Method for accurately detecting density of recombined bamboo board

Abstract

The invention relates to the technical field of recombined bamboo board processing, in particular to a recombined bamboo board density accurate detection method. Comprising the following steps: S1, collecting a vibration signal and cooperatively suppressing in different frequency bands, and synchronously collecting a dual-band image in a stable window; s2, dynamically determining a spectrum identification wave band based on the slab temperature, and accurately segmenting the bamboo wood body through a multi-task network; s3, performing multi-scale transformation on the image, calculating texture entropy features and mapping the texture entropy features into texture density contribution values; s4, performing multi-factor coupling compensation on the segmented contour to obtain geometric density, and performing Bayesian fusion on the geometric density and the texture density contribution value to obtain optimal estimation density; and S5, performing double-track judgment in combination with the confidence coefficient of the density estimation and the stability of the production process. Through multi-physical field sensing, macro and micro information complementation and error model and probability fusion cooperation, high-precision and high-robustness online density detection and process monitoring are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of reconstituted bamboo board processing, and particularly relates to a reconstituted bamboo board density accurate detection method. BACKGROUND

[0002] As a structural bio-based composite material, the density uniformity of the reconstituted bamboo board is a core index for determining the reliability of its mechanical properties. At present, the online density detection technology based on machine vision has become a research hotspot. The conventional scheme usually adopts a convolutional neural network to segment the material profile, and then combines the pixel area and weight to calculate the density. However, the present inventors found in the in-depth industrial scene research and cross-disciplinary analysis that even if an advanced visual algorithm is adopted, the existing technical scheme still has several fundamental technical bottlenecks that have been long ignored in the field and have a profound impact: the vibration spectrum in the production line environment is complex, including low-frequency conveyor belt jitter, medium-frequency motor harmonic, and high-frequency structure resonance. Traditional rigid anti-vibration or single-frequency compensation cannot cover the full frequency band. In particular, when a specific vibration frequency is coupled with the camera exposure timing or the line array scanning frequency, resonance imaging stripes are caused, which are difficult to distinguish from the bamboo texture in the spatial domain, leading to systematic misjudgment of the segmentation network and directly distorting the area calculation result; the density information of the reconstituted bamboo board is not only embodied in the macro profile, but also encoded in the mesoscopic (bundled bamboo level) and microscopic (fiber level) texture features, such as the orientation consistency of bamboo bundles and the distribution uniformity of inter-bundle gaps. The existing method only uses a neural network "black box" to extract features for classification or segmentation, losing the physical interpretability of the texture and its quantitative correlation model with density and mechanical properties. This leads to the system being unable to identify hidden defective products with "qualified profile but internal texture disorder"; during the cooling process after hot pressing, the adhesive in the board blank undergoes complex phase changes (from glassy to high-elastic state, etc.), and its near-infrared spectrum absorption peak at different temperature points will dynamically shift. Using a fixed waveband (such as 1550nm) for adhesive layer identification will cause feature waveband mismatch when the board blank temperature changes, resulting in fluctuation of adhesive layer segmentation accuracy at different positions on the production line (corresponding to different cooling temperatures), introducing unstable detection errors; when multiple physical compensation factors are introduced at the same time, the error sources corrected by each factor are not completely independent in physics. Simply performing multiplication or addition coupling will ignore the interaction coupling effect between factors, which may cause over-compensation or under-compensation, even amplify errors under certain working conditions, and lack the support of a rigorous error propagation model. SUMMARY

[0003] The technical problem to be solved by the present application is to improve the reconstituted bamboo board density accurate detection method, to systematically solve the problems of wideband vibration interference, texture density correlation disappearance, dynamic spectrum mismatch, and compensation factor coupling error from the nature of multi-physical field coupling.

[0004] To solve the above technical problems, the technical scheme adopted by the present application is: The method for accurately detecting the density of reconstituted bamboo boards comprises the following steps: S1: Collect the broadband vibration signal during the conveying process of the slab, separate it into low-frequency, medium-frequency and high-frequency components through wavelet packet decomposition, and perform cooperative suppression on the vibration of each frequency band; within the stable time window after suppression, synchronously trigger the double-waveband camera to collect the visible light image and short-wave infrared image of the slab; S2: Real-time measure the surface temperature of the slab, dynamically determine the optimal identification waveband according to the pre-established adhesive characteristic absorption peak position-temperature relationship; extract the dynamic waveband image from the short-wave infrared image, and input it and the visible light image into a multi-task learning network to output an accurate bamboo body mask; S3: Within the bamboo body mask, perform multi-scale and multi-directional transform decomposition on the visible light image; calculate the direction entropy and energy entropy of the mesoscale subband, and the complexity entropy of the microscale subband to form a texture entropy feature vector; input the feature vector into a pre-trained density regression model to obtain a texture density contribution value; S4: Calculate the basic projection area based on the bamboo body mask and the camera calibration parameters; combine the thermal deformation factor, the glue layer correction factor and the vibration residual error factor, and compensate the basic projection area through a pre-calibrated multi-factor coupling compensation model to obtain a compensated area and calculate the geometric density; based on Bayes' theorem, fuse the geometric density and the texture density contribution value to obtain an optimal estimated density; S5: Calculate the confidence of the current density estimation, and evaluate the process stability of the current production batch; combine the optimal estimated density, the confidence and the process stability for comprehensive judgment to output a qualified, suspicious or unqualified signal.

