Terahertz wave-based IV-type bottle defect detection method

Abstract

The invention belongs to the field of nondestructive testing and structural health monitoring, and relates to an IV-type bottle defect detection method based on terahertz waves. According to the method, a single-point multi-reflection type terahertz detection system is constructed, multi-point terahertz time domain signals of IV-type bottles in different states are collected, a defect real state label is obtained in combination with contact type measurement, and signal preprocessing and data set division are carried out; extracting a flight time physical feature and a machine learning signal feature from the time domain signal to form a fusion feature vector; a physical information neural network containing a feature fusion layer is constructed, a multi-constraint loss function is designed, and a PINN inversion model is obtained through staged training; and finally, in combination with a time-of-flight method, a traditional machine learning model and the output of a PINN model, based on verification set error dynamic weighted fusion, realizing accurate evaluation of a defect state. According to the method, defects such as cracks, layering and pores in the IV-type bottle can be effectively identified, and the detection precision and the system robustness are improved.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing and structural health monitoring, and in particular to a method for detecting defects in Type IV bottles based on terahertz waves. Background Technology

[0002] Type IV hydrogen storage cylinders, as core equipment in the field of hydrogen energy storage and transportation, are formed by winding multiple layers of composite materials, including an inner liner, a glass fiber layer, and a carbon fiber layer. The glass fiber layer plays a crucial role in inner liner protection, stress buffering, and insulation, ensuring the structural integrity and operational safety of the cylinder. However, defects in the glass fiber layer, such as delamination, porosity, fiber breakage, or loose winding, arising during manufacturing or service, will significantly weaken its mechanical load-bearing capacity and interface transfer capabilities, leading to localized stress concentration, inner liner exposure, and even an increased risk of hydrogen permeation. In severe cases, it may induce premature failure of the carbon fiber layer, drastically shortening the service life of the Type IV cylinder. With the large-scale application of hydrogen energy in transportation and energy storage, the need for full life-cycle safety monitoring of Type IV cylinders is increasingly urgent. Accurate identification and location of glass fiber layer defects are a core prerequisite for assessing the structural health of the cylinder and predicting potential failure risks, directly impacting the safe and stable operation of the hydrogen energy storage and transportation system.

[0003] Currently, the detection of defects in the fiberglass layer of Type IV bottles faces numerous technical bottlenecks. Traditional contact testing methods, such as visual inspection and tapping, are highly subjective and have low sensitivity, making it difficult to detect minute or hidden internal defects. While ultrasonic testing is sensitive to some defects, it is easily affected by reflection and scattering interference from complex interfaces in multilayer composite structures, and its ability to resolve microscale defects in thin-layer fiberglass is limited. X-ray testing, although providing high-resolution images, poses safety hazards due to ionizing radiation, is expensive and complex to operate, and lacks sufficient contrast for defects in low-density non-metallic materials. Therefore, there is an urgent need to develop an efficient, safe, and highly sensitive non-destructive testing method for the accurate identification and assessment of defects in the fiberglass layer of Type IV bottles.

[0004] Terahertz technology, with its unique advantages of being non-contact, non-ionizing, and highly penetrating non-metallic composite materials, has become an ideal technique for non-destructive testing of multilayer structures in Type IV bottles. When terahertz waves penetrate the multilayer structure of Type IV bottles, they form characteristic reflection peaks at the interfaces of different media. The time of flight of these peaks is strongly correlated with the defect loss in the corresponding media layer, and loss inversion can be achieved by analyzing the time difference of the reflection peaks in the time-domain signal. However, the time-of-flight method for defect loss inversion is susceptible to interference from factors such as signal baseline drift, overlapping reflection peaks between layers, and rough scattering at interfaces. Especially when the signal characteristics in the defect region are blurred, this can lead to deviations in defect estimation, thus affecting the accuracy of defect loss quantification. While some terahertz detection schemes combined with machine learning can improve signal fitting and defect prediction capabilities, they lack explicit constraints on the physical laws of electromagnetic wave propagation. Therefore, when facing complex defect morphologies or variable operating conditions, it is difficult to guarantee the reliability and generalization of defect loss assessment results. Thus, combining the traditional terahertz time-of-flight method with a fusion physical information model can not only improve the robustness of loss defect inversion but also ensure that the defect loss assessment process conforms to the basic physical mechanisms of electromagnetic wave propagation and material response. Compared with traditional defect detection methods, this strategy has significant advantages in terms of convenience, rationality, quantitative accuracy and anti-interference ability, providing reliable technical support for the accurate quantification of internal defects in the glass fiber layer of Type IV hydrogen storage cylinders and the assessment of their health status.

