A deep learning-based multi-source heterogeneous time series data evolution prediction method and system

Abstract

The application relates to the technical field of machine learning, and discloses a multi-source heterogeneous time series data evolution prediction method and system based on deep learning, which comprises the following steps: acquiring multi-source signal data containing high-frequency signals, low-frequency signals and discrete instructions; projecting the signal data to a latent space continuous time axis and calculating an evolution rate norm; correcting high-frequency feature weights by using a weight matrix generated by a super network and combining the discrete instructions to extract a heterogeneous feature tensor; representing latent space curvature by using a Jacobian spectral radius; adaptively switching between implicit and explicit operators according to the latent space curvature; and outputting a prediction result. The application suppresses cross-scale sampling feature flooding, guarantees the causality integrity under heterogeneous time series, and improves the integral stability of a neural differential equation when solving a rigid equation.

Description

Technical Field

[0001] This invention relates to a method and system for predicting the evolution of multi-source heterogeneous time-series data based on deep learning, belonging to the field of machine learning technology. Background Technology

[0002] Currently, using deep learning architectures in the continuous time domain to process non-uniformly sampled time-series data has become a common technical approach. By constructing the time derivative function of the hidden state in the latent space, the topological distortion caused by the resampling of heterogeneous data by traditional discrete-time grids is avoided, thereby maintaining the continuous evolution characteristics of the original physical signal in the time domain. In actual industrial scenarios, the stable operation of the system is limited by the non-ideal environmental constraints at the level of sensor physical deployment. There are also technical bottlenecks at the level of subsequent data evolution models and control algorithms. For example, Chinese invention patent application with publication number CN119861664A discloses an abnormal time-series data detection method and device based on a multi-view spatiotemporal interactive attention network. It models the correlation of industrial data through multi-view feature extraction and spatiotemporal correlation learning. Such technology strongly relies on discrete-time window feature reconstruction and pattern recognition, and is essentially in the category of static feature fitting. When facing highly dynamic industrial equipment such as aero-engines, the data to be predicted contains multi-source modes with huge differences in sampling frequency. High-frequency transient signals and low-frequency background features are deeply coupled in the latent space, causing drastic changes in the local curvature of the feature flow field, exhibiting numerical rigidity characteristics.

[0003] However, when the data to be predicted contains multi-source modes with huge differences in sampling frequency, the deep coupling between high-frequency transient signals and low-frequency background features in the latent space can cause drastic changes in the local curvature of the feature flow field, exhibiting obvious numerical rigidity. In order to maintain the absolute stability of the computational trajectory in the rigid flow field, conventional inference operators adopt a small time integration step size, which makes the depth of the computational graph in the forward propagation process show a nonlinear growth trend. When faced with the surge in computing power caused by the above-mentioned evolutionary rigidity, the industry has tried to alleviate numerical fluctuations by adopting gradient pruning or increasing model depth, but such improvement paths have not addressed the underlying contradiction between the latent space flow field and the numerical stability domain. Simple gradient constraints often lead to the loss of high-frequency feature information, while increasing the depth will further worsen the forward inference latency while improving the model's expressive power, making it difficult for continuous-time evolution models to achieve a balance between computational efficiency and physical realism when processing complex industrial data.

[0004] Therefore, the technical problem to be solved by this invention is to utilize the existing computing resources in the system to establish a response mechanism within the neural network architecture that can dynamically sense the stiffness of the flow field and autonomously switch computing operators in order to achieve stable evolution prediction of multi-source heterogeneous time series data. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of this invention is as follows: A method for predicting the evolution of multi-source heterogeneous time-series data based on deep learning, comprising the following steps:

[0006] Step 101: Obtain multi-source time-series signal data characterizing the physical state of the target object. The multi-source time-series signal data includes high-frequency dynamic features, low-frequency environmental features, and discretely triggered non-periodic commands.

[0007] Step 102: Project the multi-source time-series signal data onto the continuous time axis of the latent space, calculate the derivative magnitude of each modal feature with respect to time at a given moment, and obtain the evolution rate norm.

[0008] Step 103: Input the evolution rate norm into the super network. The super network generates a dynamic weight matrix corresponding to the high-frequency feature tensor based on the sampling frequency difference. The dynamic weight matrix is ​​used to correct the scoring weight of the high-frequency feature tensor in the attention mechanism. Combined with non-periodic instructions, the decoupled multi-source heterogeneous feature tensor is extracted.

[0009] Step 104: Input the multi-source heterogeneous feature tensor into the neural differential evolution network, and calculate the Jacobian matrix spectral radius by taking the partial derivative of the hidden state evolution function of the neural differential evolution network to characterize the curvature of the latent space.

[0010] Step 105: Perform adaptive routing of the prediction operator based on the latent space curvature. When the latent space curvature exceeds the preset curvature threshold, perform implicit numerical integration based on the latent space fixed point iteration and output the evolution prediction result of the target object.

[0011] Preferably, in step 105, the adaptive routing of the prediction operator satisfies the following determination rule: , where S is the selected prediction operator path; ρ(J) is the spectral radius of the Jacobian matrix; γ is the preset curvature threshold, which is a value preset based on the mean convergence curvature of the neural differential evolution network during the training phase.

