Data report generation method based on deep learning

Abstract

The invention relates to the technical field of report generation, and discloses a data report generation method based on deep learning, which comprises the following steps: S1, collecting structured data, unstructured data and multi-modal data, cleaning and preprocessing different types of data, and using a uniform data representation mode; and S2, performing feature extraction on different data modals by using a multi-modal deep learning model, fusing multi-modal features, and performing comprehensive processing on the features through a deep neural network or a self-attention mechanism. By introducing a multi-modal deep learning model, the system can better process and fuse structured data, non-structured data and multi-modal data, which means that the generated report can more comprehensively and accurately reflect information from different data sources, and provides more insight analysis, and the feature extraction and fusion technology can improve the accuracy of the system. The system can capture more detailed and important features, so that report generation is more accurate and customized.

Description

Technical Field

[0001] The present invention relates to the technical field of report generation, and in particular to a data report generation method based on deep learning. Background Art

[0002] With the rapid development and widespread application of data technology, data report generation has become an important tool in areas such as corporate decision-making, scientific research analysis, and market forecasting. Traditional data report generation methods rely on manual writing or rule-based systems. These methods have significant limitations when processing complex and diverse data, including the following:

[0003] Existing data report generation systems usually process structured data, but have limited processing capabilities for unstructured data and multimodal data, resulting in the inability of generated reports to fully reflect the information contained in different types of data and to provide comprehensive and accurate analysis.

[0004] Current report generation systems rely on fixed templates and rules, lack the ability to dynamically perceive context, and are unable to accurately adjust content based on domain-specific knowledge and historical trends, and are unable to effectively meet the needs of specific domains;

[0005] The existing report generation system lacks the ability of adaptive learning and online learning. The model is deployed in the production environment and cannot be effectively adjusted dynamically according to new data and new tasks, and cannot cope with changes in data distribution, resulting in reduced report quality and stability.

[0006] Therefore, those skilled in the art provide a data report generation method based on deep learning to solve the problems raised in the above background technology. Summary of the invention

[0007] In view of the shortcomings of the prior art, the present invention provides a data report generation method based on deep learning to solve the problems raised in the above background technology.

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions: A data report generation method based on deep learning, comprising the following steps:

[0009] S1. Collect structured data, unstructured data and multimodal data, and clean and preprocess different types of data to use a unified data representation method;

[0010] S2. Use multimodal deep learning models to extract features from different data modalities, fuse multimodal features, and comprehensively process the features through deep neural networks or self-attention mechanisms to select the most representative features for report generation;

[0011] S3. Use knowledge enhancement models to integrate domain-specific knowledge bases, industry rules, and historical data trends into feature processing, and dynamically adjust report generation content through context-aware mechanisms so that relevant reports can be generated based on data context;

[0012] S4. Deploy deep learning models that support adaptive learning, so that the model can adapt to new data and new tasks in real time, and use online learning mechanisms and user feedback to continuously adjust model weights and generation strategies;

[0013] S5. Design an interactive generation framework that allows users to provide real-time feedback during the report generation process, so as to combine the reinforcement learning mechanism and dynamically optimize the generation model using user feedback;

[0014] S6. Combine generative adversarial networks or variational autoencoder generation models in the report generation process to explore innovative content generation;

[0015] S7. Use multilingual pre-trained models to process multilingual data and dynamically adjust the expression of reports according to different cultural backgrounds;

[0016] S8. Use multiple evaluation metrics to evaluate the quality of reports and measure model performance by comparing with manually generated reports to collect user feedback and continuously iterate and optimize the model.

[0017] S9. Deploy the optimized model to the production environment, support automatic and interactive report generation, and combine it with API or Web services. At the same time, regularly monitor the performance of the model in actual applications and update the model in a timely manner according to new data and changes in demand.

[0018] Preferably, in the step S1, data of different scales are standardized by a Z-score standardization formula, and the Z-score standardization formula is as follows:

[0019]

[0020] Where x is the original data point, μ is the mean of the data set, σ is the standard deviation of the data set, and x′ is the standardized data;

[0021] In the step S1, when processing unstructured text data through the Skip-gram model, words are converted into high-dimensional vectors, and the objective function is:

[0022]

[0023] Where T is the number of words in the text, c is the size of the context window, and w t is the current word, w t+j is a word in the context, P(w t+j ∣wt ) is a given word w t In the case of generating word w t+j The probability is usually expressed using the Soft-max function:

[0024]

[0025] Among them, w O and w I are the word vectors of the output and input words, and v is the size of the vocabulary;

[0026] In the step S1, when processing image data, feature extraction is performed through a convolutional neural network, and the formula for the convolution operation is:

[0027]

[0028] Among them, x(i,j) is the pixel value of the input image, k(m,n) is the weight of the convolution kernel, y(i,j) is the output value after the convolution operation, and M and N are the sizes of the convolution kernel.