[0005] Further, in the above method for accurately detecting the density of reconstituted bamboo boards, in S1, the cooperative suppression specifically comprises: For the low-frequency component, dynamically adjust the rotating speed of the conveying roller through a model predictive control algorithm to cancel the phase; For the medium-frequency component, adjust the input current of the magneto-rheological vibration isolator to change the inherent frequency of the imaging system to avoid the peak frequency of the vibration energy; For the high-frequency component, use a numerical control delay phase-locked loop to real-time fine-tune the pixel clock trigger timing of the line array camera according to the vibration phase to realize scanning synchronization and vibration resistance.

[0006] Further, in the above method for accurately detecting the density of reconstituted bamboo boards, in S1, the double-waveband camera is a snapshot area array camera using a global shutter with an exposure time less than 200 microseconds; the determination condition of the stable time window is that the root mean square values of the vibration speeds of each frequency band are simultaneously lower than a preset threshold and last for more than 2 milliseconds.

[0007] Further, in the above-mentioned method for accurately detecting the density of reconstituted bamboo board, in S2, the multi-task learning network comprises a shared feature extraction backbone, a segmentation task head and a regression task head; the segmentation task head is used to output a bamboo body probability map, and the regression task head is used to output a glue layer phase change state index; the two task heads share features through a channel attention mechanism, and the glue layer phase change state index is used to assist in improving the segmentation accuracy in the training stage.

[0008] Further, in the above-mentioned method for accurately detecting the density of reconstituted bamboo board, in S3, the multi-scale multi-direction transform is decomposed into a non-subsampled contourlet transform; the mesoscale corresponds to a subband with a physical size of 2-10 mm, and the microscale corresponds to a subband with a physical size of less than 2 mm; the density regression model is a machine learning model based on gradient boosting tree or support vector regression, and the real density value and the corresponding texture entropy feature vector obtained by the micro-loss sampling method are used for training.

[0009] Further, in the above-mentioned method for accurately detecting the density of reconstituted bamboo board, in S4, the multi-factor coupling compensation model is calibrated by a central composite design response surface method; the thermal deformation factor is calculated based on the slab temperature, the glue layer correction factor is determined based on the proportion of the glue layer pixels in the bamboo body mask, and the vibration residual error factor is determined based on the root mean square value of the suppressed vibration in S1; the compensation model is a mathematical expression containing second-order interaction terms between the factors.

[0010] Further, in the above-mentioned method for accurately detecting the density of reconstituted bamboo board, in S4, the specific process of Bayesian fusion is as follows: the prior distribution of the real density is assumed, and the likelihood distributions of the geometric density observation value and the texture density contribution value observation value are both normal distributions; based on the prior distribution and the likelihood distribution, the mean and variance of the posterior normal distribution are calculated, the mean is taken as the optimal estimated density, and the reciprocal of the variance is normalized and used to calculate the confidence.

[0011] Further, in the above-mentioned method for accurately detecting the density of reconstituted bamboo board, in S5, the evaluation method of process stability is as follows: in a sliding window containing the latest 30 to 100 samples, the exponential weighted moving average and the exponential weighted moving standard deviation of the optimal estimated density are calculated; when the moving average falls within a preset range of the target center value and the moving standard deviation is lower than a set proportion of the historical long-term standard deviation, the process is determined to be stable.

[0012] Further, in the above-mentioned method for accurately detecting the density of reconstituted bamboo board, in S5, the logic of comprehensive judgment is as follows: If the optimal estimated density is within a preset qualified range, the confidence is higher than a confidence threshold, and the process stability shows stability, it is determined to be qualified; If the optimal estimation density is within the preset qualified range, but the confidence is lower than the confidence threshold or the process stability shows an abnormality, it is determined as suspicious and an alarm is triggered; If the optimal estimation density is not within the preset qualified range, it is determined as unqualified.

[0013] The beneficial effects of the present application are that: S1's frequency division domain active vibration suppression eliminates the interference of wideband vibration on imaging quality at the physical source through multi-modal actuators (model predictive control, magneto-rheological vibration isolation, time sequence synchronization), providing a geometrically accurate "steady state" image basis for subsequent analysis. This "steady state" basis is combined with S2's adaptive spectral adhesive layer identification based on temperature feedback, which dynamically adjusts the optical perception band on the stable image to ensure the "time-invariant" accuracy of material boundary segmentation under different working conditions. The two work together in the "time-space-spectrum" domain, realizing the leap from "passive collection of noisy images" to "active acquisition of high-quality feature images", which is the prerequisite for the effectiveness of all subsequent advanced analysis.

[0014] S3's cross-scale texture entropy analysis first extracts entropy features (directional entropy, energy entropy, complexity entropy) directly related to bamboo bundle arrangement and fiber distribution uniformity from the image, which have clear physical meaning and are mapped as "texture density contribution values". This microscopic information reflecting the uniformity of the internal structure of the material forms a strong complementary relationship with the "geometric density" calculated based on the contour in S4, which is a macroscopic information reflecting the overall size. The Bayesian fusion model in S4 is not a simple average, but an optimal estimation based on the uncertainty of the two, achieving information fusion so that the final density estimation has both macroscopic accuracy and microscopic sensitivity to internal hidden defects.

