Near-infrared polarization spectrum detection probe and detection method based on multi-measuring-point cooperation

Abstract

The invention provides a near-infrared polarization spectrum detection probe based on multi-measuring-point cooperation, the probe is suitable for multi-measuring-point high-precision online detection of solid materials in a continuous production line, the probe comprises a spectrum acquisition module, the spectrum acquisition module comprises a light source, a reference plate, an upper rotating device and a collimating lens which are symmetrically installed in parallel; the light source is fixed through a left light source clamping plate and a right light source clamping plate, a certain angle is formed between the light source and a detection optical axis, and the reference plate is connected with the upper rotating device and is driven by the upper rotating device to rotate to the position below the collimating lens so as to collect a background spectrum; wherein the probe is integrated with a multi-channel parallel calibration system, each measuring point is independently provided with a reference channel, and the baseline difference between the measuring points is eliminated based on a dynamic compensation model by alternately collecting reference signals and material spectral signals of the measuring points; according to the scheme, the influence of stray light caused by calibration mismatch due to multi-measuring-point signal mode difference is reduced, and the online spectrum detection precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of optical detection automation control, and in particular to a near-infrared polarization spectroscopy detection probe and detection method based on multi-point measurement collaboration. Specifically, it relates to a near-infrared polarization spectroscopy detection probe and detection method for fixed materials based on multi-point measurement collaboration. Background Art

[0002] Near-infrared spectroscopy detection technology has been widely used in the detection process of solid materials due to its efficient and non-destructive characteristics. However, due to the complexity of the material system of solid materials, their spectral signals are easily affected by the external environment and scattered light noise, which contains both target characteristic signals and non-target signals such as particle scattering and environmental stray light.

[0003] In a continuous production scenario, the spectral signals of materials at different measurement points will show diverse characteristics due to significant differences in their physical forms (particle size / pore structure) and chemical properties (carbonization degree / oxygen-containing functional groups). For example: the material surface at measurement point 1 is rough and rich in fiber pores, and the scattered light is mainly diffuse reflection; the material at measurement point 2 forms a transition interface due to local carbonization, and the scattering mode shows a mixed reflection characteristic; the material surface at measurement point 3 forms a smooth microcrystalline structure, and the scattered light is mainly specular reflection. The spectral signals of each measurement point not only have a gradient change in intensity distribution, but also their baseline drift characteristics show obvious differences with the change of the material surface characteristics.

[0004] Traditional reference calibration methods usually adopt globally unified reference parameters through a timing calibration method. This method, which optimizes the calibration model for strong scattering signals, will cause overcompensation or undercompensation due to the mismatch of signal distribution patterns when applied to weak absorption signals, reducing the consistency of multi-point measurement collaboration detection. Summary of the Invention

[0005] In view of the deficiencies of the prior art, the present invention proposes a near-infrared polarization spectroscopy detection probe based on multi-point measurement collaboration. The probe is suitable for high-precision online detection of multiple measurement points of solid materials in a continuous production line. The probe includes:

[0006] A spectral acquisition module, which includes a light source, a reference plate, an upper rotation device, and a collimating lens that are symmetrically and parallelly installed. The light source is fixed by a left light source clamping plate and a right light source clamping plate, and is arranged at a certain angle with respect to the detection optical axis. The reference plate is connected to the upper rotation device and is driven by the upper rotation device to rotate below the collimating lens to collect background spectra;

[0007] Among them, the probe integrates a multi-channel parallel calibration system, and each measurement point is independently configured with a reference channel. By alternately collecting the reference signals and material spectral signals of each measurement point, the baseline difference between the measurement points is eliminated based on a dynamic compensation model.

[0008] In some of these embodiments, the dynamic compensation model includes the following compensation terms:

[0009] The particle size compensation term (ΔC g ), which is calculated based on the median particle size (D i ) of the material at the measurement point and the reference particle size (A); the surface reflection compensation term (ΔC s ), which is calculated based on the surface roughness (R i ) of the material at the measurement point and the reference roughness (B); the density attenuation compensation term (ΔC d ), which is calculated based on the bulk density (ρ i ) of the material at the measurement point and the reference density (C); the formulas for the compensation terms are respectively:

[0010]

[0011] The calibration equation is:

[0012]

[0013] Note: I ref is the reference signal corresponding to the measurement point, I ref ∝I out ; I dark is the dark noise signal; I sample =I target +I out ; ΔC i includes ΔC g , ΔC d , ΔC s .

[0014] In some of these embodiments, the spectral acquisition module further includes a polarizer, and the polarizer uses an electric rotating polarizer, an integrated rotary encoder or other real-time reading systems to provide real-time feedback on the rotation angle θ of the polarizer.

