A method for intelligently removing SPR noise data based on a generative adversarial network

Abstract

The application discloses a SPR noise data intelligent removal method based on a generative adversarial network, and comprises the following steps: S1, collecting a response signal for preprocessing and outputting noise data; S2, inputting the noise data into a self-encoder to obtain sparse latent variables; S3, weighting and combining the sparse latent variable features to output spectrum enhancement features; S4, inputting the spectrum enhancement features into a diffusion model to output an SPR signal sequence; S5, inputting the SPR signal sequence into a generative adversarial network to output a pseudo-truth enhanced SPR signal; S6, inputting the pseudo-truth enhanced SPR signal into a decomposition and reconstruction module to obtain a reconstructed SPR signal; S7, comparing the reconstructed SPR signal with an original labeled signal to output optimized model parameters; and S8, performing SPR detection on the optimized model parameters to output an intelligent denoising SPR response curve. The application realizes intelligent removal of SPR noise data.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and biosensor signal technology, and in particular to an intelligent method for removing SPR noise data based on generative adversarial networks. Background Technology

[0002] In the current field of SPR (surface plasmon resonance) detection, as a highly sensitive label-free biosensing technology, SPR is widely used in many key areas such as life sciences, medical diagnostics, environmental monitoring, and food safety. Traditional SPR detection procedures mainly rely on optical systems to acquire and analyze the intensity of reflected signals, determining the binding behavior between the sample and a surface-fixed probe by detecting the response curve generated by the change in reflected light intensity with refractive index. However, in practical applications, SPR signals are highly susceptible to periodic background interference, local abrupt noise, and non-ideal responses of the sensor system, resulting in incomplete response curve structure, waveform distortion, and unstable noise intensity. This interference not only weakens the readability of the SPR response but also seriously affects the accuracy and stability of subsequent analytical models.

[0003] Current research has attempted to introduce convolutional neural networks, autoencoders, filter banks, or Fourier analysis methods to preprocess and denoise raw SPR signals. While these methods can alleviate signal noise problems to some extent, they have several major shortcomings: First, most methods lack structural modeling for periodic interference and high-frequency abrupt signal changes, making it difficult to effectively separate the main response components; second, traditional denoising methods often rely on manually constructed rules or shallow network structures, resulting in poor generalization ability when faced with complex nonlinear noise distributions; third, current model training mostly lacks a fusion optimization mechanism between tasks, with reconstruction and classification, generation and discrimination processes being isolated from each other, thus limiting the overall improvement in detection performance.

[0004] Therefore, how to provide an intelligent method for removing SPR noise data based on generative adversarial networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent SPR noise removal method based on generative adversarial networks (GANs). This method integrates an autoencoder, a spectrum enhancement mechanism, a diffusion model, and a residual reconstruction structure to accurately separate and effectively suppress noise components in the original SPR signal. By constructing sparse feature representations and adversarial generative mechanisms, this method can perform targeted modeling and reconstruction of complex noise structures such as periodic interference and high-frequency abrupt changes, improving the structural integrity and response realism of the signal. After point-by-point comparison and joint optimization training, the final model can be stably deployed in actual SPR detection processes, achieving high-precision and robust intelligent denoising of SPR signals.

[0006] According to an embodiment of the present invention, an intelligent method for removing SPR noise data based on generative adversarial networks includes the following steps:

[0007] S1. Acquire the original surface plasmon resonance sensing response signal and perform normalization preprocessing. Use a fixed-length sliding window to divide the sensing response signal into multiple sub-sequence samples, construct a noise data training set containing temporal structure, and output labeled windowed SPR noise sub-sequence data.

[0008] S2. Input the windowed SPR noise subsequence data into the sparse coding variational autoencoder for encoding and compression processing. Extract latent structural features through sparsity constraints and variational regularization terms to obtain the sparse latent variable representation after preliminary denoising.

[0009] S3. Perform a short-time Fourier transform on the sparse latent variable representation to extract local spectral features within the time window, and use an adaptive Fourier feature fusion module to weight and combine different frequency components to output fused multi-scale spectral enhancement features.

[0010] S4. Input the fused multi-scale spectral enhancement features into the deconvolution diffusion model to simulate the reverse diffusion process of noise perturbation. Generate target signal candidates layer by layer through multiple reverse denoising steps, and output the preliminarily restored SPR signal sequence.

[0011] S5. Input the initially restored SPR signal sequence into the generator module, and at the same time input the manually labeled real SPR signal into the discriminator module. Optimize the generator output quality through the adversarial training mechanism of the generative adversarial network and output an SPR signal with enhanced authenticity.

[0012] S6. Input the pseudo-authenticity enhanced SPR signal into the recursive residual decomposition and reconstruction module to separate the main response component, periodic interference component and high-frequency mutation component, and then fuse them after structural reconstruction to obtain a structurally complete reconstructed SPR signal.

[0013] S7. Compare the reconstructed SPR signal with the original labeled real signal point by point, calculate the structural similarity index and gradient fidelity loss function, and jointly perform reverse update training on the entire generative adversarial network, sparse encoder and residual decomposition network to output the final optimized model parameters.

[0014] S8. Deploy the final optimized model parameters into the SPR detection process, input the real-time acquired raw SPR signal, and output the intelligently denoised SPR response curve.

[0015] Optionally, S1 specifically includes:

[0016] S11. The sample under test is continuously monitored by a surface plasmon resonance sensor to obtain the original response signal sequence that changes over time, and the response amplitude data at equal time intervals are recorded at a fixed sampling frequency to form a complete time series dataset.

[0017] S12. Standardize the original response signal sequence by subtracting the average of all values ​​in the sequence from each response value in the signal, and then dividing the result by the standard deviation of the values ​​to obtain a standardized signal sequence with a mean of zero and consistent fluctuation scale, so as to eliminate the bias caused by the difference in dimensions.

[0018] S13. The standardized signal sequence is divided into multiple fixed-length subsequence samples using the sliding window method. Each subsequence contains signal values ​​at consecutive time points. The sliding window moves gradually on the sequence with a set step size to form the original labeled real signal with partial overlap, which is used to preserve the time dependence and local features of the original signal.

[0019] S14. For each original labeled real signal, based on the physical conditions, interference type and labeling basis provided in the experimental design, perform automated label allocation, determine whether it contains noise mutation points, periodic interference fluctuations or stable signal states, and generate training sample pairs with semantic attributes.

[0020] S15. Organize all training sample pairs and their corresponding labels into a training dataset with a uniform structure, ensuring that the sample lengths are consistent and the label types are distinguishable, so as to meet the requirements of neural network training for input format consistency and label supervision. Divide the samples according to the preset batch size, with each batch containing several subsequence samples and label information. Reconstruct and pad the batch samples according to the input dimensions required by the network to generate labeled windowed SPR noise subsequence data.

