Method and system for automatically classifying heterogeneous data based on business deep learning

Abstract

The invention discloses a method and a system for automatically classifying heterogeneous data based on business deep learning, particularly relates to the technical field of data classification, and is used for solving the problems of limited classification accuracy and insufficient interpretability caused by cross-modal feature entanglement in the existing method. The method comprises the following steps: decoupling and separating an independent feature and a redundant feature vector through a modal implicit feature, generating a noise correction instruction based on orthogonality residual evaluation, carrying out cross-modal noise distribution consistency correction on the redundant feature, quantifying a gradient direction conflict rate in combination with a vector space projection relation, and dynamically adjusting a back propagation path. Fusion path switching is triggered through semantic matching degree and contribution degree balance monitoring, a self-adaptive classification decision boundary function is constructed, and high-precision classification of heterogeneous data and strong generalization adaptation of service scenes are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of data classification, and more specifically, to a method and system for automatically classifying heterogeneous data based on business deep learning. Background Art

[0002] Automatic classification methods for heterogeneous data have been widely used in multiple business scenarios. Existing technologies usually map unstructured data from different sources (such as text, images, and time series signals) into a unified feature space through multimodal data fusion and joint training, and use deep neural networks to extract common features to complete classification tasks. While improving classification accuracy, they also rely on the collaborative expression of cross-modal features and the design of interaction mechanisms to achieve semantic consistency modeling of heterogeneous data.

[0003] However, in practical applications, due to the complexity of implicit features of unstructured data and the uncontrollable correlation between modalities, existing methods easily cause the underlying implicit features of different modalities to become unexpectedly entangled in the representation space during the feature fusion process, making it difficult for the classifier to distinguish between the essential attributes of the data and noise interference, and then relying on non-causal feature combinations to make decisions, resulting in reduced interpretability of the classification results and limited generalization capabilities for business scenarios. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a method and system for automatic classification of heterogeneous data based on business deep learning to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions: A method for automatic classification of heterogeneous data based on business deep learning includes the following steps: S1. Perform implicit feature decoupling within the modality of unstructured data such as text, images, and time series signals, separating the independent feature vectors and redundant feature vectors of each modality; S2, evaluating the effectiveness of decoupling based on the orthogonality residuals of the independent eigenvectors and the redundant eigenvectors and generating noise correction instructions; S3, performing cross-modal noise distribution consistency correction on the redundant feature vector according to the noise correction instruction to obtain a corrected redundant feature vector; S4. Based on the corrected redundant eigenvectors and independent eigenvectors, the gradient direction conflict rate of each mode is determined through the vector space projection relationship and the weight distribution deviation is generated; S5. Dynamically adjust the gradient back propagation path according to the weight distribution deviation; S6. Based on the adjusted gradient backpropagation path, monitor the matching degree between text and image at the semantic abstraction level and the balance of contribution of temporal signals in cross-modal fusion; S7. If the matching degree is lower than the preset threshold or the contribution balance deviates from the preset interval, the fusion path switching instruction is triggered to construct a classification decision boundary function to output the classification result.

[0006] In a preferred embodiment, the unstructured data of text, image and time series signals are subjected to modality-intramodal implicit feature decoupling to separate the independent feature vectors and redundant feature vectors of each modality, including: Perform word vector embedding and syntactic dependency parsing on text data to generate the initial implicit feature vector of the text modality; Perform multi-scale perceptual decomposition of convolution kernels on image data to extract the initial implicit feature vector of the image modality; Perform windowed Fourier transform and nonlinear activation mapping on the time series signal to generate the initial implicit feature vector of the time series modality; The initial implicit feature vectors of each modality are input into the sparse autoencoder respectively, and the independence weights between the features of each modality are calculated through the intra-modal orthogonal constraint function to separate the independent feature vectors and redundant feature vectors of each modality.

[0007] In a preferred embodiment, evaluating the decoupling effectiveness and generating noise correction instructions based on the orthogonality residual of the independent eigenvector and the redundant eigenvector includes: Calculate the orthogonality residual between the independent eigenvectors and redundant eigenvectors of each mode, and quantify the decoupling error by the difference between the vector inner product mean and the preset dynamic threshold; Generate noise correction instructions based on the orthogonality residual distribution. When the orthogonality residual exceeds the preset residual threshold, perform weight decay on the redundant feature vectors and trigger iterative decoupling of the sparse autoencoder of the corresponding mode. The orthogonality constraint term is updated based on the corrected redundant eigenvectors, and the residual cross-modal correlation in the independent eigenvectors is eliminated through the linear transformation matrix. The updated independent eigenvectors and redundant eigenvectors are re-orthogonally projected to generate the final decoupled output.

[0008] In a preferred embodiment, performing cross-modal noise distribution consistency correction on the redundant feature vector according to the noise correction instruction to obtain the corrected redundant feature vector includes: The cross-modal noise distribution of redundant feature vectors of each modality is extracted based on the noise correction instruction, and the noise distribution map is constructed by measuring the sliding window mean and distribution difference. The noise distribution maps of different modes are aligned across modalities, the distribution offset is eliminated through covariance matrix decomposition, and noise consistency constraints are generated; Dynamically adjust the weight coefficients of redundant eigenvectors according to the noise consistency constraint term, and perform linear projection correction on components whose cross-modal correlation is higher than the preset condition; The corrected redundant eigenvector is iteratively orthogonally verified with the independent eigenvector until the distribution difference measure of the cross-modal noise distribution is lower than the preset threshold, and the corrected redundant eigenvector is output.

[0009] In a preferred embodiment, based on the corrected redundant eigenvectors and independent eigenvectors, the gradient direction conflict rate of each mode is determined through the vector space projection relationship and the weight distribution deviation is generated, including: Calculate the gradient direction conflict rate between the corrected redundant eigenvectors and independent eigenvectors of each modality, and quantify the cross-modal gradient difference through cosine similarity and directional angle; The weight distribution deviation matrix is generated based on the gradient direction conflict rate, and the gradient distribution consistency of each mode is statistically analyzed through the probability density function; Dynamically adjust the fusion weight coefficients of each modal feature according to the weight distribution deviation matrix, and perform exponential decay compensation on the modalities whose gradient direction conflict rate is higher than the dynamic threshold; The adjusted feature fusion weight coefficient is substituted into the vector space projection relationship for iterative verification until the variance of the conflict rate of each modal gradient direction is lower than the preset convergence condition, and the final weight distribution result is output.

