Multi-source heterogeneous data fusion and intelligent analysis system and method

Abstract

The application is suitable for the technical field of intelligent analysis of multi-source heterogeneous data, and provides a multi-source heterogeneous data fusion and intelligent analysis system and method, wherein the multi-source heterogeneous data fusion and intelligent analysis method comprises the following steps: S1, obtaining multi-source heterogeneous original data, performing dimensionless standardization processing on the multi-source heterogeneous original data containing different physical dimensions and sampling frequencies, and mapping and projecting the multi-source heterogeneous original data to a unified dimension feature subspace to obtain a dimensionless unified dimension feature vector; S2, calculating an adaptive manifold distance between different heterogeneous data sources based on the unified dimension feature vector, and then constructing a high-order interaction covariance across modalities. By calculating the adaptive manifold distance between the heterogeneous data sources, the high-order interaction covariance matrix containing all the outer products of two-by-two combinations is constructed, and the fusion contribution of each data source pair is dynamically adjusted in an exponential weighting manner.

Description

Technical Field

[0001] This invention belongs to the field of intelligent analysis technology for multi-source heterogeneous data, and particularly relates to a system and method for multi-source heterogeneous data fusion and intelligent analysis. Background Technology

[0002] With the widespread application of IoT and Industrial Internet technologies, the data generated in various monitoring scenarios is experiencing explosive growth, encompassing multimodal forms such as physical dimension data collected by sensors, equipment operation logs, and environmental monitoring information. These data sources differ significantly in physical dimensions, time scales, and forms of expression, making it difficult for traditional data processing frameworks to effectively integrate their inherent relationships.

[0003] In current technical practices, data fusion solutions mainly rely on strategies such as feature splicing, weighted averaging, or decision integration. However, these methods face three structural obstacles when dealing with multi-source heterogeneous data.

[0004] First, existing fusion mechanisms lack the ability to effectively model the nonlinear high-order interaction relationships between data sources. When the number of data sources increases, the second-order and higher-order cross-coupling effects cannot be fully represented by linear combination, resulting in a lack of information in the description of the overall system state by the fusion features.

[0005] Secondly, the fusion process is disconnected from the dynamic evolution of the system. Most schemes directly input the fusion results into the static prediction model, ignoring the temporal dependency between the fusion characteristics as an external driving force and the system state transition, resulting in prediction results that violate the inherent evolutionary constraints of the physical system.

[0006] Finally, the feature fusion module and the prediction model adopt a separate optimization architecture. The adjustment target of the fusion parameters is not directly related to the prediction accuracy, and the prediction error cannot correct the fusion strategy in reverse, forming an open-loop breakpoint in parameter optimization. This fragmented design limits the overall system performance to local optima, making it difficult to adapt to the dynamic interaction characteristics of multi-source data under complex working conditions.

[0007] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0008] The purpose of this invention is to provide a method for multi-source heterogeneous data fusion and intelligent analysis, aiming to solve the above-mentioned problems.

[0009] This invention is implemented as follows: a method for multi-source heterogeneous data fusion and intelligent analysis, comprising the following steps: S1. Obtain multi-source heterogeneous raw data, perform dimensionless standardization on the multi-source heterogeneous raw data containing different physical dimensions and sampling frequencies, and map and project it onto a feature subspace of a unified dimension to obtain a dimensionless unified dimension feature vector. S2. Based on the unified dimension feature vector, calculate the adaptive manifold distance between different heterogeneous data sources, and then construct a cross-modal high-order interactive covariance to realize dynamic interactive fusion between multi-source information and output a global interactive fusion feature matrix. S3. After reducing the dimensionality of the global interactive fusion feature matrix, use it as the external driving input force and substitute it into the discretized nonlinear dynamic model. Combined with the internal state transition self-evolution mechanism of the system, predict and output the dimensionless vector of the predicted system state at the next moment. S4. Obtain the dimensionless vector of the real observation system state provided by the external observation platform, and combine the dimensionless vector of the predicted system state with the adaptive constraint of the global interactive fusion feature matrix to construct a joint objective function to optimize the closed-loop parameters of the discretized nonlinear dynamic model. The closed-loop parameter optimization adopts the gradient descent method, and the learning rate decays exponentially with the number of iterations. The optimization stops when the decrease of the objective function in several consecutive iterations is less than a preset threshold.

[0010] In a further technical solution, in step S1, the difference vector is obtained by subtracting the mathematical expectation vector of the data source within the historical sliding window from the engineering feature vector with the original physical dimensions input at time t. Each component of the standard deviation vector of the data source within the historical sliding window is added with a minimal normal number introduced to prevent the denominator from being zero, resulting in a sum vector. The difference vector is then divided by the sum to obtain a standardized vector. Finally, the standardized vector is multiplied on the left by the orthogonal projection matrix corresponding to the data source to obtain the dimensionless uniform dimension feature vector of the data source at time t.

[0011] In a further technical solution, step S2, the specific method for outputting the global interactive fusion feature matrix is ​​as follows:

[0012] For all heterogeneous data sources, a double summation is performed in pairs. For each pair of data sources, first, the dimensionless unified dimension eigenvector of the first data source is multiplied by the transpose of the dimensionless unified dimension eigenvector of the second data source to obtain the outer product matrix. Then, the square of the Euclidean distance between the eigenvectors of the pair of data sources is calculated and the negative half is taken as the exponent. The exponential function value is calculated with the natural constant e as the base to obtain the distance weight. The outer product matrix is ​​multiplied by the distance weight to obtain the interaction contribution matrix of the pair of data sources. The interaction contribution matrices of all data source pairs are summed to obtain the dimensionless global interaction fusion feature matrix at time t.

[0013] A further technical solution, wherein the specific method for using the dimensionality-reduced global interactive fusion feature matrix as the external driving input force and predicting the dimensionless vector of the system state at the next moment is as follows:

[0014] First, calculate the dimensionless vector of the system state at the current moment and multiply it by the dimensionless internal state transition matrix that characterizes the self-evolution law of the system to obtain the self-evolution contribution term. Then, calculate the global interactive fusion feature matrix and multiply it by the learnable attention vector to obtain the weighted fusion vector. Use a dimensionless linear transformation matrix to reduce the dimension of the weighted fusion vector to vector form. Then, take the hyperbolic tangent function value of each element of the vector to obtain the nonlinear driving term. Multiply the nonlinear driving term by the dimensionless external control gain matrix to obtain the external driving contribution term. Add the self-evolution contribution term and the external driving contribution term, multiply it by the discretized dimensionless time step, and finally add the dimensionless vector of the system state at the current moment to obtain the dimensionless vector of the predicted system state at the next moment.