[0021] Optionally, S2 specifically includes:

[0022] S21. The labeled windowed SPR noise subsequence data is evenly divided according to the preset spatial size to obtain image sub-blocks of the same size. Each sub-block is arranged in row priority order for independent processing and to preserve the spatial local structure of the image.

[0023] S22. Input each image sub-block into the feature extraction network of the shared structure, use local perception convolution operation to obtain the representation vector of the sub-block, add the representation vector to the embedding vector corresponding to the two-dimensional position in the original image to preserve spatial position information and form enhanced sub-block features.

[0024] S23. Combine all the enhanced sub-block features in sequence into a sequence structure, input it into the multi-head self-attention module to perform long-distance modeling between features, generate a global feature representation with context-dependent relationships, and maintain the stability of gradient flow and fuse with the original features through residual connection and normalization operation.

[0025] S24. Input the global feature representation into the feedforward neural network module, and perform linear mapping, nonlinear transformation and dimension reduction processing in sequence. Then, perform residual connection and normalization operations again to improve the representation capability and obtain the final sequence features used for encoding.

[0026] S25. Input the sequence features into the sparse coding variational autoencoder module, obtain the parameter representation of the latent space through the encoder function, use the reparameter technique for sampling, and introduce sparse constraints and prior regularization terms to construct a regularization function. Input the latent variables into the decoder, reconstruct the total sparse coding loss, and minimize the total sparse coding loss function to obtain the sparse latent variable representation after preliminary denoising.

[0027] Optionally, S3 specifically includes:

[0028] S31. Based on the block structure of the original image in the row and column directions, the sparse latent variable representation after preliminary denoising is restored according to the original sub-block order and sequentially filled into the two-dimensional matrix to generate a two-dimensional feature map consistent with the original image block structure. The size of the two-dimensional matrix is ​​consistent with the number of image blocks, and the channel dimension is kept as the feature dimension retained after sparse reconstruction.

[0029] S32. Perform compression operation on the two-dimensional feature map in the channel dimension. Use a linear mapping module with trainable parameters to reduce the dimension of the feature channel at each spatial location, compressing the high-dimensional features into a more compact low-dimensional representation to reduce the subsequent computation and enhance the expression of key information, while keeping the original spatial structure unchanged.

[0030] S33. Perform context modeling operations on the low-dimensional representation after channel compression, construct an information interaction mechanism between positions, use a global two-dimensional attention structure to weightedly fuse the features of each spatial position, construct the degree of dependence of each position on all other positions, and use a similarity scoring function to calculate the degree of response between positions to obtain the attention-enhanced two-dimensional feature map.

[0031] S34. Input the attention-enhanced 2D feature map into the channel mapping module, perform dimensionality upscaling on the feature channels at each location, and restore the compressed dimension to the same fused multi-scale spectral enhancement feature as the initial feature dimension through a learnable linear transformation, ensuring that subsequent modules can seamlessly connect and use the spatial features with full expressive power.

[0032] Optionally, S4 specifically includes:

[0033] S41. Arrange the multi-scale spectral enhancement features in the order of image sub-blocks and input them into the discriminant structure for preliminary analysis. Use the feature enhancement module composed of linear transformation and nonlinear activation to further process the feature vector at each position in the sequence and extract discriminative expressions to assist in the identification of target information in downstream classification or generation tasks.

[0034] S42. Perform order-preserving aggregation operation on discriminative representations. By setting a global query vector and constructing a weight distribution based on the correlation between features at each location, information from all locations is fused into a unified vector representation, capturing the overall global semantic features of the image.

[0035] S43. Input the global semantic features into the deconvolutional diffusion model as the initial state vector to participate in the progressively iterative reverse generation process. At each diffusion step, perform noise repair and image prior guidance on the current state to obtain the diffusion output result.

[0036] S44. Map the diffusion output back to the category space, pass through the linear projection module to obtain the final prediction output, and compare it with the label to obtain the generation error. Calculate the difference through the cross-entropy loss function for supervised training.

[0037] S45. The generation error and the total loss of sparse coding are weighted and combined to construct an overall training optimization objective, so that the entire network can maintain a unified learning strategy between compressed reconstruction and objective-oriented generation, and achieve effective reverse modeling and robust prediction.

[0038] Optionally, S5 specifically includes:

[0039] S51. The image result generated in the last round of the overall training optimization target is used as a pseudo sample and input into the discriminator network. At the same time, an auxiliary discriminator network with learnable parameters is constructed. The auxiliary discriminator network receives the generated result as input and outputs a scalar value to indicate whether the input sample comes from the real data distribution, which is used to measure the closeness between the model-generated image and the real image.

[0040] S52. Select corresponding samples from the real sample set and input them into the discriminator network along with pseudo samples. During the training phase, drive the discriminator to learn to distinguish the distribution differences between generated data and real data by marking whether the input is real. At the same time, apply a target to the generator so that the output is judged as real by the discriminator to the greatest extent.

[0041] S53. Construct an adversarial target in the discriminator module and establish independent loss functions for the generator and discriminator respectively. The generator aims to maximize the probability that a fake sample is considered a real sample by the discriminator. The discriminator aims to maximize the confidence in judging real samples as real and fake samples as fake. The two are iteratively trained in an adversarial optimization manner to form a game mechanism.

[0042] S54. Based on the game mechanism, the parameters of the independent loss functions are updated separately. Backpropagation and gradient descent optimization methods are used to adjust the weight parameters of the generator network and the discriminator network separately to obtain the classification loss. The generator relies on the error of the pseudo sample being judged as real to update its own strategy, so as to gradually improve the semantic and structural realism of the generated content. At the same time, the discriminator adjusts the discrimination boundary according to the discrimination bias to improve the discrimination ability.

[0043] S55. Introduce the generator loss term into the overall training optimization objective, and combine it with the sparse coding total loss and classification loss in a weighted combination to form the pseudo-realism enhanced SPR signal. By setting the weight factor of the generation loss, adjust the proportion of its influence on the overall optimization objective, realize the dynamic balance between generation, reconstruction and classification, and improve the overall performance of the model in image understanding and generation tasks.

[0044] Optionally, S6 specifically includes:

[0045] S61. The SPR signal processed by the pseudo-realism enhancement module is used as input and sent to the structured recursive residual decomposition reconstruction module. The recursive residual decomposition module performs multi-scale component deconstruction operation based on the overall response structure of the signal. The initial input is the original enhanced signal form, maintaining timing consistency and channel matching.

[0046] S62. The recursive residual decomposition and reconstruction module sequentially unfolds multi-level feature separation paths, performs cascaded residual updates on the input signal, extracts the main components, and obtains the separated signal components, including stable response components, interference fluctuation components with periodic characteristics, and high-frequency components with local mutation characteristics. The module outputs feature representations in different frequency domains and structural domains at each layer to capture the inherent multi-structure information of the signal.