[0015] In some of these embodiments, the transmission axis direction of the polarizer is orthogonal to the initial polarization state of the incident light, and is adjusted to according to the polarization deflection angle of the material to achieve physical separation of the target absorption signal and the scattered light. The intensity of the target signal after passing through the polarizer is:

[0016]

[0017] In some of these embodiments, the probe further includes:

[0018] A parameter acquisition module, including: a particle size sensor, a roughness sensor and a bulk density sensor of the material; the parameter acquisition module is used to read the relevant required parameters in real time.

[0019] The present invention also provides a near-infrared polarization spectroscopy detection method based on multi-measurement point collaboration. Using the probe as described above, the detection method includes the following steps:

[0020] Uniformly irradiate the material with the symmetric halogen lamps of the multi-measurement point probe, and alternately open the shutter to collect the reference signal and the material spectrum signal of each measurement point;

[0021] Use a polarizer to separate the target absorption signal and the scattered light, where the direction of the transmission axis is adjusted to Suppress the scattered light noise;

[0022] Calculate the dynamic compensation term based on the particle size, surface roughness and density parameters of the material, and map the multi-measurement point spectral data to a unified reference;

[0023] Adopt a hybrid neural network model to fuse multi-source features, including spectral signals, polarization parameters and physical property parameters, and optimize the model parameters through incremental learning.

[0024] In some embodiments, adopting a hybrid neural network model to fuse multi-source features, including spectral signals, polarization parameters and physical property parameters, and optimizing the model parameters through incremental learning, includes:

[0025] Input the physical property parameters (ΔC g , ΔC s , ΔC d ) in the compensation equation into the fully connected layer to generate the physical constraint vector V phy ;

[0026] Adopt a convolutional layer (CNN) to extract the local features of the spectral signal, and capture the response relationship in the key wavelength region through an attention mechanism, and output the feature vector V spec ;

[0027] Use a bidirectional long short-term memory network (Bi-LSTM) to model the temporal dependence of the polarization parameters, and output the feature vector V pol ; Dynamically fuse V phy , V spec , V pol , and the formula is:

[0028] V = σ(W[V phy ; V spec ; V pol )

[0029] Note: σ is the Sigmoid function, and W is the learnable parameter matrix.

[0030] In some embodiments, the compensated spectral signals (I calibrated)Perform multi-dimensional fusion with polarization characteristic parameters (polarization deflection angle absorption attenuation coefficient α);

[0031] Utilize the polarization filter transmission intensity formula to invert the polarization response characteristics of the material and generate a polarization characteristic matrix P;

[0032] Based on the physical property parameters of the measurement points (particle size D i , roughness R i , density ρ i ), calculate the dynamic weight coefficient ω i , and the formula is:

[0033]

[0034] Generate a global feature vector F by weighted fusion of the spectral data of each measurement point.

[0035] In some of these embodiments, the incremental learning to optimize the model parameters includes: by monitoring the prediction error distribution in real time, adjusting the compensation term coefficient in the calibration equation, and using a sliding window to update the historical data to maintain the stability of the model.

[0036] The present invention also provides a computer-readable storage medium, in which at least one program code is stored, and the at least one program code is loaded and executed by a processor to implement the above-mentioned near-infrared polarization spectroscopy detection method based on multi-measurement point collaboration.

[0037] As can be seen from the above solutions, the advantages of the present invention are:

[0038] The near-infrared polarization spectroscopy detection probe and detection method based on multi-measurement point collaboration provided by the present invention are applicable to multi-terminal reference calibration and scattered light suppression technologies in solid material property detection scenarios. The solution provided by the present invention reduces the calibration mismatch caused by the signal mode differences of multiple measurement points, reduces the influence of stray light, and improves the accuracy of spectral on-line detection.

[0039] The technical solution provided by the present invention forms a multi-channel parallel calibration system by integrating a reference spectrum real-time acquisition module in the detection optical paths of multiple measurement points respectively, and directly integrates a polarization filter at the light outlet of the spectrometer. By utilizing the orthogonality difference between the characteristic signal of the material and the polarization state of the scattered light, pre-separation of the scattered light at the physical level is achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a schematic structural diagram of the near-infrared polarization spectroscopy detection probe based on multi-measurement point collaboration shown in the embodiments of the present invention;

[0041] Figure 2This is another schematic diagram of the near-infrared polarization spectrum detection probe based on multi-measurement point cooperation shown in the embodiments of the present invention;