[0015] In a further technical solution, the specific calculation method of the joint objective function in step S4 is as follows:

[0016] For each sample in the batch, first calculate the squared Euclidean distance between the dimensionless vector of the predicted system state and the dimensionless vector of the actual observed system state, as the prediction error term; then calculate the transpose of the global interactive fusion feature matrix of the sample, multiply it by the global interactive fusion feature matrix, take the trace of the product matrix, and multiply it by the dimensionless regularization weight coefficient, as the regularization term; add the prediction error term and the regularization term of the sample to obtain the single-sample loss of the sample; add the single-sample losses of all samples and divide by the total number of samples in the batch to obtain the joint objective function value.

[0017] A further technical solution involves introducing a comprehensive external environmental impact factor in step S4 and constructing a joint objective function with adaptive penalty weights; the specific method is as follows: First, calculate the dimensionless comprehensive impact factor: For each external influencing factor, first calculate the original physical quantity observation value of the influencing factor in the current sample, subtract the expected value of the influencing factor, and then divide by the standard deviation of the influencing factor to obtain the standardized value; take the Sigmoid activation function value of the standardized value, and then multiply it by the attention weight of the influencing factor; sum up the weighted results of all influencing factors to obtain the comprehensive impact factor of the sample. Next, calculate the joint objective function value with adaptive penalty weights: For each sample in the batch, first calculate the exponential function value with the natural constant e as the base and the comprehensive influence factor of the sample as the exponent. Multiply the exponential function value by the prediction error term of the sample to obtain the weighted prediction error term. Then calculate the regularization term of the sample. Add the weighted prediction error term and the regularization term to obtain the weighted single-sample loss of the sample. Sum the weighted single-sample losses of all samples and divide by the total number of samples in the batch to obtain the joint objective function value with adaptive penalty weights.

[0018] A multi-source heterogeneous data fusion and intelligent analysis system, applied to the above methods, includes: The dimensionless standardization module is used to perform dimensionless standardization processing on multi-source heterogeneous raw data and map it to a unified dimension feature subspace, outputting a dimensionless unified dimension feature vector; the dimensionless standardization module also includes a multi-source data time synchronization submodule, which is used to perform time alignment and resampling on raw data with different sampling frequencies. The cross-modal dynamic fusion module is used to calculate the adaptive manifold distance based on the unified dimension feature vector, construct a high-order interactive covariance for dynamic fusion, and output a global interactive fusion feature matrix. The dynamics prediction module is used to reduce the dimensionality of the global interactive fusion feature matrix and use it as an external driving input force, which is then substituted into the discretized nonlinear dynamics model and combined with the state transition self-evolution mechanism to predict the system state vector at the next moment. The closed-loop optimization module is used to obtain the state vector of the real observed system, and construct a joint objective function by combining the predicted state and the adaptive constraints of the fused feature matrix to optimize the closed-loop parameters of the dynamic model.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0020] This invention constructs a high-order interactive covariance matrix containing all pairwise cross products by calculating the adaptive manifold distance between heterogeneous data sources, and dynamically adjusts the fusion contribution of each data source pair in an exponentially weighted manner. Compared with traditional feature concatenation or weighted averaging, this mechanism can explicitly capture the high-order cross-coupling effect between data sources, and can still fully retain nonlinear interactive information when the number of data sources increases. The fused features provide a more comprehensive and accurate description of the overall system state.

[0021] This invention uses the dimensionality-reduced global interactive fusion feature matrix as an external driving force, embedding it into a discretized nonlinear dynamic model. This force, together with the system's internal state transition self-evolution mechanism, determines the state prediction for the next time step. This design allows the fusion result to directly participate in the state update equation, rather than being independent of the temporal process. This ensures that the modulation effect of external observation information on the system state follows physical evolution constraints, and the predicted trajectory more closely reflects the intrinsic change law of the real system.

[0022] This invention constructs a joint objective function, unifying the prediction error and the regularization constraints of the fused feature matrix into the optimization process. This allows the prediction bias to update the orthogonal projection matrices of each data source and the dynamic model parameters through gradient descent, thereby indirectly optimizing the feature representations in the standardization and fusion stages. The closed-loop parameter optimization mechanism breaks the limitation of independent tuning of each module in the traditional separate architecture, ensuring that the adjustment target of the fusion parameters is directly related to the prediction accuracy, and the overall system performance approaches the global optimum.

[0023] This invention introduces a comprehensive external environmental influence factor to construct a joint objective function with adaptive penalty weights. When external environmental disturbances are severe, the comprehensive influence factor nonlinearly amplifies the weight of the prediction error term through an exponential function, driving the model to prioritize prediction bias under adverse conditions during gradient descent optimization. This mechanism enables the system to adaptively adjust its learning strategy, maintaining high prediction accuracy and robust generalization performance even under complex and changing external conditions. Attached Figure Description

[0024] Figure 1 A flowchart illustrating the steps of a multi-source heterogeneous data fusion and intelligent analysis method; Figure 2 Logic diagram for predicting dynamic models; Figure 3 Construct a logic graph for the joint objective function with adaptive penalty; Figure 4 This is a system architecture diagram. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0026] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0027] like Figures 1-3 As shown, a multi-source heterogeneous data fusion and intelligent analysis method provided by an embodiment of the present invention includes the following steps: S1. Obtain multi-source heterogeneous raw data, perform dimensionless standardization on the multi-source heterogeneous raw data containing different physical dimensions and sampling frequencies, and map and project it onto a feature subspace of a unified dimension to obtain a dimensionless unified dimension feature vector. Among them, multi-source heterogeneous raw data refers to data sets originating from different sensors, databases, or information systems, possessing different physical dimensions, sampling frequencies, data formats, or semantic structures. Dimensionless standardization refers to mathematical transformations applied to the raw data to eliminate the influence of its original physical dimensions, making it numerically comparable. A unified-dimensional feature subspace refers to a common low- or high-dimensional space to which feature vectors from different data sources are mapped through projection or transformation. In this subspace, features from different data sources can be directly compared and fused. Dimensionless unified-dimensional feature vectors refer to the data representation after dimensionless standardization and mapping to a unified-dimensional feature subspace.

