Method for measuring mechanical properties of materials based on terahertz spectral characteristic parameters

Abstract

The application discloses a material mechanical property determination method based on terahertz spectrum characteristic parameters, performs time-frequency domain distribution analysis on the obtained terahertz spectrum of a sample material, extracts terahertz spectrum time-frequency characteristics according to selected time-frequency domain distribution parameters, performs signal similarity calculation on the terahertz spectrum time-frequency characteristics of the sample material, and inputs the signal similarity calculation result into a probability distribution model or a data fitting model to obtain the mechanical property parameters of the sample material. The application performs time-frequency domain distribution analysis, characteristic parameter representation, similarity calculation, data fitting and probability neural network training, establishes a fast matching model of the terahertz spectrum time-frequency characteristics of the material and the mechanical property parameters, and realizes fast acquisition of the mechanical property of the sample material under nondestructive conditions.

Description

Technical Field

[0001] This invention belongs to the field of terahertz science and technology, and relates to the accurate, reliable and rapid characterization of terahertz spectrum characteristic parameters for signal estimation and their matching with material mechanical properties. Background Technology

[0002] Terahertz waves can cover the characteristic spectra of semiconductors, plasmas, and organic matter and biological macromolecules. Utilizing electromagnetic radiation in this band can deepen and expand our understanding of fundamental scientific questions in physics, chemistry, astronomy, informatics, and life sciences. Terahertz technology has wide applications in radar, remote sensing, homeland security, highly secure data communication and transmission, atmospheric and environmental monitoring, real-time biological information extraction, and medical diagnostics.

[0003] Shore hardness, also known as Shore scale hardness, is a method for testing and representing the hardness of materials. It is widely used in research, production, and metrological testing in industries such as machinery manufacturing, metallurgy, and chemicals. For example, it is used to test the hardness properties of finished and semi-finished products, thereby achieving quality control of rubber products. Currently, the main Shore hardness testers are of two types: A and D. JJG304-2003 "Type A Shore Hardness Tester" specifies that Type A is mainly used for soft rubber, elastomers, neoprene rubber, and silicone resin, while JJG1039-2008 "Type D Shore Hardness Tester" is suitable for harder materials such as hard rubber, epoxy resin, and hard plastics. The two types of Shore hardness testers described above consist of a micrometer, an indicator, and standard weights. When measuring hardness, pressure needs to be applied to the material to create a certain degree of penetration strength, which can cause destructive effects on the material itself. (Generally, the Shore hardness measurement process involves pressing a fixed quantity of the test sample perpendicularly into a fixed groove with an indenter. When the sample is subjected to pressure, a small plastic deformation occurs at the bottom of the groove. When the deformation of the sample reaches a level that disrupts the original equilibrium state, the reading of the Shore hardness tester will change.)

[0004] Terahertz technology has been reported in testing methods for rubber products, but the method of characterizing parameters can affect the accuracy of the test results. For example, Chinese patent CN114397266A establishes a functional relationship between the aging time of silicone rubber and the absorption peak value of a certain frequency terahertz wave, but this patent cannot detect quality control or product performance indicators related to aging. Chinese patent CN216117319U points out that the change in dielectric constant caused by the aging of silicone rubber has a good linear relationship with the return loss or insertion loss of terahertz waves, but this patent requires a large number of parameters and the acquisition of parameters based on professional simulations, which is difficult to implement. Summary of the Invention

[0005] The purpose of this invention is to provide a method for determining the mechanical properties of materials based on terahertz spectrum characteristic parameters. This invention utilizes a series of signal processing methods targeting the terahertz spectrum, including time-frequency characteristic parameter characterization, similarity calculation, and state matching, to obtain the determination results of mechanical property parameters such as Shore hardness. This avoids complex metrological testing procedures and the inherent destructive nature of the tests themselves, ultimately achieving rapid knowledge of the mechanical properties of materials under non-destructive conditions.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] 1) Obtain the terahertz spectrum of the test sample material under different material mechanical conditions (e.g., aging time from low to high), where the terahertz spectrum of the i-th test sample material (i∈N and i>1) is the terahertz wave echo signal collected after the test sample material completes the state change from the initial material mechanical conditions (e.g., no aging) to the i-th material mechanical conditions (formed by reflection or scattering after the test sample material is irradiated by terahertz waves), that is, the terahertz spectrum of the i-th sequence collected.

