Multi-batch variable error system identification method based on noise enhancement contrast learning
By employing a noise-enhanced contrastive learning approach, an NSCL algorithm framework is constructed. High-quality pseudo-denoised data is generated using XLSTM and Kalman filtering techniques, which solves the noise contamination problem in multi-batch variable error systems and achieves higher identification accuracy and stability, making it suitable for complex industrial processes.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIANGNAN UNIV
- Filing Date
- 2026-01-16
- Publication Date
- 2026-05-12
AI Technical Summary
Existing nonparametric identification methods face problems such as bias in the estimation of input noise variance and amplification of data conversion noise when dealing with input noise contamination, resulting in insufficient identification accuracy and cross-batch generalization ability of multi-batch variable error systems.
We adopt a noise-enhanced contrastive learning approach, which constructs an input generation model and a noise-enhanced auxiliary sequence contrastive learning (NSCL) algorithm framework, including feature representation learning, denoiser-driven contrastive learning, and prediction regression modules. We utilize XLSTM structure and Kalman filtering technology to generate high-quality pseudo-denoised data and optimize the contrastive loss function to improve feature learning quality and model stability.
It significantly improves the identification accuracy and stability of multi-batch variable error systems, can adapt to the dynamic changes of different batches and operation points, is suitable for complex industrial processes, especially batch processing in the pharmaceutical and chemical industries, and enhances the model's cross-batch generalization performance.
Smart Images
Figure CN122021776A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of system identification, specifically relating to a method for identifying multi-batch variable error systems based on noise-enhanced contrastive learning. Background Technology
[0002] System identification is a crucial research area in process modeling, where input and output data are fundamental elements. Most system identification studies assume that the input data is noise-free and easy to use. However, in real-world industrial processes, measurement data is frequently affected by noise, such as sensor errors and manual data acquisition errors. Directly using this noisy data can adversely affect identification results and reduce control accuracy. Therefore, research on error-in-variable (EIV) systems is gaining increasing attention because it considers scenarios where measured variables are contaminated by noise, thus enabling more robust and accurate system identification.
[0003] Batch processing is a primary production method in the pharmaceutical and chemical industries, characterized by discrete batch production, where each batch meets specific requirements. Environmental factors can lead to variations in process dynamics between different batches. Parametric identification relies on prior knowledge of the system's dynamic characteristics or structure, as well as assumptions about the predefined model structure or order. However, in multi-batch processes, pre-trained models designed for single-batch identification often fail to capture the system's dynamic characteristics across different batches due to variations in operating points or trajectories. For example, when using the Expectation-Maximization (EM) algorithm to construct a Linear Parametric Variation (LPV) model describing a nonlinear system, its performance is highly dependent on the data distribution within the training set. Therefore, if the trajectory of a new batch changes, the pre-trained model may fail to generalize effectively. As industrial processes scale up and become more complex, model distortion due to inaccurate pre-trained model assumptions becomes increasingly severe.
[0004] To address the challenges posed by variations in operational trajectories across different batches, a nonparametric identification algorithm is introduced. By utilizing a nonparametric, data-driven approach, it can effectively learn the input-output relationships of the system, thereby adapting to highly complex scenarios and systems that are difficult to model or describe. Nonparametric identification provides adaptability to dynamically changing systems by extracting information between sequences. However, existing nonparametric identification methods still face challenges in handling input noise contamination, particularly the bias in input noise variance estimation and noise amplification during data transformation. This limits their application effectiveness in identifying multi-batch variable error systems. Summary of the Invention
[0005] [Technical Issues] The technical problem to be solved by this invention is how to design a noise-robust nonparametric identification algorithm to overcome the problems of input noise variance estimation bias and data conversion noise amplification, thereby improving the identification accuracy and cross-batch generalization ability of multi-batch variable error systems.
[0006] [Technical Solution] To address the above problems, this invention provides a method, system, electronic device, and computer-readable storage medium for identifying multi-batch variable error systems based on noise-enhanced contrastive learning.