[0010] In a preferred embodiment, dynamically adjusting the gradient back propagation path according to the weight distribution deviation includes: The influence factor of each modal gradient on the back propagation path is calculated based on the weight distribution deviation matrix, and the path deviation intensity is quantified by the Hadamard product of the deviation and the gradient; Generate a gradient path mask matrix based on the influencing factors, and perform probability threshold filtering on path nodes whose deviation intensity is higher than the dynamic threshold; Dynamically adjust the backpropagation path weights based on the gradient path mask matrix, and retain the gradient update channels of low-deviation nodes through the sparse connection matrix; The adjusted back propagation path is substituted into the network feedforward calculation for iterative verification until the variance of the path weight change is lower than the convergence threshold, and the optimized gradient propagation path is output.

[0011] In a preferred embodiment, based on the adjusted gradient backpropagation path, monitoring the matching degree between text and image at the semantic abstraction level and the balance of contribution of time series signals in cross-modal fusion includes: Calculate the matching degree between text modality and image modality at the semantic abstraction level, and quantify the cross-modal semantic alignment error using the attention weight matrix and cosine similarity. Analyze the contribution balance of time series signals in cross-modal fusion, and compare the joint distribution matrix of time series features and text image features through correlation coefficients and distribution offsets; Generate cross-modal balance constraints based on cross-modal semantic alignment error and contribution balance, and eliminate inter-modal contribution deviations through matrix decomposition and dynamic weight adjustment; The parameters of the gradient backpropagation path are updated based on the cross-modal balance constraint term to ensure that the monitoring results of semantic matching and contribution deviation meet the preset conditions.

[0012] In a preferred embodiment, if the matching degree is lower than a preset threshold or the contribution balance deviates from a preset interval, the fusion path switching instruction is triggered to construct a classification decision boundary function to output the classification result, including: Monitor whether the semantic matching degree between text and image modalities is lower than a preset threshold, and detect whether the balance of time series signal contribution deviates from a preset range; When the trigger matching degree is lower than the preset threshold or the contribution balance deviates from any condition in the preset range, a fusion path switching instruction is generated and the cross-modal fusion path is switched; Based on the switched fusion path, a classification decision boundary function is constructed, and a classification hyperplane is generated through support vector machine kernel function mapping and logistic regression probability calibration; The classification results are substituted into the validation set for a decision boundary stability test. If the classification confidence variance exceeds the convergence condition, the path switching is retriggered until the classification result that meets the preset accuracy is output.

[0013] In a preferred embodiment, switching cross-modal fusion paths is achieved through dynamic weight reorganization and feature channel masking.

[0014] In another aspect, the present invention provides a system for automatic classification of heterogeneous data based on business deep learning, comprising: Feature decoupling module: performs implicit feature decoupling of unstructured data such as text, images, and time series signals within the modality, separating the independent feature vectors and redundant feature vectors of each modality; Residual evaluation module: It evaluates the decoupling effectiveness based on the orthogonality of the independent eigenvectors and the redundant eigenvectors and generates noise correction instructions; Noise correction module: performs cross-modal noise distribution consistency correction on the redundant feature vector according to the noise correction instruction to obtain the corrected redundant feature vector; Weight allocation module: Based on the corrected redundant eigenvectors and independent eigenvectors, the gradient direction conflict rate of each mode is determined through the vector space projection relationship and the weight allocation deviation is generated; Path adjustment module: dynamically adjusts the gradient back propagation path according to the weight distribution deviation; Balance monitoring module: Based on the adjusted gradient backpropagation path, it monitors the matching degree between text and image at the semantic abstraction level and the balance of contribution of time series signals in cross-modal fusion; Classification decision module: If the matching degree is lower than the preset threshold or the contribution balance deviates from the preset interval, the fusion path switching instruction is triggered to construct a classification decision boundary function to output the classification result.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. Through multi-stage feature decoupling and dynamic path optimization mechanisms, the system achieves synergistic enhancement of cross-modal semantic expression and feature interaction during the automatic classification of heterogeneous data. Based on implicit feature decoupling within modalities and orthogonality residual evaluation, it can effectively separate essential attribute features from cross-modal noise components in unstructured data, construct a feature space expression with clear semantic orientation, and enable the classifier to make decisions based on high-purity feature vectors. Through noise distribution consistency correction and gradient direction conflict quantification mechanisms, it solves the information interference problem caused by feature entanglement during the multimodal data fusion process, enabling the classification model to maintain sensitivity to the essential attributes of the data and noise resistance in complex business scenarios, significantly improving the business interpretability and decision credibility of the classification results. 2. A multi-level fusion control strategy based on dynamic path adjustment can perceive the state offset of cross-modal feature interaction in real time through semantic matching monitoring and contribution balance analysis, and adaptively trigger fusion path switching and decision boundary reconstruction. It can dynamically balance the semantic contribution weights of different modalities according to data characteristics, while ensuring the stability of the classification model and enhancing the adaptability to changes in heterogeneous data distribution. By constructing a fault-tolerant classification decision function, it achieves dual isolation of data noise interference and modal correlation fluctuations in business scenarios, so that the classification system has stronger generalization performance and scenario migration capabilities while maintaining high accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of a method for automatic classification of heterogeneous data based on business deep learning of the present invention; Figure 2 This is a structural diagram of a system for automatic classification of heterogeneous data based on business deep learning in the present invention. DETAILED DESCRIPTION

[0017] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0018] Example 1: Figure 1 The present invention provides a method for automatic classification of heterogeneous data based on business deep learning, which includes the following steps: S1. Perform implicit feature decoupling within the modality of unstructured data such as text, images, and time series signals, separating the independent feature vectors and redundant feature vectors of each modality; S2, evaluating the effectiveness of decoupling based on the orthogonality residuals of the independent eigenvectors and the redundant eigenvectors and generating noise correction instructions; S3, performing cross-modal noise distribution consistency correction on the redundant feature vector according to the noise correction instruction to obtain a corrected redundant feature vector; S4. Based on the corrected redundant eigenvectors and independent eigenvectors, the gradient direction conflict rate of each mode is determined through the vector space projection relationship and the weight distribution deviation is generated; S5. Dynamically adjust the gradient back propagation path according to the weight distribution deviation; S6. Based on the adjusted gradient backpropagation path, monitor the matching degree between text and image at the semantic abstraction level and the balance of contribution of temporal signals in cross-modal fusion; S7. If the matching degree is lower than the preset threshold or the contribution balance deviates from the preset interval, the fusion path switching instruction is triggered to construct a classification decision boundary function to output the classification result.

