403 results about "Data imbalance" patented technology

The invention discloses a medical image segmentation method fusing multi-modal graphic and text information, and the method comprises the steps: carrying out the preprocessing and coding of a metadata text of a patient through a pre-trained BERT model, extracting the feature representation, and carrying out the attention-guided data enhancement operation of a pathological image. And finally, constructing a medical image segmentation network composed of an image encoder, a text encoder and a decoder, carrying out pathological image segmentation based on the feature representation extracted in the step 1 and the pathological image after data enhancement, and visualizing a segmentation result. In the segmentation network, a multi-modal feature fusion module is designed to fuse information of two modals, the module adopts a Transform architecture, a cross-modal attention mechanism and a self-attention mechanism are introduced, and the effectiveness of information fusion is ensured. According to the method, a more accurate segmentation result is obtained, the generalization ability on new data is improved, and the common problem of data imbalance in the medical image is effectively solved.

Disclosed in the present invention is an improved deep learning model-based refrigeration unit fault detection method. The method uses an LOF algorithm to remove outliers from a fault dataset, and then uses ADASYN technology to solve the problem of data imbalance. In addition, in respect of the problems that existing refrigeration unit fault diagnosis deep learning models are prone to network degradation, and refrigeration unit fault diagnosis models generally lack weighting critical features, the present invention first alleviate, on the basis of ResNet, the problem of network performance degradation which is prone to occur in deep neural network training processes, and then integrates a CBAM for capturing critical features in fault data, so as to improve the feature extraction capability of a network. Experimental results show that the LOF-ADASYN-ResNet-CBAM method provided by the present invention effectively diagnoses refrigeration unit faults.

The invention discloses an industrial part defect sample accurate generation method based on a conditional diffusion model, and belongs to the field of image processing and artificial intelligence. The method forms a closed-loop cooperative system by constructing four deep coupling modules of physical constraint noise scheduling, multi-scale feature coupling, double-domain feedback optimization and adaptive weight adjustment; a defect physical forming mechanism is converted into a dynamic noise scheduling strategy, deep interaction between condition information and a feature map is established at multiple levels of a diffusion network, quality closed-loop optimization is achieved through dual evaluation of a pixel domain and a frequency domain, and training weight is dynamically adjusted according to defect scarcity. And multi-scale accurate control is realized, a quality guarantee closed loop is established, the problem of data imbalance is effectively solved, and the performance of an industrial defect detection model is remarkably improved.

The invention discloses a maritime accident prediction method and device based on interpretable integrated machine learning, and relates to the technical field of maritime affair safety risk analysis, and the method comprises the steps: obtaining accident investigation data, carrying out the preprocessing, balancing the data through a ten-fold layered oversampling method, and carrying out the cross verification training, and determining a performance optimal model by using the test set and carrying out interpretable analysis to explain the influence of the characteristics on the accident prediction result. By constructing a closed-loop'data processing-model optimization-explanation output 'process and adopting SMOTE oversampling and ten-fold layered cross validation training and a heterogeneous base model ensemble learning strategy, the processing capacity of the data imbalance problem of accident categories is improved, the data leakage problem of oversampling is avoided, and the possible bias of a single model is overcome. The interpretability analysis of the model prediction result can quantitatively display the contribution degree of each feature to prediction globally and locally, reveal the nonlinear relationship and interaction effect between the features, and provide transparent interpretation of model decision.

The invention relates to the field of bioinformatics and drug development, and discloses an efficient siRNA effectiveness prediction method and system based on an XGBOOST algorithm. The method comprises the following steps: acquiring a training data set containing siRNA sequence information, chemical modification information and corresponding validity data; carrying out cleaning, feature extraction and data balance processing on the data; training the prediction model by using an XGBOOST algorithm, and optimizing hyper-parameters; performing model evaluation and optimization through cross validation; a new siRNA sequence is input for effectiveness prediction; explaining a prediction result by using an SHAP value analysis technology; the actual interference efficiency of the efficient siRNA sequence is verified through an in-vitro cell experiment, and the model is further optimized based on an experiment result. According to the method, the accuracy and efficiency of siRNA effectiveness prediction are improved, the influence of chemical modification is comprehensively considered, the problem of data imbalance is solved, the model interpretation is enhanced, and the calculation efficiency is remarkably improved.

