CN-121980149-A - Time sequence generation method and system based on uncertainty quantization and frequency domain constraint

CN121980149ACN 121980149 ACN121980149 ACN 121980149ACN-121980149-A

Abstract

The invention discloses a time sequence generation method and a system based on uncertainty quantization and frequency domain constraint, which belong to the technical field of artificial intelligence and data mining, and are characterized by obtaining irregular time sequence data, predicting missing values of sequences by utilizing a probability complement network, outputting a mean value of the complement values and a variance representing uncertainty of the complement results, constructing an entropy weight self-adaptive mask generation mechanism, generating a soft mask matrix of continuous values according to the variance and local observation density, wherein high uncertainty corresponds to low mask weight, constructing a mixed loss function comprising time domain weighting loss and frequency domain consistency loss, and training a diffusion model based on the soft mask matrix and the mixed loss function to generate a regularized high-fidelity time sequence. The invention solves the problems of stiffness and incapability of perceiving the complement error of the binary mask mechanism in the prior art by introducing the Bayesian uncertainty quantization and the soft mask mechanism, and simultaneously ensures the periodic characteristics of the generated data by combining the frequency domain constraint.

Inventors

LIU DONGSHENG
SUN RENJIE
Jiao ting
GAN QIYUN
CHEN HAO
WANG XINGEN
WU TONG
LI YING
JIN RUI
XIAO JUN
GUO FEIPENG
SHEN ZHONGTAO

Assignees

浙江工商大学

Dates

Publication Date: 20260505
Application Date: 20260331

Claims (10)

1. The time sequence generation method based on uncertainty quantization and frequency domain constraint is characterized by comprising the following steps: acquiring irregular industrial time series data of a target object based on an industrial sensor, preprocessing by using a complement network, taking an output mean value as a complement point predicted value, and taking a variance as an uncertainty measure; Generating a soft mask weight matrix by utilizing an entropy weight self-adaptive mechanism based on the uncertainty measure and the local observation density; Constructing a diffusion model, driving the diffusion model to train under the constraint of time-frequency double-domain consistency based on the mixed loss comprising the time-domain weighting loss and the frequency-domain consistency loss, and restricting to generate a new industrial time sequence which is consistent with the original industrial time sequence in the frequency domain; And feeding the new industrial time sequence serving as prompt information back to the completion network to iteratively correct the industrial time sequence generated by the diffusion model.
2. The method of claim 1, wherein a dynamic mask generator is constructed, an entropy weight adaptive soft mask mechanism is introduced, a ratio of one to the uncertainty measure is summed with the local observation density through a balance parameter to generate a soft mask weight of continuous values, when the uncertainty measure of a complement point is extremely high, the soft mask weight is close to 0, so that the complement point is ignored by a diffusion model in subsequent calculation, and when the uncertainty measure is low and the local observation density is high, the soft mask weight is close to 1, so that the complement point is taken as effective context information by the diffusion model.
3. The method for generating a time series based on uncertainty quantization and frequency domain constraint according to claim 2, wherein the local observation density is a local observation density, a sliding time window centered at a certain time point is set, the number of real observation values in the window is counted, and the ratio of the number of the real observation values to the window length is used as the local observation density at the certain time point to represent the sparseness of industrial time series data near the certain time point.
4. The method of claim 1, wherein the loss function of the frequency domain consistency loss is obtained by performing fast Fourier transform on the industrial time sequence and the original industrial time sequence or the complement industrial time sequence generated by the diffusion model, respectively, and calculating the Euclidean distance between the obtained diffusion frequency domain and the original or complement frequency domain.
5. The method of claim 4, wherein the loss function of the time domain weight loss is obtained by summing the product of the reconstruction error of the original industrial time sequence or the completed industrial time sequence and the new industrial time sequence and the soft mask weight of the corresponding time point in time.
6. The method of generating a time series based on uncertainty quantization and frequency domain constraints as set forth in claim 5, wherein said loss function of mixed loss is summed with a loss function of said consistency loss and a loss function of time domain weighted loss based on a balance coefficient.
7. The method for generating a time series based on uncertainty quantization and frequency domain constraint according to claim 1, wherein the iterative correction is to initially generate the new industrial time series based on a trained diffusion model, feed back the new industrial time series as prompt information to the completion network, recalculate an uncertainty metric and a corresponding soft mask weight matrix, execute a new round of diffusion based on the updated soft mask matrix until the generated industrial time series reaches low uncertainty, and finally obtain the industrial time series through the trained diffusion model.
8. The method for generating a time sequence based on uncertainty quantization and frequency domain constraint according to claim 1, wherein an industrial time sequence with missing values or non-uniform sampling and corresponding time stamps thereof are obtained, a probability complement network is constructed for capturing bidirectional time sequence dependent characteristics of the time sequence, an output layer comprises a full connection layer for mapping hidden layers into distribution parameters of predicted values, forward sampling is performed on the probability complement network for a plurality of times in an inference stage, statistical characteristics of the probability complement network are calculated for each missing time point, an obtained average value is used as the predicted value, and an obtained variance is used as the uncertainty measure.
9. The time sequence generation system based on uncertainty quantization and frequency domain constraint comprises an uncertainty quantization module, a diffusion generation module, a loss function and a constraint module, and is characterized in that the time sequence generation method based on uncertainty quantization and frequency domain constraint is adopted, and the completion of irregular industrial time sequence data and soft mask weight generation, the construction of a diffusion model and the training of the diffusion model based on the loss function and the constraint are sequentially executed.
10. The system for generating a time series based on uncertainty quantization and frequency domain constraints of claim 9, wherein said uncertainty quantization module comprises a variance density estimator and said loss function and constraint module comprises a dynamic mask generator; The variance density estimator takes the variance of the industrial time sequence as uncertainty measurement, sets a sliding time window taking a certain time point as a center, counts the number of real observation values in the window, takes the ratio of the number of the real observation values to the window length as the local observation density of the certain time point, and represents the sparseness of the industrial time sequence data near the certain time point; The dynamic mask generator sums a ratio of one to the uncertainty measure and the local observation density through balance parameters based on an entropy weight self-adaptive soft mask mechanism to generate soft mask weights with continuous values, when the uncertainty measure of a complement point is extremely high, the soft mask weights are close to 0 so that the complement point is ignored by a diffusion model in subsequent calculation, and when the uncertainty measure is low and the local observation density is high, the soft mask weights are close to 1 so that the complement point is used as effective context information by the diffusion model.