[0007] In a first aspect, the present invention provides a method for identifying multi-batch variable error systems based on noise-enhanced contrastive learning, comprising: Step 1: Obtain the raw observation data of the multi-batch variable error EIV system. The raw observation data includes the known input source signal, the actual observed input variables contaminated by measurement noise, and the system output. Step 2: Construct an input generation model based on the original observation data. The input generation model is used to describe the dynamic generation process of pseudo-noise-free input signals. Step 3: Construct and apply the Noise Augmentation Auxiliary Sequence Contrast Learning (NSCL) algorithm framework, which includes a feature representation learning module, a denoiser-driven contrastive learning module, and a prediction and regression module. The feature representation learning module includes mapping pseudo-noise-free input signals into high-dimensional latent representations through an encoder; The denoiser-driven contrastive learning module includes generating pseudo-denoised data as positive samples and generating noise-enhanced data as negative samples; mapping positive samples, negative samples, and pseudo-noise-free input signals to high-dimensional latent representations through an encoder to construct triples; using the high-dimensional latent representation of the pseudo-noise-free input signal as an anchor point, optimizing the contrastive loss function to make the anchor point representation approach the high-dimensional latent representation of the positive samples and move away from the high-dimensional latent representation of the negative samples. The prediction regression module includes performing linear regression prediction on the high-dimensional latent representation of the optimized pseudo-noise-free input signal to obtain the predicted value output by the multi-batch variable error EIV system.
[0008] Optionally, the input generation model in step 2 is described by the following state-space equations:
[0009]
[0010] in, Given the known input source signal from the previous moment. For process noise, To measure noise, For the estimated pseudo-noise-free input signal, The pseudo-noise-free input signal estimated at the previous moment. The actual observed input variables for the measured noise pollution, and the model parameters. , , Based on known input source signal and The identification method is used for estimation.
[0011] Optionally, the encoder in step 3 adopts an XLSTM structure, which consists of multiple stacked residual network blocks. Each residual network block integrates a causal convolutional layer, a layer normalization unit, an sLSTM unit, and an mLSTM unit. The encoder transforms the pseudo-noise-free input signal into a signal with a noise-free input. Mapping to a high-dimensional latent representation The calculation method is as follows:
[0012] in, , For time steps, For the hidden layer dimension, This indicates the encoder.
[0013] Optionally, the process of generating pseudo-denoised data in step 3 is as follows: first, a subspace identification algorithm is used to optimize the parameters of the input generation model. , , Then, a Kalman filter is used to evaluate the actual observed input variables. The filtering process is performed, and the output of the filtering process is the pseudo-denoised data.
[0014] Optionally, the noise enhancement data generation process in step 3 is as follows: estimating the measurement noise using a subspace recognition algorithm. Prior variance Generate synthesized noise that satisfies a Gaussian distribution. The synthesized noise Superimposed with pseudo-noise-free input signal Obtain noise enhancement data ,Right now .
[0015] Optionally, the contrastive loss function in step 3 Error term Regularization term With sparsity terms The weighted sum, its expression is:
[0016] in, This is a balancing factor used to balance regression accuracy and representation learning quality; Error term The calculation method is as follows:
[0017] in, The predicted values output by the multi-batch variable error EIV system are... This is the system output from the original observation data; Regularization term The calculation method is as follows:
[0018] in, A high-dimensional latent representation of a pseudo-noise-free input signal. High-dimensional latent representation of positive samples High-dimensional latent representation of negative samples; , , The calculation method is as follows:
[0019] in, Indicates encoder; sparsity terms The calculation method is as follows:
[0020] in, As a sparsity factor, This refers to temperature hyperparameters.
[0021] Optionally, the multi-batch variable error EIV system is a cascaded water tank system in pharmaceutical or chemical production processes; When the system is a cascaded water tank system, the known input source signal of the original observation data The pump voltage signal of the water pump that transports water from the reservoir to the upper water tank is an actual observed input variable subject to measurement noise pollution. For measurements affected by noise, the system output... This refers to the water level in the lower water tank.
[0022] Secondly, the present invention provides a multi-batch variable error system identification system based on noise-enhanced contrastive learning, comprising: The data acquisition module is configured to acquire raw observation data of the multi-batch variable error EIV system, the raw observation data including known input source signals, actual observed input variables contaminated by measurement noise, and system output; The input generation model building module is configured to build an input generation model based on the original observation data. The input generation model is used to describe the dynamic generation process of the pseudo-noise-free input signal. The Noise Enhancement Auxiliary Sequence Contrast Learning (NSCL) module is configured to identify multi-batch variable error systems. The NSCL module includes: a feature representation learning submodule, a denoiser-driven contrastive learning submodule, and a prediction regression submodule. The feature representation learning submodule, including an encoder, is configured to map pseudo-noise-free input signals into high-dimensional latent representations; The denoiser-driven contrastive learning submodule is configured to generate pseudo-denoised data as positive samples and generate noise-enhanced data as negative samples. The encoder maps the positive samples, negative samples, and pseudo-noise-free input signals into high-dimensional latent representations to construct triples. Using the high-dimensional latent representation of the pseudo-noise-free input signal as an anchor point, the contrastive loss function is optimized to make the anchor point representation approach the high-dimensional latent representation of the positive samples and move away from the high-dimensional latent representation of the negative samples. The predictive regression submodule is configured to perform linear regression prediction on the high-dimensional latent representation of the optimized pseudo-noise-free input signal to obtain the predicted value output by the multi-batch variable error EIV system.