The invention belongs to the technical field of rainfall prediction, and discloses a short and temporary rainfall prediction method based on multi-model random scheduling integration, which develops a robust training and pushing framework based on a continuous rolling prediction strategy, and decomposes long-sequence prediction into manageable stages. According to the method, training is carried out through teacher forcing and planned sampling, error propagation is relieved, and the training process is stabilized. The invention further designs asymmetric encoder-decoders (DSE and AFD) that achieve lower FLOPs than competitive baselines under standardized assessment, where DSE selectively compresses significant features and AFD stepwise reconstructs details to mitigate excessive smoothing problems. Finally, an intensity weighted Gaussian KL divergence loss function is designed, and the key problem of data balance is solved by modeling and predicting on a distribution level and endowing a large weight to a meteorological important heavy rainfall event.

The invention discloses an information transmission attack event intelligent prediction model training optimization method, which comprises the steps of a data preparation stage, a feature engineering stage, a model selection and construction stage, a model training stage and a model evaluation and optimization stage, and is characterized in that data are collected from multiple channels, noise is cleaned and removed, missing values and abnormal values are processed, and accurate labeling is carried out; then key features are extracted, redundant irrelevant features are removed through selection, standardized transformation is carried out, a proper model is selected based on data features, an integrated learning combination model is adopted, then a training set, a verification set and a test set are divided, a proper optimizer is selected, parameters are adjusted to be over, training is carried out through the training set, and over-fitting is prevented through the verification set; and finally, using index evaluation of accuracy and recall rate, and adjusting hyper-parameters and continuous optimization of the model structure through cross validation. According to the method, data collection, data imbalance processing and data noise interference processing are adopted, data diversity is improved, and interference of data noise is reduced.

The invention relates to the field of Internet of Vehicles security, and particularly discloses a lightweight Internet of Vehicles intrusion detection method integrating multi-teacher distillation and boundary perception enhancement, which comprises the following steps: firstly, acquiring in-vehicle CAN bus flow data and out-vehicle network flow data, and dividing the data into a training set, a test set and a verification set; and the traffic data is converted into two-dimensional image data by using a Gramian Angular Field technology. Then, constructing a lightweight student model LAD Net, and adopting a multi-path feature extraction structure, a CBAM convolution attention module and a time convolution network TCN module; meanwhile, a multi-teacher fusion model is constructed based on a ResNet32 network, and optimization training is performed on the LAD Net through a boundary perception dynamic feature enhancement mechanism and an adaptive distillation temperature adjustment mechanism. And finally, the trained LAD Net is used for performing prediction and performance evaluation on the test set, so that the problems of high model complexity, data imbalance and insufficient generalization ability in the prior art are effectively solved.

The invention discloses a dynamic multivariate time sequence-oriented infrastructure load prediction method, which comprises the following steps of: predicting an infrastructure load by adopting a trained prediction model, segmenting acquired multivariate time sequence data, querying based on time steps and nodes of input segmented data to obtain a space-time pattern vector, and predicting the load of the infrastructure by adopting a dynamic multivariate time sequence-oriented infrastructure load prediction model. A sequence mode vector is obtained based on input segmented data query, the space-time mode vector and the sequence mode vector are spliced into a dual-mode vector output by a dual-mode pool module, and then in a dynamic MTS encoder, the dual-mode vector and segment embedding representation are spliced to obtain a final feature vector; and finally, inputting the final feature vector into a multi-scale decoder for decoding to obtain a prediction result. Aiming at challenges such as data imbalance, data missing and space-time dependence erosion, the accuracy of a prediction result is improved under the condition that repeated training is not needed, repeated training needed during topological change is avoided, and the robustness, prediction precision and calculation efficiency of the model are improved.

The invention discloses a grouting parameter intelligent feedback method based on rock mass groutability evaluation, which comprises the following steps: collecting geological condition data, rock mass quality data, field test data and grouting process data, and constructing a rock mass groutability grading data set; an SMOTE algorithm is adopted to process data imbalance, input indexes are screened through Pearson correlation analysis, and data are normalized through deviation standardization; an XGBoost classification model is constructed, hyper-parameters are optimized through grid search, the model is trained through K-fold cross validation, and the rock mass groutability grade is evaluated; constructing an XGBoost regression model by taking a rock mass groutability grade, grouting construction data and monitoring data as input, and predicting the maximum value and the minimum value of grouting pressure and slurry density by using an optimal model; and generating a grouting parameter interval according to the predicted maximum value and minimum value, and dynamically feeding back to the grouting equipment to regulate and control parameters. According to the method, scientificity and accuracy of rock mass groutability evaluation and grouting parameter regulation and control can be effectively improved.

