A rapid method for detecting the content of active collagen

Abstract

This invention relates to the field of biological detection signal analysis technology, and discloses a method for rapidly detecting the content of active collagen. The method includes acquiring the raw optical response signal of the reaction between the test solution and a specific reagent; constructing a spatiotemporally continuous primary reaction trajectory through multi-level slicing and cross-domain splicing in the time and frequency domains; separating a set of activity-related signal patterns using convolutional kernel groups for depth scanning; constructing a graph association network between signal patterns based on this set; and generating a steady-state feature map characterizing the reaction kinetics through iterative calculation. This feature map is mapped to a virtual reaction space defined by standard data, and quantitative determination is achieved by calculating the distance between the projection vector and a preset concentration anchor point. This method enhances the sensitivity and anti-interference capability of detection through deep analysis of the reaction dynamic signal and association network modeling.

Description

Technical Field

[0001] This invention relates to the field of biological detection signal analysis technology, specifically a method for rapidly detecting the content of active collagen. Background Technology

[0002] Currently, the detection of active collagen content largely relies on colorimetric methods, enzyme-linked immunosorbent assays (ELISA), or single-probe-based fluorescence spectrometry. These methods typically collect optical signals at specific reaction endpoints or limited time points as the basis for quantification. Essentially, they simplify complex dynamic biochemical reactions into one or a few static data points for analysis. This approach utilizes only fragmentary information from reaction kinetics and cannot fully capture all the dynamic characteristics generated by the specific binding of active collagen to reagents throughout the continuous process from reaction initiation to equilibrium, especially those weak or transient signals that only appear at specific time-frequency dimensions. This limits the sensitivity and early identification ability for detecting targets in low-concentration samples or complex matrices.

[0003] To address interference from complex sample matrices, existing techniques often employ physical separation, background subtraction, or correction methods based on fixed wavelength ratios. These methods typically presuppose that the interfering signal and the target signal can be easily separated in the spectral or time domains, or rely on prior mathematical models for noise reduction. However, in actual detection, signals generated by environmental noise and non-specific binding of impurities often exhibit high overlap with the target active ingredient's signal in both the time and frequency domains, displaying a non-linear coupling relationship. Simply filtering out specific frequency bands or performing linear background subtraction can easily distort the target signal or leave residual interference, making it difficult to thoroughly and faithfully separate the signal pattern purely related to the concentration of active collagen from the original response signal, thus affecting the accuracy and repeatability of quantitative results. Summary of the Invention

[0004] The purpose of this invention is to provide a method for rapidly detecting the content of active collagen, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a method for rapidly detecting the content of active collagen, the method comprising:

[0006] The original response signal of the reagent that undergoes a specific optical reaction with the collagen solution under excitation light is collected, and the dynamic signal spectrum formed by the combined effect of solution state and active ingredients is captured.

[0007] The original response signal is sliced ​​in multiple levels, and signal segments are extracted from the time domain, frequency domain, and time-frequency domain respectively. The signal segments from different domains are spliced ​​and superimposed according to the preset phase to construct a primary response trajectory with spatiotemporal continuity.

[0008] The primary reaction trajectory is imported into the convolution kernel group for depth scanning. The signal patterns related to the changes in active collagen concentration and the background patterns composed of environmental noise and impurities are separated through the parallel operation of the convolution kernel group, thus obtaining a set of background-free active signal patterns.

[0009] Based on the set of active signal patterns, a graph structure is used to construct an association network between different signal patterns. Nodes represent specific signal patterns, and edge weights represent the transition probability and intensity association between patterns. The network node states are updated through iterative calculation until convergence, generating a steady-state feature map that characterizes the reaction dynamics within the sample.

[0010] The steady-state feature map is mapped to a virtual reaction space, the coordinate axes of which are defined based on the historical reaction trajectories of the standard sample library, including: reading the set of historical steady-state feature maps generated by standard samples of known concentrations under the same detection conditions from the standard sample library;

[0011] Calculate the pairwise distance matrix between all feature maps in the historical steady-state feature map set, where the distance is a standardized Euclidean distance.

[0012] The distance matrix is ​​centered, and the covariance matrix is ​​calculated.

[0013] Calculate the eigenvalues ​​and eigenvectors of the covariance matrix, sort the eigenvalues ​​from largest to smallest, and select the eigenvectors corresponding to the largest and second largest eigenvalues ​​as the directions of the two principal coordinate axes of the virtual reaction space;

[0014] Define the origin and scale of the virtual reaction space;

[0015] The projection vector of the steady-state characteristic map in the virtual reaction space is calculated, and the content level of active collagen is determined based on the distance between the projection vector and the preset standard concentration anchor point.

[0016] Preferably, the step of multi-level slicing the original response signal, extracting signal segments from the time domain, frequency domain, and time-frequency domain respectively, and splicing and superimposing the signal segments from different domains according to a preset phase to construct a primary response trajectory with spatiotemporal continuity includes:

[0017] The original response signal is segmented by a sliding window to generate time-domain signal slices, and each time-domain signal slice is normalized to eliminate baseline drift.

[0018] Wavelet transform is applied to each normalized time-domain signal slice to decompose it into frequency domain coefficients of multiple frequency bands. Frequency band coefficients with energy higher than the threshold are selected to form frequency domain signal segments.

[0019] Hilbert-Huang transform is performed synchronously on each time-domain signal slice to obtain the instantaneous frequency and instantaneous amplitude of the signal, forming a joint time-frequency domain distribution segment;

[0020] A fusion time scale is established to time-align time-domain signal slices, frequency-domain signal segments, and time-frequency domain joint distribution segments. Based on the phase information revealed by the signal in the time-frequency domain joint distribution segments, the signal data of the three domains are cross-indexed and spliced ​​to generate a continuous primary response trajectory.

[0021] Preferably, the primary reaction trajectory is imported into a convolutional kernel group for depth scanning. Through parallel computation of the convolutional kernel group, signal patterns related to changes in active collagen concentration and background patterns composed of environmental noise and impurities are separated to obtain a background-free set of active signal patterns, including:

[0022] Construct a multi-scale convolution kernel group, which contains convolution kernels targeting high-frequency details, mid-frequency contours, and low-frequency trends respectively, and use the multi-scale convolution kernel group to perform parallel convolution operations on the primary response trajectory;

[0023] Correlation analysis was performed on the convolutional feature maps output by the multi-scale convolutional kernel group. The feature channels that are highly correlated with the overall change trend of the primary reaction trajectory were classified as target feature channels, and the remaining feature channels were classified as background feature channels.

