Method for denoising flame emission spectrum by using improved CNN (Convolutional Neural Network) combined with channel attention mechanism

Abstract

The invention discloses a method for denoising a flame emission spectrum by using an improved CNN (Convolutional Neural Network) combined with a channel attention mechanism, and provides a CBC (Convolutional Neural Network) architecture based on a convolutional neural network, and the number of spectrum channels is dynamically adjusted by integrating a channel heuristic attention mechanism and a self-adaptive channel selection module so as to balance the complexity and noise reduction performance of a model. A channel attention module is introduced after each layer of convolution operation, so that the model can perform adaptive denoising processing on a spectral signal with a low signal-to-noise ratio. According to the mechanism, weighting adjustment is carried out on an input feature map by calculating the weight of each channel, so that a network can focus on more important signal parts, and meanwhile, the influence of noise and irrelevant information is inhibited. The self-adaptive adjustment capability greatly enhances the performance of the network in processing a complex noise environment, and especially, the denoising effect is remarkably improved when short-exposure spectral data with low SNR is processed.

Description

Technical Field

[0001] The present invention belongs to the technical field of flame emission spectra, and in particular relates to a flame emission spectrum denoising method using an improved CNN combined with a channel attention mechanism. Background Art

[0002] Flame emission spectroscopy (FES) technology plays a pivotal role in energy research and industrial combustion. By precisely measuring the characteristic spectra of excited-state emission from elements during combustion, it provides real-time insights into the combustion state, playing an irreplaceable role in combustion monitoring, gas composition analysis, and temperature estimation. Especially in high-pressure combustion environments, FES is crucial for precise control and optimization of the combustion process, providing crucial data support for improving combustion efficiency and reducing pollutant emissions.

[0003] In practical applications, flame emission spectrum data collection is fraught with difficulties. The high-pressure combustion environment is complex, with turbulent flow, large and frequent temperature fluctuations, and continuous chemical reactions interfering with each other, seriously interfering with the quality of the spectral signal. To address these issues, traditional denoising methods, such as Gaussian filtering, wavelet transform, and Fourier transform, are widely used for spectral data preprocessing. Gaussian filtering smoothes the data by setting appropriate filter parameters, which can remove low-frequency noise to a certain extent. However, when processing spectral data, it will inevitably produce a blurring effect on the high-frequency portion, resulting in the loss of key information in the signal. For example, during the combustion process, instantaneous changes in temperature and gas composition are usually reflected in the high-frequency components of the spectral signal. While Gaussian filtering removes noise, it also weakens these high-frequency features, affecting the accurate analysis of the combustion process.

[0004] In recent years, deep learning technology has developed rapidly, and convolutional neural networks (CNNs) have made significant progress in image processing and signal denoising. With their unique structures, such as convolutional and pooling layers, CNNs can adaptively learn the characteristics of input data and automatically extract useful information from complex data. They have demonstrated superior performance to traditional methods in processing complex noise and highly nonlinear signals. In the task of flame emission spectrum denoising, CNN-based methods, such as DnCNN and FFDNet, learn the spatial features of the spectrum through multi-layer convolution operations. Without explicitly setting a noise model in advance, they can effectively remove common noise sources such as Gaussian noise and quantization noise, thereby improving the quality of spectral data to a certain extent.

[0005] Existing convolutional neural network denoising technology has defects when processing infrared spectral data. On the one hand, traditional networks often ignore the potential correlation between different channels when processing multi-channel spectral data. Although methods such as FFDNet can reduce noise, they do not fully mine channel information and their effects are limited in complex noisy environments. On the other hand, existing technologies mostly use fixed channel configurations and cannot dynamically adjust the number of spectral channels based on the characteristic distribution of input data. It is difficult to balance model complexity and denoising performance, and may result in insufficient feature expression or waste of computing resources. In addition, due to the complex noise structure of actual spectral data, existing convolutional neural networks are based on simple noise distribution assumptions and have limited adaptability to complex noise.

