Instrument analysis migration method and system based on feature decoupling and physical residual calibration

Abstract

The invention discloses an instrument analysis migration method and system based on feature decoupling and physical residual calibration, and relates to the technical field of artificial intelligence and instrument analysis data processing, and the method comprises the steps: constructing a double-flow deep neural network model, and carrying out the feature decoupling pre-training; calculating characteristic correlation loss by adding an orthogonality penalty term and carrying out back propagation; acquiring an original signal of an instrument to be analyzed, inputting the original signal into the trained double-flow deep neural network model to extract corresponding general chemical characteristics and response deviation characteristics, acquiring physical signal intensity characterization through weighted aggregation based on the general chemical characteristics, and simultaneously acquiring physical signal intensity characterization through weighted aggregation based on a linear response principle; and constructing a dynamic response factor by combining the general chemical characteristics and the response deviation characteristics, and calculating to obtain a quantitative analysis result. According to the method and the device, while the dependence of the marked sample is reduced, the accuracy of cross-instrument analysis is improved, and the cross-equipment migration of the instrument analysis model with low sample cost and high robustness is realized.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and instrumental analysis data processing technology, and in particular to an instrumental analysis migration method and system based on feature decoupling and physical residual calibration. Background Technology

[0002] Instrumental analysis technology is widely used in fields such as chemical engineering, pharmaceuticals, environmental monitoring, and food safety. It uses instruments such as spectrometers, mass spectrometers, and chromatograms to acquire characteristic data from samples, enabling qualitative and quantitative analysis of substances. In practical applications, the same analytical task often requires switching between multiple instruments or replacing instruments of different models and service lives. Due to differences in instrument hardware, fluctuations in the working environment, and component aging, the detection data for the same substance from different instruments exhibit significant distribution differences, meaning that analytical models trained on the source instrument cannot be directly used on the target instrument.

[0003] Traditional instrumental analysis transfer methods mainly rely on a large number of labeled samples for model retraining or on simple mathematical transformations to achieve data alignment. However, obtaining a large number of labeled samples is time-consuming and costly, especially for trace analysis and complex matrix sample analysis scenarios. Simple mathematical transformations can only achieve surface data distribution matching without considering the differences in the physical characteristics of the instruments themselves, resulting in insufficient accuracy and robustness of the transfer model, making it difficult to meet actual analytical needs.

[0004] Feature decoupling techniques can separate data features into core features and interference features, and transfer learning techniques can achieve knowledge transfer across domain models. However, existing technologies do not combine feature decoupling with instrument physical models, making it impossible to accurately capture system response differences between instruments, and residual calibration lacks a physical basis, thus affecting the transfer effect. Therefore, there is an urgent need for an instrument analysis transfer method that takes into account both data feature decoupling and physical residual calibration, reducing dependence on labeled samples while improving the accuracy of cross-instrument analysis, and achieving low-sample-cost, highly robust instrument analysis model transfer across devices. Summary of the Invention

[0005] The purpose of this application is to provide an instrument analysis migration method and system based on feature decoupling and physical residual calibration, which can improve the accuracy of cross-instrument analysis while reducing the dependence on labeled samples, and realize the migration of instrument analysis models across devices with low sample cost and high robustness.

[0006] To achieve the above objectives, this application provides the following solution.

[0007] In a first aspect, this application provides an instrument analysis migration method based on feature decoupling and physical residual calibration, which includes the following steps.

[0008] A two-stream deep neural network model is constructed. The two-stream deep neural network model includes a shared feature branch and a private bias branch. The shared feature branch is used to extract general chemical features across devices from instrument signals, and the private bias branch is used to extract response bias features of the corresponding devices from instrument signals. The shared feature branch and the private bias branch are configured asymmetrically. The parameters of the shared feature branch are frozen during the device-side adaptive evolution phase, and the private bias branch is initialized to a zero-bias state, a near-zero small random state, or its output contribution to the quantitative analysis results is suppressed in the initialization state.

[0009] The dual-stream deep neural network model is pre-trained with feature decoupling. During the training process, the feature correlation loss is calculated by adding an orthogonality penalty term to the total loss function and backpropagation is performed to minimize the statistical correlation between the general chemical feature and the response deviation feature, thereby obtaining the trained dual-stream deep neural network model.

[0010] The raw signal of the instrument to be analyzed is acquired and input into the trained dual-stream deep neural network model to extract the corresponding general chemical features and response deviation features. Based on the general chemical features, a weighted aggregation is used to obtain the physical signal intensity characterization. Simultaneously, based on the linear response principle, a dynamic response factor is constructed by combining the general chemical features and the response deviation features. The quantitative analysis result is calculated based on the physical signal intensity characterization and the dynamic response factor. The quantitative analysis result is the concentration value of the analyte corresponding to the raw signal of the instrument to be analyzed.

[0011] Optionally, the orthogonality penalty term is configured to calculate the cosine similarity, mutual information, or Frobenius norm between the general chemical feature matrix and the response bias feature matrix.

[0012] Optionally, when performing feature decoupling pre-training on the dual-stream deep neural network model, the pre-training data used is a constructive synthetic dataset based on a physical dynamics model. The constructive synthetic dataset is constructed by superimposing simulated baseline drift, noise, and nonlinear deformation onto an ideal signal generated based on a Gaussian function, a Lorentz function, or an exponentially modified Gaussian function.

[0013] Optionally, the weighted aggregation employs soft integral calculation based on an attention mechanism, specifically including the following steps.

[0014] Based on the general chemical characteristics, an attention weight sequence is generated that is time-domain aligned with the original signal of the instrument to be analyzed.

[0015] The attention weight sequence is multiplied point by point with the original signal of the instrument to be analyzed and then summed to obtain the physical signal intensity characterization.

[0016] Optionally, the raw signal of the instrument to be analyzed is acquired and input into the trained dual-stream deep neural network model to extract the corresponding general chemical features and response deviation features respectively. Based on the general chemical features, the physical signal intensity characterization is obtained through weighted aggregation. At the same time, based on the linear response principle, a dynamic response factor is constructed by combining the general chemical features and the response deviation features. The quantitative analysis results are calculated based on the physical signal intensity characterization and the dynamic response factor. The specific steps include the following steps.

[0017] Obtain the raw signals from the instrument to be analyzed.

[0018] The original instrument signal to be analyzed is input into the trained dual-stream deep neural network model to extract the general chemical features and response deviation features corresponding to the original instrument signal to be analyzed.

[0019] An attention weight sequence is generated based on the general chemical characteristics, and the original instrument signal to be analyzed is weighted and summed according to the attention weight sequence to obtain the physical signal intensity value characterizing the peak area, which is used as the physical signal intensity characterization.

[0020] The baseline response factor is predicted using the general chemical characteristics, and the deviation response factor is predicted using the response deviation characteristics. The baseline response factor and the deviation response factor are then superimposed to generate the dynamic response factor under the current instrument condition.

[0021] Based on the principle of linear response, calculations are performed on the physical signal intensity characterization and the dynamic response factor to obtain quantitative analysis results characterizing the concentration of the analyte.

