Intelligent detection method for antibiotic residue spectrum based on time sequence feature decoupling

Abstract

The present application relates to the technical field of spectral analysis, in particular to an antibiotic residue spectral intelligent detection method based on time sequence characteristic decoupling, which uses a light source with sinusoidal modulation initial light intensity to irradiate a sample to make it produce nonlinear photobleaching depletion, collects harmonic comb data to construct a time sequence spectral matrix, synchronously demodulates the matrix to generate a harmonic dispersion tensor, extracts second-order and above high-order slices to eliminate matrix background and scattering noise, calculates concentration-independent harmonic ratio figures by eliminating concentration through the amplitude ratio of different order high-order slices, matches kinetic fingerprints represented by photochemical rate constants according to the ratio figures, and separates amplitude channel data; based on the time evolution weight corresponding to the kinetic fingerprint, the amplitude channel data is inversely mapped, and the concentration scalar result is output; high specificity quantitative detection of antibiotic residues under strong background interference is realized.

Description

Technical Field

[0001] This invention relates to the field of spectral analysis technology, specifically to an intelligent spectral detection method for antibiotic residues based on temporal feature decoupling. Background Technology

[0002] Current techniques for the spectral detection of antibiotic residues typically involve irradiating the sample with constant light intensity, collecting the absorption or fluorescence spectra of the target antibiotic under static conditions, and then comparing these spectra with standard spectra or using chemometric or machine learning models to perform regression analysis on the spectral intensity to deduce the concentration of the target antibiotic. This method based on static spectral intensity can achieve a certain degree of accuracy in qualitative and quantitative detection in scenarios where the sample matrix composition is relatively simple and the target characteristics are prominent, effectively improving the overall efficiency of residue screening in practical applications.

[0003] However, existing technologies have limitations in addressing trace antibiotic residues in complex fluid matrices. The static spectral characteristics of the target antibiotic highly overlap with the background absorption of the complex matrix and the broadband scattering noise caused by suspended particles in the spectral space. Furthermore, existing static spectral measurements rely on the existing spectral intensity dimension, resulting in a strong intertwining and coupling between the specific identification information characterizing the intrinsic properties of the substance and the absolute concentration information characterizing its content in the amplitude dimension. Under these circumstances, the system cannot effectively separate the inherent specific characteristics of the target from the fluctuating background interference. This makes the weak spectral characteristics of trace targets easily submerged by the overlapping strong matrix background and scattering noise, leading to misjudgments in qualitative identification and deviations in quantitative inversion, making it difficult to guarantee the specificity and accuracy of trace residue detection. Summary of the Invention

[0004] To address the problems in related technologies, this invention provides an intelligent detection method for antibiotic residue spectra based on temporal feature decoupling, thereby overcoming the aforementioned technical problems existing in existing related technologies.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a smart detection method for antibiotic residue spectra based on time-series feature decoupling, comprising the following steps:

[0006] The sample to be tested is irradiated with an excitation source whose initial light intensity is sinusoidally modulated at a preset frequency. The target antibiotic with irreversible photolysis characteristics at the wavelength of the excitation source is depleted by nonlinear photobleaching under the cumulative light dose. The spontaneously generated harmonic comb data is continuously collected to construct the modulation excitation time-series spectral matrix.

[0007] The modulation excitation time-series spectral matrix is ​​synchronously demodulated according to each harmonic of the preset frequency to generate a harmonic dispersion tensor containing multiple harmonic orders.

[0008] Filter out the fundamental frequency tensor slices of the first order in the harmonic dispersion tensor, and retain the higher-order harmonic tensor slices of the second order or higher, to obtain higher-order harmonic tensor slices that have eliminated the matrix background with linear response.

[0009] Using the amplitude ratios between the higher-order harmonic tensor slices of different orders, a concentration-independent harmonic ratio diagram, which is determined solely by the photochemical rate constant of the target antibiotic after eliminating absolute concentration, is calculated.

[0010] The concentration-independent harmonic ratio plot is matched with the kinetic fingerprint characterized by the photochemical rate constant, and the amplitude channel data carrying concentration information is separated from the higher-order harmonic tensor slice based on the kinetic fingerprint.

[0011] Based on the time evolution weights corresponding to the kinetic fingerprint, the amplitude channel data is inverted and mapped to output the scalar concentration result of the target antibiotic.

[0012] By employing the above technical solution, the present invention provides an intelligent detection method for antibiotic residue spectra based on temporal feature decoupling, which has at least the following beneficial effects:

[0013] 1. This invention introduces an active sinusoidal light intensity fabrication on the time axis as a perturbation source, forcing the target antibiotic to undergo irreversible nonlinear photobleaching depletion under cumulative light dose, spontaneously generating a harmonic comb signal distinct from the linear matrix. After removing the linear matrix background and broadband scattering noise in the frequency domain, the absolute concentration is eliminated using the higher-order harmonic amplitude ratio, retaining the kinetic fingerprint determined by the photochemical rate constant. The qualitative identification characterizing the inherent properties of the substance is decoupled from the quantitative concentration inversion characterizing the content in a multidimensional way, thereby achieving highly specific quantitative detection of antibiotic residues under strong background interference.

[0014] 2. This invention actively transforms the photobleaching effect, which is usually regarded as interference or signal loss, into a signal source carrying identity information: the linearly responsive matrix background is pinned at the first-order fundamental frequency, and the broadband scattering noise lacking coherence cannot form narrowband stacking at the higher harmonic frequencies. Therefore, it is only necessary to cut off the first-order fundamental frequency and retain the second-order and above higher harmonic slices to remove the matrix background and scattering noise in the physical frequency domain, thereby improving the ability to suppress strong matrix background.

[0015] 3. This invention performs element-wise division on the amplitudes of higher harmonics of different orders, so that the common multipliers composed of absolute concentration and static spectral characteristics are canceled in the fraction, resulting in a concentration-independent harmonic ratio diagram characterized only by the modified Bessel function ratio; thus, the discrimination of target identity is no longer affected by concentration fluctuations and optical path intensity fluctuations, thereby improving the specificity and robustness of identification under trace conditions.

