Thermal analysis data noise reduction method based on generative adversarial network

Abstract

The invention provides a thermal analysis data noise reduction method based on a generative adversarial network, and the method comprises the steps: constructing a generator and a discriminator, optimizing parameters through adversarial training, and generating noiseless heat flow data. According to the method, thermal analysis data is subjected to formatting and normalization preprocessing, the generator generates noise reduction data, the discriminator discriminates the authenticity of the data, and the performance of the generator is improved. And through joint optimization of anti-loss and mean square error, a high-quality noise reduction effect is realized. Through verification, the method effectively removes noise, retains a key characteristic curve, remarkably improves the signal-to-noise ratio and analysis precision of data, and provides theoretical and technical support for data noise reduction of a thermal analysis instrument.

Description

Technical Field

[0001] The present invention relates to a thermal analysis data denoising method based on a generative adversarial network, belonging to the technical field of combining calorimetry and data processing. Background Art

[0002] Thermal analyzer is an instrument commonly used to measure the thermal properties of materials. It is widely used in many fields such as materials science, chemical engineering, and pharmaceuticals. Current thermal analyzers can collect heat flow signals of samples during temperature changes, namely differential thermal analysis (DTA) data. These signals are very weak DC signals and need to be amplified by hardware circuits after being detected by the sensor. However, noise is inevitably introduced during the amplification process. Common noises include thermal noise, 1 / f noise, and white noise. Thermal noise is generated by the thermal motion of electrons in resistive materials. It is proportional to temperature and resistance value, but has nothing to do with frequency. This noise exists in all resistive elements and is inevitable. 1 / f noise is common in semiconductor devices, resistors, power supplies, etc. It is low-frequency noise. Its spectrum intensity is inversely proportional to frequency, so it is usually most significant in the low-frequency band. White noise is a type of noise with a uniform power density across the entire spectrum. Its source is usually from external electromagnetic interference, power supply noise, and random thermal motion inside electronic devices.

[0003] There are several defects in the existing DTA data processing methods, the most significant of which is that most of the existing filtering technologies can only smooth high-frequency or low-frequency noise, and cannot flexibly adjust the filtering strategy according to the local characteristics of the data curve. The existing technology usually adopts a unified filtering scheme for the entire data curve, and fails to fully consider the trend changes of the curve in different regions, so it is difficult to achieve an ideal balance between retaining the effective components of the signal and eliminating noise. This defect limits the accuracy and adaptability of data processing, especially when the noise distribution is complex or the signal spectrum span is large, the existing methods fail to effectively improve the reliability of the data and the accuracy of the analysis results.

[0004] These defects make the accuracy and processing efficiency of thermal analysis data low, affecting the accurate evaluation of the thermal properties of materials. Therefore, a more efficient and accurate data processing method is urgently needed to improve the utilization value of thermal analysis data. Summary of the invention

[0005] The present invention provides a thermal analysis data denoising method based on a generative adversarial network, and the technical solutions adopted are mainly as follows:

[0006] Step 1: Format the original thermal analysis experimental data, extract the temperature and heat flow data in the experimental data, and normalize them;

[0007] Step 2: Build a generative adversarial network, including a generator and a discriminator;

[0008] Step 3: Use measured data to conduct adversarial training on the generative adversarial network, optimize the parameters of the generator and discriminator, and finally use the generator to generate denoised heat flow data, which is compared with the original data to verify the denoising effect.

[0009] Further, the specific steps of step 1 are as follows:

[0010] Step 1.1: Extract temperature and heat flow data by reading the raw data file obtained by the thermal analysis equipment.

[0011] Step 1.2: Normalize the extracted data and map the data to a fixed range (such as 0 to 1). The normalization formula is: Where x is the original data, min(x) and max(x) represent the minimum and maximum values ​​of the data respectively, and x′ is the normalized data.

[0012] Furthermore, the specific steps of step 2 are as follows:

[0013] Step 2.1: Construct the generator G using a multi-layer fully connected neural network, including an input layer, a hidden layer, and an output layer. The hidden layer uses batch normalization (BN) and a leaky ReLU activation function to enhance the generation quality.

[0014] Step 2.2: Construct the discriminator D, use a multi-layer fully connected neural network, combine the regularization layer and the batch normalization BN layer to improve the model's robustness to noise, and use the cross entropy loss function for optimization, which is defined as: where y i represents the true label of the sample, x i Represents input data.

[0015] Step 2.3: Generator’s loss function L G It consists of adversarial loss and mean square error (MSE), which is defined as: L G =λ1L adv +λ2MSE(Q gen , Q true ),in Q gen is the generated heat flow data, Q true The real heat flow data.

[0016] Further, the specific steps of step 3 are as follows:

[0017] Step 3.1: The generator G and the discriminator D are trained alternately. The generator optimizes its adversarial loss function L GTo generate data that can deceive the discriminator, and the discriminator optimizes the loss function L D Improve classification accuracy for both real and generated data.

[0018] Step 3.2: After training is completed, use the generator G to generate the corresponding denoised heat flow data Q for the input temperature data gen And compare it with the original data to verify the denoising effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of the generative adversarial network algorithm of the present invention.

[0020] Figure 2 It is an example diagram of the training phase of the present invention.

[0021] Figure 3 It is an example diagram of the optimization stage of the present invention. DETAILED DESCRIPTION

[0022] The present invention provides a thermal analysis data denoising method based on a generative adversarial network. The present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0023] 1. Data preprocessing

[0024] Import the raw experimental data files generated by thermal analysis instruments (such as DSC) into the system and extract the required temperature (TE) and heat flow (DSC) column data. The extracted raw data is normalized and the data range is mapped to [0, 1] for easy input into the generative adversarial network model.