[0047] S63. Construct independent structural reconstruction paths for each type of separated signal component. While maintaining the original characteristic distribution, introduce a deep residual iteration mechanism. Enhance the morphology of the main response part through multi-layer nonlinear mapping, calibrate the waveform of periodic interference, and smooth the edges and repair the details of high-frequency mutations. Each sub-path completes the target-oriented structural completion within its own dimension.

[0048] S64. The signal components reconstructed from all sub-paths are uniformly fused in the time dimension and amplitude space. They are then merged and aligned according to the initial signal arrangement. The weight adjustment mechanism controls the proportion of each component in the final output to obtain a structurally complete reconstructed SPR signal that is structurally continuous, has a consistent response, and retains key signal features. This ensures that the signal quality, structural integrity, and response accuracy meet the input standards under the multi-task fusion framework, supporting the steady-state inference capability and terminal performance of the overall system.

[0049] Optionally, S7 specifically includes:

[0050] S71. Align the reconstructed SPR signal with the original labeled real signal on the time axis point by point to ensure that the two are consistent in the sampling frequency, start time and signal length dimension. After the alignment is completed, construct a one-to-one corresponding point-to-point matching structure.

[0051] S72. In the point-to-point matching structure, for each sliding time window, the local average value of the real signal and the reconstructed signal within the sliding time window is extracted to characterize the energy center trend of the current region. The local amplitude variance of the two types of signals within the window is calculated to reflect the stability of the signal intensity distribution. At the same time, the covariance information of the two sets of signals within the window range is calculated to measure the degree of synchronous change between the waveforms. The local mean, variance and covariance are then combined to obtain an index score for measuring the structural similarity of the signals. The index score is calculated segment by segment within the entire window range to form the structural similarity estimate of the entire signal.

[0052] S73. Based on the sequence form of the real signal and the reconstructed signal in the time domain, the amplitude difference between each sampling point and the previous sampling point is calculated in turn as the approximate gradient of the current point. The absolute difference between the corresponding gradient values ​​of the real signal and the reconstructed signal at the same time point is calculated to measure the consistency of the local change trend. The errors of all sampling points are accumulated and averaged to obtain an overall gradient fidelity error used to describe whether high-frequency details are preserved. The error is used in training to constrain the model's ability to recover abrupt changes.

[0053] S74. The structural similarity index score obtained above and the overall gradient fidelity error are included as constraints in the training optimization objective. The adversarial loss included in the generator path and the main loss function used for data reconstruction in the sparse encoder are fused together to construct a multi-loss jointly driven training objective. In the training objective, adjustable weight coefficients are introduced for the error terms of different tasks to control the degree of influence of each part of the loss on the overall network parameter update, thereby achieving dynamic balance adjustment of the task objectives of different modules.

[0054] S75. Based on the training objective driven by multiple losses, all error information is propagated along the reverse path to the corresponding modules, including the generator structure, sparse encoder structure and residual reconstruction structure. The gradient propagation and weight update mechanism in deep networks are used to synchronously iteratively update the parameters of each module. During the training process, all modules are collaboratively adjusted in the same optimization cycle until the overall network converges under multiple error indicators and outputs the final optimized model parameters.

[0055] Optionally, S8 specifically includes:

[0056] S81. The final optimized model parameters obtained during the joint training phase will be persistently saved and loaded into the deployment environment of the SPR signal processing system. The deployment environment includes a front-end sensing device for SPR raw signal acquisition, a signal preprocessing module, and an inference processing engine built based on the training model.

[0057] S82. During deployment and runtime, the raw, unprocessed signal output by the SPR sensor is collected in real time, input into the inference engine in time series format, and data format standardization is performed, including signal length pruning, amplitude normalization and dimension rearrangement, to ensure that the input signal meets the model structure requirements.

[0058] S83. The standardized original SPR signal is passed as the model input to the trained end-to-end intelligent denoising network. The end-to-end intelligent denoising network includes a sparse encoder module, a recursive residual decomposition and reconstruction module, and an inverse diffusion generation module. Forward inference calculation is performed using the parameters of each sub-module determined in the previous training. Inside the end-to-end intelligent denoising network, the main response component extraction, periodic interference term suppression, high-frequency mutation feature smoothing, and structural feature repair processing steps are completed in sequence to directionally eliminate various noise and unstructured components in the original input signal, and obtain the final output result of the network, while maintaining the amplitude shape, edge transition, and fluctuation trend of the real response.

[0059] S84. The final output of the network is returned to the main control module of the detection system as the SPR response curve with complete structure and noise suppression, for subsequent analysis, identification or diagnostic tasks, to ensure that the SPR detection process has high robustness, high resolution and good response accuracy under real operating conditions.

[0060] The beneficial effects of this invention are:

[0061] 1. This invention employs a sparse autoencoder to perform latent variable compression and sparse representation on the original noise signal, enabling high-dimensional complex signals to be modeled in a more discriminative feature form, providing a stable low-dimensional abstract expression for the subsequent generation and reconstruction process.

[0062] 2. By introducing a spectrum enhancement mechanism and a recursive residual decomposition module, the structural separation and fine reconstruction of the main response component, periodic interference component and high-frequency abrupt change component in the SPR signal are effectively realized, thereby greatly improving the integrity and recognizability of the signal structure.

[0063] 3. By deploying the optimized model parameters into the SPR detection process, this invention not only improves the intelligence level and practicality of the SPR sensing system, but also has good versatility and cross-environment adaptability, significantly enhancing the robustness and practical value of SPR detection in complex backgrounds. Attached Figure Description

[0064] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0065] Figure 1 This is a flowchart of an intelligent SPR noise data removal method based on generative adversarial networks proposed in this invention.

[0066] Figure 2 This is a schematic diagram of the algorithm structure of an intelligent SPR noise data removal method based on generative adversarial networks proposed in this invention. Detailed Implementation

[0067] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0068] refer to Figure 1-2 A method for intelligent removal of SPR noise data based on generative adversarial networks includes the following steps:

[0069] S1. Acquire the original surface plasmon resonance sensing response signal and perform normalization preprocessing. Use a fixed-length sliding window to divide the sensing response signal into multiple sub-sequence samples, construct a noise data training set containing temporal structure, and output labeled windowed SPR noise sub-sequence data.

[0070] S2. Input the windowed SPR noise subsequence data into the sparse coding variational autoencoder for encoding and compression processing. Extract latent structural features through sparsity constraints and variational regularization terms to obtain the sparse latent variable representation after preliminary denoising.

[0071] S3. Perform a short-time Fourier transform on the sparse latent variable representation to extract local spectral features within the time window, and use an adaptive Fourier feature fusion module to weight and combine different frequency components to output fused multi-scale spectral enhancement features.