[0016] 4. This invention uses the modified Bessel function ratio distribution as a strong physical prior to guide the attention mechanism network to judge each spatial unit, shield residual interference and confirm the dynamic fingerprint, and then inverts the amplitude channel based on the time evolution weights parsed from the fingerprint to obtain the correlation quantity that is linearly corresponding to the absolute concentration and regresses it into a concentration scalar; it combines physically interpretable prior constraints with data-driven separation inversion, taking into account both anti-interference ability and quantitative accuracy. Attached Figure Description

[0017] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0018] Figure 1 The flowchart shows the intelligent detection method for antibiotic residue spectra based on time-series feature decoupling provided by the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Exemplary method:

[0021] In the detection of antibiotic residues in complex matrices such as milk and aquatic products, the target antibiotics are often at trace levels. Their static absorption / fluorescence characteristics highly overlap with the matrix background and broadband scattering in the spectral spatial domain, and the spectral intensity simultaneously contains both identity and concentration information. Existing technologies mainly rely on comparison or regression of static spectral intensity under constant illumination, which can only process information in the spectral spatial domain. This approach cannot separate submerged targets from strong backgrounds, nor can it decouple identity and concentration, leading to misjudgment and quantitative bias under trace and highly interfering conditions.

[0022] Faced with the above problems, the core concept of this application lies in "trading time for space and transforming linearity into nonlinearity": since the target antibiotic and the matrix background and scattering noise are indistinguishable in the spectral spatial domain, an active sinusoidal light intensity fabrication is introduced as a perturbation source on the time axis. The unique and irreversible nonlinear photobleaching depletion effect of the target antibiotic is used to force it to spontaneously generate high-order harmonic signals that are different from those of the linear matrix. Then, the first order is removed in the frequency domain and the higher order is retained to eliminate the background and noise. The common multiplier of absolute concentration is eliminated by the quotient of the higher-order harmonics, and the kinetic fingerprint determined by the photochemical rate is extracted, thus achieving accurate separation of identity confirmation and concentration information.

[0023] To address this, this application proposes a smart spectral detection method for antibiotic residues based on temporal feature decoupling. For example... Figure 1 As shown, the method includes the following steps:

[0024] The following is a complete calculation process demonstration using the detection of enrofloxacin (a quinolone antibiotic with well-defined photodegradation properties) in milk matrix as an example.

[0025] S1. Irradiate the sample to be tested with an excitation source that sinusoidally modulates the initial light intensity at a preset frequency, so that the target antibiotic with irreversible photolysis characteristics at the wavelength of the excitation source will undergo nonlinear photobleaching depletion under the cumulative light dose, and continuously collect spontaneously generated harmonic comb data to construct a modulation excitation time-series spectral matrix.

[0026] It should be noted that this step encodes the intrinsic kinetic characteristics of the target antibiotic into the time-series spectral data through active modulation, providing source data carrying nonlinear features for subsequent frequency domain decoupling.

[0027] Specifically, the output light intensity of the excitation light source is no longer constant, but is sinusoidally modulated according to a preset frequency. The formula for calculating its instantaneous light intensity is as follows:

[0028] ;

[0029] in, for The instantaneous intensity of time, This is the initial light intensity (i.e., the light intensity at the modulation center). The modulation depth (ranging from 0 to 1, representing the sinusoidal fluctuation amplitude of light intensity relative to the initial light intensity) is used. To modulate the angular frequency, This refers to the preset frequency. The rate at which the excitation light intensity sinusoidal fluctuations are characterized needs to be set in a way that considers two points: first, it should be far from the frequency bands of ambient stray light and power supply noise to facilitate subsequent phase-locked extraction; second, it should be higher than the characteristic rate of photobleaching depletion to ensure that the different spectral lines of the harmonic comb are separated from each other in the frequency domain. In this embodiment, the following is taken: =1000Hz, that is .

[0030] Furthermore, the target antibiotic undergoes irreversible photochemical decomposition under continuous light irradiation, resulting in a decrease in the number of unbleached molecules. The cumulative light dose decreases according to a first-order kinetic law, i.e. ,in The photochemical rate constant of the target antibiotic (a comprehensive characterization of its absorption cross-section and photobleaching quantum yield, serving as an intrinsic identity parameter distinguishing different antibiotics). For Integrating over time yields:

[0031] ;

[0032] in, This represents the initial number of unbleached molecules (proportional to the absolute concentration of the target antibiotic). The oscillation modulation depth (which, as the independent variable of the modified Bessel function, determines the relative weights of each order of the harmonic comb; its physical meaning and setting method are detailed in step S12) is shown. It can be seen that the time evolution of the number of molecules contains... This nonlinear oscillation factor causes the characteristic spectrum of the target antibiotic to no longer linearly follow the modulated light source, but instead to be distorted into a harmonic comb containing higher-order components. Using the generating function expansion of the modified Bessel function, this nonlinear oscillation factor can be expanded into the sum of all harmonics:

[0033] ;

[0034] in, for The first-order modified Bessel function of the order, Therefore, the first The amplitude of the first harmonic is proportional to That is, the target antibiotic spontaneously generates a frequency with a preset fundamental frequency and amplitude values ​​of various orders under modulation excitation. Weighted harmonic comb. In contrast, the matrix background, which only responds linearly to light intensity, only responds at the first order (fundamental frequency) and does not generate higher-order harmonics, thus naturally distinguishing it from the target antibiotic in the frequency domain. This step transforms a target that is difficult to resolve in the spectral spatial domain into time-series data with a separable nonlinear harmonic structure in the frequency domain.