[0042] Figure 3 This is the overall flow schematic diagram of the near-infrared polarization spectrum detection method based on multi-measurement point cooperation shown in the embodiments of the present invention;

[0043] Among them, reference numerals:

[0044] 1 - Light source;

[0045] 2 - Reference plate;

[0046] 3 - Upper rotation device;

[0047] 4 - Left light source clamping plate;

[0048] 5 - Right light source clamping plate;

[0049] 6 - Polarizer;

[0050] 7 - Particle size sensor;

[0051] 8 - Roughness sensor;

[0052] 9 - Material bulk density sensor;

[0053] 10 - Left lamp holder;

[0054] 11 - Right lamp holder;

[0055] 12 - Shutter;

[0056] 13 - Collimating lens holder;

[0057] 14 - Upper rotation robotic arm;

[0058] 15 - Brush;

[0059] 16 - Lower rotation device;

[0060] 17 - Lower rotation robotic arm;

[0061] 18 - Electronic control board;

[0062] 19 - Step-down module;

[0063] 20 - Outer shell;

[0064] 21 - Integrated rotary encoder. Detailed implementation manners

[0065] To make the above features and effects of the present invention more clearly and understandably described, specific embodiments are hereinafter given and detailed descriptions are made in conjunction with the accompanying drawings of the specification as follows.

[0066] See Figure 1-2, an embodiment of the present invention provides a near-infrared polarization spectroscopy detection probe based on multi-point measurement collaboration. The probe is suitable for multi-point high-precision on-line detection of solid materials in a continuous production line. The probe includes: a spectral acquisition module, and the spectral acquisition module includes a light source 1, a reference plate 2, an upper rotation device 3, and a collimating lens (not shown in the figure) that are symmetrically and parallely installed; the light source 1 is fixed by a left light source clamping plate 4 and a right light source clamping plate 5, and is arranged at a certain angle with the detection optical axis to uniformly irradiate the material. The reference plate 2 is connected to the upper rotation device 3 and is driven by the upper rotation device 3 to rotate below the collimating lens to collect the background spectrum;

[0067] Among them, the probe integrates a multi-channel parallel calibration system. Each measurement point is independently configured with a reference channel. By alternately collecting the reference signals and material spectral signals of each measurement point, and based on a dynamic compensation model, the baseline differences between the measurement points are eliminated.

[0068] In this embodiment, the dynamic compensation model includes the following compensation items:

[0069] Particle size compensation item (ΔC g ), calculated based on the median particle size (D i ) of the material at the measurement point and the reference particle size (A); surface reflection compensation item (ΔC s ), calculated based on the surface roughness (R i ) of the material at the measurement point and the reference roughness (B); density attenuation compensation item (ΔC d ), calculated based on the bulk density (ρ i ) of the material at the measurement point and the reference density (C); the formulas for the compensation items are respectively:

[0070]

[0071] The calibration equation is:

[0072]

[0073] Note: I ref is the reference signal corresponding to the measurement point, I ref ∝I out ; I dark is the dark noise signal; I sample =I target +I out ; ΔC i includes ΔC g , ΔC d , ΔC s .

[0074] Among them, the spectral acquisition module further includes a polarizer 6. The probe can be equipped with an electric rotating polarizer (such as an optoelectronic polarization sensor lamp), an integrated rotary encoder or other real-time reading systems to add the polarizer, and the rotation angle θ of the polarizer is fed back in real time.

[0075] Among them, the transmission axis direction of the polarizer is orthogonal to the initial polarization state of the incident light, and is adjusted according to the polarization deflection angle of the material to achieve physical separation of the target absorption signal and the scattered light. The intensity of the target signal after passing through the polarizer is:

[0076]

[0077] Among them, the probe further includes: a parameter acquisition module, including: a particle size sensor 7, a roughness sensor 8, and a material bulk density sensor 9, but not limited thereto; the parameter acquisition module is used to read relevant required parameters in real time.

[0078] In this embodiment, the multi-channel parallel calibration system retains the timing data through a sliding window mechanism and adopts an incremental learning strategy to dynamically adjust the compensation item coefficients (k1, k2, k3) to adapt to the continuous change of material characteristics.

[0079] Furthermore, the probe in this embodiment further includes a lens cleaning module and an electrical control module. The probe is mainly composed of four parts: a spectral acquisition module, a lens cleaning module, an electrical control module, and a parameter acquisition module. In this embodiment, the spectral acquisition module further includes a left lamp holder 10, a right lamp holder 11, a shutter 12, a collimating lens holder 13, and an upper rotating robotic arm 14. The shutter 12 is used to control the opening and closing of the acquisition window; the lens cleaning module includes a brush 15, a lower rotating device 16, and a lower rotating robotic arm 17; the electrical control module includes an electronic control board 18 and a step-down module 19; the above components are integrally wrapped by a box-shaped hardware housing 20 and suspended above the production line feeding conveyor belt.