[0072] S4. Input the fused multi-scale spectral enhancement features into the deconvolution diffusion model to simulate the reverse diffusion process of noise perturbation. Generate target signal candidates layer by layer through multiple reverse denoising steps, and output the preliminarily restored SPR signal sequence.

[0073] S5. Input the initially restored SPR signal sequence into the generator module, and at the same time input the manually labeled real SPR signal into the discriminator module. Optimize the generator output quality through the adversarial training mechanism of the generative adversarial network and output an SPR signal with enhanced authenticity.

[0074] S6. Input the pseudo-authenticity enhanced SPR signal into the recursive residual decomposition and reconstruction module to separate the main response component, periodic interference component and high-frequency mutation component, and then fuse them after structural reconstruction to obtain a structurally complete reconstructed SPR signal.

[0075] S7. Compare the reconstructed SPR signal with the original labeled real signal point by point, calculate the structural similarity index and gradient fidelity loss function, and jointly perform reverse update training on the entire generative adversarial network, sparse encoder and residual decomposition network to output the final optimized model parameters.

[0076] S8. Deploy the final optimized model parameters into the SPR detection process, input the real-time acquired raw SPR signal, and output the intelligently denoised SPR response curve.

[0077] This invention constructs a windowed subsequence training set, extracts latent variable features using sparse coding and variational regularization, performs multi-scale feature fusion in the spectral domain, and introduces an inverse diffusion model for candidate signal generation. It combines generative adversarial networks to enhance signal realism and further reconstructs the main response, periodic interference, and high-frequency abrupt change components through a recursive residual mechanism. By jointly training the entire network model using structural similarity and gradient fidelity loss, the reconstruction accuracy and robustness of SPR signals are significantly improved. Ultimately, it can be deployed in actual SPR detection processes to achieve high-quality intelligent denoising output.

[0078] In this embodiment, S1 specifically includes:

[0079] S11. The sample under test is continuously monitored by a surface plasmon resonance sensor to obtain the original response signal sequence that changes over time, and the response amplitude data at equal time intervals recorded at a fixed sampling frequency are used to form a complete time series dataset.

[0080] S12. Standardize the original response signal sequence by subtracting the average of all values ​​in the sequence from each response value in the signal, and then dividing the result by the standard deviation of the values ​​to obtain a standardized signal sequence with a mean of zero and consistent fluctuation scale, so as to eliminate the bias caused by the difference in dimensions.

[0081] S13. The standardized signal sequence is divided into multiple fixed-length subsequence samples using the sliding window method. Each subsequence contains signal values ​​at consecutive time points. The sliding window moves gradually on the sequence with a set step size to form the original labeled real signal with partial overlap, which is used to preserve the time dependence and local features of the original signal.

[0082] S14. For each original labeled real signal, based on the physical conditions, interference type and labeling basis provided in the experimental design, perform automated label allocation, determine whether it contains noise mutation points, periodic interference fluctuations or stable signal states, and generate training sample pairs with semantic attributes.

[0083] S15. Organize all training sample pairs and their corresponding labels into a training dataset with a uniform structure, ensuring that the sample lengths are consistent and the label types are distinguishable, so as to meet the requirements of neural network training for input format consistency and label supervision. Divide the samples according to the preset batch size, with each batch containing several subsequence samples and label information. Reconstruct and pad the batch samples according to the input dimensions required by the network to generate labeled windowed SPR noise subsequence data.

[0084] This step standardizes the continuously sampled raw response signals to eliminate dimensional differences and amplitude shifts between different measurements, enhancing data consistency and stability. While preserving the temporal structure, a sliding window strategy is used to segment the entire sequence into local sub-segments with overlapping regions, ensuring that the dynamic characteristics of the signal are fully preserved. Combining experimental conditions and annotation information, corresponding noise attribute labels are assigned to each segment, enabling accurate identification of multiple types of semantic interference signals.

[0085] In this embodiment, S2 specifically includes:

[0086] S21. The labeled windowed SPR noise subsequence data is evenly divided according to the preset spatial size to obtain image sub-blocks of the same size. Each sub-block is arranged in row priority order for independent processing and to preserve the spatial local structure of the image.

[0087] S22. Input each image sub-block into the feature extraction network of the shared structure, use local perception convolution operation to obtain the representation vector of the sub-block, add the representation vector to the embedding vector corresponding to the two-dimensional position in the original image to preserve spatial position information and form enhanced sub-block features.

[0088] S23. Combine all the enhanced sub-block features in sequence into a sequence structure, input it into the multi-head self-attention module to perform long-distance modeling between features, generate a global feature representation with context-dependent relationships, and maintain the stability of gradient flow and fuse with the original features through residual connection and normalization operation.

[0089] S24. Input the global feature representation into the feedforward neural network module, and perform linear mapping, nonlinear transformation and dimension reduction processing in sequence. Then, perform residual connection and normalization operations again to improve the representation capability and obtain the final sequence features used for encoding.

[0090] S25. Input the sequence features into the sparse coding variational autoencoder module, obtain the parameter representation of the latent space through the encoder function, perform sampling using the reparameter repetition technique, and introduce sparse constraints and prior regularization terms to construct a regularization function. The expression of the regularization function is as follows:

[0091] ;

[0092] in, For regularization terms, To input latent variables, To constrain latent variables, Let L1 norm be the sparse term. for The weight, This is the distribution consistency regularization coefficient, used to control... Weights of divergence ,for Divergence is a function used to measure the difference between two probability distributions. For the prior distribution, The posterior distribution calculated by the encoder. Let be the encoder parameter set; input the latent variables into the decoder, reconstruct the total sparse coding loss, and minimize the total sparse coding loss function to obtain the sparse latent variable representation after preliminary denoising.

[0093] This step achieves efficient feature extraction and denoising representation of SPR noise signals by constructing a sparse coding structure that integrates spatial awareness and attention mechanisms. The input signal is divided into sub-blocks while preserving temporal and spatial structure, and local semantic consistency is improved by combining positional embedding. Global dependencies are modeled through a self-attention mechanism to capture cross-regional interference correlations and structural changes. The extracted global features undergo further nonlinear transformation in a feedforward network before being input into a sparse variational autoencoder. Under sparsity constraints and distribution regularization, key latent features are compressed and expressed, enhancing the model's ability to model noisy interference structures.

[0094] In this embodiment, S3 specifically includes:

[0095] S31. Based on the block structure of the original image in the row and column directions, the sparse latent variable representation after preliminary denoising is restored according to the original sub-block order and sequentially filled into the two-dimensional matrix to generate a two-dimensional feature map consistent with the original image block structure. The size of the two-dimensional matrix is ​​consistent with the number of image blocks, and the channel dimension is kept as the feature dimension retained after sparse reconstruction.

