Thermal diffusivity high-precision inversion method based on multi-terminal deep neural network

Abstract

The invention relates to a thermal diffusivity high-precision inversion method based on a multi-terminal deep neural network, belongs to the technical field of infrared thermal wave nondestructive testing and material thermophysical property measurement, and particularly relates to the thermal diffusivity high-precision inversion method based on the multi-terminal deep neural network. The invention aims to solve the problems of low calculation efficiency, strong noise sensitivity and insufficient robustness in a traditional thermal diffusivity measurement method, and provides a thermal diffusivity high-precision inversion method based on a multi-end deep neural network for materials with damage defects. Specifically, the amplitude and phase of thermal waves are extracted by using a phase-locked thermal imaging technology, and stable characteristics describing the thermal diffusion process are obtained. A deep learning network taking space coordinates, excitation frequency, surface temperature information, amplitude and phase as input is constructed, and efficient and accurate measurement of the thermal diffusivity of a complex material is realized through multi-modal data fusion, an advanced direct current component removal technology and a multi-end deep neural network architecture.

Description

Technical Field

[0001] The present invention belongs to the technical field of infrared thermal wave non-destructive testing and material thermal property measurement, and particularly relates to a high-precision inversion method for thermal diffusivity based on a multi-terminal deep neural network. Background Art

[0002] Thermal diffusivity is a key parameter describing the dynamic heat transfer characteristics of materials, and its magnitude directly affects the thermal behavior of materials in a dynamic temperature field. Therefore, accurately measuring thermal diffusivity is of great significance for the development of new materials, non-destructive testing, and industrial applications. Traditional thermal diffusivity measurement methods can be divided into three categories. The first is the steady-state method: such as the ASTM E-1225 plate method, which measures the thermal conductivity by establishing a temperature gradient, but is limited by long measurement time and heat loss. The second is the transient method: such as the laser flash method, which calculates the thermal diffusivity through transient temperature changes. Although it has high efficiency, it is easily affected by equipment response delay and noise. The third is the periodic method: such as lock-in thermography technology, which is based on the phase and amplitude characteristics of thermal waves and is suitable for measuring in-plane and out-of-plane thermal diffusivity of complex materials, but the accuracy depends on the signal processing quality and calculation efficiency. Although the above methods can achieve good results under certain conditions, these methods have the following problems:

[0003] 1. High computational cost: Traditional numerical methods rely on iterative optimization, have low computational efficiency, and are difficult to converge under complex boundary conditions.

[0004] 2. Strong noise sensitivity: Noise in experimental data easily affects the inversion result and reduces the measurement accuracy.

[0005] 3. Insufficient robustness: Existing methods are prone to large errors in prediction accuracy when dealing with material inhomogeneity, complex structures, and internal damage.

[0006] In recent years, the development of deep learning technology has provided new possibilities for solving complex physical problems. However, there is still little research on the inversion problem of thermal diffusivity with damage in existing deep learning, and there is a lack of an effective unified method for combining numerical simulation and experimental data. Summary of the Invention

[0007] The present invention aims to solve the problems of low computational efficiency, strong noise sensitivity, and insufficient robustness in traditional thermal diffusivity measurement methods, and proposes a high-precision inversion method for thermal diffusivity based on a multi-terminal deep neural network for materials containing damage defects. Specifically, the present invention uses lock-in thermography technology to extract the amplitude and phase of thermal waves and obtains stable features describing the thermal diffusion process. A deep learning network with spatial coordinates, excitation frequency, surface temperature information, amplitude, and phase as inputs is constructed, and through multi-modal data fusion, an advanced DC component removal technology, and a multi-terminal deep neural network architecture, efficient and accurate measurement of the thermal diffusivity of complex materials is realized.

[0008] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0009] A high-precision inversion method for thermal diffusivity based on a multi-terminal deep neural network, characterized by comprising the following processes:

[0010] S1: Numerically simulate the diffusive thermal wave signal with defects and establish a numerical model;

[0011] S2: Based on the numerical model obtained in S1, establish a numerical model data set;

[0012] S3: Based on the data obtained from the residual module and S2, construct a three-channel deep neural network for training and prediction;

[0013] S4: Based on the neural network constructed in S3, predict the thermal diffusivity of real materials under experimental conditions.