[0024] Spatial pooling is performed on the data classified as target feature channels to compress their dimensionality while preserving their statistical significance, forming the target feature vector;

[0025] The data classified as background feature channels are reverse reconstructed to generate a simulated background signal trajectory. The simulated background signal trajectory is subtracted from the original primary response trajectory to obtain the preliminary denoised signal trajectory.

[0026] The signal trajectories that have been initially denoised are then subjected to feature clustering again. The signal waveforms represented by the cluster centers are defined as active signal patterns. All cluster centers are then aggregated to obtain a set of active signal patterns with background removed.

[0027] Preferably, the method involves constructing a network of connections between different signal patterns based on a set of active signal patterns using a graph structure. Nodes represent specific signal patterns, and edge weights represent the transition probabilities and intensity correlations between patterns. The network node states are updated iteratively until convergence, generating a steady-state feature map characterizing the internal reaction dynamics of the sample. This includes:

[0028] Each signal pattern in the set of active signal patterns is treated as a network node, and the state value of all nodes is initialized to the initial energy amplitude of its corresponding signal pattern.

[0029] Calculate the conditional probability of the signal patterns represented by any two network nodes appearing successively in the original primary response trajectory, and use the product of the conditional probability and the average energy of the two patterns as the initial weight of the edge connecting the two nodes.

[0030] The adjacency matrix of the network is constructed based on the edge weights, and the degree centrality of each node is calculated based on the adjacency matrix. The degree centrality is then normalized and used as the update coefficient of the node state.

[0031] Design a diffusion-aggregation iteration rule. In each iteration, each node diffuses part of its state value to its neighboring nodes according to the outgoing edge weights, while aggregating the incoming state values ​​from the neighboring nodes and updating the state values ​​in combination with its own update coefficient.

[0032] When the change in the state values ​​of all nodes is less than the convergence threshold for two consecutive iterations, the iteration stops, and the final state values ​​of all nodes are arranged in the temporal order of their corresponding signal modes to generate a steady-state feature map.

[0033] Preferably, the calculation of the projection vector of the steady-state feature map in the virtual reaction space, and the determination of the content level of active collagen based on the distance between the projection vector and the preset standard concentration anchor point, includes:

[0034] The steady-state feature map of the sample to be tested is projected onto the two principal coordinate axes of the virtual reaction space, and its projection coordinates are calculated to form a two-dimensional projection vector.

[0035] In the virtual reaction space, standard concentration anchor points corresponding to the steady-state characteristic maps of standard samples at different concentration levels are pre-marked;

[0036] Calculate the Euclidean distance from the two-dimensional projection vector of the sample to be tested to each standard concentration anchor point, and determine the concentration level corresponding to the standard concentration anchor point with the smallest distance as the content level of active collagen in the sample to be tested.

[0037] Preferably, the step of performing sliding window segmentation on the original response signal to generate time-domain signal slices, and normalizing each time-domain signal slice to eliminate baseline drift, includes:

[0038] A series of time-domain signal slices of equal length are generated by non-overlapping sliding of the original response signal using a Hamming window of fixed length.

[0039] For each time-domain signal slice, calculate the moving average of its signal values ​​to obtain the local baseline estimate of the time-domain signal slice;

[0040] The local baseline estimate is subtracted from each signal value of the time-domain signal slice to achieve zero mean of the signal;

[0041] The time-domain signal slice after zero-mean normalization is normalized by dividing it by the standard deviation of its signal value, thus generating a time-domain signal slice with a stable baseline.

[0042] Preferably, the correlation analysis between channels of the convolutional feature map output by the multi-scale convolutional kernel group, classifying the feature channels highly correlated with the overall trend of the primary response trajectory as target feature channels, and classifying the remaining feature channels as background feature channels, includes:

[0043] Calculate the Pearson correlation coefficient between the signal sequence of each feature channel output by the multi-scale convolution kernel group and the original signal sequence of the primary response trajectory;

[0044] Set a correlation threshold and mark feature channels with Pearson correlation coefficients greater than the correlation threshold as high correlation channels;

[0045] Fourier transform was performed on the signal sequences of highly correlated channels, and their power spectra were analyzed. Channels whose main energy was concentrated in the frequency range of the characteristic response of active collagen were classified as target characteristic channels.

[0046] The remaining feature channels whose Pearson correlation coefficients do not exceed the correlation threshold, as well as the channels in the high correlation channels whose main energy is not in the feature response frequency range, are uniformly classified as background feature channels.

[0047] Preferably, the design employs a diffusion-aggregation iteration rule. In each iteration, each node diffuses a portion of its state value to its neighboring nodes according to the outgoing edge weights. Simultaneously, it aggregates the incoming state values ​​from its neighboring nodes and updates the state values ​​using its own update coefficient, including:

[0048] At the beginning of each iteration, each network node calculates the state value components that diffuse to each neighbor node based on the proportion of the weights of all its outgoing edges to the total outgoing edge weights.

[0049] Each network node receives state value components from all incoming edges that are diffused from its neighboring nodes, and sums the received components as the input value for this round of aggregation;

[0050] Each network node multiplies its previous state value by a retention coefficient, adds the current aggregated input value by an absorption coefficient, and obtains an intermediate state value.

[0051] Multiply the intermediate state value by the node's update coefficient to obtain the node's state value after this iteration.

[0052] The sum of the retention coefficient and the absorption coefficient is one, and it is preset by the network's global hyperparameters.

[0053] Preferably, defining the origin and scale of the virtual reaction space includes:

[0054] All historical steady-state feature maps are projected onto two principal coordinate axes. The average value of the projected values ​​is used as the origin of the virtual reaction space coordinate axis, and the standard deviation of the projected values ​​is used as the unit scale of the coordinate axis.

[0055] Preferably, the method further includes a verification step after determining the content level:

[0056] Record the distribution of all sample projection vectors in the virtual reaction space within the current detection period to form a real-time distribution cloud map;

[0057] The overlap between the real-time distribution cloud map and the historical distribution cloud map of the standard sample library under the same conditions is compared, and the area ratio of the overlapping region is calculated.

[0058] If the area ratio is lower than the preset consistency threshold, a system calibration command is triggered, and the entire process from acquiring the original response signal to generating the steady-state feature map is re-executed using the known concentration standard samples of the current batch, so as to update the coordinate axis definition of the virtual reaction space and the standard concentration anchor point position.

[0059] Compared with the prior art, the beneficial effects of the present invention are:

[0060] This technique involves multi-level slicing of the original response signal in the time and frequency domains, and then splicing and superimposing the slices according to a preset phase to construct a primary response trajectory. This overcomes the limitations of traditional methods that only use a single-domain static or simplified dynamic signal. The technique reconstructs the chaotic initial signal of the reaction into a continuous trajectory containing complete spatiotemporal evolution information. This trajectory retains weak response features that may be missed by traditional methods, existing in specific frequency components or specific time phases, thus upgrading the description of reaction dynamics from a simple amplitude-time curve to a high-dimensional feature surface. Richer and more specific kinetic fingerprint information can be captured from the initial stage of the sample reaction, improving the detection sensitivity and recognition ability for active collagen, especially samples with low concentrations or in a weak reaction state.