[0023] Thirdly, the present invention provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method for identifying multi-batch variable error systems based on noise-enhanced contrastive learning.
[0024] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-described method for identifying multi-batch variable error systems based on noise-enhanced contrastive learning.
[0025] [Beneficial Effects] (1) In step 3 of this invention, a feature representation learning module is used. This module uses an XLSTM structure as the encoder. The XLSTM structure consists of multiple residual blocks and integrates causal convolution, layer normalization, sLSTM units and mLSTM units to effectively capture long-term temporal dependencies. The residual blocks integrate causal convolutional layers and layer normalization units, which maintain input stability while ensuring temporal causality. A denoising-driven contrastive learning module is used. This module generates high-quality pseudo-denoised data as positive samples by combining a state-space model with subspace recognition and Kalman filtering techniques. This provides a cleaner and more realistic noise-free learning target for contrastive learning, thereby improving the quality of feature learning. By jointly optimizing the error term, regularization term and sparsity term, the model achieves a balance between regression accuracy and representation learning quality, improving the overall training stability and convergence efficiency, and avoiding overfitting or underfitting problems caused by single-objective optimization.
[0026] (2) Step 3 of the present invention introduces the noise-enhanced auxiliary sequence contrastive learning NSCL algorithm framework, constructs a "pseudo-denoising-original-noise-enhancing" triplet and performs contrastive learning, so that the model actively learns noise-invariant features in the latent representation space, significantly reduces the dependence on accurate noise variance prior information, and improves the identification accuracy and stability of the system in a strong noise environment.
[0027] This invention has been experimentally verified in a cascaded water tank system, demonstrating its good applicability in typical industrial processes and its potential for widespread application to other complex industrial system identification scenarios involving multiple batches and severe noise pollution. The method of this invention does not rely on a fixed parametric model structure; it extracts sequence features through a data-driven approach, enabling it to adapt to dynamic changes across different batches and operating points. It is particularly suitable for the variable operating environments of batch processes in industries such as pharmaceuticals and chemicals, significantly improving the model's cross-batch generalization performance. Attached Figure Description
[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0029] Figure 1 The input generation model structure diagram provided by this invention.
[0030] Figure 2 The framework diagram of the noise-enhanced auxiliary sequence contrastive learning (NSCL) algorithm provided by this invention is shown.
[0031] Figure 3A diagram of a cascaded water tank control system provided for this invention.
[0032] Figure 4 The timing curves of the input and output data provided by this invention.
[0033] Figure 5 Scatter plot of prediction results from different algorithms provided in this invention.
[0034] Figure 6 A comparison chart of the prediction residual sequences of different algorithms provided in this invention.
[0035] Figure 7 The image shows the prediction results for the test set provided by this invention. Detailed Implementation
[0036] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0037] Example 1 This embodiment uses a cascaded water tank control system (see...). Figure 3 As a multi-batch variable error EIV system, this paper provides a method for identifying multi-batch variable error systems based on noise-enhanced contrastive learning, including the following steps: Step 1: Obtain raw observation data for a complete production batch from the cascaded water tank control system. The raw observation data includes known input source signals, actual observed input variables contaminated by measurement noise, and system output. Known input source signals... The pump voltage signal of the water pump that transports water from the reservoir to the upper water tank is an actual observed input variable subject to measurement noise pollution. For measurements affected by noise, the system output... This represents the water level in the lower tank. Data was sampled at 1-minute intervals, for a total of 1000 time steps (see...). Figure 4 ).right and Outlier removal and normalization are performed to map the values to the [0,1] interval.