[0096] S32. Perform compression operation on the two-dimensional feature map in the channel dimension. Use a linear mapping module with trainable parameters to reduce the dimension of the feature channel at each spatial location, compressing the high-dimensional features into a more compact low-dimensional representation to reduce the subsequent computation and enhance the expression of key information, while keeping the original spatial structure unchanged.

[0097] S33. Perform context modeling on the low-dimensional representation after channel compression to construct an information interaction mechanism between locations. Employ a global two-dimensional attention structure to weightedly fuse the features of each spatial location, constructing the dependency degree between each location and all other locations. Then, use a similarity scoring function to calculate the responsiveness between locations, obtaining the attention-enhanced two-dimensional feature map. The expression for the similarity scoring function is:

[0098] ;

[0099] in, , These represent the positions of the i-th row and j-th column, and the p-th row and q-th column, respectively. express , Similarity score between two positions This represents the compressed feature vector at the i-th row and j-th column. This represents the compressed feature vector at row p and column q. The projection matrix representing the query and key vectors. This is the scaling dimension in the attention mechanism, used to control the response intensity between different regions;

[0100] S34. Input the attention-enhanced 2D feature map into the channel mapping module, perform dimensionality upscaling on the feature channels at each location, and restore the compressed dimension to the same fused multi-scale spectral enhancement feature as the initial feature dimension through a learnable linear transformation, ensuring that subsequent modules can seamlessly connect and use the spatial features with full expressive power.

[0101] This step achieves spatial reconstruction and spectral enhancement of sparse latent variables by constructing a structure restoration and global attention fusion mechanism. The sparse latent variables are restored to two-dimensional feature maps according to the original image block order, preserving their spatial location information. Linear compression is performed in the channel dimension to extract low-dimensional features of key expressions, reducing computational complexity. A global two-dimensional attention mechanism is introduced to model the dependencies between information at different spatial locations, strengthening cross-regional feature connections through a similarity scoring function and enhancing global context awareness.

[0102] In this embodiment, S4 specifically includes:

[0103] S41. Arrange the multi-scale spectral enhancement features in the order of image sub-blocks and input them into the discriminant structure for preliminary analysis. Use the feature enhancement module composed of linear transformation and nonlinear activation to further process the feature vector at each position in the sequence and extract discriminative expressions to assist in the identification of target information in downstream classification or generation tasks.

[0104] S42. Perform order-preserving aggregation operation on discriminative representations. By setting a global query vector and constructing a weight distribution based on the correlation between features at each location, information from all locations is fused into a unified vector representation, capturing the overall global semantic features of the image.

[0105] S43. The global semantic features are input into the deconvolutional diffusion model as the initial state vector to participate in the iterative reverse generation process. At each diffusion step, noise repair and image prior guidance are performed on the current state to obtain the diffusion output result. The structure of the deconvolutional diffusion model is defined as follows:

[0106] ;

[0107] in, For the time step in the diffusion process, For the first The reconstructed state of the step, For the first The reconstructed state of the step, For the diffusion scheduling coefficient, This is the cumulative decay factor. For noise prediction networks, The semantic vector guide term generated by the preceding module. The sampling standard deviation Gaussian sampling noise is used to ensure sample diversity and model stability;

[0108] S44. Map the diffusion output back to the category space, pass through the linear projection module to obtain the final prediction output, and compare it with the label to obtain the generation error. Calculate the difference through the cross-entropy loss function for supervised training.

[0109] S45. The generation error and the total loss of sparse coding are weighted and combined to construct an overall training optimization objective, so that the entire network can maintain a unified learning strategy between compressed reconstruction and objective-oriented generation, and achieve effective reverse modeling and robust prediction.

[0110] This step achieves high-quality restoration of spectral features and target semantic guidance by constructing a generative mechanism that integrates discriminative enhancement and diffusion modeling. Input features are processed by a discriminative structure to extract discriminative representations, and then a global semantic vector is formed through a sequence-sensitive aggregation operation. The diffusion model uses this semantic vector as a condition to progressively restore the signal in multi-step reverse denoising, repairing structural details and removing residual interference, ensuring the stability and consistency of the generation process. The generated result is compared with the label to obtain the error signal, which is then jointly optimized with the encoding loss to achieve efficient end-to-end training and denoising modeling from latent variable compression and feature enhancement to target generation.

[0111] In this embodiment, S5 specifically includes:

[0112] S51. The image result generated in the last round of the overall training optimization target is used as a pseudo sample and input into the discriminator network. At the same time, an auxiliary discriminator network with learnable parameters is constructed. The auxiliary discriminator network receives the generated result as input and outputs a scalar value to indicate whether the input sample comes from the real data distribution, which is used to measure the closeness between the model-generated image and the real image.

[0113] S52. Select corresponding samples from the real sample set and input them into the discriminator network along with pseudo samples. During the training phase, drive the discriminator to learn to distinguish the distribution differences between generated data and real data by marking whether the input is real. At the same time, apply a target to the generator so that the output is judged as real by the discriminator to the greatest extent.

[0114] S53. Construct an adversarial target in the discriminator module and establish independent loss functions for the generator and discriminator respectively. The generator aims to maximize the probability that a fake sample is considered a real sample by the discriminator. The discriminator aims to maximize the confidence in judging real samples as real and fake samples as fake. The two are iteratively trained in an adversarial optimization manner to form a game mechanism.

[0115] S54. Based on the game mechanism, the parameters of the independent loss functions are updated separately. Backpropagation and gradient descent optimization methods are used to adjust the weight parameters of the generator network and the discriminator network separately to obtain the classification loss. The generator relies on the error of the pseudo sample being judged as real to update its own strategy, so as to gradually improve the semantic and structural realism of the generated content. At the same time, the discriminator adjusts the discrimination boundary according to the discrimination bias to improve the discrimination ability.

[0116] S55. Introduce the generator loss term into the overall training optimization objective, and combine it with the sparse coding total loss and classification loss in a weighted combination to form the pseudo-realism enhanced SPR signal. By setting the weight factor of the generation loss, adjust the proportion of its influence on the overall optimization objective, realize the dynamic balance between generation, reconstruction and classification, and improve the overall performance of the model in image understanding and generation tasks.

[0117] This step guides the model to learn high-order feature representations of the real data distribution by constructing an adversarial optimization mechanism between the generator and discriminator. The pseudo-samples output by the generator network are evaluated by the auxiliary discriminator module to determine their resemblance to real samples, and simultaneously input into the main discriminator along with real samples to construct a discriminative target for their difference. By setting adversarial losses for the generator and discriminator, their generation capabilities and discrimination accuracy are optimized respectively, allowing the model to continuously approach the real data distribution during the game process. During training, the generator gradually enhances the structural realism and semantic consistency of its output through error feedback, while the discriminator strengthens its identification boundaries for pseudo-samples. Finally, the generation loss, encoding loss, and classification loss are weighted and fused to form a unified optimization objective, achieving high-fidelity SPR signal modeling with enhanced pseudo-realism.