[0006] In their earlier research, Yoon et al. proposed a CNN framework based on an improved FFDNet for processing flame emission spectrum data. This framework integrated key components such as a reversible downsampling-upsampling (DU) operator, CNN layers, and a loss function based on the POD coefficient. This design, to a certain extent, addressed the challenges of noise interference and feature extraction in combustion spectra. However, in complex spectral environments and dynamic noise distributions, the framework's feature expression capabilities were limited, making it difficult to accurately extract weak or hidden features. In terms of channel information processing, the framework ignored differences in channel importance and was unable to fully explore potential connections between channels. Furthermore, Yoon et al.'s approach employed a fixed network structure and channel configuration, failing to balance model complexity with noise reduction performance, which could easily lead to wasted computing resources or poor performance. Summary of the Invention

[0007] Purpose of the Invention: To address the above challenges, this paper proposes a flame emission spectrum denoising method using an improved CNN combined with a Channel Attention Mechanism (CAM). This method overcomes the challenge of denoising short-exposure flame emission spectrum data in high-pressure combustion environments and significantly improves the accuracy of gas property estimation. This innovative architecture not only provides more reliable data support for high-pressure combustion environment monitoring but also promotes efficient optimization and precise control of the combustion process.

[0008] Technical solution: To achieve the purpose of the present invention, the technical solution adopted by the present invention is: a flame emission spectrum denoising method using an improved CNN combined with a channel attention mechanism, comprising the following steps:

[0009] Obtain a spectral dataset and divide it into training set, validation set and test set;

[0010] A flame emission spectrum denoising model was constructed. The model incorporates a channel attention mechanism after each convolution operation of a convolutional neural network layer and adaptively adjusts the weights of each channel.

[0011] The input of the model is a plurality of downsampled spectral data, which are input into the model after preprocessing; the output of the model is the generated denoised spectrum;

[0012] Use the training set to train the model, use the validation set to evaluate the model and tune hyperparameters, and then use the test set to evaluate the model's performance and measure the model's applicability and robustness in the application scenario.

[0013] Furthermore, the model includes several convolution blocks, batch normalization, activation functions, and adaptive channel number selection modules;

[0014] Convolution outputs multi-channel feature map F∈R C×H×W , where C is the number of channels, H×W is the spatial size, H is the height of the feature map, and W is the width of the feature map;

[0015] The multi-channel feature map F output by the convolution is batch normalized. The model processes the multi-channel feature map F, and each channel undergoes a global average pooling operation to compress the spatial dimension into a feature vector z c ∈R C :

[0016]

[0017] Among them, F(c,i,j) represents the value of the c-th channel at the spatial position (i,j), z c Encodes the global semantic information of the c-th channel;

[0018] Eigenvector z c The input is fed into a multilayer perceptron, a two-layer fully connected network, for nonlinear mapping to generate channel importance weights. The mapping process is as follows:

[0019] y=ReLU(W1·z c +b1) where y is the output result after the first layer operation, W1∈R d×C and b1∈R d are the weight and bias of the first layer, d is the intermediate dimension, and ReLU is the activation function;

[0020] w c =σ(W2·y+b2)

[0021] Among them, w c is the output result after the second layer operation, W2∈R C×d and b2∈R C are the weights and biases of the second layer respectively. The Sigmoid function σ maps the output to the range [0,1], and the generated weight vector w c Indicates the importance of each channel;

[0022] Using the generated channel weight w c , the model applies dynamic weighting to each channel of the feature map F:

[0023]

[0024] in, Represents the weighted value of the c-th channel at the spatial position (i, j), and the weighted feature map is

[0025] The adaptive channel number selection module dynamically adjusts the number of channels according to input features, where the input features are feature maps output by the convolutional layer;

[0026] After multiple layers of convolution, adaptive weighting, and channel number selection, the model generates optimized feature maps. The final output is the spectral characteristics.

[0027] Furthermore, the adaptive channel number selection module dynamically adjusts the number of output channels C′ of each convolutional layer, specifically:

[0028] The adaptive channel number selection module adopts a strategy of gradually increasing the number of channels, with each increase in the number of channels as a stage. After each stage of channel number increase, the model evaluates the performance indicators;

[0029] If the performance reaches its peak, stop adjusting the number of channels and select the number of channels C that provided the best performance in the previous round. opt As the final configuration.