[0015] In view of the complex situation where multiple physical error sources (heat, glue, vibration) coexist and interact, the multi-factor coupling compensation model (established by response surface method) in S4 first describes and corrects the interaction effect between each compensation factor, overcoming the coupling error of traditional linear or multiplication models. On this basis, the Bayesian probability fusion framework further unifies the outputs of the compensated geometric measurement model and the texture statistical model. The two models work together at two levels of "deterministic error correction" and "uncertainty information fusion", improving the error control of the entire system from "empirical formula combination" to the scientific height of "systematic error modeling and uncertainty management".

[0016] The double-track judgment mechanism of S5 creates a new paradigm for quality control. It not only relies on whether the single-point density value exceeds the threshold, but also innovatively introduces the "confidence" of this detection (derived from Bayesian posterior uncertainty) and the "process stability" of the production batch (based on sliding window statistics) as a double decision aid. When the single-point measurement value is on the edge of qualification but the confidence is low, the system avoids misjudgment and marks it as "suspicious"; when the process stability index shows an abnormal trend, the system gives an early warning. This "point-surface combined" decision-making cooperation upgrades the system function from a single "qualified product filter" to an intelligent decision center with "quality evaluation", "process monitoring" and "risk warning", greatly improving the engineering practical value and reliability of the detection system. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 A flowchart of a density accurate detection method of a reed bamboo board according to an embodiment of the present application. DETAILED DESCRIPTION

[0018] To explain the technical content, purposes and effects of the present application in detail, the following will be described in conjunction with the embodiments and the accompanying drawings.

[0019] Please refer to Figure 1 The present embodiment takes a production line for producing reed bamboo boards for high-strength outdoor structures as the application scenario. The present embodiment relates to a density accurate detection method of a reed bamboo board, based on a density accurate detection system of a reed bamboo board, and specifically as follows: Vibration suppression and imaging module: at the detection station, install a German Polytec LV-S01 laser Doppler vibration meter (corresponding to "vibration signal acquisition"); the imaging unit (camera and light source) is installed on a US Lord RD-8041 variable stiffness magneto-rheological vibration isolation platform (corresponding to "variable stiffness vibration isolation platform"); a Canadian FLIR ORX-10G-123S5 dual-band snapshot area array camera (corresponding to "dual-band snapshot area array camera") is used, which is equipped with a light splitting prism and can synchronously acquire visible light (VIS) channel (1280x1024) and short-wave infrared (SWIR, 1300-1700nm) channel (640x512) images; the camera exposure time is set to 100μs, meeting the "less than 200 microseconds" short exposure requirement of claim 3; Temperature sensing module: a US FLIR A615 infrared thermal imager is used to monitor the surface temperature field of the board in real time.

[0020] Weight and thickness measurement module: a Mettler-Toledo IND570 online scale (accuracy ±10g) and a Keyence LK-H050 laser ranging sensor (accuracy ±0.01mm) are used.

[0021] Data processing and calculation module: the core is an industrial computer equipped with NVIDIA RTX A5000 GPU, on which a complete set of software algorithms for implementing the methods of claims 1-9 are deployed.

[0022] Determination and output module: including PLC controller, sound and light alarm and interface for communication with production line MES system.

[0023] Before formal detection, system calibration needs to be completed, including: high-precision intrinsic and extrinsic parameter calibration of camera (Zhang Zhengyou method) for area calculation; establishing adhesive characteristic absorption peak-temperature lookup table (for S2) through a large number of samples; training multi-task learning network and texture density regression model (for S2, S3); and calibrating multi-factor coupling compensation model (for S4) through central composite design experiment.

[0024] When a piece of reconstituted bamboo blank with specifications of 2440 mm x 1220 mm x 20 mm enters the detection station, the system works according to the following process: S1: collect broadband vibration signals in the process of blank conveying, separate them into low, medium and high frequency components through wavelet packet decomposition, and perform cooperative suppression on the vibration of each frequency band; in the stable time window after suppression, trigger the dual-waveband camera to collect the visible light image and short-wave infrared image of the blank synchronously; Specifically, use a non-contact laser Doppler vibration meter (such as Polytec LV-S01) to collect the vibration velocity signal v(t) of the blank surface, with a sampling frequency fs=20 kHz; pre-process v(t): use a 5th order Butterworth band-pass filter (passband 1 Hz - 1 kHz) to filter out direct current and high-frequency noise; perform 5-layer wavelet packet decomposition: wavelet base: 'db4' (Daubechies 4); decomposition obtains 32 frequency bands (nodes (5,0) to (5,31)); According to prior knowledge, divide: low frequency band LF: 0 - 12.5 Hz, corresponding to nodes (5,0)-(5,3); medium frequency band MF: 12.5 - 250 Hz, corresponding to nodes (5,4)-(5,15); high frequency band HF: 250 - 1000 Hz, corresponding to nodes (5,16)-(5,31); reconstruct the time domain signals of each frequency band: v_LF(t), v_MF(t), v_HF(t).