[0061] This method utilizes a graph structure to construct a correlation network between different active signal patterns and generates a steady-state feature map through iterative calculation until convergence. This departs from the traditional approach of relying on preset mathematical transformations or statistical features for signal denoising and feature extraction. The method treats the separated signal patterns as interconnected dynamic state nodes, with edge weights defining the probability and strength of state transitions. The iterative convergence process essentially simulates the dynamic equilibrium reached by various interactions within the reaction system, automatically strengthening stable pattern correlation paths strongly correlated with target concentrations while suppressing non-steady-state correlations caused by random noise and transient interference. The resulting steady-state feature map is a highly stable topological structure emerging from the complex signal dynamics, rather than a simple arithmetic calculation. Therefore, it can more robustly characterize the unique reaction dynamics of active collagen in the face of environmental fluctuations and matrix differences, improving the anti-interference and repeatability of quantitative analysis. Attached Figure Description

[0062] Figure 1 This is a schematic diagram illustrating the working principle of the rapid detection method for active collagen content described in this invention.

[0063] Figure 2 A flowchart for constructing the primary reaction trajectory;

[0064] Figure 3 A flowchart for generating steady-state feature maps;

[0065] Figure 4 A comparison chart of Euclidean distances for determining the content of active collagen;

[0066] Figure 5 For the detection of active collagen - correlation network and steady-state characteristic map. Detailed Implementation

[0067] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0068] Please see Figure 1This invention provides a method for rapidly detecting the content of active collagen. The method includes: acquiring the original response signal of a reagent that undergoes a specific optical reaction with the collagen solution under excitation light. This signal is a dynamically changing signal spectrum, and its dynamic characteristics are formed by the combined effects of the solution state and active ingredients. The original response signal is subjected to multi-level slicing, extracting signal segments from the time domain, frequency domain, and time-frequency domain respectively. These signal segments from different domains are spliced ​​and superimposed according to a preset phase relationship to construct a primary reaction trajectory with spatiotemporal continuity. The constructed primary reaction trajectory is imported into a pre-designed convolutional kernel group for depth scanning. Through the parallel computing power of the convolutional kernel group, signal patterns closely related to changes in active collagen concentration are separated from the primary reaction trajectory, and background patterns composed of environmental noise and solution impurities are also separated, ultimately obtaining a set of active signal patterns free from background interference.

[0069] Based on this set of active signal patterns, a graph structure is used to construct a network connecting different signal patterns. In this network, each node represents a specific signal pattern, and the weights of the edges connecting the nodes represent the transition probability and strength of the correlation between patterns. The state of each node in the network is continuously updated through iterative calculation until the entire network converges and stabilizes, thereby generating a steady-state feature map that characterizes the reaction dynamics within the sample. The generated steady-state feature map is mapped into a virtual reaction space. The coordinate axes of this virtual reaction space are defined mathematically based on historical reaction trajectory data from a standard sample library. The projection vector of the steady-state feature map in the virtual reaction space is calculated. By measuring the distance between this projection vector and preset standard concentration anchor points corresponding to different concentration levels in the space, the specific content level of active collagen in the sample is determined.

[0070] Example 1: See Figure 2The acquired raw response signal is first fed into the signal slicing module. In this module, a fixed-length Hamming window is used to perform non-overlapping sliding segmentation of the raw response signal, generating a series of equal-length time-domain signal slices. For each generated time-domain signal slice, the moving average of its internal signal values ​​is calculated; this moving average is used as the local baseline estimate for that signal slice. The calculated local baseline estimate is subtracted from each raw signal value of the time-domain signal slice to achieve zero-mean normalization, thus initially eliminating baseline drift. For the time-domain signal slices that have undergone zero-mean normalization, they are further scaled using the standard deviation of their internal signal values ​​to achieve normalization, generating time-domain signal slices with stable baselines. For each normalized time-domain signal slice, wavelet transform is applied to decompose it into multiple frequency domain coefficients of different frequency bands. By setting an energy threshold, frequency band coefficients with energy higher than the threshold are selected, and these coefficients are combined to form frequency domain signal segments. Simultaneously, a Hilbert-Huang transform is performed on each time-domain signal slice to obtain the instantaneous frequency and amplitude information of the signal within the slice's time range. This information collectively forms a time-frequency domain joint distribution segment. A unified fusion time scale is established to precisely align the time-domain signal slices, frequency-domain signal segments, and corresponding time-frequency domain joint distribution segments on the timeline. Based on the instantaneous phase information of the signal revealed by the time-frequency domain joint distribution segment, cross-indexing and data concatenation are performed on the three signal data segments from the time domain, frequency domain, and time-frequency domain to generate a spatiotemporally continuous primary response trajectory.

[0071] In practice, the acquired raw response signal is presented as a voltage sequence that varies over time. This sequence is sent to the signal slicing module for processing. The module uses a Hamming window with 256 sampling points to perform non-overlapping sliding segmentation of the raw response signal. Each sliding step is equal to the window length, thus truncating the continuous raw response signal into a series of equal-length time-domain signal slices. For each segmented time-domain signal slice, the arithmetic mean of all voltage values ​​within it is calculated. This arithmetic mean is defined as the local baseline estimate of the time-domain signal slice. The calculated local baseline estimate is subtracted from each raw voltage value in the time-domain signal slice to achieve zero mean normalization of the signal. For the time-domain signal slice that has undergone zero mean normalization, the standard deviation of its voltage values ​​is calculated. Each voltage value in the slice is then divided by this standard deviation to complete the normalization process, generating a baseline-stable time-domain signal slice. The normalization process can be expressed by the following formula:

[0072] ;

[0073] in: This represents the normalized sequence of signal values. This represents a slice of the time-domain signal after zero-mean normalization. The local baseline estimate representing a slice of the time-domain signal is the arithmetic mean of the voltage values. The standard deviation of the voltage values ​​of a time-domain signal slice.