[0012] Preferably, step 102 includes the following sub-steps: Step 1021, mapping discrete non-periodic instructions, low-frequency environmental features, and high-frequency dynamic features in multi-source time-series signal data to the latent space using a continuous-time neural operator; Step 1022, performing derivative operations based on automatic differentiation on the continuous hidden states in the latent space to obtain the time. Step 1023: Calculate the 2-norm of the instantaneous evolution gradient tensor to obtain the evolution rate norm.

[0013] Preferably, in step 103, the supernetwork transforms the feature vectors of the low-frequency modes into dynamic weight matrices, and uses the dynamic weight matrices to perform parameter reconstruction on the transformation operator of the high-frequency feature tensor to establish asymmetric constraint relationships between modes.

[0014] Preferably, in step 104, the neural differential evolution network uses a neural network to fit the time derivative of the state variable and performs continuous time-domain integration based on ordinary differential equations in the latent space.

[0015] Preferably, in step 105, when performing implicit numerical integration, an adaptive damping term is introduced into the neural differential evolution network to suppress the increase of numerical entropy during the calculation process by limiting the single-step change of the hidden state tensor in the latent space.

[0016] Preferably, in step 101, the sampling frequency range of the multi-source timing signal data covers 1Hz to 10kHz, and includes non-periodic triggered pulse logic signals.

[0017] Preferably, in step 103, an asymmetric computation graph is constructed using the difference in the magnitude of the evolution rate norm. By isolating the backpropagation gradient flow of different frequency modes, the submergence of low-frequency mode features caused by the dominance of high-frequency mode gradients is eliminated.

[0018] Preferably, the method further includes step 106: after obtaining the evolution prediction results, calculate the residual between the prediction results and the actual observations, and use the residual to synchronously correct the initial value of the hidden state of the neural differential evolution network. The method is used to realize the continuous time-domain feature extraction and trend prediction of industrial equipment operation status data.

[0019] A deep learning-based system for predicting the evolution of multi-source heterogeneous time-series data includes:

[0020] The data acquisition module is used to acquire multi-source time-series signal data that characterizes the physical state of the target object;

[0021] The evolution rate calculation module is used to project multi-source time-series signal data onto a continuous time axis in the latent space and calculate the modulus of the derivative of each modal feature with respect to time at a given moment to obtain the evolution rate norm.

[0022] The feature decoupling and fusion module has its input end connected to the output end of the evolution rate calculation module. It is used to input the evolution rate norm into the super network, use the dynamic weight matrix generated by the super network to correct the scoring weight of the high-frequency feature tensor in the attention mechanism, and combine non-periodic instructions to extract the decoupled multi-source heterogeneous feature tensor.

[0023] The manifold curvature calculation module is used to input multi-source heterogeneous feature tensors into the neural differential evolution network, and characterize the latent space curvature by calculating the Jacobian matrix spectral radius of the latent state evolution function of the neural differential evolution network.

[0024] The prediction execution module is used to perform adaptive route switching of prediction operators based on the latent space curvature, and when the latent space curvature exceeds the preset curvature threshold, it performs implicit numerical integration through fixed point iteration to output the evolution prediction result of the target object.

[0025] Compared with the prior art, the beneficial effects of the present invention are:

[0026] 1. In the evolution prediction of multi-source heterogeneous time series data, this invention utilizes the dynamic representation of the curvature of the latent space manifold by the spectral radius of the Jacobian matrix. It establishes an adaptive implicit and explicit operator routing mechanism. When the system encounters rigid features caused by high-frequency transient inputs, the operator routing layer switches to the implicit solution subgraph and completes fixed-point iteration in the latent space, ensuring that the tensor integral during the forward propagation process is always in the absolute stable domain. This mechanism eliminates the instability of numerical computation from the bottom layer of the algorithm, enabling continuous-time neural networks to maintain reliable inference logic when dealing with cross-scale time series data, and realizes the dynamic optimization allocation of computing resources in the time domain.

[0027] 2. Based on the asymmetric computation graph topology of the evolution rate norm, this invention achieves asymmetric decoupling and fusion of multi-source heterogeneous features. By converting low-frequency environmental modes into parameter generators for high-frequency transient modes, the method blocks the dot product competition between fast and slow variables within the same activation function, solves the gradient flooding phenomenon commonly found in deep learning models, and ensures that when the model processes heterogeneous data containing high-frequency vibrations and low-frequency temperatures, it can restore the constraints of the macroscopic environment on microscopic mutations, thereby enhancing the feature representation layer's ability to analyze the causal logic of the system. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating the overall process of decoupling and evolution prediction of multi-source heterogeneous time-series data features in this invention.

[0029] Figure 2 This is a logic diagram of operator adaptive routing and hidden state correction for latent space curvature in this invention.

[0030] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0031] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0032] A deep learning-based method for predicting the evolution of multi-source heterogeneous time-series data includes the following steps:

[0033] Step 101: Obtain multi-source time-series signal data characterizing the physical state of the target object. The multi-source time-series signal data includes high-frequency dynamic features, low-frequency environmental features, and discretely triggered non-periodic commands.

[0034] Step 102: Project the multi-source time-series signal data onto the continuous time axis of the latent space, calculate the derivative magnitude of each modal feature with respect to time at a given moment, and obtain the evolution rate norm.