[0038] Step 2: Construct an input generation model based on the original observation data. The input generation model describes the dynamic generation process of the pseudo-noise-free input signal; Input Generation Model (see...) Figure 1 It is described by the following state-space equations:
[0039]
[0040] in, Given the known input source signal from the previous moment. For process noise, To measure noise, For the estimated pseudo-noise-free input signal, The pseudo-noise-free input signal estimated at the previous moment. The actual observed input variables for the measured noise pollution, and the model parameters. , , Based on known input source signal and The identification method is used for estimation.
[0041] Step 3: Construct and apply the Noise-Enhanced Auxiliary Sequence Contrast Learning (NSCL) algorithm framework (see...) Figure 2 The Noise Enhancement Auxiliary Sequence Contrast Learning (NSCL) algorithm framework includes a feature representation learning module, a denoiser-driven contrastive learning module, and a prediction and regression module. The feature representation learning module specifically maps pseudo-noise-free input signals into high-dimensional latent representations through an encoder; The encoder employs an XLSTM structure, which consists of three stacked residual network blocks. Each residual network block integrates a causal convolutional layer (kernel size = 3), a layer normalization unit, an sLSTM unit (64 hidden units), and an mLSTM unit (64 hidden units). The encoder transforms the pseudo-noise-free input signal... Mapping to a high-dimensional latent representation The calculation method is as follows:
[0042] in, , For the hidden layer dimension, This indicates the encoder.
[0043] The denoiser-driven contrastive learning module specifically generates pseudo-denoised data as positive samples and generates noise-enhanced data as negative samples. The process of generating pseudo-denoised data is as follows: First, a subspace recognition algorithm is used to optimize the parameters of the input generation model. , , Then, a Kalman filter is used to evaluate the actual observed input variables. The filtering process is performed, and the output of the filtering process is the pseudo-denoised data. ; The process of generating noise-enhanced data is as follows: the measured noise is estimated using a subspace identification algorithm. Prior variance Generate synthesized noise that satisfies a Gaussian distribution. ( ), synthesized noise Superimposed with pseudo-noise-free input signal Obtain noise enhancement data ,Right now .
[0044] The encoder maps positive samples, negative samples, and pseudo-noise-free input signals into high-dimensional latent representations, thereby constructing triplet groups. Extracted using the same XLSTM encoder , High-dimensional potential representation , , Specifically:
[0045] in, Indicates encoder; And construct a set of triplet representations:
[0046] pseudo-noiseless input signal High-dimensional potential representation Using anchor points, the contrastive loss function is optimized to make the anchor point representations approximate positive samples. High-dimensional potential representation and far from negative samples High-dimensional potential representation This improves the robustness and discriminativeness of the representation; Contrast loss function Error term Regularization term With sparsity terms The weighted sum, its expression is:
[0047] in, Here, a balancing factor is used to balance regression accuracy and representation learning quality. ; Error term The calculation method is as follows:
[0048] in, This is the predicted value output by the cascaded water tank control system. This is the system output from the original observation data; Regularization term The calculation method is as follows:
[0049] in, A high-dimensional latent representation of a pseudo-noise-free input signal. High-dimensional latent representation of positive samples High-dimensional latent representation of negative samples; sparsity terms The calculation method is as follows:
[0050] in, As the sparsity factor, here ; For temperature hyperparameters, here .
[0051] Using the Adam optimizer (with a learning rate of 100%) Minimize total loss The training process is iterated 800 times until convergence.
[0052] The prediction and regression module specifically works by taking the optimized pseudo-noise-free input signal. High-dimensional potential representation Input a single-layer linear regression network (output dimension 1) to obtain the final concentration prediction value. .
[0053] This embodiment compares different algorithms, including LSTM, XLSTM, a baseline model with contaminated input variables, and a scatter plot of the prediction results of the NSCL method proposed in this invention (see...). Figure 5 In the graph, the horizontal axis represents the true value, and the vertical axis represents the predicted value. The degree of deviation of the scatter points from the diagonal directly reflects the prediction error—the closer the distribution is to the diagonal, the higher the prediction accuracy. It can be observed that, under the same training conditions, the predicted points of the NSCL method of this invention are significantly more concentrated near the diagonal, indicating that it has higher prediction accuracy.
[0054] This embodiment further compares the prediction residual sequence diagrams of each algorithm (see...). Figure 6 The residuals of the NSCL algorithm are significantly smaller and fluctuate more stably around zero, indicating that the method of the present invention not only has high accuracy but also better stability and noise tolerance.