[0035] It should be noted that the attenuation envelope in the above formula... The bleaching depletion of a single fluid element within the light spot over time is characterized by the fact that a single element may be bleached in less than one modulation cycle. However, because the sample under test is in a continuous flow state (see S11), fresh, unbleached molecules are constantly replenished within the light spot region, causing the molecular ensemble within the light spot to establish a steady-state spectral response with a modulation period driven by sinusoidal modulation. Therefore, the acquired and demodulated harmonic comb is the steady-state quantity of this flow ensemble, which can be phase-locked averaged over multiple modulation cycles. Its steady-state nature originates from the flow replenishment rather than the survival period of a single molecule. The strength of the higher-order components of the harmonic comb depends on the nonlinearity of the ensemble's bleaching response, and is characterized by the following photobleaching modulation factor.

[0036] It should be further explained that the detector is at wavelength , Real-time measured spectral signal (This can be absorbance or fluorescence intensity) Under trace-level approximation, it is proportional to the product of the number of unbleached molecules at that moment and the static spectral characteristics, i.e. ,in This refers to the static spectral characteristics (such as absorption coefficient) of the target antibiotic. Therefore... The harmonic structure is directly inherited from Harmonic comb.

[0037] The construction of the modulation excitation time-series spectral matrix includes:

[0038] S11. Use the fluid sample in a natural flow state as the sample to be tested;

[0039] Specifically, naturally flowing fluid samples (such as pretreated milk filtrate or water samples) are used as the test samples, and the samples are passed through the excitation spot area at a certain flow rate. The purpose of using flowing sample introduction instead of a static cell is to ensure that each sample segment stays in the spot for a limited time, thereby allowing precise control of the cumulative light dose through the flow rate; at the same time, flowing can avoid the interference of local accumulation of bleaching products on subsequent samples in static measurements.

[0040] S12. Control the dwell time of the fluid sample as it flows through the light spot area of ​​the excitation source and the modulation depth of the light intensity of the excitation source, so that the photobleaching modulation factor determined by the dwell time, the modulation depth, the initial light intensity and the preset frequency reaches a preset threshold, so as to ensure that the target antibiotic produces nonlinear photobleaching depletion and that the harmonic amplitude of the order greater than or equal to the second order is higher than the preset resolvable lower limit.

[0041] It is important to emphasize that the generation and resolvability of harmonic combs are determined by two independent factors working together: one is the oscillation modulation depth introduced by sinusoidal modulation on the cumulative optical dose. It is determined by the initial light intensity Modulation depth With preset frequency The decision was made jointly, and the Bézier weights were adjusted accordingly. The relative intensities of each order of the harmonic comb are determined by two factors: firstly, the relative intensity of each order; and secondly, the cumulative optical dose due to the residence time of the fluid element within the light spot. It is determined by the initial light intensity Duration of stay Together, they determine the sufficiency of ensemble bleaching depletion, i.e., whether the harmonic comb can be effectively generated.

[0042] Taking into account the above two factors, the photobleaching modulation factor is defined as a quantity characterizing the relative generation intensity of the second harmonic in the steady-state harmonic comb: ,in For the cumulative light dose Determined harmonic generation factor, For the depth of oscillation modulation Determined harmonic shape factor. Photobleaching modulation factor. It is determined by four factors: dwell time, modulation depth, initial light intensity, and preset frequency.

[0043] Among them, the preset threshold The minimum photobleaching modulation factor required to make the second harmonic amplitude exactly reach the preset resolvable lower limit is characterized by: setting the amplitude of the second harmonic at a given detector noise level, and obtaining it by inverse solution by making the amplitude of the second harmonic equal to the resolvable lower limit (e.g., taking a number of times the noise standard deviation). When the system determines that higher-order harmonics are indistinguishable, it will prioritize reducing the flow velocity to prolong the residence time. (promote (and supplemented by increasing the modulation depth or increasing the initial light intensity), (promote and Improve through means such as until If the bleaching is too fast and causes harmonic structure distortion, then the adjustment is reversed, thereby achieving closed-loop control of the cumulative light dose.

[0044] For example, in this embodiment, the modulation depth is taken. =0.8, oscillation modulation depth =0.5 (correspondingly) ), stay time Then the cumulative light dose Harmonic generation factor At this point, the modified Bessel functions of each order take the values... , , , Therefore, photobleaching modulation factor Set a preset threshold. ,but It can generate and distinguish second-order and higher harmonics.

[0045] Therefore, in this step, "ensuring the target antibiotic produces nonlinear photobleaching depletion" is achieved by accumulating light dose to a level that fully bleachs the ensemble, while "ensuring that the amplitude of harmonics of order greater than or equal to second order is higher than the preset resolvable lower limit" is achieved by the photobleaching modulation factor. Reaching the preset threshold To achieve this; the two are not isolated statements of effect, but are respectively achieved by and Functional conditions defined by two calculable and controllable quantities.

[0046] It should be noted that this step is used to set the data collection conditions. The values ​​are derived from the pre-characterized values ​​of the target antibiotic (e.g., calibrated from its pure standards under the same modulation conditions); because the target antibiotic is a specific antibiotic, its... It can be known in advance, so it is necessary to conduct research before collection. Has it been achieved? The determination uses the pre-characteristic The estimation does not depend on the inversion results of the test sample.

[0047] S13. At a preset sampling rate greater than or equal to the preset highest interest harmonic order corresponding to the Nyquist frequency, continuously collect the harmonic comb data spontaneously generated under the nonlinear photobleaching depletion effect to construct a two-dimensional modulation excitation time-series spectral matrix.

[0048] Specifically, to ensure that the highest-order harmonic does not experience spectral aliasing, the preset sampling rate must satisfy the Nyquist sampling theorem, i.e., it must not be lower than the highest-order harmonic. Twice the corresponding frequency, The preset highest harmonic order represents the highest harmonic order involved in subsequent decoupling. Its setting needs to be a compromise between "higher-order harmonics are more sensitive to the rate constant" and "higher-order harmonics have weaker amplitudes", and is usually set to 3 to 5.

[0049] This embodiment takes The sampling rate should be no less than In this embodiment, =20000Hz. The collected data, arranged along two dimensions—preset wavelength channels and timestamps—constitute a two-dimensional modulation-excitation time-series spectral matrix with time on the horizontal axis and wavelength channels on the vertical axis. It contains a time-varying decay spectral flow modulated by nonlinear photobleaching. As an alternative, fluid injection can be replaced by microfluidic chip quantitative injection, and the excitation source can be an LED or a laser. In addition to a sine wave, the modulation waveform can also be a calibrated quasi-sine wave.