[0008] 2) Perform time-frequency domain distribution analysis on the terahertz spectra obtained in step 1 to obtain time-frequency matrices. After characterizing the terahertz spectrum feature parameters using each time-frequency matrix obtained from the analysis, extract the corresponding time-frequency information on the selected feature frequency bands as the terahertz spectrum features of the test sample material under different material mechanical conditions.

[0009] 3) Select one of the test sample materials under different material mechanical conditions as the reference sample material. Calculate the signal similarity between the terahertz spectrum characteristics of the reference sample material and the terahertz spectrum characteristics of the test sample materials under different material mechanical conditions. Based on the mechanical property parameters of some or all of the test sample materials under different material mechanical conditions (e.g., the results of measuring the corresponding test sample materials using a Shore hardness tester under each material mechanical condition) and the corresponding signal similarity calculation results, establish a rapid matching model of material mechanical properties based on the correlation between material terahertz spectrum characteristics and mechanical property parameters.

[0010] 4) For the terahertz spectrum of the test sample material under unknown material mechanical conditions obtained after irradiation with the same terahertz wave as in step 1, perform time-frequency domain distribution analysis and terahertz spectrum feature extraction as in step 2. Calculate the signal similarity between the terahertz spectrum features of the reference sample material and the terahertz spectrum features of the test sample material under unknown material mechanical conditions. Input the signal similarity calculation results corresponding to the test sample material under unknown material mechanical conditions into the above-mentioned fast matching model for material mechanical properties. Use the output of this model to obtain the measurement results of the mechanical property parameters of the test sample material under unknown material mechanical conditions.

[0011] Preferably, the specific processing method of the time-frequency domain distribution analysis is selected according to the type of time-frequency transformation (e.g., time scale decomposition, empirical mode decomposition, wavelet packet decomposition, etc.) and the frequency dimension of the characteristic frequency band, so as to use the optimal time-frequency domain distribution analysis to amplify the characteristics of the signal in the time-frequency domain.

[0012] Preferably, the reference sample material is selected from the test sample materials with the maximum or minimum mechanical property parameter among the test sample materials collected under different material mechanical conditions (e.g., at least five material mechanical conditions that increase or decrease sequentially).

[0013] Preferably, the determination of the optimal time-frequency domain distribution analysis is based on the principle that the monotonicity of the terahertz spectrum of different sequences is best after being characterized by time-frequency characteristic parameters. For example, the monotonicity of the terahertz spectrum characteristic parameters of the test sample material under at least five successively increasing or decreasing material mechanical conditions is the most significant as the corresponding mechanical property parameters of the test sample material change. The indicator for measuring the significance of monotonicity is the mean of the relative change rate of the terahertz spectrum characteristic parameters of the test sample material under two adjacent material mechanical conditions (referring to the relative change amplitude of the terahertz spectrum characteristic parameters of the test sample material under the latter material mechanical condition relative to the former material mechanical condition when the mechanical property parameters of the test sample material undergo a unit relative change). The higher the mean value, the more significant the monotonicity is determined to be.

[0014] Preferably, the determination of the optimal time-frequency domain distribution analysis specifically includes the following steps: Summing the time-frequency matrix obtained from the time-frequency domain distribution analysis of the terahertz spectrum of the test sample material under each material mechanical condition to obtain the terahertz spectrum characteristic parameters of the test sample material under each material mechanical condition; if the summed terahertz spectrum characteristic parameters of the test sample material under different material mechanical conditions show a monotonically increasing or monotonically decreasing relationship with the change of the corresponding mechanical performance parameters of the test sample material, then the time-frequency domain distribution analysis is determined to be usable (i.e., the time-frequency transformation type and corresponding characteristic frequency band of the time-frequency domain distribution analysis are usable); if it is determined that multiple time-frequency transformation types and corresponding characteristic frequency bands make the characterized characteristic parameters all satisfy a monotonic relationship, then the relative change rate of the terahertz spectrum characteristic parameters of the test sample material obtained under the corresponding time-frequency domain distribution analysis is calculated, and the time-frequency transformation type and corresponding characteristic frequency band of the optimal time-frequency domain distribution analysis are determined by referring to the case with the highest average relative change rate.