[0080] See Figure 3 , another embodiment of the present invention provides a near-infrared polarization spectroscopy detection method based on multi-point measurement collaboration, using the probe as described in the foregoing embodiment. The detection method includes the following steps:

[0081] S1. Uniformly irradiate the material with the symmetric halogen lamps of the multi-point measurement probe, and alternately open the shutter to collect the reference signal and the material spectral signal of each measurement point;

[0082] S2. Use the polarizer to separate the target absorption signal and the scattered light, where the transmission axis direction is adjusted to to suppress the scattered light noise;

[0083] S3. Calculate the dynamic compensation term based on the particle size, surface roughness, and density parameters of the material, and map the multi-measurement point spectral data to a unified reference;

[0084] S4. Use a hybrid neural network model to fuse multi-source features, including spectral signals, polarization parameters, and physical property parameters, and optimize the model parameters through incremental learning.

[0085] In this embodiment, the hybrid neural network model includes: a fully connected layer for generating a physical constraint vector (V phy ); a feature extraction layer for extracting local and global features of the spectral signal (V spec ); a bidirectional long short-term memory network (Bi-LSTM) for modeling the temporal dependence of the polarization parameters (V pol ); and a gating mechanism for dynamically fusing the above feature vectors to output the final prediction result.

[0086] In this embodiment, using a hybrid neural network model to fuse multi-source features, including spectral signals, polarization parameters, and physical property parameters, and optimizing the model parameters through incremental learning includes:

[0087] Input the physical property parameters (ΔC g , ΔC s , ΔC d ) in the compensation equation into the fully connected layer to generate a physical constraint vector V phy ;

[0088] Use a convolutional layer (CNN) to extract the local features of the spectral signal, and capture the response relationship in the key wavelength region through an attention mechanism, and output a feature vector V spec ;

[0089] Use a bidirectional long short-term memory network (Bi-LSTM) to model the temporal dependence of the polarization parameters, and output a feature vector V pol ; Dynamically fuse V phy , V spec , V pol through a gating mechanism, and the formula is:

[0090] V = σ(W[V phy ; V spec ; V pol )

[0091] Note: σ is the Sigmoid function, and W is a learnable parameter matrix.

[0092] Furthermore, in this embodiment, the compensated spectral signals (I calibrated ) of each measurement point and the polarization characteristic parameters (polarization deflection angle absorption attenuation coefficient α) are fused in multiple dimensions;

[0093] Using the transmission intensity formula of a polarizer Invert the polarization response characteristics of the material to generate a polarization characteristic matrix P;

[0094] Based on the physical property parameters of the measurement points (particle size D i , roughness R i , density ρ i ), calculate the dynamic weight coefficient ω i , and the formula is:

[0095]

[0096] Generate a global feature vector F by weighted fusion of the spectral data of each measurement point.

[0097] In this embodiment, the parameters of the incremental learning optimization model include: by real-time monitoring of the prediction error distribution, adjusting the compensation term coefficient in the calibration equation, and using a sliding window to update historical data to maintain the stability of the model.

[0098] The near-infrared polarization spectroscopy detection probe and detection method based on multi-measurement point collaboration provided by the embodiments of the present invention are applicable to high-precision online detection of multiple measurement points of solid materials in a continuous production line, including but not limited to the characteristic analysis of carbonized materials, fiber porous materials, and microcrystalline structure materials.

[0099] The following further details the near-infrared polarization light detection device and detection method based on multi-measurement point collaboration disclosed in the embodiments of the present invention.

[0100] For a single detection probe, two symmetrically and parallelly installed light sources 1 are used, and power is separately taken from the control cabinet through a step-down module 19. The light source 1 is clamped on the left lamp holder 10 by the left light source clamp plate 4 and forms an angle with the detection optical axis to ensure uniform illumination of the target material. To improve the stability of the light source, the halogen lamp in the probe remains constantly on after the system is turned on; the shutter 12 is fixed on the collimating lens bracket 13 and can close or open the acquisition window within a specified period; the reference plate 2 is embedded with the upper rotating robotic arm 14. Polarizers 6 and integrated rotary encoders 21 are respectively integrated before the probe optical fiber receives light at the light receiving port of the probe, and the rotation angle θ of the polarizer is real-time feedback, and the direction of its transmission axis is adjusted to be orthogonal to the polarization state of the incident light, so as to allow the intrinsic absorption signal parallel to the transmission axis to pass through, eliminate the light that directly enters the optical fiber without reaching the sample from the light source, and the light reflected by the sample mirror surface, so as to receive most of the diffuse reflection light from the sample.

