A high-precision inversion method for thermal diffusivity based on a multi-terminal deep neural network

Abstract

The application relates to the technical field of infrared thermal wave nondestructive testing and material thermal physical property measurement, in particular to a high-precision thermal diffusivity inversion method based on a multi-terminal deep neural network. The application aims to solve the problems of low calculation efficiency, strong noise sensitivity and insufficient robustness in the traditional thermal diffusivity measurement method, and proposes a high-precision thermal diffusivity inversion method based on a multi-terminal deep neural network for materials containing damage defects. Specifically, the application uses a phase-locked thermal imaging technology to extract the amplitude and phase of the thermal wave, and obtains stable characteristics describing the thermal diffusion process. A deep learning network is constructed with spatial coordinates, excitation frequency, surface temperature information, amplitude and phase as inputs. Through multi-modal data fusion, advanced direct current component removal technology and multi-terminal deep neural network architecture, efficient and accurate measurement of the thermal diffusivity of complex materials is realized.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of infrared thermal wave nondestructive testing and material thermal physical property measurement, and particularly relates to a high-precision thermal diffusivity inversion method based on a multi-end deep neural network. BACKGROUND

[0002] Thermal diffusivity is a key parameter that describes the dynamic heat transfer characteristics of materials, and its size directly affects the thermal behavior of materials in a dynamic temperature field. Therefore, accurate measurement of thermal diffusivity is of great significance for the development of new materials, nondestructive 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 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 thermal diffusivity from transient temperature changes, although it is more efficient, but is easily affected by device response delay and noise. The third is the periodic method: such as the lock-in thermography technique, 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 its precision depends on the quality of signal processing 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, which is low in computational efficiency, and is difficult to converge under complex boundary conditions.

[0004] 2. Strong noise sensitivity: Noise in experimental data can affect the inversion results and reduce measurement accuracy.

[0005] 3. Lack of robustness: Existing methods are prone to large errors in prediction accuracy when dealing with material heterogeneity and complex structures, as well as internal damage.

[0006] In recent years, the development of deep learning technology has provided new possibilities for solving complex physical problems. However, existing deep learning research on thermal diffusivity inversion with damage is still relatively limited, and there is a lack of unified methods that effectively combine numerical simulation and experimental data. SUMMARY

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

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

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

[0010] S1: Numerical simulation of a diffusion thermal wave signal containing defects to establish a numerical model;

[0011] S2: Based on the numerical model obtained in S1, a numerical model dataset is established;

[0012] S3: Based on the residual module and the data obtained in S2, a three-channel deep neural network is constructed, trained and predicted;

[0013] S4: Based on the neural network constructed in S3, the real material thermal diffusivity under experimental conditions is predicted.

[0014] As a preferred scheme, the S1 of the present application comprises the following steps:

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

[0016] S1.2 The spatial domain of the thermal wave model is discretized using a numerical solution finite element method to construct a spatial discrete grid model;

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

[0018] As another preferred scheme, in S1.2 of the present application, the spatial discrete grid model is selected from structured quadrilateral grid and unstructured Delaunay triangular grid to divide the two-dimensional model; unstructured Delaunay triangular grid is used to divide the geometric region to ensure the adaptability of complex boundaries;

[0019] As another preferred scheme, in S1.3 of the present application, the time domain is discretized by using a numerical solution "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 scheme, the S2 of the present application comprises the following steps:

[0021] S2.1 Based on the numerical model obtained in S1, the thermal wave signals under different modulation frequencies are numerically simulated to generate a large-scale initial thermal wave dataset under different frequencies;

[0022] S2.2, a nonlinear fitting method is used to remove the DC component signal from the initial thermal wave data set generated in S2.1, to obtain the thermal wave signal after removing the DC component; the DC component signal in the thermal wave signal is effectively filtered out, the accuracy of the thermal wave signal is improved, and high-quality input is provided for the subsequent deep learning model; the nonlinear fitting method can significantly improve the DC component removal accuracy and improve the extraction effect of small signal amplitude and phase;

[0023] S2.3, based on the phase-locked thermal imaging algorithm, the thermal wave signal obtained in S2.2 is used to extract the frequency domain amplitude and phase information data of the thermal wave signal under different modulation frequencies;

[0024] S2.4, a data enhancement method is used to introduce random noise in the data obtained in S2.3; the robustness of the model to experimental data is improved;

[0025] S2.5, the amplitude and phase data in the frequency domain are decentered and normalized to obtain a multi-modal 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 (density, specific heat capacity, thermal conductivity) of the material to be measured, and a small batch of data sets are randomly extracted as a validation set; the influence of the dimensionless difference of the thermal wave data eigenvalue on the subsequent deep neural network model is eliminated.