[0025] Low frequency suppression (model predictive control, MPC): control target, make the roller produce displacement opposite to v_LF(t); MPC parameters: prediction model: G(s) = 0.8 / (s^2 + 8.8s + 157.9) (approximate transfer function); Prediction horizon Np = 20, control horizon Nc = 5, sampling period Ts = 10 ms; cost function: J = Σ (y(k+i) - r(k+i))^T Q (y(k+i) - r(k+i)) + Σ Δu(k+j)^T R Δu(k+j), where Q=1.0, R=0.1; output: real-time velocity compensation command Δu(t) for servo roller; Medium-frequency suppression (magneto-rheological variable-stiffness damping): calculate power spectral density (PSD) of v_MF(t), identify energy peak frequency f_peak; query current-natural frequency mapping table: f_n(I) = 140 - 40 I (Hz / A, 0≤I≤1A); control law: calculate current command I_cmd that makes f_n away from f_peak. For example, if f_peak = 125 Hz, set f_target = 100 Hz, calculate I_cmd = (140 - 100) / 40 = 1.0 A; High-frequency suppression (pixel clock synchronous anti-vibration): zero-crossing detection on v_HF(t); dynamically adjust camera pixel clock trigger edge through digital-controlled delay-locked loop (DLL); DLL parameters: phase detector: XOR gate (XOR); loop filter: first-order low-pass, fc = 2 kHz; voltage-controlled delay line (VCDL) resolution: 0.1 ns; control target: align each row exposure start time with zero-crossing point of v_HF(t) (instantaneous velocity is zero); Stable window judgment and dual-band imaging; judgment conditions: calculate root mean square (RMS) values of vibrations in each frequency band after suppression, when the following conditions are met simultaneously and last for t_stable ≥ 5 ms, it is judged to enter the stable window: RMS(v_LF)<5 μm / s; RMS(v_MF)<20 μm / s; RMS(v_HF)<50 μm / s; imaging: in the stable window, trigger snapshot area array camera (such as FLIR ORX-10G-123S5) for global shutter exposure, exposure time t_exp = 100 μs; synchronous acquisition: visible light image I_vis (resolution 1280×1024); short-wave infrared image I_swir (resolution 640×512, waveband 1300-1700 nm); Specifically, after the slab triggers the photoelectric sensor, the system starts; Laser Doppler Vibrometer (LDV) collects 100 ms of slab surface vibration velocity signal v(t); adopts 'db4' wavelet basis for 5-layer wavelet packet decomposition, and divides it into low frequency band LF (0-12.5 Hz), medium frequency band MF (12.5-250 Hz), and high frequency band HF (250-1000 Hz) according to energy; Identify the main frequency (e.g. 2 Hz) and phase of the LF component. Through the Model Predictive Controller (MPC), calculate and instruct the downstream servo roller to perform a reverse speed fine-tuning within the next 200 ms to offset the expected fluctuations of the slab; Calculate the power spectral density of the MF component, and find that there is an energy peak at 125 Hz. The controller adjusts the current of the magnetorheological vibration isolator to switch its natural frequency from 140 Hz to 100 Hz, avoiding the resonance point at 125 Hz; High-frequency scanning synchronization: according to the real-time zero-crossing point of the HF component, dynamically adjust the phase of the camera line trigger signal through the digital delay lock loop (DLL), so that the exposure starting time of each line is aligned with the instantaneous speed zero point of the vibration; The system calculates the root mean square values of the vibrations in the three frequency bands in real time; when the state of RMS(LF)<5μm / s, RMS(MF)<20μm / s, and RMS(HF)<50μm / s lasts for 5 ms, it is determined that it has entered the "stable time window"; immediately trigger the dual-band camera to perform global shutter exposure and synchronously capture the visible light image I_vis and the short-wave infrared image I_swir; S2: Real-time measurement of slab surface temperature, dynamic determination of the optimal identification band according to the pre-established adhesive characteristic absorption peak-temperature relationship; extraction of the dynamic band image from the short-wave infrared image, and input of the visible light image and the dynamic band image into the multi-task learning network to output the accurate bamboo body mask; Specifically, use an infrared thermal imager (such as FLIR A615) to measure the average temperature T of the slab surface; through linear interpolation, calculate the optimal identification center wavelength λ_center(T) and bandwidth Δλ(T) at the current temperature T. For example, when T=58.7°C, λ_center ≈ 1528 nm, Δλ=30 nm; extract the data of the band [λ_center - Δλ / 2, λ_center + Δλ / 2] from I_swir, upsample to 1280×1024, and obtain the dynamic band image I_band(T); Multi-task learning network segmentation; Network input: concatenate I_vis (3 channels) and I_band(T) (1 channel) in the channel dimension to form a 4-channel input tensor; Network architecture: Main trunk: HRNet-W32; output multi-scale feature maps {F1, F2, F3, F4} with resolutions of 1 / 4, 1 / 8, 1 / 16, 1 / 32 of the input, respectively; Segmentation head (main task): Input: F4 (resolution H / 32 x W / 32 x 256); Operation: 1 x 1 Conv (256→32) → BatchNorm → ReLU → 1 x 1 Conv (32→1) → Sigmoid → bilinear up-sampling to the original image size; Output: bamboo body probability map P_wood (H x W x 1).