[0014] As a preferred solution, S1 of the present invention includes the following steps:

[0015] S1.1 Based on the Fourier heat conduction equation, establish a thermal wave model with defects and boundary conditions, and simulate the spatio-temporal characteristics of the thermal diffusion process of the material from the non-steady state to the steady state under different boundary conditions;

[0016] S1.2 Use the numerical solution of the finite element method to discretize the thermal wave model in the spatial domain and construct a spatial discretization grid model;

[0017] S1.3 Discretize the thermal wave model in the time domain to obtain the discrete time domain numerical model of the thermal wave model at the current time step.

[0018] As another preferred solution, in S1.2 of the present invention, the spatial discretization grid model selects structured quadrilateral grids and unstructured Delaunay triangular grids to divide the two-dimensional model; the unstructured Delaunay triangular grids are used to divide the geometric region to ensure the adaptability of complex boundaries;

[0019] As another preferred solution, in S1.3 of the present invention, the time domain discretization uses the numerical solution of the "fourth-order Runge-Kutta method" to integrate the time to ensure the high precision and numerical stability of the time step.

[0020] As another preferred solution, S2 of the present invention includes the following steps:

[0021] S2.1 Based on the numerical model obtained in S1, numerically simulate the thermal wave signals at different modulation frequencies to generate a large-scale initial thermal wave data set at different frequencies;

[0022] S2.2 adopts the non - linear fitting method to process the initial thermal wave data set generated in S2.1 to remove the DC component signal, and obtain the thermal wave signal after removing the DC component; effectively filter the DC component signal in the thermal wave signal, improve the accuracy of the thermal wave signal, and provide high - quality input for the subsequent deep learning model; the non - linear fitting method can significantly improve the accuracy of removing the DC component and improve the extraction effect of the amplitude and phase of small signals;

[0023] S2.3 Based on the lock - in thermography algorithm, extract the frequency - domain amplitude and phase information data of the thermal wave signal at different modulation frequencies for the thermal wave signal obtained in S2.2;

[0024] S2.4 Adopt the data augmentation method to introduce random noise into the data obtained in S2.3; improve the robustness of the model to experimental data;

[0025] S2.5 Perform decentralization and normalization processing on the amplitude and phase data in the frequency domain to obtain the final multi - modal large - scale training data set including spatial position, different modulation frequencies, time - domain surface temperature information, frequency - domain amplitude and phase information, and thermal physical parameters (density, specific heat capacity, thermal conductivity) of the material to be measured, and randomly extract a small batch of data sets as the validation set; eliminate the influence of the dimension difference of the thermal wave data eigenvalue on the subsequent deep neural network model.

[0026] As another preferred solution, in the present invention, in the S2.1, the time step of the numerical simulation calculation is from 1×10 -7 to 5×10 -7 .

[0027] As another preferred solution, in the present invention, in the S2.1, the frequency change range of the numerically simulated modulation heat source is from 0.01 Hz to 1 Hz.

[0028] As another preferred solution, in the present invention, in the S2.2, the non - linear fitting method numerical fitting model is constructed as:

[0029]

[0030] Among them, is the AC component, and the value of the constant is obtained by using the mini - batch gradient descent method.

[0031] As another preferred solution, in the present invention, in the S2.4, use various noise types such as Gaussian noise, impulse noise, and salt - and - pepper noise to simulate the noise environment in actual experiments by disturbing the simulated data, thereby improving the generalization ability of the model.

[0032] As another preferred solution, in the present invention, in the S2.4, the thermal wave response with random noise is constructed as follows:

[0033]

[0034] Among them, is the percentage of the random noise intensity, and the range of the random noise factor λ satisfies λ ∈ [0, 1].

[0035] As another preferred solution, S3 described in the present invention includes the following steps:

[0036] S3.1 Construct a three-channel deep learning model based on the residual module; use the spatial position coordinates generated by the numerical solution, modulation frequency, time-domain surface temperature information, frequency-domain amplitude and phase information, and the thermal physical parameters of the material itself as input features, and the thermal diffusivity as the output target.

[0037] S3.2 Train the three-channel deep learning model network. When the prediction error of the network model tends to be stable, verify the prediction accuracy of the model through experimental data.