[0015] Preferably, the signal similarity calculation can assess the degree of similarity between two things. Specific processing methods (e.g., Euclidean distance, dynamic time warping, Pearson correlation coefficient, etc.) are selected according to the principle of comprehensively measuring both global and detailed signal properties. This ensures that the monotonicity of the terahertz spectrum of the test sample material under at least five sequentially increasing or decreasing material mechanical conditions, after signal similarity calculation, with changes in the corresponding mechanical property parameters of the test sample material, is maintained. For example, if different sequences of terahertz spectra, after being measured by the selected optimal signal similarity calculation, show a significant change in the monotonicity with changes in the corresponding mechanical property parameters of the test sample material... The monotonicity of the parameter changes is most significant (where the measure of monotonicity is the mean of the relative rate of change of the terahertz spectrum of the test sample material under two adjacent material mechanical conditions directly calculated by signal similarity). Therefore, the optimal signal similarity calculation should also ensure that the monotonicity of the characteristics of the terahertz spectrum of different sequences (obtained through the above optimal time-frequency domain distribution analysis) after being measured by the signal similarity calculation is most significant with the change of the corresponding mechanical property parameters of the test sample material (where the measure of monotonicity is the mean of the relative rate of change of the terahertz spectrum characteristics of the test sample material under two adjacent material mechanical conditions calculated by signal similarity).

[0016] Preferably, the rapid matching model for material mechanical properties specifically adopts a probability distribution model or a data fitting model. The probability distribution model is a trained probabilistic neural network. By using the mechanical property states and their probabilities directly output by the probability distribution model, the mechanical properties of the test sample under unknown material mechanical conditions can be obtained. Alternatively, the mechanical properties of the test sample under unknown material mechanical conditions can be directly obtained using the data fitting model.

[0017] Preferably, during the training and establishment of the probability distribution model, if the number of test sample materials under the aforementioned different material mechanical conditions is sufficient and reflects the changes in the mechanical property parameters among the corresponding test sample materials on a scale not greater than 0.1 (e.g., not greater than 0.1A for Shore hardness), then the probability distribution model outputs the mechanical property parameter 'a' according to the classification problem. i and the probability y of the corresponding mechanical performance parameters iWhere i = 1, 2…n, and n is the number of some or all of the test sample materials under different material mechanical conditions, the determination result of the mechanical performance parameters of the test sample materials under unknown material mechanical conditions is determined according to the mechanical performance parameter corresponding to the maximum probability value of the output. That is, when the terahertz spectrum characteristics of the test sample materials under different material mechanical conditions are calculated with the terahertz spectrum characteristics of the reference sample materials without grouping, the mechanical performance parameter corresponding to the maximum probability value of the output is taken as the determination result. However, when the terahertz spectrum characteristics of the test sample materials under different material mechanical conditions are calculated with the terahertz spectrum characteristics of the reference sample materials in sequence according to the frequency dimension, the material mechanical performance parameter corresponding to the maximum probability value of most (e.g., three out of five consecutive groups) in the grouped frequency dimension and which is the same is taken as the determination result. If the number of test sample materials under the above different material mechanical conditions can only reflect the changes in the mechanical performance parameters between the corresponding test sample materials on a scale greater than 0.1, then the probability distribution model outputs the mechanical performance parameter a according to the regression problem. i and the probability y of the corresponding mechanical performance parameters i Where i = 1, 2…n, and n is the number of some or all of the test sample materials under different material mechanical conditions, then for the case of ungrouped signal similarity calculation above, the determination results of the mechanical property parameters of the test sample materials under unknown material mechanical conditions are calculated according to the following formula:

[0018]

[0019] The beneficial effects of this invention are reflected in:

[0020] This invention enables the determination of mechanical properties of materials under non-destructive conditions by characterizing terahertz spectrum feature parameters, calculating similarity, and matching mechanical property parameters by constructing an association model. It also takes into account both linear relationship-based fitting models and machine learning-based probability distribution models, thus enabling the establishment of a method for determining the mechanical properties of materials with very little sample data. This provides a feasible technical approach for the rapid acquisition of the mechanical properties of materials under non-destructive conditions.

[0021] Furthermore, by selecting the optimal time-frequency domain distribution analysis and the optimal signal similarity calculation (for example, for non-destructive testing of Shore hardness, wavelet packet decomposition and dynamic time warping are used to process the terahertz spectrum), this invention can achieve rapid matching between the mechanical property parameters such as Shore hardness of materials and the characteristics of the terahertz spectrum (as long as a monotonic relationship can be formed with the change of mechanical properties of materials through corresponding time-frequency domain distribution analysis and signal similarity calculation under different material mechanical conditions), thus avoiding the sample destructiveness and process complexity of traditional mechanical property testing. Attached Figure Description

[0022] Figure 1 This is a flowchart of a method for determining the mechanical properties of materials based on terahertz spectrum characteristic parameters in an embodiment of the present invention.

[0023] Figure 2 Terahertz spectra of materials under different aging time conditions (Shore hardness of 81.9A and 91.2A).