[0026] Regression head (auxiliary task): Input: F4; Operation: global average pooling (GAP) → Flatten → FC (256→128) → ReLU → Dropout (p=0.5) → FC (128→1); Output: adhesive layer curing degree scalar s_cure; Attention mechanism: an SENet module is connected after F4, with compression ratio r=16, to generate a channel weight vector w, which is used to weight the feature channels of F4 before inputting into the segmentation head; Network inference: Forward propagation to obtain P_wood and s_cure (the latter is only used for training assistance); Apply threshold th=0.5 to P_wood to obtain binary bamboo body mask M_wood; Post-processing: perform morphological closing operation (3 x 3 circular structure element) on M_wood to fill small holes; Specifically, the infrared thermal imager measures the average temperature T=58.7℃ of the slab detection area; query the pre-set lookup table to determine the current best recognition center wavelength λ_center=1528nm and bandwidth Δλ=30nm by interpolation; Extract [1513nm, 1543nm] band data from I_swir, up-sample and align with I_vis to generate a fused input image; Input the fused image into the multi-task learning network, which takes HRNet-W32 as the main trunk and contains 4 stages, outputting feature maps with resolutions of 1 / 4, 1 / 8, 1 / 16, 1 / 32 of the input, and channel numbers of 32, 64, 128, 256, respectively; through repeated multi-resolution feature fusion, high-resolution representation is maintained; after sharing the features, two heads are divided: Segmentation head: outputs a bamboo body probability map P_wood, which is thresholded (0.5) to obtain a binary mask M_wood; Regression head: outputs a scalar representing the adhesive curing degree index. In the training phase, the index is calculated with the loss of the adhesive curing degree true value measured in the laboratory; at the same time, through a channel attention module, the segmentation head can focus on the feature channels concerned by the regression head, so as to improve the segmentation accuracy in the inference phase even without regression true value (i.e. “auxiliary improvement of segmentation accuracy”); Morphological closing operation is performed on M_wood to obtain the final accurate bamboo area contour for calculation; S3: In the bamboo body mask, the visible light image is decomposed by multi-scale and multi-direction transform; the direction entropy and energy entropy of the mesoscale subband and the complexity entropy of the microscale subband are calculated to form a texture entropy feature vector; the feature vector is input into a pre-trained density regression model to obtain a texture density contribution value; Specifically, in the M_wood region, the gray-scale image of I_vis is subjected to non-subsample contourlet transform (NSCT); a ‘pyrexc’ filter set is used to decompose into 3 scales, 8 directions at the 1st and 2nd scales, and 16 directions at the 3rd scale; The texture entropy feature vector is calculated, including: Directional entropy H_dir (mesoscale, l=2): for each pixel p in the mask, the energy of the pixel p in the 8 direction subbands is calculated: E_p^k = |H_2^k(p)|^2, k=1,...,8; normalization: P_p^k = E_p^k / Σ_{j=1}^8 E_p^j; find the dominant direction k_p of pixel p: k_p = argmax_k(P_p^k); count the histogram of all pixel dominant directions k_p in the entire region, and calculate the Shannon entropy of the histogram: H_dir = - Σ_{i=1}^8 p_i log2(p_i), where p_i is the proportion of the i-th direction; Energy entropy H_energy (mesoscale, l=2): calculate the total energy of each direction subband: E_total^k =Σ_{p in mask} |H_2^k(p)|^2; normalization: q_k = E_total^k / Σ_{j=1}^8 E_total^j; calculate the Shannon entropy: H_energy = - Σ_{k=1}^8 q_k log2(q_k); Complexity entropy H complex (microscale, l = 3): select the 4 highest energy subbands in the 3rd scale {H 3 k1,..., H 3 k4}; for each selected subband, unfold the coefficient matrix into a one-dimensional sequence X = {x1, x2,..., xN} in raster scan order; calculate the fuzzy entropy (FuzzyEn) of the sequence: parameters: embedding dimension m = 2, tolerance r = 0.15 std(X), exponential n = 2; define fuzzy membership function: μ(d, r) = exp(-(d / r)^n); for i = 1 to N-m, construct vector: X_i^m = {x_i, x_{i+1},..., x_{i+m-1}}; calculate distance d_{ij} = max(|X_i^m - X_j^m|), j = 1 to N-m, j≠i; calculate similarity B_i^m(r) = (1 / (N-m-1)) Σ_{j≠i} μ(d_{ij},r); define Φ^m(r) = (1 / (N-m)) Σ_{i=1}^{N-m} B_i^m(r); calculate Φ^{m+1}(r) in the same way; fuzzy entropy FuzzyEn(m, r, N) = ln(Φ^m(r)) - ln(Φ^{m+1}(r)); H complex takes the average of the fuzzy entropies of the 4 subbands; Get feature vector F = [H_dir, H_energy, H_complex]; Input F into the pre-trained LightGBM regression model, and the model parameters are as follows: params = { 'boosting_type': 'gbdt', 'objective':'regression', 'metric': 'rmse', 'num_leaves': 31, 'learning_rate': 0.05, 'feature_fraction': 0.9, 'bagging_fraction': 0.8, 'bagging_freq': 5, 'min_data_in_leaf': 20, 'lambda_l1': 0.1, 'lambda_l2': 0.2 } The model is trained using 500 samples of density truth values obtained by micro-damage sampling method (drilling Φ8mm holes in non-critical positions of the slab, measuring the sample density as the local true value) and their F vectors. The model output is the texture density contribution value ρ_texture = 1.232 g / cm³; S4: Calculate the basic projection area based on the bamboo material body mask and camera calibration parameters; combine the thermal deformation factor, glue layer correction factor and vibration residual error factor, and compensate the basic projection area through the