[0029] Preferably, in the step S2, in multimodal deep learning, data of different modalities are mapped into the same feature space by joint embedding space, and the algorithm formula is:

[0030] h t =f text (x t ),

[0031] h i =f image (x i ),

[0032] Among them, x t and x i Represents the input data of text and image, f text and f image It is a deep neural network used to extract text and image features. t and h i is the feature vector extracted from text and image;

[0033] In the S2 step, feature fusion is achieved by concatenating, weighted averaging or fusing features of different modalities through an attention mechanism. The algorithm formula is:

[0034] h fusion =α·h t +β·h i ,

[0035] Among them, h fusion is the fused feature representation, α and β are the fusion weights;

[0036] In the step S2, the feature representation is adjusted by calculating the correlation between different features. The formula is as follows:

[0037]

[0038] Where Q is the query matrix, K is the key matrix, V is the value matrix, and d k is the dimension of the key vector and softmax is the activation function used to calculate the attention weights.

[0039] Preferably, in the step S3, the knowledge enhancement model enhances text representation by introducing entity information in the knowledge graph based on BERT, and its algorithm formula is:

[0040]

[0041] in, is the input text representation vector, is the entity enhancement vector obtained by knowledge graph embedding, is the enhanced text representation vector;

[0042] In the S3 step, the dynamic context-aware mechanism enables the model to dynamically adjust the output according to the context, and its formula is as follows:

[0043]

[0044] Where Q is the query matrix, K is the key matrix, V is the value matrix, B is the bias matrix used to encode the context, and d k is the dimension of the key vector, softmax is used to calculate the attention weight, and ContextVec represents the context vector;

[0045] In the S3 step, in order to incorporate the historical data trend into the generation process, a time series model or a recurrent neural network can be used to capture the temporal dependency, and the formula is as follows:

[0046] h t =σ(W h x t +U h h t-1 +b h ),

[0047] Among them, x t is the input feature vector at time t, h t is the hidden state vector at time t, h t-1 is the hidden state at the previous time, W h and U h is the weight matrix, b h is the bias vector.

[0048] Preferably, in step S4, the learning rate is dynamically adjusted by an adaptive learning algorithm in deep learning, and the algorithm formula is:

[0049] m t =β1m t-1 +(1-β1)g t ,

[0050]

[0051] Among them, g t is the gradient at time step t, m t and v t are the first and second order momentum of the gradient, β1 and β2 are the decay rates, and is the bias-corrected momentum, θ t is the parameter vector, α is the learning rate, and ∈ is a decimal;

[0052] In the S4 step, online learning usually involves updating model parameters sample by sample or in small batches, and its algorithm formula is:

[0053]

[0054] Among them, θ t is the parameter vector of the model at time step t, η is the learning rate, In the sample (x t ,y t ) is calculated on the parameter θ t The gradient of

[0055] In the step S4, in reinforcement learning, user feedback can be used as a reward signal to optimize the strategy. The formula of the policy gradient method is:

[0056]

[0057] Among them, π θ (a t ∣s t ) is in state s t Take action a t The strategy probability, R(τ) is the total return of trajectory τ, is the policy gradient, which is used to adjust the model parameters θ.

[0058] Preferably, in the step S5, in reinforcement learning, the strategy of generating the model can be optimized by PPO, and its algorithm formula is:

[0059]

[0060] in, is the strategy ratio, which measures the difference between the current strategy and the old strategy, is the advantage function, ∈ is a hyperparameter;

[0061] In the step S5, in the interactive generation model, the generative adversarial network can be used to generate high-quality report content, and the loss function of the discriminator in the generative adversarial network is defined as:

[0062]

[0063] Among them, D(x) is the output of the discriminator, and G(z) is the output of the generator;

[0064] In the step S5, in interactive generation, the uncertainty of the model can be used to dynamically decide whether to use the strategy. The exploration strategy based on uncertainty can be implemented by calculating the information gain:

[0065]

[0066] Among them, Ψ(θ; x) represents the information gain of parameter θ after given data x, is the posterior distribution of the parameters under the existing data D, is the parameter posterior distribution after incorporating the new data x.

