmitting the training features and labels to a preset prediction model for model training;

[0029] The audio search step of the graph structure is predicted based on the trained prediction model, and the audio search step is ended or continued according to the prediction result.

[0030] Optionally, after searching the audio data to be searched based on the approximate nearest neighbor search algorithm, the method further includes:

[0031] Use self-supervised learning algorithms to train deep learning models and collect user feedback based on reinforcement learning reward functions to dynamically adjust model parameters;

[0032] Training the deep learning model according to a preset search term recommendation loss function;

[0033] When a search term is received, the search term is recommended according to the deep learning model, and the recommendation result is dynamically optimized in combination with user behavior data;

[0034] Optionally, the reinforcement learning reward function formula is as follows:

[0035] ;

[0036] in, represents the total reward, represents the time step or number of decisions, Indicates that at time step Instant rewards;

[0037] The search term recommendation loss function formula is as follows:

[0038] ;

[0039] in, Indicates The actual label of a sample, usually 0 or 1, indicates whether the sample is the correct search term for recommendation. Indicates The predicted probability of a sample indicates the predicted value of the sample as a recommended search term.

[0040] In the second aspect, the present application provides an audio data search device based on a deep neural network of big data, which adopts the following technical solutions, including:

[0041] A data processing module is used to pre-process the original audio data to obtain key features;

[0042] A feature extraction module, used to extract the key features based on a preset feature extraction algorithm to obtain audio features;

[0043] An audio search module, used to construct a graph structure based on the audio features and establish an approximate nearest neighbor search algorithm based on the graph structure;

[0044] The search display module is used to search the audio data to be searched based on the approximate nearest neighbor search algorithm, and to display the audio data to be searched and the search results through visualization technology.

[0045] In a third aspect, the present application further provides a control device, the device comprising:

[0046] It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executes the above-mentioned audio data search method based on a deep neural network of big data.

[0047] In a fourth aspect, the present application also provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above for constructing an audio data search method based on a deep neural network based on big data.

[0048] In summary, in the present application, the system performs denoising, segmentation and other preprocessing on the original audio data to obtain key features, and the system then extracts the key features through a feature extraction algorithm to obtain audio features; then the system constructs a graph structure based on the audio features and establishes an approximate neighbor search algorithm based on the graph structure; when the system receives the audio to be searched, the system searches the audio data to be searched according to the graph structure of the approximate neighbor search algorithm, and recommends approximate search terms through a deep learning model, and finally predicts the search steps corresponding to the search terms through a prediction model to decide whether to continue or terminate the search in advance, and finally obtains the search results. The system displays the audio data to be searched, the search terms and the search results through visualization technology, thereby combining graph embedding technology to improve the efficiency of large-scale audio data search, so as to achieve efficient indexing and rapid retrieval of audio data, and increase the accuracy of the search, and at the same time use a deep learning model to recommend search terms, improve the relevance and accuracy of user queries, and decide whether to stop the search in advance based on the prediction model, which greatly saves search steps and reduces the occurrence of invalid searches. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flowchart of an audio data search method constructed using a deep neural network based on big data.

[0050] Figure 2 This is the overall process of the approximate nearest neighbor search method based on graph structure.

[0051] Figure 3 This is a flow chart of the query-aware dynamic feature early stopping strategy based on the prediction model.

[0052] Figure 4 It is a schematic diagram of the network sample architecture and prior constraints.

[0053] Figure 5 The present invention is a structural block diagram of an audio data search device constructed based on a deep neural network of big data.

[0054] Explanation of reference numerals: 510, data processing module; 520, feature extraction module; 530, audio search module; 540, search display module. DETAILED DESCRIPTION

[0055] The following combination Figure 1 - Figure 5 This application is described in further detail.

[0056] The current audio search technology in the industry has the following main problems:

[0057] Low search accuracy: Traditional audio search methods perform poorly in handling the diversity and complexity of audio data, resulting in low accuracy and relevance of search results.

[0058] Low processing efficiency: Faced with massive amounts of audio data, existing search systems have bottlenecks in processing speed and response time, making it difficult to meet the needs of real-time search.

[0059] Lack of intelligence and personalization: Most audio search systems fail to fully utilize users’ search history and behavior data, resulting in insufficient personalized recommendations and search experience.