[0074] It is understandable that a discrete wavelet transform is applied to each baseline-stable time-domain signal slice after normalization. In some embodiments, the DB4 wavelet is selected as the basis function, and each baseline-stable time-domain signal slice is decomposed into 5 levels to obtain a series of wavelet coefficients corresponding to different frequency bands. The sum of squares of the wavelet coefficients of each frequency band is calculated as the energy of that frequency band. An energy threshold is set, and wavelet coefficients corresponding to frequency bands with energy higher than the energy threshold are selected. These selected wavelet coefficients are combined to form a frequency domain signal segment. Simultaneously, empirical mode decomposition is performed on the same baseline-stable time-domain signal slice to obtain a set of eigenmode functions. A Hilbert transform is performed on these eigenmode functions to solve for the analytic signal of each eigenmode function, thereby obtaining the instantaneous frequency and instantaneous amplitude of the signal. The joint distribution of instantaneous frequency and instantaneous amplitude in the time dimension constitutes a time-frequency joint distribution segment.

[0075] Example 2: The constructed primary response trajectory is fed into the feature separation module. In this module, a multi-scale convolutional kernel group is constructed, containing different kernels specifically designed to capture high-frequency details, mid-frequency contours, and low-frequency trends. This multi-scale convolutional kernel group is used to perform parallel convolution operations on the primary response trajectory, generating multi-channel convolutional feature maps. For each feature channel output by the multi-scale convolutional kernel group, the Pearson correlation coefficient between its signal sequence and the original signal sequence of the primary response trajectory is calculated. A correlation threshold is set, and feature channels with Pearson correlation coefficients greater than this threshold are marked as high-correlation channels. Fourier transforms are performed on the signal sequences of all high-correlation channels to analyze their power spectrum energy distribution. Channels whose main energy in the power spectrum is concentrated in the frequency range overlapping with the known active collagen characteristic response frequency range are classified as target feature channels. Feature channels whose Pearson correlation coefficients do not exceed the preset correlation threshold, and channels that are high-correlation channels but whose main energy in their power spectrum is not in the characteristic response frequency range, are uniformly classified as background feature channels. Spatial pooling is performed on data classified as target feature channels to compress their dimensionality while preserving their statistical significance, forming a set of target feature vectors. Inverse reconstruction is performed on data classified as background feature channels to generate a simulated background signal trajectory. This simulated background signal trajectory is subtracted from the original primary response trajectory to obtain a preliminary denoised signal trajectory. Feature clustering analysis is then performed on this preliminary denoised signal trajectory. The signal waveforms represented by each cluster center are defined as different active signal patterns. All cluster centers are then aggregated to obtain the final set of background-removed active signal patterns.

[0076] In practice, the constructed primary response trajectory is input into the feature separation module for processing. The primary response trajectory is a continuous data sequence containing multi-domain features. The feature separation module incorporates a multi-scale convolutional kernel group, which contains three different sizes of one-dimensional convolutional kernels. For example, a kernel of size 3 is used to capture high-frequency details, a kernel of size 7 is used to capture mid-frequency contours, and a kernel of size 15 is used to capture low-frequency trends. The multi-scale convolutional kernel group performs parallel convolution operations on the primary response trajectory. Each convolutional kernel independently slides and calculates on the primary response trajectory, generating a corresponding convolutional feature map. The multi-scale convolutional kernel group ultimately outputs a multi-channel set of convolutional feature maps.

[0077] In practice, for the signal sequence of each feature channel output by the multi-scale convolutional kernel group, it is necessary to calculate its Pearson correlation coefficient with the original signal sequence of the primary response trajectory to assess the correlation. The formula for calculating the Pearson correlation coefficient is as follows:

[0078] ;

[0079] in: Represents the Pearson correlation coefficient. The first characteristic channel signal sequence The value of each point, This represents the average value of all values ​​in the characteristic channel signal sequence. The first signal in the original signal sequence representing the primary response trajectory The value of each point, This represents the average value of all values ​​in the original signal sequence of the primary response trajectory. This represents the length of the signal sequence. A correlation threshold is set, which can be 0.6. Feature channels with calculated Pearson correlation coefficients greater than this threshold are marked as highly correlated channels.

[0080] In some embodiments, a Fast Fourier Transform (FFT) is performed on each signal sequence labeled as a highly correlated channel to convert the time-domain signal into a frequency-domain power spectrum. The energy distribution of the power spectrum is analyzed to identify the main frequency band with the highest energy. When active collagen undergoes a specific optical reaction with the reagent, its characteristic reaction frequency band is pre-determined through standard experiments, for example, the 10Hz to 50Hz band. Highly correlated channels whose main energy is concentrated in the frequency band that overlaps with the characteristic reaction frequency band are classified as target characteristic channels. It can be understood that characteristic channels whose Pearson correlation coefficient does not exceed the correlation threshold are directly classified as background characteristic channels. At the same time, channels that are highly correlated but whose main energy in their power spectrum is not in the characteristic reaction frequency band are also classified as background characteristic channels.

[0081] In some embodiments, spatial pooling is performed on the data categorized as target feature channels. The spatial pooling operation employs max pooling, with a pooling window length of 5 and a stride of 5. By sliding the pooling window across the data sequence of the target feature channels and selecting the maximum value within the window, the data dimensionality is compressed while preserving salient features, thereby forming a set of target feature vectors. Optionally, a reverse reconstruction operation is performed on the data categorized as background feature channels. The reverse reconstruction operation utilizes transposed convolution to upsample and combine the background feature channel data, generating a simulated background signal trajectory with the same length as the original signal sequence of the primary response trajectory. The simulated background signal trajectory is subtracted point-by-point from the original signal sequence of the primary response trajectory to obtain a preliminary denoised signal trajectory.

[0082] Example 3: See Figure 3 In the network construction module, a graph structure network is built based on the obtained set of active signal patterns. Each independent signal pattern in the set is defined as a network node, and the initial state value of all nodes is set to the energy amplitude of its corresponding signal pattern at the initial time. The conditional probability of any two network nodes representing signal patterns appearing successively in the original primary response trajectory is calculated. The product of this conditional probability and the average energy amplitude of the two signal patterns is used as the initial weight of the edge connecting the two nodes. Based on the edge weights between all nodes, the adjacency matrix of the entire network is constructed. Based on this adjacency matrix, the degree centrality index of each node in the network is calculated. The calculated degree centrality is normalized, and the normalized result is used as the update coefficient of the node's state value.

[0083] A diffusion-aggregation iterative rule is designed to update node states. At the beginning of each iteration, each network node calculates the state value component to be diffused to each neighbor node based on the proportion of the weights of all its outgoing edges to the total outgoing edge weights. Simultaneously, each network node receives the state value components diffused from each of its neighbor nodes from all its incoming edges and sums all received components as the input value for aggregation from the outside in this iteration. Each network node multiplies its state value from the previous iteration by a preset retention coefficient, adds it to the input value obtained from aggregation in this round multiplied by a preset absorption coefficient, and obtains an intermediate state value. This intermediate state value is multiplied by the node's own update coefficient to obtain the final state value of the node after the update in this iteration. The retention coefficient and absorption coefficient are preset global hyperparameters of the network, and their sum is one. The iteration stops when the change in the state values ​​of all nodes in two consecutive iterations is less than a preset convergence threshold. The final state values ​​of all nodes are arranged according to the temporal order of their corresponding signal patterns in the original reaction trajectory, generating a steady-state feature map characterizing the internal reaction dynamics of the sample.