[0067] Preferably, in step S6, the generative adversarial network is composed of a generator G and a discriminator D, which are mutually optimized through a game process. The loss function of the generative adversarial network is usually defined as:

[0068]

[0069] Among them, D(x) is the output of the discriminator, which represents the probability that the input x comes from the real data, G(z) is the sample generated by the generator from the noise z, and p lats (x) is the distribution of real data, p z (z) is the noise distribution;

[0070] In step S6, the conditional generative adversarial network introduces additional conditional information in the generation process, and the loss function of the conditional generative adversarial network is:

[0071]

[0072] Among them, y is the conditional information, D(x|y) represents the probability of judging whether the input x is real data under condition y, and G(z|y) represents the sample generated by the generator under condition y.

[0073] Preferably, in step S7, self-supervised learning technology is widely used in multi-language pre-training models, and its loss function is:

[0074]

[0075] Among them, P(w i ∣w 1:i-1 ,w i+1:n ) is the model in a given context w 1:i-1 and w i+1:n In the case of i probability;

[0076] In step S7, in order to achieve vocabulary alignment between different languages, Procrustes analysis can be used to align the embedding space, and the formula is:

[0077]

[0078] Among them, X and Y are the word embedding matrices of the two languages, and R is the rotation matrix used for alignment.

[0079] Preferably, in step S8, in order to improve the robustness of the model under adversarial samples, a gradient penalty term may be added to the loss function. The gradient penalty formula in adversarial training is:

[0080]

[0081] Among them, λ is the weight of the gradient penalty, is the interpolation sample between the generated data and the real data, D Is the critic in the sample Output on;

[0082] In step S8, multi-task learning allows the model to be trained on multiple related tasks to improve overall performance by sharing representations. Assume that there are N groups of tasks and the loss of each group of tasks is The total loss is:

[0083]

[0084] Among them, λ i is the weight of task i.

[0085] Preferably, in step S9, in order to process the concept drift in the data stream, a sliding window method can be used to update the model. Assuming that the window size is W, when new data arrives, the window update formula is:

[0086]

[0087] in, is the new subset of data within the window.

[0088] The present invention provides a data report generation method based on deep learning. It has the following beneficial effects:

[0089] 1. By introducing a multimodal deep learning model, the present invention enables the system to better process and fuse structured data, unstructured data and multimodal data, which means that the generated reports can more comprehensively and accurately reflect information from different data sources and provide more insightful analysis. The feature extraction and fusion technology enables the system to capture more detailed and important features, making report generation more accurate and customized.

[0090] 2. Through the introduction of context-aware mechanism and knowledge enhancement model, the present invention can dynamically adjust the content of report generation to make it more in line with domain-specific knowledge and historical trends, improve the relevance and accuracy of report content, and make the generated report more in line with actual needs and application scenarios. Combined with the domain knowledge base and industry rules, the report content can be more in-depth and professional, thereby enhancing the credibility and application value of the report.

[0091] 3. The present invention introduces adaptive learning and online learning mechanisms, so that the model can be dynamically adjusted and optimized according to the latest data and tasks, which means that the system can better cope with changes in data distribution and maintain the high quality and stability of report generation. By combining the interactive generation framework and reinforcement learning mechanism of user feedback, the system can continuously adjust the generation strategy according to the real-time feedback of users, thereby improving user satisfaction and the personalization of reports. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0093] In order to make the technical personnel in the technical field understand the scheme of the present invention, the technical scheme in the embodiment of the present invention will be clearly and completely described below in combination with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a partial embodiment of the present invention, not a complete embodiment. Based on the embodiment of the present invention, other embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present invention.