[0038] As another preferred solution, in S3.1 of the present invention, the first branch inputs the surface temperature information in the time domain to capture the time characteristics of material heat diffusion; the second branch inputs the amplitude and phase information in the frequency domain to extract the frequency domain characteristics of the material; the third branch inputs the physical property parameters (density, specific heat capacity, thermal conductivity), and extracts the physical property characteristics through the fully connected layer.

[0039] As another preferred solution, in S3.1 of the present invention, each branch network uses the residual module for feature extraction, and adopts skip connection to alleviate the gradient vanishing problem in the deep network, accelerate convergence and improve the accuracy of the model.

[0040] As another preferred solution, in S3.1 of the present invention, after the output of each branch network, the weighted average method is adopted for feature fusion; the feature vectors of each branch are weighted and combined according to the dynamically calculated weights to generate the final fused feature vector; in the feature fusion process, the position attention module and the channel attention module are introduced to weight and adjust the features in the spatial dimension and channel dimension respectively; the position attention module focuses on the importance of features at different spatial positions, and the channel attention module focuses on the relationship between different channels.

[0041] As another preferred solution, in S3.2 of the present invention, the particle swarm optimization algorithm is used to optimize the initial weights and bias values during the training process; the Adam training algorithm is used for adaptive learning rate during the training process, and the training speed is increased.

[0042] Secondly, S4 described in the present invention includes the following steps:

[0043] S4.1 Adopt the lock-in thermography method with a periodically modulated excitation heat source to collect the time-domain surface temperature information data at different frequencies;

[0044] S4.2 Based on S2.2, remove the DC component signal in the experimental data, and based on S2.3, obtain the amplitude and phase information data in the frequency domain;

[0045] S4.3 The processed experimental data is input into the multi-terminal neural network that has been trained in S3 to obtain the predicted thermal diffusivity result of the material to be tested.

[0046] In addition, in the S4.1 of the present invention, the thermal excitation signal of the lock-in thermography method adopts a periodic square wave with a 50% duty cycle. For the periodic thermal excitation signal, compared with the sine and cosine waveforms, the periodic square wave excitation signal with a 50% duty cycle can effectively improve the signal-to-noise ratio.

[0047] Advantages of the present invention.

[0048] 1. The present invention combines the numerical model with the deep neural network for the first time. The model can cope with the challenges under different materials and different environmental conditions, provide a more accurate solution for measuring thermal physical properties, and achieve high-precision prediction of the thermal diffusivity of defective media.

[0049] 2. The present invention introduces a data augmentation method to perturb the numerical simulation data, enhancing the model's adaptability to noise in the experimental data, thereby improving the stability and generalization ability of the model in different experimental environments.

[0050] 3. The non-linear fitting method for removing the DC component proposed by the present invention can effectively filter out the DC component in the thermal wave signal, provide a more accurate result of the AC signal component, provide more accurate input features for the subsequent deep learning model, and improve the overall prediction performance.

[0051] 4. Through multi-modal data fusion (time-domain surface temperature information, frequency-domain amplitude and phase information, and thermal physical properties parameters) of the present invention, the advantages of each type of data can be fully utilized, improving the accuracy of thermal diffusivity prediction. Especially when dealing with complex materials (including defects), the thermal characteristic information of the materials can be effectively extracted.

[0052] 5. The large-scale data set generated by numerical simulation combined with actual experimental data helps the model to adapt to more diverse test conditions and material types, reduces the model's dependence on specific data sets, and enhances the robustness in practical applications. Description of the Drawings

[0053] The present invention will be further described below in conjunction with the drawings and specific embodiments. The protection scope of the present invention is not limited only to the description of the following content.

[0054] Appendix Figure 1 : Overall flowchart of the method proposed by the present invention.

[0055] Appendix Figure 2 : Thermal wave signal diagram after DC component removal processing.

[0056] Appendix Figure 3 : (a) Amplitude and (b) phase diagrams in space of a defective numerical model at a frequency of 0.02 Hz.

[0057] Appendix Figure 4 : Schematic diagram of the multi-terminal deep neural network model architecture.

[0058] Appendix Figure 5 : Error result diagram after the model is trained to stability.

[0059] Appendix Figure 6 : Thermal diffusivity verification result diagram after the experimental data with defects is input into the model.