[0050] S2. The modulation excitation time-series spectral matrix is ​​synchronously demodulated according to each harmonic of the preset frequency to generate a harmonic dispersion tensor containing multiple harmonic orders.

[0051] It should be noted that this step transforms the time-domain modulation-excitation time-series spectral matrix into a frequency-domain harmonic dispersion tensor. The purpose of this transformation is to separate the harmonics aliased in the time series, resulting in a three-dimensional tensor. This lays the data structure foundation for subsequent order-based physical filtering and quotient calculation. This step is implemented using phase-locked demodulation technology, and the specific implementation of the subordinate steps is given below.

[0052] The process of synchronously demodulating the modulation excitation time-series spectral matrix according to each harmonic of the preset frequency to generate a harmonic dispersion tensor containing multiple harmonic orders includes:

[0053] S21. Obtain the known sinusoidal modulation waveform data corresponding to the preset frequency as a phase-locked reference signal;

[0054] Specifically, since the modulation waveform is actively generated by the system and its frequency and phase are known, the known sine waveform and its harmonics can be directly used as the phase-locked reference signal without the need for additional phase estimation.

[0055] S22. Using the phase-locked reference signal, perform phase-locked demodulation operation on the modulation excitation timing spectrum matrix along the time axis in each preset wavelength channel to extract the amplitude data and phase data corresponding to each harmonic.

[0056] Specifically, for each wavelength channel time series By performing correlation integration with each order of reference components, the in-phase and quadrature components are obtained, and then the first-order reference component is calculated. Amplitude and phase of first harmonics:

[0057] ;

[0058] ;

[0059] in, , The first The in-phase and quadrature components of the first harmonic, i.e., the measured signal. respectively with the first cosine reference component Sine reference component The result obtained by performing relevant integration (taking the time average); and Wavelength First The amplitude and phase of the first harmonic. This represents the average value over an integer number of modulation periods along the time axis. This amplitude is proportional to the value in step S1. .

[0060] S23. From the extracted amplitude data and phase data, subtract the known mixing crosstalk term caused by the modulation illumination component of the excitation light source, obtain the corrected amplitude data and phase data, and construct the harmonic dispersion tensor based on the corrected amplitude data and phase data.

[0061] Specifically, since the modulation illumination of the excitation source itself introduces known mixing crosstalk terms on each harmonic, this step uses the pre-calibrated modulation characteristics of the light source to subtract the crosstalk terms from the amplitude and phase of each order, obtaining a correction result that only reflects the sample response, and organizing it into a harmonic dispersion tensor according to the three dimensions of wavelength, order, and amplitude / phase.

[0062] S3. Filter out the fundamental frequency tensor slice with the first order in the harmonic dispersion tensor, and retain the higher-order harmonic tensor slice with the order greater than or equal to the second order, to obtain the higher-order harmonic tensor slice with the matrix background that has been removed and has a linear response.

[0063] It should be noted that this step removes two types of interference in the physical frequency domain in one step: the matrix background with a linear response and incoherent broadband scattering noise. Since this processing directly acts on the harmonic order axis, it is a frequency domain filtering with clear physical meaning, rather than statistical denoising that relies on empirical parameters, and therefore has good interpretability and robustness.

[0064] Specifically, the matrix background (such as fat and protein in milk) mainly exhibits a linear absorption / scattering response to the modulated light intensity. Its response strictly follows the modulated light source, thus concentrating energy at the first-order fundamental frequency and producing almost no higher-order harmonics. Therefore, by filtering out the entire slice of the first-order fundamental frequency tensor, the linear matrix background can be removed at the source of the data.

[0065] The retention of higher-order harmonic tensor slices with an order greater than or equal to second includes: retaining nonlinear harmonic components that form narrowband coherent stacks at the corresponding frequencies of higher-order harmonics, and removing broadband scattering noise that does not form narrowband coherent stacks at the corresponding frequencies of higher-order harmonics.

[0066] Specifically, broadband scattering noise exhibits energy dispersion in the frequency domain and lacks phase coherence with the modulation frequency, making it unable to form narrowband coherent stacking at discrete frequency points corresponding to second-order and higher harmonics; while the nonlinear harmonics of the target antibiotic coherently concentrate precisely at these discrete frequency points. Therefore, order preservation is crucial. Higher-order harmonic tensor slices This is equivalent to a set of narrowband filters centered on each harmonic frequency, naturally eliminating diffuse broadband scattering noise. After this step, the linear basis of the data is mathematically zeroed out, retaining only the high-order harmonic slices that carry the nonlinear dynamic information of the target antibiotic.

[0067] S4. Using the amplitude ratios between the higher-order harmonic tensor slices of different orders, calculate the concentration-independent harmonic ratio diagram, which is determined only by the photochemical rate constant of the target antibiotic after eliminating the absolute concentration.

[0068] It should be noted that this step removes the concentration information still coupled to the higher harmonic amplitudes and extracts the dynamic fingerprint that characterizes the target.

[0069] Specifically, the amplitude of each harmonic can be written as a product of "absolute concentration × static spectral characteristics × Bessel weight". Absolute concentration and static spectral characteristics are the same common multipliers for different orders. Only the Bessel weight changes with the order. Therefore, the common multipliers can be canceled out by dividing by different orders.

[0070] The concentration-independent harmonic ratio plot, calculated and determined solely by the photochemical rate constant of the target antibiotic after eliminating absolute concentration, includes:

[0071] S41. Extract the target order higher harmonic amplitude and the reference order higher harmonic amplitude from the higher harmonic tensor slice.

[0072] Specifically, from the preserved high-order harmonic tensor slices, a target order (such as the third order) and a reference order (such as the second order) are selected, and their harmonic amplitudes in each wavelength channel are extracted respectively, which are used as the numerator and denominator of the subsequent quotient.