[0024] Figure 3a The characteristic parameters were obtained by summing the time-frequency matrices of the materials obtained under the terahertz spectrum of the material under the inherent time scale decomposition of 24h aging time (Shore hardness of 65.1A) and other aging time conditions (Shore hardness of 69.1A, 74.1A, 81.9A, and 91.2A respectively).

[0025] Figure 3b The characteristic parameters were obtained by summing the time-frequency matrices of the materials obtained after applying spectral kurtosis decomposition of the terahertz spectra of the materials under the 24-hour aging time condition (Shore hardness of 65.1A) and other aging time conditions (Shore hardness corresponding to 69.1A, 74.1A, 81.9A, and 91.2A).

[0026] Figure 3c The characteristic parameters are obtained by summing the time-frequency matrices obtained after decomposing the terahertz spectra of materials under 24h aging time (Shore hardness of 65.1A) and other aging time conditions (Shore hardness of 69.1A, 74.1A, 81.9A, and 91.2A) using 4-3 node (2.75GHz-3.37GHz band) wavelet packet decomposition.

[0027] Figure 3d The characteristic parameters are obtained by summing the time-frequency matrices obtained after decomposing the terahertz spectra of materials under 24h aging time (Shore hardness of 65.1A) and other aging time conditions (Shore hardness of 69.1A, 74.1A, 81.9A, and 91.2A) using 4-4 ​​node (3.37GHz-4.58GHz band) wavelet packet decomposition.

[0028] Figure 4 The relative average rate of change of the terahertz spectrum of the material under 24h aging time (Shore hardness of 65.1A) and the terahertz spectrum of the material under other aging time conditions (Shore hardness of 69.1A, 74.1A, 81.9A, and 91.2A) is calculated using three similarity calculations (Pearson correlation coefficient, Euclidean distance, and dynamic time warping). The relative average rate of change of the three similarity calculations is also calculated after performing 4-3 node (2.75GHz-3.37GHz band) wavelet packet decomposition on the above terahertz spectra.

[0029] Figure 5 To obtain the similarity calculation results obtained by applying wavelet packet decomposition and dynamic time warping to the terahertz spectrum of a material under a certain aging time condition (Shore hardness of 81.9A), and inputting them into a trained probabilistic neural network (trained for Shore hardness of 81.9A), the output probability distribution is obtained.

[0030] Figure 6a This is a data fitting graph showing the similarity calculation results between the Shore hardness (69.1A, 74.1A, 91.2A) measured under different aging time conditions and the corresponding terahertz spectra of the materials, obtained by wavelet packet decomposition and non-grouped dynamic time warping.

[0031] Figure 6b To obtain the similarity calculation results obtained by applying wavelet packet decomposition to the terahertz spectrum of a material under a certain aging time condition (Shore hardness of 81.9A) and performing dynamic time warping without grouping, and inputting the results into a trained probabilistic neural network (untrained for the case of Shore hardness of 81.9A), the output probability distribution is obtained.

[0032] Figure 6c To obtain the similarity calculation results obtained by applying wavelet packet decomposition and dynamic time warping to the terahertz spectrum of a material under a certain aging time condition (Shore hardness of 81.9A), and inputting them into a trained probabilistic neural network (untrained for Shore hardness of 81.9A), the output probability distribution is obtained. Detailed Implementation

[0033] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. These embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0034] First, taking nitrile rubber as an example, we will specifically explain the establishment of a method for determining the mechanical properties of materials based on terahertz spectral characteristic parameters:

[0035] Step 1: Obtain the terahertz spectrum (reflection signal) of the material in all bands within the 0-15THz range under different aging time conditions (T1=24h, T2=72h, T3=168h, T4=336h, T5=672h). i (i = 1, 2, 3, 4, 5), where i is the terahertz spectrum index; the corresponding Shore hardness (65.1A, 69.1A, 74.1A, 81.9A, 91.2A) of the material under different aging time conditions (T1, T2, T3, T4, T5) were measured using a Shore hardness tester. Figure 2As shown, the terahertz spectra of materials under two aging time conditions corresponding to Shore hardness 81.9A and 91.2A (referred to as Shore hardness 81.9A and Shore hardness 91.2A) were selected and compared. They were found to be very similar. Therefore, time-frequency characteristic analysis was considered to amplify the characteristics of terahertz spectra of different numbers.

