A method for measuring the thickness of cigarette glue based on deep learning neural network

Abstract

The present application relates to the technical field of nondestructive testing, and particularly relates to a cigarette adhesive thickness measurement method based on a deep learning neural network, which comprises the following steps: acquiring a terahertz intensity image by using a point frequency source terahertz imaging system; screening a to-be-measured region; detecting the to-be-measured region by using a thickness measurement instrument to obtain a real thickness; detecting the to-be-measured region by using a terahertz time-domain spectroscopy system and performing preprocessing to obtain a corresponding terahertz time-domain waveform; establishing a cigarette adhesive data set; constructing a hybrid network, training the hybrid network by using the cigarette adhesive data set, and obtaining a trained hybrid network; for a to-be-detected cigarette adhesive sample, acquiring a to-be-measured region in a terahertz intensity image by using the point frequency source terahertz imaging system; and inputting the trained hybrid network to obtain a thickness prediction value of the cigarette adhesive, so that the present application can detect the real thickness of the adhesive on a rough cigarette paper substrate.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology, specifically to a terahertz thickness measurement method for cigarette adhesive based on deep learning neural networks. Background Technology

[0002] In cigarette manufacturing, the attachment of filter rods to cigarette paper requires bonding with a specific tobacco adhesive. This adhesive is typically applied in extremely fine lines, and its coating quality, especially its thickness and uniformity, are key process parameters affecting the air permeability, burning rate, draw resistance stability, and final sensory quality of cigarette products. Therefore, achieving accurate, rapid, and non-destructive testing of the tobacco adhesive coating thickness is crucial for ensuring product quality and optimizing production processes.

[0003] Currently, the industry primarily relies on non-contact optical methods, such as the laser triangulation method, to measure coating thickness. This method involves emitting a laser beam onto the surface of the object being measured and receiving the reflected light spot, then using triangulation to calculate the height information. However, when applied to measuring cigarette adhesives, this method has inherent limitations: First, it measures the physical height of the adhesive line relative to the cigarette paper substrate, rather than the actual physical thickness of the adhesive material itself. For transparent or semi-transparent adhesives, the measurement value is easily affected by the substrate's color, texture, and optical properties. Second, the laser spot size is relatively large compared to the adhesive line width, and edge effects can affect measurement accuracy. Furthermore, this method struggles to handle noise caused by the porosity and surface roughness of the cigarette paper itself.

[0004] Terahertz time-domain spectroscopy is an emerging non-destructive detection technique. Terahertz waves, located between microwaves and infrared light, possess excellent penetrating power through many non-polar dielectric materials, such as paper, plastics, and adhesives. More importantly, a terahertz pulse signal, when passing through or reflected from a multilayer structure, generates reflected sub-pulses at the interfaces of different materials. By analyzing the time intervals between these sub-pulses and combining this with the refractive index of the materials, the thickness of each layer can be calculated, with theoretical measurement accuracy reaching micrometers or even submicrometers. This makes it demonstrates great potential for thickness measurement of thin materials such as coatings and films.

[0005] Chinese invention patent application CN110186930A discloses a method and apparatus for measuring coating thickness based on terahertz technology. This method extracts multiple echoes from the terahertz reflected signal and calculates the time difference to invert the coating thickness. However, this traditional terahertz thickness measurement method, based on a physical model and manual peak extraction, suffers from signal overlap problems, substrate interference, and noise sensitivity, making it unsuitable for achieving the desired high accuracy and reliability in industrial testing scenarios involving tobacco adhesives. Summary of the Invention

[0006] In view of the above problems, the present invention provides a terahertz thickness measurement method for cigarette adhesive based on deep learning neural network, which solves the technical problems that the existing technology cannot accurately measure the true thickness of ultra-thin cigarette adhesive on rough cigarette paper substrate, and that the traditional terahertz inversion algorithm has poor robustness and low measurement accuracy under strong noise and signal overlap.