[0060] Appendix Figure 7 : System block diagram of the high-precision thermal diffusivity inversion method proposed by the present invention. Detailed implementation manners

[0061] To better illustrate the technical solution of the present invention, the technical solution of the present invention will be further described below in conjunction with the accompanying drawings, but it is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention shall be covered within the protection scope of the present invention.

[0062] Figure 1 is the overall flowchart of the detailed implementation manners of a high-precision thermal diffusivity inversion method based on a multi-terminal deep neural network provided by the present invention. As Figure 1 shown, a high-precision thermal diffusivity inversion method based on a multi-terminal deep neural network in this embodiment, for a metal powder compacted thin plate with defects, based on the lock-in thermography measurement method, is carried out according to the following steps:

[0063] S1: Numerical model of the diffusive thermal wave signal of a two-dimensional model of a metal powder compacted thin plate with defects:

[0064] S1.1 Based on the Fourier heat conduction equation, construct a thermal wave model with defects and define the boundary conditions of the sample during the heat conduction process to further simulate the spatio-temporal characteristics of the material from the non-steady state to the transient heat diffusion process under different boundary conditions.

[0065] Before establishing the thermal wave numerical model, it is necessary to precisely define the geometric parameters of the metal powder compacted thin plate sample: such as depth (length), width, and height. Specifically, the geometric parameters of the two-dimensional model of the metal powder compacted thin plate in this embodiment are depth (length) L = 20 mm, height H = 4 mm, and the geometric parameters of the internal defect size are depth (length) L N = 6 mm, height H N = 1 mm.

[0066] Furthermore, considering the density of the metal powder compacted thin plate ρ = 7150 kg / m 3 , specific heat capacity C p = 900 J / kg·K, thermal diffusivity α xy = 4.33×10 -6 m 2 / s. The thermal diffusivity of the damage α N xy = 6.33×10 -6 m 2 / s.

[0067] S1.2 Based on the principle of virtual work, the two-dimensional thermal wave model established in step S1.1 can be further expressed as a thermal wave model related to the boundary conditions. Using the numerical solution finite element method to discretize the thermal wave model in the spatial domain and construct a spatial discrete grid model, the matrix numerical solution of the thermal wave model after spatial discretization can be obtained:

[0068]

[0069] Among them, is the temperature vector, is the derivative vector of temperature with respect to time, G is the stiffness matrix, and F is the thermal load matrix.

[0070] Furthermore, the spatial discrete grid model can select a structured quadrilateral grid or an unstructured Delaunay triangular grid to divide the two-dimensional model. Specifically, through calculation, it is found that when the number of grids in this embodiment is selected as 1500, using the unstructured Delaunay triangular grid to divide the geometric region has better convergence and can ensure the adaptability of complex boundaries.

[0071] When discretizing the thermal wave model in the time domain, the "fourth-order Runge-Kutta method" is used for numerical integration of time to ensure the high precision and numerical stability of the time step, and the discrete time domain numerical solution of the thermal wave model at the current time step is obtained:

[0072]

[0073] where the superscript t+Δt represents the current time step. are the coefficients of the fourth-order Runge-Kutta method, and they have the following relationship with the matrix numerical solution after spatial discretization of the heat wave model:

[0074]

[0075] Establishment of the S2 numerical model dataset:

[0076] S2.1 Based on the numerical solution of the discrete domain of the two-dimensional model already established in S1, under this condition, numerical simulations of heat wave signals at different modulation frequencies are carried out, and large-scale initial heat wave dataset results are generated. The frequency change range of the modulated heat source in this embodiment is from 0.01 Hz to 1 Hz, and the time step of numerical calculation can generally be selected as 1×10 -7 to 5×10 -7 , in the current embodiment, the step size is 5×10 -7 which can meet the calculation accuracy of the model.

[0077] S2.2 In order to effectively filter out the DC component signals in each group of heat wave signal data obtained in S2.1, a DC component removal signal fitting model is further constructed. Further, in this embodiment, a non-linear fitting method is used to process the DC component removal of the heat wave signal.

[0078] Among them, the non-linear fitting method proposed by the present invention can significantly improve the DC component removal accuracy and improve the extraction effect of the small signal amplitude and phase. The numerical fitting model of the non-linear fitting method is constructed as:

[0079]

[0080] Among them, is the AC component, and the value of the constant is obtained using the mini-batch gradient descent method. In this embodiment, the calculation results are 33.7349, 0.01315, 111.8414.