[0036] Step two attempts to process the terahertz spectra of different indices using intrinsic time-scale decomposition, spectral kurtosis decomposition, and wavelet packet decomposition, and then sums the time-frequency matrices of each terahertz spectrum obtained from each method. The results are as follows: Figure 3a , Figure 3b , Figure 3c As shown, only the summation results of the terahertz spectrum time-frequency matrices obtained after wavelet packet decomposition in band 4-3 (2.75GHz-3.37GHz) satisfy the monotonicity condition with the change of the corresponding Shore hardness. The summation results of the terahertz spectrum time-frequency matrices obtained after intrinsic time-scale decomposition or spectral kurtosis decomposition do not satisfy the monotonicity condition with the change of the corresponding Shore hardness. Furthermore, when the wavelet basis is rbio3.1, selecting other wavelet packet decomposition bands, such as after wavelet packet decomposition in band 4-4 (3.37GHz-4.58GHz), and summing the obtained terahertz spectrum time-frequency matrices separately, yields the following results: Figure 3d As shown, compared with the selected frequency band 4-3 (2.75GHz-3.37GHz), it was found that the summation results of the time-frequency matrices of each terahertz spectrum under the 4-4 frequency band are not monotonic with the change of the corresponding Shore hardness. Therefore, not all frequency bands can satisfy the monotonicity condition in wavelet packet decomposition. The following steps use the wavelet packet 4-3 frequency band for time-frequency feature analysis.

[0037] Step 3: The terahertz spectrum of the material at Shore hardness 65.1A is sequentially compared with the terahertz spectra of the material at Shore hardness 69.1A, 74.1A, 81.9A, and 91.2A. Pearson correlation coefficient, Euclidean distance, and dynamic time warping are calculated. The relative average rate of change of the obtained similarity calculation results is taken (i.e., the similarity calculation is performed on 5 sets of terahertz spectrum data at Shore hardness 65.1A and 65.1A, 65.1A and 69.1A, 65.1A and 74.1A, 65.1A and 81.9A, and 65.1A and 91.2A. The relative rate of change of the obtained results, i.e. the measurement value, is calculated according to the subsequent aging time relative to the previous aging time as a function of Shore hardness, and then the average is taken). The terahertz spectrum of the material at a Shore hardness of 65.1A was decomposed using wavelet packets, and then compared with the terahertz spectra of the materials at Shore hardnesses of 69.1A, 74.1A, 81.9A, and 91.2A after wavelet packet decomposition. Pearson correlation coefficient, Euclidean distance, and dynamic time warping were calculated, and the relative average rate of change was used to calculate the similarity results. The results are as follows... Figure 4As shown, the metric obtained by dynamic time warping after wavelet packet decomposition has the highest relative average rate of change. Moreover, compared with not performing wavelet packet decomposition, the overall effect of different similarity calculations after wavelet packet decomposition is better. The following steps use dynamic time warping after wavelet packet decomposition to process the extracted features.

[0038] Step four involves performing wavelet packet decomposition on the terahertz spectra at Shore hardness levels of 69.1A, 74.1A, 81.9A, and 91.2A. The resulting time-frequency matrices are then divided into multiple groups or not grouped, and dynamically time-warped with the terahertz spectrum time-frequency matrix obtained from wavelet packet decomposition at Shore hardness level of 65.1A. The resulting metrics and corresponding labels (Shore hardness 69.1A, 74.1A, 81.9A, 91.2A) are input into a probabilistic neural network for training, yielding a result based on the material's terahertz spectrum characteristics and Shore hardness. A probabilistic distribution model of the correlation between material mechanical properties; or performing wavelet packet decomposition on the terahertz spectra at Shore hardness 69.1A, 74.1A, 81.9A, and 91.2A, and performing dynamic time warping calculations on the obtained time-frequency matrices without grouping them with the time-frequency matrices of the terahertz spectrum obtained by wavelet packet decomposition at Shore hardness 65.1A, and fitting some or all of the metric values ​​with the corresponding Shore hardness labels to obtain a material mechanical property data fitting model based on the correlation between material terahertz spectrum characteristics and Shore hardness.

[0039] The specific steps of the above-mentioned non-grouped training are as follows: The entire time-frequency matrix after 4-3 band wavelet packet decomposition of the terahertz spectrum under Shore hardness 65.1A and the entire time-frequency matrix after 4-3 band wavelet packet decomposition of the terahertz spectrum under Shore hardness 69.1A, 74.1A, 81.9A, and 91.2A are dynamically time-warped (equivalent to obtaining the same value under Shore hardness labels 69.1A, 74.1A, 81.9A, and 91.2A) are substituted into the probabilistic neural network for training.