[0028] For example, in an industrial equipment monitoring scenario, vibration, temperature, and pressure data can be acquired. This raw data needs preprocessing before direct use. Specifically, dimensionless standardization is performed on the raw data to eliminate the influence of different physical dimensions, making the data numerically comparable. For example, all data can be scaled to the range of 0 to 1. Subsequently, the standardized data is mapped and projected onto a feature subspace of uniform dimension. This mapping process can be achieved through simple linear transformations, such as principal component analysis (PCA) to project high-dimensional data into a lower-dimensional space, thereby obtaining dimensionless feature vectors of uniform dimension.

[0029] S2. Based on the unified dimension feature vector, calculate the adaptive manifold distance between different heterogeneous data sources, and then construct a cross-modal high-order interactive covariance to realize dynamic interactive fusion between multi-source information and output a global interactive fusion feature matrix.

[0030] Among them, adaptive manifold distance refers to a measure of the similarity between feature vectors of two heterogeneous data sources in the feature space. This distance is adaptively calculated using an exponential function, so that data sources that are closer in distance have higher similarity and are therefore given greater weight in the fusion process; conversely, data sources that are farther apart have similarity approaching zero, and their conflicting information is automatically suppressed. Higher-order interaction covariance refers to a fusion feature matrix used to capture nonlinear correlations between cross-data sources. This matrix is ​​obtained by performing an outer product operation on the dimensionless feature vectors of any two data sources from multiple heterogeneous data sources and then weighted summing over all possible combinations. Global interaction fusion feature matrix refers to the comprehensive feature representation output by achieving dynamic interaction fusion of multi-source information after calculating adaptive manifold distance and constructing higher-order interaction covariance. This matrix contains rich nonlinear interaction information between multi-source heterogeneous data.

[0031] S3. After reducing the dimensionality of the global interactive fusion feature matrix, use it as the external driving input force and substitute it into the discretized nonlinear dynamic model. Combined with the internal state transition self-evolution mechanism of the system, predict and output the dimensionless vector of the predicted system state at the next moment. Discretized nonlinear dynamics refers to a mathematical model that describes the discrete evolution of a system's state over time. This model can capture the nonlinear changes within the system and is influenced by external driving forces. The state transition self-evolution mechanism refers to the process in the discretized nonlinear dynamics model where the system state evolves according to its inherent laws without external input. The dimensionless vector predicting the system state refers to the dimensionless representation of the system state at the next moment, predicted by the discretized nonlinear dynamics model based on the current system state and external driving forces.

[0032] S4. Obtain the dimensionless vector of the real observation system state provided by the external observation platform, and combine the dimensionless vector of the predicted system state with the adaptive constraint of the global interactive fusion feature matrix to construct a joint objective function to optimize the closed-loop parameters of the discretized nonlinear dynamic model. The closed-loop parameter optimization adopts the gradient descent method, and the learning rate decays exponentially with the number of iterations. The optimization stops when the decrease of the objective function in several consecutive iterations is less than a preset threshold.

[0033] Among them, the dimensionless vector of the real observed system state refers to the dimensionless representation of the true state of the system at a certain moment, actually measured or acquired by an external observation platform. The joint objective function refers to a function constructed during model training and optimization that comprehensively considers multiple optimization objectives (such as prediction error and fusion feature complexity). By minimizing this function, different aspects of the model can be optimized simultaneously. Closed-loop parameter optimization refers to a model parameter adjustment strategy where the model's output (e.g., prediction results) is compared with the real observations, and the resulting error is fed back to the model for iterative adjustment of model parameters to reduce error and improve performance. Gradient descent is a commonly used optimization algorithm for minimizing the objective function. This method iteratively adjusts the model parameters along the direction of gradient descent of the objective function, gradually approaching the optimal solution.

[0034] This embodiment provides a method for multi-source heterogeneous data fusion and intelligent analysis, which achieves effective processing of complex data and accurate prediction of system status through a series of steps.

[0035] In a preferred embodiment of the present invention, in step S1, the mathematical formula for calculating the dimensionless vector of the predicted system state at the next moment by using the dimensionality-reduced global interactive fusion feature matrix as the external driving input force is as follows: in, This represents the unified-dimensional feature vector of the m-th data source at time t after dimensionless transformation and spatial mapping; Let m be the orthogonal projection matrix corresponding to the m-th data source; where m is the orthogonal projection matrix. Let be a column orthogonal matrix, satisfying ,in, To unify the dimension of the feature subspace, For the first The original feature dimensions of each data source for An identity matrix of order 1.

[0036] In this embodiment, the mathematical formula defines the specific process of converting raw data into a feature vector of uniform dimension. Specifically, This represents the engineering feature vector with original physical dimensions input from the m-th data source at time t, such as sensor readings, equipment operating parameters, etc.

[0037] This represents the mathematical expectation vector of the m-th data source within the historical sliding window. It reflects the average level of that data source, and its physical dimensions are consistent with the original data.

[0038] Let represent the standard deviation vector of the m-th data source within the historical sliding window. It reflects the volatility of the data source, and its physical dimensions remain consistent with the original data. By subtracting the expectation and dividing by the standard deviation, the data is transformed into a dimensionless form with a mean of 0 and a standard deviation of 1. To avoid division-by-zero errors when the standard deviation is extremely small or zero, a dimensionless minimum positive constant is introduced. .

[0039] Orthogonal projection matrix This matrix is ​​used to map and project dimensionless eigenvectors onto a eigenspace of uniform dimension. It is a column orthogonal matrix with dimensions of . ,in It is the dimension of the unified feature subspace. It is the original feature dimension of the m-th data source. It satisfies... The properties of this projection ensure linear independence and information preservation in the projection process, facilitating the establishment of consistent feature representations across different data sources. Through this projection, even if the original data sources have different dimensions, they can be transformed into a common low- or high-dimensional feature space, facilitating subsequent fusion and analysis. This orthogonal projection matrix can be obtained through various methods, such as learning the main directions of data variation through principal component analysis (PCA), or training it through linear layers in a neural network to effectively map features from different data sources to a unified space.