[0060] Limited cross-modal search capabilities: Existing technologies mainly focus on single-modal data search and are unable to achieve the fusion and collaborative search of multi-modal data such as audio, text, and images.

[0061] Improvements or innovations to address the shortcomings of this point:

[0062] This application proposes a method for searching audio data using a deep neural network based on big data. The main innovations include:

[0063] Design of efficient algorithm based on graph approximate neighbor search architecture: Propose an approximate neighbor search algorithm based on graph structure to improve the efficiency of large-scale audio data search. Combined with graph embedding technology, it can achieve efficient indexing and fast retrieval of audio data.

[0064] Search term recommendation method based on deep learning: Use deep learning models to recommend search terms and improve the relevance and accuracy of user queries. Combine user behavior data to dynamically adjust and optimize recommendation results.

[0065] Research on neural network architecture search for large-scale graph data: Research and design neural network architecture for large-scale graph data, optimize audio data feature extraction and representation. Use neural architecture search (NAS) technology to automatically search and optimize deep neural network structures.

[0066] Research on audio data visualization methods: Develop audio data visualization technology to help users understand and analyze audio content more intuitively. Use graphics processing units (GPUs) to accelerate the visualization process and improve real-time and interactivity.

[0067] Based on this, the embodiment of the present application discloses a method for searching audio data by building a deep neural network based on big data. The execution subject is a control system. The system combines graph embedding technology to improve the efficiency of large-scale audio data search, so as to achieve efficient indexing and rapid retrieval of audio data. At the same time, it uses a deep learning model to recommend search terms, which improves the relevance and accuracy of user queries. In addition, it decides whether to stop the search in advance based on the prediction model, which greatly saves the search steps and reduces the occurrence of invalid searches.

[0068] Reference Figure 1 and Figure 2 , the embodiment of the present application at least includes step S10 to step S40.

[0069] S10, preprocessing the original audio data to obtain key features.

[0070] S20, extracting key features based on a preset feature extraction algorithm to obtain audio features.

[0071] S30, constructing a graph structure based on the audio features, and establishing an approximate neighbor search algorithm based on the graph structure.

[0072] S40, searching the audio data to be searched based on an approximate nearest neighbor search algorithm, and displaying the audio data to be searched and the search results through visualization technology.

[0073] Specifically, the system performs pre-processing such as denoising and segmentation on the original audio data to obtain key features. The system then extracts the key features through a feature extraction algorithm to obtain audio features. The system then constructs a graph structure based on the audio features and establishes an approximate neighbor search algorithm based on the graph structure.

[0074] When the system receives the audio to be searched, it searches the audio data according to the graph structure of the approximate nearest neighbor search algorithm, and recommends approximate search terms through the deep learning model. Finally, the prediction model is used to predict the search steps corresponding to the search terms to decide whether to continue the search or terminate it early. Finally, the search results are obtained. The system uses visualization technology to display the audio data to be searched, search terms, and search results, thereby combining graph embedding technology to improve the efficiency of large-scale audio data search, so as to achieve efficient indexing and fast retrieval of audio data.

[0075] In actual situations, audio features are extracted using methods such as Mel-frequency cepstral coefficients (MFCC) and spectrogram features, and high-dimensional feature embedding and representation is performed through deep learning models (such as convolutional neural networks CNN, recurrent neural networks RNN, self-attention mechanism Transformer, etc.). In this embodiment, convolutional neural network features are used to extract audio features, and the formula is as follows:

[0076] ;

[0077] in, : audio features, : Convolutional Neural Network, : Input audio data.

[0078] In some embodiments, considering that a large amount of audio data needs to be manually collected, the corresponding processing steps are as follows: training a preset self-supervised learning model according to a self-supervised learning loss function; generating a training data set according to the trained self-supervised learning model and audio features; constructing a graph structure based on the audio features and the training data set; the formula of the self-supervised learning loss function is as follows:

[0079] ;

[0080] in, refers to the self-supervised learning loss function, represents the number of samples, represents the number of categories or features, Indicates The sample in The actual labels on the categories; Indicates The sample in Predicted probabilities for the categories.

[0081] Specifically, the system uses self-supervised learning technology to generate training data sets, allowing the model to be trained in an unsupervised environment, greatly improving the richness of sample data and reducing dependence on manually collected audio samples. At the same time, the self-supervised learning model is trained through a self-supervised learning loss function, reducing dependence on manually labeled data and improving the model's generalization ability.