[0007] This invention provides a terahertz thickness measurement method for cigarette adhesive based on deep learning neural networks, comprising the following steps: Step S1: Use a point frequency source terahertz imaging system to detect multiple tobacco adhesive samples coated on standard cigarette paper and obtain the corresponding terahertz intensity images; then screen out the areas to be tested in the terahertz intensity images. Step S2: Use a thickness measuring instrument to detect the area to be measured and obtain the true thickness; The area to be tested is detected and preprocessed using a terahertz time-domain spectroscopy system to obtain the corresponding terahertz time-domain waveform; the area to be tested and the terahertz time-domain waveform in the actual thickness and terahertz intensity images are used as samples to establish a tobacco adhesive dataset from multiple samples; Step S3: Construct a hybrid network, which is used to receive the region to be measured and the terahertz time-domain waveform in the terahertz intensity image and output the predicted value of the thickness of the tobacco adhesive. Step S4: Train the hybrid network using the tobacco glue dataset to obtain the trained hybrid network; Step S5: For the tobacco adhesive sample to be tested, the test area in the terahertz intensity image is obtained using a point frequency source terahertz imaging system. Terahertz time-domain waveforms are obtained by using a terahertz time-domain spectroscopy system; the region to be measured in the terahertz intensity image and the terahertz time-domain waveform are input into a trained hybrid network to obtain the predicted value of the thickness of the tobacco adhesive.

[0008] Preferably, step S1 specifically includes: Step S1-1: Obtain multiple tobacco adhesive samples coated on standard cigarette paper, wherein the adhesive thickness of the tobacco adhesive samples ranges from 1 to 1000 μm; Step S1-2: The point frequency source terahertz imaging system is constructed by using a continuous wave terahertz source with a working frequency of 340 GHz and the corresponding focal plane detector array or two-dimensional scanning platform. Steps S1-3: Place multiple tobacco adhesive samples on the imaging plane of the point frequency source terahertz imaging system and scan to obtain the terahertz intensity image of each scanning point at a frequency of 340 GHz. Steps S1-4: Select regions with uniform thickness or varying thickness gradients from the terahertz intensity image as the regions to be tested.

[0009] Preferably, in step S2, the thickness measuring instrument is a white light interferometer or a step gauge; The steps of detecting and preprocessing the region under test using a terahertz time-domain spectroscopy system to obtain the corresponding terahertz time-domain waveform specifically include: The sample of tobacco adhesive is placed in the transmission or reflection measurement optical path, and point measurements are performed in the area to be measured. The original terahertz time-domain waveform signal corresponding to each measurement point is collected. For the original terahertz time-domain waveform signal, a time window is extracted, background signals with system noise are removed, and amplitude normalization is performed to form one-dimensional time-series data, which serves as the terahertz time-domain waveform.

[0010] Preferably, in step S3, the input to the hybrid network is the fusion information of the pixel intensity information of the region to be measured in the terahertz intensity image and the terahertz time-domain waveform.

[0011] Preferably, the hybrid network in step S3 includes a CNN convolutional module, a Transformer encoder module, and a feature coupling unit; The CNN convolutional module receives the input signal and processes it through multiple convolutional sub-modules to obtain the output features of the CNN convolutional module. The convolutional sub-module includes a convolutional layer, an activation function, and a pooling layer connected in sequence. The output features of the CNN convolutional module are positionally encoded and then input into the Transformer encoder module. After processing by the self-attention mechanism, the output features of the Transformer are obtained. The output features of the CNN convolutional module and the output features of the Transformer are fused by the feature coupling unit, and then output as a predicted value of the thickness of the tobacco glue in micrometers through a fully connected layer.

[0012] Preferably, step S4 specifically includes: The tobacco adhesive dataset is divided into training and testing sets. The pixel intensity information of the test area in the preprocessed terahertz intensity image in the training set and the fusion information of the terahertz time-domain waveform are used as input and fed into the hybrid network for forward propagation calculation, outputting the corresponding predicted thickness value of tobacco adhesive. The mean square error loss function between the predicted thickness value and the actual thickness of tobacco adhesive is calculated. The gradient of the mean square error loss function with respect to the network parameters of each layer of the hybrid network is calculated through the backpropagation algorithm. The network parameters are iteratively updated using the Adam adaptive optimizer. Training stops when the mean square error loss function is minimized or when the preset number of iterations is reached, resulting in the trained hybrid network.

[0013] Preferably, the point frequency source terahertz imaging system and the terahertz time-domain spectroscopy system are integrated in hardware to form an integrated two-stage detection device.