[0035] Step 103: Input the evolution rate norm into the super network. The super network generates a dynamic weight matrix corresponding to the high-frequency feature tensor based on the sampling frequency difference. The dynamic weight matrix is ​​used to correct the scoring weight of the high-frequency feature tensor in the attention mechanism. Combined with non-periodic instructions, the decoupled multi-source heterogeneous feature tensor is extracted.

[0036] Step 104: Input the multi-source heterogeneous feature tensor into the neural differential evolution network, and calculate the Jacobian matrix spectral radius by taking the partial derivative of the hidden state evolution function of the neural differential evolution network to characterize the curvature of the latent space.

[0037] Step 105: Perform adaptive routing of the prediction operator based on the latent space curvature. When the latent space curvature exceeds the preset curvature threshold, perform implicit numerical integration based on the latent space fixed point iteration and output the evolution prediction result of the target object.

[0038] Preferably, in step 105, the adaptive routing of the prediction operator satisfies the following determination rule: , where S is the selected prediction operator path; ρ(J) is the spectral radius of the Jacobian matrix; γ is the preset curvature threshold, which is a value preset based on the mean convergence curvature of the neural differential evolution network during the training phase.

[0039] Preferably, step 102 includes the following sub-steps: Step 1021, mapping discrete non-periodic instructions, low-frequency environmental features, and high-frequency dynamic features in multi-source time-series signal data to the latent space using a continuous-time neural operator; Step 1022, performing derivative operations based on automatic differentiation on the continuous hidden states in the latent space to obtain the time. Step 1023: Calculate the 2-norm of the instantaneous evolution gradient tensor to obtain the evolution rate norm.

[0040] Preferably, in step 103, the supernetwork transforms the feature vectors of the low-frequency modes into dynamic weight matrices, and uses the dynamic weight matrices to perform parameter reconstruction on the transformation operator of the high-frequency feature tensor to establish asymmetric constraint relationships between modes.

[0041] Preferably, in step 104, the neural differential evolution network uses a neural network to fit the time derivative of the state variable and performs continuous time-domain integration based on ordinary differential equations in the latent space.

[0042] Preferably, in step 105, when performing implicit numerical integration, an adaptive damping term is introduced into the neural differential evolution network to suppress the increase of numerical entropy during the calculation process by limiting the single-step change of the hidden state tensor in the latent space.

[0043] Preferably, in step 101, the sampling frequency range of the multi-source timing signal data covers 1Hz to 10kHz, and includes non-periodic triggered pulse logic signals.

[0044] Preferably, in step 103, an asymmetric computation graph is constructed using the difference in the magnitude of the evolution rate norm. By isolating the backpropagation gradient flow of different frequency modes, the submergence of low-frequency mode features caused by the dominance of high-frequency mode gradients is eliminated.

[0045] Preferably, the method further includes step 106: after obtaining the evolution prediction results, calculate the residual between the prediction results and the actual observations, and use the residual to synchronously correct the initial value of the hidden state of the neural differential evolution network. The method is used to realize the continuous time-domain feature extraction and trend prediction of industrial equipment operation status data.

[0046] A deep learning-based system for predicting the evolution of multi-source heterogeneous time-series data includes:

[0047] The data acquisition module is used to acquire multi-source time-series signal data that characterizes the physical state of the target object;

[0048] The evolution rate calculation module is used to project multi-source time-series signal data onto a continuous time axis in the latent space and calculate the modulus of the derivative of each modal feature with respect to time at a given moment to obtain the evolution rate norm.

[0049] The feature decoupling and fusion module has its input end connected to the output end of the evolution rate calculation module. It is used to input the evolution rate norm into the super network, use the dynamic weight matrix generated by the super network to correct the scoring weight of the high-frequency feature tensor in the attention mechanism, and combine non-periodic instructions to extract the decoupled multi-source heterogeneous feature tensor.

[0050] The manifold curvature calculation module is used to input multi-source heterogeneous feature tensors into the neural differential evolution network, and characterize the latent space curvature by calculating the Jacobian matrix spectral radius of the latent state evolution function of the neural differential evolution network.

[0051] The prediction execution module is used to perform adaptive route switching of prediction operators based on the latent space curvature, and when the latent space curvature exceeds the preset curvature threshold, it performs implicit numerical integration through fixed point iteration to output the evolution prediction result of the target object.

[0052] Example 1: In the scenario of long-period cross-domain state evolution prediction for aero-engines, there exists a cross-scale coupling phenomenon between high-frequency vibration abrupt changes and low-frequency temperature rise signals. The data acquisition module acquires multi-source time-series signal data characterizing the physical state of the target object. This multi-source time-series signal data includes high-frequency dynamic features, low-frequency environmental features, and discretely triggered non-periodic commands. The sampling frequency range of this multi-source time-series signal data spans from 1Hz to 10kHz. When facing heterogeneous data sources with drastic differences in sampling rates, traditional tensor calculations require mandatory time grid alignment and data resampling. This transformation operation inevitably destroys the original physical topology of the high-frequency dynamic features and introduces discretization interpolation noise, causing local time step information density imbalance when the deep learning model extracts cross-modal joint features. To eliminate the topological break caused by gradient explosion, the continuous-time neural operator establishes a sliding window buffer of 100ms before projection. Within this buffer, the system performs downsampling preprocessing on the high-frequency signal with a sampling frequency of 10kHz, averaging every 10 sampling points. Simultaneously, it performs upsampling supplementation on the low-frequency environmental features with a sampling frequency of 1Hz based on linear interpolation, ensuring that the time step of all modes on the latent space axis is uniformly anchored at 1ms. The projection mapping function is implemented using a lookup table method based on 128 discrete mapping points, linearly mapping the signal amplitude normalized to the interval of 0.0 to 1.0 to the continuous hidden state tensor of the latent space. The lookup table step precision is set to 0.0078. The evolution rate calculation module projects the multi-source time-series signal data onto the continuous time axis of the latent space.