[0019] S1. Perform implicit feature decoupling within the modality of unstructured data such as text, images, and time series signals, separating the independent and redundant feature vectors of each modality, including: In step S1, implicit feature decoupling within the modality is performed on unstructured data such as text, images, and time series signals, separating the independent and redundant feature vectors of each modality. Specifically, word vector embedding and syntactic dependency parsing are performed on the text data to generate an initial implicit feature vector for the text modality. Word vector embedding uses a pre-trained language model to convert words in the text into vector representations with a fixed dimension, such as the Word2Vec model, where the word vector dimension is set to 300.

[0020] Syntactic dependency parsing extracts grammatical structural features by analyzing the subject-verb-object relationships and modifier relationships between words in a sentence. For example, a dependency tree structure is generated using the Stanford Parser tool. Word vectors and syntactic structural features are weighted and fused based on positional information. The weight coefficients are dynamically adjusted based on the word's hierarchical depth in the dependency tree. For example, the root node weight is set to 0.5, and the weight increases by 0.1 with each descending level. This forms the initial implicit feature vector of the text modality.

[0021] The image data is subjected to multi-scale perceptual decomposition using convolution kernels to extract the initial implicit feature vector of the image modality. Features are extracted from the input image using multiple convolution kernels of different sizes in parallel, for example, 3×3, 5×5, and 7×7 kernel sizes, with 64 kernels per size. The feature maps output by the multi-scale convolution are concatenated in the channel dimension, compressed to a quarter of their original size through pooling, and then restored to their original resolution through an interpolation algorithm, such as bilinear interpolation for upsampling. The final output feature vector has 192 channels, corresponding to the product of the number and size of convolution kernels, forming the initial implicit feature vector of the image modality.

[0022] Perform a windowed Fourier transform and nonlinear activation mapping on the time series signal to generate the initial implicit feature vector of the time series modality. The time series data is divided into fixed-length window segments, for example, each window contains 256 sampling points, and adjacent windows overlap by 128 sampling points. A fast Fourier transform is performed on the signal within each window, retaining the amplitude spectrum of the first 128 frequency components. Noise components are suppressed using a nonlinear function, such as a ReLU activation function, to filter frequency components with amplitudes below 10% of the maximum value. The windowed spectral features are average-pooled along the time axis and compressed into a 128-dimensional vector to form the initial implicit feature vector of the time series modality.

[0023] The initial implicit feature vectors for each modality are fed into a sparse autoencoder. Independent and redundant feature vectors are separated using an intra-modality orthogonal constraint function. The sparse autoencoder employs a three-layer, fully connected network structure. The encoder compresses the input features to half their original dimension, and the decoder restores the original dimension through deconvolution. During training, an L1 regularization constraint is applied to the hidden layer activations, for example, setting the regularization coefficient to 0.01 to enhance feature sparsity.

[0024] The intra-modal orthogonality constraint function calculates the covariance matrix of the initial eigenvectors and constrains the amplitude of their off-diagonal elements, for example, calculating the Frobenius norm of the covariance matrix as an orthogonality loss term. The orthogonality loss and reconstruction loss are weighted and summed according to a preset ratio, for example, the orthogonality loss weight is 1 and the reconstruction loss weight is 2. The network parameters are optimized through backpropagation so that the cosine similarity between independent eigenvectors is less than 0.1 and the spatial angle between redundant eigenvectors and independent eigenvectors is greater than 80 degrees. Ultimately, the independent eigenvectors retain the feature dimensions in the covariance matrix with a variance greater than 0.8, and the redundant eigenvectors are generated by projecting the remaining dimensions through a linear transformation matrix.

[0025] S2. Evaluate the decoupling effectiveness based on the orthogonality residual of the independent eigenvector and the redundant eigenvector and generate noise correction instructions, including: In step S2, the effectiveness of decoupling is evaluated based on the orthogonality residuals of the independent eigenvectors and the redundant eigenvectors and a noise correction instruction is generated. Specifically, the orthogonality residuals between the independent eigenvectors and the redundant eigenvectors of each mode are calculated, and the decoupling error is quantified by the difference between the vector inner product mean and the preset dynamic threshold. The vector inner product mean is calculated by aligning the independent eigenvectors and the redundant eigenvectors of the same mode according to the dimension and then multiplying them element by element, taking the average of the products of all elements as the inner product mean. For example, when the dimension of the independent eigenvector is 128, the sum of the products of the corresponding elements is calculated and then divided by the dimension value. The preset dynamic threshold is dynamically adjusted according to the statistical characteristics of the orthogonality residuals during the historical iteration process. For example, the threshold range is determined by a weighted combination of the moving average and the standard deviation. The initial threshold is set to 0.1 based on experimental experience.

[0026] Noise correction instructions are generated based on the orthogonality residual distribution. When the orthogonality residual exceeds a preset residual threshold, weight decay is applied to the redundant feature vectors, triggering iterative decoupling of the sparse autoencoder for the corresponding modality. Weight decay is implemented by adding an L2 regularization term to the sparse autoencoder's loss function and linearly increasing the regularization coefficient in proportion to the residual excess. For example, when the residual exceeds the threshold, the regularization coefficient increases proportionally to the excess. Iterative decoupling is triggered when the residual exceeds the threshold for multiple consecutive iterations of the same modality. At this point, the feature decoupling process is re-executed, retaining the previous independent feature vector as the initial input to accelerate convergence.

[0027] Specifically, the initial value of the preset residual threshold is determined based on the median of the orthogonality residual distribution of historical training data. For example, during the training phase, the residual values of the first 100 iterations are collected, and the 50th percentile is taken as the initial threshold. The dynamic adjustment rule is as follows: after each full iteration, the moving average of the current residual value is calculated (the window length is the most recent 5 iterations). If the moving average is lower than 80% of the current threshold, the threshold is decreased by 90% of the current value; if it is higher than 120% of the current threshold, it is increased by 110% of the current value. This adjustment process is implemented through a sliding window mechanism to ensure that the threshold changes dynamically with model convergence. The threshold application condition is: weight decay is triggered when the orthogonality residual exceeds 15% of the current threshold. This ratio is determined based on the balance between model stability and convergence speed in experimental verification.

[0028] The orthogonality constraints are updated based on the corrected redundant eigenvectors, and any remaining cross-modal correlations in the independent eigenvectors are eliminated using a linear transformation matrix. The linear transformation matrix is generated by orthogonalizing the corrected redundant eigenvectors and the cross-modal eigenvectors, extracting orthogonal basis vectors to construct the transformation matrix. For example, the Gram-Schmidt orthogonalization method is used to generate orthogonal basis vectors. The dimensions of the transformation matrix match those of the independent eigenvectors, and matrix multiplication is used to map the independent eigenvectors to an orthogonal subspace, eliminating correlations with eigenvectors from other modalities.