[0082] In some embodiments, step S30 specifically includes the following steps: mapping audio features into nodes in a graph structure; calculating the similarity between audio features according to a preset similarity algorithm, and establishing edges of the graph structure according to the similarity between audio features; converting the nodes and edges in the graph into vector representations in a low-dimensional space based on graph embedding technology to obtain embedded representations of the nodes in the graph structure; and establishing an approximate neighbor search algorithm based on the embedded representation.

[0083] The similarity algorithm formula is as follows:

[0084] ;

[0085] in, Representing audio features and audio features The similarity between A feature vector representing audio features.

[0086] It should be understood that the so-called distance and edge in this embodiment are concepts used to measure the similarity between two audio feature nodes. The greater the similarity, the shorter the edge and the smaller the distance.

[0087] Specifically, we propose an approximate nearest neighbor search algorithm based on a graph structure to improve the efficiency of large-scale audio data search. Combined with graph embedding technology, we can achieve efficient indexing and fast retrieval of audio data.

[0088] Reference Figure 2 and Figure 3 Further, in some embodiments, considering the problem of invalid search steps, the corresponding processing steps are as follows: the training data set is transmitted to the graph structure for indexing, and the training features and labels are output. The training features include: the distance between the search point and the query point, the distance of the k search points closest to the query point in the queue, and the distance from the nearest search point in the result queue to the query point after the current search ends. The labels include: continue the search and terminate the search; the training features and labels are transmitted to the preset prediction model for model training; the audio search step of the graph structure is predicted based on the trained prediction model, and the audio search step is terminated or continued according to the prediction result.

[0089] In practice, in order to effectively solve the redundancy problem in the search process, this embodiment proposes an early stopping strategy based on query-aware dynamic features. Based on the above analysis, the present invention can use the curve characteristics of D1 relative to D2 to identify specific areas in the search process. By analyzing the curve smoothness of D1 and D2 during the search process, the present invention designs a prediction model to decide whether to stop the search in advance. In addition, in order to eliminate the influence of inconsistencies between each query and its nearest neighbors, the present invention also considers the D3 feature. D1: The distance between the current search point and the query point in the current step; D2: The distance to the kth point closest to the query in the result queue after the current step ends; D3: The distance from the nearest point in the result queue to the query after the current step ends.

[0090] like Figure 3 As shown, no additional structure needs to be added to the graph index during the entire working process. In the training phase, the learning vector is used as the training set (which is also divided into a validation set). For each data in the training set, a complete search process is performed according to the above search algorithm, and the obtained dynamic features are used to train the prediction model. During the search process, the distance of each step (D1, D2, and D3) can be easily obtained as the training feature of the prediction model. Then, the minimum search step length is calculated and compared with the curve length to obtain the label. The label 1 means "continue the search" and the label 0 means "terminate the search". In the inference phase, the trained model is integrated into the graph index framework. Dynamic features are collected during the search process, and the remaining search steps are predicted by the trained prediction model, so that the prediction model can be used to decide whether to stop the search in advance, which greatly saves the search steps and reduces the occurrence of invalid searches.

[0091] In some embodiments, considering the problem of search term recommendation, the corresponding processing steps are as follows: use a self-supervised learning algorithm to train the deep learning model, and collect user feedback based on the reinforcement learning reward function to dynamically adjust the model parameters; train the deep learning model according to the preset search term recommendation loss function; when receiving the search term, recommend the search term according to the deep learning model, and dynamically optimize the recommendation results in combination with user behavior data.

[0092] The reinforcement learning reward function formula is as follows:

[0093] ;

[0094] in, represents the total reward, represents the time step or number of decisions, Indicates that at time step Instant rewards.

[0095] The formula for the search term recommendation loss function is as follows:

[0096] ;

[0097] in, Indicates The actual label of a sample, usually 0 or 1, indicates whether the sample is the correct search term for recommendation. Indicates The predicted probability of a sample indicates the predicted value of the sample as a recommended search term.

[0098] Specifically, we use self-supervised and unsupervised learning algorithms to train deep learning models, introduce reinforcement learning to optimize search strategies, and dynamically adjust model parameters based on user feedback. We use deep learning models to recommend search terms, and dynamically optimize recommendation results based on user behavior data.