[0073] S42. Perform element-wise division between the target order higher harmonic amplitude and the reference order higher harmonic amplitude to eliminate the common multiplier data composed of absolute concentration and static spectral characteristics, and output a ratio matrix containing only the modified Bessel function ratio determined by the photochemical rate constant, as a concentration-independent harmonic ratio map.

[0074] Specifically, combining step S1 and taking into account the steady-state average of the flow ensemble, the first The amplitude of the first harmonic can be approximately decomposed into: ,in The initial number of molecules is proportional to the absolute concentration. Static spectral characteristics (absorption coefficient) For the cumulative light dose The ensemble bleaching envelope factor is determined and is the same for all harmonics.

[0075] It should be noted that the approximate equality sign is used here because... The value originates from the average residence time distribution of the flow ensemble, which is shared across different orders and is canceled out during quotient calculation, thus not affecting subsequent conclusions. The target order... With reference order The magnitudes of (all of which are of higher order, not less than second order) are calculated by dividing element by element, using the following formula:

[0076] ;

[0077] Among them, proportional to absolute concentration Static spectral characteristics With ensemble bleaching envelope The three factors are canceled out as common multipliers in the fraction, resulting in a ratio. Only the depth of oscillation modulation The function; while , , Under the condition that all are known, Uniquely determined by the photochemical rate constant of the target antibiotic The resulting ratio matrix is ​​a concentration-independent harmonic ratio diagram, determined solely by the target's identity and independent of absolute concentration. This constitutes the kinetic fingerprint of the target antibiotic.

[0078] For example, the aforementioned =0.5, taking the target order as 3 and the reference order as 2 (both are high orders not less than second order, consistent with S31 eliminating the first-order fundamental frequency), then the concentration-independent harmonic ratio is Regardless of whether the absolute concentration of enrofloxacin being tested is high or low, as long as its Unchanged (i.e.) (Unchanged), the above ratio remains constant, thus cleanly separating the identity from the concentration. It should be noted that although the amplitude of higher-order harmonics is weaker, the constancy of their ratio does not depend on the amplitude magnitude and can be extracted by phase-sensitive phase-locked detection combined with multi-cycle averaging stability.

[0079] S5. Match the kinetic fingerprint characterized by the photochemical rate constant based on the concentration-independent harmonic ratio diagram, and separate the amplitude channel data carrying concentration information from the higher-order harmonic tensor slice based on the kinetic fingerprint.

[0080] It should be noted that this step, guided by the concentration-independent harmonic ratio diagram, identifies the kinetic fingerprint spatial unit that truly corresponds to the target antibiotic, and based on this, demixes the high-order harmonic slices that are still mixed with background residues and other components, separating the amplitude channels that only carry target concentration information.

[0081] The matching of the kinetic fingerprint characterized by the photochemical rate constant based on the concentration-independent harmonic ratio diagram includes:

[0082] First, the attention weights of each spatial unit in the concentration-independent harmonic ratio diagram are calculated using a pre-defined attention mechanism network;

[0083] Specifically, the pre-defined attention mechanism network is a network structure that selectively weights input features with attention weights, using the strong physical prior of "pre-defined modified Bessel function ratio distribution law" as the evaluation criterion. The attention mechanism network calculates an attention weight for each spatial unit (corresponding to a wavelength / position) in the concentration-independent harmonic ratio map, and the weight level represents the degree of agreement between the ratio of that unit and the theoretical Bessel ratio of the target antibiotic.

[0084] In one feasible approach, the attention mechanism network uses the spatial unit vectors of the concentration-independent harmonic ratio map as queries and the target vectors as queries. The generated theoretically modified Bessel ratio template is used as the key for each spatial unit. Calculate the deviation between the observed ratio and the theoretical template. The bias is then weighted by a Gaussian kernel and mapped as the attention weight. ,in, The scale parameter for controlling the fit tolerance is determined by the allowable deviation bandwidth of each component of the theoretical template. This network structure can be implemented by a lightweight self-attention layer or a multilayer perceptron, using a theoretical ratio template pre-calibrated with standard samples as a prior constraint. It does not require large-scale data training and is independent of the absolute concentration of the target.

[0085] It should be noted that the theoretical template mentioned It is a reference vector predetermined by the physical properties of the target antibiotic itself, rather than being fitted from the sample to be tested. First, the photochemical rate constant of the target antibiotic is obtained. (It can be pre-characterized by a pure standard sample of the target under the same modulation conditions—by measuring its higher harmonic ratios and performing an inverse solution, or by taking known photochemical parameters of the target), and calculated by combining the known initial light intensity, modulation depth, and preset frequency. Then, for a selected number of higher orders... Calculate the corrected Bessel function ratios one by one. Arrange these ratios into a vector according to their order, which is the theoretical template for the objective.

[0086] For example, the aforementioned =0.5, taking the order pairs (3,2), (4,2), (4,3), by , , The theoretical template can be obtained. The closer the observation ratio vector of a certain spatial cell is to the template, the smaller its deviation. The smaller the value, the more likely it is to be identified as the kinetic fingerprint of the target antibiotic; different antibiotics due to Different z correspond to different template vectors, thus they can be distinguished from each other.

[0087] Secondly, the attention weights corresponding to spatial units that do not conform to the preset distribution law of the modified Bessel function ratio are attenuated to below the preset first weight threshold to shield residual interference data;

[0088] Specifically, the pre-defined distribution law of the modified Bessel function ratio means that for spatial units truly originating from the target antibiotic, the ratios of each harmonic should strictly fall within the range specified by the target antibiotic. (i.e., specific) The determination of ) Near the theoretical curve. For spatial units that deviate from the theoretical distribution (mostly background remnants or heterogeneous components), the network decays their attention weights to below a preset first weight threshold, thereby masking them at the data level.