[0029] The updated independent eigenvectors and redundant eigenvectors are re-orthogonally projected to generate the final decoupled output. The orthogonal projection adopts the orthogonal basis decomposition method, with the independent eigenvectors as orthogonal basis vectors and the redundant eigenvectors as residual components for decomposition. During the projection process, if the cosine similarity between the independent eigenvectors is still higher than the preset condition, additional orthogonal constraints are imposed on the redundant eigenvectors, such as forcing the angle condition to be met by adjusting the vector direction. The judgment condition for the final decoupled output is that the mean of the orthogonality residuals of consecutive iterations is lower than the preset proportion of the dynamic threshold, and the maximum cosine similarity between the independent eigenvectors does not exceed the preset threshold. If the conditions are not met, the orthogonality evaluation and correction process is repeated until the maximum number of iterations or the convergence condition is reached.

[0030] S3. Performing cross-modal noise distribution consistency correction on the redundant feature vector according to the noise correction instruction to obtain a corrected redundant feature vector, including: In step S3, the redundant feature vector is corrected for cross-modal noise distribution consistency according to the noise correction instruction to obtain the corrected redundant feature vector. Specifically, the cross-modal noise distribution of each modal redundant feature vector is extracted based on the noise correction instruction, and the noise distribution map is constructed by the sliding window mean and the distribution difference metric. The window length of the sliding window mean is set to twice the number of previous iterations. For example, if the current iteration is the 5th iteration, the window contains the redundant feature vectors generated by the first 10 iterations. The arithmetic mean of the values of each feature dimension in the window is taken to generate the local noise distribution. The distribution difference metric is calculated using the KL divergence. Specifically, the probability density function of each modal noise distribution is discretized into a preset number of bins, for example, the numerical range is divided into 100 intervals, the frequency distribution of the data points in each bin is counted, and the product of the logarithm of the frequency ratio of the corresponding bins between the two modes and the frequency difference is calculated to obtain the KL divergence value as the distribution difference metric.

[0031] The noise distribution maps of different modalities are cross-modally aligned, and the distribution offset is eliminated through covariance matrix decomposition to generate noise consistency constraints. The implementation method of covariance matrix decomposition is as follows: calculate the joint covariance matrix of the noise distribution maps of the text modality and the image modality, perform eigenvalue decomposition on the matrix, and extract the eigenvector corresponding to the maximum eigenvalue as the distribution alignment reference direction. The elimination of distribution offset is completed by projecting the noise distribution map to the orthogonal complement space of the alignment reference direction. For example, the component of the image modality noise distribution that is in the text modality reference direction is orthogonally projected and eliminated. The noise consistency constraint is generated by a linear combination of the distribution difference measure after alignment and the sliding window mean. The combination coefficient is dynamically allocated according to the proportion of the noise energy of each modality. For example, if the noise energy of the text modality accounts for 40%, the image modality accounts for 30%, and the time series modality accounts for 30%, then the combination coefficients are 0.4, 0.3, and 0.3 respectively.

[0032] The weight coefficients of redundant eigenvectors are dynamically adjusted based on the noise consistency constraint, and linear projection correction is performed on components whose cross-modal correlation exceeds a preset condition. The dynamic adjustment rule for the weight coefficients is as follows: when the value of the noise consistency constraint exceeds a preset threshold, the weight of the corresponding dimension in the redundant eigenvector is exponentially decayed according to the excess ratio. For example, if the preset threshold is 0.2 and the constraint value is 0.25, the excess ratio is 25%, and the weight coefficient is updated to a negative exponential function of the original value multiplied by a natural constant, where the exponent is the ratio of the excess ratio to the threshold.

[0033] Linear projection correction is achieved by constructing an orthogonal projection matrix. Specifically, the Gram-Schmidt orthogonalization method is used to generate orthogonal basis vectors based on independent eigenvectors. Components of the redundant eigenvectors with a cosine similarity greater than 0.1 to the independent eigenvectors are projected onto the space spanned by these orthogonal basis vectors, retaining the vertical components after projection as the correction result. The cosine similarity threshold of 0.1 was determined through experimental verification, for example, by testing the impact of different thresholds on model accuracy on the training set and selecting the optimal value.

[0034] The corrected redundant eigenvectors are iteratively orthogonally validated against the independent eigenvectors until the cross-modal noise distribution difference metric falls below a preset threshold. The corrected redundant eigenvectors are then output. The specific process for iterative orthogonal validation is as follows: The cosine similarity between the corrected redundant eigenvectors and the independent eigenvectors of each modality is recalculated. If the cross-modal similarity exceeds the preset threshold of 0.1, the weight coefficient of the noise consistency constraint is adjusted based on the excess similarity. For example, if the similarity is 0.12, the weight coefficient is increased by 10% of the excess, and the linear projection correction process is repeated. The initial value of the preset threshold is set to the statistical median of the initial noise distribution difference metric values. For example, if the initial difference metric set is [0.3, 0.25, 0.28], the median is 0.28. The dynamic adjustment rule is: if the difference metric decreases less than 10% over three consecutive iterations, the threshold is decreased by 5% of the current value. For example, if the current threshold is 0.2, it is adjusted to 0.19. The final output of the corrected redundant eigenvector must satisfy the requirement that all cross-modal distribution difference metrics are lower than the dynamically adjusted threshold, and the orthogonality residual of the independent eigenvector decreases by more than 50% before correction, which is a preset convergence condition. This condition is determined through historical data statistics. For example, a 50% decrease in the residual in 80% of the test cases can meet the model requirements.

[0035] S4. Based on the corrected redundant eigenvectors and independent eigenvectors, the gradient direction conflict rate of each mode is determined through the vector space projection relationship and the weight distribution deviation is generated, including: In step S4, based on the corrected redundant eigenvectors and independent eigenvectors, the gradient direction conflict rate of each mode is determined by the vector space projection relationship and the weight distribution deviation is generated. Specifically, the gradient direction conflict rate of the corrected redundant eigenvectors and independent eigenvectors of each mode is calculated, and the cross-modal gradient difference is quantified by cosine similarity and directional angle. The cosine similarity is calculated by aligning the redundant eigenvectors and independent eigenvectors of the same mode according to the dimension, multiplying them element by element and summing them, and then dividing them by the product of the module lengths of the two vectors. For example, when the dimension of the redundant eigenvector is 128, the sum of the products of the corresponding elements is calculated and divided by the L2 norm product of the two vectors. The directional angle is converted from cosine similarity to an angle value by the inverse cosine function. For example, when the cosine similarity is 0.1, the corresponding angle is 84 degrees. The angle threshold is determined by experimental verification.