[0030] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0031] This paper utilizes an improved CNN architecture. By introducing a channel attention module after each convolution operation, the model is able to adaptively denoise spectral signals with low signal-to-noise ratios. This mechanism calculates weights for each channel and adjusts the input feature map, allowing the network to focus on the more important signal components while suppressing the influence of noise and irrelevant information. This adaptive adjustment capability greatly enhances the network's performance in complex noisy environments, particularly when processing short-exposure spectral data with low SNR, achieving significantly improved denoising results.

[0032] The CBC model of the present invention performs better in terms of utilizing channel correlation, dynamically adjusting the number of channels, and adapting to complex noise by introducing a channel attention mechanism and an adaptive channel selection module, thereby effectively improving the noise reduction performance of infrared spectra. In the denoising task, the channel attention mechanism dynamically learns the importance of channels by weighting each channel of the input feature map, allowing the network to automatically adjust its attention to different channels. In the processing of flame emission spectra, it accurately retains important information and suppresses noise. Since the information in different bands of the flame emission spectrum is different, some bands are related to key information such as gas temperature, reaction state, and chemical composition, while some may be noise or irrelevant information. Therefore, the introduction of the channel attention mechanism allows the network to assign higher weights to key bands and improve denoising accuracy.

[0033] By adding an adaptive channel selection module and a channel attention module, the CBC model can better adapt to complex noise and improve the noise reduction effect; the channel attention mechanism of the CBC model can dynamically learn the importance of channels and optimize the noise reduction effect; the adaptive channel selection module of the CBC model can dynamically adjust the number of channels, efficiently utilize computing resources, and enhance the model's generalization ability and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is the overall processing process of the method of the present invention.

[0035] Figure 2 This is a schematic diagram of the improved denoising neural network architecture.

[0036] Figure 3 This is a diagram of the core architecture of the CBC model.

[0037] Figure 4 It is a channel-based attention mechanism.

[0038] Figure 5 The spectrum of the noise signal is compared with the real signal and the signal after model processing. DETAILED DESCRIPTION

[0039] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0040] The improved CNN combined with the channel attention mechanism described in the present invention is used for the flame emission spectrum denoising method, such as Figure 1 As shown in the figure, it includes four parts: analyzing the training data set, dividing and processing the training and test data, training the CNN, and evaluating the model performance. The specific implementation includes the following:

[0041] (1) Network architecture.

[0042] Figure 2This is a schematic diagram of the improved neural network structure. In order to further improve the performance of the model, the present invention introduces a channel attention mechanism and embeds it after each layer of convolution operation to enhance the feature expression capability.

[0043] Figure 3 This paper demonstrates the architecture of the Convolutional Batch Normalization with Channel Heuristic Attention Layer (CBC) model, which consists of three parts: input, core processing, and output. The input consists of multiple downsampled spectral data, which are preprocessed and then fed into the model. The core processing part, the main architecture of the CBC model, includes multiple convolutional blocks, batch normalization (BN), an activation function (ReLU), and an innovative adaptive channel number selection module (ACNSM). This module dynamically adjusts the number of channels based on input features to achieve an optimal balance between performance and computing resources. Through repeated convolution operations and channel selection, the model gradually extracts and enhances the high-dimensional features of the spectrum. The output is a high-quality, denoised spectrum generated by the model. Compared to the input, the output spectrum has a significantly higher signal-to-noise ratio (SNR), effectively suppressing noise while preserving key features. This entire process demonstrates the outstanding performance of the CBC model in processing complex infrared spectra.

[0044] The CBC model is designed to improve the feature representation and noise reduction performance of spectral data in complex spectral environments. The model dynamically adjusts the importance of the convolutional layer output and integrates an adaptive channel number selection mechanism (ACNSM) to optimize feature representation. Figure 4 As shown, the following is a detailed description of its working principle.

[0045] CBC model processes multi-channel feature maps F∈R C×H×W , where C is the number of channels, H×W is the spatial size, H is the height of the feature map, and W is the width of the feature map; capturing the global importance of each channel. In order to focus on the global feature semantics, each channel undergoes a global average pooling (GAP) operation to compress the spatial dimension into a feature vector z c ∈R C :

[0046]

[0047] Among them, F(c,i,j) represents the value of the c-th channel at the spatial position (i,j), z c It encodes the global semantic information of the cth channel. GAP reduces the computational complexity while capturing the channel-level semantic information.