[0099] Furthermore, the neural architecture search (NAS) technology is used to automatically optimize the deep neural network structure and improve the accuracy of audio data feature extraction and representation.

[0100] Neural Architecture Search: ;

[0101] Benefits: This formula represents the automatic optimization of the audio feature extraction model through the neural architecture search (NAS) technology, improving the accuracy and performance of the model.

[0102] The entire search architecture here includes the original search space and attention mechanism representation of NAS. In addition, prior constraints are introduced to reduce complexity and redundancy without losing expressive power.

[0103] Reference Figure 4 ,The search process brings great challenges and contains many redundant, meaningless or homogeneous network architectures.,This application mainly proposes three constraints on the network architecture to ensure that the expressiveness of the architecture is maintained while reducing the complexity.

[0104] Constraint 1: Skeleton Chain Structure

[0105] This application constrains the network architecture to be a chain (e.g. Figure 4 As shown in ① in the figure). Such connections are called skeletons because they are the main information flow of the entire network architecture. For the addition layer, this application only needs to bind the first input connection to the previous layer; for the attention layer, since the query in the attention will be used as a reference to re-extract the information in the key and value, this application binds the query connection to the previous layer, which is also a common design in the attention network manually designed by domain experts.

[0106] Constraint 2: Consistency of keys and values

[0107] This application binds the source layers of the key and value connections in the attention layer to the same layer (such as Figure 4 This is because in previous studies, keys and values ​​are always designed as memory modules, so the two should have similar semantics.

[0108] Constraint 3: Non-zero operation limit

[0109] Furthermore, this application restricts the selection of skeleton operations to non-zero, because if the skeleton operation selects zero operation, it is very likely that all the following layers will lose their expressiveness. At the same time, the connection with the attention layer is also restricted to non-zero to avoid meaningless attention mechanism construction. (e.g. Figure 4 (as shown in ③).

[0110] Through the above three constraints, the complexity of the search space can be significantly reduced and the search efficiency can be improved while ensuring the expressiveness of the network architecture.

[0111] In some embodiments, the content and features of the audio data are displayed through visualization technology, and real-time interaction is achieved using GPU acceleration.

[0112] Visual representation of audio data: ;

[0113] Benefits: This formula is used to visualize audio data, helping users to intuitively understand and analyze audio content.

[0114] The implementation principle of the method for constructing an audio data search method based on a deep neural network of big data in the embodiment of the present application is as follows: in the present application, the system performs pre-processing such as denoising and segmentation on the original audio data to obtain key features, and then the system extracts the key features through a feature extraction algorithm to obtain audio features; then the system constructs a graph structure according to the audio features, and establishes an approximate neighbor search algorithm based on the graph structure; when the system receives the audio to be searched, the system searches the audio data to be searched according to the graph structure of the approximate neighbor search algorithm, and recommends approximate search terms through a deep learning model, and finally predicts the search steps corresponding to the search terms through a prediction model to decide whether to continue or terminate the search in advance, and finally obtains the search results, and the system displays the audio data to be searched, the search terms and the search results through visualization technology, thereby combining the graph embedding technology to improve the efficiency of large-scale audio data search, so as to achieve efficient indexing and rapid retrieval of audio data, and at the same time uses a deep learning model to recommend search terms, which improves the relevance and accuracy of user queries, and decides whether to stop the search in advance according to the prediction model, which greatly saves search steps and reduces the occurrence of invalid searches.

[0115] Figure 1 FIG. 1 is a flow chart of a method for constructing an audio data search method based on a deep neural network of big data in one embodiment. It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but the steps are not necessarily executed in the order indicated by the arrows; unless otherwise specified in this document, there is no strict order restriction for the execution of the steps, and the steps may be executed in other orders; and Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0116] Based on the same technical concept, Figure 5 The embodiment of the present application also provides an audio data search device based on a deep neural network of big data, which adopts the following technical solution. The device includes:

[0117] The data processing module 510 is used to pre-process the original audio data to obtain key features;

[0118] A feature extraction module 520 is used to extract key features based on a preset feature extraction algorithm to obtain audio features;

[0119] An audio search module 530, configured to construct a graph structure based on audio features and establish an approximate neighbor search algorithm based on the graph structure;

[0120] The search display module 540 is used to search the audio data to be searched based on the approximate nearest neighbor search algorithm, and display the audio data to be searched and the search results through visualization technology.