[0014] Preferably, step S4 further includes: The trained hybrid network is subjected to lightweight update processing, including model pruning, weight quantization, and knowledge distillation.

[0015] Compared with the prior art, the present invention has at least the following beneficial effects: (1) High precision and ultra-thin measurement capability: Convolutional neural networks can learn and analyze weak echo signals that are severely overlapping in the time domain, breaking through the limitations of traditional algorithms on echo time intervals, and realizing accurate measurement of the thickness of ultra-thin glue layers at the micrometer level.

[0016] (2) Strong robustness: Deep learning models have a natural ability to suppress the roughness and non-uniformity of cigarette paper substrates and random noise in the system. Through training, the network learns to focus on key features related to thickness while ignoring irrelevant interference, which significantly improves the measurement stability under complex working conditions.

[0017] (3) Non-contact, non-destructive and efficient: Based on terahertz technology, the measurement process is completely non-contact and will not cause contamination or damage to the sample. Once the model training is completed, the thickness prediction process can be completed in milliseconds, which is very suitable for online, high-speed detection needs.

[0018] (4) End-to-end simplified process: This invention eliminates the cumbersome signal preprocessing, peak search, physical model fitting and other steps in the traditional method, realizes end-to-end prediction from raw data to thickness value, and reduces the algorithm complexity and the technical requirements of operators.

[0019] (5) Adaptability and generalization ability: By expanding the training dataset (including data on different types of paper, different types of glue, and different environmental conditions), the trained model can have a certain generalization ability and adapt to various changes on the production line. Attached Figure Description

[0020] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.

[0021] Figure 1 The flowchart shows the terahertz thickness measurement method for cigarette adhesive based on deep learning neural networks provided by this invention.

[0022] Figure 2The images shown are imaging results of different glue layer thicknesses in the 340 GHz terahertz imaging system provided by the present invention. (a) is an imaging result of a triangular white paper without glue; (b) is an imaging result of glue with a nominal thickness of 70 micrometers; (c) is an imaging result of glue with a nominal thickness of 140 micrometers; (d) is an imaging result of glue with a nominal thickness of 210 micrometers; (e) is an imaging result of glue with a nominal thickness of 280 micrometers; and (f) is an imaging result of glue with a nominal thickness of 350 micrometers.

[0023] Figure 3 The system provided by this invention generates a terahertz reference signal and its spectrum, wherein (a) is the terahertz reference signal and (b) is the terahertz spectrum.

[0024] Figure 4 The hybrid network structure diagram provided by this invention.

[0025] Figure 5 The model thickness prediction accuracy diagram provided by this invention for different training periods and different error ranges.

[0026] Figure 6 The thickness values ​​obtained by four methods are shown in the figure provided by the present invention. Detailed Implementation

[0027] To better understand the above-described objectives, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other. Furthermore, the present invention can be implemented in other ways different from those described herein; therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0028] This invention constructs an end-to-end thickness measurement system integrating an advanced deep learning architecture, encompassing "data acquisition, feature learning, and thickness regression." This system abandons the traditional thickness inversion process based on physical models and manual feature extraction. Instead, it creatively employs a Conformer (CNN-Transformer) hybrid neural network model to automatically learn and extract deep features highly correlated with glue thickness directly from complex terahertz time-domain signals. This achieves effective decoupling of overlapping echoes and robust suppression of substrate interference.

[0029] To illustrate the effectiveness of the method proposed in this invention, the following detailed description of the above technical solution of this invention is provided through a specific embodiment.

[0030] like Figure 1 As shown, this invention discloses a terahertz thickness measurement method for cigarette adhesive based on deep learning neural networks. The specific implementation steps are as follows: Step S1: Use a point frequency source terahertz imaging system to detect multiple tobacco adhesive samples coated on standard cigarette paper and obtain the corresponding terahertz intensity images; then screen out the areas to be tested in the terahertz intensity images. In this step, the present invention first completes the qualitative analysis and sample screening of tobacco adhesive samples based on the point frequency source terahertz imaging system.

[0031] Before using a high-precision terahertz time-domain spectroscopy (THz-TDS) system for quantitative measurement and model training, this invention first uses a point-frequency source terahertz imaging system for preliminary scanning in order to quickly and intuitively evaluate the glue thickness distribution and screen typical samples.