[0022] Optionally, after the step of calculating the quantitative analysis results based on the physical signal intensity characterization and the dynamic response factor, the instrument analysis migration method based on feature decoupling and physical residual calibration further includes the step of device-side adaptive evolution.

[0023] The device-side adaptive evolution specifically includes the following steps.

[0024] During the online model operation phase, target samples are screened based on the confidence assessment of the current prediction results, and the calibration true value corresponding to the target sample is received from external input. The current prediction result is the quantitative analysis result corresponding to the original instrument signal to be analyzed, the target sample is the original instrument signal sample to be analyzed whose confidence assessment value is greater than the preset calibration threshold and requires manual provision of the true concentration value to complete the model calibration update, and the calibration true value is the true concentration value.

[0025] Based on the target sample and its calibration true value, the parameters of the private bias branch are incrementally updated, while the parameters of the shared feature branch are kept frozen.

[0026] Optionally, the confidence assessment is achieved by calculating the uncertainty index predicted by the trained two-stream deep neural network model. The uncertainty index can be calculated by performing multiple inferences using Monte Carlo loss technique to calculate the variance or by calculating the output dispersion using a deep ensemble network.

[0027] Optionally, when receiving the calibration truth value corresponding to the target sample from an external input, a dual threshold screening and interactive prompting strategy is adopted, specifically including the following steps.

[0028] Set a safety threshold and a calibration threshold, wherein the calibration threshold is greater than the safety threshold.

[0029] When the confidence assessment value of the current prediction result is less than the safety threshold, the current sample is determined to be a normal, risk-free sample, and the corresponding quantitative analysis result is directly output; the current sample is the original instrument signal sample to be analyzed that is currently acquired and input.

[0030] When the confidence assessment value of the current prediction result is greater than or equal to the safety threshold and less than or equal to the calibration threshold, the current sample is determined to be a medium-risk sample, the current sample is marked as a yellow warning signal, and the corresponding quantitative analysis result is output.

[0031] When the confidence assessment value of the current prediction result is greater than the calibration threshold, the current sample is determined to be a high-risk sample, a calibration prompt signal is generated and output, and calibration truth data input in response to the calibration prompt signal is received; the calibration prompt signal is used to instruct external devices or interfaces to input the calibration truth data for the high-risk sample.

[0032] Secondly, this application proposes an instrument analysis migration system based on feature decoupling and physical residual calibration. The instrument analysis migration system based on feature decoupling and physical residual calibration is used to implement the instrument analysis migration method based on feature decoupling and physical residual calibration as described in any of the first aspects. The instrument analysis migration system based on feature decoupling and physical residual calibration includes the following modules.

[0033] A model building module is used to construct a two-stream deep neural network model. The two-stream deep neural network model includes a shared feature branch and a private bias branch. The shared feature branch is used to extract general chemical features across instruments from instrument signals, and the private bias branch is used to extract response bias features of the corresponding instruments from instrument signals. The shared feature branch and the private bias branch are configured asymmetrically. The parameters of the shared feature branch are frozen during the adaptive evolution phase at the instrument end, and the private bias branch is initialized to a zero-bias state, a near-zero small random state, or its output contribution to the quantitative analysis results is suppressed in the initial state.

[0034] The feature decoupling pre-training module is used to perform feature decoupling pre-training on the two-stream deep neural network model. During the training process, the feature correlation loss is calculated by adding an orthogonality penalty term to the total loss function and backpropagation is performed to minimize the statistical correlation between the general chemical feature and the response deviation feature, thereby obtaining the trained two-stream deep neural network model.

[0035] The signal acquisition module is used to acquire the raw signals from the instrument to be analyzed.

[0036] The dual-stream inference module is used to input the raw signal of the instrument to be analyzed into the trained dual-stream deep neural network model to extract the corresponding general chemical features and response deviation features, respectively. Based on the general chemical features, a physical signal intensity characterization is obtained through weighted aggregation. Simultaneously, based on the linear response principle, a dynamic response factor is constructed by combining the general chemical features and the response deviation features. The quantitative analysis result is calculated based on the physical signal intensity characterization and the dynamic response factor. The quantitative analysis result is the concentration value of the analyte corresponding to the raw signal of the instrument to be analyzed.

[0037] Optionally, the instrument analysis migration system based on feature decoupling and physical residual calibration further includes the following modules.

[0038] An adaptive evolution module is used during the online model run phase to screen target samples based on the confidence assessment of the current prediction results and receive externally input calibration ground truth values ​​corresponding to the target samples; and to incrementally update the parameters of the private bias branch based on the target samples and the calibration ground truth values, while keeping the parameters of the shared feature branch frozen; wherein, the current prediction result is the quantitative analysis result corresponding to the original instrument signal to be analyzed obtained at present, the target sample is the original instrument signal sample to be analyzed whose confidence assessment value is greater than a preset calibration threshold and requires manual provision of a true concentration value to complete the model calibration update, and the calibration ground truth value is the true concentration value.

[0039] According to the specific embodiments provided in this application, this application has the following technical effects.

[0040] This application provides an instrument analysis migration method and system based on feature decoupling and physical residual calibration. On one hand, by constructing a two-stream deep neural network model including a shared feature branch for extracting general chemical features and a private bias branch for extracting device response bias features, and by adopting an asymmetric configuration between the shared feature branch and the private bias branch, the computational requirements for edge training are greatly reduced through an asymmetric strategy of freezing a large number of shared feature branches and updating only a small number of private bias branches, thus preventing catastrophic forgetting and enabling the instrument to have self-calibration and evolution capabilities throughout its entire lifecycle. On the other hand, By combining feature decoupling with physical residual calibration, feature decoupling is implemented in the pre-training stage of the two-stream deep neural network model. An orthogonality penalty term is added to minimize the statistical correlation between general chemical features and response bias features, forcibly separating general chemical features from device response bias. This allows the shared feature branch to extract a "chemical fingerprint" that does not age with the device, effectively avoiding negative migration during cross-device migration and achieving deep feature decoupling and robust migration. Physical residual calibration is implemented in the two-stream inference stage of the two-stream deep neural network model. By adopting a zero-bias initialization strategy with private bias branches, new devices can perform "zero-sample inference" based on shared feature branches without any calibration data. Simultaneously, adaptation can be completed by updating the lightweight private bias branches with only a small amount of calibration data, significantly reducing maintenance costs, solving the cold start problem, and lowering data costs. Furthermore, this method combines deep learning feature extraction with the principle of physical linear response, ensuring that the quantitative calculation process conforms to the physical principles of instrumental analysis, improving the metrological credibility of the results, enhancing the interpretability and physical consistency of the model, and reducing reliance on calibrated samples while improving the accuracy of cross-instrumental analysis. It enables low-sample-cost, highly robust instrumental analysis model transfer across devices, solving problems such as reliance on a large number of calibrated samples, lack of physical basis for calibration, low transfer accuracy, and catastrophic forgetting caused by incremental updates in existing cross-instrumental analysis transfers. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is an application environment diagram of an instrument analysis migration method based on feature decoupling and physical residual calibration, provided as an embodiment of this application.