[0094] The present invention is described in detail below in conjunction with the accompanying drawings:

[0095] Example:

[0096] Please refer to the attached Figure 1 , an embodiment of the present invention provides a data report generation method based on deep learning, comprising the following steps: S1, collecting structured data, unstructured data and multimodal data, and cleaning and preprocessing different types of data to use a unified data representation method;

[0097] S2. Use multimodal deep learning models to extract features from different data modalities, fuse multimodal features, and comprehensively process the features through deep neural networks or self-attention mechanisms to select the most representative features for report generation;

[0098] S3. Use knowledge enhancement models to integrate domain-specific knowledge bases, industry rules, and historical data trends into feature processing, and dynamically adjust report generation content through context-aware mechanisms so that relevant reports can be generated based on data context;

[0099] S4. Deploy deep learning models that support adaptive learning, so that the model can adapt to new data and new tasks in real time, and use online learning mechanisms and user feedback to continuously adjust model weights and generation strategies;

[0100] S5. Design an interactive generation framework that allows users to provide real-time feedback during the report generation process, so as to combine the reinforcement learning mechanism and dynamically optimize the generation model using user feedback;

[0101] S6. Combine generative adversarial networks or variational autoencoder generation models in the report generation process to explore innovative content generation;

[0102] S7. Use multilingual pre-trained models to process multilingual data and dynamically adjust the expression of reports according to different cultural backgrounds;

[0103] S8. Use multiple evaluation metrics to evaluate the quality of reports and measure model performance by comparing with manually generated reports to collect user feedback and continuously iterate and optimize the model.

[0104] S9. Deploy the optimized model to the production environment, support automatic and interactive report generation, and combine it with API or Web services. At the same time, regularly monitor the performance of the model in actual applications and update the model in a timely manner according to new data and changes in demand.

[0105] The benefit of S1 is to ensure that different types of data can be represented in a unified manner, eliminate noise and inconsistency in the data, lay a solid foundation for subsequent feature extraction and model training, provide more comprehensive input information, and make the generated reports more comprehensive and accurate; the benefit of S2 is to improve the analytical depth and insight of the report; the benefit of S3 is to enhance the accuracy and professionalism of the report and improve the practicality of the report; the benefit of S4 is to improve the personalization of the report and user satisfaction; the benefit of S5 is that the system can dynamically adjust the generation strategy based on user feedback to improve the personalization and quality of the report; the benefit of S6 is to avoid template generation, provide new perspectives and insights, and make the generated reports more attractive and unique; the benefit of S7 is that the report content can be adjusted according to the cultural background to improve the acceptance and accuracy of the report in different cultural environments; the benefit of S8 is to collect user feedback and continuous optimization mechanism to ensure that the model performance continues to improve over time and the report quality is continuously improved; the benefit of S9 is to meet high concurrency requirements and maintain the efficiency and stability of report generation.

[0106] In step S1, the data of different scales are standardized by the Z-score standardization formula. The Z-score standardization formula is as follows:

[0107]

[0108] Where x is the original data point, μ is the mean of the data set, σ is the standard deviation of the data set, and x′ is the standardized data;

[0109] In step S1, the Skip-gram model is used to convert words into high-dimensional vectors when processing unstructured text data. Its objective function is:

[0110]

[0111] Where T is the number of words in the text, c is the size of the context window, and w t is the current word, w t+j is a word in the context, P(w t+j ∣w t ) is a given word w t In the case of generating word w t+j The probability is usually expressed using the Soft-max function:

[0112]

[0113] Among them, w O and w I are the word vectors of the output and input words, and v is the size of the vocabulary;

[0114] In step S1, when processing image data, feature extraction is performed through a convolutional neural network, and the formula for the convolution operation is:

[0115]

[0116] Among them, x(i,j) is the pixel value of the input image, k(m,n) is the weight of the convolution kernel, y(i,j) is the output value after the convolution operation, and M and N are the sizes of the convolution kernel.

[0117] Z-score standardization ensures uniform scale of data and optimizes model performance;

[0118] When processing text data, the Skip-gram model can generate high-quality word vectors and improve the model's ability to understand text semantics.

[0119] When processing image data, convolutional neural networks extract local features through convolution operations, providing strong support for the model's image recognition.