[0036] A weighted distribution deviation matrix is generated based on the gradient direction conflict rate, and the gradient distribution consistency of each modality is calculated using a probability density function. The probability density function is constructed by dividing the gradient direction conflict rate of each modality into bins according to a preset numerical range. For example, the conflict rate range of 0 to 1 is divided into 50 equal intervals, and the frequency of conflict rate within each bin is calculated as the probability density. The weighted distribution deviation matrix is generated by performing a cross-modal comparison of the gradient distribution probability density of text, image, and time series modalities, and calculating the sum of the absolute differences in the corresponding bin probability densities between each modality. For example, if the probability densities of text and image modalities in the bin interval [0, 0.02] are 0.05 and 0.08, respectively, the difference is 0.03. The deviation matrix is then accumulated from all bin differences. The number of bins and the range of intervals are set based on the distribution characteristics of historical training data.

[0037] The weight coefficients for each modal feature fusion are dynamically adjusted based on the weighted distribution deviation matrix. Exponential decay compensation is applied to modalities whose gradient conflict rates exceed a dynamic threshold. The dynamic threshold is set by taking the weighted sum of the moving average and standard deviation of the gradient conflict rates over the past iterations. For example, if the mean conflict rate for the last five iterations is 0.15 and the standard deviation is 0.03, the dynamic threshold is set to 0.15 plus twice the standard deviation, or 0.21.

[0038] The specific steps for weight adjustment are as follows: When the conflict rate of a modality exceeds the dynamic threshold, its weight coefficient is multiplied by a negative exponential function with a natural constant as the base, where the exponent is the ratio of the excess value to the threshold. For example, if the conflict rate is 0.25 and the threshold is 0.21, the excess ratio is (0.25 - 0.21) / 0.21 ≈ 0.19. The compensation amplitude of the exponential decay has a nonlinear relationship with the excess ratio, ensuring rapid convergence of the weights of high-conflict modalities.

[0039] The adjusted feature fusion weight coefficients are substituted into the vector space projection relationship for iterative verification until the variance of the gradient direction conflict rate of each modality is lower than the preset convergence condition, and the final weight distribution result is output. The verification method of the vector space projection relationship is: the independent feature vectors of each modality are weighted and summed according to the adjusted weight coefficients to generate a fusion feature vector, and the cosine similarity between the fusion vector and the redundant feature vectors of each modality is calculated. If the similarity variance exceeds the preset convergence condition, for example, the variance threshold is set to 0.01, the gradient direction conflict rate is recalculated and the deviation matrix is updated. The iterative termination condition is that the variance values of three consecutive iterations are all lower than the threshold, or the maximum number of iterations is 10 times, which is set according to the scale of the training data, for example, the data volume of 10,000 corresponds to an upper limit of 10 times. The output requirements of the final weight distribution result are that the adjustment range of the weight coefficient of each modality is no more than 20% compared with the initial value, and the maximum value of the cross-modal conflict rate is lower than 80% of the dynamic threshold. This ratio is determined by cross-validation.

[0040] S5. Dynamically adjust the gradient back propagation path according to the weight distribution deviation, including: In step S5, the gradient back propagation path is dynamically adjusted according to the weight distribution deviation. Specifically, the influence factor of each modal gradient on the back propagation path is calculated based on the weight distribution deviation matrix, and the path deviation strength is quantified by the Hadamard product of the deviation and the gradient. The Hadamard product is calculated by multiplying the weight distribution deviation matrix and the gradient matrix element by element. For example, for a deviation matrix with a dimension of 128×128 and a gradient matrix of the same dimension, the elements at corresponding positions are multiplied to generate an influence factor matrix. Each element value in the matrix represents the deviation strength of the corresponding path node. The normalization of the influence factor is achieved by dividing by the maximum absolute value of the matrix elements. For example, when the maximum absolute value is 0.5, all element values are scaled to the interval [-1,1].

[0041] A gradient path mask matrix is generated based on the impact factors, and probabilistic threshold filtering is performed on path nodes with deviation strengths above a dynamic threshold. The dynamic threshold is set by taking the median of the moving average of the impact factor matrix over the past iterations as the baseline threshold. For example, if the median of the last five iterations is 0.3, the dynamic threshold is set to 0.3.

[0042] Probabilistic threshold filtering is implemented by generating a binary mask for nodes in the influence factor matrix whose values exceed a dynamic threshold. For example, positions with values greater than 0.3 are marked as 0 (masked), and all others are marked as 1 (retained). The frequency of mask generation is adjusted based on the depth of the network layer. For example, shallow networks update the mask every two iterations, while deep networks update it every iteration.

[0043] Backpropagation path weights are dynamically adjusted based on the gradient path mask matrix, preserving the gradient update paths for low-deviation nodes through a sparse connection matrix. The sparse connection matrix is generated by element-wise multiplication of the mask matrix with the original gradient matrix, masking the gradient values of high-deviation nodes. For example, the gradient values corresponding to the positions marked as 0 in the mask matrix are set to zero.

[0044] The specific steps for dynamically adjusting weights are: multiply the gradient values of the retained low-deviation nodes by a decay factor. The decay factor is set based on the depth of the network layer in which the node is located. For example, the decay factor for the first layer is 0.9, and the decay factor decreases by 0.1 with each deeper layer, with a minimum of 0.5. The relationship between the decay factor and network layer depth is verified through experimental verification, for example, by testing the effect of different decay factors on the model convergence speed on the training set and then selecting the optimal ratio.

[0045] The adjusted backpropagation path is substituted into the network's feedforward calculation for iterative validation until the variance of the path weight change falls below a convergence threshold. The optimized gradient propagation path is then output. The iterative validation process is as follows: During the feedforward calculation, the change in the weight matrix of each layer is recorded, its variance is calculated, and compared with a convergence threshold, for example, 10% of the initial weight change variance. If the variance exceeds the threshold, the influence factor matrix is recalculated and the mask matrix is updated. For example, recalculation is triggered when the variance reaches 15% of the initial value. The iteration terminates when the variance of three consecutive validations falls below the threshold or when a preset maximum number of iterations (for example, 10) is reached. This number is set based on the number of network layers, for example, a 10-layer network has a maximum of 10 iterations. The final optimized gradient propagation path must ensure that the weight change of each layer does not exceed 15% compared to the initial value, and the proportion of nodes retained in the mask matrix is at least 20% of the total number of nodes, as determined through cross-validation.