[0026] As another preferred scheme, in the S2.1 of the present application, the time step of the numerical simulation calculation is 1x10 -7 to 5x10 -7 .

[0027] As another preferred scheme, in the S2.1 of the present application, the frequency variation range of the numerical simulation modulation heat source is 0.01Hz to 1Hz.

[0028] As another preferred scheme, in the S2.2 of the present application, the numerical fitting model of the nonlinear fitting method is constructed as follows:

[0029]

[0030] wherein, is an alternating component, the constant The value of the constant is obtained using the small batch gradient descent method.

[0031] As another preferred scheme, in the S2.4 of the present application, Gaussian noise, impulse noise, and salt and pepper noise are used to simulate the noise environment in actual experiments by perturbing the simulation data, thereby improving the generalization ability of the model.

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

[0033]

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

[0035] As another preferred solution, the S3 of the present application comprises the following steps:

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

[0037] S3.2 training the three-channel deep learning model network, and when the prediction error of the network model tends to be stable, verifying the prediction accuracy of the model through experimental data.

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

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

[0040] As another preferred solution, in the S3.1 of the present application, 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; in the feature fusion process, a position attention module and a channel attention module are introduced to weight and adjust the features in the spatial dimension and the channel dimension, respectively; the position attention module focuses on the importance of features in different spatial positions, and the channel attention module focuses on the relationship between different channels.

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

[0042] Secondly, the S4 of the present application comprises the following steps:

[0043] S4.1 adopts a phase-locked thermography mode with a periodically modulated heat source to collect time-domain surface temperature information data at different frequencies;

[0044] S4.2 removes the direct current 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;

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

[0046] In addition, in S4.1, the heat excitation signal of the phase-locked thermography mode adopts a periodic square wave with a duty cycle of 50%. Compared with the sine and cosine wave, the periodic square wave excitation signal with a duty cycle of 50% can effectively improve the signal-to-noise ratio.

[0047] Advantages of the present application.

[0048] 1. The present application first combines numerical models with deep neural networks, and the model can cope with different materials and different environmental conditions, providing a more accurate thermal property measurement solution and realizing high-precision prediction of the thermal diffusivity of a defective medium.

[0049] 2. The present application introduces a data enhancement method to perturb the numerical simulation data, enhances the model's adaptability to noise in experimental data, and improves the stability and generalization ability of the model in different experimental environments.

[0050] 3. The nonlinear fitting direct current component removal processing method proposed in the present application can effectively filter out the direct current component in the thermal wave signal, provide more accurate alternating current signal component results, and provide more accurate input features for the subsequent deep learning model, improving the overall prediction performance.

[0051] 4. The present application can fully utilize the advantages of each type of data by fusing multi-modal data (time-domain surface temperature information, frequency-domain amplitude and phase information, and thermal property parameters), improving the accuracy of thermal diffusivity prediction, and effectively extracting the thermal characteristic information of the material, especially when dealing with complex materials (containing defects).

[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. BRIEF DESCRIPTION OF DRAWINGS

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

[0054] Figure 1 shows a schematic diagram of a thermal wave signal. Figure 1 Figure 2 shows a schematic diagram of a method according to the present application.

[0055] Figure 3 shows a schematic diagram of a thermal wave signal after DC removal. Figure 2 Figure 4 shows a schematic diagram of (a) amplitude and (b) phase plots of a numerical model of a defective sample at a frequency of 0.02 Hz.

[0056] Figure 5 shows a schematic diagram of a multi-terminal deep neural network model architecture. Figure 3 Figure 6 shows a schematic diagram of error results after model training to stability.