[0101] Specifically, when collecting the spectrum of solid materials at the measurement point, the light source irradiates the solid materials through the window, and the diffuse reflection spectrum is transmitted to the spectrometer through the optical fiber. At the same time, the shutter 12 under the collimating mirror is closed. After passing through the polarizer 6, due to the specific polarization response characteristics of the intrinsic absorption signal of the target material (such as the vibration absorption peak of oxygen-containing functional groups in straw), its polarization direction is consistent with the initial polarization state of the incident light (assumed to be the linear polarization direction θ). The scattered light generated on the surface of the solid material (such as particle scattering, interface reflection, etc.) has a random distribution characteristic of polarization state due to the multiple scattering effect. By adjusting the light transmission axis direction of the polarizer 6 to (where is the polarization deflection angle of the solid material), the following separation effects can be achieved:

[0102] a. The intrinsic absorption signal maintains the same polarization state θ as the incident light when penetrating the material. However, since the light transmission axis direction of the polarizer is the theoretical transmission intensity of the target signal should approach zero. However, due to the anisotropic absorption and secondary polarization modulation effects existing in the actual material, some components parallel to the light transmission axis direction in the target signal are retained (such as direction), and thus are selectively transmitted.

[0103] b. Among the random polarization states of the scattered light, only the components consistent with the light transmission axis direction can pass through the polarizer. Due to the extremely low statistical correlation between the polarization direction of the scattered light and the light transmission axis direction its transmission intensity is significantly attenuated.

[0104] Assume that the initial polarization state of the incident light is linearly polarized in the θ direction, and its Jones vector is expressed as:

[0105]

[0106] Note: E in is the electric field vector of the incident light; E0 is the initial polarization of the incident light.

[0107] After the target signal acts on the material, the polarization state deflects due to anisotropic absorption and can be decomposed into:

[0108]

[0109] Note: E target is the electric field vector of the target material after transmission; α is the absorption attenuation coefficient; is the polarization deflection angle, which is determined by the characteristics of the material.

[0110] The polarization state of the scattered light is randomly distributed and can be expressed as a coherent superposition:

[0111]

[0112] Note: I is the polarization state intensity of the scattered light; I0 is the initial intensity of the incident light.

[0113] The transmission axis direction of the polarized light is Its Jones matrix is:

[0114]

[0115] Note: J is the Jones matrix.

[0116] The target signal intensity after passing through the polarizer is:

[0117]

[0118] The statistical mean of the scattered light projection intensity follows Malus' law:

[0119]

[0120] Note: I out The statistical average intensity of the scattered light after passing through the polarizer.

[0121] Furthermore, in the multi - measurement point continuous detection network of the continuous production line, an independent reference channel is embedded in each measurement point probe, and a composite calibration module integrating a white board and a reference board 2 is used. The white board is used to collect the scattered reference drift data in real time; the reference board 2 is used to monitor the light source intensity fluctuation and dark current noise. When collecting the background spectral data, the upper rotation device 3 drives the upper rotation robotic arm 14 to drive the reference board 2 to rotate 90 degrees under the collimating lens, and at the same time, the shutter 12 of the probe closes; after the background spectral detection is completed, the reference board 2 swings back to its original position with the upper rotation device 3. The multi - measurement point probes are respectively embedded with shutters 12, and the spectral signals of the solid materials at each measurement point can be alternately collected by controlling the shutter 12.

[0122] Multiple points alternately sample to collect reference information and material information. The collection interval time of the reference information can be set according to the working conditions, and a calibration equation for the measurement point is established:

[0123]

[0124] Note: I calibrated is the calibrated spectral signal; -log10 converts the spectral signal into absorbance; I ref is the reference signal corresponding to the measurement point, measured on the reference board, reflecting the environmental scattered light I out and light source fluctuation, so I ref ∝I out ; I dark is the dark noise signal; I sample is the sample spectral signal, including the target signal I target and the scattered light transmission intensity I out, that is: I sample = I target + I out .

[0125] In the continuous production process, due to the differences in the solid materials at each measuring point, it is necessary to perform physical property compensation on the calibration equation:

[0126] a. Particle size effect: Large particle solid materials have positive compensation. As the light scattering increases, the signal will weaken; small particle solid materials have negative compensation. As the light absorption increases, the signal will increase. When the brush 15 swings above the solid material, the median particle size of the solid material at this moment is read in real time through the particle size sensor 7 and substituted into the compensation term formula for calculation:

[0127]

[0128] Note: D i is the median particle size of the material at measuring point i, and A is the reference particle size.