[0089] Among them, the first weight threshold corresponds to a larger deviation boundary. :when When the attention weight is below the first weight threshold, it is blocked. This deviation boundary (And the corresponding first weight threshold) is usually set based on the statistical upper limit of the deviation of each unit in the blank sample without target. For example, the mean of the deviation of the blank sample is taken plus several times the standard deviation, so that the background residue is judged to be outside the neighborhood with a high probability. The lower the threshold, the more strict the shielding, the purer the units participating in demixing are retained, but the fewer the number. Conversely, it may introduce residual interference.

[0090] Finally, spatial units that conform to the preset modified Bessel function ratio distribution law are assigned attention weights greater than a preset second weight threshold, and the features of spatial units assigned attention weights greater than the preset second weight threshold are identified as dynamic fingerprints.

[0091] Specifically, corresponding to the first weight threshold, the second weight threshold corresponds to a smaller deviation boundary. ( ):when When the attention weight is higher than the second weight threshold, the unit is identified as the kinetic fingerprint of the target antibiotic. The setting method of the second weight threshold is symmetrical to that of the first weight threshold: it is usually determined based on the distribution of the deviation (or weight) of the real target units in the known standard or spiked samples containing the target. For example, the weight corresponding to the upper limit of the statistical upper limit of the deviation of the real target units, or the weight corresponding to the precision of the confirmation result reaching a preset level (such as 95%), is taken as the threshold, so that the vast majority of real target units are confirmed, while occasional interfering units with matching results are excluded; the higher the threshold is set, the only units with extremely high matching degree are confirmed, the more specific the identification, but the lower the recall.

[0092] Depend on and (i.e., the first and second weighted thresholds) jointly define The interval serves as a transition zone. Spatial cells falling into this zone are neither immediately shielded nor directly confirmed. Instead, they can be further determined by the consistency of subsequent tensor unmixing, thereby achieving a balance between anti-interference and integrity.

[0093] Specifically, the amplitude channel data carrying concentration information extracted from the higher-order harmonic tensor slice based on the dynamic fingerprint includes:

[0094] First, using the attention mechanism network, tensor decomposition is performed on the higher-order harmonic tensor slices after weighting with the attention weights in a bilinear domain composed of orthogonal dynamic harmonic axes and spectral axes, to obtain a decomposed feature tensor containing several rank-1 components.

[0095] Specifically, this step involves unmixing within a bilinear domain spanned by the dynamic harmonic axis and the spectral axis. First, the higher-order harmonic tensor slices are weighted with the obtained attention weights to suppress interfering units and highlight fingerprint units. Then, tensor decomposition is performed on the weighted tensor, approximating it as the sum of several rank-one components. ,in For the components on the dynamic harmonic axis, The components on the spectral axis, This represents the outer product. Since the dynamic axis is orthogonal to the spectral axis, different components can be effectively separated within this bilinear domain, resulting in the decomposed characteristic tensor.

[0096] Secondly, based on the feature index of the kinetic fingerprint, the rank-one component whose load on the kinetic harmonic axis matches the kinetic fingerprint is selected as the target rank-one component, and it is restored to the amplitude distribution of the target rank-one component on the spectral axis, as the amplitude channel data corresponding to the target antibiotic.

[0097] Specifically, based on the feature index corresponding to the kinetic fingerprint in the decomposed feature tensor, the rank-one component belonging to the target antibiotic is located as the target rank-one component. The target rank-one component is then re-expanded along the wavelength dimension (i.e., the spectral axis) of the original tensor. The amplitude distribution on the original dimension is restored from the compressed one-dimensional representation, resulting in amplitude channel data containing only target concentration information after the background and fingerprint have been removed.

[0098] In one feasible approach, for each rank-one component Calculate its dynamic harmonic shaft load inner product with fingerprint template Take the one with the largest inner product. For the target component, its corresponding spectral axis loading The amplitude is the amplitude channel data carrying concentration information. Tensor decomposition can be achieved using methods such as CP decomposition and Tucker decomposition; in this decomposition, the dynamic harmonic axis (organized by harmonic order) and the spectral axis (organized by wavelength) belong to different physical bases and are approximately orthogonal, ensuring that different components can be separated. The attention mechanism network can also be replaced by a weighted matching module with equivalent physical prior constraints, and tensor decomposition can also be replaced by methods with equivalent separation capabilities, such as nonnegative matrix decomposition.

[0099] S6. Based on the time evolution weight corresponding to the kinetic fingerprint, perform inversion mapping on the amplitude channel data and output the concentration scalar result of the target antibiotic.

[0100] It should be noted that this step first eliminates the system disturbance caused by light source drift, and then, based on the time evolution weights obtained from the dynamic fingerprint analysis, the amplitude channel data is inverted into a correlation quantity that is linearly corresponding to the absolute concentration, and finally regressed into a single concentration scalar and displayed.

[0101] Furthermore, before performing the inversion mapping on the amplitude channel data based on the time evolution weights corresponding to the dynamic fingerprint, the method further includes:

[0102] First, extract the fundamental frequency amplitude value corresponding to the preset non-absorbing reference band in the matrix of the sample to be tested;

[0103] Specifically, a preset non-absorption reference band, in which the target antibiotic does not absorb, is selected within the same matrix as the target antibiotic, and the first-order fundamental frequency amplitude of this band is extracted. Since this band does not contain the target antibiotic signal, its fundamental frequency amplitude only maps the slow drift of the excitation source itself and can be used as a real-time reference for the light source intensity. The preset non-absorption reference band characterizes the band used to monitor light source drift, and its setting principle is that this band has neither target antibiotic absorption nor is affected by strong matrix absorption as much as possible.

[0104] Secondly, the amplitude channel data is normalized using the fundamental frequency amplitude to eliminate system disturbances caused by the initial light intensity drift of the excitation source, thereby obtaining normalized amplitude channel data.

[0105] Specifically, the separated amplitude channel data is divided by the fundamental frequency amplitude of the reference band to achieve normalization correction of the initial light intensity drift of the light source: ,in For the separated amplitude channel data, The reference band fundamental frequency amplitude.