[0055] Example 2 In this embodiment, a new batch of cascaded water tank system data is selected as the application object to further verify the universality of the method of the present invention.
[0056] Step 1: Build a cascaded water tank control system in a laboratory environment. The input source signal is known. The pump voltage signal of the water pump that transports water from the reservoir to the upper water tank is an actual observed input variable subject to measurement noise pollution. For measurements affected by noise, the system output... This represents the water level in the lower tank. The variance is actively injected into the input measurement. Additive white Gaussian noise was used to simulate severe noise pollution. Batch data were collected at two different operating points, one for training and one for testing.
[0057] Step 2: Similar to Step 2 in Example 1, based on training batches: and A state-space model was established using the subspace identification method, and the noise characteristics were initially estimated.
[0058] Step 3: This is basically the same as Step 3 in Example 1, with the key difference being: (1) The training set and the test set are two independent batches from different operation points, which are specifically used to test the generalization of the model.
[0059] (2) The number of hidden units in the XLSTM encoder is set to 32, the number of training iterations is 600, and the loss function balance factor is... .
[0060] Based on the test set prediction results of this embodiment (i.e., the test set prediction graph, see...), see... Figure 7 As can be seen, the method proposed in this invention can accurately predict the numerical changes of test batches, and its predicted trajectory is highly consistent with the actual value. The results of this embodiment demonstrate that the method of this invention can effectively overcome the challenges posed by input noise and operating point variations, and is applicable to various complex industrial scenarios.
[0061] Example 3 This embodiment provides a multi-batch variable error system identification system based on noise-enhanced contrastive learning, including: The data acquisition module is configured to acquire raw observation data from the multi-batch variable error EIV system. The raw observation data includes known input source signals, actual observed input variables contaminated by measurement noise, and system output. The input generation model building module is configured to build an input generation model based on the original observation data. The input generation model is used to describe the dynamic generation process of pseudo-noise-free input signals. The Noise Enhancement Auxiliary Sequence Contrast Learning (NSCL) module is configured to identify multi-batch variable error systems. The NSCL module includes: a feature representation learning submodule, a denoiser-driven contrastive learning submodule, and a prediction regression submodule. The feature representation learning submodule, including an encoder, is configured to map pseudo-noise-free input signals into high-dimensional latent representations; The denoiser-driven contrastive learning submodule is configured to generate pseudo-denoised data as positive samples and noise-enhanced data as negative samples. The encoder maps the positive samples, negative samples, and pseudo-noise-free input signals into high-dimensional latent representations, thereby constructing triples. Using the high-dimensional latent representation of the pseudo-noise-free input signal as the anchor point, the contrastive loss function is optimized to make the anchor point representation approach the high-dimensional latent representation of the positive samples and move away from the high-dimensional latent representation of the negative samples. The predictive regression submodule is configured to perform linear regression prediction on the high-dimensional latent representation of the optimized pseudo-noise-free input signal to obtain the predicted value output by the multi-batch variable error EIV system.
[0062] Example 4 This embodiment provides an electronic device, including: The processor and memory, the memory storing the computer program, the processor executing the computer program to implement the above-mentioned multi-batch variable error system identification method based on noise-enhanced contrastive learning.
[0063] Example 5 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for identifying multi-batch variable error systems based on noise-enhanced contrastive learning.
[0064] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.
Claims
1. A method for identifying multi-batch variable error systems based on noise-enhanced contrastive learning, characterized in that, include: Step 1: Obtain the raw observation data of the multi-batch variable error EIV system. The raw observation data includes the known input source signal, the actual observed input variables contaminated by measurement noise, and the system output. Step 2: Construct an input generation model based on the original observation data. The input generation model is used to describe the dynamic generation process of pseudo-noise-free input signals. Step 3: Construct and apply the Noise Augmentation Auxiliary Sequence Contrast Learning (NSCL) algorithm framework, which includes a feature representation learning module, a denoiser-driven contrastive learning module, and a prediction and regression module. The feature representation learning module includes mapping pseudo-noise-free input signals into high-dimensional latent representations through an encoder; The denoiser-driven contrastive learning module includes generating pseudo-denoised data as positive samples and generating noise-enhanced data as negative samples; mapping positive samples, negative samples, and pseudo-noise-free input signals to high-dimensional latent representations through an encoder to construct triples; using the high-dimensional latent representation of the pseudo-noise-free input signal as an anchor point, optimizing the contrastive loss function to make the anchor point representation approach the high-dimensional latent representation of the positive samples and move away from the high-dimensional latent representation of the negative samples. The prediction regression module includes performing linear regression prediction on the high-dimensional latent representation of the optimized pseudo-noise-free input signal to obtain the predicted value output by the multi-batch variable error EIV system.