[0032] In some embodiments, a 340GHz point-source terahertz imaging system can be used to rapidly scan and detect multiple tobacco adhesive samples coated on standard cigarette paper to achieve qualitative investigation and sample screening. Specifically, this includes the following steps: Sample preparation: Obtain multiple tobacco adhesive samples coated on standard cigarette paper. The adhesive thickness of the tobacco adhesive samples covers the expected measurement range, which may be 1-1000 μm.

[0033] Equipment preparation: Prepare a continuous wave terahertz source with a working frequency of 340 GHz and a corresponding focal plane detector array or two-dimensional scanning platform to form a point frequency source terahertz imaging system.

[0034] Data Acquisition and Qualitative Analysis: A series of cigarette adhesive samples of varying thicknesses were placed on the imaging plane. The system acquired terahertz intensity images of each pixel or scan point at a frequency of 340 GHz by scanning. Since the absorption and interference effects of terahertz waves are directly related to the thickness of the adhesive, regions of different thicknesses will exhibit significant differences in grayscale or contrast in the image.

[0035] like Figure 2 As shown, Figure 2 In the figure, (a), (b), (c), (d), (e), and (f) represent the imaging results of the area of ​​the triangular white paper without glue, and the glue-coated areas with nominal thicknesses of 70 micrometers, 140 micrometers, 210 micrometers, 280 micrometers, and 350 micrometers, respectively.

[0036] By observing the images, the uniformity of the glue coating can be qualitatively determined, and areas that are too thin, too thick, or have varying thickness gradients can be identified.

[0037] Sample Screening: By observing and analyzing the acquired terahertz intensity images, the uniformity of the adhesive coating can be qualitatively determined, and areas with uniform thickness or varying thickness gradients can be identified. Combined with the imaging results, images of representative test areas with varying thicknesses can be precisely screened from a large-area sample, including images of typical areas such as uniform thickness regions and regions with specific thickness gradients.

[0038] These selected test areas will serve as precise location guidance for the next stage of quantitative measurement and data collection, ensuring that the subsequent training dataset has good coverage and representativeness, thereby significantly improving the efficiency and quality of data collection.

[0039] Step S2: Use a thickness measuring instrument to detect the area to be measured and obtain the true thickness; The area to be tested is detected and preprocessed using a terahertz time-domain spectroscopy system to obtain the corresponding terahertz time-domain waveform; the actual thickness and the terahertz time-domain waveform are used as samples to establish a tobacco adhesive dataset from multiple samples; After selecting the test areas, it is necessary to calibrate the actual thickness of the test areas and acquire terahertz time-domain signals. First, a high-precision thickness measuring instrument is used to measure the thickness of the test areas selected in the previous steps to obtain the actual thickness as the reference value for the training labels.

[0040] In some embodiments, the thickness measuring instrument can be a white light interferometer or a profilometer. The measurement range covers 1-1000 μm, ensuring the accuracy and reliability of the calibrated thickness values. For each area to be measured, its corresponding true thickness value is strictly recorded, and this value will serve as the true thickness supervision signal for subsequent model training.