[0053] For the continuous hidden states in the latent space on the continuous time axis, the evolution rate calculation module uses an automatic differentiation engine to perform derivative operations, outputting the instantaneous evolution gradient tensor at time t, and calculating the L2 norm of the instantaneous evolution gradient tensor to obtain the evolution rate norm. Addressing the phenomenon in physical evolution where high-frequency transient excitations correspond to huge gradient norms, leading to the submergence of low-frequency background feature gradients, the feature decoupling and fusion module blocks the symmetric cross-computation flow between features, constructing an asymmetric computation graph. The feature decoupling and fusion module inputs the evolution rate norm into the supernetwork, which directly generates a dynamic weight matrix corresponding to the high-frequency feature tensor based on the sampling frequency difference. This dynamic weight matrix is ​​used to apply parameter-level corrections to the scoring weights of the high-frequency feature tensor in the attention mechanism, and combined with non-periodic instructions to extract the decoupled multi-source heterogeneous feature tensors, thus obtaining the evolution rate norm. The input is a one-dimensional scalar, which is mapped to a latent vector through two layers of multilayer perceptrons containing ReLU activation functions. The output layer linear transformation unit generates a tensor of the corresponding attention mechanism dimension, forming a dynamic weight matrix. The feature decoupling and fusion process superimposes the dynamic weight matrix onto the self-attention operator scoring function term, corrects the integral distribution of the query vector Q and the key vector K, suppresses the masking of low-frequency feature gradients by high-frequency modes, and ensures that the causal sequential relationship of the multi-source heterogeneous feature tensor in the latent space satisfies the dynamic constraints of the original physical process.

[0054] To address the situation where discretely triggered, non-periodic instructions are injected into the latent space of multi-source time-series signal data, causing discontinuities in physical states, the system monitors the arrival timestamps of the non-periodic instructions on a continuous time axis. When the latent space features are extrapolated to this timestamp The time prediction execution module instantaneously suspends the continuous temporal integration process of the neural differential evolution network, and the feature decoupling and fusion module extracts the previous hidden state tensor in the latent space at that moment. The eigenvectors of the non-periodic instruction mapping are input together to the discrete event response unit of the system. The internal multi-layer feedforward network calculates and outputs the state transition tensor Δh corresponding to the discrete operation excitation. The matrix operation unit then combines the state transition tensor Δh with the preceding hidden state tensor. Perform matrix addition to obtain the updated post-hidden state tensor The system will use this post-hidden state tensor By setting new initial values ​​for the integral boundary and reactivating the neural differential evolution network, the underlying computing power control link constrains the reset action of discrete pulse signals in multi-source heterogeneous time-series data to specific isolated time nodes, thereby eliminating the interference of non-periodic logic inputs on the topology of solving continuous ordinary differential equations.

[0055] The manifold curvature calculation module inputs multi-source heterogeneous feature tensors into the neural differential evolution network. The neural differential evolution network uses an internal multilayer perceptron to fit the time derivatives of the state variables and performs continuous time-domain integration based on ordinary differential equations within the latent space. Addressing the abrupt changes in local flow field curvature and numerical stiffness caused by the interaction of multi-scale dynamic features, the manifold curvature calculation module calculates the Jacobian matrix spectral radius ρ(J) by taking the partial derivative of the latent state evolution function of the neural differential evolution network. This Jacobian matrix spectral radius ρ(J) is used to quantify the current latent space curvature. The prediction execution module performs adaptive routing switching of the prediction operator based on the latent space curvature. When the Jacobian matrix spectral radius ρ(J) exceeds a preset curvature threshold γ, the prediction execution module blocks the explicit numerical integration calculation path and triggers an implicit numerical integration branch containing an independent Newton-Raphson root-seeking subgraph, performing implicit numerical integration based on latent space fixed-point iteration. γ is a preset curvature threshold, whose value is obtained by statistically analyzing the mean convergence curvature obtained during the early model training phase of the neural differential evolution network. The Jacobian matrix spectral radius ρ is calculated using a power-law iteration algorithm to quickly estimate the eigenvalues ​​of the Jacobian matrix of the hidden state evolution function. The maximum eigenvalue modulus is output through 3 to 5 iterations to characterize the curvature of the latent space. The preset curvature threshold γ is calibrated by statistically analyzing the distribution histogram of the convergence state spectral radius during offline training, and taking the 95th quantile of the distribution as the baseline value. In engineering deployment, it is fine-tuned according to the floating-point operation throughput per unit cycle of the processor to balance the calculation accuracy of explicit paths and the computational load of implicit paths. During the execution of implicit numerical integration, the prediction execution module introduces an adaptive damping term into the neural differential evolution network to limit the single-step change of the hidden state tensor in the latent space, thereby suppressing the numerical entropy increase phenomenon in the rigid flow field. The stability control of the implicit numerical integration branch is achieved by injecting regularization coefficients into the main diagonal of the Jacobian matrix. An adaptive damping term is constructed, with its coefficient positively correlated with the deviation of the current spectral radius ρ from the preset curvature threshold γ. Implicit numerical integration is performed, introducing the damping term into the neural differential evolution network to limit the single-step change of the hidden state tensor in the latent space, suppressing the numerical entropy increase and gradient explosion caused by nonlinear abrupt changes, so that the prediction result is within the physically defined convergence boundary during the long-period time-series extrapolation process. After performing numerical integration, the prediction execution module outputs the evolution prediction result of the target object. The prediction execution module calculates the residual between the evolution prediction result and the actual sensor observation value, and uses the residual to synchronously backcorrect the initial value of the hidden state of the neural differential evolution network. This scheme reconstructs the continuous-time domain feature evolution logic through asymmetric parameter generation and operator dynamic routing mechanism, isolates the backpropagation perturbation of different evolution rate features, and maintains the computational stability and physical convergence during the long-period integration process of high-dimensional heterogeneous tensors.