[0081] As Figure 2 shown, after removing the DC component signals from the heat wave data using the non-linear fitting method proposed by the present invention, the resulting AC component amplitude curve results. Obviously, the AC component performance effect obtained by the non-linear fitting method adopted in this embodiment is very stable.

[0082] S2.3 After removing the DC component signals in each group of heat wave signal data at different modulation frequencies in S2.2, based on the lock-in thermography algorithm, the frequency domain amplitude and phase information data of the DC component removed heat wave signals at different modulation frequencies are extracted. As Figure 3As shown, in this embodiment, the (a) amplitude and (b) phase diagrams in space of the defective numerical model at a modulation frequency of 0.02 Hz are shown after removing the DC component information.

[0083] S2.4 To further improve the robustness of the model to experimental data, a data augmentation method is adopted to introduce random noise into the data. Multiple types of random noise such as Gaussian noise, impulse noise, and salt-and-pepper noise are used to simulate the noise environment in actual experiments by perturbing the simulated data, thereby enhancing the generalization ability of the model.

[0084] Furthermore, in this embodiment, random noise with noise intensities of 1%, 5%, 10%, and 15% is considered in the thermal wave response. The thermal wave response with random noise can be constructed as follows:

[0085]

[0086] where is the percentage of the random noise intensity, and the range of the random noise factor λ satisfies λ ∈ [0, 1].

[0087] S2.5 To eliminate the influence of the dimensional difference of the thermal wave data eigenvalues on the subsequent deep neural network model, the amplitude and phase data in the frequency domain are de-centered and normalized to obtain a final multi-modal large-scale training dataset containing spatial position, different modulation frequencies, surface temperature information in the time domain, amplitude and phase information in the frequency domain, and thermal physical parameters (density, specific heat capacity, thermal conductivity) of the material to be measured, and a small batch of datasets are randomly extracted as the validation set.

[0088] Furthermore, the value ranges of the training set parameters are as follows. The value range of the spatial position satisfies x ∈ (2, 20] mm; y ∈ (0, 4] mm; the modulation frequency range is [0.01, 1] Hz; the thermal diffusivity is set to α xy ∈ (0, 12] × 10 -6 m 2 / s.

[0089] S3 Construction, training and prediction of a three-channel deep neural network based on the residual module:

[0090] S3.1 As Figure 4As shown in the figure, a three-channel deep learning model based on a residual module is constructed. The spatial coordinates, modulation frequency, time-domain surface temperature information, frequency-domain amplitude and phase information generated by the numerical solution, and the thermal physical parameters of the material itself are used as input features, and the thermal diffusivity is used as the output target. Further, the first branch inputs the surface temperature information in the time domain to capture the time characteristics of the material's heat diffusion; the second branch inputs the amplitude and phase information in the frequency domain to extract the frequency-domain characteristics of the material; the third branch inputs the physical property parameters (density, specific heat capacity, thermal conductivity), and extracts the physical property characteristics through a fully connected layer. Each branch network uses a residual module for feature extraction, and skip connections are used to alleviate the vanishing gradient problem in the deep network, accelerate convergence, and improve the accuracy of the model. To standardize the numerical values after convolution, a normalization layer is added after each convolutional layer. The pooling method is selected as max pooling. At the same time, to reduce the impact of overfitting on the prediction model, a Dropout layer is added to the model, and the probability of the Dropout layer is set to 0.2.

[0091] In this embodiment, the number of convolutional layers for each branch is constructed as 15 layers, the convolutional kernel sizes are [2,1], [3,1], and [5,1], the number of convolutional kernels per layer is 18, 36, and 72, the convolutional stride is 1, and the padding model is zero padding. After the output of each branch network, a weighted average method is used for feature fusion. The feature vectors of each branch are weighted and combined according to the dynamically calculated weights to generate the final fused feature vector. Further, in the feature fusion process, a position attention module and a channel attention module are introduced to perform weighted adjustment on the features in the spatial dimension and the channel dimension respectively. The position attention module focuses on the importance of features at different spatial positions, and the channel attention module focuses on the relationship between different channels.