[0005] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to address the technical problems existing in the background technology. To this end, it provides a terahertz wave-based defect detection method for Type IV hydrogen storage cylinders. This method utilizes terahertz technology to detect defects in the glass fiber layer of Type IV hydrogen storage cylinders, integrating physical and data features. It combines the time-of-flight method, machine learning models, and physical information neural networks, and outputs the final defect state through dynamic weight fusion, significantly improving detection reliability. This technical solution is an effective non-destructive testing method that can flexibly and accurately process terahertz time-domain signal data, achieving high-precision detection of glass fiber layer defects in Type IV hydrogen storage cylinders. This improves the existing technical system and solves the technical pain points of traditional detection methods, such as insufficient accuracy and poor robustness.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for detecting defects in Type IV bottles based on terahertz waves includes the following steps: S1: Construct a reflective terahertz detection system, which includes a terahertz time-domain spectrometer, a focusing lens group, a three-dimensional displacement platform, a standard defect calibration block, and a data acquisition module; select a type IV bottle sample and plan the distribution rules of its circumferential and axial detection points; S2: Perform multiple terahertz time-domain signal acquisitions at each preset detection point, and obtain a standardized signal by averaging; simultaneously use a contact detection device to record the actual defect status of the point as a true label; S3: Preprocessing the terahertz time-domain signal by baseline correction, wavelet denoising, and truncation; extracting two types of features from the preprocessed signal: time-of-flight physical features and machine learning signal features; S4: Construct a physical information neural network model with a feature fusion layer, design a total loss function that integrates data loss and physical constraints, and train it in stages to obtain the PINN defect inversion model; at the same time, train a traditional machine learning model and calculate the defect reference value based on the time-of-flight method. S5: Based on the error on the validation set, dynamically allocate the weights of the time-of-flight method, the traditional machine learning model, and the PINN model, and obtain the final defect assessment result through weighted fusion; use new samples to iteratively optimize the model parameters and weights to enhance the model's generalization and anti-interference ability.

[0008] The following are further technical solutions of the present invention: the terahertz time-domain spectrometer has an output bandwidth of 0.1–3 THz and a sampling rate of not less than 10 GS / s; the focusing lens group realizes the vertical incidence of terahertz waves and the diameter of the focused spot is less than 2 mm; the three-dimensional displacement platform has a positioning accuracy of ±0.01 mm and supports the movement of samples along a preset path; the standard defect calibration block is a Type IV bottle standard part with known defect type and degree; the data acquisition module synchronously stores signals, coordinates and environmental parameters.

[0009] The following is a further defined technical solution of the present invention: Terahertz time-domain signals are repeatedly acquired K times at each preset detection point, where K ≥ 10. During the acquisition process, K sets of original signal time-domain waveforms S are synchronously recorded by the data acquisition module. k (t)(k=1,2,...K); After acquisition, the 3σ criterion is used to remove one field of signal to obtain K. * The effective signal is obtained and averaged to obtain the normalized terahertz time-domain signal S at that sampling point. * (t), and the formula for averaging is: .

[0010] The following is a further defined technical solution of the present invention: K is obtained by eliminating one field signal using the 3σ criterion. * Valid signals include: setting the mean S of the original signal. * (t), with standard deviation σ(t), if Sk (t) satisfies: If the signal is abnormal, it will be rejected.

