Sample concentration detection methods, systems, equipment and readable storage media

Abstract

This invention provides a sample concentration detection method, system, device, and readable storage medium. The method includes: irradiating a sample containing the analyte with a laser of switchable excitation wavelength, and determining the laser wavelength switching time as a clock reference point; configuring and locking each state component as a target value in a first time window before the clock reference point; acquiring spectral data based on the locked state components in a second time window after the clock reference point; constructing a multidimensional homologous spectral tensor based on the spectral data and the corresponding state components; and inputting the multidimensional homologous spectral tensor into an evaluation model to obtain the concentration of the analyte in the sample output by the evaluation model. Compared with existing technologies, this invention achieves multidimensional physical homologous co-sampling, significantly reducing multivariate coupling errors in the sampling scheme.

Description

Technical Field

[0001] This invention belongs to the field of photoelectric detection technology, specifically relating to a sample concentration detection method, system, device, and readable storage medium. Background Technology

[0002] Raman spectroscopy provides rich molecular vibrational fingerprint information and is an important tool for non-invasive biomedical detection. However, in in vivo detection of biological tissues, the effective Raman signal is extremely weak, and it faces multiple challenges such as strong fluorescence background interference, signal attenuation due to tissue scattering, and baseline drift caused by physiological motion.

[0003] To address these issues, the academic community has proposed three main technical approaches: frequency-shifted excited Raman differential spectroscopy (SERDS), spatially offset Raman spectroscopy (SORS), and time-gated Raman detection. These approaches utilize differences in wavelength response, spatial separation, and temporal response to suppress background interference, respectively. However, in existing systems, these three techniques are typically executed as independent modules. Even with the introduction of a global clock for timestamping, each module remains physically independent—wavelength switching, spatial scanning, and time gating are performed separately at different times, and the data is then aligned along the time axis during post-processing.

[0004] This "temporally adjacent" implementation method inherently cannot guarantee that the spatial bias state and time-gated state remain strictly unchanged when calculating the SERDS difference; nor can it guarantee that the wavelength state and time-gated state remain strictly unchanged when calculating the SORS difference. The noise introduced by independent sampling in each dimension will superimpose, severely limiting the detection accuracy of weak Raman signals.

[0005] Therefore, to address the aforementioned technical problems, it is necessary to provide a sample concentration detection method, system, device, and readable storage medium based on multidimensional co-sampling.

[0006] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0007] The purpose of this invention is to provide a sample concentration detection method, system, device, and readable storage medium, which solves the problem of large multivariate coupling error in sampling schemes.

[0008] To achieve the above objectives, the technical solution provided by the present invention is as follows:

[0009] In a first aspect, the present invention provides a sample concentration detection method, comprising:

[0010] The sample containing the analyte is irradiated with a laser with switchable excitation wavelength, and the laser wavelength switching time is determined as the clock reference point.

[0011] In the first time window before the clock reference point, each state component is configured and locked as a target value; in the second time window after the clock reference point, spectral data is acquired based on the locked state components; the state components include spatial offset channel state, time gating parameters, and modulation coding state.

[0012] Based on the spectral data and the corresponding state components, a multidimensional homologous spectral tensor is constructed; wherein the multidimensional homologous spectral tensor contains at least one pair of spectral data samples that vary along a single target dimension and are locked in the other dimensions, and the target dimension is the laser wavelength or any state component.

[0013] The multidimensional homology spectral tensor is input into the evaluation model to obtain the concentration of the analyte in the sample output by the evaluation model.

[0014] In one or more embodiments of the present invention, the first time window and the second time window between adjacent clock reference points are configured to not overlap on the time axis.

[0015] In one or more embodiments of the present invention, the end point of the first time window is configured as the clock reference point;

[0016] Configure the starting point of the second time window as the clock reference point; configure the ending point of the second time window as the starting point of the next first time window.

[0017] In one or more embodiments of the present invention, determining a second time window located after the clock reference point includes:

[0018] A third time window is determined starting from the clock reference point. Within the third time window, each state component is locked. The duration of the third time window is greater than or equal to the duration required for the parasitic oscillation to decay to the first threshold.

[0019] The second time window is determined after the end point of the third time window and before the start point of the next first time window.

[0020] In one or more embodiments of the present invention, determining the second time window located after the clock reference point includes:

[0021] The starting point of the second time window is determined as the ending point of the third time window; the ending point of the second time window is the starting point of the next first time window; wherein the first time window, the second time window, and the third time window do not overlap with each other on the time axis.

[0022] In one or more embodiments of the present invention, a multidimensional homologous spectral tensor is constructed based on the spectral data and the state components corresponding to the spectral data, including:

[0023] The spectral data is preprocessed, including determining whether the state components corresponding to the spectral data are homologous samples, and / or normalizing the spectral data based on the laser power;

[0024] The preprocessed spectral data and the corresponding state components of the spectral data are written into the corresponding positions of the multidimensional homogeneous spectral tensor.

[0025] In one or more embodiments of the present invention, determining whether the state component corresponding to the spectral data is a homologous sample includes:

[0026] Obtain the lock start time for each state component, as well as the calibration value of the lock time required;

[0027] Calculate the time difference between the start of locking for each state component and the next clock reference point;

[0028] If the time difference is greater than or equal to the calibration value, then the state component corresponding to the spectral data is a homologous sample;

[0029] If the time difference is less than the calibration value, then the state component corresponding to the spectral data is a non-homologous sample.

[0030] In one or more embodiments of the present invention, determining whether the state component corresponding to the spectral data is a homologous sample includes:

[0031] Calculate the deviation of each state component when it is sampled in the second time window relative to when it is locked in the first time window;

[0032] If the deviation is less than the preset upper limit of error, then the state component corresponding to the spectral data is a homologous sample.

[0033] If the deviation is greater than or equal to the preset upper limit of error, then the state component corresponding to the spectral data is a non-homogeneous sample.

[0034] In one or more embodiments of the present invention, determining whether the state component corresponding to the spectral data is a homologous sample includes:

[0035] Acquire spectral data sample pairs collected under lasers of different wavelengths with the same state components;

[0036] Separate the background region of the spectral data and count the photon count sequence in the background region of the two samples;

[0037] Calculate the Pearson correlation coefficient value of the photon counting sequence;

[0038] If the Pearson correlation coefficient value is greater than or equal to the coefficient threshold, then the state component corresponding to the spectral data is a homologous sample;

[0039] If the Pearson correlation coefficient is less than the coefficient threshold, then the state component corresponding to the spectral data is a non-homologous sample.

[0040] In one or more embodiments of the present invention, determining whether the state component corresponding to the spectral data is a homologous sample includes:

[0041] The laser beam is split into a first beam that illuminates the sample and a second beam that is guided into a reference optical path;

[0042] Multiple sets of power values ​​of the second beam are acquired within the second time window for acquiring the current sample spectral data;

[0043] Calculate the relative fluctuation value of the power value, where the relative fluctuation value is the quotient of the standard deviation and the average value of the power sampled values ​​within the second time window;

[0044] If the relative fluctuation value of the power value is less than the power fluctuation threshold, then the state component corresponding to the spectral data is a homologous sample;

[0045] If the relative fluctuation value of the power value is greater than or equal to the power fluctuation threshold, then the state component corresponding to the spectral data is a non-homogeneous sample.

[0046] In one or more embodiments of the present invention, if the sample is a non-homologous sample, the non-homologous sample is discarded, and / or the spectral data acquisition under the corresponding state component is re-executed, and / or the start time of the second time window is adjusted.

[0047] In one or more embodiments of the present invention, the normalization of the spectral data based on laser power includes:

[0048] The laser beam is split into a third beam to illuminate the sample and a fourth beam to guide the reference optical path;

[0049] Determine the average power of the fourth beam within the second time window for collecting the current sample spectral data;

[0050] The spectral data of the current sample is normalized based on the power mean to obtain the spectral data under unit excitation power.

[0051] In one or more embodiments of the present invention, the multidimensional homology spectral tensor is input into the evaluation model to obtain the concentration of the analyte in the sample output by the evaluation model, including:

[0052] For the spectral data in the multidimensional homogeneous spectral tensor, calculate the gradient features along at least one dimension;

[0053] Based on the spectral data and the gradient features, a joint input vector is generated;

[0054] Based on the joint input vector, the concentration of the analyte in the sample is determined and output.

[0055] In one or more embodiments of the present invention, generating a joint input vector based on the spectral data and the gradient features includes:

[0056] Obtain the preset fluctuation calibration values ​​of the spectral data and the gradient features, and perform standardization processing on the spectral data and the gradient features based on the fluctuation calibration values;

[0057] The normalized spectral data and the gradient features are spliced ​​together to generate the joint input vector.

[0058] In one or more embodiments of the present invention, the gradient feature along at least one dimension includes: a first gradient feature and a second gradient feature.

[0059] The first gradient feature is determined by the ratio of the difference between two frames of spectral data that differ in the target dimension but are the same in the other dimensions in the multidimensional homogeneous spectral tensor, to the dimensional interval of the target dimension; the dimensional interval refers to the difference between two different values ​​of the same dimension; and / or

[0060] The second gradient feature is determined by the ratio of the difference between the two first gradient features to the dimensional interval of the other dimension, wherein the two first gradient features are different in the other dimension but are the same in the other dimensions.