[0056] Example 2: This example constructs a coupled simulation platform for transient thermodynamics and rotor dynamics of an aero-engine. This platform solves the Navier-Stokes control equations to output aerodynamic and thermodynamic parameters, integrates the rotor dynamics model to calculate the physical displacement of the main shaft, and generates multi-source time-series signal data characterizing the physical state of the target object. The acquisition module extracts high-frequency dynamic characteristics of the main shaft support position, setting the sampling frequency to 10kHz; it also extracts low-frequency environmental characteristics of the combustion chamber outlet, setting the sampling frequency to 1Hz. To simulate electromagnetic interference and measurement errors in a real industrial environment, Gaussian white noise with a signal-to-noise ratio of 20dB is superimposed on the acquired original multi-source time-series signal data, and a 50Hz power frequency interference harmonic is injected. A preset curvature threshold γ is set, and the technical considerations for setting this parameter are... The consideration is to balance the continuous time-domain integration accuracy of the neural differential evolution network with the computational load of single-step inference. When multi-source time-series signal data input causes the characteristic flow field of the latent space to enter the rigid abrupt change region, the Jacobian matrix spectral radius ρ(J) expands. If the parameter value is too low, the network frequently blocks the explicit feedforward operation path and calls the implicit numerical integration including the inversion of the Hessian matrix, resulting in nonlinear saturation of the system's computational load. If the parameter value is too high, the network maintains the explicit feedforward operation path in the rigid region, and the local truncation error is amplified exponentially, causing the prediction results to diverge. Based on the above judgment rules, the system extracts the Jacobian spectral radius convergence sequence of the latent state evolution function of the neural differential evolution network in the standard working condition training stage, calculates the mean of the sequence, and thus determines the preset curvature threshold γ as 1.2.

[0057] The present invention establishes a sample group to perform asymmetric parameter reconstruction and adaptive operator routing based on latent space curvature. A comparative sample group one uses a symmetric self-attention network structure to process multi-source time-series signal data; a comparative sample group two uses a fixed explicit ordinary differential equation solver. The same thrust step test sequence is input to all three sample groups. The system projects this test sequence onto the continuous time axis of the latent space and performs derivative operations based on automatic differentiation. Observing the characteristics of the original input data, the 2-norm of the instantaneous evolution gradient tensor corresponding to high-frequency dynamic features reaches 45.2, while the evolution rate norm corresponding to low-frequency environmental features is only 0.03. In the symmetric computational structure of comparative sample group one, the high-frequency evolution rate norm dominates, causing the attention weight gradient of low-frequency environmental features to decay to below 0.001 during backpropagation, resulting in gradient flooding. The present invention's sample group inputs the evolution rate norm into a supernetwork, which generates a dynamic weight matrix based on the sampling frequency difference. It then uses this matrix to perform parameter reconstruction using a transformation operator on the high-frequency feature tensor. The results show that the present invention's sample group... The gradient of the low-to-medium frequency environmental features is maintained in the range of 0.45 to 0.65. The system outputs a decoupled multi-source heterogeneous feature tensor. The manifold curvature calculation module inputs the decoupled multi-source heterogeneous feature tensor into the neural differential evolution network to continuously calculate the Jacobian matrix spectral radius ρ(J). During the preceding steady cruise phase, the measured Jacobian matrix spectral radius ρ(J) fluctuates in the range of 0.4 to 0.7. The prediction execution module maintains an explicit feedforward operation path. The root mean square error of the prediction for the sample group of this invention remains at 0.05. After the thrust step change time window, the superimposed 20dB Gaussian white noise and power frequency disturbance caused a sudden increase in the curvature of the latent space flow field. The monitoring module recorded that the Jacobian matrix spectral radius ρ(J) jumped to 3.8, 4.5, and 4.2 in three consecutive time steps. At this time, the comparison sample group 2 continued to use an explicit feedforward calculation path. Since the integration step size could not constrain the abrupt change gradient, its prediction root mean square error quickly rose to 1.85, and the predicted trajectory completely deviated from the actual observation value. The sample group of this invention determined the Jacobian matrix spectral radius. If the curvature exceeds the preset curvature threshold of 1.2, the prediction execution module immediately blocks the explicit feedforward calculation path, introduces an adaptive damping term, and performs implicit numerical integration based on the iteration of the latent space fixed point. The measured root mean square error of the prediction of the sample group of the present invention converges to 0.08 within the abrupt change window.