[0040] In a preferred embodiment of the present invention, the mathematical formula for calculating the output global interactive fusion feature matrix in step S2 is as follows: in, This indicates that all [elements] are merged at time t. The dimensionless global interactive fusion feature matrix output after multiple heterogeneous data sources; This represents the total number of heterogeneous data sources; m and n represent the indices of the first and second heterogeneous data sources that participated in the interaction, respectively. and Let represent the dimensionless, uniform-dimensional feature vectors of the m-th and n-th data sources at time t, respectively; The expression represents the transpose of a matrix; exp represents an exponential function with the natural constant as its base. It represents the square of the Euclidean distance between two dimensionless eigenvectors.

[0041] In this embodiment, the global interactive fusion feature matrix This is the result of a comprehensive representation of information from all heterogeneous data sources at a specific time t. It not only includes the independent information from each data source, but more importantly, it captures the complex nonlinear correlations between data sources through high-order interactive covariance. This matrix serves as the external driving force for subsequent dynamic model predictions, and its quality directly affects the accuracy of system state predictions.

[0042] Summation term in the formula The goal is to compute the outer product between all possible pairs of heterogeneous data sources. (Outer product operation) It can expand the information of two feature vectors into a matrix, thereby capturing all the second-order interaction information between them, which reveals deeper connections between data sources than simple vector concatenation or dot product.

[0043] Exponential function term It acts as an adaptive weighting factor. This term is based on two dimensionless, uniform-dimensional feature vectors. and The square of the Euclidean distance between them The similarity is calculated using an exponential function. The closer two feature vectors are, the more similar or consistent their information is, resulting in a larger exponential function value and a greater weight assigned to their interaction terms. Conversely, the greater the distance, the greater the information difference or potential conflict, causing the exponential function value to approach zero, thus automatically suppressing the contribution of their interaction terms. This mechanism effectively filters out irrelevant or conflicting information, enhancing the robustness of the fusion process.

[0044] Total number of heterogeneous data sources The size of the data sources participating in the fusion is determined. m and n serve as indices to iterate through all data source pairs, ensuring that all possible pairwise interactions are considered. The matrix transpose operation T is a necessary step for performing the outer product operation to ensure dimension matching.

[0045] When the total number of heterogeneous data sources When the summation is large, the above double summation can be optimized by engineering: only when m and n satisfy a preset correlation threshold (e.g., ,in The calculation is performed for a data source combination with a preset dimensionless threshold; the exponential terms of other combinations are close to zero and can be ignored, thus reducing the computational complexity from... Reduce to ,in The number of nearest neighbor data sources to meet the threshold condition.

[0046] The solution in this application achieves dynamic interactive fusion of multi-source heterogeneous data through the aforementioned mathematical calculation formula. First, at time t, dimensionless, unified-dimensional feature vectors from different heterogeneous data sources... and The data is retrieved. Subsequently, the system iterates through all possible heterogeneous data source pairs (m and n).

[0047] For each pair of data sources, first calculate the outer product of their eigenvectors. This outer product forms a matrix that can meticulously characterize all second-order interactions between the two data sources, thus laying the foundation for higher-order interaction covariance. Simultaneously, the squared Euclidean distance between the two eigenvectors is calculated. This distance quantifies their similarity in the feature space. Then, this distance is passed through an exponential function. This is transformed into an adaptive weighting factor. Finally, the outer product of each data source pair is multiplied by its corresponding adaptive weighting factor, and all weighted interaction matrices are summed to obtain the global interaction fusion feature matrix. This mechanism ensures that the fusion process not only considers the direct interaction between data sources, but also dynamically adjusts the contribution of each interaction item according to their inherent similarity, thereby generating a comprehensive and robust system state representation, effectively solving the problem of how to accurately quantify and adaptively fuse heterogeneous data.

[0048] In a preferred embodiment of the present invention, the mathematical formula for calculating the dimensionless vector of the predicted system state at the next moment, after dimensionality reduction of the global interactive fusion feature matrix as the external driving input force, is as follows: in, This represents the dimensionless vector of the system state at time t+1; This represents the dimensionless vector of the current state of the system at time t. Represents the dimensionless time step of the discretization; Represents the dimensionless internal state transition matrix that characterizes the self-evolution law of the system; This represents the dimensionless external driving gain matrix; This represents the dimensionless global interactive fusion feature matrix at time t; This is a learnable attention vector used to differentially weight each dimension of the fused feature matrix; its values ​​are adaptively learned through training.

[0049] In this embodiment, the discretized nonlinear dynamics model is a mathematical framework used to describe the evolution of a system state over discrete time steps, characterized by the nonlinear relationship between the state and the input. This model aims to capture the inherent complex, non-proportional relationships and behaviors of real-world systems, which are difficult to fully express using linear models. It can be implemented through a series of nonlinear functions or structures, such as deep learning models using recurrent neural networks or long short-term memory networks, or by discretizing nonlinear differential equations.

[0050] The predictive output is a dimensionless vector of the system state at the next time step. Its purpose is to provide a quantitative, dimensionless representation of the system's expected state at the next discrete time step, based on the current state and fused input information. This vector is the output of the dynamic model and can represent key variables in the system, such as temperature, pressure, location, or health indicators, presented in a standardized form. The system's internal state transition self-evolution mechanism refers to the way the system state changes naturally over time, independent of external inputs and governed by the system's inherent properties and dynamic laws. Its purpose is to model the system's intrinsic time dependence and internal dynamics.

[0051] In the above formula, this is usually determined by The term is represented as follows, where A is a matrix used to transform the current state to reflect its natural evolution. The dimensionless internal state transition matrix A is a matrix that defines how the current system state vector contributes linearly to the internal evolution of the next state. Its role is to mathematically capture the inherent dynamics of the system and how the current state influences its future state without external intervention. This matrix can be randomly initialized and learned through an optimization process, or derived from a linearized approximation of the system's underlying physics.