[0061] In one or more embodiments of the present invention, the method further includes:

[0062] Construct an evaluation model and set the relevant parameters of the evaluation model;

[0063] Obtain a training sample set, which includes the multidimensional homology spectral tensor and the concentration of the component to be measured in the corresponding sample;

[0064] The evaluation model is trained based on the training sample set, and the parameters of the evaluation model are corrected until the deviation value of the concentration of the component to be tested in the sample obtained by the evaluation model output is less than a preset threshold.

[0065] In one or more embodiments of the present invention, the evaluation model includes one or more machine learning models selected from deep neural networks, Gaussian process regression, or physical information neural networks.

[0066] In one or more embodiments of the present invention, the method further includes:

[0067] Get the current cumulative number of samples participating in model training;

[0068] Based on the number of samples, select one from a set of preset machine learning models as the current model to be trained and evaluated, or

[0069] Based on the number of samples, the contribution ratio of each of the multiple machine learning models in the final concentration estimate is determined.

[0070] In one or more embodiments of the present invention, the method further includes:

[0071] Obtain the physiological background parameters of the sample, which include at least one of water peak intensity, temperature-sensitive Raman peak shift, collagen characteristic peak intensity, and fat characteristic peak intensity;

[0072] An initial concentration estimate is obtained based on the multidimensional homology spectral tensor, and the initial concentration estimate is calibrated based on the physiological background parameter to obtain the concentration of the analyte.

[0073] In one or more embodiments of the present invention, the laser wavelength serves a dual function of time-domain reference and frequency-domain difference: the switching period of the laser wavelength serves as a reference dimension for the global clock, driving each state component to enter the locking sequence within the first time window; the value of the laser wavelength serves as one of the target dimensions, and by differentiating the spectral data collected at adjacent wavelengths, fluorescence background is eliminated, thereby realizing frequency-shifted excitation Raman differential.

[0074] Secondly, the present invention provides a sample concentration detection system, comprising:

[0075] The irradiation module is used to irradiate a sample containing the analyte with a laser based on a switchable excitation wavelength, and to determine the laser wavelength switching time as a clock reference point.

[0076] The configuration module is used to configure and lock each state component as a target value in a first time window before the clock reference point; and to perform spectral data acquisition based on the locked state components in a second time window after the clock reference point; the state components include spatial offset channel state, time gating parameters, and modulation coding state.

[0077] The analysis module is used to construct a multidimensional homologous spectral tensor based on the spectral data and the state components corresponding to the spectral data; wherein the multidimensional homologous spectral tensor contains at least one pair of spectral data samples that vary along a single target dimension and are locked in the other dimensions, and the target dimension is the laser wavelength or any state component;

[0078] The output module is used to input the multidimensional homology spectral tensor into the evaluation model to obtain the concentration of the analyte in the sample output by the evaluation model.

[0079] Thirdly, the present invention provides an electronic device comprising:

[0080] At least one processor; and a memory storing instructions that, when executed by the at least one processor, cause the at least one processor to perform the sample concentration detection method.

[0081] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to perform the sample concentration detection method.

[0082] Compared with existing technologies, the sample concentration detection method provided by this invention achieves physical co-sampling of four dimensions—excitation wavelength, spatial bias, temporal gating, and modulation coding—through a pre-defined mechanism of each time window and a global clock reference point. This ensures that when calculating the difference in any dimension, the physical states of the other dimensions remain locked and unchanged. The background noise in the difference operation is highly correlated, achieving noise cancellation rather than error superposition. This fundamentally solves the problem of multivariate coupling error amplification in existing "temporally adjacent" independent sampling schemes, thereby enabling efficient and accurate acquisition of the concentration of the analyte in the sample. Attached Figure Description

[0083] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0084] Figure 1This is a timing diagram of a sample concentration detection method in one embodiment of the present invention;

[0085] Figure 2 This is a flowchart of a sample concentration detection method in one embodiment of the present invention;

[0086] Figure 3 This is a structural block diagram of a sample concentration detection system according to another embodiment of the present invention;

[0087] Figure 4 This is a structural block diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0088] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0089] Unless otherwise expressly stated, throughout the specification and claims, the term "comprising" or its variations such as "including" or "comprises" shall be understood to include the stated elements or components without excluding other elements or other components.

[0090] In one implementation, the photoelectric detection method for sample concentration detection can employ multiple independent modules, such as frequency-shifted excited Raman differential spectroscopy (SERDS), spatially biased Raman spectroscopy (SORS), and time-gated detection. These modules sequentially perform wavelength switching, spatial bias channel scanning, and time-gated acquisition within different time windows. Finally, a global clock timestamps each frame of data, and alignment is performed in the post-processing stage to calculate the differential spectrum and eliminate background interference. However, since the sampling in each dimension does not occur simultaneously physically, the spatial bias state and time-gated state are difficult to maintain strictly constant between two adjacent wavelength samples when calculating the SERDS differential. Similarly, when calculating the SORS differential, the wavelength state and time-gated state can introduce unexpected fluctuations. Therefore, this technical solution inevitably leads to the superposition of multi-dimensional sampling noise during differential calculations, resulting in error propagation and accumulation. According to the error propagation law, the noise from independently sampled dimensions cannot be eliminated through post-processing alignment, severely limiting the detection accuracy of weak Raman signals.

[0091] The inventors have discovered that the main drawback of the prior art is that the sampling of each dimension is essentially temporally adjacent rather than physically simultaneous. Even if a global clock is introduced for timestamp alignment, it cannot be guaranteed that the physical state of the other dimensions is truly locked at the instant of differential operation.

[0092] Based on this shortcoming, the inventors proposed a novel technical approach: using the laser wavelength switching event as the reference event for global sampling. During the hard-lock period before the event occurs, the spatial bias channel, time gating parameters, and modulation coding state are pre-configured and locked by a unified modulation clock to ensure that all other dimensions are in a deterministic and stable physical state at the instant the wavelength switching occurs, thereby achieving co-sampling of SERDS, SORS, and time gating at the same physical moment.

[0093] Building upon this foundation, the system further constructs rigorous multidimensional partial derivative approximation conditions, resulting in a high correlation between background noise in the difference operations and proactively achieving noise cancellation rather than error accumulation. This technical approach overturns the traditional paradigm of first performing time-separated independent measurements and then aligning the differences, transforming the multidimensional sampling problem from posterior data alignment to prior physical state construction. Its core idea is to ensure physical simultaneity through proactive pre-setting locking rather than passive timestamp synchronization. This idea is not publicly disclosed or implied in existing technologies and cannot be derived simply by combining existing modules.

[0094] Please refer to Figure 2 The diagram shown is a schematic flowchart of sample concentration detection according to an embodiment of the present invention. The sample concentration detection method specifically includes the following steps:

[0095] S201: Irradiate a sample containing the analyte with a laser based on a switchable excitation wavelength, and determine the laser wavelength switching time as a clock reference point;

[0096] In one embodiment of the present invention, at least one of the following embodiments can be combined, that is, any one of the following methods can be executed sequentially or alternately within the same sampling period.

[0097] Firstly, by utilizing the physical property that the position of the Raman scattering peak shifts with the excitation wavelength while the fluorescence background hardly shifts with the wavelength, two slightly different excitation wavelengths are used alternately to collect spectra. The two spectra are then subtracted to eliminate the fluorescence background, thereby achieving noise suppression in frequency-shifted excitation Raman differential spectroscopy.

[0098] Secondly, by spatially separating the laser excitation point and the Raman signal collection point by a certain distance, the spectra at multiple offset distances are collected and differential operations are performed to enhance the Raman scattering contribution from deep tissues, suppress shallow fluorescence and scattering interference, and achieve depth resolution of spatially offset Raman spectra.

[0099] Third, a single-photon detector is activated to collect Raman photons within an extremely short time window after the pulsed laser reaches the sample, while the detector is turned off during the subsequent period of continuous fluorescence emission. By utilizing the difference between the picosecond-level instantaneous response of Raman scattering and the nanosecond-level long lifetime of fluorescence, Raman signals are selectively collected, thereby achieving fluorescence suppression in time-gated Raman detection.

[0100] Fourth, the modulation coding state of the laser is pre-configured and locked within the first time window before wavelength switching. Under the premise that other conditions are constant, different modulation states are switched sequentially or alternately and spectra are collected respectively. By utilizing the difference in sensitivity of the sample Raman response and fluorescence background to the modulation mode under different modulation states, differential operation is performed on the spectra of two frames under different modulation states. This can eliminate common-mode noise background that is unrelated to the modulation mode and selectively extract Raman signal components that are sensitive to specific modulation dimensions, thereby realizing signal separation and signal-to-noise ratio improvement in modulation-coded differential Raman detection.

[0101] This invention aims to eliminate the problem of multidimensional differential error propagation and accumulation caused by independent sampling, and to achieve true multidimensional co-sampling at the physical layer. Therefore, in one embodiment of this invention, the laser wavelength switching event is used as the physical trigger signal for the entire multidimensional co-sampling process. By explicitly defining the wavelength switching moment as a clock reference point, the system can perform state preset locking and effective data acquisition before and after this reference point, thereby ensuring that the states of each dimension (spatial bias, time gating, modulation coding) are simultaneously determined at the moment of wavelength switching at the physical layer.