[0057] Figure 7 shows a schematic diagram of a thermal diffusivity verification result plot after inputting experimental data into the model. Figure 4 Figure 8 shows a schematic diagram of a system block diagram of a high-precision inversion method of thermal diffusivity according to the present application.

[0058] Figure 9 shows a schematic diagram of a specific implementation of a high-precision inversion method of thermal diffusivity based on a multi-terminal deep neural network according to the present application. Figure 5 Figure 10 shows a schematic diagram of error results after model training to stability.

[0059] Figure 11 shows a schematic diagram of a thermal diffusivity verification result plot after inputting experimental data into the model. Figure 6 Figure 12 shows a schematic diagram of a system block diagram of a high-precision inversion method of thermal diffusivity according to the present application.

[0060] DETAILED DESCRIPTION Figure 7 In order to better illustrate the technical solutions of the present application, the technical solutions of the present application will be further described below in conjunction with the accompanying drawings, but are not limited thereto, and any modifications or equivalent replacements to the technical solutions of the present application without departing from the spirit and scope of the technical solutions of the present application shall be encompassed within the protection scope of the present application.

[0061] Figure 1 shows a schematic diagram of a thermal wave signal.

[0062] Figure 2 shows a schematic diagram of a method according to the present application. Figure 1 Figure 3 shows a schematic diagram of a thermal wave signal after DC removal. Figure 1 Figure 4 shows a schematic diagram of (a) amplitude and (b) phase plots of a numerical model of a defective sample at a frequency of 0.02 Hz.

[0063] S1: A numerical model of a two-dimensional model of a defective metal powder compacted sheet:

[0064] S1.1 Based on the Fourier heat conduction equation, a thermal wave model containing defects is constructed, and the boundary conditions of the sample in the heat conduction process are defined to further simulate the space-time characteristics of the material from the non-steady state to the transient thermal diffusion process under different boundary conditions.

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

[0066] Furthermore, considering the density ρ of the compacted metal powder sheet is 7150 kg / m³ 3 Specific heat capacity C p = 900 J / kg·K, thermal diffusivity α xy =4.33×10 -6 m 2 / s. Thermal diffusivity α of the damage N xy =6.33×10 -6 m 2 / s.

[0067] Based on the principle of virtual work, the two-dimensional heat wave model established in step S1.1 can be further represented as a heat wave model related to boundary conditions. Using the numerical solution finite element method, the heat wave model is spatially discretized, and a spatially discrete mesh model is constructed. This yields the numerical solution of the heat wave model's matrix after spatial discretization.

[0068]

[0069] in, It is a temperature vector. It is the temperature derivative vector with respect to time, G is the stiffness matrix, and F is the thermal load matrix.

[0070] Furthermore, the spatial discrete mesh model can be divided into a structured quadrilateral mesh or an unstructured Delaunay triangular mesh to partition the two-dimensional model. Specifically, calculations have shown that when the number of meshes in this embodiment is selected as 1500, using an unstructured Delaunay triangular mesh to partition the geometric region has better convergence and can ensure adaptability to complex boundaries.

[0071] S1.3 When discretizing the heat wave model in the time domain, the fourth-order Runge-Kutta method is used to integrate over time to ensure high accuracy and numerical stability of the time step, thus obtaining the discrete-time numerical solution of the heat wave model at the current time step:

[0072]

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

[0074]

[0075] Establishment of the S2 numerical model dataset:

[0076] S2.1 The discrete-domain numerical solution of the two-dimensional model has been established in S1. Based on this condition, numerical simulations are performed on the thermal wave signals under different modulation frequencies, and a large-scale initial thermal wave dataset is generated. In this embodiment, the frequency variation range of the modulated heat source is 0.01Hz to 1Hz, and the time step for numerical calculation is generally selected as 1×10⁻⁶. -7 Up to 5×10 -7 In the current embodiment, the step size is 5×10. -7 It can meet the computational accuracy requirements of the model.

[0077] S2.2 To effectively filter out the DC component signal from each group of thermal wave signal data obtained in S2.1, a DC component removal fitting model is further constructed. Furthermore, this embodiment employs a nonlinear fitting method to process the thermal wave signal for DC component removal.