pre-calibrated multi-factor coupling compensation model to obtain the compensated area and calculate the geometric density; based on Bayes theorem, the geometric density and the texture density contribution value are fused to obtain the optimal estimated density; Specifically, based on the M_wood profile and camera calibration parameters, the orthographic projection area A_base = 2.970 m² is calculated; Thermal deformation factor η = 1 + α (T - T_ref); where, α = 2.5e-6 / °C (coefficient of linear expansion of bamboo), T_ref = 25°C; Glue layer correction factor γ = 1 - k_glue Ratio_glue; where, Ratio_glue is the proportion of glue layer pixels estimated by the network, k_glue = 0.05 (obtained by calibration); Vibration residual error factor ξ = 1 / (1 + β RMS_total). Where, RMS_total is the RMS value of the total vibration after suppression, β = 1e-4 (calibration coefficient); Model form: δ = β0 + β1 η + β2 γ + β3 ξ + β12 η γ + β13 η ξ + β23 γ ξ + β11 η² + β22 γ² + β33 ξ² β0=-0.0185, β1=0.0121, β2=0.0098, β3=0.0012, β12 = -0.0055, β13 = -0.0008, β23 = -0.0011, β11 = 0.0001, β22 = 0.0003, β33 = 0.00005; Calculate the predicted relative area error δ; Compensated area: A_comp = A_base (1 - δ); Specifically, according to the temperature T = 58.7℃, calculate the thermal deformation factor η = 1 + 2.5e-6 (58.7-25) ≈1.000084; According to the estimation of the proportion of adhesive layer pixels in the middle feature map of the segmentation network, the adhesive layer correction factor γ = 0.982 is obtained; According to the vibration RMS value after suppression, the vibration residual error factor ξ = 1.001 is calculated; Substitute (η, γ, ξ) into the second-order coupling compensation model calibrated by central composite design response surface method: A_comp = A_base K(η, γ, ξ); Where the compensation coefficient matrix K has been pre-fitted; Calculate A_comp =2.967 m²; The online scale obtains the slab weight m (kg), the laser range finder obtains the actual thickness of the slab h (m), and the geometric density: ρ_geo = m / (A_comp h) (g / cm³, note the unit conversion); Specifically, the online scale reading m = 72.56 kg, the laser thickness h = 20.02 mm. Calculate the geometric density: ρ_geo = m / (A_comp h) = 72.56 / (2.967 0.02002) ≈ 1.221 g / cm³; Calculate the posterior distribution, posterior mean (optimal estimated density): μ_post = (μ_prior / σ_prior² + ρ_geo / σ_geo² + ρ_texture / σ_texture²) / (1 / σ_prior² + 1 / σ_geo² + 1 / σ_texture²), Posterior variance: σ_post² = 1 / (1 / σ_prior² + 1 / σ_geo² + 1 / σ_texture²); Output the most estimated density: ρ_fused = μ_post; Estimate uncertainty: σ_post; Specifically, in Bayesian fusion, we assume the true density prior ρ_true ~ N(μ=1.250, σ_prior=0.015); let the geometric density observation likelihood ρ_geo ~ N(ρ_true, σ_geo=0.010), and the texture density observation likelihood ρ_texture ~ N(ρ_true, σ_texture=0.015) (σ value is based on historical error statistics); according to Bayes' theorem, we calculate the posterior distribution N(μ_post, σ_post); substituting ρ_geo=1.221 and ρ_texture=1.232, we obtain: μ_post ≈ 1.226 g / cm³ (this is the optimal estimated density ρ_fused); σ_post ≈ 0.008 g / cm³; S5: Calculate the confidence level of this density estimate and evaluate the process stability of the current production batch; combine the optimal estimated density, the confidence level, and the process stability to make a comprehensive judgment and output a qualified, doubtful, or unqualified signal; Specifically, the precision of this density estimation is calculated as: precision_post = 1 / σ_post²; the theoretical maximum precision (under ideal observation) is calculated as: precision_max = 1 / (1 / σ_prior² + 1 / σ_geo_min² + 1 / σ_texture_min²), where σ_geo_min = 0.008, σ_texture_min = 0.012; and the confidence threshold C_th is set to 0.80. Maintain a sliding window containing the most recent N=50 ρ_fused values; calculate the exponentially weighted moving average (EWMA) and moving standard deviation (EWMSD): EWMA_t = λ ρ_fused_t + (1-λ) EWMA_{t-1}, where λ=0.2; EWMSD_t= sqrt(λ (ρ_fused_t - EWMA_t)^2 + (1-λ) EWMSD_{t-1}^2); Control limit: UCL / LCL = μ_target ± 3 σ_long sqrt(λ / (2-λ)); Where μ_target=1.250, σ_long=0.010 (long-term process standard deviation); If three consecutive EWMA points fall within [LCL, UCL], and EWMSD < 0.8 σ_long, then the process is determined to be stable (S = 1); otherwise, the process is determined to be unstable (S = 0); Dual-track judgment logic: input: ρ_fused, C, S, pass range [ρ_low = 1.22, ρ_high = 1.28]; determination rule: IF ρ_low ≤ ρ_fused ≤ ρ_high AND C ≥ C_th AND S == 1 THEN output “pass”; ELSE IF ρ_low ≤ ρ_fused ≤ ρ_high AND (C < C_th OR S == 0) THEN output “suspect - to be checked”, trigger alarm; ELSE output “fail”; Specifically, the posterior precision is normalized, C = (1 / σ_post^2) / (1 / σ_post^2)_max ≈ 0.85; The system maintains a sliding window of ρ_fused of the last 50 plates; calculates its exponential weighted moving average (EWMA) and moving standard deviation (EWMSD); the current EWMA = 1.238 (within the target center 1.250 ± 0.01), EWMSD = 0.006 (less than 0.8 times the long-term standard deviation 0.010), so the process stability S = 1 (stable); Satisfy all “pass” conditions, the system outputs “pass” signal to MES through the judgment module, turns on the green indicator light, and archives the data; If the detection ρ_fused = 1.225 g / cm³ (pass) at this time, but C = 0.78 < C_th due to image blur, the system determines it as “suspect”, triggers the yellow alarm light, and marks the plate ID for manual review in MES, effectively preventing misjudgment.