[0043] Figure 2 This is a flowchart illustrating an instrument analysis migration method based on feature decoupling and physical residual calibration, provided as an embodiment of this application.

[0044] Figure 3 This is a schematic diagram of the overall adaptive migration process provided in an embodiment of this application.

[0045] Figure 4 This is a schematic diagram of a dual-stream network architecture and feature decoupling mechanism provided in an embodiment of this application.

[0046] Figure 5 This is a schematic diagram of the quantitative calculation logic based on the physical response model provided in an embodiment of this application.

[0047] Figure 6 This is a schematic diagram of the device-side adaptive evolution and human-computer interaction process provided in an embodiment of this application.

[0048] Figure 7 This is a schematic diagram of the structure of an instrument analysis migration system based on feature decoupling and physical residual calibration, provided as an embodiment of this application. Detailed Implementation

[0049] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0050] Instrumental analytical techniques (such as gas chromatography, liquid chromatography, mass spectrometry, and spectroscopy) are the cornerstone of modern chemistry, medicine, and environmental monitoring. In traditional quantitative analysis procedures, the analytical results are highly dependent on the instrument's condition and the calibration of the standard curve. According to the Lambert-Beer Law or the principle of chromatographic linear response, the concentration of the analyte is usually linearly related to the instrument's response signal (such as peak area or absorbance).

[0051] However, in practical applications, instrument response characteristics exhibit significant device specificity and time-varying characteristics. Even two analytical instruments of the same model often show non-negligible deviations in their response signals to the same sample due to differences in light source aging, column packing material, detector sensitivity drift, and thermal noise of electronic components. Therefore, traditional methods require the reconstruction of the standard curve using a series of standard substances before each new instrument is put into use, or after the instrument has been running for a period of time. This frequent downtime calibration and manual calibration not only incurs expensive standard costs and severely reduces detection efficiency, but also fails to meet the needs of large-scale, real-time online analysis at the industrial edge.

[0052] In recent years, deep learning technologies, represented by convolutional neural networks (CNN) and recurrent neural networks (RNN), have been applied to instrument signal processing, attempting to achieve automated qualitative and quantitative analysis through end-to-end learning. Although deep learning models perform well in feature extraction, they still face severe challenges in cross-device model migration scenarios, with the following technical problems: (1) Negative transfer problem caused by feature entanglement: Existing deep learning models often couple the general chemical characteristics of substances with the specific response bias of devices, resulting in a significant performance drop when the model is migrated to a new device due to domain drift. (2) Cold start and data dependency problem: New devices usually lack sufficient calibration data, and traditional fine-tuning methods require expensive standard data and are difficult to achieve zero-risk cold start. (3) Lack of physical interpretability: Purely data-driven "black box" models ignore the physical linear response principle behind instrument analysis, resulting in a lack of interpretability and metrological trust in quantitative analysis results. (4) Catastrophic forgetting caused by incremental updates: Full parameter updates of the model at the edge can easily overwrite the learned general knowledge, reducing the long-term robustness of the model.

[0053] The purpose of this application is to provide an instrument analysis migration method and system based on feature decoupling and physical residual calibration. It is mainly applied to intelligent quantification and long-term maintenance scenarios of analytical instruments such as chromatography, spectroscopy and mass spectrometry. In the process of instrument analysis migration, it takes into account both data feature decoupling and physical residual calibration. It can improve the accuracy of cross-instrument analysis while reducing the dependence on calibrated samples. It can realize the migration of instrument analysis models across devices with low sample cost and high robustness, and solve the problems of existing cross-instrument analysis migration, such as dependence on a large number of calibrated samples, lack of physical basis for calibration, low migration accuracy, and catastrophic forgetting caused by incremental updates.

[0054] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0055] The instrument analysis migration method based on feature decoupling and physical residual calibration provided in this application can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or placed in the cloud or on another server. Terminal 102 can send the raw instrument signal to be analyzed to server 104. After receiving the raw instrument signal, server 104 constructs a dual-stream deep neural network (DNN) model and performs feature decoupling pre-training. It calculates the feature correlation loss by adding an orthogonality penalty term and performs backpropagation. The raw instrument signal is then input into the trained dual-stream deep neural network model to extract the corresponding general chemical features and response bias features. Based on the general chemical features, a weighted aggregation is used to obtain the physical signal intensity characterization. Simultaneously, based on the linear response principle, a dynamic response factor is constructed by combining the general chemical features and response bias features to calculate the quantitative analysis results. Server 104 can then feed back the obtained quantitative analysis results to terminal 102. In addition, in some embodiments, the instrument analysis migration method based on feature decoupling and physical residual calibration can also be implemented by the server 104 or the terminal 102 separately. For example, the terminal 102 can directly perform instrument analysis migration processing on the original signal of the instrument to be analyzed, or the server 104 can obtain the original signal of the instrument to be analyzed from the data storage system and perform instrument analysis migration processing on the original signal of the instrument to be analyzed.

[0056] The terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, and IoT devices. The server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers, or it can be a cloud server.

[0057] In one exemplary embodiment, such as Figure 2 As shown, an instrument analysis migration method based on feature decoupling and physical residual calibration is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the following steps are included.

[0058] S1: Construct a two-stream deep neural network model. This model includes a shared feature branch and a private bias branch. The shared feature branch extracts cross-device general chemical features from instrument signals, while the private bias branch extracts response bias features specific to the respective device from the instrument signals. The shared feature branch and the private bias branch are configured asymmetrically. During the device-side adaptive evolution phase, the shared feature branch maintains parameter freeze, while the private bias branch is initialized to a zero-bias state, a near-zero small random state, or its output's contribution to the quantitative analysis results is suppressed in the initial state. This ensures that the initial inference results of the two-stream deep neural network model on new devices are primarily determined by the shared feature branch.

[0059] S2: Perform feature decoupling pre-training on the dual-stream deep neural network model, and calculate the feature correlation loss by adding an orthogonality penalty term to the total loss function and perform backpropagation during the training process to minimize the statistical correlation between the general chemical feature and the response deviation feature, thereby achieving feature decoupling and obtaining the trained dual-stream deep neural network model.

[0060] S3: Acquire the raw signal of the instrument to be analyzed and input it into the trained dual-stream deep neural network model to extract the corresponding general chemical features and response deviation features. Based on the general chemical features, obtain the physical signal intensity characterization through weighted aggregation. Simultaneously, based on the linear response principle, construct a dynamic response factor by combining the general chemical features and the response deviation features. Calculate the quantitative analysis result based on the physical signal intensity characterization and the dynamic response factor. The quantitative analysis result is the concentration value of the analyte corresponding to the raw signal of the instrument to be analyzed.

[0061] By implementing steps S1 to S3 above, and taking into account both data feature decoupling and physical residual calibration during the instrument analysis migration process, the accuracy of cross-instrument analysis can be improved while reducing the dependence on labeled samples. This enables low-sample-cost, highly robust instrument analysis model migration across devices, solving problems such as reliance on a large number of labeled samples, lack of physical basis for calibration, low migration accuracy, and catastrophic forgetting caused by incremental updates in existing cross-instrument analysis migration.