[0118] In this embodiment, S6 specifically includes:

[0119] S61. The SPR signal processed by the pseudo-realism enhancement module is used as input and sent to the structured recursive residual decomposition reconstruction module. The recursive residual decomposition module performs multi-scale component deconstruction operation based on the overall response structure of the signal. The initial input is the original enhanced signal form, maintaining timing consistency and channel matching.

[0120] S62. The recursive residual decomposition and reconstruction module sequentially unfolds multi-level feature separation paths, performs cascaded residual updates on the input signal, extracts the main components, and obtains the separated signal components, including stable response components, interference fluctuation components with periodic characteristics, and high-frequency components with local mutation characteristics. The module outputs feature representations in different frequency domains and structural domains at each layer to capture the inherent multi-structure information of the signal.

[0121] S63. Construct independent structural reconstruction paths for each type of separated signal component. While maintaining the original characteristic distribution, introduce a deep residual iteration mechanism. Enhance the morphology of the main response part through multi-layer nonlinear mapping, calibrate the waveform of periodic interference, and smooth the edges and repair the details of high-frequency mutations. Each sub-path completes the target-oriented structural completion within its own dimension.

[0122] S64. The signal components reconstructed from all sub-paths are uniformly fused in the time dimension and amplitude space. They are then merged and aligned according to the initial signal arrangement. The weight adjustment mechanism controls the proportion of each component in the final output to obtain a structurally complete reconstructed SPR signal that is structurally continuous, has a consistent response, and retains key signal features. This ensures that the signal quality, structural integrity, and response accuracy meet the input standards under the multi-task fusion framework, supporting the steady-state inference capability and terminal performance of the overall system.

[0123] This step involves constructing a recursive residual decomposition and reconstruction module to achieve structured reconstruction of the pseudo-fidelity-enhanced SPR signal. This module utilizes multi-level residual paths to decompose the input signal, extracting key components such as the main response, periodic interference, and high-frequency abrupt changes. For each type of component, an independent reconstruction path is constructed, achieving refined completion of the waveform structure and precise repair of noise components.

[0124] In this embodiment, S7 specifically includes:

[0125] S71. Align the reconstructed SPR signal with the original labeled real signal on the time axis point by point to ensure that the two are consistent in the sampling frequency, start time and signal length dimension. After the alignment is completed, construct a one-to-one corresponding point-to-point matching structure.

[0126] S72. In the point-to-point matching structure, for each sliding time window, the local average value of the real signal and the reconstructed signal within the sliding time window is extracted to characterize the energy center trend of the current region. The local amplitude variance of the two types of signals within the window is calculated to reflect the stability of the signal intensity distribution. At the same time, the covariance information of the two sets of signals within the window range is calculated to measure the degree of synchronous change between the waveforms. The local mean, variance and covariance are then combined to obtain an index score for measuring the structural similarity of the signals. The index score is calculated segment by segment within the entire window range to form the structural similarity estimate of the entire signal.

[0127] S73. Based on the sequence form of the real signal and the reconstructed signal in the time domain, the amplitude difference between each sampling point and the previous sampling point is calculated in turn as the approximate gradient of the current point. The absolute difference between the corresponding gradient values ​​of the real signal and the reconstructed signal at the same time point is calculated to measure the consistency of the local change trend. The errors of all sampling points are accumulated and averaged to obtain an overall gradient fidelity error used to describe whether high-frequency details are preserved. The error is used in training to constrain the model's ability to recover abrupt changes.

[0128] S74. The structural similarity index score obtained above and the overall gradient fidelity error are included as constraints in the training optimization objective. The adversarial loss included in the generator path and the main loss function used for data reconstruction in the sparse encoder are fused together to construct a multi-loss jointly driven training objective. In the training objective, adjustable weight coefficients are introduced for the error terms of different tasks to control the degree of influence of each part of the loss on the overall network parameter update, thereby achieving dynamic balance adjustment of the task objectives of different modules.

[0129] S75. Based on the training objective driven by multiple losses, all error information is propagated along the reverse path to the corresponding modules, including the generator structure, sparse encoder structure and residual reconstruction structure. The gradient propagation and weight update mechanism in deep networks are used to synchronously iteratively update the parameters of each module. During the training process, all modules are collaboratively adjusted in the same optimization cycle until the overall network converges under multiple error indicators and outputs the final optimized model parameters.

[0130] This step introduces a dual constraint mechanism of structural similarity and gradient fidelity to achieve precise alignment and evaluation of the reconstructed SPR signal with the real signal at the structural and detail levels. A sliding window is used to calculate local statistical features, including mean, variance, and covariance, to measure the structural consistency of the overall signal. The adversarial loss of the joint index generator and the reconstruction loss of the sparse encoder together constitute multiple training objectives, and a weight adjustment strategy is introduced to achieve joint optimization of each sub-module.

[0131] In this embodiment, S8 specifically includes:

[0132] S81. The final optimized model parameters obtained during the joint training phase will be persistently saved and loaded into the deployment environment of the SPR signal processing system. The deployment environment includes a front-end sensing device for SPR raw signal acquisition, a signal preprocessing module, and an inference processing engine built based on the training model.

[0133] S82. During deployment and runtime, the raw, unprocessed signal output by the SPR sensor is collected in real time, input into the inference engine in time series format, and data format standardization is performed, including signal length pruning, amplitude normalization and dimension rearrangement, to ensure that the input signal meets the model structure requirements.

[0134] S83. The standardized original SPR signal is passed as the model input to the trained end-to-end intelligent denoising network. The end-to-end intelligent denoising network includes a sparse encoder module, a recursive residual decomposition and reconstruction module, and an inverse diffusion generation module. Forward inference calculation is performed using the parameters of each sub-module determined in the previous training. Inside the end-to-end intelligent denoising network, the main response component extraction, periodic interference term suppression, high-frequency mutation feature smoothing, and structural feature repair processing steps are completed in sequence to directionally eliminate various noise and unstructured components in the original input signal, and obtain the final output result of the network, while maintaining the amplitude shape, edge transition, and fluctuation trend of the real response.

[0135] S84. The final output of the network is returned to the main control module of the detection system as the SPR response curve with complete structure and noise suppression, for subsequent analysis, identification or diagnostic tasks, to ensure that the SPR detection process has high robustness, high resolution and good response accuracy under real operating conditions.

[0136] This step involves model deployment, including parameter persistence and inference environment loading, combined with standardized data format conversion to ensure compatibility between the input signal and the model structure. Based on an end-to-end network structure generated by sparse coding, residual decomposition, and diffusion, the model can perform multi-dimensional modeling and progressive removal of the main response, periodic interference, and high-frequency noise in the input signal, while preserving the structural characteristics and dynamic trends of the real signal.