[0027] Experiment 1: Overall precision and anti-interference ability experiment, test the anti-interference ability of the above embodiment detection method under active vibration suppression; Near the detection station, artificially introduce periodic interference sources: ① low frequency (2Hz) simulated roll cylinder out of round; ② medium frequency (125Hz) pneumatic knocking simulating mechanical impact; ③ random high frequency vibration; Randomly select 120 plates from the production line, take them off immediately after detection, and measure their absolute true density in a constant temperature laboratory (23±2℃) using the drainage method (Volumetric Method) in ASTM D2395-14 “Standard Test Methods for Density of Wood” as the “gold standard” value ρ_true; After each slab is detected by the detection method of the above embodiment, the detection density ρ_det of each slab is recorded respectively; The absolute error | ρ_det-ρ_true | of each slab under each set of systems is calculated; The mean absolute error (MAE) is 0.0041 g / cm³, the root mean square error (RMSE) is 0.0053 g / cm³, and the maximum absolute error is 0.0102 g / cm³.

[0028] Experiment two: glue layer segmentation robustness and hidden defect detection experiment, verify the stability of adaptive spectral glue layer identification at different temperatures, and the detection ability of cross-scale texture entropy analysis to internal hidden defects; 60 slabs are intercepted from the production line, and they are respectively stabilized at five temperature points (12 pieces each) of 40℃, 50℃, 60℃, 70℃ and 80℃ in the laboratory temperature box, and then quickly moved to the detection station for image acquisition and segmentation. Cooperate with the process personnel, prepare 30 known hidden defect boards with "flat surface, qualified profile but internal texture disorder area" by deliberately adjusting the group slab process (such as local uneven laying, adding a small amount of abnormal bamboo bundle with water content), at the same time, randomly select 70 normal boards as the control.

[0029] For temperature adaptability samples, three experienced quality inspectors manually label the collected images at the pixel level as fine, as segmentation ground truth, and calculate the intersection over union (IoU) between the bamboo mask output by the network at different temperatures and the manually labeled images; 100 slabs (30 defects + 70 normal) are detected by the detection scheme of the above embodiment; the combined results of subsequent ultrasonic propagation speed detection and destructive section observation in the laboratory are used as the final basis for defect existence (ground truth), and the detection rate and false positive rate of the above embodiment method are calculated; In the experimental results, the glue layer segmentation IoU (mean ± standard deviation) is 0.923 ± 0.012; the IoU value at 40℃ is 0.918, and the IoU value at 80℃ is 0.927; the hidden defect detection rate is 30 / 30 (100%); the hidden defect false positive rate is 2 / 70 (2.9%).

[0030] Experiment three: error compensation and information fusion model effectiveness experiment, quantitatively evaluate the contribution of multi-factor coupled compensation model and Bayesian fusion respectively, select 50 slabs covering high, medium and low density range, use a high-precision three-dimensional scanner to obtain their real three-dimensional point cloud model, and calculate the "real surface area" A—_true_3D as the geometric reference of this experiment; at the same time, the real density ρ_true is measured by the drainage method; The relative error mean of the coupling compensation area A_comp in the above embodiment is -0.28%, and the absolute error mean of the density p_fused of the Bayesian fusion is 0.0045 g / cm3.

[0031] The above only describes the embodiments of the present application, and does not limit the patent scope of the present application, and any equivalent transformation or direct or indirect application in the related technical field by using the content of the specification and drawings of the present application is also included in the patent protection scope of the present application.