2. The method according to claim 1, characterized in that, The input generation model in step 2 is described by the following state-space equations: in, Given the known input source signal from the previous moment. For process noise, To measure noise, For the estimated pseudo-noise-free input signal, The pseudo-noise-free input signal estimated at the previous moment. The actual observed input variables for the measured noise pollution, and the model parameters. , , Based on known input source signal and The identification method is used for estimation.
3. The method according to claim 1, characterized in that, The encoder in step 3 employs an XLSTM structure, which consists of multiple stacked residual network blocks. Each residual network block integrates a causal convolutional layer, a layer normalization unit, an sLSTM unit, and an mLSTM unit. The encoder transforms the pseudo-noise-free input signal... Mapping to a high-dimensional latent representation The calculation method is as follows: in, , For time steps, For the hidden layer dimension, This indicates the encoder.
4. The method according to claim 1, characterized in that, The process of generating pseudo-denoised data in step 3 is as follows: First, a subspace identification algorithm is used to optimize the parameters of the input generation model. , , Then, a Kalman filter is used to evaluate the actual observed input variables. The filtering process is performed, and the output of the filtering process is the pseudo-denoised data.
5. The method according to claim 1, characterized in that, The noise enhancement data generation process in step 3 is as follows: the measurement noise is estimated using a subspace recognition algorithm. Prior variance Generate synthesized noise that satisfies a Gaussian distribution. The synthesized noise Superimposed with pseudo-noise-free input signal Obtain noise enhancement data ,Right now .
6. The method according to claim 1, characterized in that, The contrast loss function in step 3 For error terms Regularization term With sparsity terms The weighted sum, its expression is: in, This is a balancing factor used to balance regression accuracy and representation learning quality; Error term The calculation method is as follows: in, The predicted values output by the multi-batch variable error EIV system are... This is the system output from the original observation data; Regularization term The calculation method is as follows: in, A high-dimensional latent representation of a pseudo-noise-free input signal. High-dimensional latent representation of positive samples High-dimensional latent representation of negative samples; , , The calculation method is as follows: in, Indicates encoder; sparsity terms The calculation method is as follows: in, As a sparsity factor, This refers to temperature hyperparameters.
7. The method according to claim 1, characterized in that, The multi-batch variable error EIV system is a cascaded water tank system in pharmaceutical or chemical production processes; When the system is a cascaded water tank system, the known input source signal of the original observation data The pump voltage signal of the water pump that transports water from the reservoir to the upper water tank is an actual observed input variable subject to measurement noise pollution. For measurements affected by noise, the system output... This refers to the water level in the lower water tank.
8. A multi-batch variable error system identification system based on noise-enhanced contrastive learning, characterized in that, The system includes: The data acquisition module is configured to acquire raw observation data of the multi-batch variable error EIV system, the raw observation data including known input source signals, actual observed input variables contaminated by measurement noise, and system output; The input generation model building module is configured to build an input generation model based on the original observation data. The input generation model is used to describe the dynamic generation process of the pseudo-noise-free input signal. The Noise Enhancement Auxiliary Sequence Contrast Learning (NSCL) module is configured to identify multi-batch variable error systems. The NSCL module includes: a feature representation learning submodule, a denoiser-driven contrastive learning submodule, and a prediction regression submodule. The feature representation learning submodule, including an encoder, is configured to map pseudo-noise-free input signals into high-dimensional latent representations; The denoiser-driven contrastive learning submodule is configured to generate pseudo-denoised data as positive samples and generate noise-enhanced data as negative samples. The encoder maps the positive samples, negative samples, and pseudo-noise-free input signals into high-dimensional latent representations to construct triples. Using the high-dimensional latent representation of the pseudo-noise-free input signal as an anchor point, the contrastive loss function is optimized to make the anchor point representation approach the high-dimensional latent representation of the positive samples and move away from the high-dimensional latent representation of the negative samples. The predictive regression submodule is configured to perform linear regression prediction on the high-dimensional latent representation of the optimized pseudo-noise-free input signal to obtain the predicted value output by the multi-batch variable error EIV system.
9. An electronic device comprising a processor and a memory, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.