[0025] 2. Generate adversarial network structure

[0026] In step 2, the DSC curve after preprocessing can be used to design the generator and discriminator. The specific steps are as follows:

[0027] 2.1 Design of the Generator

[0028] Select the DTA curve as the analysis curve according to the requirements, and select "connecting method", "vertical method" or "inverse curve method" as the analysis method.

[0029] The generator consists of a multi-layer fully connected neural network. The input layer receives the normalized temperature data, the multi-layer neural network processes it through batch normalization and activation functions to enhance the stability and continuity of data generation, and the output layer generates denoised heat flow data corresponding to the input temperature data.

[0030] 2.2 Discriminator Design

[0031] The discriminator is also composed of a multi-layer fully connected neural network, and its main function is to determine whether the input data is real heat flow data or pseudo heat flow data generated by the generator.

[0032] The input layer receives both real heat flow data and data generated by the generator. The hidden layer uses the Leaky ReLU activation function and the Dropout layer to enhance robustness to noise. The output layer outputs the true and false discrimination results through the binary cross entropy loss function.

[0033] 3. Model training and optimization

[0034] In step 3, the model training and optimization process includes alternating training of the generator and the discriminator, as well as model performance verification, which specifically includes the following steps:

[0035] 3.1 Alternating training of generator and discriminator

[0036] The training and optimization of the model adopts the method of alternating training of the generator and the discriminator to generate noise-free, high-quality heat flow data. First, in the training of the discriminator, the real heat flow data (DSC data) and the pseudo data generated by the generator (denoised DSC data) are input. The discriminator optimizes its classification ability by calculating the classification loss of the real data and pseudo data to ensure that it can accurately judge the authenticity of the input data. To this end, the binary cross entropy loss function (BCEWithLogitsLoss) is used, the real data label is set to 1, the generated data label is set to 0, and the parameters of the discriminator are updated according to the loss result, thereby improving its discrimination accuracy.

[0037] In the training of the generator, temperature data (TE) is input, and the generator outputs the corresponding pseudo DSC data, which is then input into the discriminator. The goal of the generator is to deceive the discriminator through adversarial loss, causing the discriminator to mistakenly believe that the generated data is real (labeled as 1). In order to further improve the quality of the generated data, the generator introduces the mean square error loss (MSELoss) to constrain the error between the generated data and the real DSC data. Adversarial loss and constraint loss are used jointly to optimize the parameters of the generator so that the data it generates is closer to the real distribution. During the training process, the generator and the discriminator are optimized alternately, the ability of the discriminator is continuously improved, and the generator is gradually improved through feedback information until the two reach a balance and the generator can output high-quality denoised data.

[0038] 3.2 Noise reduction effect verification

[0039] After the model training is completed, the trained generator is used to input the temperature data (TE) to generate the corresponding noise-free heat flow data (denoised DSC data). Subsequently, the generated denoised DSC data is compared with the original DSC data to analyze the denoising effect. The performance of the model is verified by calculating evaluation indicators such as mean square error (MSE) and signal-to-noise ratio (SNR). The comparison results show that the denoised data generated by the generator can effectively remove the noise in the heat flow data while retaining the key characteristic curves in the original data, thereby achieving high-quality thermal analysis data denoising.

Claims

1. This invention patent provides a thermal analysis data denoising method based on a generative adversarial network, which is characterized by The following steps are involved: Step 1: Format the original thermal analysis experimental data, extract the temperature and heat flow data in the experimental data, and normalize them; Step 2: Build a generative adversarial network, including a generator and a discriminator; Step 3: Use measured data to conduct adversarial training on the generative adversarial network, optimize the parameters of the generator and discriminator, and finally use the generator to generate denoised heat flow data, which is compared with the original data to verify the denoising effect.

2. According to the thermal analysis data denoising method based on generative adversarial network in claim 1, the specific steps of step 1 are as follows: Step 1.1: Extract temperature and heat flow data by reading the raw data file obtained by the thermal analysis equipment. Step 1.2: Normalize the extracted data and map the data to a fixed range (such as 0 to 1). The normalization formula is: Where x is the original data, min(x) and max(x) represent the minimum and maximum values ​​of the data respectively, and x′ is the normalized data.

3. The thermal analysis data denoising method based on generative adversarial network according to claim 1 is characterized in that: The specific steps of step 2 are as follows: Step 2.1: Construct the generator G using a multi-layer fully connected neural network, including an input layer, a hidden layer, and an output layer. The hidden layer uses batch normalization (BN) and a leaky ReLU activation function to enhance the generation quality. Step 2.2: Construct the discriminator D, use a multi-layer fully connected neural network, combine the regularization layer and the batch normalization BN layer to improve the model's robustness to noise, and use the cross entropy loss function for optimization, which is defined as: where y i represents the true label of the sample, x i Represents input data. Step 2.3: Generator loss function L G It consists of adversarial loss and mean square error (MSE), which is defined as: L G =λ1L adv +λ2MSE(Q gen , Q true ),in Q gen is the generated heat flow data, Q true The real heat flow data.

4. The thermal analysis data denoising method based on generative adversarial network according to claim 1, characterized in that: The specific steps of step 3 are as follows: Step 3.1: The generator G and the discriminator D are trained alternately. The generator optimizes its adversarial loss function L G To generate data that can deceive the discriminator, and the discriminator optimizes the loss function L D Improve classification accuracy for both real and generated data. Step 3.2: After training is completed, use the generator G to generate the corresponding denoised heat flow data Q for the input temperature data gen And compare it with the original data to verify the denoising effect.

Images