[0040] The specific steps of the above group training are as follows: The time-frequency matrix of the terahertz spectrum under Shore hardness 65.1A, after 4-3 band wavelet packet decomposition, is divided into two or more frequency dimensions (specifically six frequency dimensions). This is then compared with the time-frequency matrix of the terahertz spectrum under Shore hardness 69.1A, 74.1A, 81.9A, and 91.2A, after 4-3 band wavelet packet decomposition, to calculate the value through dynamic time warping. (This is equivalent to obtaining six values ​​for each of the Shore hardness labels 69.1A, 74.1A, 81.9A, and 91.2A, i.e., the number of groupings by frequency dimension.) Figure 5 The sample number (corresponding to the sample number in the original text) is substituted into the probabilistic neural network for training.

[0041] like Figure 5As shown, the probability distribution model for the material's mechanical properties was tested. When Shore hardness labels 69.1A, 74.1A, 81.9A, and 91.2A were all present, the probability distribution of the output Shore hardness label 81.9A was determined. The results showed that because Shore hardness label 81.9A and its corresponding metric were input in groups during training, Shore hardness label 81.9A consistently produced the optimal output compared to other labels (i.e., outputting Shore hardness label 81.9A with the highest probability). This result suggests that when the number of Shore hardness labels is sufficiently large (e.g., the scale for obtaining Shore hardness parameters is no greater than 0.1A, meaning adjacent Shore hardness labels vary by a scale no greater than 0.1A), the above probability distribution model for the material's mechanical properties can accurately and quickly match the corresponding label based on the input metric, and output it with the highest probability, even when the aging time conditions of the corresponding label are unknown. This allows for the non-destructive determination of the Shore hardness of the material under the given aging time conditions.

[0042] like Figure 6a As shown, the Shore hardness values ​​of 69.1A, 74.1A, and 91.2A are fitted with their corresponding Shore hardness labels, resulting in a fitted curve formula of y = 0.4883x + 64.19(r 2 =1), substituting the Shore hardness value of 36.29 corresponding to the Shore hardness label 81.9A into the fitting curve formula for calculation, the result is 81.91A, which is very close to the actual value of 81.9A. This result shows that the data fitting model is not affected by the number of Shore hardness labels when used to determine the Shore hardness of materials under unknown aging time conditions.

[0043] For the above probability distribution model, if a certain Shore hardness label (e.g., 81.9A) and its corresponding metric value are omitted during the non-grouping training process, then when this metric value is substituted into the probabilistic neural network trained with this input data for prediction, only different probabilities of output under labels 69.1A, 74.1A, and 91.2A can be obtained. Figure 6b However, if the probability distributions of the obtained labels 69.1A, 74.1A, and 91.2A are calculated using a similar weighting method, i.e., 0.272×69.1+0.384×74.1+0.342×91.2=78.589, then a measurement result closer to the actual value 81.9A (i.e., 78.589A) can be obtained. Similarly, if a certain Shore hardness label (e.g., 81.9A) and its corresponding measurement value are omitted during the group training process of the above probability distribution model, then when this measurement value is substituted into the probabilistic neural network trained with this input data for prediction, although only the different probabilities output according to six different samples under labels 69.1A, 74.1A, and 91.2A can be obtained (…), the probability distributions of the obtained labels 69.1A, 74.1A, and 91.2A can be obtained (…). Figure 6cHowever, by averaging the obtained probabilities (meaning averaging the probabilities of each label across six samples) and then calculating the average probability distribution of the labels 69.1A, 74.1A, and 91.2A using a similar weighting method, i.e., 0.2735×69.1+0.6860×74.1+0.0405×91.2=78.5742, a measurement result (78.5742A) that is closer to the actual value 81.9A can be obtained. Although these two calculated measurement results are not as accurate as those calculated using the data fitting model, they provide a feasible approach to obtaining relatively accurate Shore hardness measurement results when the amount of aging experiment data is limited.

[0044] The features of this invention are as follows:

[0045] (1) This invention employs multiple methods, including time-frequency domain distribution analysis, similarity calculation, data fitting, and probabilistic neural network training, and ultimately utilizes emerging artificial intelligence techniques to form quantitative relationships and probability distribution models to output Shore hardness. This avoids the destructive and cumbersome nature of Shore hardness tester measurement experiments and significantly improves the efficiency of obtaining Shore hardness parameters (see [link]). Figure 1 ).