[0102] It is important to note that the laser wavelength plays a dual role in both the time and frequency domains in this invention. In the time domain, the switching period of the laser wavelength is used as the reference dimension for the global modulation clock. The wavelength switching moment is the clock reference point, and the system uses this to drive the spatial bias, time gating, and modulation coding state components to be preset and locked within the first time window before the wavelength switching. In the frequency domain, the laser wavelength itself switches between at least two excitation wavelengths during adjacent sampling cycles. Utilizing the characteristic that the Raman scattering peak shifts with the excitation wavelength while the fluorescence background hardly shifts with the wavelength, the spectra acquired at adjacent wavelengths are differentially analyzed to eliminate the fluorescence background, achieving frequency-shifted excitation Raman differential. Because the laser wavelength changes abruptly at the moment of switching, rather than being pre-locked before switching as in the other three dimensions, this invention defines the spatial bias channel state, time gating parameters, and modulation coding state as state components locked within the first time window, while using the laser wavelength as a separate trigger dimension for defining the clock reference point and as one of the target dimensions that can participate in the differential analysis. The three state components are ready before the wavelength switching, and the superimposed wavelength reaches a definite value at the moment of switching, thereby realizing the simultaneous determination of wavelength, spatial offset, time gating, and modulation coding in four dimensions at the clock reference point.

[0103] Specifically, in one embodiment, spatial offset can be achieved using a fixedly arranged multi-channel collecting fiber array, with different offset distances. For each fixed fiber optic channel in the array, channel switching is accomplished by electronic switches, without any mechanical moving parts. The activation sequence of each channel is uniformly preset by a global modulation clock determined by a clock reference point, ensuring that the spatial bias state remains stable and determined at the wavelength switching moment.

[0104] Time gating can be achieved by configuring the detector shutter width and opening delay. Examples include single-photon avalanche diode (SPAD) gating detection, and this embodiment of the invention does not limit this. The time gating parameters are pre-configured by a global modulation clock determined by a clock reference point before the wavelength switching event, ensuring that the gating state and wavelength state are determined simultaneously, achieving physical co-sampling of the time and wavelength dimensions.

[0105] In this embodiment of the invention, the modulation coding state may include, but is not limited to, amplitude modulation, phase modulation, polarization modulation, and spatial mode modulation. Amplitude modulation can achieve periodic changes in laser power through driving current modulation; phase modulation can change the phase of the optical field through an electro-optic modulator; polarization modulation can switch the polarization state through a liquid crystal waveplate or a rotating waveplate; and spatial mode modulation can switch the laser beam mode (Gaussian mode, ring mode, etc.). All modulation methods together constitute the modulation coding space. The modulation and coding state is preset and stabilized by the global modulation clock before wavelength switching, ensuring that the coding state and wavelength state are determined simultaneously.

[0106] In one embodiment, the system not only supports switchable output of at least two excitation wavelengths, but also further performs amplitude modulation on the laser power. Specifically, the change of laser power over time can be expressed as:

[0107]

[0108] in, Average power; Modulation amplitude; This is the amplitude modulation frequency. satisfy This size relationship ensures that the time interval between two adjacent wavelength switching events contains multiple complete amplitude modulation cycles, providing support for power normalization of the reference optical path in subsequent steps.

[0109] For example, take to This ensures that during the stable output of each wavelength segment, the laser power has undergone tens to hundreds of sinusoidal fluctuations.

[0110] S202: In the first time window before the clock reference point, configure and lock each state component as the target value; in the second time window after the clock reference point, perform spectral data acquisition based on the locked state components; the state components include spatial offset channel state, time gating parameters, and modulation coding state.

[0111] This invention, based on a clock reference point configuration, further divides each sampling period into multiple time windows. The first time window, located before the clock reference point, is used to preset and lock the target values ​​of each state component; the second time window, located after the clock reference point, is used to perform actual spectral data acquisition in the locked state. The state components include at least the spatial offset channel state, time gating parameters, and modulation coding state.

[0112] A global clock, established based on a clock reference point, serves as the highest-priority instruction at the physical layer. Through a first time window, it ensures absolute consistency of the multi-dimensional sampling state at the instant of wavelength transitions. The system uses the wavelength switching period as the reference unit for the global clock, achieving frequency alignment and master-slave drive. Before each wavelength switching event occurs, the global clock drives each dimension into a locking sequence.

[0113] Specifically, the dimensions for entering the locked sequence can include, but are not limited to, spatial bias channel state, temporal gating parameters, and modulation coding state. The specific selection is related to the spectral processing method to be fused within the sampling period.

[0114] Taking an embodiment that integrates multiple processing methods as an example, before wavelength switching is triggered, the system pre-activates the electronic switch of the corresponding spatial bias channel to ensure that the switch setup time is fully contained within the first time window, thereby ensuring that the spatial dimension is in a physically steady state at the moment of wavelength switching. Simultaneously, the system pre-configures the timing reference and gating parameters of the single-photon avalanche diode to ensure that the time-to-digital converter has completed timing alignment and is in a ready-to-fire state at the moment of wavelength switching at a preset resolution. Furthermore, the system pre-solidifies the phase or amplitude modulation state to ensure that the modulation coding features and wavelength components are highly coupled at the physical layer.

[0115] It should be noted that, since the first time window is defined as a state preset window that is forcibly executed by the global modulation clock before the wavelength switching event, its duration must meet the following timing constraints:

[0116]

[0117] in, The setup time for the high-speed analog switch of the j-th spatial channel; The preset time for the latest dimension to begin.

[0118] This constraint guarantees that, when When a wavelength transition event occurs, All three dimensions (spatial bias, temporal gating, and modulation coding) have completed the physical transition from the old state to the target state and reached a steady state; no dimension is in a transitional state. The first time window is the only quantifiable means of verifying the physical homology claim of this invention at the temporal level. If the above constraints cannot be met, then it is impossible to... The four-dimensional state is determined at every moment, and the physical homology condition does not hold mathematically.

[0119] Following the wavelength switching event, the system enters a second time window. Within this window, all locked state components remain stable, and the system initiates effective integration acquisition while simultaneously performing power normalization of the reference optical path. Typically, the second time window is on the order of milliseconds, allowing for the accumulation and integration of millions of picosecond pulse trains to obtain raw spectral data with a sufficient signal-to-noise ratio.

[0120] It should also be noted that, to avoid timing confusion between two adjacent clock reference points, the first time window within each sampling period and the second time window of the previous period do not overlap on the time axis. In the higher-level embodiment of this invention, the following time window configuration rules are adopted: the end point of the first time window is the clock reference point; the start point of the second time window is the same clock reference point; and the end point of the second time window is the start point of the first time window corresponding to the next clock reference point. In other words, adjacent sampling periods are consecutive, and the first time window of the next period begins immediately after the second time window of the previous period ends. However, the first time window and the second time window within the same period are located on opposite sides of the clock reference point and do not overlap.

[0121] Through the above timing configuration, the system has a dedicated first time window for state presetting and locking before each wavelength switching event occurs, and immediately enters the second time window for effective spectral acquisition after the switching event occurs. This ensures that at the physical instant of wavelength jump, the spatial bias channel, time gating parameters, and modulation coding state are all stable at the target values, thus providing strict physical partial derivative conditions for the subsequent construction of multidimensional homologous spectral tensors.

[0122] It is understandable that transient parasitic oscillations will inevitably occur during wavelength switching. To avoid their impact on the sampling process, a third time window, or stabilization window, is configured within the sampling period in a preferred embodiment. The third time window is located after the wavelength switching moment and before the start of the second time window, serving as a quiet period to isolate and eliminate parasitic oscillations caused by hardware switching, ensuring the quality of subsequently acquired spectral data.

[0123] The third time window is defined as Then, the second time window begins ( The duration of the preceding silent period must simultaneously satisfy the following bilateral constraints:

[0124]

[0125] in, Let be the decay time constant of the parasitic oscillations at the instant of electronic switch switching. This represents a single spectral integration period. The left-hand constraint ensures that parasitic oscillations have sufficiently decayed before acquisition begins, while the right-hand constraint ensures that the third time window's occupation of the integration time is negligible. The third time window and the first time window together constitute the physical boundary of the second time window: the first time window ensures readiness before switching, and the third time window ensures stability after switching.

[0126] In one implementation, a third time window can be determined starting from a clock reference point, and each state component is kept locked within the third time window. The duration of the third time window is greater than or equal to the duration required for the parasitic oscillation to decay to a first threshold. After the end point of the third time window and before the start point of the next first time window, a second time window is determined.

[0127] Furthermore, to avoid timing confusion between two adjacent clock reference points, in one embodiment, the starting point of the second time window can be determined as the ending point of the third time window; the ending point of the second time window is the starting point of the next first time window; wherein, the first time window, the second time window, and the third time window do not overlap with each other on the time axis.

[0128] That is, refer to Figure 1 The timing diagram is shown below. Combining the constraints of the first time window mentioned above, the relationship between the two forms a complete chain of timing inequalities: ,in[ , [This is the first time window,] , [This is the third time window] , [This is the second time window.]

[0129] For example, in one specific embodiment, the system uses two switchable excitation wavelengths, 830 nm and 832 nm, with a wavelength switching period of 100 ms. The specific implementation process of step S201 is as follows:

[0130] The laser driver circuit receives periodic trigger signals from the global modulation clock. At the beginning of each sampling period, the system first enters the first time window. Within the first time window, the system sequentially or in parallel performs the following state preset operations: sending a channel selection command to the spatial bias module to activate the electronic switch corresponding to the collection fiber channel with an offset distance of 0.5 mm. The setup time of this switch is 15 ns, and it must be stably turned on before the wavelength switching event; configuring the current gating parameters to the time gating module, setting the SPAD opening delay to 100 ps, ​​the gating window width to 200 ps, ​​and aligning the TDC timing reference with the global clock; and simultaneously sending a command to the modulation and coding module to set the laser polarization state to horizontal linear polarization and wait for the liquid crystal waveplate to complete rotational relaxation.