[0078] The nonlinear fitting method proposed in this invention can significantly improve the accuracy of DC component removal and enhance the extraction effect of small signal amplitude and phase. The numerical fitting model for the nonlinear fitting method is constructed as follows:

[0079]

[0080] in, It is an AC component, a constant. The value is obtained using the mini-batch gradient descent method. In this embodiment, The calculated results are 33.7349, 0.01315, and 111.8414.

[0081] like Figure 2 As shown, the AC component amplitude curve is obtained after removing the DC component signal from the thermal wave data using the nonlinear fitting method proposed in this invention. Clearly, the AC component performance obtained by the nonlinear fitting method used in this embodiment is very stable.

[0082] S2.3, after removing the DC component from the thermal wave signal data at different modulation frequencies in S2.2, extracts the frequency domain amplitude and phase information data of the thermal wave signals at different modulation frequencies after removing the DC component, based on the phase-locked-in thermal imaging algorithm. For example... Figure 3As shown, in this embodiment, the defect-containing numerical model under a modulation frequency of 0.02 Hz is shown in the spatial (a) amplitude and (b) phase maps after removing the direct current component information.

[0083] S2.4 To further improve the robustness of the model to experimental data, a data enhancement method is used to introduce random noise in the data. Various 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 simulation data, thereby improving the generalization ability of the model.

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

[0085]

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

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

[0088] Further, the value range of the training set parameters satisfies the following, 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 as α xy ∈(0,12]×10 -6 m 2 / s.

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

[0090] S3.1 As Figure 4As shown, a three-channel deep learning model based on residual modules is constructed. The spatial coordinates, modulation frequency, time-domain surface temperature information, frequency-domain amplitude and phase information generated by numerical solution, and the thermal physical parameters of the material itself are taken as input features, and the thermal diffusivity is taken as output target. Further, the first branch inputs the time-domain surface temperature information, which is used to capture the time characteristics of material thermal diffusion; the second branch inputs the frequency-domain amplitude and phase information, which is used to extract the frequency-domain characteristics of the material; and the third branch inputs the physical property parameters (density, specific heat capacity, thermal conductivity), which are extracted through a fully connected layer to extract the physical property characteristics. Each branch network uses a residual module for feature extraction, uses a skip connection to alleviate the gradient vanishing problem in the deep network, accelerates convergence, and improves the accuracy of the model. In order to standardize the numerical values after convolution, a normalization layer is added after each convolution layer. The maximum pooling method is selected. At the same time, in order to reduce the influence 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 convolution layers of each branch is constructed as 15 layers, the convolution kernel size is [2, 1], [3, 1] and [5, 1], the number of convolution kernels in each layer is 18, 36 and 72, the convolution step is 1, and the padding model is 0 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 merged 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 weight and adjust the features in the spatial and channel dimensions respectively. The position attention module focuses on the importance of features in 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 taken as input features to input the deep neural network, and the thermal diffusivity is taken 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, in which the value of y is fixed as 2 mm; the value of x ranges from 1 to 15 mm with an increment of 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 weight and bias values of the neural network model. In this embodiment, the boundary values of the initial weight and bias values satisfy [-1, 1]; the global and individual learning factors are set to 3; the population update times are 30; and the population size is 15.

[0094] Figure 5This embodiment demonstrates the fitting trend of a multi-terminal deep neural network with a large number of training samples, and uses the root mean square error (RMSE) algorithm to evaluate the degree of matching between the true and predicted values. It can be observed that the RMSE and loss function gradually stabilize with the increase of iteration steps. The stabilized RMSE is 0.82813.

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

[0096] S4.1 employs phase-locked thermal imaging technology with a periodically modulated square wave excitation heat source and a 50% duty cycle to acquire time-domain surface temperature data of unsintered metal powder blanks at different frequencies. In this embodiment, the selected modulation frequencies are 0.01Hz, 0.02Hz, and 0.05Hz. Furthermore, the non-steady-state thermal wave response exhibits a nonlinear DC component and is spatially dependent. Therefore, analyzing the amplitude and phase based on the steady-state thermal wave field can further reduce computational costs. In this embodiment, the thermal load application time is set to 1300s, the sampling frame rate of the mid-infrared camera is 15Hz, and sampling begins after 1000s.