[0046] S6. Based on the adjusted gradient backpropagation path, monitor the matching degree between text and image at the semantic abstraction level and the balance of contribution of temporal signals in cross-modal fusion, including: In step S6, based on the adjusted gradient backpropagation path, the matching degree between text and image at the semantic abstraction level and the balance of the contribution of time series signals in cross-modal fusion are monitored. Specifically, the matching degree between the text modality and the image modality at the semantic abstraction level is calculated, and the cross-modal semantic alignment error is quantified using the attention weight matrix and cosine similarity. The attention weight matrix is generated by aligning the independent feature vectors of the text modality with the independent feature vectors of the image modality according to their dimensions and then multiplying them element-by-element to generate the original correlation matrix. The matrix is then normalized along the rows using the Softmax function. The temperature coefficient during normalization is set to the square root of the feature vector dimension to avoid overly sharp probability distributions after normalization. The semantic alignment error is calculated by taking the weighted sum of the mean of the main diagonal elements in the attention weight matrix and the cosine similarity, with a weight ratio of 3:1. For example, when the main diagonal mean is 0.6 and the cosine similarity is 0.8, the semantic alignment error is (0.6×3+0.8×1) / 4=0.65.

[0047] To analyze the balanced contribution of time series signals in cross-modal fusion, the joint distribution matrix of time series features and text image features is compared using correlation coefficients and distribution offsets. The correlation coefficient is calculated using the Pearson correlation coefficient. Specifically, the covariance between the time series feature vector and the text image fusion feature vector is calculated and then divided by the product of their standard deviations. For example, if the covariance is 0.12, the standard deviation of the text image features is 0.4, and the standard deviation of the time series features is 0.3, then the correlation coefficient is 0.12 / (0.4×0.3)=1.0. The distribution offset is calculated using the KL divergence. The joint distribution matrix of time series features and text image features is discretized according to a preset number of bins. The number of bins is dynamically adjusted based on the feature dimension. For example, when the feature dimension is 256, it is divided into 20 bins. The value interval of each bin is uniformly distributed. After calculating the joint probability of the data points in each bin, the sum of the logarithmic difference of the corresponding bin probability values of the two distributions and the product of the probabilities is calculated.

[0048] Based on the cross-modal semantic alignment error and contribution balance, cross-modal balance constraints are generated. Inter-modal contribution deviations are eliminated through matrix decomposition and dynamic weight adjustment. The matrix decomposition involves performing singular value decomposition (SVD) on the joint distribution matrix of text and image. The left and right singular vectors corresponding to the first k singular values are extracted to construct a low-rank approximation matrix. The value k is set to 50% of the matrix rank. For example, if the original matrix rank is 100, k=50.

[0049] The dynamic weight adjustment rule is as follows: when the semantic alignment error exceeds the dynamic threshold, the fusion weights of the text and image modalities are asymmetrically attenuated according to the error excess ratio. For example, when the error exceeds the threshold by 20%, the text weight is attenuated by 0.85, and the image weight is attenuated by 0.9. The difference in attenuation factors is 0.05, which is determined by the proportion of modality contribution deviations in historical training. The dynamic threshold is initially set to the moving average of the semantic alignment error of the first 10 iterations and is updated every 5 iterations thereafter, with the new threshold being 90% of the current moving average.

[0050] The parameters of the gradient backpropagation path are updated based on the cross-modal balance constraint to ensure that the monitoring results of the semantic match and contribution deviation meet the preset conditions. The parameter update is implemented by adding the balance constraint as an additional regularization term to the loss function. The regularization coefficient is dynamically adjusted based on the variance of the semantic match of the current iteration. For example, when the variance is 0.1, the coefficient is 0.05, and the coefficient increases by 0.005 for every 0.01 increase in variance.

[0051] Verification of the gradient backpropagation path is accomplished by recording the changes in parameters at each layer during the feedforward calculation process. The variance of the parameter changes is calculated by taking the element-wise squared difference of the updated values of the weight matrix at each layer, taking the mean, and comparing it with the initial variance. For example, if the initial variance is 0.2, a current variance of 0.15 would satisfy the convergence condition. The pre-defined criteria are: the semantic match variance must be less than 50% of the initial value, and the contribution deviation coefficient must be less than 15% of the cross-modal weight mean. The weight mean is calculated by averaging the results of the last five iterations using a sliding window.

[0052] S7. If the matching degree is lower than the preset threshold or the contribution balance deviates from the preset range, the fusion path switching instruction is triggered to construct a classification decision boundary function to output the classification result, including: In step S7, if the matching degree is lower than the preset threshold or the contribution balance deviates from the preset interval, the fusion path switching instruction is triggered to construct a classification decision boundary function to output the classification result. Specifically, the semantic matching degree of the text and image modality is monitored to see if it is lower than the preset threshold, and the contribution balance of the time series signal is detected to see if it deviates from the preset interval. The preset threshold is set as follows: the semantic matching threshold is the moving average of the results of the first 10 iterations. For example, if the average of the last 10 matching degrees is 0.65, the preset threshold is 0.65; the preset interval of contribution balance is determined based on ±15% of the historical mean of the KL divergence of the joint distribution matrix of the time series features and the text image features. For example, when the KL divergence mean is 0.25, the interval is [0.2125, 0.2875]. During the monitoring process, if the semantic matching degree is lower than the threshold for two consecutive iterations or the contribution balance deviates from the interval once, the trigger condition is determined to be met.

[0053] When the trigger condition is met, a fusion path switching instruction is generated, switching the cross-modal fusion path through dynamic weight reorganization and feature channel masking. Dynamic weight reorganization is implemented by redistributing weights based on the historical contribution ratio of each modality. For example, if the historical contributions of text, image, and time series modalities are 35%, 40%, and 25%, respectively, the new weight coefficients are normalized proportionally to [0.35, 0.40, 0.25]. The feature channel mask is generated as follows: a binary mask is generated for channels in the redundant feature vector whose cross-modal correlation exceeds a dynamic threshold. The correlation threshold is determined by the 90th percentile of the historical correlation coefficient. For example, when the historical correlation quantile is 0.18, the threshold is set to 0.18. The number of channels marked as 0 in the mask matrix is the number of channels in the top 15% of the correlation.