[0041] The region under test was detected using a terahertz time-domain spectroscopy system. A high-resolution terahertz time-domain spectroscopy (THz-TDS) system was used to measure each calibrated standard sample. The system typically includes a femtosecond laser, a terahertz emitter (such as a photoconductive antenna), a terahertz receiver, and a precision optical delay line. The pump beam of the experimental setup for the terahertz generation and detection system was generated by an industrial ytterbium-doped femtosecond laser supplied by Ultron Photonics. This laser has a wavelength of 1030 nm, a pulse width adjustable from 800 to 2000 fs, and a repetition rate adjustable from 1 Hz to 100 kHz. In this experiment, it can provide a maximum single-pulse output energy of 2 mJ and a maximum output power of 200 W at a repetition rate of 100 kHz. A combination of a half-wave plate and a polarizing beam splitter (not shown in the figure) was placed after the laser output to adjust the laser intensity. Most of the laser intensity comes from the reflected light from the polarizing beam splitter, serving as the pump beam for the strong-field terahertz generation while simultaneously protecting the high-reflectivity mirror from damage. A beam expander consisting of a plano-concave lens (L1: f = -50 mm) and a plano-convex lens (L2: f = 100 mm) is inserted into the optical path, magnifying the spatial size of the pump beam by a factor of 2 (corresponding to the focal length ratio). This beam expansion helps prevent damage to the lithium tantalate crystal caused by high pump power density. The expanded beam illuminates a reflection grating (Spectrogon: 715. 707. 510 PC 1400 50×50×10 NIR, 1400 lines / mm) at an incident angle of approximately 41.37°. The grating has a negative first-order diffraction efficiency of 95%. Subsequently, a tilted pulse leading-edge pump is generated by negative first-order diffraction (m = -1), with an angle of approximately 10° to the incident beam. This tilted pulse leading-edge pump is further transmitted to a polarized 4f imaging system, which uses a third lens (L3: f = 370 mm) and a fourth lens (L4: f = 200 mm). A half-wave plate, rotating the polarization state to s-polarization, is placed between the two lenses. The beam is ultimately imaged into a lithium tantalate crystal of the same composition with a beam-shrinkage ratio of approximately 1.85. This bulk lithium tantalate crystal is 5% doped with MgO, has a 1030 nm antireflection coating on its input surface, and is cut into an isosceles prism (20 mm × 20 mm × 30 mm) with a 63° apex angle to match the angle of the tilted pulse leading edge pump. The terahertz radiation generated in the crystal is collimated by off-axis parabolic mirrors (OAPs) and indium tin oxide (ITO) elements, causing it to overlap with the probe beam in time and space for the electro-optic sampling process. This sampling employs a conventional balanced detection technique, using a 0.5 mm thick ZnTe electro-optic crystal, a quarter-wave plate, a Wollaston prism, and two photodetectors.For the probe beam, a small portion of the fundamental frequency laser, split by a polarization beam splitter, is fed into an optical parametric amplifier and then compressed to generate an 800 nm probe pulse with a pulse width of approximately 60 fs. A mechanical optical delay line composed of highly reflective mirrors is mounted on an electrically driven linear displacement stage with a travel range of 100 mm to finely adjust the optical path length of the probe beam, ensuring its precise temporal overlap with the terahertz pulse. The two beams are confocally focused onto a ZnTe crystal by an off-axis parabolic mirror OAP4 and precisely aligned with the terahertz beam to detect the terahertz time-domain waveform. The experiment was conducted at room temperature (23°C) and humidity (53%).

[0042] In this step, the sample to be tested is placed in the transmission or reflection measurement optical path, and point measurements are performed in the area to be tested, acquiring the original terahertz time-domain waveform signal corresponding to each measurement point. For example... Figure 3 The figure shows the terahertz reference signal and its spectrum of the terahertz time-domain spectral system, where (a) is the terahertz reference signal and (b) is the terahertz spectrum.

[0043] The acquired raw terahertz time-domain waveform needs to undergo necessary preprocessing operations, including: extracting an effective signal time window, removing background signals from system noise, and performing amplitude normalization, to form regular one-dimensional time-series data, which serves as the terahertz time-domain waveform.

[0044] The preprocessed terahertz time-domain waveform and its corresponding true thickness value constitute a complete training data pair. By performing the above measurements and calibrations on multiple test areas of multiple samples with different thicknesses covering the expected measurement range, a large number of "terahertz time-domain waveform-true thickness" label pairs are collected, establishing a complete dataset for tobacco adhesive. In some embodiments, the dataset can be proportionally divided into a training set, a validation set, and a test set, where the training set is used for model parameter learning, the validation set is used to monitor the training process and prevent overfitting, and the test set is used for final performance evaluation.

[0045] Step S3: Construct a hybrid network, which is used to receive the region to be measured and the terahertz time-domain waveform in the terahertz intensity image and output the predicted value of the thickness of the tobacco adhesive. In this step, the present invention constructs a Conformer hybrid network model for predicting the thickness of tobacco adhesive. The Conformer model innovatively integrates a convolutional neural network with a Transformer module, fully leveraging the complementary advantages of the two architectures. This hybrid network is specifically designed to receive the region to be measured and the terahertz time-domain waveform from a terahertz intensity image and output the predicted thickness.

[0046] like Figure 4 As shown, the hybrid network of the present invention includes a CNN convolutional module and a Transformer encoder module connected in sequence.