[0058] An out-of-range control group targeting a preset curvature threshold γ was set up to verify the nonlinear impact of parameter boundaries on computational latency and prediction accuracy. Five hundred time-step sequences containing extreme surge characteristics were input in parallel to three models with thresholds set to 0.8, 1.2, and 2.5, respectively. Test data showed that the control group using a lower threshold of 0.8 frequently triggered implicit numerical integration in the small perturbation region, resulting in a global average computational latency of 42ms, exceeding the real-time monitoring cycle limit for industrial control, and exhibiting nonlinear overload characteristics in the computational load. The control group using an upper threshold of 2.5 failed to switch operators when crossing moderately rigid regions, and the accumulated truncation error caused a peak in the local root mean square error of the prediction. A spike with a value of 1.05 results in a loss of state evolution prediction accuracy. Using the sample group of this invention with a baseline value of 1.2, the average computational delay remains at 18ms, and the root mean square error of global prediction does not exceed 0.08. The above test data show that the system reconstructs the scoring weights of high-frequency feature tensors based on sampling frequency differences, suppresses gradient imbalance in the joint processing of heterogeneous features, and performs operator adaptive routing switching based on the latent space curvature quantized by the Jacobian matrix spectral radius. When encountering strong disturbance transient conditions, it absorbs the numerical rigidity generated by the cross-scale fusion of multi-source heterogeneous time series data, and maintains the physical convergence and computational stability of long-period physical state evolution prediction in the continuous time domain.

[0059] Example 3: The current gas turbine rotor full life cycle degradation simulation environment provides an offline training sample set for system parameter calibration. This set includes real historical data of high-frequency dynamic features and low-frequency environmental features recorded during the rotor's complete operating cycle. Before the system is officially deployed online, a model building module is configured to establish the neural network topology and quantify the baseline thresholds required for adaptive routing of the prediction operator. The computational path of multi-source heterogeneous time-series signals in feature reconstruction and implicit numerical integration is clarified. The model building module establishes the internal hardware execution architecture of the hypernetwork, which includes two layers of multilayer perceptrons with ReLU activation functions. The input of the hypernetwork receives the evolution rate... The rate norm scalar is mapped to a one-dimensional flattened vector by the multilayer perceptron and output at the output end. The tensor reshaping unit receives the one-dimensional flattened vector and reshapes it into a two-dimensional target dimension according to the feature dimension of the multi-source heterogeneous feature tensor, forming a dynamic weight matrix. The system uses this dynamic weight matrix to directly replace the static key-value query matrix used for feature extraction in the original attention mechanism. The model building module configures a neural differential evolution network and sets it as a residual network derivative fitting operator containing three cascaded fully connected layers. This derivative fitting operator receives the continuous hidden state tensor in the latent space as input, calculates and outputs the first-order partial derivative of the hidden state tensor with respect to time.

[0060] During the offline training phase, the training execution module inputs the offline training sample set into the neural differential evolution network and continuously updates the network weight parameters using the backpropagation algorithm. The system monitors the rate of decrease of the loss function value according to the training iteration cycle. The set judgment rule is: when the rate of decrease of the loss function is less than 0.001 for 5 consecutive training iteration cycles, the neural differential evolution network is determined to have reached the convergence state. After satisfying this convergence state, the curvature statistics module continuously extracts the hidden state tensors generated in the latent space for the next 100 training batches. For each hidden state tensor, the automatic differentiation engine calculates the Jacobian matrix of the derivative fitting operator and solves for the largest absolute eigenvalue of the Jacobian matrix as the single-order spectral radius sample. The system integrates... The 100 sampled data sets were used to construct a convergent curvature sequence. The parameter calibration module calculated the arithmetic mean of this convergent curvature sequence. The arithmetic mean output by the test platform was 1.18. The parameter calibration module fixed this value of 1.18 as the preset curvature threshold γ. The network hierarchy mapping flow direction and structural parameter dimension definition of the supernetwork and neural differential evolution network were established. The system generates a high-dimensional dynamic weight matrix that constrains the high-frequency feature weights based on discrete scalar input. Using the convergence state criterion based on the continuous loss function decline rate and the statistical algorithm of the mean radius of the cross-batch Jacobian matrix spectrum, the predictive operator switching benchmark for dealing with rigid abruptly changing flow fields was quantitatively established, ensuring the continuous execution of the operator adaptive routing mechanism in the online running state.