[0054] Based on the switched fusion path, a classification decision boundary function is constructed. A classification hyperplane is generated through kernel mapping of the support vector machine and logistic regression probability calibration. A Gaussian kernel function is used for kernel mapping, and the bandwidth parameter is automatically adjusted based on the variance of the feature vector. For example, when the feature variance is 0.2, the bandwidth parameter is set to 50% of the inverse of the variance. Logistic regression probability calibration is implemented by converting the support vector machine decision values into probabilities using a sigmoid function. The temperature coefficient is adjusted based on the classification accuracy of the validation set. For example, when the accuracy is 85%, the temperature coefficient is set to 1.0, and the temperature coefficient is reduced by 0.1 for every 5% decrease in accuracy. The optimization goal of the classification hyperplane is to maximize the margin while minimizing the probability calibration error. The hyperplane parameters are solved using the sequential minimization algorithm, with a maximum number of iterations set to 1000.

[0055] The classification results are substituted into the validation set for a decision boundary stability test. If the classification confidence variance exceeds the convergence criterion, path switching is retriggered until the output classification result meets the preset accuracy. The specific process of the decision boundary stability test is as follows: the classification confidence variance of the validation set samples is calculated, for example, the variance of the probability values of 1000 samples is calculated. If the variance exceeds 0.03, the classification is considered unstable. The convergence criterion is set to 40% of the initial variance. For example, if the initial variance is 0.1, the convergence criterion is 0.04. The rule for retriggering path switching is as follows: if the variance exceeds the limit for three consecutive tests, the system returns to the dynamic weight reorganization step and retains the previous optimal weight coefficients as the initial values. The final classification result output must meet the validation set accuracy requirement of at least 90% of the historical best value. For example, if the historical best accuracy is 88%, the final result must reach at least 79.2%.

[0056] Example 2: Figure 2 The present invention provides a structural diagram of a system for automatic classification of heterogeneous data based on business deep learning, which includes: Feature decoupling module: performs implicit feature decoupling of unstructured data such as text, images, and time series signals within the modality, separating the independent feature vectors and redundant feature vectors of each modality; Residual evaluation module: It evaluates the decoupling effectiveness based on the orthogonality of the independent eigenvectors and the redundant eigenvectors and generates noise correction instructions; Noise correction module: performs cross-modal noise distribution consistency correction on the redundant feature vector according to the noise correction instruction to obtain the corrected redundant feature vector; Weight allocation module: Based on the corrected redundant eigenvectors and independent eigenvectors, the gradient direction conflict rate of each mode is determined through the vector space projection relationship and the weight allocation deviation is generated; Path adjustment module: dynamically adjusts the gradient back propagation path according to the weight distribution deviation; Balance monitoring module: Based on the adjusted gradient backpropagation path, it monitors the matching degree between text and image at the semantic abstraction level and the balance of contribution of time series signals in cross-modal fusion; Classification decision module: If the matching degree is lower than the preset threshold or the contribution balance deviates from the preset interval, the fusion path switching instruction is triggered to construct a classification decision boundary function to output the classification result.

[0057] The above formulas are all dimensionless and numerical calculations. The formula is a formula that is closest to the actual situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.

[0058] It should be noted that the present invention can be deployed on the device itself to implement embedded applications, and can also be run on a PC or other terminal with a user interface, thereby meeting various hardware environments and usage requirements.

[0059] The above embodiments can be implemented in whole or in part via software, hardware, firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product comprises one or more computer instructions or computer programs. When loaded or executed on a computer, the processes or functions described in the embodiments of this application are fully or partially performed. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired means (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium accessible by a computer or a data storage device such as a server or data center that contains a collection of one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, tapes), optical media (e.g., DVDs), or semiconductor media. The semiconductor media can be a solid-state drive.

[0060] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and modules described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0061] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the modules is only a logical function division. In actual implementation, there may be other division methods, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0062] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, and may be located in one place or distributed across multiple network modules. Some or all of the modules may be selected to achieve the purpose of this embodiment according to actual needs.

[0063] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0064] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as 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.

[0065] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

[0066] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for automatic classification of heterogeneous data based on business deep learning, characterized in that: The steps include: S1. Perform implicit feature decoupling within the modality of unstructured data such as text, images, and time series signals, separating the independent feature vectors and redundant feature vectors of each modality; S2, evaluating the effectiveness of decoupling based on the orthogonality residuals of the independent eigenvectors and the redundant eigenvectors and generating noise correction instructions; S3, performing cross-modal noise distribution consistency correction on the redundant feature vector according to the noise correction instruction to obtain a corrected redundant feature vector; S4. Based on the corrected redundant eigenvectors and independent eigenvectors, the gradient direction conflict rate of each mode is determined through the vector space projection relationship and the weight distribution deviation is generated; S5. Dynamically adjust the gradient back propagation path according to the weight distribution deviation; S6. Based on the adjusted gradient backpropagation path, monitor the matching degree between text and image at the semantic abstraction level and the balance of contribution of temporal signals in cross-modal fusion; S7. If the matching degree is lower than the preset threshold or the contribution balance deviates from the preset interval, the fusion path switching instruction is triggered to construct a classification decision boundary function to output the classification result.

2. The method for automatic classification of heterogeneous data based on business deep learning according to claim 1 is characterized in that: Perform implicit feature decoupling of unstructured data such as text, images, and time series signals within the modality, separating the independent and redundant feature vectors of each modality, including: Perform word vector embedding and syntactic dependency parsing on text data to generate the initial implicit feature vector of the text modality; Perform multi-scale perceptual decomposition of convolution kernels on image data to extract the initial implicit feature vector of the image modality; Perform windowed Fourier transform and nonlinear activation mapping on the time series signal to generate the initial implicit feature vector of the time series modality; The initial implicit feature vectors of each modality are input into the sparse autoencoder respectively, and the independence weights between the features of each modality are calculated through the intra-modal orthogonal constraint function to separate the independent feature vectors and redundant feature vectors of each modality.

3. The method for automatic classification of heterogeneous data based on business deep learning according to claim 1 is characterized in that: The decoupling effectiveness is evaluated based on the orthogonality residual of the independent eigenvectors and the redundant eigenvectors, and noise correction instructions are generated, including: Calculate the orthogonality residual between the independent eigenvectors and redundant eigenvectors of each mode, and quantify the decoupling error by the difference between the vector inner product mean and the preset dynamic threshold; Generate noise correction instructions based on the orthogonality residual distribution. When the orthogonality residual exceeds the preset residual threshold, perform weight decay on the redundant feature vectors and trigger iterative decoupling of the sparse autoencoder of the corresponding mode. The orthogonality constraint term is updated based on the corrected redundant eigenvectors, and the residual cross-modal correlation in the independent eigenvectors is eliminated through the linear transformation matrix. The updated independent eigenvectors and redundant eigenvectors are re-orthogonally projected to generate the final decoupled output.