The invention provides a diabetic retinopathy grading method based on a vision-language pre-training model and a ranking perception prompt strategy. According to the method, an eye ground color image and a corresponding DR grading label are converted into image features and text embedding through an image and text dual-coding mode, matching is carried out through the similarity between the image and the text, and accurate DR grading is achieved. In order to solve the problems of class imbalance and natural sequence modeling, a ranking perception prompt module and a similarity matrix smoothing module are introduced, ranking perception prompt ensures that the model can learn and maintain the sequence relation between different DR levels, and an SMS module effectively relieves the influence caused by data imbalance. The method disclosed by the invention can show relatively high accuracy and robustness in various DR grading tasks, especially under the condition of few samples or unbalanced categories. The method not only improves the grading accuracy, but also provides an effective auxiliary tool for clinical diagnosis, and has a wide application prospect.

The invention provides a self-adaptive sampling method for fault diagnosis under data multi-class imbalance and related equipment, and effectively solves the problems of low sample generation quality, high parameter dependence, high calculation complexity, poor adaptability to complex working conditions and the like. The method comprises the steps of coping with different data feature scenes through a parameter adaptive calculation mechanism, then exploring a global optimal solution in multi-classification modeling accuracy model solution optimization by an evolution mechanism through employing a Newton-Raphson optimizer thought, and finally converting an optimal solution set into various fault samples by using a feature recombination mechanism. Therefore, high-quality sample equalization is realized, small sample multi-class imbalance fault diagnosis is carried out by combining MAESTE and multi-class LS-SVM, and the interpretability of the model is improved.

The invention discloses an electric energy metering data anomaly detection method based on a generative adversarial network and multi-task optimization, which solves the problem of data imbalance in an electric energy metering data anomaly detection scene by utilizing a dynamic resampling strategy, and trains the generative adversarial network in combination with electric energy metering characteristics. Pseudo samples consistent with real distribution are generated through adversarial training of a generator and a discriminator, and an abnormal sample set is expanded, so that the detection capability of the model on abnormal data is enhanced. By constructing a multi-task learning framework, sharing a feature extraction module and jointly optimizing an anomaly detection task and a load prediction task, the accuracy of anomaly detection and the precision of load prediction are remarkably improved. The method specifically comprises the following steps: a data preprocessing step, a dynamic resampling step, a generative adversarial network training step, a multi-task joint optimization step, and an anomaly detection and load prediction step.

The invention provides a spatio-temporal data-based early warning method and a spatio-temporal data-based early warning system for an extreme weather-induced power failure, and aims to accurately and timely predict and warn possible influences and failures of an extreme weather event on a power system so as to improve the reliability and the toughness of the power system. The method comprises the following steps: firstly, collecting spatio-temporal data and other influence factors observed by a meteorological station, including but not limited to meteorological data, power grid structure and operation data, geographical and time data and the like; secondly, a ConvLstm-based deep space-time residual network is utilized to fuse space-time data and external factors to realize time-space two-dimensional prediction of meteorological characteristics, and an SMOTE-NC-NCL mixed sampling method is adopted to relieve data imbalance; afterwards, a complex relation between an extreme weather event and a power system fault is finally predicted and identified by using a LightGBM algorithm, early warning information is provided for a power grid in time, an electric power company is assisted to make a quick response and a decision, and the stability of power grid operation is ensured. Therefore, economic loss is reduced and social public safety is improved.

The invention discloses an enhanced contrast learning method and device for highly unbalanced text classification, and the method comprises the steps: carrying out the preprocessing of text data in a training data set and a test data set, and obtaining corresponding original training texts and test texts; respectively inputting the original training text and the test text into the RoBERTa model to obtain corresponding text embedding representations; constructing a text classification model based on data enhancement and comparative learning, wherein the text classification model comprises a text enhancement module, a text representation module, a comparative learning module and a text classification module; and applying the trained and tested text classification model to classification processing of to-be-processed text data, and outputting a classification result. The invention provides a new self-constrained stacked data enhancement method and a new double-weighted contrast loss function, the characterization problem caused by data imbalance and data lack can be solved, and a better text classification effect is achieved.