[0061] Example 4: When the system faces the specific condition of deploying the cross-domain state evolution prediction architecture to a new batch of aero-engine entities, the external electromagnetic environment of the target object and the installation tolerance of physical sensors will cause nonlinear initial zero-point drift error in the multi-source time-series signal data before it is projected onto the continuous time axis of the latent space. Before the main prediction task starts, the data acquisition module opens a reference no-load test channel for 300 seconds to collect the high-frequency dynamic characteristics and low-frequency environmental characteristics of the target object in the non-combustion driven state, calculates its steady-state noise root mean square value and the static environmental bias of the low-frequency environmental characteristics, and injects the steady-state noise root mean square value and the static environmental bias as calibration reference values ​​into the hardware preprocessing port of the data acquisition module. The underlying differential amplifier circuit is used to subtract the static environmental bias from the subsequently continuously acquired multi-source time-series signal data. At the same time, a digital bandpass filter is used to remove high-frequency clutter components whose signal amplitude exceeds the set multiple of the steady-state noise root mean square value, so that the continuous latent state entering the latent space is forcibly anchored to the physical reference zero plane that eliminates environmental differences.

[0062] After eliminating the bias of the underlying sensor physical signals, the evolution rate calculation module initiates a fine-tuning program for the dynamic parameters of the specific mechanical entity. A test step excitation signal with an amplitude of 0.5 MPa is applied to the target object, and the instantaneous evolution gradient tensor in the latent space is recalculated. The manifold curvature calculation module extracts the Jacobian matrix under this step excitation response and calculates the initial Jacobian matrix spectral radius. ,in To quantify the scalar parameters characterizing the numerical stiffness of the current physical equipment foundation, the system calculates the spectral radius of the initial Jacobian matrix. The dimensionless scaling factor between the supernetwork and the preset curvature threshold γ determined in the early training is used by the parameter calibration module to reverse reset the output bias node parameters of the multilayer perceptron inside the supernetwork. This ensures that the dynamic weight matrix generated by the supernetwork for the current hardware platform is aligned with the basic evolution rate norm in terms of numerical magnitude. After the baseline calibration and tensor mapping relationship synchronization program is completed, the neural differential evolution network enters an online working state of continuously receiving multi-source heterogeneous feature tensors and outputting the evolution prediction results of the target object based on adaptive routing.

[0063] Example 5: When the aero-engine online monitoring platform is dealing with the continuous increase in the curvature of the latent space flow field caused by the compressor stall sudden change condition, the prediction execution module receives a multi-source heterogeneous feature tensor containing extremely high frequency distortion vibration characteristics. The manifold curvature calculation module measures that the Jacobian matrix spectral radius continuously exceeds the preset curvature threshold, triggering an implicit numerical integration branch. The Newton-Raphson root-seeking subgraph inside this implicit numerical integration branch generates nonlinear oscillations when dealing with extreme rigidity characteristics, causing the single-step integration calculation time to approach the maximum hardware cycle threshold of 20ms limited by the predictive control system.

[0064] The predictive execution module initiates an online adaptive damping adjustment procedure for extreme rigid flow fields. The curvature monitoring operator extracts the Jacobian matrix spectral radius ρ(J) and the preset curvature threshold γ at the current time step. The system constructs a dynamic penalty function based on curvature deviation to calculate the adaptive damping coefficient λ. The specific calculation logic satisfies λ=κ⋅(ρ(J)-γ), where κ is a dimensionless penalty constant preset based on the upper limit of the floating-point operation throughput of the underlying processor. The matrix operation unit superimposes the adaptive damping coefficient onto the main diagonal elements of the implicit numerical integral Jacobian matrix, forcibly compressing the update step size of the hidden state tensor during the iterative solution process. The system sets the residual convergence judgment threshold to 0.005. When the L2 norm of the hidden state residual vector generated by two consecutive iterations is lower than this judgment threshold, the controller terminates the internal iterative loop and outputs the hidden state update value at the current time step.

[0065] The implicit computational branch that injects adaptive damping coefficients changes the condition number of the system's Jacobian matrix and converges the numerical divergence boundary of the high-dimensional feature space, keeping the number of single-step iterations of the internal inversion process within 5. The hardware bus monitoring node records that the global computation time of a single implicit integral has fallen back and is maintained at a constant 12ms, eliminating the risk of task thread blocking caused by computing power exhaustion. The prediction execution module sends the evolution prediction results of the target object to the external interface based on the continuously updated implicit state, maintaining the computational convergence state and continuous operation capability of the multi-source heterogeneous time series data evolution prediction system under non-stationary disaster conditions.