[0048] Eigenvector z cThe input is fed into a multilayer perceptron (MLP), a two-layer fully connected network, for nonlinear mapping to generate channel importance weights. The mapping process includes the following steps:

[0049] y=ReLU(W1·z c +b1)

[0050] Among them, y is the output result after the first layer operation, W1∈R d×C and b1∈R d are the weights and biases of the first layer, and d is the intermediate dimension. The ReLU activation function introduces nonlinearity to the model.

[0051] w c =σ(W2·y+b2)

[0052] Among them, w c is the output result after the second layer operation, W2∈R C×d and b2∈R C are the weights and biases of the second layer. The Sigmoid function σ maps the output to the range [0,1]. The resulting weight vector w c Indicates the importance of each channel.

[0053] Using the generated channel weight w c , the CBC model applies dynamic weighting to each channel of the feature map F. This step strengthens important features and suppresses irrelevant noise:

[0054]

[0055] in, Represents the weighted value of the c-th channel at the spatial position (i, j), and the weighted feature map is Weighted feature map Focus on the most significant channels and filter out features that have little impact on the final task or contain high noise levels.

[0056] To optimize computational efficiency, CBC integrates an adaptive channel number selection module (ACNSM), which dynamically adjusts the number of output channels C′ of each convolutional layer. The module follows a strategy of gradually increasing the number of channels, for example, from 16 to 32 to 64 to 128. At each stage, the model evaluates performance metrics such as SNR or MAE. The specific process is as follows:

[0057] After the number of channels is increased, the performance indicators are evaluated. Once the performance reaches a peak (for example, no significant improvement in SNR), stop adjusting the number of channels and select the channel number C that provides the best performance in the previous round. opt This dynamic adjustment prevents unnecessary computational redundancy and ensures efficient resource utilization without sacrificing performance.

[0058] After multiple layers of convolution, adaptive weighting, and channel number selection, the model generates an optimized feature map F out These feature maps have higher SNR while retaining key spectral feature information in the spectral data, thus providing reliable support for subsequent spectral analysis and processing:

[0059]

[0060] The final output effectively restores the highly fluctuating spectral features in complex spectral environments, improving the clarity and accuracy of the spectrum.

[0061] The CBC model combines dynamic channel weighting, adaptive channel number selection, and efficient feature extraction to improve the denoising performance of spectral data in complex spectral environments. This method outperforms traditional models by adapting to complex noise distributions and optimizing computational efficiency while retaining key spectral information. Consequently, in complex spectral data, the model can more accurately and efficiently filter and utilize features, optimizing the quality of the overall feature representation. This lays a solid foundation for subsequent denoising and spectral property prediction tasks. In practical applications, the channel-based attention mechanism can effectively enhance the model's focus on key features, reduce the impact of redundant information, and further improve prediction accuracy and generalization, particularly in scenarios such as multi-task learning and few-shot learning.

[0062] In the CBC model of the present invention, first, through multi-layer convolution, low-level simple features are gradually converted into high-level complex features, providing rich feature information for subsequent analysis. Batch normalization (BN) works in conjunction with the activation function (ReLU): batch normalization is performed after convolution to standardize the feature map data output by the convolution, reduce internal covariate offset, accelerate model training, and make the data distribution more stable. Then, the ReLU activation function introduces nonlinearity into the model, enhances the expressiveness of the model, enables the model to learn more complex feature relationships, further highlights effective features, and improves the feature extraction effect. Channel-based attention mechanism: through the global average pooling (GAP) operation, the spatial dimension of the multi-channel feature map is compressed into a feature vector, encoding the global semantic information of each channel. Afterwards, the feature vector is nonlinearly mapped by a multi-layer perceptron (MLP) to generate channel importance weights. These weights are used to dynamically weight each channel of the feature map, strengthen key feature channels, suppress noise and irrelevant information, and achieve efficient focusing and extraction of key spectral features. Adaptive Channel Number Selection Module (ACNSM): This module dynamically adjusts the number of channels in the convolutional layer output based on input features. Starting with an initial number of channels, it gradually increases the number, evaluating performance metrics under different channel configurations and selecting the optimal number of channels. This avoids the degradation of feature representation due to insufficient channels and the waste of computational resources due to excessive channels. While maintaining feature extraction integrity, it also improves computational efficiency, ensuring that the model can efficiently extract key features from spectral data.