[0011] The following is a further defined technical solution of the present invention. The features incorporated into the machine learning model include time-of-flight physical features and machine learning signal features. The time-of-flight physical feature data includes three-dimensional feature parameters: the reflection peak Δt1 at the incident surface and the liner-glass fiber interface, Δt2 at the glass fiber interface and the carbon fiber interface, and the theoretical thickness d, wherein: ; Where c is the speed of light, Δt is the terahertz wave flight time, and n is the refractive index of the material.

[0012] The machine learning signal feature data includes 12 features: time domain features: time delay, peak intensity, peak-to-peak time, time width, and energy width; frequency domain features: frequency domain area, peak frequency, centroid frequency, and signal-to-noise ratio; and power spectrum features: average spectral intensity, absorption peak intensity, and resonance peak intensity.

[0013] The following is a further defined technical solution of the present invention: the total loss function for the PINN model design includes a data loss function and a physical constraint loss function, wherein the data loss function L... data The mean squared error between the model's predicted defect loss and the actual loss label is calculated using the following formula: ; where N train The number of training samples. The predicted defect severity value. This is a real label; The physical constraint loss function includes wave speed constraint loss L. v and defect loss conservation constraint loss L d ,in: ; ; in, d represents the model inversion depth, and d represents the actual defect impact depth; pre To predict the defect depth for the model, d max This represents the total thickness of the material layer.

[0014] Therefore, the total loss function is a weighted fusion loss of the data loss function and the physical constraint loss function, and the calculation formula is as follows: ; in, , ; and It is an adjustable hyperparameter that can be adaptively adjusted based on the performance of the validation set.

[0015] The following are further technical solutions defined in this invention. The final defect evaluation result is obtained by a weighted fusion of the time-of-flight method, a machine learning model, and the PINN model, including: The defect loss value of the fiberglass layer is calculated by obtaining time-domain data and the difference between the model defect loss and the physical time-of-flight method. The defect loss result of the fiberglass layer is obtained by using a machine learning model with large sample data and the PINN physical constraint model. The weights are then designed based on the error. The final prediction result of the fiberglass layer defect loss is as follows: ; in , , These are the weighting coefficients for the time-of-flight method, the traditional machine learning model, and the PINN model for predicting defect losses, respectively, and their sum is equal to 1.

[0016] The following is a further defined technical solution of the present invention, which uses the inverse error method to calculate the weight coefficients of three types of methods: the time-of-flight method, the traditional machine learning model, and the PINN model: ; ; ; Among them, MAE TOF MAE ML MAE PINN These are the mean absolute errors of three methods on the validation set: the time-of-flight method, the traditional machine learning model, and the PINN model.

[0017] Compared with the prior art, the present invention has the following technical effects: This invention integrates the time-of-flight method, machine learning model method, and PINN model fusion to detect defect losses in the glass fiber layer of Type IV hydrogen storage cylinders. The key technological breakthrough lies in the integration of the physical and data characteristics of terahertz signals to construct an inversion model that combines physical plausibility with accurate data fitting.

[0018] This invention belongs to non-contact non-destructive testing, which uses multi-source data and combines PINN-enhanced machine learning algorithm for data fusion and neural network regression. It can make comprehensive and accurate predictions and is suitable for detecting defects and losses in the glass fiber layer of Type IV hydrogen storage cylinders.

[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of a reflective terahertz system. Figure 3 This is a circular section diagram of a type IV bottle; Figure 4 This is a diagram of a reflection-based terahertz time-domain spectroscopy detection method; Figure 5 This is a flowchart illustrating the specific implementation of the present invention. Detailed Implementation

[0022] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0023] Example 1: like Figure 1 As shown, this embodiment provides a method for detecting defects in Type IV bottles based on terahertz waves, including: A reflective terahertz detection system was constructed, comprising a terahertz time-domain spectrometer, a focusing lens group, a three-dimensional displacement platform, a standard defect calibration block, and a data acquisition module. A type IV bottle sample was selected, and the distribution rules of its circumferential and axial detection points were planned. For example... Figure 2 As shown, the reflective terahertz detection system also consists of a femtosecond laser, a reflector, a beam splitter, a delay platform, a terahertz transmitter and receiver, a sample, and a computer.