[0084] In implementation, a set of active signal patterns de-contaminated is input into the association network construction module. This module treats each independent signal pattern in the active signal pattern set as a network node, with the total number of network nodes equal to the number of signal patterns in the active signal pattern set. The initial state value of each network node is set to the absolute value of the energy amplitude of its corresponding signal pattern at the first occurrence point in the original primary response trajectory. All initial state values ​​of the network nodes constitute the initial state vector. The conditional probability of any two network nodes representing signal patterns appearing consecutively in the original primary response trajectory is calculated. This probability is calculated by counting the number of consecutive occurrences of the two signal patterns and dividing by the total number of occurrences of the first signal pattern. The average energy amplitude of the two signal patterns at all occurrence positions in the original primary response trajectory is obtained, and the product of these two averages is calculated. The product of the calculated conditional probability and the product of the two averages is used as the initial weight of the edge connecting these two network nodes. If two signal patterns have never appeared consecutively in the original primary response trajectory, the initial weight of the edge connecting these two network nodes is zero.

[0085] In practical implementation, a square adjacency matrix is ​​constructed based on the initial weights of the edges connecting all network nodes. The rows and columns of the adjacency matrix correspond to network nodes, and the element in the i-th row and j-th column represents the initial weight of the edge pointing from network node i to network node j. The degree centrality of each network node is calculated based on the adjacency matrix. Degree centrality includes out-degree centrality and in-degree centrality. Out-degree centrality is the sum of the weights of all outgoing edges of the network node, and in-degree centrality is the sum of the weights of all incoming edges of the network node. In some embodiments, the out-degree centrality and in-degree centrality of a network node are added together to obtain a degree centrality index. The degree centrality indices of all network nodes are then normalized by dividing the degree centrality index of each network node by the sum of the degree centrality indices of all network nodes. The result of the normalization process is used as the update coefficient for the state value of that network node.

[0086] It is understandable that a diffusion-aggregation iterative rule is designed to update the state values ​​of network nodes, and this rule is executed sequentially in each iteration. At the beginning of each iteration, each network node calculates the state value component to be diffused to each neighbor node based on the ratio of the weights of all its outgoing edges to the total outgoing edge weights. The state value component diffused to each neighbor node is equal to the network node's current state value multiplied by the ratio of the weight of the outgoing edge pointing to that neighbor node to the total outgoing edge weights. Each network node receives the state value components diffused from each of its neighbors from all its incoming edges, and sums all the received components. The sum is used as the input value for aggregation from the outside in this iteration. The update of the network node state value follows the following formula:

[0087] ;

[0088] in: Represents network nodes In the The state value after round of iterations, Represents network nodes The update coefficient, Represents the preset global retention coefficient of the network. Represents network nodes In the The state values ​​of the round of iterations, This represents the preset global absorption coefficient of the network. Represents network nodes In the The input values ​​obtained through aggregation in each iteration. Retained coefficients. With absorption coefficient The sum is 1, and it is preset by the network's global hyperparameters, such as the retention coefficient. The absorption coefficient is 0.3. It is 0.7.

[0089] In some embodiments, the iterative calculation stops when the absolute value of the change in the state values ​​of all network nodes in two consecutive iterations is less than a preset convergence threshold, which can be set to 1e-6. The final state values ​​of all network nodes are arranged according to the temporal order of the first appearance of the signal pattern corresponding to the network node in the original primary reaction trajectory, generating a steady-state feature map characterizing the internal reaction dynamics of the sample. Optionally, the diffusion calculation can also be performed based on the proportion of the incoming edge weights, and the aggregation process is adjusted accordingly to diffusion based on the outgoing edge weights. Optionally, the retention coefficient and absorption coefficient in the node state update formula can be set as learnable parameters. It can be understood that through iterative updates, the state values ​​of the network nodes eventually converge to a stable distribution, which reflects the steady-state weights of different signal patterns in the reaction dynamics, thus forming a steady-state feature map.

[0090] Example 4: In the content determination module, a virtual reaction space for mapping analysis is constructed. Historical steady-state feature maps generated and processed under the same detection conditions from all standard samples of known concentrations in the stored standard sample library are read, forming a set of historical steady-state feature maps. The standardized Euclidean distance between each pair of feature maps in this set is calculated to form a distance matrix. This distance matrix is ​​centered, and the covariance matrix of the centered matrix is ​​calculated. The eigenvalues ​​and corresponding eigenvectors of this covariance matrix are calculated. The eigenvalues ​​are sorted from largest to smallest, and the eigenvectors corresponding to the largest and second-largest eigenvalues ​​are selected, with their directions used as the directions of the two principal coordinate axes of the virtual reaction space. All feature maps in the historical steady-state feature map set are projected onto these two principal coordinate axes, and the average value of all projected point coordinates is calculated. This average value is set as the origin of the coordinate axes of the virtual reaction space, and the standard deviation of all projected point coordinates is calculated. This standard deviation is used as the unit scale of the coordinate axes, thus completing the definition of the virtual reaction space coordinate system. In the virtual reaction space, points corresponding to the steady-state characteristic maps of standard samples at different concentration levels are pre-marked and defined as standard concentration anchor points for that concentration level. The steady-state characteristic map of the sample to be tested is projected onto the two principal coordinate axes of the predefined virtual reaction space, and its projected coordinates are calculated to form a two-dimensional projection vector. The Euclidean distance from this two-dimensional projection vector to each pre-defined standard concentration anchor point in the virtual reaction space is calculated, and the minimum value is found among all distances. The active collagen concentration level corresponding to the standard concentration anchor point that produces the minimum distance is determined as the content level of the sample to be tested.

[0091] In practice, the steady-state feature maps generated by the correlation network construction module are sent to the content determination module for analysis. The content determination module has a built-in virtual reaction space for mapping and determination. A series of historical steady-state feature maps generated and processed under the same detection conditions from a stored standard sample library of known concentrations are read. These historical steady-state feature maps constitute a set, and each historical steady-state feature map in the set is a numerical vector. The standardized Euclidean distance between all pairs of historical steady-state feature maps in the set is calculated. The calculation of the standardized Euclidean distance requires first calculating the standard deviation of all historical steady-state feature map data in each feature dimension. The specific calculation formula is as follows:

[0092] ;

[0093] in: Represents historical steady-state characteristics Historical steady-state characteristic map The standardized Euclidean distance between them The total number of dimensions representing the steady-state feature map. Represents historical steady-state characteristics In the eigenvalues ​​of dimension Represents historical steady-state characteristics In the eigenvalues ​​of dimension Representing all samples in the historical steady-state feature map set at the th The standard deviation over the eigenvalues. A distance matrix is ​​formed by calculating the distances between all historical steady-state eigenmap pairs.