[0120] In step S2, in multimodal deep learning, data of different modalities are mapped into the same feature space through the joint embedding space. The algorithm formula is:

[0121] h t =f text (x t ),

[0122] h i =f image (x i ),

[0123] Among them, x t and x i Represents the input data of text and image, f text and f image It is a deep neural network used to extract text and image features. t and h i is the feature vector extracted from text and image;

[0124] In step S2, feature fusion is achieved by concatenating, weighted averaging or fusing features of different modalities through an attention mechanism. The algorithm formula is:

[0125] h fusion =α·h t +β·h i ,

[0126] Among them, h fusion is the fused feature representation, α and β are the fusion weights;

[0127] In step S2, the feature representation is adjusted by calculating the correlation between different features. The formula is as follows:

[0128]

[0129] Where Q is the query matrix, K is the key matrix, V is the value matrix, and d k is the dimension of the key vector and softmax is the activation function used to calculate the attention weights.

[0130] The joint embedding space aligns data from different modalities in the same feature space, enhancing the model’s cross-modal understanding capabilities.

[0131] Feature fusion uses information from different modalities in a variety of ways to improve the comprehensiveness and flexibility of report generation;

[0132] Feature relevance adjustment dynamically adjusts feature weights through the attention mechanism, allowing the model to more accurately generate reports that meet actual needs.

[0133] In step S3, the knowledge enhancement model enhances text representation by introducing entity information in the knowledge graph based on BERT. The algorithm formula is:

[0134]

[0135] in, is the input text representation vector, is the entity enhancement vector obtained by knowledge graph embedding, is the enhanced text representation vector;

[0136] In step S3, the dynamic context-aware mechanism enables the model to dynamically adjust the output according to the context. The formula is as follows:

[0137]

[0138] Where Q is the query matrix, K is the key matrix, V is the value matrix, B is the bias matrix used to encode the context, and d k is the dimension of the key vector, softmax is used to calculate the attention weight, and ContextVec represents the context vector;

[0139] In step S3, in order to incorporate historical data trends into the generation process, a time series model or a recurrent neural network can be used to capture temporal dependencies. The formula is as follows:

[0140] h t =σ(W h x t +U h h t-1 +b h ),

[0141] Among them, x t is the input feature vector at time t, h tis the hidden state vector at time t, h t-1 is the hidden state at the previous time, W h and U h is the weight matrix, b h is the bias vector.

[0142] The knowledge enhancement model improves the semantic richness and accuracy of text representation by combining entity information from external knowledge graphs, making the report content more professional and in-depth;

[0143] Dynamic context-aware mechanism enables the model to dynamically adjust output according to real-time context, ensuring that the report content accurately reflects the current environment and needs;

[0144] Time series models enhance the predictive power of reports by capturing temporal dependencies, allowing the generated reports to be based on current data and provide insights into future trends.

[0145] In step S4, the learning rate is dynamically adjusted through an adaptive learning algorithm in deep learning, and its algorithm formula is:

[0146] m t =β1m t-1 +(1-β1)g t ,

[0147]

[0148] Among them, g t is the gradient at time step t, m t and v t are the first and second order momentum of the gradient, β1 and β2 are the decay rates, and is the bias-corrected momentum, θ t is the parameter vector, α is the learning rate, and ∈ is a decimal;

[0149] In step S4, online learning usually involves updating model parameters sample by sample or in small batches. The algorithm formula is:

[0150]

[0151] Among them, θ t is the parameter vector of the model at time step t, η is the learning rate, In the sample (x t ,y t ) is calculated on the parameter θ t The gradient of

[0152] In step S4, in reinforcement learning, user feedback can be used as a reward signal to optimize the strategy. The formula of the policy gradient method is:

[0153]

[0154] Among them, π θ (a t ∣s t ) is in state s t Take action a t The strategy probability, R(τ) is the total return of trajectory τ, is the policy gradient, which is used to adjust the model parameters θ.

[0155] The adaptive learning algorithm improves the efficiency and stability of model training by dynamically adjusting the learning rate, ensuring that the generated reports have good generalization capabilities;

[0156] Online learning enables the model to adapt to new data in real time, maintaining the accuracy and relevance of report generation, and is used in scenarios where data changes frequently;

[0157] The policy gradient method of reinforcement learning is combined with user feedback to optimize the generation strategy, enhance the personalization of reports and user satisfaction, while optimizing long-term returns to ensure the overall performance of the report.