[0131] After the above preset operations are completed, the system confirms that the first time window constraint is met—that is, the time difference from the latest preset completion time to the next wavelength switching event is not less than the maximum value of the setup time of all switched devices. Subsequently, at the same physical moment when the wavelength switching event occurs, the system executes a jump of the laser wavelength from the current state (e.g., 830 nm) to another state (e.g., 832 nm). At this time, the spatial bias channel is maintained in the 0.5 mm conduction state, the SPAD gating parameters are latched and the timing reference is aligned, and the polarization modulation state is solidified—that is, at the instant of wavelength switching, the four-dimensional state is determined simultaneously, and the physical co-sampling condition is achieved at this moment.

[0132] After wavelength switching is complete, the system enters the third time window, during which spectral acquisition is prohibited to allow for the sufficient attenuation of parasitic charge injection spikes caused by electronic switch switching and transient power overshoot from the laser. After the third time window ends, the system opens the effective integration acquisition window (second time window). The SPAD gate accumulates Raman scattered photons under set parameters, with an integration time of 50 ms. Simultaneously, the average laser power within this integration window is recorded by the photodetector in the reference optical path.

[0133] After this round of data acquisition is completed, the system will tagged the raw spectrum obtained in the current four-dimensional state with a timestamp and a homology determination label along with the reference power reading, and temporarily store them in the data buffer. Subsequently, the system will switch to the next wavelength state or other non-wavelength dimension state combinations and repeat the above process in the same time sequence.

[0134] S203: Based on the spectral data and the state components corresponding to the spectral data, construct a multidimensional homogeneous spectral tensor; wherein, the multidimensional homogeneous spectral tensor contains at least one pair of spectral data samples that vary along a single target dimension and are locked in the other dimensions, and the target dimension is the laser wavelength or any state component;

[0135] It should be noted that the multidimensional homologous spectral tensor is the core data structure of this invention for systematically organizing spectral data obtained after physical homologous co-sampling and power normalization. The data dimensions that make up the multidimensional homologous spectral tensor are also related to the spectral processing method to be fused within the sampling period.

[0136] For example, in one specific embodiment, the dimensions of the multidimensional homogeneous spectral tensor may include: a wavelength dimension, a spatial bias dimension, a time-gated dimension, and a modulation coding dimension. The wavelength dimension corresponds to the multiple excitation wavelength states used in frequency-shifted excited Raman differential spectroscopy. The system can generate at least two different excitation wavelengths using a built-in switchable laser, and the wavelength switching event simultaneously serves as the reference trigger signal for the global modulation clock. Each index position along this dimension of the tensor corresponds to a specific excitation wavelength value.

[0137] The spatial bias dimension corresponds to different spatial bias distances between the excitation spot and the end face of the collecting fiber in a spatially biased Raman spectrum. The system employs a fully electronic multi-channel collecting fiber array, with each channel corresponding to a fixed bias distance. Switching between channels is completed by a high-speed electronic switch within a hard-lock period, without any mechanical moving parts. Each index position of the tensor along this dimension corresponds to a specific spatial bias channel.

[0138] The time-gated dimension corresponds to different opening delays or window widths of the single-photon detector gating window in time-gated Raman detection. The system pre-configures the SPAD gating parameters before wavelength switching using a global modulation clock, ensuring the detector responds to incident photons only within the preset time window. Each index position of the tensor along this dimension corresponds to a defined set of time-gated parameters.

[0139] The modulation coding corresponds to a modulation state that can be actively switched under the laser excitation conditions. Modulation methods include, but are not limited to, amplitude modulation, phase modulation, polarization modulation, or spatial mode modulation. Each modulation state is preset and locked by a global clock during the hard-lock period. Each index position of the tensor along this dimension corresponds to a specific modulation coding state.

[0140] It should also be noted that although the system provides hardware guarantees for simultaneous locking of various dimensions at the timing level by configuring the first and third time windows, in actual engineering operation, there are problems such as the electronic switch setup time may deviate from the factory calibration value due to temperature drift or aging, and the phase or polarization state of the modulation device may undergo slight relaxation due to environmental disturbances. These factors may cause the physical state of some non-target dimensions to not be strictly locked at the moment of wavelength switching or within the effective acquisition window, or to deviate slightly beyond the allowable range.

[0141] Therefore, in one embodiment of the present invention, before constructing a multidimensional homogeneous spectral tensor based on the spectral data and the corresponding state components, homogeneity verification can be performed on the spectral data after each frame is acquired and the actual physical states of each state component at the acquisition time. This determines whether the current acquisition satisfies the homogeneity condition that the non-target dimension has the same physical state between the two frames participating in the difference.

[0142] In this context, when the system aims to calculate the difference along a specific dimension, that dimension is called the target dimension, while the other three dimensions are called non-target dimensions. For example, when the system performs a difference operation to eliminate fluorescence background, it can compare two adjacent excitation wavelengths. and The spectrum collected below. At this time, the excitation wavelength... The dimensions that change are the target dimensions; while the spatial bias channel state r, the time gating parameter t, and the modulation coding state m must remain strictly consistent and stable during the two frames of spectral acquisition and must not change. Therefore, these three dimensions are the non-target dimensions.

[0143] Understandably, when a system integrates differential processing capabilities across multiple dimensions, including excitation wavelength, spatial bias, temporal gating, and modulation coding, the same set of acquired spectral data may be used to calculate different types of gradient features. For example, a particular frame of spectrum may be paired with the spectrum of another wavelength to calculate the wavelength gradient, or it may be paired with the spectrum of another spatial bias channel to calculate the spatial gradient. In different differential operations, the role of this frame of spectrum differs, and the set of its non-target dimensions also changes accordingly.

[0144] Specifically, in one implementation, the homology time series index can be defined as:

[0145]

[0146] in, This represents the actual deviation of the i-th non-target dimension within the partial derivative sampling window; This represents the upper bound of the calibration error for the corresponding dimension.

[0147] like If ≤ 1, the sample is marked as a homologous sample, and partial derivative calculation or model training can proceed; if If the value is greater than 1, the sample group is marked as a non-homologous sample and cannot be used for partial derivative solving or model training. The system may choose to discard the sample or trigger a resampling process.

[0148] Furthermore, in another implementation, the homology determination system can be expanded into multiple independent criteria, which together constitute a complete acceptance framework. These independent criteria may include: temporal homology criteria, amplitude homology criteria, noise correlation criteria, and reference optical path consistency criteria.

[0149] Specifically, the temporal origin criterion is used to verify whether the timing difference between the state locking time and the wavelength switching event of each dimension meets the constraint of the first time window. The system records the state locking start time of each non-target dimension and calculates the time difference from that time to the next clock reference point. If the time difference is greater than or equal to the calibrated value of the locking time required for that dimension, then that dimension meets the temporal origin requirement.

[0150] For example, the system is configured with three spatial bias channels, whose high-speed analog switch calibration setup times are 12 nanoseconds, 15 nanoseconds, and 18 nanoseconds, respectively. The wavelength switching period of the global modulation clock is 100 milliseconds, and the first time window duration is 2 microseconds. Within the first time window, the system sequentially sends preset commands to each channel, i.e., the locking start times of each channel are as follows: , , For the channel with the longest setup time (18 nanoseconds), its locking start time... To the next clock reference point Time difference The duration of the entire first time window is set to 2 microseconds (2000 nanoseconds), which is much greater than 18 nanoseconds. Therefore, the time sequence origin criterion is passed.

[0151] If the actual setup time of a switch is extended to 25 nanoseconds due to temperature drift, the criterion still passes because the time difference ΔT between the start of the lockout and the reference point is still 2000 nanoseconds, which is much greater than 25 nanoseconds. This indicates that the system timing design has sufficient margin to tolerate a certain degree of device parameter drift.

[0152] However, if the system compresses the first time window to 20 nanoseconds due to a fault or special configuration, and the locking start time of the channel happens to be at the beginning of the window, then ΔT = 20 nanoseconds, while the actual setup time requires 25 nanoseconds. In this case, 20 nanoseconds is less than 25 nanoseconds, so the criterion fails, and the sample is marked as time-series non-homogeneous.

[0153] The amplitude co-origin criterion is used to verify whether the deviations of physical quantities in each non-target dimension within the effective integration window are within the upper bound of the calibration error. The system continuously monitors the actual physical quantities of each state component within the effective acquisition window, such as the actual on-resistance of the spatial bias channel, the actual opening delay of the time-gated system, and the actual phase or amplitude value of the modulation code, and calculates the maximum deviation relative to the locked target value. If the ratio of the maximum deviation of each non-target dimension to the corresponding upper bound of the error is not greater than 1, the criterion passes.

[0154] For example, the upper bound of the system calibration time gating error. The value is 5 picoseconds. Within the effective integration window, the actual value of the SPAD gating delay was observed to fluctuate around the target value of 100 picoseconds, with a maximum deviation of 3 picoseconds. The amplitude source criterion is passed. If the deviation reaches 7 picoseconds due to power supply noise, the ratio 1.4 > 1, the criterion is not passed, and the sample cannot be used for subsequent differencing.