[0094] The obtained distance matrix is ​​centered using a double-centering method, which involves subtracting both the row and column means and then adding the overall mean. The covariance matrix of the centered matrix is ​​then calculated, along with all its eigenvalues ​​and corresponding eigenvectors. These eigenvalues ​​are then sorted from largest to smallest. The eigenvector corresponding to the largest eigenvalue is selected as the direction vector for the first principal coordinate axis, and the eigenvector corresponding to the second largest eigenvalue is selected as the direction vector for the second principal coordinate axis. All historical steady-state feature maps are then projected onto these two principal coordinate axes. The projection operation involves taking the dot product of the historical steady-state feature map vectors and the principal coordinate axis direction vectors to obtain the projected coordinates of each historical steady-state feature map on the two principal axes.

[0095] In some embodiments, the arithmetic mean of the projected coordinate values ​​of all historical steady-state feature maps on the first principal coordinate axis is calculated, and the arithmetic mean of the projected coordinate values ​​of all historical steady-state feature maps on the second principal coordinate axis is calculated. The point determined by these two averages is set as the origin of the virtual reaction space coordinate system. The standard deviation of the projected coordinate values ​​of all historical steady-state feature maps on the first principal coordinate axis is calculated and used as the unit scale of the first principal coordinate axis. The standard deviation of the projected coordinate values ​​of all historical steady-state feature maps on the second principal axis is calculated and used as the unit scale of the second principal axis. In the virtual reaction space, position points corresponding to the steady-state feature maps of standard samples at different concentration levels are pre-marked. The correspondence between the concentration levels of the standard samples and the position points is obtained through pre-calibration, and these position points are defined as standard concentration anchor points. Refer to Table 1 for examples of standard concentration anchor points and their coordinates.

[0096] Table 1: Standard Concentration Anchor Point Table

[0097] Concentration level Projected coordinates on the first principal coordinate axis Projected coordinates on the second principal coordinate axis low concentration -1.85 0.32 medium and low concentrations -0.92 -0.15 medium concentration 0.05 -0.41 medium to high concentration 1.12 -0.08 high concentration 2.03 0.28

[0098] In practice, the steady-state feature vector of the sample to be tested is projected onto the two principal coordinate axes of a predefined virtual reaction space. The projection is calculated using a dot product operation. The steady-state feature vector of the sample to be tested is dot-producted with the direction vector of the first principal coordinate axis to obtain the projected coordinate values ​​on the first principal coordinate axis. The dot product is then performed with the direction vector of the second principal coordinate axis to obtain the projected coordinate values ​​on the second principal axis. These two projected coordinate values ​​constitute a two-dimensional projection vector. The Euclidean distance from the two-dimensional projection vector to each preset standard concentration anchor point in the virtual reaction space is calculated. The Euclidean distance is calculated by taking the square root of the sum of the squares of the differences between the coordinates of the two-dimensional projection vector and the coordinates of the standard concentration anchor points. The distance with the smallest value is selected from all calculated Euclidean distances. The active collagen concentration level corresponding to the standard concentration anchor point that produces this smallest distance is determined as the active collagen content level in the sample to be tested.

[0099] Optionally, the principal coordinate analysis can also select the first three principal coordinate axes to construct a three-dimensional virtual reaction space. Optionally, the distance calculation can also use Mahalanobis distance instead of standardized Euclidean distance. In some embodiments, for cases where it is difficult to determine the concentration level near the midpoint of the line connecting two standard concentration anchor points, the K-nearest neighbor algorithm can be used to vote on the concentration levels of the K nearest standard sample anchor points to the projected point of the sample. It can be understood that by projecting a high-dimensional steady-state feature map onto a low-dimensional virtual reaction space, a mapping from complex features to intuitive judgment is achieved, and the method of determining the concentration level based on distance has clear geometric meaning.

[0100] See Figure 4 This is a comparison chart of Euclidean distances for determining the content of active collagen. This chart corresponds to the "distance calculation" stage of active collagen content detection, used to quantify the degree of matching between the sample and each standard concentration anchor point: the smaller the distance, the closer the characteristics of the sample are to the corresponding concentration level of the standard sample; it clarifies the content determination result: the "minimum distance" directly determines the concentration level of the sample as "medium concentration"; it verifies the reliability of the determination: the distance between the sample and the "medium concentration" anchor point is much smaller than other levels, indicating a high degree of discrimination in the determination result. This chart visually shows that the active collagen content level of the sample is "medium concentration"; the confidence level of the determination result is high.

[0101] Example 5: After determining the content level, the system executes a verification step. This step records the distribution of the projection vectors of all tested samples (including standard samples and test samples) in the virtual reaction space within the current testing cycle or batch, forming a real-time distribution cloud map. This real-time distribution cloud map is compared with historical distribution cloud maps generated by the standard sample library under the same experimental conditions, and the area ratio of the overlapping region in the virtual reaction space is calculated. If the calculated area ratio is lower than a preset consistency threshold, the system automatically triggers a calibration command. After triggering the calibration command, the system uses standard samples of known concentrations provided in the current testing batch to re-execute the entire process from acquiring the original response signal, through constructing the primary reaction trajectory, convolutional scanning separation mode, and building an association network to generate a steady-state feature map. Based on the newly generated steady-state feature map data of the standard samples, the definition of the virtual reaction space coordinate axes and the determination of the standard concentration anchor point positions are re-executed to update the system's judgment criteria.

[0102] In practice, after determining the level of active collagen content, the system executes an independent verification step to assess the overall reliability of the current batch's test results. This verification step first records the distribution of all sample projection vectors in the virtual reaction space within the current testing period. All samples in the current testing period include standard samples of known concentrations used for quality control and samples of unknown concentrations to be tested. For example, a testing batch may contain 3 standard samples and 7 samples to be tested. The system records the two-dimensional projection coordinates of these 10 samples in the virtual reaction space, and the set of all projection coordinate points forms a real-time distribution cloud map.