[0097] The data acquired in S4.2 is based on the DC component signal removed from the experimental data in S2.2, and the amplitude and phase information data in the frequency domain are obtained based on S2.3.

[0098] The processed experimental data in S4.3 is input into the multi-terminal deep neural network that has been trained in S3 to obtain the predicted thermal diffusivity of the unsintered metal powder compact material to be tested.

[0099] like Figure 6 This embodiment presents the experimental data used to verify the prediction accuracy of the model. The test sample used was an unsintered metal powder compact with a length, width, and height of 31 mm, 12 mm, and 6 mm, respectively. The experiment selected spatial distribution data of thermal wave amplitude and phase under modulation frequencies of 0.01 Hz, 0.02 Hz, and 0.05 Hz. The coordinate information, amplitude, and phase values ​​of 15 randomly selected pixels within the sample range were used as input parameters for training the deep neural network. The prediction results are shown below. Figure 6 As shown, at locations far from the damaged area, the predicted value for the healthy area is close to the true value; as the test location gets closer to the damaged area, the predicted value for thermal diffusivity gradually approaches the true value for the damaged area. Comparing the prediction results for the two areas, it can be found that the prediction result for the healthy area is significantly affected by the damaged area. This is because the thermal diffusivity of the damaged area is greater than that of the healthy area, thus the diffusion of heat waves is mainly dominated by the damaged area with its high thermal diffusivity. Figure 6 The results show fluctuations in the predictions, with an average value of 5.2 × 10⁻⁶. -6 m 2s. In comparison with the thermal diffusivity values measured in the literature, the predicted values obtained by the current model are in the same range as the reported values.

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

[0101] Figure 7 is a block diagram of the high-precision thermal diffusivity inversion method system device based on a multi-terminal deep neural network shown in the embodiments of the application. Please refer to Figure 7 The system device realizes the technical solutions described in the application in a software manner, and the device comprises:

[0102] The numerical simulation module 701 is used for simulating to generate a required time-domain numerical model data set. A two-dimensional numerical model is designed from the material parameters to be measured, and a boundary condition is set. The two-dimensional model is further divided into a grid, and a numerical calculation time step is set. Surface temperature signal original data sets under different modulation frequencies are calculated and generated.

[0103] The data preprocessing module 702 is used for removing a direct current component to 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. A nonlinear fitting method is used to remove the direct current component of the data to obtain stable alternating current component data. Based on a phase-locked thermal imaging algorithm, the amplitude and phase results of the alternating current component data in the frequency domain are calculated to obtain amplitude and phase information under different modulation frequencies.

[0104] The data enhancement module 703 is used for introducing random noise of 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. Random noise with noise intensities of 1%, 5%, 10% and 15% is considered in the thermal wave response data. The application uses various random noise types such as Gaussian noise, impulse noise and salt and pepper noise. The noise environment in the actual experiment is simulated 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 data set, train the network and extract feature information and fusion, output the predicted material thermal diffusivity result. After processing by the numerical simulation module 701, the data preprocessing module 702 and the data enhancement module 703, the multi-modal large-scale data containing time domain surface temperature information data, frequency domain amplitude and phase information data, spatial coordinates, modulation frequency and material thermal parameters are input 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, the predicted material thermal diffusivity result value is output;

[0106] Experimental data verification module 705: used for inputting 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 material thermal diffusivity result.

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

[0108] The present application is particularly suitable for high-precision prediction of thermal diffusivity in complex media and can be widely applied in the fields of non-destructive testing, material thermal property analysis, industrial production process monitoring, etc.

[0109] Although the present application includes specific implementation details, these should not be interpreted as limiting the scope of any invention or the claimed scope, but mainly for describing the features of the specific embodiments of the particular invention. Some features described in the embodiments of the present application can also be implemented in a single embodiment. On the other hand, various features described in a single embodiment can also be implemented separately in multiple embodiments or 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 variation of a sub-combination.

[0110] Similarly, although the operations are depicted in a particular order in the drawings, this should not be understood as requiring the operations to be performed in the particular 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 can be advantageous. In addition, the separation of various system modules and components in the above embodiments should not be understood as requiring such separation 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 descriptions are only the preferred embodiment of the application, not intended to limit the application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the application shall be included in the protection scope of the application.