[0158] In step S5, in reinforcement learning, the strategy of generating the model can be optimized by PPO, and its algorithm formula is:

[0159]

[0160] in, is the strategy ratio, which measures the difference between the current strategy and the old strategy, is the advantage function, ∈ is a hyperparameter;

[0161] In step S5, in the interactive generation model, the generative adversarial network can be used to generate high-quality report content. The loss function of the discriminator in the generative adversarial network is defined as:

[0162]

[0163] Among them, D(x) is the output of the discriminator, and G(z) is the output of the generator;

[0164] In step S5, in interactive generation, the uncertainty of the model can be used to dynamically decide whether to use the strategy. The exploration strategy based on uncertainty can be achieved by calculating the information gain:

[0165]

[0166] Among them, Ψ(θ; x) represents the information gain of parameter θ after given data x, is the posterior distribution of the parameters under the existing data D, is the parameter posterior distribution after incorporating the new data x.

[0167] The PPO optimization strategy provides a stable and efficient strategy update mechanism, ensuring that the model can continuously improve the generation strategy and avoid local optimality;

[0168] The GAN discriminator loss function drives the generator to generate high-quality, realistic, and creative report content, improving the realism and diversity of the report;

[0169] The uncertainty-based exploration strategy optimizes the balance between exploration and utilization in the generation process through intelligent decision-making, further improving the efficiency and accuracy of report generation.

[0170] In step S6, the generative adversarial network consists of a generator G and a discriminator D, which optimize each other through a game process. The loss function of the generative adversarial network is usually defined as:

[0171]

[0172] Among them, D(x) is the output of the discriminator, which represents the probability that the input x comes from the real data, G(z) is the sample generated by the generator from the noise z, and p lats (x) is the distribution of real data, p z (z) is the noise distribution;

[0173] In step S6, the conditional generative adversarial network introduces additional conditional information in the generation process, and the loss function of the conditional generative adversarial network is:

[0174]

[0175] Among them, y is the conditional information, D(x|y) represents the probability of judging whether the input x is real data under condition y, and G(z|y) represents the sample generated by the generator under condition y.

[0176] Generative adversarial networks significantly improve the quality and realism of generated content through adversarial optimization of the generator and the discriminator, making the generated reports more realistic and credible.

[0177] The conditional generative adversarial network generates customized content by introducing conditional information and ensures that the generated content is highly relevant to specific conditions, thereby enhancing the pertinence and practicality of the generated reports.

[0178] In step S7, self-supervised learning technology is widely used in multilingual pre-training models, and its loss function is:

[0179]

[0180] Among them, P(w i ∣w 1:i-1 ,wi+1:n ) is the model in a given context w 1:i-1 and w i+1:n In the case of i probability;

[0181] In step S7, in order to achieve vocabulary alignment between different languages, Procrustes analysis can be used to align the embedding space. The formula is:

[0182]

[0183] Among them, X and Y are the word embedding matrices of the two languages, and R is the rotation matrix used for alignment.

[0184] Self-supervised learning technology promotes the model's learning on multilingual semantic consistency through its loss function, improves the system's generalization ability in different languages, and ensures high-quality report generation in a multilingual environment;

[0185] Procrustes analysis aligns word embedding matrices of different languages ​​to enhance the model’s vocabulary consistency and comprehension capabilities across multiple languages, allowing the report content to maintain a high degree of semantic consistency and accuracy in different language environments.

[0186] In step S8, in order to improve the robustness of the model under adversarial samples, a gradient penalty term can be added to the loss function. The gradient penalty formula in adversarial training is:

[0187]

[0188] Among them, λ is the weight of the gradient penalty, is the interpolation sample between the generated data and the real data, Is the critic in the sample Output on;

[0189] In step S8, multi-task learning allows the model to be trained on multiple related tasks to improve overall performance by sharing representations. Assume there are N groups of tasks and the loss of each group of tasks is The total loss is:

[0190]

[0191] Among them, λ i is the weight of task i.

[0192] The gradient penalty term enhances the robustness of the model under adversarial samples, ensuring that the generated report content remains robust and reliable in the face of various data perturbations and attacks, thereby improving the credibility and security of the report.

[0193] Multi-task learning improves the generalization ability and overall performance of the model by sharing representations, enabling the model to perform well on multiple related tasks and further optimizing the applicability and versatility of report generation.

[0194] In step S9, in order to deal with the concept drift in the data stream, the sliding window method can be used to update the model. Assuming that the window size is W, when new data arrives, the window update formula is:

[0195]

[0196] in, is the new subset of data within the window.