[0062] As an optional implementation, in step S2, the orthogonality penalty term is configured to calculate the cosine similarity, mutual information, or Frobenius norm between the general chemical feature matrix and the response deviation feature matrix.

[0063] As an optional implementation, in step S2, when performing feature decoupling pre-training on the dual-stream deep neural network model, the pre-training data used is a constructive synthetic dataset based on a physical dynamics model. The constructive synthetic dataset is constructed by superimposing simulated baseline drift, noise, and nonlinear deformation onto an ideal signal generated based on a Gaussian function, Lorentz function, or exponentially modified Gaussian function (EMG).

[0064] As an optional implementation, in step S3, the weighted aggregation is calculated using soft integration based on an attention mechanism, specifically including the following steps.

[0065] (1) Based on the general chemical characteristics, generate an attention weight sequence that is time-domain aligned with the original signal of the instrument to be analyzed.

[0066] (2) Multiply the attention weight sequence with the original signal of the instrument to be analyzed point by point and sum them to obtain the physical signal intensity characterization.

[0067] As an optional implementation, step S3 specifically includes the following steps.

[0068] S31: Acquire the raw signal from the instrument to be analyzed.

[0069] S32: Input the original signal of the instrument to be analyzed into the trained dual-stream deep neural network model, and extract the general chemical features and the response deviation features corresponding to the original signal of the instrument to be analyzed.

[0070] S33: Generate an attention weight sequence based on the general chemical characteristics, and perform a weighted summation operation on the original instrument signal to be analyzed according to the attention weight sequence to obtain a physical signal intensity value that characterizes the peak area, which is used as the physical signal intensity characterization.

[0071] S34: Predict the baseline response factor using the general chemical characteristics, and simultaneously predict the deviation response factor using the response deviation characteristics, and then superimpose the baseline response factor and the deviation response factor to generate the dynamic response factor under the current instrument state.

[0072] S35: Based on the principle of linear response, perform calculations on the physical signal intensity characterization and the dynamic response factor to obtain quantitative analysis results characterizing the concentration of the analyte.

[0073] As an optional implementation, after step S3, the instrument analysis migration method based on feature decoupling and physical residual calibration further includes step S4: device-side adaptive evolution, which specifically includes the following steps.

[0074] S41: During the online model operation phase, target samples are screened based on the confidence assessment of the current prediction results, and the calibration true value corresponding to the target sample is received from external input; wherein, the current prediction result is the quantitative analysis result corresponding to the original instrument signal to be analyzed obtained at present, the target sample is the original instrument signal sample to be analyzed whose confidence assessment value is greater than the preset calibration threshold and requires manual provision of the true concentration value to complete the model calibration update, and the calibration true value is the true concentration value.

[0075] S42: Based on the target sample and its calibration true value, incrementally update the parameters of the private bias branch while keeping the parameters of the shared feature branch frozen.

[0076] As an optional implementation, in step S41, the confidence assessment is achieved by calculating the uncertainty index predicted by the trained two-stream deep neural network model. The uncertainty index can be calculated by using Monte Carlo Dropout technology to perform multiple inference calculations to calculate the variance or by using a deep ensemble network to calculate the output dispersion.

[0077] As an optional implementation, in step S41, when receiving the calibration truth value corresponding to the target sample from an external input, a dual threshold screening and interactive prompting strategy is adopted, which specifically includes the following steps.

[0078] (1) Set a safety threshold and a calibration threshold, wherein the calibration threshold is greater than the safety threshold.

[0079] (2) When the confidence assessment value of the current prediction result is less than the safety threshold, the current sample is determined to be a normal and risk-free sample, and the corresponding quantitative analysis result is directly output; the current sample is the original instrument signal sample to be analyzed that is currently acquired and input.

[0080] (3) When the confidence assessment value of the current prediction result is greater than or equal to the safety threshold and less than or equal to the calibration threshold, the current sample is determined to be a medium-risk sample, the current sample is marked as a yellow warning signal and the corresponding quantitative analysis result is output.

[0081] (4) When the confidence assessment value of the current prediction result is greater than the calibration threshold, the current sample is determined to be a high-risk sample, a calibration prompt signal is generated and output, and calibration truth data in response to the calibration prompt signal is received; the calibration prompt signal is used to instruct external devices or interfaces to input the calibration truth data for the high-risk sample.

[0082] To make the technical solution of this application clearer, the specific implementation process of the technical solution of this application will be explained in detail below with examples.

[0083] This embodiment proposes an instrument analysis transfer method based on feature decoupling and physical residual calibration. First, it clarifies the specific construction method of the two-stream deep neural network model. For example... Figure 3 As shown, the network architecture is designed as an asymmetric two-branch structure, intended to process one-dimensional instrument signals (such as chromatograms and spectra) and output quantitative analysis results. The construction process of the two-stream deep neural network model includes the following.

[0084] (1) Input data preprocessing.

[0085] In this embodiment, the input received by the two-stream deep neural network model is a preprocessed one-dimensional time-series signal tensor X∈R. B×1×L Where B is the batch size and L is the signal length (e.g., L = 3000 sampling points). The preprocessing process includes the following operations.

[0086] Baseline correction processing: The asymmetric least squares smoothing (AsLS) algorithm is used to remove low-frequency background drift.

[0087] Normalization: The signal intensity is normalized to the maximum value (Min-Max Scaling) and mapped to the [0,1] interval to accelerate network convergence.

[0088] (2) Shared feature branches.

[0089] like Figure 4 As shown, the shared feature branch is used to extract common chemical features (i.e., the "fingerprint" information of a substance) across devices.

[0090] Network structure: A 5-layer stacked one-dimensional convolutional neural network (1D-CNN) is used.

[0091] First layer: kernel size k=7, stride s=2, output channels C=32, activation function is ELU. Larger kernels are used to capture broad-peak features.

[0092] Layers 2 through 4: kernel size k=3, stride s=1, and the number of channels doubles sequentially (64, 128, 256). Each layer is followed by batch normalization and max pooling.

[0093] The fifth layer: Global Average Pooling, which compresses temporal features into a fixed-dimensional general chemical feature vector. .

[0094] Parameter freezing strategy: After the model completes pre-training and is deployed to edge devices, all weight parameters of this branch will be frozen, meaning the gradient ▽θ will be frozen during backpropagation. shared =0. This ensures that the model will not forget the essential characteristics of matter due to deviations in adapting to new equipment.

[0095] (3) Private bias branch.

[0096] like Figure 4 As shown, the private bias branch is used to extract response bias characteristics specific to the device (such as tailing caused by column aging or decreased detector sensitivity). The network structure employs a lightweight multilayer perceptron (MLP).

[0097] Input: Receive the dimensionality-reduced features of the original signal X, which are the same as those of the shared feature branch (or directly reuse the output of the shallow layer of the shared feature branch).

[0098] Hidden layers: These consist of two fully connected layers with 64 and 32 nodes respectively, using the Tanh activation function. The Tanh output range is [-1, 1], suitable for simulating positive or negative instrument bias.

[0099] Output layer: Outputs the bias feature vector.