Description

Time sequence generation method and system based on uncertainty quantization and frequency domain constraint Technical Field The invention belongs to the technical field of artificial intelligence and data mining, and particularly relates to a time sequence generation method and system based on uncertainty quantization and frequency domain constraint. Background Time series analysis plays a vital role in industrial monitoring, medical diagnostics, and economic prediction. However, real world acquired time series data tends to have a high degree of irregularity, mainly manifested by non-uniformity of sampling intervals and missing values at a large number of time points. Diffusion Models (Diffusion Models) have achieved great success in the field of image generation as a powerful generative model by adding noise corrupted data step by step and denoising the reconstructed original samples in the reverse process. However, standard diffusion models typically rely on regular grid data structures and cannot handle the irregular sequences directly and efficiently. Conventional fault diagnosis or sequence repair methods typically employ a two-stage strategy of "interpolation complement + mask training". However, the related art has the following significant drawbacks: Firstly, a mask mechanism is rigidified, the existing method generally only adopts a static binary mask (namely, an observation point is marked as 1, a complement point is marked as 0), and the non-black and white mechanism ignores the potential value of the complement data and cannot distinguish interpolation information with high reliability and low reliability; the existing completion module generally only outputs a single certainty predicted value, and cannot quantify the Uncertainty (Uncertinty) of the completion result, so that the model is easily misled by the wrong completion value; finally, the prior art processes the time sequence based on visual angle, mainly focuses on point-to-point reconstruction in the time domain, and easily loses the periodic characteristics of the time sequence in the frequency domain, so that the Long-range dependence (Long-TERM DEPENDENCY) of the generated data is insufficient. Disclosure of Invention In order to solve the defects in the prior art, realize the combination of probability uncertainty quantization and frequency domain constraint, effectively adapt to the time sequence diffusion of irregular sampling characteristics, and generate and repair irregular time sequence data with high fidelity, the invention adopts the following technical scheme: the time sequence generation method based on uncertainty quantization and frequency domain constraint comprises the following steps: Acquiring irregular industrial time series data of a target object based on an industrial Internet of things sensor, preprocessing by using a probability complement network, taking the output mean value as a complement point predicted value, and taking the variance as an uncertainty measure; Generating a soft mask weight matrix by utilizing an entropy weight self-adaptive mechanism based on the uncertainty measure and the local observation density; constructing a diffusion model, based on mixed loss comprising time domain weighting loss and frequency domain consistency loss, driving the diffusion model to train under the constraint of time-frequency double-domain consistency, so as to generate a new industrial time sequence which is consistent with the original industrial time sequence in frequency domain distribution through Fast Fourier Transform (FFT) constraint, aligning the forcedly generated data with real data on frequency spectrum characteristics, effectively solving the problem of lack of periodicity of the sequence, enhancing the information integrity of the generated data, and the mixed loss function not only can overcome the defect that only attention is paid to time domain reconstruction in the prior art, but also can avoid the defect that how long the visual view angle only attention is paid to time domain point-to point reconstruction in the prior art by the loss of frequency domain consistency, thereby causing lack of periodicity and long-range dependence of the generated data; And feeding the new industrial time sequence serving as prompt information back to the complement network to realize closed-loop optimization so as to iteratively correct the industrial time sequence generated by the diffusion model. Further, a dynamic mask generator is constructed, an entropy weight self-adaptive soft mask mechanism is introduced, and the ratio of one to the uncertainty measure and the local observation density are summed through balance parameters to generate soft mask weights with continuous values so as to make the uncertainty measure of the complement pointsExtremely high, soft mask weightsApproaching 0 so that the diffusion model ignores the complement point in subsequent calculations, soft mask weights when uncertainty met