[0024] A circular section diagram of a type IV bottle, as shown below. Figure 3 As shown, it consists of a glass fiber layer, a carbon fiber layer and a metal substrate in sequence. Multiple terahertz time-domain signal acquisitions are performed on each preset detection point, and the average is taken to obtain a standardized signal. At the same time, a contact detection device is used to record the actual defect state of the point as a real label.

[0025] The acquired terahertz time-domain signal is preprocessed by baseline correction, wavelet denoising, and signal truncation; signal features are extracted from the preprocessed signal, including time-of-flight physical features and machine learning signal features.

[0026] A physical information neural network model with a feature fusion layer is constructed, and a total loss function that integrates data loss and physical constraints is designed. The PINN defect inversion model is trained in stages. At the same time, a traditional machine learning model is trained, and the defect reference value is calculated based on the time-of-flight method.

[0027] Based on the errors of the three methods on the validation set, the weights of the time-of-flight method, the traditional machine learning model, and the PINN model are dynamically allocated, and the final defect assessment result is obtained through weighted fusion. The model parameters and weights are iteratively optimized using new samples to enhance the model's generalization and anti-interference capabilities.

[0028] The calculation principle of the time-of-flight method is as follows: Figure 4 As shown, the reflected data will have a time difference, which is the time it takes for the terahertz wave to travel through the fiberglass layer. The final defect loss prediction result is obtained through weighted fusion. The model is iteratively trained using new Type IV bottle defect loss sample data. The inversion accuracy of "multi-model fusion" and a single model is compared, and the model parameters and fusion weights are optimized to enhance the model's generalization ability and anti-interference capability.

[0029] Example 2: like Figure 5 As shown, a reflective terahertz detection system was built, and the circumferential and axial sampling point distribution rules of each bottle were determined by selecting type IV bottles. Terahertz time-domain signals were repeatedly acquired at the sampling points. Each preset sampling point required K times (K≥10) of repeated terahertz time-domain signal acquisition. During the acquisition process, K sets of original signal time-domain waveforms S were synchronously recorded through the data acquisition module. k (t); After acquisition, the 3σ criterion is used to remove one field of signal to obtain K. * The effective signal is obtained and averaged to obtain the normalized terahertz time-domain signal S at that sampling point. * (t), and the formula for averaging is: ; The 3σ criterion is used to eliminate abnormal signals and obtain K. * A group of valid signals is the set mean S of the original signal. * (t), with standard deviation σ(t), if S k (t) satisfies: ; If the measured signal satisfies the above formula, it is determined to be an abnormal signal and is removed.

[0030] Subsequently, the material defect loss measured by a contact thickness gauge was used as the true label of the glass fiber layer. The obtained standardized terahertz time-domain signal is then preprocessed with baseline correction, positive wavelet denoising, and signal truncation.

[0031] Features are extracted and incorporated into the machine learning model, including time-of-flight physical features and machine learning signal features, specifically including: The time-of-flight physical characteristic data includes three dimensions of characteristic parameters, specifically: the reflection peak Δt1 at the incident surface and the liner-glass fiber interface, Δt2 at the glass fiber interface and the carbon fiber interface, and the theoretical thickness d, where: ; Where c is the speed of light, Δt is the terahertz wave flight time, and n is the refractive index of the material.

[0032] The input machine learning signal feature data includes 12 specific features: time domain features: time delay, intensity peak, time domain peak-to-peak value, time width, energy width; frequency domain features: frequency domain area, frequency domain peak, centroid frequency, signal-to-noise ratio; power spectrum features: average spectral intensity, absorption peak intensity, resonance peak intensity.

[0033] The acquired multidimensional feature vectors are input into the machine learning model and the PINN model with a total data loss function. The total loss function designed for the PINN model includes a data loss function and a physical constraint loss function. The data loss function L... data The mean squared error between the model's predicted defect loss and the actual loss label is calculated using the following formula: ; Where N train The number of training samples. The predicted defect severity value. This is a real label; The physical constraint loss function includes wave speed constraint loss L. v and defect loss conservation constraint loss L d ;in: ; ; in d represents the model inversion depth, and d represents the actual defect impact depth; pre To predict the defect depth for the model, d max This represents the total thickness of the material layer.