[0063] (2) Adaptive channel number selection module.

[0064] To optimize model performance and computational efficiency, this paper proposes an adaptive channel selection module, designed to find the optimal balance between the number of spectral channels and noise reduction performance. The core concept of this module is to gradually increase the number of channels in the network, dynamically monitor changes in noise reduction performance indicators, and automatically select the most appropriate number of channels based on performance trends, thereby achieving a balance between performance and resource consumption.

[0065] Module design starts with the initial number of channels C init Starting with C = 16, the number of channels was gradually increased, for example, C = 16, 32, 64, and 128. After each adjustment, the model was trained and validated, and corresponding noise reduction performance metrics were recorded, including the SNR improvement and MAE reduction. Specifically, the SNR improvement measures the model's ability to improve spectral quality, while the MAE reduction reflects the model's effectiveness in suppressing noise. Model complexity metrics (such as the number of parameters and computational cost) were also monitored.

[0066] When it is found that the performance indicators (such as SNR or MAE) no longer improve or even deteriorate after increasing the number of channels, the module stops increasing the number of channels and selects the number of channels that provided the best performance in the previous round as the final configuration. This dynamic adjustment prevents unnecessary computational redundancy and ensures efficient resource utilization without sacrificing performance.

[0067] (3) Training and testing data.

[0068] The experimental data set follows the results of previous studies, and the data collection process is rigorous and systematic. In a stable spectral environment, the experimental conditions were carefully set, and data were collected using a professional spectral acquisition device - a portable USB-connected spectrometer. By precisely controlling the spectral parameters, the characteristic distribution continuously changes within the specified range. Using a high-precision control system, the environmental spectral parameters are gradually increased from 1 unit to 10 units. At the same time, the observation time of the spectrometer is flexibly adjusted, and multiple time options are set to 0.05 seconds, 0.2 seconds, 0.4 seconds and 2 seconds. For each combination of conditions, short observation (instantaneous) low signal-to-noise ratio spectra and long observation (average) high signal-to-noise ratio spectra are collected, totaling 80,000 sets of spectral data.

[0069] During data preprocessing, dark spectra collected under signal-free conditions are first removed from all acquired spectral data to effectively eliminate background noise. All spectral data are then normalized to ensure uniform dimensionality and comparability across different operating conditions. This series of steps provides a high-quality, standardized dataset for subsequent model training.

[0070] (4) Training and verification.

[0071] According to established rules, the complete dataset is divided into training, validation, and test sets. The training set, which accounts for approximately 70% of the data, is used to learn the model's core parameters and establish its feature extraction capabilities. The validation set, which accounts for approximately 15%, is used for critical model evaluation and hyperparameter tuning during the training process. By regularly monitoring model performance metrics (such as loss and accuracy) on the validation set, the network architecture, learning rate, optimizer parameters, and other hyperparameters are adjusted in a timely manner. This helps effectively avoid overfitting and guides the model towards optimal generalization performance. The remaining 15% of the data constitutes the test set, which is specifically used for a final, fair, and objective performance evaluation on unseen data after model training and optimization. This helps to accurately measure the applicability and robustness of the model in real-world application scenarios.

[0072] During training, the Adam optimizer was selected as the primary algorithm for updating model parameters. This optimizer adaptively adjusts the learning rate based on the first- and second-order moment estimates of the gradient, ensuring a balance between rapid convergence and stable training. The initial learning rate was set to [X], and a cosine annealing strategy was used to dynamically adjust the learning rate during training. Data was input in batches of 128 during each training cycle, ensuring efficient balance between memory utilization and gradient computation stability. The model was iteratively optimized for 50 cycles on the training set, gradually improving its ability to understand and process spectral data features. Training concluded when performance on the validation set stabilized, indicating that the model had reached optimality.

[0073] (5)Evaluation indicators.