[0046] (2) This invention compares different similarity calculation methods and combines similarity calculation methods with time-frequency domain distribution analysis methods. The time-frequency domain distribution analysis method amplifies the feature similarity calculation results of different data, making it easier to train features and further improve the prediction ability under small sample conditions.

[0047] (3) The signal processing framework proposed in this invention integrates multiple emerging technologies, including time-frequency domain distribution analysis, feature parameter characterization, and similarity calculation that takes into account both global and detailed signal measurement effects, as well as artificial intelligence methods such as probabilistic neural networks. It proposes specific evaluation indicators and judgment methods for time-frequency transformation and similarity measurement, thereby achieving accurate determination of mechanical performance parameters. This includes determining the type of time-frequency transformation and its parameters for optimal time-frequency domain distribution analysis from the dimension of monotonicity significance, determining the similarity calculation method from the principle of monotonicity significance, and determining the optimal fast matching model method from the mechanical performance measurement error.

[0048] (5) The input signal of the signal similarity calculation method proposed in this invention is the time-frequency characteristics of the terahertz spectrum data of the test sample and the selected reference sample. The time-frequency characteristics are the simplest and most effective information extracted from the terahertz spectrum data after time-frequency distribution analysis. Their differences are more significant than those of the original terahertz spectrum data. The similarity calculation method can better achieve the optimal measurement effect from the two dimensions of local details and global trends.

[0049] (6) Traditional similarity measurement methods, such as Euclidean distance, face problems such as unequal time series lengths and phase drift. Dynamic time warping can better match similar local features between time series, not only eliminating the deficiency of lockstep measurement methods in measuring unequal time series, but also effectively dealing with characteristics such as phase shift and amplitude variation of time series. Compared with Euclidean distance and Pearson correlation coefficient, the dynamic time warping method used in this invention can further amplify the differences between terahertz spectrum features, making it easier to substitute into probability distribution models and data fitting models to obtain the best data prediction and fitting relationship.

[0050] (7) In practical application scenarios, terahertz spectrum data may be affected by factors such as the type of material to be tested and the scale of mechanical performance parameter testing state, which may result in complex and diverse sample state situations. The mechanical performance fast matching model (probability distribution model, data fitting model) proposed in this invention can effectively solve the mechanical performance parameter classification problem under large sample training conditions, while also effectively solving the regression problem of continuous and accurate prediction of mechanical performance parameters under small sample training conditions.

[0051] In summary, this invention effectively filters terahertz spectra through time-frequency domain distribution analysis. With accurate spectral distribution as the goal and transformation form and parameters as constraints (e.g., the transformation form is the type of time-frequency transformation used, and the parameters are the frequency band selection), it can be applied to various material mechanical conditions. It constructs a database based on the terahertz spectrum distribution characterization results, builds a probability distribution model of material mechanical properties and spectral characteristics based on machine learning algorithms, and constructs a data fitting model through mathematical relationships, ultimately achieving rapid knowledge of material mechanical properties under non-destructive conditions.

Claims

1. A non-destructive method for determining the mechanical properties of materials based on terahertz spectrum characteristic parameters, characterized in that: Includes the following steps: 1) Obtain the terahertz spectrum of the test samples under different material mechanical conditions; 2) Perform time-frequency domain distribution analysis on the terahertz spectra obtained in step 1, characterize the terahertz spectrum characteristic parameters, and extract the corresponding time-frequency information on the selected characteristic frequency bands as the terahertz spectrum characteristics of the test samples under different material mechanical conditions. 3) Calculate the similarity between the terahertz spectral characteristics of the reference sample and the terahertz spectral characteristics of the test samples under different material mechanical conditions. Based on the material mechanical property parameters of some or all of the test samples under different material mechanical conditions and the corresponding similarity calculation results, i.e., the terahertz spectral characteristic metrics of these samples, establish a rapid matching model of material mechanical properties based on the correlation between material terahertz spectral characteristics and mechanical property parameters. The reference sample is a test sample selected from the test samples under different material mechanical conditions. 4) Obtain the terahertz spectrum of the test sample under unknown material mechanical conditions and extract the corresponding time-frequency information on the selected characteristic frequency band after time-frequency domain distribution analysis as the terahertz spectrum feature of the test sample under unknown material mechanical conditions. Calculate the similarity between the terahertz spectrum feature of the reference sample and the terahertz spectrum feature of the test sample under unknown material mechanical conditions. Input the similarity calculation result corresponding to the test sample under unknown material mechanical conditions, i.e., the terahertz spectrum feature metric value of the sample, into the material mechanical property fast matching model. Use the output of the model to obtain the material mechanical property parameter measurement results of the test sample under unknown material mechanical conditions. The specific processing method for the time-frequency domain distribution analysis is selected according to the type of time-frequency transformation and the frequency dimension of the characteristic frequency band; The specific processing method for the similarity calculation is selected according to the principle of comprehensive global and detailed measurement, so that the monotonicity of the terahertz spectrum of the test sample under at least five successively increasing or decreasing material mechanical conditions, after signal similarity calculation and measurement, with the change of the corresponding material mechanical property parameters of the test sample, is maintained after the terahertz spectrum characteristics of the test sample are measured by similarity calculation. The rapid matching model for material mechanical properties specifically adopts a material mechanical property probability distribution model or a material mechanical property data fitting model based on the correlation between material terahertz spectrum characteristics and mechanical properties. Therefore, the probability distribution model is a trained probabilistic neural network.