[0100] Zero-bias initialization: To achieve a "cold start" for the new device, the final layer weight W of this branch... last and bias b last Initialized to a local minimum, such as a Gaussian distribution N(0, 10) -6 This means that before any calibration is performed, the output of this branch is close to zero, the model degenerates into a standard model, and the prediction noise caused by random initialization is avoided.

[0101] After constructing the two-stream deep neural network model, this embodiment begins the generation of constructive synthetic data and orthogonal decoupling training. To train the two-stream deep neural network model to achieve feature decoupling, this embodiment constructs a synthetic data generator based on physical equations and designs a loss function that includes orthogonal penalties, specifically including the following:

[0102] (1) Constructed synthetic datasets.

[0103] To address the problem of insufficient training data, this embodiment utilizes a physical model to generate massive amounts of labeled data. When constructing a constructive synthetic dataset, it superimposes simulated baseline drift, noise, and nonlinear deformation onto an ideal signal generated based on a Gaussian function, Lorentz function, or exponentially modified Gaussian function.

[0104] In this embodiment, an exponentially modified Gaussian function is used to generate the ideal chromatographic peak when the ideal signal is generated, as expressed in the following formula.

[0105] .

[0106] in, Represents the peak height function. The signal represents the intensity parameter (desired peak area) when synthesizing an ideal signal; t is time; μ is the peak center retention time; σ is the peak width; This is a tailing factor. A simulated response factor is introduced when generating the constructed synthetic dataset. Set the true concentration label for this sample. This allows for the construction of paired training data. .

[0107] Then, device deviation injection is performed, including the following:

[0108] Nonlinear response: Simulating detector saturation effect, for Transform = ,in, .

[0109] Baseline drift: Superimposed with a low-frequency sine wave, represented as y drift =asin(2πft).

[0110] Among them, y drift This represents the numerical value of the baseline drift signal. This is to simulate low-frequency fluctuations in the background signal (such as baseline undulations caused by temperature changes) during long-term instrument operation. 'a' represents the drift amplitude, which determines the maximum intensity of the baseline fluctuation. π is the mathematical constant pi; in this embodiment, π is taken as 3.14159. 'f' represents the drift frequency. Due to the low-frequency sine wave mentioned, the value of 'f' here is usually very small, representing slowly changing fluctuations. 't' represents time, corresponding to the time coordinate of the instrument signal at the sampling point.

[0111] Noise superposition: Gaussian white noise with a signal-to-noise ratio (SNR) between 10dB and 60dB is added.

[0112] Through the above steps, pairs of data are generated. This covers a wide range of extreme instrument conditions. Among them, This represents the synthesized instrument input signal. It is achieved by superimposing the ideal signal (EMG function) with the baseline drift y. drift "Dirty" data generated by random noise and nonlinear deformation is used as input to a neural network to simulate instrument spectra under various harsh conditions in the real world. This indicates the true concentration label, which corresponds to the intensity parameter when synthesizing the ideal signal. Divide by the simulated response factor As the prediction target / label of a neural network, the model needs to learn how to interpret complex... Accurately reproduce .

[0113] (2) Feature decoupling loss function.

[0114] The goal of training is to force It contains only chemical information, while It only includes equipment deviation. Therefore, a total loss function is defined. .in, This represents a universal chemical feature vector. It is extracted from a shared feature branch. Physically, it represents the "chemical fingerprint" information (such as retention time and characteristic peak shape) of the analyte itself. It is designed to be device-independent, meaning that the same substance can be analyzed on any machine. They should be the same. This represents the response bias feature vector. It is extracted from the Private Branch. Its physical meaning represents the instrument status of a specific device (e.g., column aging, detector sensitivity degradation, electronic noise level). It is specifically designed to capture error information unique to this particular machine.

[0115] Total loss function The expression is as follows.

[0116] .

[0117] in, Indicates the prediction error, and =( - ) 2 This ensures quantitative accuracy. This represents the concentration value predicted by the model. This is the final quantitative analysis result output by the deep neural network (i.e., the value calculated using C=I / K). Its function is to represent the concentration value of the analyte in the sample as predicted by the model. This represents the actual concentration value (label). Source: The actual concentration parameter set when generating this sample in the synthetic dataset, or the true value manually entered during the calibration phase. It serves as the standard answer, used to measure the accuracy of the model's predictions. and These are the corresponding weights. For sparsity constraints, Applying L1 regularization encourages the model to automatically close private bias branches when the instrument is in good condition (no bias). The orthogonality penalty is represented by the cosine similarity or Frobenius norm of two feature matrices, calculated to remove correlation between features. This is expressed as the following formula.

[0118] .

[0119] in, Let Frobenius norm be denoted, and T denote the transpose. The orthogonality penalty term forces two eigenvectors to be perpendicular to each other in the feature space, i.e., statistically independent.

[0120] Then, this embodiment performs quantitative inference based on physically-aware residual calibration. This embodiment clarifies how the model combines deep learning features with physical laws (such as Lambert-Beer law or chromatographic linearity principles) for calculation during the inference phase. Unlike traditional "black-box" models that directly regress concentrations, this application adopts a "white-box" approach of "physical parameter prediction," including the following:

[0121] (1) Physical signal intensity extraction.

[0122] like Figure 5 As shown, this embodiment utilizes shared features. Generate attention mask sequence .

[0123] In this embodiment, when extracting the physical signal strength, the physical signal strength is first calculated. (e.g., peak area), the calculation formula is as follows.

[0124] .

[0125] in, This represents the original signal strength of the instrument at time t. Specifically, it refers to the signal strength at time t. At each sampling point, the instrument acquires the voltage or absorbance value (which has been preprocessed). It is a time-series signal tensor. The first in Each element represents the raw data (including both signal and noise) seen by the instrument. This represents the attention mask value (weight) at time t. It is a value between 0 and 1. ∈[0, 1]. It is automatically generated by the model based on "general chemical characteristics". The representative model considers the first... The probability that each sampling point is a valid signal. If A value close to 1 indicates that the model considers this point to be a peak of matter, which should be retained. If... A value close to 0 indicates that the model considers this point to be noise or baseline, and it needs to be filtered out.

[0126] Compared to traditional hard thresholding, this soft integration method can handle overlapping peaks and noise interference more robustly.

[0127] (2) Construction of dynamic response factors.

[0128] In this embodiment, the baseline response factor Predicted by the shared feature branch, representing the unit concentration response value on an ideal standard instrument. Reference Response Factor The calculation formula is as follows.

[0129] .

[0130] in, This represents a multilayer perceptron (regression prediction head) on the shared feature branch. It's a small, fully connected neural network module connected at the end of the "shared feature branch." Its task is to perform a mapping, transferring the abstract chemical feature vectors... Translated into specific physical parameters (i.e., response factors under ideal conditions) ).

[0131] In this embodiment, the deviation response factor Predicted by the private bias branch, representing the current sensitivity drift of the device. Bias response factor. The calculation formula is as follows.

[0132] .

[0133] in, This represents a multilayer perceptron (regression prediction head) on the private bias branch. This is a small neural network module connected to the end of the private bias branch. Its task is to process the device bias feature vector... Translated into specific numerical values, namely the amount of sensitivity drift.