[0129] b. Surface reflection effect: The surface of the solid material is rough and rich in fiber pores, and the scattered light is mainly diffuse reflection; the surface of the solid material forms a smooth microcrystalline interface due to high-temperature carbonization, and the scattered light is mainly specular reflection. When the brush 15 swings above the solid material, the surface roughness of the solid material at this moment is read in real time through the roughness sensor 8 and substituted into the compensation term formula for calculation:

[0130]

[0131] Note: R i is the surface roughness of the material at measuring point i, and B is the reference roughness.

[0132] c. Density attenuation effect: Low-density solid materials have positive compensation, large voids, and virtual increase in the optical path; high-density solid materials have negative compensation, small voids, and signal saturation. When the brush 15 swings above the solid material, the bulk density of the solid material at this moment is read in real time through the bulk density sensor 9 of the material and substituted into the compensation term formula for calculation:

[0133]

[0134] Note: ρ i is the bulk density of the material at measuring point i, and C is the reference density.

[0135] Establish a calibration equation with a unified reference for multiple measuring points:

[0136]

[0137] Note: ΔC i includes ΔC g , ΔC d , ΔCs 。

[0138] Furthermore, after eliminating the baseline differences between multiple detection points through multi-measurement point calibration, the spectral data of multiple measurement points are mapped to the same scale, and a prediction model is established to achieve high-precision detection of multi-measurement points collaboration of solid materials.

[0139] The establishment of the standard model in the embodiments of the present invention is as follows:

[0140] (1) Multi-source feature fusion and dynamic weight allocation

[0141] The compensated spectral signals (I calibrated ) of each measurement point are fused with polarization feature parameters (polarization deflection angle absorption attenuation coefficient α) in multiple dimensions. Using the polarization sheet transmission intensity formula to invert the polarization response characteristics of the material and generate a polarization feature matrix P.

[0142] Based on the physical property parameters of the measurement point (particle size D i , roughness R i , density ρ i ), calculate the dynamic weight coefficient ω i , and the formula is:

[0143]

[0144] Generate a global feature vector F by weighted fusion of the spectral data of each measurement point.

[0145] (2) Hybrid neural network architecture design

[0146] Construct a multi-branch attention network, including: inputting the physical property parameters (ΔC g , ΔC s , ΔC d ) in the compensation equation into the fully connected layer to generate a physical constraint vector V phy . Use a convolutional layer (CNN) to extract the local features of the spectral signal, and capture the response relationship in the key wavelength region through the attention mechanism, and output a feature vector V spec . Use a bidirectional long short-term memory network (Bi-LSTM) to model the temporal dependence of the polarization parameters, and output a feature vector V pol . Dynamically fuse V phy , V spec , V pol through a gating mechanism, and the formula is:

[0147] V = σ(W[V phy ; V spec ; V pol )

[0148] Note: σ is the Sigmoid function, and W is the learnable parameter matrix.

[0149] (3) Adaptive incremental learning and online optimization

[0150] In response to the dynamic changes in the material properties in the continuous production scenario, an incremental model update strategy is designed: the latest spectral data and detection results of each measurement point are stored in real time, and a sliding window mechanism is adopted to retain the temporal correlation. By monitoring the cumulative distribution of the prediction error, the compensation term coefficients (such as k1, k2, k3) in the calibration equation are dynamically adjusted to ensure the long-term stability of the model.

[0151] The following takes the online detection applied to the biomass carbonization production line as an example to introduce the technical solution provided by the embodiment of the present invention.

[0152] (1) Detection content

[0153] Measurement point 1: Detect the raw material straw (with a moisture content of 15%-25%, high fiber content, rough surface, and developed pores).

[0154] Measurement point 2: Detect the finished biochar (carbonization temperature of 600°C, a dense microcrystalline structure is formed on the surface, and the oxygen-containing functional groups are reduced).

[0155] (2) Probe device

[0156] Light source: Halogen lamps are symmetrically installed on both sides (wavelength 350 - 2500 nm), at an angle of 45° with the detection optical axis, ensuring uniform illumination of the straw surface.