[0052] The dimensionless external drive gain matrix B is a matrix that scales and transforms the processed external drive input, which then affects the system state. Its function is to adjust the strength and direction of the external input's influence on the system state; it can be identified through least-squares system analysis or determined based on prior physical knowledge (such as actuator gain, coupling coefficients, etc.).

[0053] This represents a dimensionless linear transformation matrix that reduces a matrix to a vector. This transformation can be achieved through principal component analysis (PCA), linear discriminant analysis (LDA), or autoencoder neural networks to fuse historical feature matrices. To acquire through learning.

[0054] Hyperbolic tangent activation function It is a non-linear activation function that compresses the input value to... Within a certain range. Its function is to introduce nonlinearity into the model, enabling it to capture the complex nonlinear relationship between external inputs and their influence on the system state, which is impossible for linear models. Simultaneously, it helps stabilize the model by limiting the output range of input terms. A dimensionless linear transformation matrix that reduces a matrix to a vector. It is a high-dimensional matrix (such as The matrix is ​​projected onto a low-dimensional vector space. Its purpose is to reduce the dimensionality of the fused feature matrix, making it compatible with the input requirements of the dynamic model and potentially focusing on the most salient features.

[0055] Learnable attention vectors This is a vector whose element weights are learned during training, used to assign different importance to different dimensions of the input. Its function is to enable the model to selectively focus on the most relevant dimensions in the transformed fused feature matrix, thereby achieving differentiated weighting and enhancing the model's ability to capture key information.

[0056] The proposed solution achieves accurate prediction of the system's future state by organically combining the system's internal self-evolutionary laws with external driving input forces. Specifically, it predicts the system's state at time t+1. It is determined by two parts: one part is the current state of the system at time t. The internal state transition matrix A embodies the system's inherent self-evolution mechanism; the other part is the external driving input force, which first transforms the dimensionless global interactive fusion feature matrix at time t. After dimensionless linear transformation matrix Dimensionality reduction is performed, among which, , The feature dimension after dimensionality reduction, and And by learnable attention vectors Differential weighting is applied to each dimension to highlight key information. Then, a nonlinearity is introduced through the hyperbolic tangent activation function tanh, enabling the model to capture the complex nonlinear relationship between external inputs and system states. Finally, the model is modulated by a dimensionless external driving gain matrix B. These two contributions are reflected in the discretized dimensionless time step. The system state is updated jointly under the influence of various factors. This combined approach allows the model to not only follow the system's own evolutionary laws but also fully utilize the rich information obtained from the fusion of multi-source heterogeneous data and respond to changes in the external environment in a nonlinear manner, thereby achieving more accurate and robust state prediction in complex dynamic systems.

[0057] In a preferred embodiment of the present invention, in step S4, the mathematical formula for calculating the joint objective function is: in, This represents the dimensionless joint objective function value used in model training. This represents the total number of batch samples used to calculate the objective function; Indicates the first in the current batch One data sample index; Indicates the first The dimensionless vector of the predicted system state of a sample at time t+1; Indicates the first The dimensionless vector of the actual observed system state corresponding to each sample at time t+1; This represents the mean square error between the predicted state and the actual observed state. This represents the dimensionless regularization weight coefficient used to balance prediction error and the complexity of fused features; Represents the trace of a matrix; Indicates the first The dimensionless global interactive fusion feature matrix of each sample at time t.

[0058] In this embodiment, the joint objective function It is a core metric used during model training to measure model performance and guide parameter optimization. It typically consists of multiple terms designed to simultaneously optimize different aspects of the model, such as prediction accuracy and model complexity. This function quantifies the "goodness" of a model by calculating the difference between predicted and true values, as well as penalties imposed on certain internal properties of the model; its goal is to minimize this function value.

[0059] Total number of batch samples This refers to the number of samples used to calculate the objective function in a single iteration of model parameter updates. In machine learning, mini-batch training is typically used, meaning that each iteration processes... Instead of all training samples, the current batch uses a specific sample (the 1st) to improve training efficiency and stability. Data Sample Index Used to identify a specific sample being processed within a batch.

[0060] Predict the dimensionless vector of system state This is the model's prediction of the system state at time t+1 based on the input data and current parameters. It is a dimensionless vector reflecting the expected state of the system at future times. The actual observed system state is a dimensionless vector. These are the actual observations corresponding to the predicted system state, usually acquired by external sensors or measuring devices at time t+1 and processed to be dimensionless.

[0061] Mean square error between predicted state and actual observed state Mean squared error is a commonly used metric for measuring the accuracy of model predictions. It calculates the sum of squares of the differences between predicted and actual values, effectively reflecting the degree of prediction deviation. In addition to mean squared error, other error measures can be used, such as mean absolute error (L1 norm) or Huber loss, to accommodate different error distribution characteristics.

[0062] Dimensionless regularization weight coefficients It is a hyperparameter used to balance the importance of different terms in the joint objective function. It determines the trade-off between pursuing prediction accuracy and controlling model complexity, and is usually tuned using methods such as cross-validation.

[0063] trace of a matrix This refers to the sum of the elements on the diagonal of a square matrix. In this application, Represents the dimensionless global interactive fusion feature matrix The square of the Frobenius norm is used as a regularization term to penalize excessively large values ​​or complexities in the feature matrix, thereby preventing the model from overfitting.

[0064] Dimensionless global interactive fusion feature matrix It is a feature representation of the system's comprehensive information at time t, obtained after processing by a multi-source heterogeneous data fusion module. It captures the dynamic interaction and nonlinear correlation between different data sources.

[0065] The solution proposed in this application constructs a joint objective function that includes a prediction error term and a regularization term. This achieves closed-loop optimization of parameters for multi-source heterogeneous data fusion and intelligent analysis models. Specifically, the prediction error term... To ensure the model can accurately predict the system state at the next moment, the model parameters are adjusted to reduce prediction bias. Simultaneously, the regularization term... Then the global interactive fusion feature matrix By constraining the complexity of features and penalizing excessively large feature values ​​or overly complex feature representations, overfitting during model training can be effectively prevented.