The invention discloses a few-sample target detection method and system based on an aggregation variational prototype. The method comprises the steps of constructing a data set containing a base class and a new class, dividing the data set into a support set and a query set, generating a class prototype by utilizing a P-VAE module in combination with CLIP semantic features and a feature discriminator, realizing bidirectional fusion of the support set and the query set features by means of an MFM module, fusing the query features and the class prototype, and inputting the fused query features and the class prototype into a detection head to complete target detection. The system comprises a data set construction module, a priori variational automatic encoder P-VAE module, a mutual fusion module MFM, a feature fusion module and a target detection module. According to the scheme, by introducing semantic priori, optimizing prototype generation and feature interaction, the problems of data imbalance and insufficient new class feature representation in a few-sample scene are solved, improvement of new class detection precision is verified on PASCAL VOC, MS COCO and other data sets, and an effective solution is provided for target detection of sample scarce scenes such as medical images and rare species monitoring.

The invention discloses a test risk identification method and system based on artificial intelligence, and particularly relates to the technical field of data processing. By introducing an adaptive learning rate and a feedback mechanism, the improved test risk identification method is provided for the problem of data imbalance existing in the test process of an existing machine learning model, and the test risk identification method is provided by comprehensively collecting test data including test case execution results and test environment states. Extracting a dependency relationship abnormal value and a network delay high-frequency value between code modules, constructing a data prediction model, and comprehensively calculating a deviation of a prediction result based on the data; by dynamically adjusting the deviation of the model to the successful use case, the identification capability of the failed use case and the high-risk area is enhanced, and the high-risk area is effectively prevented from being underestimated or the normal use case is misjudged to be high in risk, so that the prediction precision is improved, the configuration of test resources is optimized, the invalid test reexamination is reduced, the test efficiency is remarkably improved, and the cost and the period are reduced.

The invention discloses a hydraulic pump fault prediction method and system, and relates to the technical field of hydraulic system fault diagnosis and prediction.The hydraulic pump fault prediction method comprises the steps that a vibration sensor signal of a hydraulic pump is collected, and the vibration sensor signal is converted into time domain vibration data according to a preset sampling frequency; performing wavelet packet decomposition on the time domain vibration data to obtain energy coefficients of different frequency bands, and selecting the frequency bands corresponding to the first three maximum energy coefficients as characteristic frequency bands; extracting characteristic indexes in the characteristic frequency band, constructing a fault characteristic vector, training the fault characteristic vector, and establishing a hydraulic pump fault prediction model; and inputting the fault feature vector of a to-be-tested hydraulic pump into the hydraulic pump fault prediction model, and predicting the fault type of the hydraulic pump. According to the method, the fault type and degree of the hydraulic pump are accurately predicted, and the challenges of data imbalance and nonlinear feature extraction in hydraulic pump fault diagnosis are effectively solved, so that early prediction and early warning of the fault of the hydraulic pump are realized.

The invention provides a nuclear power valve fault diagnosis method and system, and the method comprises the steps: collecting vibration signal data of different types of valves of a nuclear power plant, carrying out the normalization of the vibration signal data, and carrying out the drawing storage of a target picture; wherein the normal vibration signal data pictures in the target pictures are used as a training set, and the normal and abnormal vibration signal data pictures are used as a test set; a loss function of the ProGAN model is improved based on a Wasserstein distance, and a threshold determination module is introduced; and training the improved model based on the target picture. According to the nuclear power valve fault diagnosis method, the ProGAN model is improved, the nuclear power plant valve data imbalance problem and the valve diversity problem are effectively adapted, meanwhile, the loss function of the ProGAN is improved based on the Wasserstein distance, and the model training stability and the generated sample quality are improved.

