CN-122015935-A - Method and device for dynamically compensating interpretable heterogeneous sensor embedded with physical constraint

CN122015935ACN 122015935 ACN122015935 ACN 122015935ACN-122015935-A

Abstract

The application discloses an interpretable heterogeneous sensor dynamic compensation method and device embedded with physical constraints, and relates to the field of sensors. The method comprises the steps of obtaining signals to be compensated and corresponding reference signals of heterogeneous sensors, constructing a collaborative matrix based on the signals to be compensated and the reference signals, constructing an artificial convolution kernel based on the collaborative matrix, carrying out physical priori initialization by utilizing a convolution kernel of an adaptive spectrum block of an artificial convolution check TSLANet, training TSLANet after the physical priori initialization, introducing physical constraint to construct total loss containing a plurality of target loss functions, training to obtain a compensation model for dynamically compensating the heterogeneous sensors, and self-adaptively adjusting training parameters based on learning track records and a time-frequency domain energy evolution process in the training process. By the method, the interpretable features, the physical priors and the depth compensation network are subjected to depth fusion, so that an integral interpretable optimization mechanism from feature representation to training process is formed.

Inventors

XU BO
WANG SHUO
TANG HAO
YOU JIA
XU DONGYANG

Assignees

海南大学

Dates

Publication Date: 20260512
Application Date: 20260127

Claims (10)

1. A method of dynamic compensation of an interpretable heterogeneous sensor embedded in a physical constraint, the method comprising: Acquiring a signal to be compensated and a corresponding reference signal of a heterogeneous sensor, and constructing a synergy matrix integrating frequency domain energy characteristics and time domain attribution characteristics based on the signal to be compensated and the reference signal; constructing an artificial convolution kernel based on the cooperative matrix, and carrying out physical priori initialization by utilizing the convolution kernel of the self-adaptive spectrum block of the artificial convolution kernel TSLANet; Training TSLANet after physical priori initialization, introducing physical constraint to construct total loss containing multiple objective loss functions, training to obtain a compensation model for dynamically compensating the heterogeneous sensor, and self-adaptively adjusting training parameters based on learning track records and time-frequency domain energy evolution processes in the training process.
2. The method of claim 1, wherein constructing a synergy matrix integrating frequency domain energy features and time domain attribution features based on the signal to be compensated and the reference signal comprises: Respectively processing the signal to be compensated and the reference signal by utilizing short-time Fourier transformation, and correspondingly acquiring a frequency spectrum of the signal to be compensated and a frequency spectrum of the reference signal; According to the amplitude, phase and power spectrum density deviation between the signal spectrum to be compensated and the reference spectrum, obtaining error energy distribution as the frequency domain energy characteristic; Performing attribution analysis on the pre-trained compensation model based on an SHAP interpretability method to obtain attribution weights of all time steps as the time domain attribution features; constructing a frequency band weight corresponding to each sensor in the heterogeneous sensors based on priori physical knowledge; and constructing the coordination matrix according to the frequency domain energy characteristic, the time domain attribution characteristic and the frequency band weight.
3. The method of claim 1, wherein constructing an artificial convolution kernel based on the co-matrix comprises: Normalizing the synergistic matrix to obtain a normalized matrix and element sum of the normalized matrix; Setting an energy coverage threshold, and searching a rectangular target area with the smallest area on a time-frequency plane constrained by the window size of the normalized matrix time and frequency direction, wherein the sum of elements in the target area is not smaller than the product of the sum of elements of the normalized matrix and the energy coverage threshold; rearranging elements in the target area into submatrices according to time and frequency sequences; and carrying out scale normalization and smoothing on the submatrices to obtain a convolution kernel weight matrix serving as the artificial convolution kernel.
4. The method of claim 1, wherein the total loss comprises a fit loss function, a frequency characteristic constrained target loss function, a dynamics constrained target loss function, and an constitutive constraint target loss function.
5. The method of claim 4, wherein the target loss function of the frequency characteristic constraint is determined based on a variance between a standard frequency response and an output frequency response, wherein the output frequency response is obtained by Fourier transforming a compensation output of the compensation model.
6. The method of claim 4, wherein the method of constructing the dynamically constrained target loss function comprises constructing a dynamic residual from a compensation output of the compensation model, a first derivative and a second derivative of the compensation output, and constructing the dynamically constrained target loss function from the dynamic residual.
7. The method according to claim 4, wherein the constructing method of the objective loss function of the constitutive relation constraint comprises constructing a constitutive residual according to at least one parameter of equivalent stress, equivalent strain and temperature change of a sensor, and constructing the objective loss function of the constitutive relation constraint according to the constitutive residual.
8. The method of claim 1, wherein the compensation model comprises a number of convolution layers; The self-adaptive training parameter adjustment comprises the following steps of: for each convolution layer, acquiring a weight tensor of the convolution layer during each round of training to form a weight track sequence, acquiring an accumulated weight variation representing the variation degree of the weight tensor according to the weight track sequence, and adjusting the weight of the convolution layer or the convolution layer according to the accumulated weight variation; Aiming at each training round, acquiring time-frequency energy distribution according to the compensation output of the compensation model, carrying out alignment analysis on the time-frequency energy distribution and target energy distribution, acquiring a correlation coefficient, and adjusting the weight coefficient of each target loss function in the total loss according to the correlation coefficient.
9. An interpretable heterogeneous sensor dynamic compensation device embedded in a physical constraint, the device comprising: The data processing module is used for acquiring signals to be compensated and corresponding reference signals of the heterogeneous sensor, and constructing a synergy matrix integrating frequency domain energy characteristics and time domain attribution characteristics based on the signals to be compensated and the reference signals; the model construction module is used for constructing an artificial convolution kernel based on the collaborative matrix, and carrying out physical priori initialization by utilizing the convolution kernel of the self-adaptive spectrum block of the artificial convolution kernel TSLANet; the training module is used for training TSLANet after physical priori initialization, introducing physical constraint to construct total loss comprising a plurality of target loss functions, and training to obtain a compensation model for dynamically compensating the heterogeneous sensor; and the interpretability evaluation and feedback module is used for adaptively adjusting training parameters based on the learning track record and the time-frequency domain energy evolution process in the training process.
10. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the method of any one of claims 1 to 8 when the computer program is executed.