[0121] In some embodiments, the feature extraction module 520 is further used to train a preset self-supervised learning model according to the self-supervised learning loss function;

[0122] Generate a training data set based on the trained self-supervised learning model and audio features;

[0123] Build a graph structure based on audio features and training datasets;

[0124] The formula for the self-supervised learning loss function is as follows:

[0125] ;

[0126] in, refers to the self-supervised learning loss function, represents the number of samples, represents the number of categories or features, Indicates The sample in The actual labels on the categories; Indicates The sample in Predicted probabilities for the categories.

[0127] In some embodiments, the audio search module 530 is specifically used to map audio features to nodes in a graph structure;

[0128] Calculate the similarity between the audio features according to a preset similarity algorithm, and establish edges of the graph structure according to the similarity between the audio features;

[0129] Based on graph embedding technology, the nodes and edges in the graph are converted into vector representations in a low-dimensional space to obtain the embedded representation of the nodes in the graph structure;

[0130] Establish an approximate nearest neighbor search algorithm based on embedded representation.

[0131] In some embodiments, the similarity algorithm formula is as follows:

[0132] ;

[0133] in, Representing audio features and audio features The similarity between A feature vector representing audio features.

[0134] In some embodiments, the audio search module 530 is further used to transmit the training data set to the graph structure for indexing, and output training features and labels, the training features include: the distance between the search point and the query point, the distance of the k search points closest to the query point in the queue, and the distance from the nearest search point in the result queue to the query point after the current search ends, and the labels include: continue searching and terminate searching;

[0135] The training features and labels are sent to the preset prediction model for model training;

[0136] The audio search steps of the graph structure are predicted based on the trained prediction model, and the audio search steps are terminated or continued according to the prediction results.

[0137] In some embodiments, the audio search module 530 is further used to train the deep learning model using a self-supervised learning algorithm and dynamically adjust the model parameters by collecting user feedback based on a reinforcement learning reward function;

[0138] The deep learning model is trained according to the preset search term recommendation loss function;

[0139] When receiving a search term, it recommends the search term based on the deep learning model and dynamically optimizes the recommendation results based on user behavior data;

[0140] In some embodiments, the reinforcement learning reward function formula is as follows:

[0141] ;

[0142] in, represents the total reward, represents the time step or number of decisions, Indicates that at time step Instant rewards;

[0143] The formula for the search term recommendation loss function is as follows:

[0144] ;

[0145] in, Indicates The actual label of a sample, usually 0 or 1, indicates whether the sample is the correct search term for recommendation. Indicates The predicted probability of a sample indicates the predicted value of the sample as a recommended search term.

[0146] The embodiment of the present application also discloses a control device.

[0147] Specifically, the control device includes a memory and a processor, and the memory stores a computer program that can be loaded by the processor and execute the above-mentioned audio data search method constructed by a deep neural network based on big data.

[0148] The embodiment of the present application also discloses a computer-readable storage medium.

[0149] Specifically, the computer-readable storage medium stores a computer program that can be loaded by a processor and execute the above-mentioned audio data search method based on a deep neural network of big data. The computer-readable storage medium includes, for example: a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes.

[0150] The above are all preferred embodiments of the present application, and the protection scope of the present application is not limited thereto. Therefore, any equivalent changes made according to the structure, shape, and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for searching audio data using a deep neural network based on big data, characterized in that: include: Preprocess the raw audio data to obtain key features; Extract the key features based on a preset feature extraction algorithm to obtain audio features; Constructing a graph structure based on the audio features, and establishing an approximate nearest neighbor search algorithm based on the graph structure; Searching the audio data to be searched based on the approximate nearest neighbor search algorithm, and displaying the audio data to be searched and the search results through visualization technology; Wherein, extracting the key features based on the preset feature extraction algorithm to obtain the audio features also includes: training the preset self-supervised learning model according to the self-supervised learning loss function; Generate a training data set according to the trained self-supervised learning model and the audio features; Building a graph structure based on the audio features and the training data set; The self-supervised learning loss function formula is as follows: ； in, refers to the self-supervised learning loss function, represents the number of samples, represents the number of categories or features, Indicates The sample in The actual labels on the categories; Indicates The sample in Predicted probabilities on categories; The step of constructing a graph structure based on the audio features and establishing an approximate nearest neighbor search algorithm based on the graph structure includes: Mapping the audio features into nodes in the graph structure; Calculating the similarity between the audio features according to a preset similarity algorithm, and establishing the edges of the graph structure according to the similarity between the audio features; Based on the graph embedding technology, the nodes and edges in the graph are converted into vector representations in a low-dimensional space to obtain embedded representations of the nodes in the graph structure; Establishing the approximate nearest neighbor search algorithm based on the embedded representation; The similarity algorithm formula is as follows: ； in, Representing audio features and audio features The similarity between Representing audio features The characteristic vector of Representing audio features The eigenvector of Wherein, after constructing a graph structure based on the audio features and establishing an approximate nearest neighbor search algorithm based on the graph structure, the method further includes: The training data set is transmitted to the graph structure for indexing, and training features and labels are output, wherein the training features include: the distance between the search point and the query point, the distance of the k search points closest to the query point in the queue, and the distance from the nearest search point in the result queue to the query point after the current search ends, and the labels include: continue search and terminate search; Transmitting the training features and labels to a preset prediction model for model training; The audio search step of the graph structure is predicted based on the trained prediction model, and the audio search step is ended or continued according to the prediction result.