[0103] In some embodiments, a real-time distribution cloud map is compared with a historical distribution cloud map generated under the same experimental conditions from a standard sample library. The historical distribution cloud map is a statistical set of the projected coordinates of standard samples accumulated from multiple past experiments. The comparison requires calculating the area ratio of the overlapping region between the two cloud maps in the virtual reaction space. The formula for calculating the area ratio is:

[0104] ;

[0105] in: Represents the percentage of area. This represents the area of ​​the overlapping region between the real-time distribution cloud map and the historical distribution cloud map in the virtual reaction space. This represents the total area covered by the historical distribution cloud map. Calculate the area of ​​the overlapping region. One approach is to discretize the virtual reaction space into a fine grid, statistically analyze the set of grids covered by historical distribution cloud maps, and then statistically analyze the set of grids covered by real-time distribution cloud maps. The number of grids in the intersection of the two sets multiplied by the area of ​​a single grid is the result. The historical distribution cloud map is calculated by multiplying the number of grids covered by the cloud by the area of ​​a single grid. .

[0106] It's understandable that a preset consistency threshold is set, which could be 0.75. The calculated area percentage is then used. Compared with the consistency threshold, if the area ratio If the result falls below a preset consistency threshold, a system calibration command is triggered. Once triggered, the control process automatically guides the re-analysis using standard samples of known concentrations provided in the current testing batch. The re-execution process is fully automated, starting from the acquisition of the raw response signal, including excitation light irradiation and optical signal acquisition, followed by the entire process of constructing the primary response trajectory, performing deep scanning to separate signal patterns using convolutional kernel groups, and building a correlation network to generate a steady-state feature map.

[0107] In some embodiments, a new set of steady-state feature maps regenerated using the current batch of standard samples will be used to update the system's core decision criteria. Based on the new set of steady-state feature maps, the principal coordinate axis orientation of the virtual reaction space is recalculated, i.e., all steps of principal coordinate analysis on the historical set of steady-state feature maps are re-executed, including calculating the distance matrix, centering, and solving for eigenvectors and eigenvalues. It can be understood that the origin and unit scale of the virtual reaction space are recalculated and determined based on the projected coordinates of the new set of steady-state feature maps. Simultaneously, based on the projected position of the new set of steady-state feature maps in the redefined virtual reaction space, the coordinates of the standard concentration anchor points are remarked and updated.

[0108] Optionally, the area percentage can be calculated by generating a continuous probability density distribution using kernel density estimation, and then calculating the intersection area between the two distributions. Optionally, the consistency threshold can be dynamically adjusted according to the required detection accuracy; for example, a higher consistency threshold can be set for high-precision detection scenarios. In specific implementations, if the area percentage... If the result is higher than or equal to the consistency threshold, the verification passes, and the test results and judgments for the current batch are valid, without the need to trigger calibration. The verification step provides a periodic verification mechanism for the long-term stability and reliability of the detection system by monitoring the consistency of the overall distribution. The triggering of system calibration commands ensures that the judgment benchmark can adapt to possible minor conditional drifts.

[0109] See Figure 5This is a graph illustrating the detection of active collagen, its associated network, and its steady-state characteristics. Node energy intensity exhibits a stratified distribution over time, with high energy (yellow / green) concentrated in the time dimension (0-0.2) and energy intensity region (15000+), while low energy (purple) is concentrated in the time dimension (1.2-1.4) and energy intensity region (5000+). The gray trajectory shows a decreasing trend in energy intensity over time, reflecting the dynamic decay of node energy. This graph is commonly used for node state analysis in network systems, visually displaying the distribution characteristics of node energy in the time and intensity dimensions, aiding in the identification of spatiotemporal clusters of high-energy (high-load) nodes; the dynamic trend of energy changes provides data support for network resource scheduling and fault early warning.

[0110] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0111] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for rapidly detecting the active collagen content, characterized by, The method comprises: Collecting the original response signal of the reagent under excitation light, which specifically reacts with the collagen solution to be tested, and capturing the dynamic signal spectrum formed by the joint action of the solution state and the active ingredient; Multi-level slicing of the original response signal is performed to extract signal segments from time domain, frequency domain and time-frequency domain respectively, and the signal segments in different domains are spliced and superimposed according to a preset phase to construct a primary reaction trajectory with time and space continuity; The primary reaction trajectory is introduced into a convolution kernel group for deep scanning, and the signal mode related to the concentration change of active collagen and the background mode composed of environmental noise and impurity interference are separated through parallel operation of the convolution kernel group to obtain an active signal mode set without background; Based on the active signal mode set, a correlation network between different signal modes is constructed using a graph structure, a node represents a specific signal mode, and an edge weight represents the transition probability and intensity correlation between modes. The network node state is updated through iterative calculation until convergence to generate a steady-state feature map representing the internal reaction kinetics of the sample; Mapping the steady-state feature map to a virtual reaction space, the coordinate axes of the virtual reaction space are defined according to the historical reaction trajectories of the standard sample library, including: reading a set of historical steady-state feature maps generated by standard samples with known concentrations under the same detection conditions from the standard sample library; Calculate the distance matrix between all feature maps in the historical steady-state feature map set, the distance uses the standardized Euclidean distance; Centralize the distance matrix to calculate the covariance matrix; Calculate the eigenvalues and eigenvectors of the covariance matrix, sort the eigenvalues from large to small, and select the eigenvectors corresponding to the largest and second largest eigenvalues as the directions of the two principal coordinate axes of the virtual reaction space; Define the coordinate origin and scale of the virtual reaction space; Calculate the projection vector of the steady-state feature map in the virtual reaction space, and determine the content level of the active collagen according to the distance between the projection vector and the preset standard concentration anchor point.

2. The method for rapidly detecting the active collagen content according to claim 1, characterized in that, The multi-level slicing of the original response signal, the extraction of signal segments from time domain, frequency domain and time-frequency domain, and the splicing and superimposition of signal segments in different domains according to a preset phase to construct a primary reaction trajectory with time and space continuity, comprises: Sliding window segmentation of the original response signal to generate time domain signal slices, and normalization processing of each time domain signal slice to eliminate baseline drift; Apply wavelet transform to each normalized time domain signal slice to decompose it into frequency domain coefficients of multiple frequency bands, and select frequency band coefficients with energy higher than a threshold to form frequency domain signal segments; Synchronously perform Hilbert-Huang transform on each time domain signal slice to obtain the instantaneous frequency and instantaneous amplitude of the signal, forming a time-frequency domain joint distribution segment; Establish a time scale fusion to time-align the time domain signal slices, frequency domain signal segments and time-frequency domain joint distribution segments, cross-index and splice the signal data in the three domains according to the phase information revealed by the signal in the time-frequency domain joint distribution segment to generate a continuous primary reaction trajectory.