[0137] Example 1:

[0138] To verify the feasibility of this invention in practice, it was applied to an environmental biosensor experimental station in the Yangtze River Basin located in a certain city to conduct an online monitoring task for trace pesticide residues. The station was equipped with three parallel-operating surface plasmon resonance (SPR) online sensors, with monoclonal antibodies highly affinity for common organophosphorus pesticides immobilized on the sensor chip surface. Due to the monitoring point's proximity to the main shipping channel, ship exhaust discharge, high-frequency water surface fluctuations, and bubble interference introduced by thunderstorms caused periodic baseline drift and dense high-frequency noise pulses in the SPR response curve. Traditional Savitzky-Golay smoothing and wavelet threshold filtering were insufficient to stably denoise the signal without damaging the target binding signal, thus affecting the accuracy of identifying instantaneous pesticide concentration changes.

[0139] In this scenario, maintenance personnel continuously import the raw SPR signals acquired 24 hours a day into the end-to-end intelligent denoising platform of this invention. The platform first standardizes the raw curve to zero mean and converts it into labeled subsequence data using a sliding window with a length of 800 points and a step size of 200 points, significantly improving the sensitivity of subsequent models to local noise patterns. Subsequently, the windowed subsequences are fed into a pre-trained sparse coding variational autoencoder to extract latent variables, and combined with an adaptive Fourier feature fusion module to obtain multi-scale spectral enhancement representations. A deconvolutional diffusion model uses these representations to generate candidate denoised sequences, and a generative adversarial network further enhances the realism of the reconstructed sequences. A recursive residual decomposition reconstruction module subdivides the generated sequences into three components: main response, periodic interference, and high-frequency mutations, and reconstructs and fuses them, outputting a structurally complete reconstructed SPR curve. Finally, the system compares the structural similarity of the reconstructed curve with historical "clean" sample curves in real time and dynamically adjusts the discrimination threshold, controlling the noise elimination time to within an average of 46 milliseconds per sequence without affecting the on-site sampling rhythm, thus achieving online continuous inference. To visually demonstrate the effectiveness of this embodiment, the following table provides a comparison of noise reduction results from online SPR monitoring of the lower Yangtze River:

[0140] Table 1 Comparison of noise reduction results from online SPR monitoring of water bodies in the lower reaches of the Yangtze River

[0141] Average signal-to-noise ratio of the original curve / dB Signal-to-noise ratio after denoising / dB SSIM Enhancement <![CDATA[True value of pesticide concentration / ng mL⁻ 1 > Traditional method error / % 11.4 21.0 +0.29 2.7 15.2 10.8 20.2 +0.33 4.1 17.5 12.1 22.4 +0.30 3.6 19.3 9.9 19.4 +0.28 5.8 20.7 10.3 20.6 +0.32 1.9 16.4 11.7 21.9 +0.31 6.2 22.0

[0142] As shown in the table above, during the entire monitoring period, the instantaneous concentration of trace pesticides in the Yangtze River water fluctuated between 0 ng / mL. -1 Up to 6.2 ng / mL -1 The SPR curve after treatment according to this invention is at 4.8 ng / mL. -1The peak exhibits a clear binding-dissociation biphasic characteristic, which traditional wavelet filtering schemes cannot distinguish due to residual high-frequency noise interference. Actual laboratory comparison tests show that the average relative error between the concentration value calculated using this invention and the concentration measured by gas chromatography-mass spectrometry (GC-MS) decreased from 18.6% to 5.2%, the structural similarity index improved by an average of 0.31, and the signal-to-noise ratio improved by approximately 9.7 dB.

[0143] The end-to-end model of this invention can be deployed once and inferred on three sensors in parallel streams, with CPU utilization not exceeding 42% and no significant latency. It also exhibits good robustness to sudden rainstorms and low-temperature nighttime environments in the Yangtze River. This embodiment fully verifies the comprehensive suppression capability of this invention against high-frequency abrupt pulses and low-frequency baseline drift noise in complex open water environments, significantly improving the real-time sensing resolution of trace pesticide concentrations in the SPR online detection system, and providing a reliable technical means for environmental emergency monitoring.

[0144] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent removal of SPR noise data based on generative adversarial networks, characterized in that, Includes the following steps: S1. Acquire the original surface plasmon resonance sensing response signal and perform normalization preprocessing. Use a fixed-length sliding window to construct the original labeled real signal containing the temporal structure. The output is labeled windowed SPR noise subsequence data. S2. Input the windowed SPR noise subsequence data into the sparse coding variational autoencoder for coding and compression processing to obtain the sparse latent variable representation after preliminary denoising. S3. Perform a short-time Fourier transform on the sparse latent variable representation, and use an adaptive Fourier feature fusion module to weight and combine different frequency components to output fused multi-scale spectral enhancement features. S4. Input the fused multi-scale spectral enhancement features into the deconvolution diffusion model to simulate the reverse diffusion process of noise perturbation. Generate target signal candidates layer by layer through multiple reverse denoising steps, and output the overall training optimization target. S5. Input the overall training optimization target into the generator module, and simultaneously input the manually labeled real SPR signal into the discriminator module. Optimize the generator output quality through the adversarial training mechanism of the generative adversarial network, and output an SPR signal with enhanced pseudo-realism. S6. Input the pseudo-authenticity enhanced SPR signal into the recursive residual decomposition and reconstruction module, perform structural reconstruction and then fuse the signals to obtain a structurally complete reconstructed SPR signal. S7. Compare the reconstructed SPR signal with the original labeled real signal point by point, calculate the structural similarity index and gradient fidelity loss function, perform reverse update training, and output the final optimized model parameters. S8. Deploy the final optimized model parameters into the SPR detection process, input the real-time acquired raw SPR signal, and output the intelligently denoised SPR response curve. S2 specifically includes: S21. Divide the labeled windowed SPR noise subsequence data evenly according to the preset spatial size to obtain image sub-blocks of the same size; S22. Input each image sub-block into the feature extraction network of the shared structure, use local perception convolution operation to obtain the representation vector of the sub-block, and add the representation vector to the embedding vector corresponding to the two-dimensional position in the original image to form the enhanced sub-block feature. S23. Combine all the enhanced sub-block features in sequence into a sequence structure, input it into the multi-head self-attention module to perform long-distance modeling between features, and generate a global feature representation with context-dependent relationships. S24. Input the global feature representation into the feedforward neural network module, and perform linear mapping, nonlinear transformation and dimension reduction processing in sequence to obtain the final sequence features used for encoding. S25. Input the sequence features into the sparse coding variational autoencoder module, obtain the parameter representation of the latent space through the encoder function, use the reparameter technique for sampling, and introduce sparse constraints and prior regularization terms to construct a regularization function. Input the latent variables into the decoder, reconstruct the total sparse coding loss, and minimize the total sparse coding loss function to obtain the sparse latent variable representation after preliminary denoising.