[0155] The noise correlation criterion is used to verify whether the correlation coefficient of the differential frame with the background noise is higher than a preset threshold. After acquiring two frames of spectra at adjacent wavelengths or different spatial biases, the system selects a pure background band from the spectra that does not contain any Raman characteristic peaks, extracts the photon counting sequence of the two frames in that band, and calculates the Pearson correlation coefficient. .

[0156] The correlation coefficient between the differential frames and the background noise can be expressed mathematically as follows:

[0157]

[0158] in, and These represent the photon count sequences of the pure background region in two frames of spectral data collected under the same non-target dimension state but different target dimensions, respectively. for and covariance; for The variance.

[0159] For example, system calibration The correlation coefficient is 0.95. For two frames of spectrum acquired in a certain instance, the calculated background band correlation coefficient is 0.97, which is greater than 0.95, indicating that the background noise of the two frames is highly correlated and can be effectively canceled after differencing, thus passing the criterion. In another embodiment, if the correlation coefficient drops to 0.82 due to a sudden change in ambient light or sample movement, which is below the threshold, the background covariance structure is damaged, and the effect of differential noise cancellation cannot be guaranteed. The sample should be downgraded or discarded directly, and the system can trigger resampling.

[0160] The reference optical path consistency criterion is used to verify whether the fluctuation of the reference optical path power monitoring value within the effective integration window is lower than the calibration threshold. Within the second time window, the system continuously acquires the output value of the photodetector in the reference optical path and calculates the average value of all sampled values ​​within that window. and standard deviation Calculate the relative standard deviation of fluctuation. / And compare the relative standard deviation of volatility and the calibration threshold. That's all.

[0161] In one specific embodiment, a calibrated threshold is used. The standard deviation is 1%. Within a valid acquisition window, the average power of the reference optical path is 5.2 microwatts, the standard deviation is 0.03 microwatts, and the relative fluctuation is 0.03 / 5.2≈0.58%, which is less than 1%. Therefore, the criterion is met, and the power normalization accuracy of this frame's spectrum meets the requirements and can be used for subsequent gradient feature generation. If the power fluctuation is severe due to laser mode switching or fiber microbending, and the relative fluctuation reaches 2.5%, exceeding the threshold, the power normalization accuracy is insufficient. In this case, only the original observation value is retained for absolute amplitude analysis, and it is not used for gradient feature calculation; or the entire frame is discarded and resampled.

[0162] It should be noted that the reference optical path shares the same laser source as the main optical path irradiating the sample, and the integration windows for the reference optical path acquisition and spectral acquisition are synchronized. Since there is a fixed linear proportional relationship between the laser power received by the detector corresponding to the reference optical path and the laser power irradiating the sample surface, this proportional relationship depends only on the optical coating characteristics of the beam splitter and the optical path coupling efficiency, remaining constant under stable system operating conditions. Therefore, by monitoring the power of the reference optical path, the relative laser power change at the irradiated sample can be obtained equivalently, meeting the accuracy requirements of power monitoring for power normalization.

[0163] The quantization thresholds for the four criteria mentioned above can be determined through repeated measurements of standard reference samples during the system's factory calibration phase and stored in the system parameter table. During real-time detection, the system automatically calculates the above indicators and compares them with the thresholds for each frame or set of differential sample pairs acquired. Samples that meet all criteria can be identified as homologous samples, written into the four-dimensional spectral tensor, and used for subsequent gradient differencing and model inference; samples that fail any criterion cannot proceed to partial derivative solving or model training. The system will discard the sample, trigger resampling, or dynamically adjust the third time window and gating parameters according to a preset strategy and then attempt to acquire it again.

[0164] This multi-criteria collaborative homology verification system, on the basis of hardware timing assurance, adds physical state deviation monitoring, noise covariance quality assessment and power normalization accuracy control, which together ensure that the data entering the differential operation has true partial derivative validity from multiple dimensions, thus making this invention different from existing solutions that only rely on timestamp alignment.

[0165] In another embodiment, the sample concentration detection method of the present invention further includes: normalizing the spectral data to eliminate the multiplicative perturbation of the Raman signal amplitude caused by laser power drift. The order in which the spectral data normalization process and the aforementioned homology sample detection procedure are performed is not restricted. Preferably, to improve computational efficiency and avoid meaningless operations on invalid data, a rapid and low-cost homology determination can be performed first. After confirming the data is qualified, the relatively more computationally expensive normalization process is then performed, and finally, the data is written into a tensor.

[0166] In one exemplary embodiment, the normalization of the spectral data based on laser power includes splitting the laser beam emitted from the laser source into a main optical path and a reference optical path for irradiating the sample. The average power of the reference optical path is determined within a second time window for acquiring the spectral data of the current sample; the spectral data of the current sample is normalized based on the average power to obtain spectral data at unit excitation power.

[0167] It is understandable that the reference optical path in this step can be reused from the reference optical path consistency criterion mentioned above. Alternatively, another beam splitter can be established to split the laser source into a main optical path, a first reference optical path, and a second reference optical path. The first reference optical path is used for homology detection, and the second reference optical path is used for spectral data normalization.

[0168] The aforementioned normalization of the spectral data based on laser power can be expressed mathematically as follows:

[0169]

[0170] in, This represents the spectral data after power normalization. This refers to the raw spectral signal directly acquired within the second time window, under otherwise constant conditions. This refers to the readings from the photodetector in the reference optical path within the same second time window; For each dimension of the spectral data (laser wavelength, spatial bias, time gating, modulation coding).

[0171] Furthermore, the spectral data identified as homologous samples, and / or the spectral data after normalization based on the laser power, along with the corresponding state components of the spectral data, are written into the corresponding positions of the multidimensional homologous spectral tensor to form the multidimensional homologous spectral tensor.

[0172] The multidimensional homogeneous spectral tensor constructed based on the above implementation method satisfies the condition that the states of all dimensions are determined simultaneously at the same time. Under this condition, in the difference operation of any dimension, the physical states of the other dimensions remain unchanged simultaneously. The difference noise is correlated, thus noise cancellation can be achieved instead of error accumulation.

[0173] S204: Input the multidimensional homology spectral tensor into the evaluation model to obtain the concentration of the component to be measured in the sample output by the evaluation model.

[0174] It is understandable that each element in the multidimensional homology spectral tensor is spectral data obtained under unit excitation power after power normalization, and its intensity has a physical mapping relationship with the concentration of the analyte in the sample determined by the Raman scattering cross section. The machine learning model (evaluation model) has the ability to automatically learn complex nonlinear mappings from high-dimensional data. During the training phase, the model takes the multidimensional homology spectral tensor as input and known concentration labels as supervision signals. Through optimization algorithms, it automatically discovers implicit feature patterns related to concentration in the spectral data, and can establish an end-to-end mapping from the multidimensional homology spectral tensor to concentration values ​​without explicitly performing difference operations.

[0175] Furthermore, the time window mechanism of this invention ensures that the acquisition conditions of each frame of data in the multidimensional homologous spectral tensor are strictly controlled at the physical level. The consistency and repeatability of the data are significantly better than those of the traditional time-division sampling scheme, providing higher quality training samples for machine learning models, thereby enabling them to directly extract effective information from the multidimensional homologous spectral tensor and output concentration estimates.

[0176] It should be noted that in biological tissue detection scenarios, the fluorescence background intensity is typically several orders of magnitude higher than the target Raman signal. Machine learning models need to learn changes in the Raman signal from the background component, which may result in insufficient sensitivity to target parameters and limited generalization ability.

[0177] Therefore, in an exemplary embodiment of the present invention, the evaluation model first calculates the gradient features of the spectral data along at least one dimension of the multidimensional homologous spectral tensor, then combines these gradient features with the original spectral data into a joint input vector, and then regresses the concentration of the analyte in the sample based on the joint input vector.

[0178] The gradient features along at least one dimension may include a first gradient feature and / or a second gradient feature. The first gradient feature is the first-order rate of change along a single dimension of the multidimensional homogeneous spectral tensor, and its physical meaning is the approximation of the univariate partial derivative of the spectral response with respect to the target dimension under the condition that the other dimensions are physically locked. The second gradient feature is the second-order mixed rate of change along two different dimensions of the multidimensional homogeneous spectral tensor, and its physical meaning is the cross-dimensional coupling information obtained by successively approximating the univariate partial derivatives in two different dimensions.

[0179] Because the time window mechanism of this invention ensures strict consistency of the physical state of non-target dimensions when calculating differences in any dimension, the aforementioned first and second gradient features mathematically approximate the true partial derivatives, rather than incorporating perturbation terms from other dimensions as in traditional time-division independent sampling schemes. The calculation and use of gradient features enable the system to transform weak Raman signals submerged in a strong fluorescence background in the original spectrum into structural features sensitive to target parameters, while preserving the low-frequency amplitude information of the original spectrum. The combination of these two features constitutes a complementary multi-scale representation, significantly improving the machine learning model's ability to identify target parameters and its numerical stability.

[0180] For example, for The first characteristic gradient may include the frequency domain gradient. Spatial gradient Time gradient and modulation coding gradient One or more of the features; the second gradient features may include One or more of them.

[0181] In one implementation, the joint gradient is constructed based on the optimal weight vector obtained by solving the spatiotemporal coupling Vandermonde matrix. The joint gradient can be expressed mathematically as follows:

[0182]

[0183] in, , , , These correspond to the frequency domain gradient, spatial gradient, temporal gradient, and modulation coding gradient, respectively. This joint gradient can simultaneously suppress fluorescence background, deep scattering, and temporal drift noise under optimal weighting conditions.