[0092] Finally, the spatial coordinates, modulation frequency, amplitude, and phase generated by the numerical solution are used as input features to input into the deep neural network, and the thermal diffusivity is used as the final output target. In this embodiment, the number of training set samples is 30000, and the number of test set samples is 15 groups, where the value of y is fixed at 2 mm; the value range of x is [1,15] mm, and the increment is 1 mm.

[0093] S3.2 Train the deep neural network model, use the Adam training algorithm for adaptive learning rate, and increase the training speed. When the prediction error of the network model tends to be stable, verify the prediction accuracy of the model through experimental data. During the training process, the particle swarm optimization algorithm is used to optimize the initial weights and bias values of the neural network model. In this embodiment, the boundary values of the initial weights and bias values satisfy [-1,1]; the global and individual learning factors are set to 3; the number of population updates is 30 times; the population size is 15.

[0094] Figure 5In this embodiment, the fitting trend of the multi-terminal deep neural network under a large number of training samples is provided, and the root mean square error (RMSE) algorithm is used to evaluate the matching degree between the true value and the predicted value. It can be found that the RMSE and the loss function gradually tend to be stable as the number of iteration steps increases. The stabilized RMSE is 0.82813.

[0095] Prediction of the thermal diffusivity of the real material under S4 experimental conditions

[0096] S4.1 The lock-in thermography technology using a periodic modulation square wave excitation heat source with a 50% duty cycle is adopted to collect the time-domain surface temperature information data of the unsintered metal powder compact at different frequencies. The modulation frequencies selected in this embodiment are 0.01 Hz, 0.02 Hz, and 0.05 Hz. In addition, the thermal wave of the unsteady response has a non-linear DC component and is dependent on space. Therefore, analyzing the amplitude and phase based on the steady-state thermal wave field can further save the calculation cost. In this embodiment, the duration of the thermal load application is set to 1300 s, the sampling frame rate of the mid-infrared camera is 15 Hz, and the sampling starts after 1000 s.

[0097] The collected data removes the DC component signal in the experimental data based on S2.2 and obtains the amplitude and phase information data in the frequency domain based on S2.3.

[0098] The processed experimental data is input into the multi-terminal deep neural network that has been trained in S3 to obtain the predicted thermal diffusivity results of the unsintered metal powder compact material to be measured.

[0099] As Figure 6 is the result graph of verifying the prediction accuracy of this model through experimental data provided in this embodiment. The test sample uses an unsintered metal powder compact, and its length, width, and height are 31 mm, 12 mm, and 6 mm respectively. The experimental data of the distribution of the thermal wave amplitude and phase along the space under the modulation frequencies of 0.01 Hz, 0.02 Hz, and 0.05 Hz are selected, and the coordinate information, amplitude, and phase values of 15 pixels are randomly selected within the sample range as the input parameters of the deep neural network for training. The prediction results are as Figure 6 shown. At positions far from the damaged area, the predicted values in the healthy area are close to the true values; as the test position gradually approaches the damaged area, the predicted values of the thermal diffusivity gradually tend to the true values in the damaged area. By comparing the prediction results of the two areas, it can be found that the prediction results in the healthy area are significantly affected by the damaged area. This is because the thermal diffusivity in the damaged area is greater than that in the healthy area, so the diffusion of the thermal wave is mainly dominated by the damaged area with high thermal diffusivity. From Figure 6 it can be found that the prediction results have fluctuations, and their average value is 5.2×10 -6 m 2 / s. Comparing with the measured values of thermal diffusivity in existing literature, the predicted values obtained by the current model are within the same range as the values reported in the literature.

[0100] Corresponding to the embodiments of the foregoing method, the present invention also provides an embodiment of a high-precision inversion method system for thermal diffusivity based on a multi-terminal deep neural network.