[0047] The CNN convolutional module receives the fusion information of pixel intensity information of the test region in the terahertz intensity image and the terahertz time-domain waveform, and processes it through multiple convolutional modules. The convolutional module includes a convolutional layer, an activation function and a pooling layer connected in sequence to obtain output features.

[0048] The CNN part is responsible for efficiently extracting local detail features in terahertz time-domain signals, such as weak reflection peaks, abrupt edges, and specific interference patterns, by utilizing the local receptive field characteristics of its convolutional kernels, and reducing the sequence length through pooling layers. The inherent translation invariance of CNNs gives them a natural ability to filter random noise in the signal, effectively suppressing interference from cigarette paper substrate roughness and system noise.

[0049] Before receiving the feature sequence extracted by the CNN, the Transformer module first performs positional encoding enhancement to preserve the positional information in the sequence. The Transformer then processes the output features through a self-attention mechanism.

[0050] The hybrid network also includes a Feature Coupling Unit (FCU), which employs a multi-layer perceptron structure. Nonlinear spatial transformations facilitate the effective integration of features from CNN and Transformer branches, thereby enhancing the model's feature representation capabilities. Residual connections are introduced throughout the network to improve training efficiency and stability, enabling the model to learn deep feature representations more effectively. Layer normalization techniques are also employed to mitigate the bias of internal covariates, accelerating network convergence and enhancing the model's generalization ability.

[0051] The model ends with a fully connected layer, which is responsible for integrating the fused features from CNN and Transformer, and finally outputs the thickness prediction value in micrometers (μm).

[0052] By combining the local feature extraction capability of CNN with the global context modeling capability of Transformer, the hybrid network of this invention can predict the glue thickness from the original terahertz time-domain waveform affected by cigarette paper substrate noise and signal interference.

[0053] It should be noted that the present invention uses the fusion information of pixel intensity information of the region to be measured in the terahertz intensity image and the terahertz time-domain waveform as the input of the hybrid network.

[0054] Specifically, this invention integrates a 340GHz point-frequency source terahertz imaging system with a terahertz time-domain spectroscopy system into a unified two-stage integrated detection device.

[0055] The two-stage integrated inspection equipment first utilizes 340GHz imaging to rapidly scan a large-area sample, obtaining an overview of the sample's thickness distribution based on pixel intensity information within seconds, and identifying areas requiring focused attention. It then switches to terahertz time-domain spectroscopy measurement mode to perform precise point measurements or line scans on the key areas identified in the imaging stage, acquiring high-precision pixel intensity information to represent high-precision thickness. This combination of "area screening" and "precise point measurement" ensures both comprehensive inspection and measurement accuracy of critical areas, while significantly improving overall inspection efficiency, making it particularly suitable for online quality control in production lines.

[0056] In this embodiment, the pixel intensity information of the test area in the terahertz intensity image obtained by 340GHz imaging can be fused with the fine time-domain waveform acquired by THz-TDS and used as the input of the Conformer hybrid network, which can further improve the measurement robustness in edge regions or extremely non-uniform regions.

[0057] Step S4: Train the hybrid network using the tobacco glue dataset to obtain the trained hybrid network; In this step, the present invention uses a dataset of tobacco adhesive to train the hybrid network. The training process adopts a supervised learning method, which enables the network to automatically optimize its internal parameters by learning from a large number of known "terahertz time-domain waveform-true thickness" label pairs, and establishes a precise nonlinear mapping relationship from complex terahertz time-domain waveforms to thickness values.

[0058] The specific training process is as follows: The preprocessed hybrid network in the training set is fed with the fused information of pixel intensity information of the test area in the terahertz intensity image and the terahertz time-domain waveform. This data is then fed into the Conformer network for forward propagation calculation, and the network outputs the corresponding thickness prediction value. The loss function between the predicted thickness value and the true calibrated thickness value is calculated, typically using mean squared error (MSE) as the loss metric. The gradient of the loss function with respect to the parameters of each network layer is calculated using the backpropagation algorithm. Adaptive optimizers such as Adam are used to iteratively update the network weights, gradually minimizing the loss function.