[0034] Therefore, the total loss function is a weighted fusion loss of data loss and physical constraint loss, specifically expressed as: ;in , ,and and It is based on adaptive adjustment of validation set error.

[0035] The final defect loss prediction result is obtained by fusing the time-of-flight method, machine learning model method, and PINN model. The defect loss value of the fiberglass layer is calculated by using time-domain data and the difference between the outer layer thickness of the Type IV bottle and the physical time-of-flight method. The defect loss result of the fiberglass layer is obtained using a machine learning model with large sample data and the PINN physical constraint model. The weights are then designed based on the errors. The final fiberglass layer defect loss prediction result is as follows: ; WTOF, WML, and WPINN are the weighting coefficients of the corresponding defect loss prediction methods, which sum to 1.

[0036] The weighting coefficients for methods of detecting material defect losses are calculated using the inverse error method, i.e.: ; ; ; Among them, MAE TOF MAE ML MAE PINN These are the mean absolute errors on the revalidation sets for the three methods.

[0037] In summary, this invention constructs a single-point, multiple-reflection terahertz detection system to acquire multi-point terahertz time-domain signals from Type IV bottles in different states. It combines this with contact measurement to obtain the true state labels of defects, and performs signal preprocessing and dataset partitioning. Time-of-flight physical features and machine learning signal features are extracted from the time-domain signals to form a fused feature vector. A physical information neural network with a feature fusion layer is constructed, and a multi-constraint loss function is designed. A PINN inversion model is obtained through phased training. Finally, the outputs of the time-of-flight method, the traditional machine learning model, and the PINN model are combined, and a dynamic weighted fusion based on the validation set error is used to achieve accurate assessment of the defect state. This invention can effectively identify internal defects such as cracks, delamination, and porosity in Type IV bottles, improving detection accuracy and system robustness.

[0038] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any person skilled in the art can make many possible variations and modifications to the technical solution of the present invention, or modify it into equivalent embodiments, without departing from the scope of the present invention's technical solution. Therefore, all equivalent changes made based on the shape, structure, and principle of the present invention without departing from the scope of the present invention's technical solution should be covered within the protection scope of the present invention.

Claims

1. A method for detecting defects in Type IV bottles based on terahertz waves, characterized in that, Includes the following steps: S1: Build a reflective terahertz detection system, which includes a terahertz time-domain spectrometer, a focusing lens group, a three-dimensional displacement platform, a standard defect calibration block, and a data acquisition module; Select a Type IV bottle sample and plan the distribution rules for its circumferential and axial detection points; S2: Perform multiple terahertz time-domain signal acquisitions at each preset detection point, and obtain a standardized signal by averaging; simultaneously use a contact detection device to record the actual defect status of the point as a true label. S3: Preprocessing the terahertz time-domain signal by baseline correction, wavelet denoising, and truncation; extracting two types of features from the preprocessed signal: time-of-flight physical features and machine learning signal features; S4: Construct a physical information neural network model with a feature fusion layer, design a total loss function that integrates data loss and physical constraints, and train it in stages to obtain the PINN defect inversion model; at the same time, train a traditional machine learning model and calculate the defect reference value based on the time-of-flight method. S5: Based on the error on the validation set, dynamically allocate the weights of the time-of-flight method, the traditional machine learning model, and the PINN model, and obtain the final defect assessment result through weighted fusion; use new samples to iteratively optimize the model parameters and weights to enhance the model's generalization and anti-interference ability.

2. The method for detecting defects in Type IV bottles based on terahertz waves as described in claim 1, characterized in that, The terahertz time-domain spectrometer has an output bandwidth of 0.1–3 THz and a sampling rate of no less than 10 GS / s; the focusing lens group enables vertical incidence of terahertz waves, with a focused spot diameter of less than 2 mm; the three-dimensional displacement platform has a positioning accuracy of ±0.01 mm and supports sample movement along a preset path; the standard defect calibration block is a Type IV bottle standard part with known defect types and degrees; the data acquisition module synchronously stores signals, coordinates, and environmental parameters.