3. The method for rapidly detecting the active collagen content according to claim 1, characterized in that, The primary reaction track is introduced into the convolution kernel group for deep scanning, and the signal mode related to the concentration change of active collagen protein and the background mode composed of environmental noise and impurity interference are separated through parallel operation of the convolution kernel group, to obtain a set of active signal modes after background removal, including: A multi-scale convolution kernel group is constructed, which contains convolution kernels for high-frequency details, medium-frequency profiles, and low-frequency trends, respectively, and the multi-scale convolution kernel group is used for parallel convolution operation on the primary reaction track; The correlation analysis between channels is performed on the convolution feature map output by the multi-scale convolution kernel group, the feature channels highly related to the overall change trend of the primary reaction track are classified as target feature channels, and the remaining feature channels are classified as background feature channels; The data classified as target feature channels are spatially pooled to compress their dimensions and retain their statistical significance, forming a target feature vector; The data classified as background feature channels are reconstructed in reverse to generate a simulated background signal track, which is subtracted from the original primary reaction track to obtain a preliminary denoised signal track; The preliminary denoised signal track is again subjected to feature clustering, and the signal waveform represented by the cluster center is defined as an active signal mode. All cluster centers are collected to obtain a set of active signal modes after background removal.

4. The method for rapidly detecting the active collagen content according to claim 1, characterized in that, Based on the set of active signal modes, an association network between different signal modes is constructed using a graph structure, where nodes represent specific signal modes and edge weights represent the transition probability and intensity correlation between modes. The network node state is updated iteratively until convergence to generate a steady-state feature map representing the internal reaction dynamics of the sample, including: Each signal mode in the set of active signal modes is taken as a network node, and the state value of all nodes is initialized as the initial energy amplitude of the corresponding signal mode; The conditional probability of the signal modes represented by any two network nodes appearing successively in the original primary reaction track is calculated, and the product of the conditional probability and the average energy of the two modes is taken as the initial weight of the edge connecting the two nodes; The adjacency matrix of the network is constructed according to the edge weight, and the degree centrality of each node is calculated based on the adjacency matrix. The normalized degree centrality is used as the update coefficient of the node state; A diffusion-aggregation iteration rule is designed. In each iteration, each node diffuses its partial state value to neighbor nodes according to the outgoing edge weight, and aggregates the incoming state value from neighbor nodes. The state value is updated combined with the update coefficient of itself. When the change amount of all node state values in two consecutive iterations is less than the convergence threshold, the iteration stops. The final state values of all nodes are arranged in the time sequence order of the corresponding signal modes to generate a steady-state feature map.

5. The method for rapidly detecting the active collagen content according to claim 1, characterized in that, The projection vector of the steady-state feature map in the virtual reaction space is calculated, and the distance between the projection vector and the preset standard concentration anchor point is used to determine the content level of active collagen protein, including: The steady-state feature map of the sample to be tested is projected onto the two principal coordinate axes of the virtual reaction space, and its projection coordinates are calculated to form a two-dimensional projection vector. In the virtual reaction space, standard concentration anchors corresponding to the steady-state characteristic maps of standard samples of different concentration levels are marked in advance; The Euclidean distance of the two-dimensional projection vector of the sample to be tested to each standard concentration anchor is calculated, and the concentration level corresponding to the standard concentration anchor with the smallest distance is determined as the content level of the active collagen in the sample to be tested. 6.The method for rapidly detecting the content of active collagen according to claim 2, characterized in that, The original response signal is divided by a sliding window to generate a time-domain signal slice, and each time-domain signal slice is normalized to eliminate baseline drift, including: The original response signal is overlapped by a fixed length Hamming window to generate a series of equal length time-domain signal slices; For each time-domain signal slice, the sliding average of its signal value is calculated to obtain the local baseline estimate of the time-domain signal slice; The local baseline estimate is subtracted from each signal value of the time-domain signal slice to achieve zero mean of the signal; The time-domain signal slice after zero mean is divided by its signal value standard deviation to complete the normalization processing and generate a baseline stable time-domain signal slice.

7. The method for rapidly detecting the active collagen content according to claim 3, characterized in that, The correlation between the channels of the convolution feature map output by the multi-scale convolution kernel group is analyzed, the feature channels highly correlated with the overall change trend of the primary reaction trajectory are classified as target feature channels, and the remaining feature channels are classified as background feature channels, including: The Pearson correlation coefficient of the signal sequence of each feature channel output by the multi-scale convolution kernel group and the primary reaction trajectory original signal sequence is calculated; Set a correlation threshold, and mark the feature channels with a Pearson correlation coefficient greater than the correlation threshold as high correlation channels; Perform Fourier transform on the signal sequence of the high correlation channel to analyze its power spectrum, and classify the channels with main energy concentrated in the characteristic reaction frequency band coinciding with the active collagen as target feature channels; The remaining feature channels whose Pearson correlation coefficient does not exceed the correlation threshold, and the channels in the high correlation channel whose main energy is not in the characteristic reaction frequency band, are uniformly classified as background feature channels. 8.The method for rapidly detecting the content of active collagen according to claim 4, characterized in that, The diffusion-aggregation iteration rule is designed, in each iteration, each node diffuses its partial state value to neighbor nodes according to the edge weight, and aggregates the incoming state value from neighbor nodes, combines the update coefficient to update the state value, including: At the beginning of each iteration, each network node calculates the state value component diffused to each neighbor node according to the proportion of the weight of all outgoing edges in the total outgoing edge weight; Each network node receives the state value component diffused from its neighbor nodes from all incoming edges, and sums the received components as the input value of this round of aggregation; Each network node multiplies its state value of the last round by a retention coefficient, adds the input value of this round of aggregation multiplied by an absorption coefficient, to obtain an intermediate state value; Multiply the intermediate state value by the update coefficient of the node to obtain the updated state value of the node in this round of iteration; Wherein, the sum of the retention coefficient and the absorption coefficient is one, and is preset by the network global hyperparameter.

9. The method for rapidly detecting the active collagen content according to claim 1, characterized in that, The coordinate origin and scale of the virtual reaction space are defined, including: Projecting all historical steady-state feature maps onto two principal coordinate axes, taking the average of the projected values as the origin of the virtual response space coordinate axes, and taking the standard deviation of the projected values as the unit scale of the coordinate axes.

10. The method for rapidly detecting the active collagen content according to claim 1, characterized in that, The method further comprises a verification step after determining the content level: Record the distribution of all sample projection vectors in the virtual response space in the current detection period to form a real-time distribution cloud map; Compare the real-time distribution cloud map with the historical distribution cloud map of the standard sample library under the same conditions in terms of overlap degree, and calculate the area ratio of the overlapping area; If the area ratio is lower than a preset consistency threshold, a system calibration instruction is triggered, and the whole process from collecting the original response signal to generating the steady-state feature map is re-executed using the known concentration standard sample of the current batch to update the coordinate axis definition and standard concentration anchor position of the virtual response space.

Images