[0197] The sliding window method enables the model to be updated dynamically, ensuring that the generated reports always reflect the latest trends and changes, maintain high accuracy and real-time performance, reduce the model's reliance on old data, reduce the risk of overfitting, and make the report content more valuable for practical reference.

[0198] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A data report generation method based on deep learning, It is characterized in that The steps include: S1. Collect structured data, unstructured data and multimodal data, and clean and preprocess different types of data to use a unified data representation method; S2. Use multimodal deep learning models to extract features from different data modalities, fuse multimodal features, and comprehensively process the features through deep neural networks or self-attention mechanisms to select the most representative features for report generation; S3. Use knowledge enhancement models to integrate domain-specific knowledge bases, industry rules, and historical data trends into feature processing, and dynamically adjust report generation content through context-aware mechanisms so that relevant reports can be generated based on data context; S4. Deploy deep learning models that support adaptive learning, so that the model can adapt to new data and new tasks in real time, and use online learning mechanisms and user feedback to continuously adjust model weights and generation strategies; S5. Design an interactive generation framework that allows users to provide real-time feedback during the report generation process, so as to combine the reinforcement learning mechanism and dynamically optimize the generation model using user feedback; S6. Combine generative adversarial networks or variational autoencoder generation models in the report generation process to explore innovative content generation; S7. Use multilingual pre-trained models to process multilingual data and dynamically adjust the expression of reports according to different cultural backgrounds; S8. Use multiple evaluation metrics to evaluate the quality of reports and measure model performance by comparing with manually generated reports to collect user feedback and continuously iterate and optimize the model. S9. Deploy the optimized model to the production environment, support automatic and interactive report generation in combination with API or Web services, and regularly monitor the performance of the model in actual applications, and update the model in a timely manner based on new data and changes in demand.

2. According to the method for generating data reports based on deep learning in claim 1, It is characterized in that In the step S1, data of different scales are standardized by the Z-score standardization formula, and the Z-score standardization formula is as follows: Where x is the original data point, μ is the mean of the data set, σ is the standard deviation of the data set, and x′ is the standardized data; In the step S1, when processing unstructured text data through the Skip-gram model, words are converted into high-dimensional vectors, and the objective function is: Where T is the number of words in the text, c is the size of the context window, and w t is the current word, w t+j is a word in the context, P(w t+j ∣w t ) is a given word w t In the case of generating word w t+j The probability is usually expressed using the Soft-max function: Among them, w O and w I are the word vectors of the output and input words, and v is the size of the vocabulary; In the step S1, when processing image data, feature extraction is performed through a convolutional neural network, and the formula for the convolution operation is: Among them, x(i,j) is the pixel value of the input image, k(m,n) is the weight of the convolution kernel, y(i,j) is the output value after the convolution operation, and M and N are the sizes of the convolution kernel.

3. According to the method for generating data reports based on deep learning in claim 1, It is characterized in that In the step S2, in multimodal deep learning, data of different modalities are mapped into the same feature space by joint embedding space, and its algorithm formula is: h t =f text (x t ), h i =f image (x i ), Among them, x t and x i Represents the input data of text and image, f text and f image It is a deep neural network used to extract text and image features. t and h i is the feature vector extracted from text and image; In the S2 step, feature fusion is achieved by concatenating, weighted averaging or fusing features of different modalities through an attention mechanism. The algorithm formula is: h fusion =a·h t +β·h i , Among them, h fusion is the fused feature representation, α and β are the fusion weights; In the step S2, the feature representation is adjusted by calculating the correlation between different features. The formula is as follows: Where Q is the query matrix, K is the key matrix, V is the value matrix, and d k is the dimension of the key vector and softmax is the activation function used to calculate the attention weights.

4. The method for generating a data report based on deep learning according to claim 1, It is characterized in that In the S3 step, the knowledge enhancement model enhances text representation by introducing entity information in the knowledge graph based on BERT. The algorithm formula is: in, is the input text representation vector, is the entity enhancement vector obtained by knowledge graph embedding, is the enhanced text representation vector; In the S3 step, the dynamic context-aware mechanism enables the model to dynamically adjust the output according to the context, and its formula is as follows: Where Q is the query matrix, K is the key matrix, V is the value matrix, B is the bias matrix used to encode the context, and d k is the dimension of the key vector, softmax is used to calculate the attention weight, and ContextVec represents the context vector; In the S3 step, in order to incorporate the historical data trend into the generation process, a time series model or a recurrent neural network can be used to capture the temporal dependency, and the formula is as follows: h t =σ(W h x t +U h h t-1 +b h ), Among them, x t is the input feature vector at time t, h t is the hidden state vector at time t, h t-1 is the hidden state at the previous time, W h and U h is the weight matrix, b h is the bias vector.