[0134] In this embodiment, during dynamic synthesis, the calculation formula for the actual dynamic response factor under the current instrument state is as follows.

[0135] .

[0136] in, This represents the dynamic response factor. This is the calibration coefficient ultimately used for quantitative calculations, composed of the "ideal reference value" and the "deviation calibration value." It represents the true sensitivity of this specific instrument under its current conditions. Its ultimate application is to determine the final concentration based on physical laws (such as Lambert-Beer's law). By physical signal strength Divide by dynamic response factor The result is calculated and expressed as the following formula.

[0137] .

[0138] (3) Calculation of quantitative analysis results.

[0139] Based on the principle of linear response Calculate the final predicted concentration value. The calculation formula is as follows.

[0140] .

[0141] in, This represents the predicted concentration value. This is the final concentration value of the analyte output by the model. This represents the physical signal intensity (usually referring to the peak area or absorbance integral). It is a physical quantity representing the total amount of matter extracted from the original signal by the two-stream network through "soft integration". This is the baseline response factor. Predicted by the shared characteristic branch. It represents the signal response (i.e., standard sensitivity) that a unit concentration of substance should produce under "ideal, standard" instrument conditions. This is the bias response factor. Predicted by a proprietary bias branch. It represents how much the current, specific device has "drifted" relative to its ideal state (e.g., detector aging leading to decreased sensitivity). (Possibly a negative value). + The overall dynamic response factor represents the actual sensitivity at the current moment and under the current device conditions. The model dynamically reconstructs the true physical response coefficient K by adjusting the baseline value. This formula ensures that even if a branch of the deep neural network predicts a small fluctuation, the final result still strictly adheres to physical constraints, greatly improving the interpretability and econometric reliability of the results.

[0142] Then, this embodiment implements the edge-end adaptive evolution and human-computer interaction process. This embodiment deploys the model on an instrument embedded system (such as an edge computing box or industrial control computer) for online operation. Figure 6 As shown, this process achieves dual-threshold screening through confidence assessment, establishing a secure human-computer interaction mechanism. Specifically, it includes the following:

[0143] (1) Uncertainty assessment.

[0144] This embodiment employs the Monte Carlo dropout technique, keeping the Dropout layer enabled in the shared feature branch during the inference phase (with a dropout rate set to 0.1).

[0145] Perform forward inference N times (e.g., N=10) on the same input sample to obtain N predicted concentration values ​​{c1, c2, ..., c N The coefficient of variation of the predicted results is calculated as a confidence index. , expressed as the following formula.

[0146] .

[0147] Where c represents a set of predicted concentration values ​​obtained from N repeated inferences (e.g., 10 times) using the Monte Carlo loss technique on the same input sample. Its form is c = {c1, c2, ..., c...} N Because Dropout is enabled, the structure of the neural network undergoes slight random changes during each inference, so the outputs of these N iterations will not be exactly the same, but will follow a distribution. This represents the arithmetic mean of N predictions. It represents the model's final estimate of the substance's concentration (the most likely prediction). This represents the dispersion of the N prediction results. It measures the uncertainty of the model regarding this prediction. If... A very small value indicates that the predictions across N iterations are very close, suggesting the model is very "robust." If The large value indicates that the model's predictions fluctuate wildly, suggesting that the model is "hesitant." The confidence level indicator, specifically the coefficient of variation, is calculated by dividing the standard deviation by the mean. It eliminates the influence of absolute numerical values ​​and represents relative error. Through use... It allows for a fair comparison of confidence levels at different concentrations. The smaller the value, the lower the uncertainty and the higher the confidence level (i.e., normal sample).

[0148] (2) Dual threshold filtering and interactive prompting strategy.

[0149] In this embodiment, two thresholds are preset, including a safety threshold T. safe (e.g., 5%) and calibration threshold T calib (e.g., 15%). And corresponding to the following different scenarios.

[0150] Scenario A: U<T safe (Normal, risk-free sample).

[0151] At this point, the system determines that the instrument is stable and the sample characteristics are within the model distribution. It then directly outputs the quantitative analysis results without requiring manual intervention.

[0152] Scene B: T safe ≤U≤T calib (Medium-risk sample).

[0153] The quantitative analysis results are output at this time, but a yellow warning signal is displayed. Users are advised to run quality control samples (QC) during subsequent idle periods.

[0154] Scenario C: U > T calib(High-risk sample).

[0155] At this point, the system determines that it has encountered unseen interference or that the instrument has experienced severe drift (such as a light source malfunction). The output is then blocked, and a "Request Calibration" dialog box pops up on the human-machine interface.

[0156] (3) Online incremental update.

[0157] When the user responds to the request in "Scenario C", they input the actual concentration value of the sample. Then, the system performs the following updates: Constructs sample pairs: Lock shared feature branches and set Requires. Grad θ share =False; Unlocks the private deviation branch and sets Requires. Grad θ private =True.

[0158] Here, θ represents the model parameters, which include all the weights and biases in the neural network. These parameters determine how the network extracts features and makes predictions. share This represents all the parameters of the shared feature branch, referring to the weights of the 1D-CNN network portion used to extract "general chemical features." These parameters store knowledge of the substance's "chemical fingerprint," are cross-device universal, and should not change with device variations. θ private This represents all parameters of the private bias branch, referring to the weights of the MLP network portion used to extract "device response bias." These parameters are specifically responsible for recording the current machine's error and drift and need to be adjusted based on the latest calibration data. Requires Grad This indicates whether gradient calculation is required (i.e., whether training or updating is allowed). During the backpropagation process of a neural network, only parameters that "require gradient calculation" will be modified by the optimization algorithm. False indicates "frozen," and the parameter values ​​remain unchanged. True indicates "unfrozen" or "activated." The system automatically calculates the gradient and updates these parameters based on the error. When performing gradient descent, the parameter update module only updates the parameters of the private bias branch, not the parameters of the shared feature branch, thereby minimizing the prediction error. Since the number of parameters in the private bias branch is usually extremely small, this process can be completed within 100ms on edge devices. This means that the instrument can continuously "evolve" during use, automatically adapting to the aging characteristics of the device with minimal user interaction.

[0159] As an optional implementation, this embodiment also provides a device supporting the above method, which can be integrated into the control terminal of a liquid chromatograph (HPLC), gas chromatograph (GC), or mass spectrometer (MS). This device mainly includes the following components.

[0160] Data acquisition and preprocessing unit: Includes a high-precision analog-to-digital converter (ADC) and FPGA chip, used to acquire voltage signals from instrument detectors in real time and perform hardware-level filtering and noise reduction.

[0161] Edge inference computing unit: Employs an embedded processor with a tensor accelerator (NPU) (such as ARM Cortex-A76+NPU).

[0162] Storage partitioning: The memory is divided into a "read-only area" and a "read-write area". Pre-trained shared feature branch weights are stored in the read-only area to prevent tampering by viruses or erroneous operations; private bias branch weights are stored in the read-write area to support online updates.

[0163] Interactive calibration unit: Runs on the host computer software or touch screen UI. It is responsible for real-time monitoring of the confidence level U of the inference paper. When U exceeds the limit, it triggers an audible and visual alarm and pops up a window to guide the user to perform a single-point calibration operation.