[0106] For example, if the reference fundamental frequency amplitude drifts from 1.00 to 0.92 during calibration due to heat generation of the light source, normalization can compensate for the 8% intensity drift and eliminate its systematic impact on quantification.

[0107] Specifically, the inversion mapping of the amplitude channel data based on the time evolution weights corresponding to the dynamic fingerprint includes:

[0108] First, the photochemical rate constant contained in the kinetic fingerprint is analyzed;

[0109] Specifically, the dynamic fingerprint confirmed by matching the concentration-independent harmonic ratio map corresponds to a specific oscillation modulation depth; combined with the known initial light intensity, modulation depth, and preset frequency, ... The photochemical rate constant of the target antibiotic can then be solved. This serves as the basis for subsequent calculations of the time evolution weights at each order.

[0110] It should be noted that the separated amplitude channel data has stripped away the fingerprint shape and only retained the concentration amplitude. This step is to parse the rate parameter from the confirmed fingerprint (not the amplitude channel).

[0111] Secondly, the modified Bessel weights corresponding to each order are calculated using the photochemical rate constant, and used as time evolution weights.

[0112] Specifically, obtained from parsing Calculate the values ​​of modified Bessel functions of each order. As the first The time evolution weight of each harmonic is used to characterize the intrinsic scaling factor of the target antibiotic relative to its absolute concentration under its intrinsic photobleaching kinetics.

[0113] Finally, the normalized amplitude channel data is divided by the corresponding time evolution weight to extract the linear correlation data that has a linear relationship with the absolute concentration of the target antibiotic.

[0114] Specifically, by dividing the normalized amplitude channel data by the corresponding order of time evolution weight, the nonlinear modulation of the Bessel weights can be removed, yielding a linearly correlated quantity that corresponds linearly to the absolute concentration: Combined with step S4 It can be seen that, divided by The result after That is, proportional to It is also directly proportional to the absolute concentration.

[0115] For example, taking a second-order channel, the time evolution weights Let the normalized amplitude be... ,but This value is linearly related to the absolute concentration of enrofloxacin.

[0116] The output of the concentration scalar result of the target antibiotic includes:

[0117] First, the linear correlation data is input into the pre-trained inversion regression network layer;

[0118] Specifically, the pre-trained inversion regression network layer uses a series of known concentration standard samples obtained under the same process as training samples and a pre-fitted regression mapping. Its function is to comprehensively map multi-wavelength / multi-order linear correlation quantities into concentration values.

[0119] Secondly, the fully connected function in the inversion regression network layer is used to map the multi-dimensional linear correlation data into a single numerical result.

[0120] Specifically, the inversion regression network layer uses a fully connected function. Multidimensional linear correlation The weighted summation maps to a single concentration value, where, and For the weights and biases obtained from pre-training, This refers to the dimension of the linear correlation quantity (i.e., the order or number of wavelength channels involved in the fusion). Due to the various All values ​​have been linearly correlated with absolute concentration. This regression mainly serves as a multi-channel fusion and calibration conversion tool, achieving a robust output of concentration scalars.

[0121] Finally, the single numerical result is output as a concentration scalar result, and the corresponding numerical rendering and display instructions are triggered.

[0122] Specifically, the above single numerical value is output as a scalar result of the target antibiotic's concentration, triggering numerical rendering and display commands to present the detection result to the user in numerical or graphical form. The identification and quantification of the target antibiotic's concentration were completed under strong matrix background interference, achieving precise decoupling detection of identification and concentration.

[0123] For ease of understanding, the complete numerical flow of this embodiment is presented below: at a preset frequency =1000Hz ), modulation depth A sinusoidal light intensity of 0.8 was applied to a flowing milk sample (target substance enrofloxacin), and the residence time resulted in a cumulative light dose. ≈5.9, Oscillation Modulation Depth =0.5, photobleaching modulation factor Exceeding the threshold ;by =20000Hz sampling to construct modulation-excitation time-series spectral matrix Using a known sinusoidal waveform as a reference, phase-locked demodulation is used to obtain the harmonic dispersion tensor; the first-order fundamental frequency is removed, and second-order and higher-order slices are retained to eliminate linear matrix background and broadband scattering; the third-order and second-order tensors are then compared to obtain... The enrofloxacin dynamic fingerprint was matched; the amplitude channel was separated by attention filtering and tensor raising-rank unmixing and normalized to the reference fundamental frequency. ; obtained from reverse fingerprinting And thus, the weight of time evolution. The linear correlation quantity is obtained by inversion. Finally, the concentration scalar is mapped and rendered by the inversion regression network layer. Only by The decision is independent of concentration and serves to confirm identity. Linearly corresponding to absolute concentration, carrying capacity concentration is inverted.

[0124] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented 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.

[0125] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0126] The preferred embodiments of the invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A method for intelligent spectral detection of antibiotic residues based on time-series feature decoupling, characterized in that, Includes the following steps: The sample to be tested is irradiated with an excitation source whose initial light intensity is sinusoidally modulated at a preset frequency. The target antibiotic with irreversible photolysis characteristics at the wavelength of the excitation source is depleted by nonlinear photobleaching under the cumulative light dose. The spontaneously generated harmonic comb data is continuously collected to construct the modulation excitation time-series spectral matrix. The modulation excitation time-series spectral matrix is ​​synchronously demodulated according to each harmonic of the preset frequency to generate a harmonic dispersion tensor containing multiple harmonic orders. Filter out the fundamental frequency tensor slices of the first order in the harmonic dispersion tensor, and retain the higher-order harmonic tensor slices of the second order or higher, to obtain higher-order harmonic tensor slices that have eliminated the matrix background with linear response. Using the amplitude ratios between the higher-order harmonic tensor slices of different orders, a concentration-independent harmonic ratio diagram, which is determined solely by the photochemical rate constant of the target antibiotic after eliminating absolute concentration, is calculated. The concentration-independent harmonic ratio plot is matched with the kinetic fingerprint characterized by the photochemical rate constant, and the amplitude channel data carrying concentration information is separated from the higher-order harmonic tensor slice based on the kinetic fingerprint. Based on the time evolution weights corresponding to the kinetic fingerprint, the amplitude channel data is inverted and mapped to output the scalar concentration result of the target antibiotic.