4. The method for automatic classification of heterogeneous data based on business deep learning according to claim 1 is characterized in that: According to the noise correction instruction, the redundant feature vector is corrected for cross-modal noise distribution consistency to obtain a corrected redundant feature vector, including: The cross-modal noise distribution of redundant feature vectors of each modality is extracted based on the noise correction instruction, and the noise distribution map is constructed by measuring the sliding window mean and distribution difference. The noise distribution maps of different modes are aligned across modalities, the distribution offset is eliminated through covariance matrix decomposition, and noise consistency constraints are generated; Dynamically adjust the weight coefficients of redundant eigenvectors according to the noise consistency constraint term, and perform linear projection correction on components whose cross-modal correlation is higher than the preset condition; The corrected redundant eigenvector is iteratively orthogonally verified with the independent eigenvector until the distribution difference measure of the cross-modal noise distribution is lower than the preset threshold, and the corrected redundant eigenvector is output.

5. The method for automatic classification of heterogeneous data based on business deep learning according to claim 1 is characterized in that: Based on the corrected redundant eigenvectors and independent eigenvectors, the gradient direction conflict rate of each mode is determined through the vector space projection relationship and the weight distribution deviation is generated, including: Calculate the gradient direction conflict rate between the corrected redundant eigenvectors and independent eigenvectors of each modality, and quantify the cross-modal gradient difference through cosine similarity and directional angle; The weight distribution deviation matrix is generated based on the gradient direction conflict rate, and the gradient distribution consistency of each mode is statistically analyzed through the probability density function; Dynamically adjust the fusion weight coefficients of each modal feature according to the weight distribution deviation matrix, and perform exponential decay compensation on the modalities whose gradient direction conflict rate is higher than the dynamic threshold; The adjusted feature fusion weight coefficient is substituted into the vector space projection relationship for iterative verification until the variance of the conflict rate of each modal gradient direction is lower than the preset convergence condition, and the final weight distribution result is output.

6. The method for automatic classification of heterogeneous data based on business deep learning according to claim 1 is characterized in that: Dynamically adjust the gradient backpropagation path based on the weight distribution deviation, including: The influence factor of each modal gradient on the back propagation path is calculated based on the weight distribution deviation matrix, and the path deviation intensity is quantified by the Hadamard product of the deviation and the gradient; Generate a gradient path mask matrix based on the influencing factors, and perform probability threshold filtering on path nodes whose deviation intensity is higher than the dynamic threshold; Dynamically adjust the backpropagation path weights based on the gradient path mask matrix, and retain the gradient update channels of low-deviation nodes through the sparse connection matrix; The adjusted back propagation path is substituted into the network feedforward calculation for iterative verification until the variance of the path weight change is lower than the convergence threshold, and the optimized gradient propagation path is output.

7. The method for automatic classification of heterogeneous data based on business deep learning according to claim 1 is characterized in that: Based on the adjusted gradient backpropagation path, we monitor the matching degree between text and image at the semantic abstraction level and the balance of contribution of temporal signals in cross-modal fusion, including: Calculate the matching degree between text modality and image modality at the semantic abstraction level, and quantify the cross-modal semantic alignment error using the attention weight matrix and cosine similarity. Analyze the contribution balance of time series signals in cross-modal fusion, and compare the joint distribution matrix of time series features and text image features through correlation coefficients and distribution offsets; Generate cross-modal balance constraints based on cross-modal semantic alignment error and contribution balance, and eliminate inter-modal contribution deviations through matrix decomposition and dynamic weight adjustment; The parameters of the gradient backpropagation path are updated based on the cross-modal balance constraint term to ensure that the monitoring results of semantic matching and contribution deviation meet the preset conditions.

8. The method for automatic classification of heterogeneous data based on business deep learning according to claim 1 is characterized in that: If the matching degree is lower than the preset threshold or the contribution balance deviates from the preset range, the fusion path switching instruction is triggered to build a classification decision boundary function to output the classification result, including: Monitor whether the semantic matching degree between text and image modalities is lower than a preset threshold, and detect whether the balance of time series signal contribution deviates from a preset range; When the trigger matching degree is lower than the preset threshold or the contribution balance deviates from any condition in the preset range, a fusion path switching instruction is generated and the cross-modal fusion path is switched; Based on the switched fusion path, a classification decision boundary function is constructed, and a classification hyperplane is generated through support vector machine kernel function mapping and logistic regression probability calibration; The classification results are substituted into the validation set for a decision boundary stability test. If the classification confidence variance exceeds the convergence condition, the path switching is retriggered until the classification result that meets the preset accuracy is output.

9. The method for automatic classification of heterogeneous data based on business deep learning according to claim 8 is characterized in that: Switching cross-modal fusion paths is achieved through dynamic weight reorganization and feature channel masking.

10. A system for automatic classification of heterogeneous data based on business deep learning, used to implement the method for automatic classification of heterogeneous data based on business deep learning according to any one of claims 1 to 9, characterized in that: include: Feature decoupling module: performs implicit feature decoupling of unstructured data such as text, images, and time series signals within the modality, separating the independent feature vectors and redundant feature vectors of each modality; Residual evaluation module: It evaluates the decoupling effectiveness based on the orthogonality of the independent eigenvectors and the redundant eigenvectors and generates noise correction instructions; Noise correction module: performs cross-modal noise distribution consistency correction on the redundant feature vector according to the noise correction instruction to obtain the corrected redundant feature vector; Weight allocation module: Based on the corrected redundant eigenvectors and independent eigenvectors, the gradient direction conflict rate of each mode is determined through the vector space projection relationship and the weight allocation deviation is generated; Path adjustment module: dynamically adjusts the gradient back propagation path according to the weight distribution deviation; Balance monitoring module: Based on the adjusted gradient backpropagation path, it monitors the matching degree between text and image at the semantic abstraction level and the balance of contribution of time series signals in cross-modal fusion; Classification decision module: If the matching degree is lower than the preset threshold or the contribution balance deviates from the preset interval, the fusion path switching instruction is triggered to construct a classification decision boundary function to output the classification result.

Images