Claims

1. A method for accurately detecting the density of reconstituted bamboo boards, characterized in that, Includes the following steps: S1: Acquire broadband vibration signals during slab transport, separate them into low-frequency, mid-frequency and high-frequency components through wavelet packet decomposition, and perform coordinated suppression on vibrations in each frequency band; within the stabilization time window after suppression, simultaneously trigger a dual-band camera to acquire visible light and short-wave infrared images of the slab; S2: Real-time measurement of slab surface temperature, dynamic determination of the optimal identification band based on the pre-established characteristic absorption peak position-temperature relationship of adhesive; extraction of dynamic band images from shortwave infrared images, and input of them together with visible light images into a multi-task learning network to output an accurate bamboo body mask. S3: Within the bamboo body mask, the visible light image is decomposed into multi-scale and multi-directional transformations; the directional entropy and energy entropy of the mesoscale subband and the complexity entropy of the microscale subband are calculated to form a texture entropy feature vector; the feature vector is input into a pre-trained density regression model to obtain the texture density contribution value. S4: Calculate the base projection area based on the bamboo body mask and camera calibration parameters; combine the thermal deformation factor, adhesive layer correction factor and vibration residual error factor, and compensate the base projection area through a pre-calibrated multi-factor coupling compensation model to obtain the compensated area and calculate the geometric density; based on Bayes' theorem, fuse the geometric density with the texture density contribution value to obtain the optimal estimated density; S5: Calculate the confidence level of this density estimate and evaluate the process stability of the current production batch; combine the optimal estimated density, the confidence level, and the process stability to make a comprehensive judgment and output a qualified, doubtful, or unqualified signal.

2. The method for accurately detecting the density of reconstituted bamboo board according to claim 1, characterized in that, In S1, the cooperative inhibition specifically includes: For the low-frequency components, the phase cancellation is achieved by dynamically adjusting the speed of the conveyor rollers using a model predictive control algorithm; For the mid-frequency component, the natural frequency of the imaging system is changed by adjusting the input current of the magnetorheological vibration isolator to avoid the peak frequency of vibration energy; For the high-frequency components, the pixel clock triggering timing of the linear array camera is finely adjusted in real time according to the vibration phase by a numerically controlled delay phase-locked loop to achieve scanning synchronization and vibration resistance.

3. The method for accurately detecting the density of reconstituted bamboo board according to claim 1 or 2, characterized in that, In S1, the dual-band camera is a snapshot-type area array camera, which uses a global shutter and has an exposure time of less than 200 microseconds; the determination condition for the stable time window is that the root mean square value of the vibration velocity of each frequency band is simultaneously lower than a preset threshold and lasts for more than 2 milliseconds.

4. The method for accurately detecting the density of reconstituted bamboo board according to claim 1, characterized in that, In S2, the multi-task learning network includes a shared feature extraction backbone, a segmentation task head, and a regression task head; the segmentation task head is used to output a probability map of the bamboo body, and the regression task head is used to output the adhesive layer phase transition state index; the two task heads share features through a channel attention mechanism, and the adhesive layer phase transition state index is used to help improve segmentation accuracy during the training phase.

5. The method for accurately detecting the density of reconstituted bamboo board according to claim 1, characterized in that, In S3, the multi-scale multi-directional transformation is decomposed into a non-downsampled contour wave transformation; the mesoscale corresponds to a sub-band with a physical size of 2-10 mm, and the microscale corresponds to a sub-band with a physical size of less than 2 mm; the density regression model is a machine learning model based on gradient boosting tree or support vector regression, and is trained using the true density value obtained by the micro-loss sampling method and its corresponding texture entropy feature vector.

6. The method for accurately detecting the density of reconstituted bamboo board according to claim 1, characterized in that, In S4, the multi-factor coupled compensation model is calibrated using the central composite design response surface method; the heat deformation factor is calculated based on the slab temperature, the adhesive layer correction factor is determined based on the proportion of adhesive layer pixels in the bamboo body mask, and the vibration residual error factor is determined based on the root mean square value of the suppressed vibration in S1; the compensation model is a mathematical expression containing second-order interaction terms between the factors.

7. The method for accurately detecting the density of reconstituted bamboo board according to claim 1, characterized in that, In S4, the specific process of the Bayesian fusion is as follows: assuming that the prior distribution of the true density, as well as the likelihood distribution of the geometric density observation and the texture density contribution observation, are all normal distributions; based on the prior distribution and the likelihood distribution, the mean and variance of the posterior normal distribution are calculated, the mean is used as the optimal estimated density, and the inverse of the variance is normalized and used to calculate the confidence level.

8. The method for accurately detecting the density of reconstituted bamboo board according to claim 1, characterized in that, In S5, the method for evaluating the stability of the process is as follows: within a sliding window containing the most recent 30 to 100 samples, calculate the exponentially weighted moving average and the exponentially weighted moving standard deviation of the optimal estimated density; when the moving average falls within a preset range of the target center value and the moving standard deviation is lower than a set proportion of the historical long-term standard deviation, the process is determined to be stable.

9. The method for accurately detecting the density of reconstituted bamboo board according to claim 1, characterized in that, In S5, the logic for comprehensive judgment is as follows: If the optimal estimated density is within the preset acceptable range, the confidence level is higher than the confidence threshold, and the process stability shows stability, then it is determined to be acceptable. If the optimal estimated density is within the preset acceptable range, but the confidence level is lower than the confidence threshold or the process stability shows abnormality, it is judged as suspicious and an alarm is triggered. If the optimal estimated density is not within the preset acceptable range, it is determined to be unacceptable.

Images