The invention belongs to the technical field of sliding bearings, and particularly discloses a sliding bearing fault diagnosis method and system under data extreme imbalance, and the method comprises the steps: collecting vibration accelerations and acoustic emission signals at different positions of a sliding bearing in different states, extracting initial features of data feature vectors, and splicing the initial features to obtain fusion features; a multi-source heterogeneous data fusion fault diagnosis model is constructed and split, a multi-source heterogeneous data fusion feature generation model is constructed and trained based on CWGAN-GP, a CWGAN-GP model generator and a discriminator are subjected to confrontation adjustment, features generated by the generator each time are input into a fault diagnosis classification part of the fault diagnosis model to calculate classifier loss, and a fault diagnosis result is obtained. And feeding back to the generator network to adjust parameters until nearly real characteristics are generated, and outputting a bearing fault diagnosis result. By adopting the technical scheme, the characteristic data with higher value density is generated under the condition of fully utilizing the data information to supplement the unbalanced sample, and the fault diagnosis under the extreme imbalance of the sliding bearing data is realized.

The invention relates to a bearing unbalanced sample fault diagnosis method based on dynamic modeling and a causal interpretable GAN, and belongs to the technical field of mechanical intelligent detection. Aiming at the problems of low model diagnosis precision caused by data imbalance in the prior art and deficient physical significance and poor interpretability of existing GAN network generation samples, the invention provides a technical scheme of fusing dynamic modeling and an interpretable generative adversarial network. The method comprises the following steps: generating a fault response signal with physical significance through a two-degree-of-freedom dynamic model; constructing a GabcauACGAN network, combining a causal loss function to restrain the relevance between the generated features and fault tags, and extracting time-frequency features by adopting a learnable Gabor filtering layer. The method has the technical effects that the time-frequency similarity SSIM of a generated fault sample and a real signal reaches 99%, and the diagnosis accuracy is improved to 98% or above.

The invention discloses a federated undersampling learning method for unbalanced network intrusion detection, which comprises a federated learning server and a client, and communication connection is established between the federated learning server and the client; all models for network intrusion detection are configured in the server; the server is used for maintaining a global model and deploying the global model to the client for training and learning, and the client adopts a local data set to train a local model downloaded from the server and uploads the local model after training and learning to the server; the server aggregates the local models uploaded by the client to obtain a global model with updated parameters; and when the client adopts the local data set to train the server, undersampling processing is performed on the local data set to obtain a balanced data set to perform learning training on the client. When federated learning is carried out on an intrusion detection network model, the influence of training data imbalance in a client on the model detection performance is solved through an undersampling method.

The invention relates to the technical field of nuclear power plant safety, in particular to a nuclear power plant fault diagnosis method based on data balancing, which establishes a data set balancing technology for deep learning through SA-TimeGAN, can effectively solve the problem of data imbalance caused by large normal data volume and small fault data volume in a nuclear power plant, and improves the fault diagnosis accuracy of the nuclear power plant. According to the method, firstly, small sample fault data are preprocessed, then, SA-TimeGAN is used for conducting adversarial training on the multiple types of small sample fault data, generated fault samples are added into an original training data set to achieve balance of sample data, and finally, the balanced data set is input into a CNN-BiLSTM diagnosis model for training, so that fault diagnosis is achieved. According to the method, the recognition capability of the model on minority classes of fault samples is effectively enhanced, and real data distribution is approached through adversarial training, so that high-quality synthetic data is generated.

The invention relates to the technical field of data processing, and discloses a defect identification method and device for defect data, equipment and a medium, and the method comprises the steps: dynamically adjusting the attention weight of a defect identification model for difficult-to-classify samples based on the feature distribution and inter-class relationship of each sample in a training sample set; extracting text features and image features of the input data through a multi-mode encoder, and generating a fusion control signal; based on the fusion control signal and the attention weight, generating synthetic defect data consistent with target defect semantics through a controllable diffusion model; mixing the synthesized defect data with an original training sample set, and training a defect recognition model after a balanced data set is constructed; and carrying out defect classification and positioning on an input image by utilizing the trained defect identification model, and outputting defect category and position information. According to the method, the problems of data imbalance and sample scarcity can be effectively solved, and the accuracy and robustness of defect identification on the defect data are remarkably improved.