2. The intelligent detection method for antibiotic residues based on temporal feature decoupling according to claim 1, characterized in that, The construction of the modulation excitation time-series spectral matrix includes: Fluid samples in a state of natural flow are used as test samples; The residence time of the fluid sample flowing through the light spot area of ​​the excitation source and the modulation depth of the light intensity of the excitation source are controlled so that the photobleaching modulation factor determined by the residence time, the modulation depth, the initial light intensity and the preset frequency reaches a preset threshold, so as to ensure that the target antibiotic produces nonlinear photobleaching depletion and that the harmonic amplitude of the order greater than or equal to the second order is higher than the preset resolvable lower limit. Using a preset sampling rate greater than or equal to the preset highest interest harmonic order corresponding to the Nyquist frequency, harmonic comb data spontaneously generated under the nonlinear photobleaching depletion effect are continuously collected to construct a two-dimensional modulation excitation time-series spectral matrix.

3. The intelligent detection method for antibiotic residues based on temporal feature decoupling according to claim 1, characterized in that, The step of synchronously demodulating the modulation excitation time-series spectral matrix according to each harmonic of the preset frequency to generate a harmonic dispersion tensor containing multiple harmonic orders includes: Obtain the known sinusoidal modulation waveform data corresponding to the preset frequency as a phase-locked reference signal; Using the phase-locked reference signal, phase-locked demodulation operation is performed on the modulation excitation timing spectrum matrix along the time axis in each preset wavelength channel to extract the amplitude data and phase data corresponding to each harmonic; From the extracted amplitude data and phase data, the known mixing crosstalk term caused by the modulation illumination component of the excitation light source is subtracted to obtain the corrected amplitude data and phase data, and a harmonic dispersion tensor is constructed based on the corrected amplitude data and phase data.

4. The intelligent detection method for antibiotic residues based on temporal feature decoupling according to claim 1, characterized in that, The retention of higher-order harmonic tensor slices with an order greater than or equal to second includes: retaining nonlinear harmonic components that form narrowband coherent stacks at the corresponding frequencies of higher-order harmonics, and removing broadband scattering noise that does not form narrowband coherent stacks at the corresponding frequencies of higher-order harmonics.

5. The intelligent detection method for antibiotic residues based on temporal feature decoupling according to claim 1, characterized in that, The calculated concentration-independent harmonic ratio diagram, after eliminating absolute concentrations, is determined solely by the photochemical rate constant of the target antibiotic and includes: Extract the target order higher harmonic amplitude and the reference order higher harmonic amplitude from the higher harmonic tensor slice; The target order higher harmonic amplitude and the reference order higher harmonic amplitude are divided element-wise to eliminate the common multiplier data composed of absolute concentration and static spectral characteristics. The output is a ratio matrix containing only the modified Bessel function ratio determined by the photochemical rate constant, which serves as the concentration-independent harmonic ratio map.

6. The intelligent detection method for antibiotic residues based on temporal feature decoupling according to claim 1, characterized in that, The matching of the kinetic fingerprint characterized by the photochemical rate constant based on the concentration-independent harmonic ratio diagram includes: The attention weights of each spatial cell in the concentration-independent harmonic ratio diagram are calculated using a pre-defined attention mechanism network. The attention weights corresponding to spatial units that do not conform to the preset distribution law of the modified Bessel function ratio are attenuated to below the preset first weight threshold to block residual interference data. Spatial cells that conform to the preset modified Bessel function ratio distribution law are assigned attention weights greater than a preset second weight threshold, and the features of spatial cells assigned attention weights greater than the preset second weight threshold are identified as dynamic fingerprints.

7. The intelligent detection method for antibiotic residues based on temporal feature decoupling according to claim 6, characterized in that, The step of separating amplitude channel data carrying concentration information from the higher-order harmonic tensor slice based on the dynamic fingerprint includes: Using the attention mechanism network, tensor decomposition is performed on the higher-order harmonic tensor slices after weighting with the attention weights in a bilinear domain composed of orthogonal dynamic harmonic axes and spectral axes, to obtain a decomposed feature tensor containing several rank-1 components. Based on the feature index of the kinetic fingerprint, the rank-one component whose load on the kinetic harmonic axis matches the kinetic fingerprint is selected as the target rank-one component, and it is restored to the amplitude distribution of the target rank-one component on the spectral axis, as the amplitude channel data corresponding to the target antibiotic.

8. The intelligent detection method for antibiotic residues based on temporal feature decoupling according to claim 1, characterized in that, Before performing the inversion mapping on the amplitude channel data based on the time evolution weights corresponding to the dynamic fingerprint, the method further includes: Extract the fundamental frequency amplitude corresponding to the preset non-absorbing reference band in the matrix of the sample under test; The amplitude channel data is normalized using the fundamental frequency amplitude to eliminate system disturbances caused by the initial light intensity drift of the excitation source, thereby obtaining normalized amplitude channel data.

9. The intelligent detection method for antibiotic residues based on temporal feature decoupling according to claim 8, characterized in that, The inversion mapping of the amplitude channel data based on the time evolution weight corresponding to the dynamic fingerprint includes: The photochemical rate constant contained in the kinetic fingerprint was analyzed; The modified Bessel weights corresponding to each order are calculated using the photochemical rate constant and used as time evolution weights. Divide the normalized amplitude channel data by the corresponding time evolution weight to extract the linear correlation data that has a linear relationship with the absolute concentration of the target antibiotic.

10. The intelligent detection method for antibiotic residues based on temporal feature decoupling according to claim 9, characterized in that, The output of the concentration scalar result of the target antibiotic includes: The linear correlation data is input into the pre-trained inversion regression network layer; By utilizing the fully connected function in the inversion regression network layer, the multi-dimensional linear correlation data is mapped into a single numerical result; The single numerical result is output as a concentration scalar result, and the corresponding numerical rendering and display instructions are triggered.

Images