Search

CN-122021776-A - Multi-batch variable error system identification method based on noise enhancement contrast learning

CN122021776ACN 122021776 ACN122021776 ACN 122021776ACN-122021776-A

Abstract

The invention belongs to the technical field of system identification, and particularly relates to a multi-batch variable error system identification method based on noise enhancement contrast learning. Aiming at the problems of deviation of input noise variance estimation and noise amplification in the data conversion process, a noise enhancement auxiliary sequence comparison learning algorithm is provided. The algorithm is characterized in that a pseudo-denoising sequence, the original data and a noise enhancement sequence are represented by a triplet, and the original data is enabled to approach the pseudo-denoising sequence and be far away from the noise enhancement sequence by utilizing a contrast learning mechanism, so that dependence on accurate noise variance is reduced, and feature reconstruction capability and noise robustness are improved. The encoder can effectively extract the sequence characteristics and overcome the limitation of the adaptation of the traditional method to the changing operation points. The effectiveness of the proposed algorithm was verified by cascading the tank examples.

Inventors

  • MA JUNXIA
  • LI RONGHUAN
  • ZHANG XUHANG
  • SHI XUDONG
  • XIONG WEILI

Assignees

  • 江南大学

Dates

Publication Date
20260512
Application Date
20260116

Claims (10)

  1. 1. A multi-batch variable error system identification method based on noise enhancement contrast learning is characterized by comprising the following steps: Step 1, acquiring original observation data of a multi-batch variable error EIV system, wherein the original observation data comprises known input source signals, actual observation input variables polluted by measurement noise and system output; Step 2, an input generation model is built based on the original observation data, and the input generation model is used for describing a dynamic generation process of a pseudo noise-free input signal; Step 3, constructing and applying a noise enhancement auxiliary sequence contrast learning NSCL algorithm frame, wherein the noise enhancement auxiliary sequence contrast learning NSCL algorithm frame comprises a characteristic representation learning module, a denoising device driven contrast learning module and a prediction regression module; The feature representation learning module includes mapping, by an encoder, the pseudo noise-free input signal into a high-dimensional potential representation; The contrast learning module driven by the denoising device comprises a positive sample generation module, a negative sample generation module, a three-tuple construction module, a contrast loss function optimization module and a comparison module, wherein the positive sample generation module is used for generating pseudo-denoising data, the negative sample generation module is used for generating noise enhancement data, the positive sample generation module is used for mapping positive samples, negative samples and pseudo-noiseless input signals into high-dimensional potential representations through an encoder respectively, wherein Gao Weiqian of the pseudo-noiseless input signals are used as anchor points, and Gao Weiqian of the anchor points approaching to the positive samples is represented and far from the high-dimensional potential representations of the negative samples through optimization of the contrast loss function; The predictive regression module comprises performing linear regression prediction on the Gao Weiqian representation of the optimized pseudo-noiseless input signal to obtain a predictive value output by the multi-batch variable error EIV system.
  2. 2. The method of claim 1, wherein the input generation model in step 2 is described by the following state space equations: Wherein, the For a known input source signal at a previous time instant, In order for the process to be noisy, In order to measure the noise of the light, For an estimated pseudo-noise free input signal, For an estimated pseudo noise free input signal at a previous time instant, For actually observing input variables polluted by measured noise, model parameters 、 、 Based on known input source signals And Is estimated.
  3. 3. The method of claim 1, wherein the encoder in step 3 employs a XLSTM structure, the XLSTM structure is composed of a plurality of stacked residual network blocks, each integrated with causal convolutional layers, layer normalization units, sLSTM units, and mLSTM units, and the pseudo-noiseless input signal is encoded by the encoder Mapping to a high-dimensional potential representation The calculation mode of (a) is as follows: Wherein, the , In order to make the number of time steps, In order to hide the dimensions of the layer, Representing the encoder.
  4. 4. The method according to claim 1, wherein the pseudo-denoising data in step 3 is generated by optimizing parameters of the input generation model by a subspace recognition algorithm 、 、 The Kalman filter is used for actually observing the input variable And performing filtering processing, wherein the output of the filtering processing is pseudo-denoising data.
  5. 5. The method according to claim 1, wherein the noise enhancement data in step 3 is generated by estimating the measurement noise by a subspace recognition algorithm Prior variance of (2) Generating synthetic noise satisfying Gaussian distribution Synthesizing the noise Superimposed with pseudo-noise-free input signals Obtaining noise enhanced data I.e. 。
  6. 6. The method according to claim 1, wherein the contrast loss function in step 3 Is an error term Regularization term And sparsity term Is used in the field of the digital camera, the expression is as follows: Wherein, the Is a balance factor for balancing regression accuracy with a representation of learning quality; error term The calculation mode of (a) is as follows: Wherein, the For the predicted value output by the multi-batch variable error EIV system, Outputting the system in the original observation data; Regularization term The calculation mode of (a) is as follows: Wherein, the Gao Weiqian representing a pseudo-noise-free input signal is shown, A high-dimensional potential representation representing a positive sample, A high-dimensional potential representation representing a negative sample; 、 、 The calculation mode of (a) is as follows: Wherein, the Representing an encoder; sparsity term The calculation mode of (a) is as follows: Wherein, the As a result of the sparseness factor, Is a temperature super parameter.
  7. 7. The method of claim 1, wherein the multi-lot variable error EIV system is a cascade tank system in a pharmaceutical or chemical manufacturing process; when the system is a cascade tank system, the known input source signal of the raw observation data In order to supply water from reservoir to pump voltage signal of water pump of upper water tank, actual observation input variable polluted by measured noise As a measure of noise pollution, the system outputs Is the water level of the lower water tank.
  8. 8. A multi-batch variable error system identification system based on noise-enhanced contrast learning, the system comprising: A data acquisition module configured to acquire raw observation data of a multi-batch variable error EIV system, the raw observation data including a known input source signal, an actual observation input variable contaminated with measured noise, and a system output; An input generation model construction module configured to construct an input generation model for describing a dynamic generation process of a pseudo noise-free input signal based on the raw observation data; the noise enhancement auxiliary sequence contrast learning NSCL module is configured for identifying a multi-batch variable error system and comprises a characteristic representation learning sub-module, a contrast learning sub-module driven by a denoising device and a prediction regression sub-module; a feature representation learning sub-module comprising an encoder configured to map a pseudo-noise-free input signal into a high-dimensional potential representation; The contrast learning submodule is used for generating pseudo-denoising data as positive samples and noise enhancement data as negative samples, mapping the positive samples, the negative samples and pseudo-noiseless input signals into high-dimensional potential representations respectively through the encoder to construct triplets, enabling the Gao Weiqian of the anchor point representation approaching the positive samples to be represented and far from the high-dimensional potential representations of the negative samples through optimizing contrast loss functions by using Gao Weiqian of the pseudo-noiseless input signals as anchor points; And the prediction regression sub-module is configured to perform linear regression prediction on the Gao Weiqian representation of the optimized pseudo-noiseless input signal to obtain a predicted value output by the multi-batch variable error EIV system.
  9. 9. An electronic device comprising a processor and a memory, the memory storing a computer program, characterized in that the processor implements the method of any of claims 1 to 7 when executing the computer program.
  10. 10. A computer readable storage medium, on which a computer program is stored, which computer program, when being executed by a processor, implements the method according to any of claims 1 to 7.