[0059] In some embodiments, to improve the generalization ability and robustness of the model, various optimization strategies can be adopted during training: First, Gaussian noise is injected into the training data to simulate various disturbances in the actual measurement environment, thereby enhancing the model's ability to suppress noise; second, a cyclic learning rate scheduling strategy is adopted to dynamically adjust the learning rate to avoid getting trapped in local optima and promote the exploration of the global optimum; third, the training process is monitored in real time using a validation set, and training is stopped or hyperparameters are adjusted in a timely manner when the validation set loss no longer decreases or reaches the preset number of iterations to prevent overfitting.

[0060] like Figure 5 As shown, after 42 and 47 iterations, the model achieved thickness prediction accuracy of 99.8% within error ranges of ±20 μm and ±15 μm, respectively. After 55 iterations, the model achieved prediction accuracy of 99.6% within an error range of ±10 μm and 95% within an error range of ±7 μm. The trained network was validated using experimentally measured transmitted echo signals from adhesive layer samples of different thicknesses, further confirming the effectiveness of this deep learning-assisted measurement method. The deviation between the model output and the nominal thickness value was within ±20 μm. To improve the reliability of the experimental results, the thickness of the adhesive layer samples was measured using a thickness gauge. Figure 6 As shown, the thickness values ​​obtained by three different methods are compared: neural network prediction, direct measurement by a thickness gauge, and thickness value calculated from the time interval between two transmission pulses.

[0061] Step S5: For the tobacco adhesive sample to be tested, the terahertz time-domain waveform is obtained by using a terahertz time-domain spectroscopy system; the terahertz time-domain waveform is input into the trained hybrid network to obtain the predicted value of the tobacco adhesive thickness.

[0062] In this step, once the Conformer model has been trained and its performance verified, it can be deployed in the data processing unit of an actual production line or laboratory for rapid and accurate thickness measurement of tobacco adhesive samples with unknown thickness.

[0063] For the cigarette adhesive samples to be tested, a terahertz time-domain spectroscopy system with the same configuration as in the training phase was first used for detection. The cigarette paper sample coated with the adhesive was placed in the measurement optical path of the terahertz time-domain spectroscopy system to acquire the original terahertz time-domain waveform of the sample. This waveform underwent preprocessing operations completely identical to those in the training phase. These preprocessing steps included: extracting a time window containing effective reflection or transmission signals, removing system background noise interference, normalizing the waveform amplitude, and ensuring that the data format remained consistent with the training data. After preprocessing, well-organized one-dimensional time-series data was obtained, which had the same data dimension and feature distribution as the training samples.

[0064] The preprocessed terahertz time-domain waveform is input into the deployed and trained Conformer hybrid network model to obtain the predicted value of the thickness of the adhesive used in cigarettes.

[0065] This invention, through the above hybrid network, can learn and analyze weak echo signals that are severely overlapping in the time domain, breaking through the limitations of traditional algorithms on echo time intervals, and realizing the accurate measurement of the thickness of ultrathin glue layers at the micrometer level.

[0066] While the specific embodiments of the present invention depict actions or steps in a particular order, this should be understood as requiring such actions or steps to be performed in the shown specific order or sequential order, or requiring all illustrated actions or steps to be performed to achieve the desired result. In certain environments, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation may also be implemented individually or in any suitable sub-combination in multiple implementations. The above descriptions are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention.

[0067] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A terahertz thickness measurement method for cigarette adhesive based on deep learning neural networks, characterized in that, Includes the following steps: Step S1: Use a point frequency source terahertz imaging system to detect multiple tobacco adhesive samples coated on standard cigarette paper and obtain the corresponding terahertz intensity images; then screen out the areas to be tested in the terahertz intensity images. Step S2: Use a thickness measuring instrument to detect the area to be measured and obtain the true thickness; The area to be tested is detected and preprocessed using a terahertz time-domain spectroscopy system to obtain the corresponding terahertz time-domain waveform; the area to be tested and the terahertz time-domain waveform in the actual thickness and terahertz intensity images are used as samples to establish a tobacco adhesive dataset from multiple samples; Step S3: Construct a hybrid network, which is used to receive the region to be measured and the terahertz time-domain waveform in the terahertz intensity image and output the predicted value of the thickness of the tobacco adhesive. Step S4: Train the hybrid network using the tobacco glue dataset to obtain the trained hybrid network; Step S5: For the tobacco adhesive sample to be tested, the test area in the terahertz intensity image is obtained using a point frequency source terahertz imaging system. Terahertz time-domain waveforms are obtained by using a terahertz time-domain spectroscopy system; the region to be measured in the terahertz intensity image and the terahertz time-domain waveform are input into a trained hybrid network to obtain the predicted value of the thickness of the tobacco adhesive.