[0066] Dimensionless regularization weight coefficients This approach provides an adjustable balance between the two, allowing the model to maintain good generalization ability and stability while pursuing high prediction accuracy. During optimization, the model iteratively adjusts its internal parameters to minimize this joint objective function. The optimization process stops when the joint objective function reaches a preset convergence condition. At this point, the model parameters minimize the error between the predicted system state and the actual observed system state, and the complexity of the fused feature matrix is ​​effectively controlled. This optimization mechanism enables the model to not only learn effective prediction patterns from multi-source heterogeneous data but also ensures that the learned feature representations are robust and interpretable, thus solving the problems of overfitting and insufficient generalization ability that may occur in complex data environments.

[0067] In a preferred embodiment of the present invention, in step S4, a comprehensive external environmental influence factor is introduced, and a joint objective function with adaptive penalty weights is constructed; the specific mathematical calculation formula is as follows:

[0068] Calculate the dimensionless comprehensive impact factor: Joint objective function value with adaptive penalty weights: ;

[0069] in, This represents the dimensionless comprehensive influence factor calculated by aggregating the k-th sample at time t, used to characterize the severity of the comprehensive disturbance in the external environment; Indicates the first Dimensionless attention weights for various external influencing factors; This indicates that the k-th sample was collected at time t. The original physical quantity observations of various external influencing factors; and They represent the first Expected value and standard deviation of observed values ​​of external influencing factors; To map the standardized external influencing factors to Activation function for the interval.

[0070] In this embodiment, the dimensionless comprehensive influence factor This is used to quantify the comprehensive disturbance that the external environment may cause to the prediction of system state at a specific time and for a specific sample. Its calculation method involves weighted aggregation of multiple external influencing factors followed by a nonlinear transformation. For example, in addition to mapping the standardized external influencing factors using the Sigmoid function, the ReLU function or Tanh function can be used for nonlinear mapping to adapt to different disturbance characteristics; or, in some scenarios, a piecewise linear function determined based on expert experience or historical data analysis can be used to characterize the degree of influence.

[0071] This represents the total number of types of external influencing factors. Indicates the first External influencing factors include temperature, humidity, pressure, vibration, light intensity, wind speed, load changes, etc. These factors can be physical quantities, chemical quantities, or operating conditions.

[0072] Dimensionless attention weights This is used to measure the relative importance of different external influencing factors on system disturbances. For example, in some industrial scenarios, temperature changes may have a more significant impact on the system state than humidity changes; in this case, the attention weight corresponding to temperature can be set higher. These weights can be preset, set based on domain knowledge, or adaptively optimized as learnable parameters during model training. The learnable parameters mainly include the orthogonal projection matrices of each data source. and dynamic models , , , The prediction bias is updated using gradient descent, thereby indirectly optimizing the feature representation in the standardization and fusion stages.

[0073] raw physical quantity observation values The first sample collected at a specific time. The original physical quantity observations of various external influencing factors retain their original physical dimensions.

[0074] No. Expected values ​​of observations of external influencing factors With the Standard deviation of observed values ​​of external influencing factors This is used to standardize the observed values ​​of raw external influencing factors, eliminating differences in dimensions and numerical ranges between different factors and making them comparable. For example, these values ​​can be obtained through historical data statistics, or dynamically updated using a sliding window mean and standard deviation in real-time operation.

[0075] The Sigmoid activation function maps standardized external influencing factors to a dimensionless value between 0 and 1, allowing the degree of influence of different factors to be uniformly measured and exhibiting a smooth, non-linear trend. Besides the Sigmoid function, the Softmax function can be used to normalize multiple influencing factors, or Gaussian radial basis functions can be used to characterize the degree of influence.

[0076] The exponential function exp nonlinearly amplifies the prediction mean square error term in the joint objective function. As the dimensionless comprehensive influence factor increases, the exponential function can increase the weight of the error term more rapidly, thus significantly improving the model's sensitivity to severe external disturbances. This nonlinear amplification mechanism allows the model to respond more sensitively to prediction biases under adverse conditions during optimization.

[0077] Joint objective function with adaptive penalty weights Based on the existing prediction error and fusion feature complexity regularization terms, an adaptive penalty weight driven by a dimensionless comprehensive influence factor is introduced. This weight can dynamically adjust the importance of the prediction error term according to the degree of external environmental disturbance, ensuring that the model can prioritize learning and adapt to critical operating conditions with severe external disturbances during training.

[0078] Based on the aforementioned multi-source heterogeneous data fusion and intelligent analysis methods, this application introduces a comprehensive external environment influence factor and constructs a joint objective function with adaptive penalty weights to further improve the robustness and prediction accuracy of the model under complex and ever-changing external environments, especially when external disturbances are severe.

[0079] Specifically, before optimizing the closed-loop parameters, the system first acquires raw physical quantity observations of various external influencing factors for each sample at the current moment. These observations are standardized and mapped to a unified dimensionless interval using a sigmoid activation function. Combined with preset or learned attention weights, a dimensionless comprehensive influence factor is calculated. This factor can quantitatively characterize the severity of the comprehensive disturbance in the current external environment.

[0080] Subsequently, when constructing the joint objective function for model optimization, this dimensionless comprehensive influence factor is cleverly applied to the prediction mean squared error term. Specifically, this influence factor is converted into a non-linear penalty weight through an exponential function and multiplied onto the prediction mean squared error term. This means that the more severe the external environmental disturbance, i.e., the larger the dimensionless comprehensive influence factor, the more exponentially the weight of the prediction mean squared error term in the joint objective function will be amplified. This mechanism allows the model to pay more attention to and prioritize reducing the prediction error under conditions of severe external disturbances during gradient descent optimization.

[0081] In this way, the model can adaptively adjust its learning strategy, forcing parameters such as the internal state transition matrix and the external driving input gain matrix to better adapt to and handle the uncertainty and complexity brought about by drastic changes in the external environment during the optimization process, thereby significantly enhancing the model's generalization ability and prediction accuracy under various complex conditions.