[0157] Polarizer: The value of θ is read in real time. The transmission axis direction of measurement point 1 is adjusted to the polarization state of the incident light θ + 90°. Due to the anisotropic absorption and secondary polarization modulation effects inside the material, some components parallel to the transmission axis direction (such as the θ + 90° direction) are retained in the target signal, and thus are selectively passed through to suppress the diffuse reflection scattered light of the straw. For measurement point 2, due to the high degree of carbonization of the solid material and the tight pore structure, the polarization response has strong anisotropy, and the deflection angle of the polarization direction from the initial polarization state (θ) of the incident light can reach 75 - 85°. The transmission axis direction of the polarizer is adjusted to θ + 75°.

[0158] Reference module: A standard polytetrafluoroethylene white board, which is switched to below the collimating lens by a servo drive to collect the background spectrum at a high frequency (interval of 10 seconds).

[0159] Cleaning module: A servo-driven brush is used to clean the dust on the surface of the collimating lens every 30 minutes.

[0160] (3) Detection process

[0161] Step 1: High-frequency alternating calibration of multiple measurement points

[0162] Measurement point 1: Switch the reference plate to below the collimating lens every 10 seconds, close the Shutter, and collect the background spectrum and dark current; Measurement point 2: Close the Shutter every 8 seconds, and collect the reference signal and dark current.

[0163] Calculate the compensation terms. ① Particle size compensation term: Take D1 = 1.2 mm, A = 1 mm; ② Surface reflection compensation term: Take R1 = 15 μm, B = 10 μm; ③ Density compensation term: Take ρ1 = 0.6 g / cm 3 , C = 0.8 g / cm 3 . Perform dynamic compensation through the calibration equation:

[0164]

[0165] Use a sliding module to store the real-time data of measurement point 1 and measurement point 2, where the data includes D i , R i , ρ i and the prediction error. Perform incremental learning every 30 minutes, and adjust the compensation term coefficients by the gradient descent method.

[0166] Step 2: Polarization separation and signal acquisition

[0167] At measurement point 1, turn on the Shutter, irradiate the surface of the straw with a halogen lamp, and the diffuse reflection light is filtered by a polarizer (θ + 90°) to selectively pass the target signal, with a collection period of 2 seconds / time; At measurement point 2, the polarizer (θ + 75°) matches the polarization characteristics of the biochar to transmit the target signal, with a collection period of 1.5 seconds / time.

[0168] Step 3: Data fusion and model prediction

[0169] (1) Multi-source feature extraction

[0170] Map the calibration spectra of measurement point 1 and measurement point 2 to the same principal component space, extract the principal components, and eliminate the baseline drift. Decompose the spectral signal using Daubechies-4 basis functions to extract the high-frequency details in the wavelength domain.

[0171] (2) Hybrid neural network architecture

[0172] Input the compensation term parameters (ΔC g , ΔC s , ΔC d ), and output the physical feature vector V phy . Use a 1D convolutional layer (kernel size = 5, number of channels = 32) to extract local spectral features, and the self-attention mechanism focuses on the key wavelength regions. Input the polarization deflection angle (φ) and the absorption attenuation coefficient (α), and output the time series feature vector V pol .

[0173] (3) Dynamic weight fusion and prediction

[0174] Fusing V through a gating mechanism phy and V spec and V pol :

[0175] V = σ(W[V phy ; V spec ; V pol ).

[0176] The present invention also provides a computer-readable storage medium, in which at least one program code is stored, and the at least one program code is loaded and executed by a processor to implement the foregoing near-infrared polarization spectrum detection method based on multi-measurement point collaboration.

[0177] After considering the specification and practicing the content disclosed herein, those skilled in the art will readily think of other implementation manners of the present application. The present application is intended to cover any variations, uses or adaptations of the present application, and these variations, uses or adaptations follow the general principles of the present application and include the common general knowledge or conventional technical means in the technical field not disclosed in the present application. The specification and embodiments are only regarded as exemplary.

[0178] The above are only the preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A near-infrared polarization spectroscopy detection probe based on multi-point measurement collaboration, characterized in that: The probe is applicable to multi-point high-precision on-line detection of solid materials in a continuous production line. The probe includes: A spectral acquisition module, which includes a light source, a reference plate, an upper rotating device, and a collimating lens that are symmetrically and parallelly installed. The light source is fixed by a left light source clamping plate and a right light source clamping plate, and is arranged at a certain angle with the detection optical axis. The reference plate is connected to the upper rotating device and is driven by the upper rotating device to rotate below the collimating lens to collect background spectra; Among them, the probe integrates a multi-channel parallel calibration system. Each measurement point is independently configured with a reference channel. By alternately collecting the reference signals and material spectral signals of each measurement point, the baseline differences between the measurement points are eliminated based on a dynamic compensation model.