2. The intelligent SPR noise data removal method based on generative adversarial networks according to claim 1, characterized in that, S1 specifically includes: S11. The sample under test is continuously monitored using a surface plasmon resonance sensor to obtain the original response signal sequence that changes over time. S12. Standardize the original response signal sequence to obtain a standardized signal sequence with zero mean and consistent fluctuation scale. S13. The standardized signal sequence is divided into multiple fixed-length subsequence samples using the sliding window method. The sliding window is moved step by step on the sequence to form the original labeled real signal with partial overlap. S14. Perform automated label assignment for each original labeled real signal to generate training sample pairs with semantic attributes; S15. Organize all training sample pairs and their corresponding labels into a training dataset with a uniform structure, and divide it according to the preset batch size. Reconstruct and pad the batch samples according to the input dimension required by the network to generate labeled windowed SPR noise subsequence data.

3. The intelligent SPR noise data removal method based on generative adversarial networks according to claim 1, characterized in that, S3 specifically includes: S31. Based on the block structure of the original image in the row and column directions, the sparse latent variable representations after preliminary denoising are restored to their original positions according to the order of the sub-blocks and sequentially filled into the two-dimensional matrix to generate a two-dimensional feature map consistent with the block structure of the original image. S32. Perform compression operation on the two-dimensional feature map in the channel dimension. Use the linear mapping module to reduce the dimension of the feature channel at each spatial location, compressing the high-dimensional features into a more compact low-dimensional representation. S33. Perform context modeling on the low-dimensional representation after channel compression, use a global two-dimensional attention structure to weight and fuse the features of each spatial location, and use a similarity scoring function to calculate the degree of response between locations to obtain the attention-enhanced two-dimensional feature map. S34. Input the attention-enhanced 2D feature map into the channel mapping module, perform dimensionality upscaling on the feature channels at each location, and restore the compressed dimension to the same fused multi-scale spectral enhancement feature as the initial feature dimension through a learnable linear transformation.

4. The intelligent SPR noise data removal method based on generative adversarial networks according to claim 1, characterized in that, S4 specifically includes: S41. Arrange the multi-scale spectral enhancement features in the order of image sub-blocks and input them into the discriminative structure for preliminary analysis to extract discriminative expressions; S42. Perform order-preserving aggregation operation on discriminative representations. By setting a global query vector and constructing a weight distribution based on the correlation between features at each location, information from all locations is fused into a unified vector representation, capturing the overall global semantic features of the image. S43. Input the global semantic features into the deconvolutional diffusion model as the initial state vector to participate in the progressively iterative reverse generation process. At each diffusion step, perform noise repair and image prior guidance on the current state to obtain the diffusion output result. S44. Map the diffusion output back to the category space, pass it through the linear projection module to obtain the final prediction output, and compare it with the label to obtain the generation error; S45. The generation error and the total loss of sparse coding are weighted and combined to construct the overall training optimization objective.

5. The intelligent SPR noise data removal method based on generative adversarial networks according to claim 1, characterized in that, S5 specifically includes: S51. Input the image results generated in the last round as pseudo-samples into the discriminator network, and at the same time construct an auxiliary discriminator network with learning parameters. S52. Select corresponding samples from the real sample set and input them into the discriminator network along with the pseudo samples. During the training phase, drive the discriminator to learn to distinguish the distribution differences between generated data and real data by marking whether the input is real. S53. Construct an adversarial target in the discrimination module, establish independent loss functions for the generator and discriminator respectively, and iteratively train the two in an adversarial optimization manner to form a game mechanism. S54. Based on the game theory mechanism, the parameters of the independent loss functions are updated separately. The weight parameters of the generator network and the discriminator network are adjusted separately using backpropagation and gradient descent optimization methods to obtain the classification loss. S55. Introduce the generator's loss term into the overall training optimization objective, and combine it with the sparse coding total loss and classification loss in a weighted manner to form the pseudo-realism enhanced SPR signal.

6. The intelligent SPR noise data removal method based on generative adversarial networks according to claim 1, characterized in that, S6 specifically includes: S61. The SPR signal processed by the pseudo-authenticity enhancement module is used as input and sent to the structured recursive residual decomposition and reconstruction module. S62. Within the recursive residual decomposition and reconstruction module, multi-level feature separation paths are sequentially expanded to perform cascaded residual updates on the input signal, extract the main components, and obtain the separated signal components. The main components include stable response components, interference fluctuation components with periodic characteristics, and high-frequency components with local mutation characteristics. S63. Construct independent structural reconstruction paths for each type of separated signal component, enhance the morphology of the main response part through multi-layer nonlinear mapping, and complete the target-oriented structural completion within its own dimension. S64. The signal components reconstructed from all sub-paths are uniformly fused in the time dimension and amplitude space, and then merged and aligned according to the initial signal arrangement to obtain a structurally complete reconstructed SPR signal that is structurally continuous, has a consistent response, and retains key signal features.

7. The intelligent SPR noise data removal method based on generative adversarial networks according to claim 1, characterized in that, Specifically, S7 includes: S71. Align the structurally complete reconstructed SPR signal with the original labeled real signal point by point on the time axis. After alignment, construct a one-to-one corresponding point-to-point matching structure. S72. In the point-to-point matching structure, for each sliding time window, extract the local average value of the real signal and the reconstructed signal, and obtain an index score to measure the similarity of the signal structure. S73. Based on the sequence form of the real signal and the reconstructed signal in the time domain, calculate the absolute difference between the corresponding gradient values ​​of the real signal and the reconstructed signal at the same time point to obtain the overall gradient fidelity error. S74. The structural similarity index score and the overall gradient fidelity error are incorporated as constraints into the training optimization objective to construct a training objective jointly driven by multiple losses. S75. Based on the training objective driven by multiple losses, all error information is propagated to the corresponding modules along the reverse path. Gradient propagation and weight update mechanisms are used to update and train the parameters of each module, and the final optimized model parameters are output.

8. The intelligent SPR noise data removal method based on generative adversarial networks according to claim 1, characterized in that, S8 specifically includes: S81. The final optimized model parameters obtained during the joint training phase will be persistently saved and loaded into the deployment environment of the SPR signal processing system. S82. During deployment and operation, the raw, unprocessed signals output by the SPR sensor are collected in real time, input into the inference engine in time series format, and the data format is standardized. S83. The standardized original SPR signal is passed as the model input to the trained end-to-end intelligent denoising network to perform targeted elimination of various noise and unstructured components in the original input signal, and the final output result of the network is obtained. S84. The final output of the network is returned to the main control module of the detection system as the SPR response curve after structural integrity and noise suppression.

Images