[0066] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for predicting the evolution of multi-source heterogeneous time-series data based on deep learning, characterized in that, Includes the following steps: Step 101: Obtain multi-source time-series signal data characterizing the physical state of the target object. The multi-source time-series signal data includes high-frequency dynamic features, low-frequency environmental features, and discretely triggered non-periodic commands. Step 102: Project the multi-source time-series signal data onto the continuous time axis of the latent space, and calculate the derivative magnitude of each modal feature with respect to time at a given time to obtain the evolution rate norm; wherein, step 102 includes the following sub-steps: Step 1021: Map the discrete non-periodic instructions, low-frequency environmental features, and high-frequency dynamic features in the multi-source time-series signal data to the latent space through a continuous-time neural operator; Step 1022: Perform derivative operations based on automatic differentiation on the continuous hidden states in the latent space to obtain the instantaneous evolution gradient tensor at time t; Step 1023: Calculate the second norm of the instantaneous evolution gradient tensor to obtain the evolution rate norm; Step 103: The evolution rate norm is input into the supernetwork. The supernetwork generates a dynamic weight matrix corresponding to the high-frequency feature tensor based on the sampling frequency difference. The dynamic weight matrix is ​​then used to correct the scoring weights of the high-frequency feature tensor in the attention mechanism. Combined with non-periodic instructions, the decoupled multi-source heterogeneous feature tensor is extracted. The supernetwork transforms the feature vectors of the low-frequency mode into a dynamic weight matrix. The dynamic weight matrix is ​​then used to perform parameter reconstruction on the transformation operator of the high-frequency feature tensor to establish asymmetric constraint relationships between modes. Step 104: Input the multi-source heterogeneous feature tensor into the neural differential evolution network, and calculate the Jacobian matrix spectral radius by taking the partial derivative of the hidden state evolution function of the neural differential evolution network to characterize the curvature of the latent space. Step 105: Perform adaptive routing of the prediction operator based on the latent space curvature. When the latent space curvature exceeds the preset curvature threshold, perform implicit numerical integration based on the latent space fixed point iteration and output the evolution prediction result of the target object. Furthermore, in step 105, the adaptive routing of the prediction operator satisfies the following decision rule: , where S is the selected prediction operator path; ρ(J) is the spectral radius of the Jacobian matrix; γ is the preset curvature threshold, which is a value preset based on the mean convergence curvature of the neural differential evolution network during the training phase.

2. The method for predicting the evolution of multi-source heterogeneous time-series data based on deep learning according to claim 1, characterized in that, In step 104, the neural differential evolution network uses a neural network to fit the time derivative of the state variables and performs continuous time-domain integration based on ordinary differential equations in the latent space.

3. The method for predicting the evolution of multi-source heterogeneous time-series data based on deep learning according to claim 1, characterized in that, In step 105, when performing implicit numerical integration, an adaptive damping term is introduced into the neural differential evolution network to suppress the increase of numerical entropy during the computation process by limiting the single-step change of the hidden state tensor in the latent space.

4. The method for predicting the evolution of multi-source heterogeneous time-series data based on deep learning according to claim 1, characterized in that, In step 101, the sampling frequency range of the multi-source timing signal data covers 1Hz to 10kHz and includes non-periodic triggered pulse logic signals.

5. The method for predicting the evolution of multi-source heterogeneous time-series data based on deep learning according to claim 1, characterized in that, In step 103, an asymmetric computation graph is constructed using the magnitude difference in the evolution rate norm. By isolating the backpropagation gradient flow of different frequency modes, the feature flooding of low-frequency modes caused by the dominance of high-frequency mode gradients is eliminated.

6. The method for predicting the evolution of multi-source heterogeneous time-series data based on deep learning according to claim 1, characterized in that, It also includes step 106: after obtaining the evolution prediction results, calculate the residual between the prediction results and the actual observations, and use the residual to synchronously correct the initial value of the hidden state of the neural differential evolution network. The method is used to realize the continuous time-domain feature extraction and trend prediction of industrial equipment operation status data.

7. A deep learning-based system for predicting the evolution of multi-source heterogeneous time-series data, used to execute the deep learning-based method for predicting the evolution of multi-source heterogeneous time-series data as described in claim 1, characterized in that, include: The data acquisition module is used to acquire multi-source time-series signal data that characterizes the physical state of the target object; The evolution rate calculation module projects multi-source time-series signal data onto a continuous time axis in the latent space and calculates the modulus of the derivative of each modal feature with respect to time at a given moment to obtain the evolution rate norm. Specifically, discrete non-periodic instructions, low-frequency environmental features, and high-frequency dynamic features in the multi-source time-series signal data are mapped to the latent space through continuous-time neural operators. Derivative operations based on automatic differentiation are performed on the continuous hidden states in the latent space to obtain the instantaneous evolution gradient tensor at time t. The second norm of the instantaneous evolution gradient tensor is calculated to obtain the evolution rate norm. The feature decoupling and fusion module, whose input is connected to the output of the evolution rate calculation module, is used to input the evolution rate norm into the supernetwork. The supernetwork generates a dynamic weight matrix to correct the scoring weights of high-frequency feature tensors in the attention mechanism, and extracts the decoupled multi-source heterogeneous feature tensors by combining non-periodic instructions. The supernetwork transforms the feature vectors of low-frequency modes into dynamic weight matrices, and uses the dynamic weight matrix to perform parameter reconstruction on the transformation operators of high-frequency feature tensors to establish asymmetric constraint relationships between modes. The manifold curvature calculation module is used to input multi-source heterogeneous feature tensors into the neural differential evolution network, and characterize the latent space curvature by calculating the Jacobian matrix spectral radius of the latent state evolution function of the neural differential evolution network. The prediction execution module is used to perform adaptive route switching of prediction operators based on the latent space curvature, and when the latent space curvature exceeds the preset curvature threshold, it performs implicit numerical integration through fixed point iteration to output the evolution prediction result of the target object. Furthermore, the adaptive routing of the prediction operator satisfies the following decision rules: , where S is the selected prediction operator path; ρ(J) is the spectral radius of the Jacobian matrix; γ is the preset curvature threshold, which is a value preset based on the mean convergence curvature of the neural differential evolution network during the training phase.

Images