[0184] This invention extends the one-dimensional Vandermonde method to multi-dimensional coupling scenarios, constructing a spatiotemporal coupling extended matrix whose basis function set simultaneously includes monomials and cross-dimensional cross terms for each sampling dimension. Typical basis function forms encompass constant terms, first and second-order wavelength terms, first and second-order spatial offset distance terms, first and second-order time-gated delay terms, first-order modulation and coding state terms, as well as multi-dimensional coupling terms such as cross-products of wavelength and spatial offset, cross-products of wavelength and time-gated control, and cross-products of spatial offset and time-gated control. By solving the linear equations corresponding to this extended matrix, the optimal weight coefficients for each dimension are obtained, including wavelength dimension weights, spatial dimension weights, time dimension weights, modulation and coding dimension weights, as well as cross-dimensional weights for wavelength and spatial offset, wavelength and time-gated control, and spatial offset and time-gated control.

[0185] The above basis functions can be written as:

[0186]

[0187] It should be noted that the spatiotemporal coupling extension matrix is ​​essentially an asymmetric, non-uniform sampling basis function matrix. In the specific devices implementing this invention, the spatial offset channel layout distance is often non-uniformly distributed due to the physical arrangement limitations of the fiber array, and the time-gated delay can be dynamically adjusted according to the actual signal-to-noise ratio requirements rather than being uniformly stepped.

[0188] Traditional uniform sampling Vandermonde matrices require sampling points to be arranged at equal intervals along the target dimension, which cannot adapt to the aforementioned non-uniform sampling conditions. The extended matrix constructed in this invention directly uses the actual sampling coordinate values ​​of each dimension as the input variables of the basis functions, naturally compatible with non-uniform step sizes and asymmetric sampling grids, eliminating the need for interpolation and resampling of the collected data, thereby avoiding human error introduced by resampling.

[0189] Based on the above solution equations The obtained multidimensional weight vector ,in, It is a basis function matrix composed of the values ​​calculated by the basis functions at specific sampling points; It is a constant vector representing the coefficients of the objective differential operator.

[0190] In the multidimensional case of this invention Let w be the target derivative coefficient vector corresponding to the multidimensional basis functions. It defines the mathematical conditions that w must satisfy. For example, if the expectation is... It is an operator that achieves "wavelength direction smoothing and first derivative calculation", and the positions corresponding to 1 and λ in C will have specific non-zero values.

[0191] The system linearly combines first-order gradient features from each dimension with cross-dimensional coupled gradient features according to their corresponding weights to construct an optimal joint gradient. This joint gradient is not a simple feature concatenation or equal-weighted superposition, but rather a joint suppression of noise in each dimension based on the optimal weighting coefficients in the least-squares sense: the weighting coefficients automatically tilt towards gradient channels with high signal-to-noise ratio and good background suppression, while assigning lower weights to gradient channels with higher noise or instability under the current sample conditions. Under the action of this joint gradient, the gradual variation components of fluorescence background, the path difference effect of deep scattering, and the low-frequency perturbations introduced by time drift can be simultaneously and optimally suppressed within a unified numerical framework, providing a single high signal-to-noise ratio feature input channel for subsequent concentration estimation.

[0192] It should also be noted that the raw observation data obtained through multidimensional co-sampling differs fundamentally in physical dimensions from the gradient features of each dimension, and therefore cannot be directly mixed and input into a neural network. Taking blood glucose detection as an example: the original Raman peak intensity is measured in photon counts or radiation intensity (W / sr), while the partial derivative with respect to the excitation wavelength... The dimension is counts / nm, and the partial derivative with respect to spatial bias is... The dimension is counts / mm, and the partial derivative with respect to time gating is... The units of measurement are counts / ns. The units of measurement for each feature are different and their orders of magnitude vary greatly. Directly concatenating them will cause the features with larger units to dominate the gradient direction during model optimization, while the actual contribution of the features with smaller units will be submerged.

[0193] Therefore, in one embodiment, before performing concentration analysis on the evaluation model, it is necessary to normalize the dimensions of the feature quantities in each dimension. Preferably, before performing dimensional normalization, the multiplicative disturbance introduced by laser power fluctuations is eliminated in real time based on the set reference optical path using the corresponding method in S203. Subsequently, the preset fluctuation calibration values ​​of the spectral data and the gradient features are obtained, and the spectral data and the gradient features are standardized (dimensional normalized) based on the fluctuation calibration values.

[0194] For example, the raw observations after power normalization And gradient features of each dimension, the measurement standard deviation of each feature channel is estimated through repeated measurements during the system calibration phase. And use this as the normalization benchmark:

[0195]

[0196] in For the first One characteristic, Features The standard deviation of a single measurement under calibration conditions. Normalized characteristics. It is a dimensionless quantity.

[0197] Based on the normalization scheme corresponding to this embodiment, different physical quantities can be normalized according to their respective... Normalization to dimensionless quantities eliminates the problem of dimensional inconsistency. The normalized original observation components provide amplitude benchmarks and overall scale information, while the normalized gradient components provide structural features in parameter-sensitive differential directions. Both are expressed in terms of signal-to-noise ratio multiples, have consistent dimensions, and are comparable in magnitude, so they can be directly used as input to machine learning models without introducing dimensional bias.

[0198] At the same time, due to The signal-to-noise ratio (SNR) directly reflects the measurement noise level of the characteristic channel. In subsequent operations, channels with high SNR have relatively larger amplitudes after normalization, and can naturally be assigned higher weights in model optimization. Channels with low SNR have relatively compressed amplitudes after normalization, which can reduce their impact on the loss function.

[0199] about In one embodiment, during the system calibration phase, a uniform reference sample (such as a standard scatterer or ex vivo tissue phantom) can be repeatedly measured N times; the sample standard deviation of the N measurements for each feature channel is calculated, and the results are stored for normalization.

[0200] Preferably, Calibration should be performed under conditions similar to the actual usage environment and updated regularly to track long-term system drift.

[0201] It should be noted that the evaluation model described in this invention includes one or more machine learning models selected from deep neural networks, Gaussian process regression, or physical information neural networks. This invention does not impose any limitations on this.

[0202] Taking deep neural networks as an example, the system accumulates a large number of labeled concentration and spectrum paired samples over time. Under the condition of sufficient data, deep neural networks can approach the joint optimal solution of three theoretical levels: improved Fisher information, improved Jacobian matrix condition number, and reduced generalization error.

[0203] When using deep neural networks to estimate sample concentration, the deep neural network takes a joint input vector, which is standardized by fluctuation calibration, as input. This joint input vector is composed of the standardized original spectral components and the standardized gradient feature components, and outputs the estimated sample concentration.

[0204] In one embodiment of the present invention, the deep neural network architecture may employ an attention residual mechanism. The attention residual mechanism replaces fixed accumulation with attention aggregation of the outputs of previous layers, allowing each layer to selectively backtrack to earlier representations with learned, input-dependent weights, thus solving the problem of inter-layer information dilution.

[0205] Specifically, the deep neural network architecture divides the network layers into three processing blocks according to the physical feature hierarchy: the first processing block processes the normalized original spectral components, learning the amplitude reference and scale information of the overall optical state of the organization; the second processing block processes first-order gradient features, including wavelength-dimensional gradient, spatial-dimensional gradient, temporal-dimensional gradient, and modulation-coding gradient, learning the Raman peak structure after fluorescence cancellation, depth-resolved component differences, and time-gated lifetime features; the third processing block processes cross-dimensional gradient features, learning the second-order coupling structure between various physical dimensions. Attention residual mechanisms are used to aggregate the processing blocks, and the representations of the three processing blocks are adaptively weighted during the final concentration estimation, resulting in better expressive power than a fixed-weight fusion layer.

[0206] Preferably, the query vector for the attention residual mechanism is generated based on physiological background slow variables, including water peak intensity, temperature-sensitive Raman peak shift, collagen characteristic peak intensity, and fat characteristic peak intensity. This allows the attention weights between processing blocks to adaptively adjust according to the individual's current physiological state, such as skin hydration, tissue temperature, collagen background, and subcutaneous fat thickness.

[0207] For example, for samples with high fluorescence background, the system automatically increases the backtracking weight for the second processing block; for samples with thicker subcutaneous fat, the system automatically increases the backtracking weight for processing blocks containing spatial gradient features. The acquisition of slow variables related to the physiological background is completed by a separate low-speed sampling channel, requiring no additional hardware.

[0208] In the continuous monitoring implementation, the attention residual mechanism can be further extended to the temporal dimension. That is, each processing block representation of each acquisition frame is considered a temporal layer, and a temporal attention residual mechanism is used to selectively backtrack between adjacent frames. The concentration prediction for each new frame can be backtracked to the most relevant spectral state in historical frames based on learned content weights, rather than relying on a fixed window uniform average. The intra-frame processing block attention residual and the inter-frame temporal attention residual together constitute a dual attention residual architecture to fully utilize the temporal data advantages in wearable continuous monitoring scenarios.

[0209] In large-scale user deployment scenarios, a meta-learning framework can be used to train and quickly adapt initialization parameters using multi-user data, enabling devices to complete individualized calibration for new users with only a small number of labeled samples. Furthermore, self-supervised pre-training can be implemented on a large number of unlabeled spectral frames: the network is reconstructed from the remaining dimensions after partially masking gradient features as a pre-training task. This task naturally forces the network to learn the physical coupling relationships between dimensions, encoding the homology constraints of physical co-sampling into the network representation space.