2. The terahertz thickness measurement method for cigarette adhesive based on deep learning neural networks according to claim 1, characterized in that, Step S1 specifically includes: Step S1-1: Obtain multiple tobacco adhesive samples coated on standard cigarette paper, wherein the adhesive thickness of the tobacco adhesive samples ranges from 1 to 1000 μm; Step S1-2: The point frequency source terahertz imaging system is constructed by using a continuous wave terahertz source with a working frequency of 340 GHz and the corresponding focal plane detector array or two-dimensional scanning platform. Steps S1-3: Place multiple tobacco adhesive samples on the imaging plane of the point frequency source terahertz imaging system and scan to obtain the terahertz intensity image of each scanning point at a frequency of 340 GHz. Steps S1-4: Select regions with uniform thickness or varying thickness gradients from the terahertz intensity image as the regions to be tested.

3. The terahertz thickness measurement method for cigarette adhesive based on deep learning neural networks according to claim 2, characterized in that, In step S2, the thickness measuring instrument is a white light interferometer or a step gauge; The steps of detecting and preprocessing the region under test using a terahertz time-domain spectroscopy system to obtain the corresponding terahertz time-domain waveform specifically include: The sample of tobacco adhesive is placed in the transmission or reflection measurement optical path, and point measurements are performed in the area to be measured. The original terahertz time-domain waveform signal corresponding to each measurement point is collected. For the original terahertz time-domain waveform signal, a time window is extracted, background signals with system noise are removed, and amplitude normalization is performed to form one-dimensional time-series data, which serves as the terahertz time-domain waveform.

4. The terahertz thickness measurement method for cigarette adhesive based on deep learning neural network according to claim 3, characterized in that, In step S3, the input to the hybrid network is the fusion information of the pixel intensity information of the region to be tested in the terahertz intensity image and the terahertz time-domain waveform.

5. The terahertz thickness measurement method for cigarette adhesive based on deep learning neural network according to claim 4, characterized in that, The hybrid network in step S3 includes a CNN convolutional module, a Transformer encoder module, and a feature coupling unit; The CNN convolutional module receives the input signal and processes it through multiple convolutional sub-modules to obtain the output features of the CNN convolutional module. The convolutional sub-module includes a convolutional layer, an activation function, and a pooling layer connected in sequence. The output features of the CNN convolutional module are positionally encoded and then input into the Transformer encoder module. After processing by the self-attention mechanism, the output features of the Transformer are obtained. The output features of the CNN convolutional module and the output features of the Transformer are fused by the feature coupling unit, and then output as a predicted value of the thickness of the tobacco glue in micrometers through a fully connected layer.

6. The terahertz thickness measurement method for cigarette adhesive based on deep learning neural network according to claim 5, characterized in that, Step S4 specifically includes: The tobacco adhesive dataset is divided into training and testing sets. The pixel intensity information of the test area in the preprocessed terahertz intensity image in the training set and the fusion information of the terahertz time-domain waveform are used as input and fed into the hybrid network for forward propagation calculation, outputting the corresponding predicted thickness value of tobacco adhesive. The mean square error loss function between the predicted thickness value and the actual thickness of tobacco adhesive is calculated. The gradient of the mean square error loss function with respect to the network parameters of each layer of the hybrid network is calculated through the backpropagation algorithm. The network parameters are iteratively updated using the Adam adaptive optimizer. Training stops when the mean square error loss function is minimized or when the preset number of iterations is reached, resulting in the trained hybrid network.

7. The terahertz thickness measurement method for cigarette adhesive based on deep learning neural networks according to claim 6, characterized in that, The point frequency source terahertz imaging system and the terahertz time-domain spectroscopy system are integrated into a single two-stage integrated detection device.

8. The terahertz thickness measurement method for cigarette adhesive based on deep learning neural network according to claim 7, characterized in that, Step S4 also includes: The trained hybrid network is subjected to lightweight update processing, including model pruning, weight quantization, and knowledge distillation.

Images