[0101] Figure 7 It is a block diagram of a high-precision inversion method system device for thermal diffusivity based on a multi-terminal deep neural network shown in the embodiments of the present invention. Please refer to Figure 7 , this system device implements the technical solution of the present invention in a software manner, and this device includes:

[0102] Numerical simulation module 701: used to simulate and generate the required time-domain numerical model data set. Design a two-dimensional numerical model and set boundary conditions from the parameters of the material to be measured, further divide the grid of the two-dimensional model, set the time step of numerical calculation, and calculate and generate the original data set of surface temperature signals at different modulation frequencies;

[0103] Data preprocessing module 702: used to remove the DC component and generate frequency-domain amplitude and phase data. The original data set obtained from the numerical simulation module 701 is input into the data preprocessing module 702, and the data is processed to remove the DC component using a non-linear fitting method to obtain stable AC component data. Based on the lock-in thermography algorithm, the amplitude and phase results of the AC component data are calculated in the frequency domain to obtain the amplitude and phase information at different modulation frequencies;

[0104] Data enhancement module 703: used to introduce random noises with different intensities into the data set data. In order to further improve the robustness of the multi-terminal deep neural network model to experimental data, the data processed by the data preprocessing module 702 is input into the data enhancement module 703, and random noises with noise intensities of 1%, 5%, 10%, and 15% are considered in the thermal wave response data. The present invention uses various types of random noises such as Gaussian noise, impulse noise, and salt-and-pepper noise to simulate the noise environment in actual experiments by perturbing the simulation data, thereby improving the generalization ability of the model;

[0105] Multi-terminal neural network module 704: Input the generated large-scale multi-modal dataset, train the network, extract feature information and fuse it, and output the predicted result of the thermal diffusivity of the material. After being processed by the numerical simulation module 701, the data preprocessing module 702, and the data augmentation module 703, a large-scale multi-modal data including time-domain surface temperature information data, frequency-domain amplitude and phase information data, spatial coordinates, modulation frequency, and thermal physical parameters of the material has been generated. Input it into the multi-terminal neural network module 704 for network training. When the multi-terminal deep neural network model is trained to a stable state, output the predicted result value of the thermal diffusivity of the material;

[0106] Experimental data verification module 705: Used to input the experimental data of the actual material to be measured in the real environment into the trained multi-terminal deep neural network module to generate the predicted result of the thermal diffusivity of the material.

[0107] In summary, this embodiment confirms the effectiveness of the method for predicting the thermal diffusivity of materials by constructing a multi-terminal deep neural network. The thermal diffusivity coefficient of the unsintered metal powder compact obtained in this embodiment is within the same numerical range as the existing result value of this material, which confirms the innovation and effectiveness of the high-precision inversion method for thermal diffusivity based on a multi-terminal deep neural network proposed by the present invention.

[0108] The present invention is particularly suitable for high-precision prediction of thermal diffusivity in complex media and can be widely applied to fields such as non-destructive testing, thermal physical property analysis of materials, and industrial production process monitoring.

[0109] Although the present invention includes specific implementation details, these should not be construed as limiting the scope of any invention or the scope of the claimed invention, but are mainly used to describe the features of specific embodiments of a particular invention. Certain features described in the embodiments of the present invention can also be combined and implemented in a single embodiment. On the other hand, various features described in a single embodiment can also be separately implemented in multiple embodiments or implemented in any suitable sub-combination. In addition, although the features can function in certain combinations as described above, one or more features from the claimed combination can be removed from the combination in some cases, and the claimed combination can refer to a sub-combination or a variant of the sub-combination.

[0110] Similarly, although the operations are depicted in a specific order in the drawings, this should not be construed as requiring these operations to be performed in the specific order shown or sequentially, or requiring all of the illustrated operations to be performed to achieve the desired result. In some cases, multi-tasking and parallel processing may be advantageous. In addition, the separation of various system modules and components in the above embodiments should not be construed as required in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0111] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A high-precision inversion method for thermal diffusivity based on a multi-terminal deep neural network, characterized in that: The following processes are included: S1: Numerical simulation of diffusion heat wave signals containing defects and establishment of numerical model; S2: Based on the numerical model obtained in S1, establish the numerical model data set; S3: Based on the residual module and the data obtained in S2, a three-channel deep neural network is constructed for training and prediction; S4: Based on the neural network constructed in S3, the thermal diffusivity of real materials under experimental conditions is predicted.

2. A high-precision inversion method for thermal diffusivity based on a multi-terminal deep neural network according to claim 1, characterized in that The S1 comprises the following steps: S1.1 Based on the Fourier heat conduction equation, a thermal wave model with defects and boundary conditions are established to simulate the spatiotemporal characteristics of the material's thermal diffusion process from non-steady state to steady state under different boundary conditions; S1.2 Use the numerical finite element method to discretize the thermal wave model in the spatial domain and construct a spatial discrete grid model; S1.3 discretizes the thermal wave model in the time domain to obtain a discrete time domain numerical model of the thermal wave model at the current time step.