5. The method for generating a data report based on deep learning according to claim 1, It is characterized in that In step S4, the learning rate is dynamically adjusted by an adaptive learning algorithm in deep learning, and the algorithm formula is: Among them, g t is the gradient at time step t, m t and v t are the first and second order momentum of the gradient, β 1 and β 2 is the decay rate, m t and v t is the bias-corrected momentum, θ t is the parameter vector, α is the learning rate, and ∈ is a decimal; In the S4 step, online learning usually involves updating model parameters sample by sample or in small batches, and its algorithm formula is: Among them, θ t is the parameter vector of the model at time step t, η is the learning rate, In the sample (x t ,y t ) is calculated on the parameter θ t The gradient of In the step S4, in reinforcement learning, user feedback can be used as a reward signal to optimize the strategy. The formula of the policy gradient method is: Among them, π θ (a t ∣s t ) is in state s t Take action a t The strategy probability, R(τ) is the total return of trajectory τ, is the policy gradient, which is used to adjust the model parameters θ.

6. The method for generating a data report based on deep learning according to claim 1, It is characterized in that In the step S5, in reinforcement learning, the strategy of generating the model can be optimized by PPO, and its algorithm formula is: in, is the strategy ratio, which measures the difference between the current strategy and the old strategy, is the advantage function, ∈ is a hyperparameter; In the step S5, in the interactive generation model, the generative adversarial network can be used to generate high-quality report content, and the loss function of the discriminator in the generative adversarial network is defined as: Among them, D(x) is the output of the discriminator, and G(z) is the output of the generator; In the step S5, in interactive generation, the uncertainty of the model can be used to dynamically decide whether to use the strategy. The exploration strategy based on uncertainty can be implemented by calculating the information gain: Among them, Ψ(θ; x) represents the information gain of parameter θ after given data x, is the posterior distribution of the parameters under the existing data D, is the parameter posterior distribution after incorporating the new data x.

7. The method for generating a data report based on deep learning according to claim 1, It is characterized in that In step S6, the generative adversarial network consists of a generator G and a discriminator D, which optimize each other through a game process. The loss function of the generative adversarial network is usually defined as: Among them, D(x) is the output of the discriminator, which represents the probability that the input x comes from the real data, G(z) is the sample generated by the generator from the noise z, and p lats (x) is the distribution of real data, p z (z) is the noise distribution; In step S6, the conditional generative adversarial network introduces additional conditional information in the generation process, and the loss function of the conditional generative adversarial network is: Among them, y is the conditional information, D(x|y) represents the probability of judging whether the input x is real data under condition y, and G(z|y) represents the sample generated by the generator under condition y.

8. The method for generating a data report based on deep learning according to claim 1, It is characterized in that In step S7, self-supervised learning technology is widely used in multi-language pre-training models, and its loss function is: Among them, P(w i ∣w 1:i-1 ,w i+1:n ) is the model in a given context w 1:i-1 and w i+1:n In the case of i probability; In step S7, in order to achieve vocabulary alignment between different languages, Procrustes analysis can be used to align the embedding space, and the formula is: Among them, X and Y are the word embedding matrices of the two languages, and R is the rotation matrix used for alignment.

9. The method for generating a data report based on deep learning according to claim 1, It is characterized in that In step S8, in order to improve the robustness of the model under adversarial samples, a gradient penalty term can be added to the loss function. The gradient penalty formula in adversarial training is: Among them, λ is the weight of the gradient penalty, is the interpolation sample between the generated data and the real data, D Is the critic in the sample Output on; In step S8, multi-task learning allows the model to be trained on multiple related tasks to improve overall performance by sharing representations. Assume that there are N groups of tasks and the loss of each group of tasks is The total loss is: Among them, λ i is the weight of task i.

10. The method for generating a data report based on deep learning according to claim 1, It is characterized in that In step S9, in order to handle the concept drift in the data stream, a sliding window method can be used to update the model. Assuming that the window size is W, when new data arrives, the window update formula is: in, is the new subset of data within the window.

Images