2. According to claim 1, a method for searching audio data using a deep neural network based on big data is characterized in that: After searching the audio data to be searched based on the approximate nearest neighbor search algorithm, the method further includes: Use self-supervised learning algorithms to train deep learning models and collect user feedback based on reinforcement learning reward functions to dynamically adjust model parameters; Training the deep learning model according to a preset search term recommendation loss function; When a search term is received, a search term recommendation is performed based on the deep learning model, and the recommendation result is dynamically optimized in combination with user behavior data.

3. According to claim 2, a method for searching audio data using a deep neural network based on big data is characterized in that: The search term recommendation loss function formula is as follows: ； in, Indicates The actual label of a sample, usually 0 or 1, indicates whether the sample is the correct search term for recommendation. Indicates The predicted probability of a sample indicates the predicted value of the sample as a recommended search term.

4. An audio data search device based on a deep neural network of big data, characterized in that: The device comprises: A data processing module is used to pre-process the original audio data to obtain key features; A feature extraction module, used to extract the key features based on a preset feature extraction algorithm to obtain audio features; An audio search module, used to construct a graph structure based on the audio features and establish an approximate nearest neighbor search algorithm based on the graph structure; A search display module, used to search the audio data to be searched based on the approximate nearest neighbor search algorithm, and display the audio data to be searched and the search results through visualization technology; The feature extraction module is also used to: train a preset self-supervised learning model according to a self-supervised learning loss function; Generate a training data set according to the trained self-supervised learning model and the audio features; Building a graph structure based on the audio features and the training data set; The self-supervised learning loss function formula is as follows: ； in, refers to the self-supervised learning loss function, represents the number of samples, represents the number of categories or features, Indicates The sample in The actual labels on the categories; Indicates The sample in Predicted probabilities on categories; The audio search module is specifically used to map the audio features into nodes in the graph structure; Calculating the similarity between the audio features according to a preset similarity algorithm, and establishing the edges of the graph structure according to the similarity between the audio features; Based on the graph embedding technology, the nodes and edges in the graph are converted into vector representations in a low-dimensional space to obtain embedded representations of the nodes in the graph structure; Establishing the approximate nearest neighbor search algorithm based on the embedded representation; The similarity algorithm formula is as follows: ； in, Representing audio features and audio features The similarity between Representing audio features The characteristic vector of Representing audio features The eigenvector of The audio search module is also used to transmit the training data set to the graph structure for indexing, and output training features and labels, wherein the training features include: the distance between the search point and the query point, the distance of the k search points closest to the query point in the queue, and the distance from the nearest search point in the result queue to the query point after the current search ends, and the labels include: continue search and terminate search; Transmitting the training features and labels to a preset prediction model for model training; The audio search step of the graph structure is predicted based on the trained prediction model, and the audio search step is ended or continued according to the prediction result.

5. A control device, characterized in that: The device comprises: The method comprises a memory and a processor, wherein the memory stores a computer program which can be loaded by the processor and execute the method according to any one of claims 1 to 3.

6. A computer-readable storage medium, characterized in that: A computer program is stored which can be loaded by a processor and execute the method as claimed in any one of claims 1 to 3.

Images