Claims

1. A high-precision inversion method for thermal diffusivity based on a multi-terminal deep neural network, characterized in that, The process includes the following steps: S1: Numerical simulation of the diffused thermal wave signal containing defects to establish a numerical model; S2: Based on the numerical model obtained in S1, establish the numerical model dataset; S3: Based on the residual module and the data obtained in S2, construct a three-channel deep neural network for training and prediction; S4: Based on the neural network constructed in S3, predict the thermal diffusivity of real materials under experimental conditions; S2 includes the following steps: Based on the numerical model obtained in S1, S2.1 performs numerical simulations on thermal wave signals at different modulation frequencies to generate large-scale initial thermal wave datasets at different frequencies. S2.2 uses a nonlinear fitting method to process the initial heat wave dataset generated in S2.1 to remove the DC component signal, thus obtaining the heat wave signal after removing the DC component. S2.3 Based on the phase-locked thermal imaging algorithm, frequency domain amplitude and phase information data of the thermal wave signal under different modulation frequencies are extracted from the thermal wave signal obtained in S2.2; S2.4 employs a data augmentation method to introduce random noise into the data obtained in S2.3; S2.5 performs decentralization and normalization on the amplitude and phase data in the frequency domain to obtain the final multimodal large-scale training dataset containing spatial location, different modulation frequencies, time-domain surface temperature information, frequency-domain amplitude and phase information, and thermophysical parameters of the material under test, and randomly extracts small batch datasets as validation sets. In step S2.2, the numerical fitting model constructed using the nonlinear fitting method is as follows: ； in, It is an AC component, a constant. The value is obtained using the mini-batch gradient descent method.

2. The high-precision inversion method for thermal diffusivity based on a multi-terminal deep neural network according to claim 1, characterized in that... S1 includes 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 unsteady state to steady state under different boundary conditions; S1.2 The thermal wave model is spatially discretized using the numerical solution finite element method to construct a spatial discrete mesh model; S1.3 Discretizes 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.

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

4. The high-precision inversion method for thermal diffusivity based on a multi-terminal deep neural network according to claim 1, characterized in that... S3 includes the following steps: S3.1 Construct a three-channel deep learning model based on residual modules; take the spatial location coordinates, modulation frequency, time-domain surface temperature information, frequency-domain amplitude and phase information, and the material's own thermal parameters generated by the numerical solution as input features, and thermal diffusivity as the output target; S3.2 Train the three-channel deep learning model network. Once the prediction error of the network model tends to stabilize, verify the prediction accuracy of the model through experimental data.

5. The high-precision inversion method for thermal diffusivity based on a multi-terminal deep neural network according to claim 4, characterized in that... In S3.1, the first branch inputs the surface temperature information in the time domain to capture the time characteristics of material thermal 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 takes input physical property parameters and extracts physical features through a fully connected layer.

6. The high-precision inversion method for thermal diffusivity based on a multi-terminal deep neural network according to claim 4, characterized in that... In S3.1, each branch network uses a residual module for feature extraction and employs skip connections to alleviate the gradient vanishing problem in deep networks, thereby accelerating convergence and improving the accuracy of the model.

7. The high-precision inversion method for thermal diffusivity based on a multi-terminal deep neural network according to claim 4, characterized in that... In S3.1, 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 merged according to dynamically calculated weights to generate the final fused feature vector. During feature fusion, a position attention module and a channel attention module are introduced to perform weighted adjustments on features in the spatial and channel dimensions, respectively. The position attention module focuses on the importance of features at different spatial locations, while the channel attention module focuses on the relationships between different channels.

8. The high-precision inversion method for thermal diffusivity based on a multi-terminal deep neural network according to claim 1, characterized in that... S4 includes the following steps: S4.1 employs a phase-locked thermal imaging method with a periodically modulated excitation heat source to acquire time-domain surface temperature information data at different frequencies; S4.2 removes the DC component signal from 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 is input into the multi-terminal neural network that has been trained in S3 to obtain the predicted thermal diffusivity of the material under test.

Images