[0210] In another implementation, Gaussian process regression can be used as the evaluation model. This implementation is more suitable for the initial deployment phase or scenarios with limited sample size.

[0211] Gaussian process regression requires no large amount of data and has a built-in probabilistic posterior, directly outputting estimates and confidence intervals. Its kernel function can directly encode physical priors, and its tensor product kernel naturally matches the four-dimensional tensor structure of the system. Within the Gaussian process regression framework, the use of gradient features differs fundamentally from that in deep neural networks. Gradient observations are not directly concatenated as input features, but rather incorporated as derivative observations into the joint likelihood of Gaussian process regression.

[0212] Since the derivative of a Gaussian process is still a Gaussian process, gradient observations in each dimension can directly constrain the curvature of the posterior distribution, which is equivalent to replacing some labeled samples with physical gradient information, significantly improving sample efficiency in small sample scenarios. Of course, it is understandable that the mathematical consistency of this mechanism is based on the premise that the gradient has true partial derivative semantics, that is, it requires physical co-sampling.

[0213] In another implementation, a Physical Information Neural Network (PIN) can be used as the evaluation model. The photon diffusion equation is embedded into the network loss function as a hard constraint, and the spatial and temporal gradients correspond to the spatial and temporal partial derivatives of the diffusion equation, respectively, and are directly used as constraint terms in the partial differential equation. The PIN architecture achieves separation between the physical and data layers. The physical layer is described by the analytical structure of the diffusion equation without requiring labeled data, while the data layer only learns the residuals, reducing the required number of labeled samples by one to two orders of magnitude. Cross-dimensional gradients have the strongest physical constraint significance within the PIN framework, and their accurate expression presupposes physical co-sampling.

[0214] Furthermore, sparse Bayesian learning can be used as an evaluation model. Its sparsity mechanism assigns independent relevance weights to each input feature, while the weights of irrelevant features automatically shrink to zero. The sparse solution can directly diagnose which physical dimension gradients truly carry concentration information, providing a data-driven basis for system hardware design optimization. Alternatively, partial least squares regression can be used to replace the traditional single-dimensional spectral input with an augmented joint input vector. By comparing it with traditional partial least squares regression, the Fisher information gain brought by physical co-sampling data can be directly quantified, serving as a benchmark evaluation method to verify the technical effectiveness of this invention.

[0215] Because this invention can be configured with a variety of different evaluation models, an adaptive algorithm switching framework is further proposed to combine the strengths of each available model.

[0216] Specifically, in an exemplary embodiment, the current cumulative number of samples participating in model training can be obtained; based on the number of samples, one of a plurality of preset machine learning models can be selected as the current model to be trained and evaluated, or based on the number of samples, the contribution ratio of each of the plurality of machine learning models in the final concentration estimate can be determined.

[0217] For example, in the initial deployment phase (number of labeled individuals < 100), a physical constraint prior-driven evaluation model (GPR or PINN) is adopted to compensate for insufficient data by utilizing physical priors and fully leveraging the derivative constraints of gradient observations or PDE loss constraints. As the number of samples accumulates (100–10,000 labeled individuals), the model transitions to a meta-learning DNN framework. In the data-rich phase (> 10,000 labeled individuals), the AttnRes temporal architecture and self-supervised pre-training mechanism are enabled to achieve joint optimization at the three theoretical levels. All phases share the same... The normalized joint input interface makes algorithm switching transparent to the hardware layer.

[0218] It should also be noted that the homology verification in S203 is actually a verification of sample quality. The same method can be used to determine the quality of the samples used in training during model training evaluation. Similar indicators that can serve as quality checks include, for example, the minimum length of the third time window and the correlation coefficient of the difference frames to background noise.

[0219] In one embodiment, each training sample may be accompanied by one or more of the following metadata tags before being written into the training set: a homology tag recording whether the sample passes the temporal homology determination and amplitude homology determination; an error upper bound tag recording the ratio of the actual deviation of each dimension to the corresponding threshold; a stability window tag recording the duration of the third time window actually used in this acquisition frame to distinguish the sample quality under different temporal conditions during training; a reference optical path normalization tag recording the relative value of the reference optical path power fluctuation in this frame to indicate whether the power normalization accuracy meets the standard; and a noise correlation coefficient tag recording the correlation coefficient of the difference frame to the background to indicate the quality of the noise covariance structure.

[0220] Furthermore, based on the aforementioned metadata tags, the system executes the following data entry rules: if the homology determination fails, the sample is excluded from the training set, regardless of whether its concentration label value is reliable; if the reference optical path power fluctuation exceeds a preset threshold within the stable window, the sample retains only the original spectral data after power normalization, does not participate in the generation of gradient features, and enters the training set as a sample specifically for original observations; if the background correlation coefficient is lower than a preset threshold, the differential features of the sample are downgraded and labeled, and are given lower weights or removed from the high-quality training subset during training.

[0221] In an exemplary embodiment, the sample concentration detection method of the present invention may further include: obtaining physiological background parameters of the sample, wherein the physiological background parameters include at least one of water peak intensity, temperature-sensitive Raman peak shift, collagen characteristic peak intensity, and fat characteristic peak intensity; obtaining an initial concentration estimate based on the multidimensional homology spectral tensor, and calibrating the initial concentration estimate based on the physiological background parameters to obtain the concentration of the component to be measured.

[0222] It should be noted that wavelength switching, spatial bias channel switching, time gating parameter switching, and modulation coding state switching must complete a complete timing process of preset locking, wavelength transition, stabilization waiting, and effective integration within a single sampling period. These dimensions change at high rates and must complete state switching and locking in an extremely short time. Therefore, they must be incorporated into a hard-lock period and stabilization window mechanism strictly driven by the global modulation clock to ensure that the physical states of each dimension are determined simultaneously at the moment of wavelength switching.

[0223] On the other hand, the aforementioned physiological background parameters reflect physiological states such as skin moisture content, tissue temperature, connective tissue density, and subcutaneous fat thickness. These physiological parameters remain almost unchanged over a timescale of several seconds to several minutes during a single test, and their variation period is much longer than the single sampling period of the main link.

[0224] Therefore, in one embodiment of the present invention, an independent low-speed sampling channel is used to detect the above-mentioned physiological background parameters.

[0225] Please refer to Figure 3 As shown, based on the same inventive concept as the aforementioned sample concentration detection method, one embodiment of the present invention provides a sample concentration detection system 300, including: an irradiation module 301, a configuration module 302, an analysis module 303, and an output module 304.

[0226] Specifically, the irradiation module 301 is used to irradiate a sample containing the component to be tested with a laser based on a switchable excitation wavelength, and the laser wavelength switching time is determined as the clock reference point.

[0227] The configuration module 302 is used to configure and lock each state component as a target value in a first time window before the clock reference point; and to perform spectral data acquisition based on the locked state components in a second time window after the clock reference point; the state components include spatial offset channel state, time gating parameters, and modulation coding state.

[0228] The analysis module 303 is used to construct a multidimensional homologous spectral tensor based on the spectral data and the state components corresponding to the spectral data; wherein the multidimensional homologous spectral tensor contains at least one pair of spectral data samples that vary along a single target dimension and are locked in the other dimensions, and the target dimension is the laser wavelength or any state component;

[0229] The output module 304 is used to input the multidimensional homology spectral tensor into the evaluation model to obtain the concentration of the component to be measured in the sample output by the evaluation model.

[0230] Please refer to Figure 4 As shown, embodiments of the present invention also provide an electronic device 400, which includes at least one processor 401, a memory 402 (e.g., non-volatile memory), a main memory 403, and a communication interface 404, and the at least one processor 401, memory 402, main memory 403, and communication interface 404 are connected together via an internal bus 405. The at least one processor 401 is used to invoke at least one program instruction stored or encoded in the memory 402, so that the at least one processor 401 performs various operations and functions of the sample concentration detection method described in the various embodiments of this specification.

[0231] In the embodiments of this specification, electronic device 400 may include, but is not limited to: personal computer, server computer, workstation, desktop computer, laptop computer, notebook computer, mobile electronic device, smartphone, tablet computer, cellular phone, personal digital assistant (PDA), handheld device, messaging device, wearable electronic device, consumer electronic device, etc.

[0232] This invention also provides a computer-readable medium carrying computer-executable instructions. When executed by a processor, these instructions can be used to implement various operations and functions of the sample concentration detection methods described in the various embodiments of this specification.

[0233] The computer-readable medium in this invention can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0234] In this invention, the computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. This propagated data signal may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wireline, optical fiber, RF, etc., or any suitable combination thereof.

[0235] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0236] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus, systems, and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0237] The foregoing description of specific exemplary embodiments of the invention is for illustrative and explanatory purposes. These descriptions are not intended to limit the invention to the precise forms disclosed, and it will be apparent that many changes and variations can be made in accordance with the foregoing teachings. The exemplary embodiments were chosen and described in order to explain the specific principles of the invention and its practical application, thereby enabling those skilled in the art to implement and utilize various different exemplary embodiments of the invention, as well as various different choices and variations. The scope of the invention is intended to be defined by the claims and their equivalents.