Description

Method and device for dynamically compensating interpretable heterogeneous sensor embedded with physical constraint Technical Field The application relates to the technical field of sensors, in particular to a dynamic compensation method and device for an interpretable heterogeneous sensor embedded with physical constraints. Background Along with the deep promotion of the national 'novel industrialization' construction and 'informatization upgrading' strategy, the field of high-end equipment such as aerospace and the like has higher requirements on high-precision and high-reliability testing and state monitoring capability. In the face of engine testing, strong impact and wide-band dynamic environments, the sensor output is often accompanied by significant dynamic distortion and frequency response distortion. In order to restore the real signal, the traditional compensation method based on an empirical mode or a fixed filtering structure is used as the conventional first section of engineering for a long time. However, such methods generally assume that the dynamic characteristics of the sensor remain stable in the time and frequency domains, and it is difficult to accurately characterize the distortion mechanism under complex, nonlinear, and time-varying conditions, resulting in limited compensation capability. With the wide application of deep learning technology in recent years, the strong characteristic expression capability is introduced into the dynamic compensation task of the sensor, and the compensation precision, the frequency response recovery capability and the fitting capability to complex dynamics are remarkably improved. Nevertheless, most of the existing researches focus on optimization of indexes such as compensation precision and loss function convergence, a black box network structure with random initialization is generally adopted, and the description and the constraint on internal decision logic are lacked. The model is often non-traceable in what frequency band and based on what dynamic characteristics, so that the non-transparency of the compensation process is caused, the association relation between key characteristics and physical mechanisms is difficult to analyze, and the model is difficult to directly apply in scenes with extremely high requirements on reliability and traceability, such as aerospace, ship positioning and the like. In order to improve the credibility and usability of the model, expert students put forward an interpretable artificial intelligence (XAI, explainable ARTIFICIAL INTELLIGENCE) technology, try to make the decision process of the deep learning model as transparent as possible through methods such as feature importance assessment, model visualization, rule extraction and the like, and reversely guide model training and structure optimization by utilizing interpretation results. However, the current interpretable artificial intelligence method for the dynamic compensation of the sensor still has defects in various aspects, and mainly solves the problems that the interpretable analysis stays in the visualization of the postmortem characteristic significance, the internal representation of the model is disjointed with the physical mechanism of the dynamic response, the commonality and the dissimilarity of the heterogeneous sensors of multiple types are difficult to consider under a unified framework, and the like. Disclosure of Invention Based on the above, it is necessary to provide a method and a device for dynamic compensation of an interpretable heterogeneous sensor embedded with physical constraints, so as to improve the interpretability and the reliability of a sensor compensation model. In a first aspect, the present application provides a method of dynamic compensation of an interpretable heterogeneous sensor embedded with physical constraints. The method comprises the following steps: acquiring a signal to be compensated and a corresponding reference signal of a heterogeneous sensor, and constructing a synergy matrix integrating frequency domain energy characteristics and time domain attribution characteristics based on the signal to be compensated and the reference signal; Constructing a manual convolution kernel based on the cooperative matrix, and carrying out physical priori initialization by utilizing the manual convolution kernel of the TSLANet self-adaptive spectrum block; Training TSLANet after physical priori initialization, introducing physical constraint to construct total loss containing multiple objective loss functions, training to obtain a compensation model for dynamically compensating heterogeneous sensors, and self-adaptively adjusting training parameters based on learning track records and time-frequency domain energy evolution processes in the training process. In one embodiment, constructing a synergy matrix integrating the frequency domain energy characteristic and the time domain attribution characteristic based on the signal