2. The near-infrared polarization spectroscopy detection probe based on multi-measurement point collaboration according to claim 1, wherein: The dynamic compensation model includes the following compensation terms: Particle size compensation term (ΔC g ), calculated based on the median particle size (D i ) of the material at the measurement point and the reference particle size (A); surface reflection compensation term (ΔC s ), calculated based on the surface roughness (R i ) of the material at the measurement point and the reference roughness (B); density attenuation compensation term (ΔC d ), calculated based on the bulk density (ρ i ) of the material at the measurement point and the reference density (C); the compensation term formulas are respectively: The calibration equation is: Note: I ref is the reference signal corresponding to the measurement point, I ref ∝I out ; I dark is the dark noise signal; I sample =I target +I out ; ΔC i includes ΔC g , ΔC d , ΔC s .

3. The near-infrared polarization spectroscopy detection probe based on multi-measurement point collaboration according to claim 1, wherein: The spectral acquisition module further includes a polarizer. The polarizer adopts an electric rotating polarizer, an integrated rotary encoder, or other real-time reading systems to real-time feedback the rotation angle θ of the polarizer.

4. The near-infrared polarization spectroscopy detection probe based on multi-measurement point collaboration according to claim 3, characterized in that: The transmission axis direction of the polarizer is orthogonal to the initial polarization state of the incident light, and is adjusted according to the polarization deflection angle of the material to achieve physical separation of the target absorption signal and the scattered light. The intensity of the target signal after passing through the polarizer is:

5. The near-infrared polarization spectroscopy detection probe based on multi-measurement point collaboration according to claim 1, characterized in that: The probe further includes: A parameter acquisition module, including: a particle size sensor, a roughness sensor, and a material bulk density sensor; the parameter acquisition module is used to real-time read relevant required parameters.

6. A near-infrared polarization spectroscopy detection method based on multi-measurement point collaboration, characterized in that: Using the probe according to any one of claims 1-5, the detection method includes the following steps: Uniformly irradiate the material through the symmetric halogen lamps of the multi-point probe, and alternately open the shutter to collect the reference signals and material spectral signals of each measurement point; Use a polarizer to separate the target absorption signal from the scattered light, with the transmission axis direction adjusted to suppress scattered light noise; Calculate the dynamic compensation terms based on the particle size, surface roughness, and density parameters of the material, and map the multi-point spectral data to a unified reference; Adopt a hybrid neural network model to fuse multi-source features, including spectral signals, polarization parameters, and physical property parameters, and optimize the model parameters through incremental learning.

7. The near-infrared polarization spectroscopy detection method based on multi-measurement point collaboration according to claim 6, characterized in that: Adopt a hybrid neural network model to fuse multi-source features, including spectral signals, polarization parameters, and physical property parameters, and optimize the model parameters through incremental learning, including: Input the physical property parameters (ΔC g , ΔC s , ΔC d ) in the compensation equation into the fully connected layer to generate the physical constraint vector V phy ; The convolutional layer (CNN) is used to extract the local features of the spectral signal, and the attention mechanism is adopted to capture the response relationship in the key wavelength region, and the feature vector V is output spec ; Model the temporal dependence of polarization parameters using a bidirectional long short-term memory network (Bi-LSTM) and output the feature vector V pol ; Dynamically fuse V through a gating mechanism phy , V spec , V pol , The formula is: V = σ(W[V phy ; V spec ; V pol ) Note: σ is the Sigmoid function, and W is the learnable parameter matrix.

8. The near-infrared polarization spectroscopy detection method based on multi-measurement point collaboration according to claim 6, wherein: The compensated spectral signals (I calibrated ) of each measurement point and the polarization characteristic parameters (polarization deflection angle absorption attenuation coefficient α) are multi-dimensionally fused; Using the polarization filter transmission intensity formula Invert the polarization response characteristics of the material to generate a polarization feature matrix P; Based on the physical property parameters of the measurement points (particle size D i , roughness R i , density ρ i ), calculate the dynamic weight coefficient ω i , and the formula is: Generate a global feature vector F by weighted fusion of the spectral data of each measurement point.

9. The near-infrared polarization spectroscopy detection method based on multi-measurement point collaboration according to claim 6, wherein: The incremental learning to optimize the model parameters includes: by real-time monitoring the prediction error distribution, adjusting the compensation term coefficients in the calibration equation, and using a sliding window to update the historical data to maintain the stability of the model.

10. A computer-readable storage medium, characterized in that: At least one program code is stored in the computer-readable storage medium, and the at least one program code is loaded and executed by a processor to implement the near-infrared polarization spectral detection method based on multi-point collaboration according to any one of claims 6 to 9.

Images