2. The non-destructive testing method for material mechanical properties based on terahertz spectrum characteristic parameters according to claim 1, characterized in that: The selection principles for the time-frequency transformation types and frequency dimensions of the characteristic frequency bands in the time-frequency domain distribution analysis include: the monotonicity of the terahertz spectrum characteristic parameters of the test sample under at least five successively increasing or decreasing material mechanical conditions is the most significant as the corresponding material mechanical property parameters of the test sample change; among which, the measure index of the degree of monotonicity is the mean of the relative change rate of the terahertz spectrum characteristic parameters of the test sample under two adjacent material mechanical conditions, and the higher the mean value, the more significant the monotonicity is determined.

3. The non-destructive testing method for material mechanical properties based on terahertz spectrum characteristic parameters according to claim 2, characterized in that: The selection of the time-frequency transformation type and the frequency dimension of the characteristic frequency band in the time-frequency domain distribution analysis specifically includes the following steps: summing the time-frequency matrix obtained by the time-frequency domain distribution analysis of the terahertz spectrum of the test sample under each material mechanical condition to obtain the terahertz spectrum characteristic parameters of the test sample under each material mechanical condition. If the terahertz spectrum characteristic parameters of the test sample under different material mechanical conditions obtained by the summation calculation show a monotonically increasing or monotonically decreasing relationship with the change of the corresponding material mechanical property parameters of the test sample, then it is determined that the frequency dimension of the time-frequency transformation type and the characteristic frequency band can be used in the time-frequency domain distribution analysis.

4. The method for determining the mechanical properties of materials based on terahertz spectrum characteristic parameters according to claim 2, characterized in that: The time-frequency transform type is wavelet packet decomposition.

5. The non-destructive testing method for material mechanical properties based on terahertz spectrum characteristic parameters according to claim 1, characterized in that: The specific processing method for similarity calculation is dynamic time warping.

6. The non-destructive testing method for material mechanical properties based on terahertz spectrum characteristic parameters according to claim 1, characterized in that: During the training process, if the number of test samples under different material mechanical conditions is sufficient and reflects the changes in the material mechanical property parameters among the test samples on a scale no greater than 0.1, then the probability distribution model outputs the material mechanical property parameter 'a' according to the classification problem. i and the probability y of the corresponding material mechanical property parameters i Where i = 1, 2…n, and n is the number of some or all of the test samples under different material mechanical conditions, the determination results of the material mechanical property parameters of the test samples under unknown material mechanical conditions are determined according to the material mechanical property parameters corresponding to the maximum probability value of the output. That is, when the terahertz spectrum characteristics of the test samples under different material mechanical conditions are calculated without grouping similarity with the terahertz spectrum characteristics of the reference sample, the material mechanical property parameters corresponding to the maximum probability value of the output are taken as the determination results. However, when the terahertz spectrum characteristics of the test samples under different material mechanical conditions are calculated with the terahertz spectrum characteristics of the reference sample in a grouped similarity according to the frequency dimension, the material mechanical property parameters corresponding to the maximum probability values ​​of most outputs in the grouped frequency dimension and which are the same are taken as the determination results. Otherwise, the probability distribution model outputs the material mechanical property parameter 'a' according to the regression problem. i and the probability y of the corresponding material mechanical property parameters i Where i = 1, 2…n, and n is the number of some or all of the test samples under different material mechanical conditions, then, in the case of calculating the terahertz spectrum characteristics of the test samples under different material mechanical conditions and the terahertz spectrum characteristics of the reference sample without grouping, the measurement results of the material mechanical property parameters of the test samples under unknown material mechanical conditions are calculated according to the following formula: a 未知 =a1×y1+a2×y2+…a n ×y n 。

Images