[0164] Based on the same inventive concept, this application also provides an instrument analysis migration system based on feature decoupling and physical residual calibration for implementing the instrument analysis migration method based on feature decoupling and physical residual calibration described above. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations in the instrument analysis migration system embodiments based on feature decoupling and physical residual calibration provided below can be found in the limitations of the instrument analysis migration method based on feature decoupling and physical residual calibration described above, and will not be repeated here.

[0165] In one exemplary embodiment, such as Figure 7 As shown, an instrument analysis migration system based on feature decoupling and physical residual calibration is provided, which specifically includes the following modules.

[0166] A model building module is used to construct a two-stream deep neural network model. The two-stream deep neural network model includes a shared feature branch and a private bias branch. The shared feature branch is used to extract general chemical features across instruments from instrument signals, and the private bias branch is used to extract response bias features of the corresponding instruments from instrument signals. The shared feature branch and the private bias branch are configured asymmetrically. The parameters of the shared feature branch are frozen during the adaptive evolution phase at the instrument end, and the private bias branch is initialized to a zero-bias state, a near-zero small random state, or its output contribution to the quantitative analysis results is suppressed in the initial state.

[0167] The feature decoupling pre-training module is used to perform feature decoupling pre-training on the two-stream deep neural network model. During the training process, the feature correlation loss is calculated by adding an orthogonality penalty term to the total loss function and backpropagation is performed to minimize the statistical correlation between the general chemical feature and the response deviation feature, thereby obtaining the trained two-stream deep neural network model.

[0168] The signal acquisition module is used to acquire the raw signals from the instrument to be analyzed.

[0169] The dual-stream inference module is used to input the raw signal of the instrument to be analyzed into the trained dual-stream deep neural network model to extract the corresponding general chemical features and response deviation features, respectively. Based on the general chemical features, a physical signal intensity characterization is obtained through weighted aggregation. Simultaneously, based on the linear response principle, a dynamic response factor is constructed by combining the general chemical features and the response deviation features. The quantitative analysis result is calculated based on the physical signal intensity characterization and the dynamic response factor. The quantitative analysis result is the concentration value of the analyte corresponding to the raw signal of the instrument to be analyzed.

[0170] As an optional implementation, the dual-stream inference module includes a shared feature encoding unit storing fixed parameters and a private bias encoding unit storing updatable parameters. It is used to obtain a physical signal intensity characterization based on the general chemical features through weighted aggregation. At the same time, based on the linear response principle, it constructs a dynamic response factor by combining the general chemical features and the response bias features, and outputs quantitative analysis results according to the physical signal intensity characterization and the dynamic response factor.

[0171] As an optional implementation, the instrument analysis migration system based on feature decoupling and physical residual calibration also includes the following modules.

[0172] An adaptive evolution module is used during the online model run phase to screen target samples based on the confidence assessment of the current prediction results and receive externally input calibration ground truth values ​​corresponding to the target samples; and to incrementally update the parameters of the private bias branch based on the target samples and the calibration ground truth values, while keeping the parameters of the shared feature branch frozen; wherein, the current prediction result is the quantitative analysis result corresponding to the original instrument signal to be analyzed obtained at present, the target sample is the original instrument signal sample to be analyzed whose confidence assessment value is greater than a preset calibration threshold and requires manual provision of a true concentration value to complete the model calibration update, and the calibration ground truth value is the true concentration value.

[0173] The adaptive evolution module is mainly used to calculate the confidence level of the current prediction result, determine the target sample to be calibrated according to the dual threshold screening and interactive prompting strategy, and update the parameters of the private bias coding unit online using the supervision data of the target sample.

[0174] The instrument analysis migration method and system proposed in this application based on feature decoupling and physical residual calibration have the following advantages.

[0175] (1) Deep decoupling and robust migration of features are achieved: By forcibly separating the general chemical features and the device response deviation through orthogonality constraints, the shared feature branches can extract "chemical fingerprints" that do not age with the device, effectively avoiding the negative migration phenomenon when migrating across devices.

[0176] (2) Solved the cold start problem and reduced data costs: The zero-bias initialization strategy of private bias branch is adopted, which enables new equipment to perform "zero-sample inference" based on shared feature branch without any calibration data; at the same time, only a small amount of calibration data is needed to complete the adaptation by updating the lightweight private bias branch, which greatly reduces the operation and maintenance costs.

[0177] (3) Improved the interpretability and physical consistency of the model: Combining deep learning feature extraction with the principle of physical linear response ensures that the quantitative calculation process conforms to the physical principle of instrumental analysis and improves the metrological credibility of the results.

[0178] (4) Support for long-term adaptation at low computing power edge: By freezing large-scale shared feature branches and updating only small-scale private bias branches, the computing power requirement for edge training is greatly reduced, catastrophic forgetting is prevented, and the instrument is able to have self-calibration and evolution capabilities throughout its entire life cycle.

[0179] (5) A secure human-computer interaction closed loop was constructed: the confidence-based dual threshold screening and interactive prompting strategy enables the model to actively identify high-risk samples and request human intervention, ensuring high reliability and security in industrial application scenarios.

[0180] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0181] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. An instrument analysis migration method based on feature decoupling and physical residual calibration, characterized in that, The instrument analysis migration method based on feature decoupling and physical residual calibration includes: A two-stream deep neural network model is constructed. This model includes a shared feature branch and a private bias branch. The shared feature branch extracts cross-device general chemical features from instrument signals, while the private bias branch extracts response bias features specific to the respective device from the instrument signals. The shared feature branch and the private bias branch are configured asymmetrically. During the device-side adaptive evolution phase, the shared feature branch maintains parameter freezing. The private bias branch is initialized to a zero-bias state, a near-zero small random state, or its output's contribution to the quantitative analysis results is suppressed in the initialization state. The dual-stream deep neural network model is pre-trained with feature decoupling. During the training process, the feature correlation loss is calculated by adding an orthogonality penalty term to the total loss function and backpropagation is performed to minimize the statistical correlation between the general chemical feature and the response bias feature, thereby obtaining the trained dual-stream deep neural network model. The raw signal of the instrument to be analyzed is acquired and input into the trained dual-stream deep neural network model to extract the corresponding general chemical features and response deviation features. Based on the general chemical features, a weighted aggregation is used to obtain the physical signal intensity characterization. Simultaneously, based on the linear response principle, a dynamic response factor is constructed by combining the general chemical features and the response deviation features. The quantitative analysis result is calculated based on the physical signal intensity characterization and the dynamic response factor. The quantitative analysis result is the concentration value of the analyte corresponding to the raw signal of the instrument to be analyzed.

2. The instrument analysis migration method based on feature decoupling and physical residual calibration according to claim 1, characterized in that, The orthogonality penalty term is configured to calculate the cosine similarity, mutual information, or Frobenius norm between the general chemical feature matrix and the response bias feature matrix.