3. The method for detecting defects in Type IV bottles based on terahertz waves as described in claim 1, characterized in that, At each preset detection point, terahertz time-domain signals are repeatedly acquired K times, where K ≥ 10. During the acquisition process, K sets of original signal time-domain waveforms S are synchronously recorded by the data acquisition module. k (t)(k=1,2,...K); After acquisition, the 3σ criterion is used to remove one field of signal to obtain K. * The effective signal is obtained and averaged to obtain the normalized terahertz time-domain signal S at that sampling point. * (t), and the formula for averaging is: 。 4. The method for detecting defects in Type IV bottles based on terahertz waves as described in claim 3, characterized in that, K is obtained by eliminating one field signal using the 3σ criterion. * Valid signals include: setting the mean S of the original signal. * (t), with standard deviation σ(t), if S k (t) satisfies: If the signal is abnormal, it will be rejected.

5. The method for detecting defects in Type IV bottles based on terahertz waves as described in claim 1, characterized in that, Time-of-flight physical characteristic data includes three dimensions of characteristic parameters: the reflection peak Δt1 at the incident surface and the liner-glass fiber interface, Δt2 at the glass fiber interface and the carbon fiber interface, and the theoretical thickness d, where: ； Where c is the speed of light, Δt is the terahertz wave flight time, and n is the refractive index of the material.

6. The method for detecting defects in Type IV bottles based on terahertz waves as described in claim 1, characterized in that, The machine learning signal feature data includes 12 parameters: time delay, peak intensity, peak-to-peak time, time width, energy width, area in the frequency domain, peak in the frequency domain, centroid frequency, signal-to-noise ratio, average spectral intensity, absorption peak intensity, and resonance peak intensity.

7. The method for detecting defects in Type IV bottles based on terahertz waves as described in claim 1, characterized in that, The total loss function designed for the PINN model includes a data loss function and a physical constraint loss function. The data loss function L data The mean squared error between the model's predicted defect loss and the actual loss label is calculated using the following formula: ; where N train The number of training samples. The predicted defect severity value. This is a real label; The physical constraint loss function includes wave speed constraint loss L. v and defect loss conservation constraint loss L d ,in: ； ； in, d represents the model inversion depth, and d represents the actual defect impact depth; pre To predict the defect depth for the model, d max This represents the total thickness of the material layer.

8. The method for detecting defects in Type IV bottles based on terahertz waves as described in claim 7, characterized in that, The total loss function is a weighted fusion loss of the data loss function and the physical constraint loss function, and the calculation formula is as follows: ； in, , ; and It is an adjustable hyperparameter that adaptively adjusts based on the performance on the validation set.

9. The method for detecting defects in Type IV bottles based on terahertz waves as described in claim 1, characterized in that, The final defect assessment result is obtained by a weighted fusion of the time-of-flight method, machine learning model, and PINN model, including: The defect loss value of the fiberglass layer is calculated by obtaining time-domain data and the difference between the model defect loss and the physical time-of-flight method. The defect loss result of the fiberglass layer is obtained by using a machine learning model with large sample data and the PINN physical constraint model. The weights are then designed based on the error. The final prediction result of the fiberglass layer defect loss is as follows: ； in , , These are the weighting coefficients for the time-of-flight method, the traditional machine learning model, and the PINN model for predicting defect losses, respectively, and their sum is equal to 1.

10. The method for detecting defects in Type IV bottles based on terahertz waves as described in claim 1, characterized in that, The weight coefficients of three methods—time-of-flight method, traditional machine learning model, and PINN model—are calculated using the inverse error method. ； ； ； Among them, MAE TOF MAE ML MAE PINN These are the mean absolute errors of three methods on the validation set: the time-of-flight method, the traditional machine learning model, and the PINN model.

Images