[0082] The following is an illustration using a specific example. As a concrete implementation, consider a method for predicting the operating status of large industrial equipment (e.g., wind turbine generators). During closed-loop parameter optimization in step S4, in addition to the usual mean square error between the predicted and observed states and the regularization term of the fused feature matrix, this application introduces a comprehensive external environmental influence factor. For example, wind speed, ambient temperature, humidity, atmospheric pressure, and power grid fluctuation amplitude can be selected as external influencing factors. For each sample, the raw observation values ​​of these factors are first collected; for example, the current wind speed is 15 m / s and the ambient temperature is -5°C. These raw values ​​are standardized based on their historical statistical data (expectation and standard deviation) and then mapped to a range of 0 to 1 using the Sigmoid function. Next, based on the experience or preset weights of these factors' influence on the wind turbine generator's operating status (e.g., the weight of wind speed may be higher than that of humidity), they are weighted and summed to obtain a dimensionless comprehensive influence factor. It is assumed that the calculated comprehensive influence factor is higher under strong wind and low temperature conditions. At this point, when constructing the joint objective function, this higher overall influence factor will non-linearly amplify the mean squared error term of the current sample's prediction through an exponential function. This means that if the model makes a prediction error under severe conditions such as strong winds and low temperatures, the penalty it incurs in the objective function will be much greater than the same prediction error under mild weather conditions. Through this mechanism, the model will be more inclined to optimize its parameters during training to ensure that it maintains high prediction accuracy even under highly disturbed environments such as strong winds and low temperatures, thereby improving the robustness and reliability of the entire prediction system under extreme conditions.

[0083] like Figure 4 As shown, the multi-source heterogeneous data fusion and intelligent analysis system, applied to the above-mentioned methods, includes a dimensionless standardization module, a cross-modal dynamic fusion module, a dynamic prediction module, and a closed-loop optimization module.

[0084] The dimensionless standardization module performs dimensionless standardization on multi-source heterogeneous raw data and maps it to a unified-dimensional feature subspace, outputting a dimensionless unified-dimensional feature vector. This module also includes a multi-source data time synchronization submodule, used for time alignment and resampling of raw data with different sampling frequencies. Specifically, the dimensionless standardization process eliminates the influence of the original physical dimensions, making data with different dimensions, such as vibration signals, temperature readings, and pressure data, numerically comparable. The time synchronization submodule aligns the time axis of high-frequency sampled vibration data with low-frequency sampled temperature data, ensuring the temporal consistency of multi-source data.

[0085] The cross-modal dynamic fusion module calculates adaptive manifold distance based on the unified dimensional feature vector, constructs a high-order interaction covariance for dynamic fusion, and outputs a global interaction fusion feature matrix. The adaptive manifold distance, based on the Euclidean distance between feature vectors from two heterogeneous data sources in the feature space, is calculated adaptively using an exponential function to obtain a similarity measure; the closer the distance, the higher the similarity and the greater the fusion weight; the farther the distance, the closer the similarity approaches zero, automatically suppressing the interaction of conflicting information. The high-order interaction covariance is obtained by performing an outer product operation on the dimensionless feature vectors of any two data sources from M heterogeneous data sources and then weighted summing all M² combinations. It is used to capture nonlinear correlations between data sources. For example, in industrial equipment monitoring, this module can quantify the dynamic correlation strength between vibration feature vectors and temperature feature vectors, and explicitly model the high-order cross-coupling effect between vibration modes and temperature anomalies through outer product operations.

[0086] The dynamics prediction module uses the dimensionality-reduced global interactive fusion feature matrix as an external driving force, which is then substituted into a discretized nonlinear dynamics model. Combined with a state transition self-evolution mechanism, this predicts the system state vector at the next moment. This discretized nonlinear dynamics model characterizes the nonlinear evolution of the system state over time, the external driving force reflects the dynamic modulation effect of the fused features on the system state, and the state transition self-evolution mechanism describes the system's intrinsic evolution process in the absence of external input. For example, embedding the dimensionality-reduced fusion feature vector as the driving force into the model enables the prediction of device health status to not only depend on historical states but also dynamically respond to real-time feature changes after multi-source data fusion.

[0087] The closed-loop optimization module obtains the state vector of the real observed system and constructs a joint objective function by combining the adaptive constraints of the predicted state and the fused feature matrix to optimize the closed-loop parameters of the dynamic model. This closed-loop parameter optimization employs gradient descent, with the learning rate decaying exponentially with the number of iterations. Optimization stops when the decrease in the objective function is less than a preset threshold for multiple consecutive iterations. The joint objective function comprehensively considers both prediction error and the complexity of the fused features. Through backpropagation of the deviation between the real observed state and the predicted state, the dynamic model parameters are dynamically adjusted to ensure coordinated optimization of the fusion strategy and the prediction task.

[0088] The core innovation of this embodiment lies in its use of a cross-modal dynamic fusion module to calculate adaptive manifold distance and construct a high-order interactive covariance, effectively addressing the problem of insufficient modeling of nonlinear high-order interactive relationships between heterogeneous data and capturing high-order cross-coupling effects between multi-source data. By embedding the global interactive fusion feature matrix as an external driving force into a discretized nonlinear dynamic model, a causal relationship between fusion features and system dynamic evolution is established, ensuring that the prediction results follow the inherent laws of system change. Furthermore, a closed-loop optimization module is used to achieve closed-loop parameter optimization based on prediction bias, overcoming the limitations of open-loop independent optimization strategies and enabling the fusion process to directly serve the prediction task. Compared to early fusion methods in the prior art that rely solely on feature concatenation or weighted averaging, this application explicitly models the implicit physical relationships between multi-source data through high-order interactive covariance, avoiding the neglect of high-order interactive effects by linear models. Simultaneously, embedding fusion features as an external driving force into the dynamic model, rather than static mapping, ensures the physical consistency between the prediction results and the system's dynamic evolution. In addition, the closed-loop parameter optimization mechanism allows prediction errors to correct the fusion strategy in reverse, avoiding the local compromise problem of independent optimality of each module in open-loop design.