3. A high-precision inversion method for thermal diffusivity based on a multi-terminal deep neural network according to claim 1, characterized in that The S2 comprises the following steps: S2.1 Based on the numerical model obtained in S1, the thermal wave signals at different modulation frequencies are numerically simulated to generate large-scale initial thermal wave data sets at different frequencies; S2.2 uses a nonlinear fitting method to remove the DC component signal from the initial thermal wave data set generated by S2.1 to obtain the thermal wave signal after the DC component is removed; S2.3 extracts the frequency domain amplitude and phase information data of the thermal wave signal under different modulation frequencies from the thermal wave signal obtained in S2.2 based on the phase-locked thermal imaging algorithm; S2.4 uses data enhancement method to introduce random noise into the data obtained in S2.3; S2.5 decentralized and normalized the amplitude and phase data in the frequency domain to obtain the final multimodal large-scale training data set containing spatial position, different modulation frequencies, time domain surface temperature information, frequency domain amplitude and phase information, and thermal parameters of the material to be tested, and randomly extracted small batch data sets as verification sets.

4. A high-precision inversion method for thermal diffusivity based on a multi-terminal deep neural network according to claim 3, characterized in that In S2.2, the nonlinear fitting method numerical fitting model is constructed as follows: in, is the AC component, constant The value of is obtained using mini-batch gradient descent.

5. A high-precision inversion method for thermal diffusivity based on a multi-terminal deep neural network according to claim 3, characterized in that In S2.4, the thermal wave response with random noise is constructed as follows: in, is the percentage of random noise intensity, and the range of random noise factor λ satisfies λ∈[0,1].

6. A high-precision inversion method for thermal diffusivity based on a multi-terminal deep neural network according to claim 1, characterized in that The S3 comprises the following steps: S3.1 builds a three-channel deep learning model based on the residual module; the spatial position coordinates, modulation frequency, time domain surface temperature information, frequency domain amplitude and phase information and the thermal parameters of the material itself generated by the numerical solution are used as input features, and the thermal diffusivity is used as the output target; S3.2 trains the three-channel deep learning model network, and when the prediction error of the network model tends to be stable, verifies the prediction accuracy of the model through experimental data.

7. A high-precision inversion method for thermal diffusivity based on a multi-terminal deep neural network according to claim 6, characterized in that In S3.1, the first branch inputs the surface temperature information in the time domain to capture the time characteristics of the thermal diffusion of the material; the second branch inputs the amplitude and phase information in the frequency domain to extract the frequency domain characteristics of the material; The third branch inputs physical property parameters and extracts physical features through the fully connected layer.

8. A high-precision inversion method for thermal diffusivity based on a multi-terminal deep neural network according to claim 6, characterized in that In S3.1, each branch network uses a residual module for feature extraction and adopts jump connections to alleviate the gradient vanishing problem in the deep network, accelerate convergence and improve the accuracy of the model.

9. A high-precision inversion method for thermal diffusivity based on a multi-terminal deep neural network according to claim 6, characterized in that In S3.1, after the output of each branch network, a weighted average method is used to perform feature fusion; the feature vectors of each branch are weighted merged according to the dynamically calculated weights to generate a final fused feature vector; In the feature fusion process, the position attention module and channel attention module are introduced to perform weighted adjustments on the features of the spatial dimension and channel dimension respectively; the position attention module focuses on the importance of features at different spatial positions, and the channel attention module focuses on the relationship between different channels.

10. A high-precision inversion method for thermal diffusivity based on a multi-terminal deep neural network according to claim 1, characterized in that The S4 comprises the following steps: S4.1 uses a phase-locked thermal imaging method with a periodically modulated excitation heat source to collect time-domain surface temperature information data at different frequencies; S4.2 removes the DC component signal in the experimental data based on S2.2, and obtains the amplitude and phase information data in the frequency domain based on S2.3; The processed experimental data in S4.3 are input into the trained multi-terminal neural network in S3 to obtain the predicted thermal diffusivity result of the material to be tested.

Images