The invention discloses a large model-based production breaking data analysis method, equipment and a medium. The method comprises the following steps: acquiring multi-channel production breaking data through a distributed data acquisition system, and realizing automatic classified storage of the data by adopting OCR and text extraction technologies; a deep learning model is utilized to extract a production breaking field entity, and an enterprise production breaking knowledge graph is constructed through a mutual attention mechanism; an oversampling algorithm is applied to process a data imbalance problem, and a multi-dimensional feature system is constructed in combination with a knowledge graph; a Qwen-plus large language model and an improved V-detector algorithm are adopted to realize dynamic risk early warning; similar historical cases are retrieved based on the knowledge graph, and a personalized disposal scheme is generated in combination with rule reasoning; according to the method, the knowledge graph and the deep learning technology are creatively fused, the dynamic monitoring, accurate early warning and intelligent decision support of the production breaking risk are realized, and the accuracy and timeliness of the production breaking data analysis are remarkably improved. The method can be widely applied to the field of financial institution, court and enterprise risk management.

The invention discloses an epilepsy early warning system based on electroencephalogram signals, and relates to the technical field of nervous systems, the epilepsy early warning system comprises an information acquisition module, a data division module, a data preprocessing module, a model training module and a real-time reasoning module, the information acquisition module is used for determining a few-channel electroencephalogram signal acquisition position and acquiring electroencephalogram signal data; the data division module is used for recording the total number of times of epileptic seizure events and defining the interval of seizure, the early stage of seizure and the period of seizure for each epileptic seizure event, and the data preprocessing module is used for data denoising and dynamically adjusting the overlapping proportion of a data window by adopting a sliding window-based adaptive overlapping data slicing method so as to obtain the data of the epileptic seizure event. The model training module is used for completing a preprocessing process to relieve the problem of data imbalance, and is used for constructing and training an early warning model based on a deep learning model of time-frequency feature fusion; the complexity and wearing burden of the equipment are effectively reduced, and efficient epilepsy early warning is realized.

The invention provides a data tamper-proofing method and system in a data sharing scene, which solves the problem of data imbalance of a training data set by using an oversampling algorithm and an undersampling algorithm, optimizes a depth anomaly detection model, shares trained model parameters among nodes participating in sharing, and improves the data tamper-proofing efficiency. According to the method and the system, the nodes can use the model with the same detection effect, and the deep anomaly detection model is used for identifying attacks such as counterfeit synthesis or tampering, so that data communication security and efficient matching are ensured, and the problem that privacy leakage and security threats are easily caused by a disclosed channel in the prior art is solved.

An unsupervised industrial abnormal sound detection method combining self-supervised learning and a Gaussian mixture model (GMM) is provided to deal with intelligent industrial equipment monitoring challenges. Traditional diagnosis depends on manpower or rules, various faults are difficult to process, and data are uneven. According to the method, TinyCRNN and GMM are used, audio is processed into a Mel spectrogram through Librosa, normal sound features are extracted through self-supervision, then abnormity is identified through GMM clustering analysis, and performance is evaluated through AUC and the like. According to the method, dependence on labeled data is reduced, the problem of data scarcity is solved, detection accuracy and robustness are improved, the application potential in industrial fault detection is large, and especially the advantage on unknown faults is remarkable.

The invention provides a method and a system for detecting network hidden attacks, which solve the problem of data imbalance of a training data set by using an oversampling algorithm and an undersampling algorithm, and highlight confused key features by capturing the long-distance dependency relationship of features and calculating the importance degree of each attribute feature, so that the network hidden attacks can be detected. A depth anomaly detection model is optimized, a node uploads trained model parameters, a control node optimizes and aggregates, and then issues the model parameters to each node, and the hidden attack is quickly identified by using the depth anomaly detection model, so that the problems that obfuscated codes cannot be effectively resisted and hidden attack means cannot be quickly and accurately identified in the prior art are solved.

The invention discloses a voice and music collaborative generation method and system based on a dynamic mixed attention and expert architecture, terminal equipment and a medium, and relates to the technical field of audio generation. The method comprises the following steps: acquiring multi-modal input containing at least one of audio, text and vision, performing embedding processing on the multi-modal input, and mapping the multi-modal input to a unified space to obtain a fusion sequence; setting a plurality of attention heads, dynamically selecting the attention heads to activate experts through attention head routing gating, and weighting and aggregating calculation results to obtain mixed attention features; generating expert probability distribution through expert set routing gating, dynamically selecting an expert set to activate experts, and calculating task adaptation features; and performing audio generation processing based on the audio language modeling head to obtain a voice or music result. According to the method, collaborative generation of voice and music, dynamic allocation of computing resources, balance of domain-specific and cross-domain general knowledge learning are realized, the problems of task conflict and data imbalance are solved, and the audio generation quality and efficiency are improved.