[0238] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0239] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for detecting sample concentration, characterized in that, include: The sample containing the analyte is irradiated with a laser with switchable excitation wavelength, and the laser wavelength switching time is determined as the clock reference point. In the first time window before the clock reference point, configure and lock each state component as the target value; In the second time window following the clock reference point, spectral data is acquired based on the locked state components; The state components include spatial bias channel state, time gating parameters, and modulation coding state, thereby enabling co-sampling of frequency-shift excited Raman differential spectroscopy (SERDS), spatial bias Raman spectroscopy (SORS), and time gating at the same physical moment; wherein, the first time window and the second time window between adjacent clock reference points do not overlap on the time axis. Based on the spectral data and the corresponding state components, a multidimensional homologous spectral tensor is constructed; wherein, the multidimensional homologous spectral tensor contains at least one pair of spectral data samples that vary along a single target dimension and are locked in the other dimensions, and the target dimension is the laser wavelength or any state component; The multidimensional homology spectral tensor is input into the evaluation model to obtain the concentration of the analyte in the sample output by the evaluation model.

2. The sample concentration detection method according to claim 1, characterized in that, Configure the end point of the first time window as the clock reference point; Configure the starting point of the second time window as the clock reference point; configure the ending point of the second time window as the starting point of the next first time window.

3. The sample concentration detection method according to claim 1, characterized in that, Determining a second time window located after the clock reference point includes: A third time window is determined starting from the clock reference point. Within the third time window, each state component is locked. The duration of the third time window is greater than or equal to the duration required for the parasitic oscillation to decay to the first threshold. The second time window is determined after the end point of the third time window and before the start point of the next first time window.

4. The sample concentration detection method according to claim 3, characterized in that, Determining the second time window located after the clock reference point includes: The starting point of the second time window is determined as the ending point of the third time window; the ending point of the second time window is the starting point of the next first time window; wherein the first time window, the second time window, and the third time window do not overlap with each other on the time axis.

5. The sample concentration detection method according to claim 1, characterized in that, Based on the spectral data and the corresponding state components, a multidimensional homologous spectral tensor is constructed, including: The spectral data is preprocessed, including determining whether the state components corresponding to the spectral data are homologous samples, and / or normalizing the spectral data based on the power of the laser. The preprocessed spectral data and the corresponding state components of the spectral data are written into the corresponding positions of the multidimensional homogeneous spectral tensor.

6. The sample concentration detection method according to claim 5, characterized in that, Determining whether the state components corresponding to the spectral data are homologous samples includes: Obtain the lock start time for each state component, as well as the calibration value of the lock time required; Calculate the time difference between the start of locking for each state component and the next clock reference point; If the time difference is greater than or equal to the calibration value, then the state component corresponding to the spectral data is a homologous sample; If the time difference is less than the calibration value, then the state component corresponding to the spectral data is a non-homologous sample.

7. The sample concentration detection method according to claim 5, characterized in that, Determining whether the state components corresponding to the spectral data are homologous samples includes: Calculate the deviation of each state component when it is sampled in the second time window relative to when it is locked in the first time window; If the deviation is less than the preset upper limit of error, then the state component corresponding to the spectral data is a homologous sample. If the deviation is greater than or equal to the preset upper limit of error, then the state component corresponding to the spectral data is a non-homogeneous sample.

8. The sample concentration detection method according to claim 5, characterized in that, Determining whether the state components corresponding to the spectral data are homologous samples includes: Acquire spectral data sample pairs collected under lasers of different wavelengths with the same state components; Separate the background region of the spectral data and count the photon count sequence in the background region of the two samples; Calculate the Pearson correlation coefficient value of the photon counting sequence; If the Pearson correlation coefficient value is greater than or equal to the coefficient threshold, then the state component corresponding to the spectral data is a homologous sample; If the Pearson correlation coefficient is less than the coefficient threshold, then the state component corresponding to the spectral data is a non-homologous sample.

9. The sample concentration detection method according to claim 5, characterized in that, Determining whether the state components corresponding to the spectral data are homologous samples includes: The laser beam is split into a first beam that illuminates the sample and a second beam that is guided into a reference optical path; Multiple sets of power values ​​of the second beam are acquired within the second time window for acquiring the current sample spectral data; Calculate the relative fluctuation value of the power value, where the relative fluctuation value is the quotient of the standard deviation and the average value of the power sampled values ​​within the second time window; If the relative fluctuation value of the power value is less than the power fluctuation threshold, then the state component corresponding to the spectral data is a homologous sample; If the relative fluctuation value of the power value is greater than or equal to the power fluctuation threshold, then the state component corresponding to the spectral data is a non-homogeneous sample.

10. The sample concentration detection method according to claim 5, characterized in that, If the sample is a non-homologous sample, then discard the non-homologous sample, and / or re-execute the spectral data acquisition under the corresponding state component, and / or adjust the start time of the second time window.

11. The sample concentration detection method according to claim 5, characterized in that, The spectral data is normalized based on the laser power, including: The laser beam is split into a third beam to illuminate the sample and a fourth beam to guide the reference optical path; Determine the average power of the fourth beam within the second time window for collecting the current sample spectral data; The spectral data of the current sample is normalized based on the power mean to obtain the spectral data under unit excitation power.

12. The sample concentration detection method according to claim 1, characterized in that, The multidimensional homology spectral tensor is input into the evaluation model to obtain the concentration of the analyte in the sample output by the evaluation model, including: For the spectral data in the multidimensional homogeneous spectral tensor, calculate the gradient features along at least one dimension; Based on the spectral data and the gradient features, a joint input vector is generated; Based on the joint input vector, the concentration of the analyte in the sample is determined and output.

13. The sample concentration detection method according to claim 12, characterized in that, The step of generating a joint input vector based on the spectral data and the gradient features includes: Obtain the preset fluctuation calibration values ​​of the spectral data and the gradient features, and perform standardization processing on the spectral data and the gradient features based on the fluctuation calibration values; The normalized spectral data and the gradient features are spliced ​​together to generate the joint input vector.

14. The sample concentration detection method according to claim 12, characterized in that, The gradient feature along at least one dimension includes: a first gradient feature and a second gradient feature; The first gradient feature is determined by the ratio of the difference between two frames of spectral data that differ in the target dimension but are the same in the other dimensions in the multidimensional homogeneous spectral tensor, to the dimensional interval of the target dimension; the dimensional interval refers to the difference between two different values ​​in the same dimension; and / or The second gradient feature is determined by the ratio of the difference between the two first gradient features to the dimensional interval of the other dimension, wherein the two first gradient features are different in the other dimension but are the same in the other dimensions.

15. The sample concentration detection method according to claim 1, characterized in that, The method further includes: Construct an evaluation model and set the relevant parameters of the evaluation model; Obtain a training sample set, which includes the multidimensional homology spectral tensor and the concentration of the component to be measured in the corresponding sample; The evaluation model is trained based on the training sample set, and the parameters of the evaluation model are corrected until the deviation value of the concentration of the component to be tested in the sample obtained by the evaluation model output is less than a preset threshold.

16. The sample concentration detection method according to claim 1, characterized in that, The evaluation model includes one or more machine learning models such as deep neural networks, Gaussian process regression, or physical information neural networks.

17. The sample concentration detection method according to claim 16, characterized in that, The method further includes: Get the current cumulative number of samples participating in model training; Based on the number of samples, select one from a set of preset machine learning models as the current model to be trained and evaluated, or Based on the number of samples, the contribution ratio of each of the multiple machine learning models in the final concentration estimate is determined.

18. The sample concentration detection method according to claim 1, characterized in that, The method further includes: Obtain the physiological background parameters of the sample, which include at least one of water peak intensity, temperature-sensitive Raman peak shift, collagen characteristic peak intensity, and fat characteristic peak intensity; An initial concentration estimate is obtained based on the multidimensional homology spectral tensor, and the initial concentration estimate is calibrated based on the physiological background parameter to obtain the concentration of the analyte.

19. The sample concentration detection method according to claim 14, characterized in that, The laser wavelength serves a dual purpose: it acts as a time-domain reference and a frequency-domain differential. The switching period of the laser wavelength serves as a reference dimension for the global clock, driving each state component to enter the locking sequence within the first time window. The value of the laser wavelength serves as one of the target dimensions, eliminating fluorescence background by differentiating spectral data collected at adjacent wavelengths, thus achieving frequency-shifted excitation Raman differential.

20. A sample concentration detection system, characterized in that, include: The irradiation module is used to irradiate a sample containing the analyte with a laser based on a switchable excitation wavelength, and to determine the laser wavelength switching time as a clock reference point. The configuration module is used to configure and lock each state component to a target value in the first time window before the clock reference point; In the second time window following the clock reference point, spectral data is acquired based on the locked state components; The state components include spatial bias channel state, time gating parameters, and modulation coding state, thereby enabling co-sampling of frequency-shift excited Raman differential spectroscopy (SERDS), spatial bias Raman spectroscopy (SORS), and time gating at the same physical moment; wherein, the first time window and the second time window between adjacent clock reference points do not overlap on the time axis. The analysis module is used to construct a multidimensional homologous spectral tensor based on the spectral data and the state components corresponding to the spectral data; wherein the multidimensional homologous spectral tensor contains at least one pair of spectral data samples that vary along a single target dimension and are locked in the other dimensions, and the target dimension is the laser wavelength or any state component; The output module is used to input the multidimensional homology spectral tensor into the evaluation model to obtain the concentration of the analyte in the sample output by the evaluation model.

21. An electronic device, characterized in that, include: At least one processor; as well as A memory that stores instructions that, when executed by the at least one processor, cause the at least one processor to perform the sample concentration detection method as described in any one of claims 1 to 19.

22. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the sample concentration detection method according to any one of claims 1-19.

Images