3. The instrument analysis migration method based on feature decoupling and physical residual calibration according to claim 1, characterized in that, When performing feature decoupling pre-training on the dual-stream deep neural network model, the pre-training data used is a constructive synthetic dataset based on a physical dynamics model. The constructive synthetic dataset is constructed by superimposing simulated baseline drift, noise, and nonlinear deformation onto an ideal signal generated based on a Gaussian function, a Lorentz function, or an exponentially modified Gaussian function.

4. The instrument analysis migration method based on feature decoupling and physical residual calibration according to claim 1, characterized in that, The weighted aggregation employs soft integral calculation based on an attention mechanism, specifically including: Based on the general chemical characteristics, an attention weight sequence is generated that is time-domain aligned with the original signal of the instrument to be analyzed; The attention weight sequence is multiplied point by point with the original signal of the instrument to be analyzed and then summed to obtain the physical signal intensity characterization.

5. The instrument analysis migration method based on feature decoupling and physical residual calibration according to claim 1, characterized in that, The raw instrument signal to be analyzed is acquired and input into the trained dual-stream deep neural network model to extract the corresponding general chemical features and response deviation features. Based on the general chemical features, a weighted aggregation is used to obtain a physical signal intensity characterization. Simultaneously, based on the linear response principle, a dynamic response factor is constructed by combining the general chemical features and the response deviation features. Quantitative analysis results are calculated based on the physical signal intensity characterization and the dynamic response factor, specifically including: Acquire the raw signals from the instrument to be analyzed; The original signal of the instrument to be analyzed is input into the trained dual-stream deep neural network model to extract the general chemical features and response deviation features corresponding to the original signal of the instrument to be analyzed. An attention weight sequence is generated based on the general chemical characteristics, and the original instrument signal to be analyzed is weighted and summed according to the attention weight sequence to obtain the physical signal intensity value characterizing the peak area, which is used as the physical signal intensity characterization. The reference response factor is predicted using the general chemical characteristics, and the deviation response factor is predicted using the response deviation characteristics. The reference response factor and the deviation response factor are then superimposed to generate the dynamic response factor under the current instrument state. Based on the principle of linear response, calculations are performed on the physical signal intensity characterization and the dynamic response factor to obtain quantitative analysis results characterizing the concentration of the analyte.

6. The instrument analysis migration method based on feature decoupling and physical residual calibration according to claim 1, characterized in that, After the step of calculating the quantitative analysis results based on the physical signal intensity characterization and the dynamic response factor, the instrument analysis migration method based on feature decoupling and physical residual calibration also includes the step of device-side adaptive evolution. The device-side adaptive evolution specifically includes: During the online model operation phase, target samples are screened based on the confidence assessment of the current prediction results, and the calibration true value corresponding to the target sample is received from external input; wherein, the current prediction result is the quantitative analysis result corresponding to the original instrument signal to be analyzed obtained at present, the target sample is the original instrument signal sample to be analyzed whose confidence assessment value is greater than the preset calibration threshold and requires manual provision of the true concentration value to complete the model calibration update, and the calibration true value is the true concentration value; Based on the target sample and its calibration true value, the parameters of the private bias branch are incrementally updated, while the parameters of the shared feature branch are kept frozen.

7. The instrument analysis migration method based on feature decoupling and physical residual calibration according to claim 6, characterized in that, The confidence assessment is achieved by calculating the uncertainty index predicted by the trained dual-stream deep neural network model. The uncertainty index is calculated by using Monte Carlo loss technique to perform multiple inferences to calculate the variance or by using a deep ensemble network to calculate the output dispersion.

8. The instrument analysis migration method based on feature decoupling and physical residual calibration according to claim 6, characterized in that, When receiving the calibration truth value corresponding to the target sample from external input, a dual-threshold filtering and interactive prompting strategy is adopted, specifically including: Set a safety threshold and a calibration threshold, wherein the calibration threshold is greater than the safety threshold; When the confidence assessment value of the current prediction result is less than the safety threshold, the current sample is determined to be a normal, risk-free sample, and the corresponding quantitative analysis result is directly output; the current sample is the original instrument signal sample to be analyzed that is currently acquired and input. When the confidence assessment value of the current prediction result is greater than or equal to the safety threshold and less than or equal to the calibration threshold, the current sample is determined to be a medium-risk sample, the current sample is marked with a yellow warning signal, and the corresponding quantitative analysis result is output. When the confidence assessment value of the current prediction result is greater than the calibration threshold, the current sample is determined to be a high-risk sample, a calibration prompt signal is generated and output, and calibration truth data input in response to the calibration prompt signal is received; the calibration prompt signal is used to instruct external devices or interfaces to input the calibration truth data for the high-risk sample.

9. An instrument analysis migration system based on feature decoupling and physical residual calibration, characterized in that, The instrument analysis migration system based on feature decoupling and physical residual calibration is used to implement the instrument analysis migration method based on feature decoupling and physical residual calibration as described in any one of claims 1-8. The instrument analysis migration system based on feature decoupling and physical residual calibration includes: A model building module is used to construct a two-stream deep neural network model. The two-stream deep neural network model includes a shared feature branch and a private bias branch. The shared feature branch is used to extract general chemical features across instruments from instrument signals, and the private bias branch is used to extract response bias features of the corresponding instruments from instrument signals. The shared feature branch and the private bias branch are configured asymmetrically. The parameters of the shared feature branch are frozen during the adaptive evolution phase at the instrument end, and the private bias branch is initialized to a zero-bias state, a near-zero small random state, or its output contribution to the quantitative analysis results is suppressed in the initial state. The feature decoupling pre-training module is used to perform feature decoupling pre-training on the dual-stream deep neural network model. During the training process, the feature correlation loss is calculated by adding an orthogonality penalty term to the total loss function and backpropagation is performed to minimize the statistical correlation between the general chemical feature and the response bias feature, thereby obtaining the trained dual-stream deep neural network model. The signal acquisition module is used to acquire the raw signals from the instrument to be analyzed. The dual-stream inference module is used to input the raw signal of the instrument to be analyzed into the trained dual-stream deep neural network model to extract the corresponding general chemical features and response deviation features, respectively. Based on the general chemical features, a physical signal intensity characterization is obtained through weighted aggregation. Simultaneously, based on the linear response principle, a dynamic response factor is constructed by combining the general chemical features and the response deviation features. The quantitative analysis result is calculated based on the physical signal intensity characterization and the dynamic response factor. The quantitative analysis result is the concentration value of the analyte corresponding to the raw signal of the instrument to be analyzed.

10. The instrument analysis migration system based on feature decoupling and physical residual calibration according to claim 9, characterized in that, The instrument analysis migration system based on feature decoupling and physical residual calibration also includes: An adaptive evolution module is used during the online model run phase to screen target samples based on the confidence assessment of the current prediction results and receive externally input calibration ground truth values ​​corresponding to the target samples; and to incrementally update the parameters of the private bias branch based on the target samples and the calibration ground truth values, while keeping the parameters of the shared feature branch frozen; wherein, the current prediction result is the quantitative analysis result corresponding to the original instrument signal to be analyzed obtained at present, the target sample is the original instrument signal sample to be analyzed whose confidence assessment value is greater than a preset calibration threshold and requires manual provision of a true concentration value to complete the model calibration update, and the calibration ground truth value is the true concentration value.

Images