Description

Multi-batch variable error system identification method based on noise enhancement contrast learning Technical Field The invention belongs to the field of system identification, and particularly relates to a multi-batch variable error system identification method based on noise enhancement contrast learning. Background System identification is a critical area of research in process modeling where input and output data are fundamental elements. Most system identification studies assume that the input data is noiseless and easy to use. However, in real-world industrial processes, measurement data is often subject to noise interference, such as sensor errors and human data acquisition errors. The direct use of these noise data adversely affects the recognition result and reduces the control accuracy. Accordingly, research into variable Error (EIV) systems is attracting increasing attention because it considers the noise-contaminated scene of the measured variable, thereby enabling more robust, accurate system identification. Batch processing is a major production method in the pharmaceutical and chemical industries and is characterized by discrete batch production, each meeting specific requirements. Environmental factors may cause variations in process dynamics from batch to batch. Parameterized identification relies on a priori knowledge of the dynamics or structure of the system, as well as assumptions on the structure or order of the predefined model. However, in a multi-lot process, due to the variation of the operation points or the trajectories, the preset model designed for single lot identification often cannot capture the system dynamics between different lots. For example, when a Linear Parameter Variation (LPV) model describing a nonlinear system is constructed using an Expectation Maximization (EM) algorithm, its performance is highly dependent on the data distribution within the training set. Thus, if the new lot of trajectories change, the pre-trained model may not be effectively generalized. As industrial processes grow in size and complexity, model distortion due to inaccurate preset model assumptions becomes more and more severe. In order to cope with challenges brought by the change of running tracks among different batches, a non-parameter identification algorithm is introduced, and the input-output relation of the system can be effectively learned by utilizing a non-parameter and data driving mode, so that the system is suitable for high-complexity scenes and systems which are difficult to model or describe. The non-parameter identification provides adaptability for a dynamically changing system by extracting information between sequences. However, the existing non-parameter identification method still faces challenges when handling input noise pollution, especially the problems of bias in input noise variance estimation and noise amplification in the data conversion process, which limit the application effect in multi-batch variable error system identification. Disclosure of Invention [ Problem ] The invention aims to solve the technical problem of how to design a noise robust non-parameter identification algorithm so as to overcome the problems of input noise variance estimation deviation and data conversion noise amplification, thereby improving the identification precision and the cross-batch generalization capability of a multi-batch variable error system. [ Technical solution ] In order to solve the above problems, the present invention provides a multi-batch variable error system identification method, system, electronic device and computer readable storage medium based on noise enhancement contrast learning. In a first aspect, the present invention provides a method for identifying a multi-batch variable error system based on noise enhancement contrast learning, including: Step 1, acquiring original observation data of a multi-batch variable error EIV system, wherein the original observation data comprises known input source signals, actual observation input variables polluted by measurement noise and system output; Step 2, an input generation model is built based on the original observation data, and the input generation model is used for describing a dynamic generation process of a pseudo noise-free input signal; Step 3, constructing and applying a noise enhancement auxiliary sequence contrast learning NSCL algorithm frame, wherein the noise enhancement auxiliary sequence contrast learning NSCL algorithm frame comprises a characteristic representation learning module, a denoising device driven contrast learning module and a prediction regression module; The feature representation learning module includes mapping, by an encoder, the pseudo noise-free input signal into a high-dimensional potential representation; The contrast learning module driven by the denoising device comprises a positive sample generation module, a negative sample generation module, a three-tuple construction