[0089] 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, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for multi-source heterogeneous data fusion and intelligent analysis, characterized in that, Includes the following steps: S1. Obtain multi-source heterogeneous raw data, perform dimensionless standardization on the multi-source heterogeneous raw data containing different physical dimensions and sampling frequencies, and map and project it onto a feature subspace of a unified dimension to obtain a dimensionless unified dimension feature vector. S2. Based on the unified dimension feature vector, calculate the adaptive manifold distance between different heterogeneous data sources, and then construct a cross-modal high-order interactive covariance to realize dynamic interactive fusion between multi-source information and output a global interactive fusion feature matrix. S3. After reducing the dimensionality of the global interactive fusion feature matrix, use it as the external driving input force and substitute it into the discretized nonlinear dynamic model. Combined with the internal state transition self-evolution mechanism of the system, predict and output the dimensionless vector of the predicted system state at the next moment. S4. Obtain the dimensionless vector of the real observation system state provided by the external observation platform, and combine the dimensionless vector of the predicted system state with the adaptive constraint of the global interactive fusion feature matrix to construct a joint objective function to optimize the closed-loop parameters of the discretized nonlinear dynamic model. The closed-loop parameter optimization adopts the gradient descent method, and the learning rate decays exponentially with the number of iterations. The optimization stops when the decrease of the objective function in several consecutive iterations is less than a preset threshold.

2. The method according to claim 1, characterized in that, In step S1, the difference vector is obtained by subtracting the mathematical expectation vector of the data source within the historical sliding window from the engineering feature vector with the original physical dimensions input at time t. Each component of the standard deviation vector of the data source within the historical sliding window is added with a minimal positive constant introduced to prevent the denominator from being zero, resulting in a sum vector. The difference vector is then divided by the sum to obtain the standardized vector. Finally, the standardized vector is multiplied on the left by the orthogonal projection matrix corresponding to the data source to obtain the dimensionless uniform dimension feature vector of the data source at time t.

3. The method according to claim 2, characterized in that, In step S2, the specific method for outputting the global interactive fusion feature matrix is ​​as follows: For all heterogeneous data sources, a double summation is performed in pairs. For each pair of data sources, first, the dimensionless uniform dimension eigenvector of the first data source is multiplied by the transpose of the dimensionless uniform dimension eigenvector of the second data source to obtain the outer product matrix. Then, the square of the Euclidean distance between the eigenvectors of the pair of data sources is calculated and the negative half is taken as the exponent. The exponent function value is calculated with the natural constant e as the base to obtain the distance weight. Multiply the outer product matrix by the distance weight to obtain the interaction contribution matrix of the data source pair; sum the interaction contribution matrices of all data source pairs to obtain the dimensionless global interaction fusion feature matrix at time t.

4. The method according to claim 3, characterized in that, The specific method for using the dimensionality-reduced global interactive fusion feature matrix as the external driving input force and predicting the dimensionless vector of the system state at the next moment is as follows: First, calculate the dimensionless vector of the system state at the current moment and multiply it by the dimensionless internal state transition matrix that characterizes the self-evolution law of the system to obtain the self-evolution contribution term. Then, calculate the global interactive fusion feature matrix and multiply it by the learnable attention vector to obtain the weighted fusion vector. Use a dimensionless linear transformation matrix to reduce the dimension of the weighted fusion vector to vector form. Then, take the hyperbolic tangent function value of each element of the vector to obtain the nonlinear driving term. Multiply the nonlinear driving term by the dimensionless external control gain matrix to obtain the external driving contribution term. Add the self-evolution contribution term and the external driving contribution term, multiply it by the discretized dimensionless time step, and finally add the dimensionless vector of the system state at the current moment to obtain the dimensionless vector of the predicted system state at the next moment.

5. The method according to claim 4, characterized in that, In step S4, the joint objective function is calculated as follows: For each sample in the batch, first calculate the squared Euclidean distance between the dimensionless vector of the predicted system state and the dimensionless vector of the actual observed system state, as the prediction error term; then calculate the transpose of the global interactive fusion feature matrix of the sample, multiply it by the global interactive fusion feature matrix, take the trace of the product matrix, and multiply it by the dimensionless regularization weight coefficient, as the regularization term; add the prediction error term and the regularization term of the sample to obtain the single-sample loss of the sample; add the single-sample losses of all samples and divide by the total number of samples in the batch to obtain the joint objective function value.

6. The method according to claim 4, characterized in that, In step S4, a comprehensive external environmental impact factor is introduced, and a joint objective function with adaptive penalty weights is constructed; the specific method is as follows: First, calculate the dimensionless comprehensive impact factor: For each external influencing factor, first calculate the original physical quantity observation value of the influencing factor in the current sample, subtract the expected value of the influencing factor, and then divide by the standard deviation of the influencing factor to obtain the standardized value; take the Sigmoid activation function value of the standardized value, and then multiply it by the attention weight of the influencing factor; sum up the weighted results of all influencing factors to obtain the comprehensive impact factor of the sample. Next, calculate the joint objective function value with adaptive penalty weights: For each sample in the batch, first calculate the exponential function value with the natural constant e as the base and the comprehensive influence factor of the sample as the exponent. Multiply the exponential function value by the prediction error term of the sample to obtain the weighted prediction error term. Then calculate the regularization term of the sample. Add the weighted prediction error term and the regularization term to obtain the weighted single-sample loss of the sample. Sum the weighted single-sample losses of all samples and divide by the total number of samples in the batch to obtain the joint objective function value with adaptive penalty weights.

7. A multi-source heterogeneous data fusion and intelligent analysis system, applied to the method described in claims 1-6, characterized in that, include: The dimensionless standardization module is used to perform dimensionless standardization processing on multi-source heterogeneous raw data and map it to a unified dimension feature subspace, outputting a dimensionless unified dimension feature vector; the dimensionless standardization module also includes a multi-source data time synchronization submodule, which is used to perform time alignment and resampling on raw data with different sampling frequencies. The cross-modal dynamic fusion module is used to calculate the adaptive manifold distance based on the unified dimension feature vector, construct a high-order interactive covariance for dynamic fusion, and output a global interactive fusion feature matrix. The dynamics prediction module is used to reduce the dimensionality of the global interactive fusion feature matrix and use it as an external driving input force, which is then substituted into the discretized nonlinear dynamics model and combined with the state transition self-evolution mechanism to predict the system state vector at the next moment. The closed-loop optimization module is used to obtain the state vector of the real observed system, and construct a joint objective function by combining the predicted state and the adaptive constraints of the fused feature matrix to optimize the closed-loop parameters of the dynamic model.

Images