[0074] In order to comprehensively and accurately evaluate the performance of the improved model architecture in flame emission spectrum denoising and gas property prediction tasks, the present invention selects signal-to-noise ratio (SNR) and mean absolute error (MAE) as core evaluation indicators.

[0075] The signal-to-noise ratio (SNR) is a classic indicator for measuring signal quality. In this paper, the purity and clarity of the spectrum are quantified by accurately calculating the ratio of the effective signal power to the noise power in the spectral signal. The calculation formula is:

[0076]

[0077] Among them, P signal Represents the signal power, P noise represents noise power. A higher SNR value indicates that the signal component in the spectrum is dominant and noise interference is effectively suppressed, providing a good foundation for subsequent accurate extraction of characteristic information in the spectrum, especially weak but critical characteristic peaks related to gas properties. Conversely, a lower SNR value indicates severe noise contamination, which may mask or distort the true spectral characteristics, thereby significantly interfering with the identification and measurement of gas properties.

[0078] The mean absolute error (MAE) is used to evaluate the average deviation between the gas property values ​​predicted by the model and the true values. Its mathematical expression is:

[0079]

[0080] Where N is the number of samples, y i represents the true gas property value of the i-th sample, is the gas property value predicted by the model for that sample. Smaller MAE values ​​indicate closer model predictions are to the true values, and the model's accuracy and reliability in quantitatively predicting gas properties are higher, providing solid data support for the precise monitoring and control of key parameters in the combustion process. Conversely, larger MAE values ​​indicate significant deviations in the model's predictions, which can lead to misjudgment and incorrect control of the combustion state, thus affecting the performance and stability of the combustion system.

[0081] Effect evaluation of this embodiment:

[0082] (1) Use the proposed model for denoising.

[0083] To fully investigate the noise reduction capabilities of the proposed model, we applied the improved model architecture to a spectral dataset. These conditions included high- and low-fluctuation spectra at varying dynamic levels (from low to high dynamics, 1-10 units), equivalence ratios (from lean to rich, 0.8-1.2), and different spectral steady-state conditions. The goal was to comprehensively assess the model's adaptability and robustness in a dynamic environment.

[0084] Figure 5 Demonstrates the denoising effect of the model under typical conditions. Figure 5 (a) in the figure shows the original low signal-to-noise ratio high fluctuation spectrum, which is affected by severe noise, resulting in significant baseline fluctuations and unclear spectral peaks, thus affecting spectral analysis and applications. Figure 5 The low-fluctuation spectrum shown in (c) has less noise interference and higher spectral quality.

[0085] After the improved model processing, Figure 5 As shown in (b), the noise is significantly and effectively suppressed, and the processed spectrum closely resembles the true low-fluctuation spectrum. The model accurately identifies and reduces the noise component, stabilizes the baseline, and significantly reduces noise fluctuations. Notably, during the denoising process, the model successfully preserves and enhances key characteristic peaks closely related to spectral properties. The precise restoration and enhancement of these characteristic peaks is achieved thanks to a channel-based attention mechanism that deeply mines and intelligently weights spectral features, ensuring that the model focuses on key features during the denoising process and avoiding information loss. This significantly improves the reliability and accuracy of spectral analysis.

[0086] (2) Adaptive channel selection module.

[0087] In experimental evaluations, the adaptive channel selection module demonstrated exceptional performance optimization capabilities. By dynamically adjusting the number of channels, the module successfully balanced the model's feature representation capabilities with computational resource utilization, achieving both improved performance and efficiency in denoising tasks. As shown in Tables 1 and 2, the adaptive channel selection module consistently identified the optimal channel configuration, enabling the network to excel in capturing key features of spectra with low signal-to-noise ratios (high fluctuations).

[0088] Specifically, the adaptive channel selection module starts with 16 channels and gradually doubles the number of channels (16→32→64→128), and continuously evaluates performance indicators such as signal-to-noise ratio (SNR) and mean absolute error (MAE) under different channel configurations. Ultimately, the module selects the number of channels that can bring the best performance. Experimental results show that after increasing the number of channels from 16 to 64, both the signal-to-noise ratio and noise reduction accuracy are significantly improved, significantly enhancing the model's ability to capture complex spectral features. Especially in cases where the spectral features are weak or the noise level is high, the adaptive module optimizes the number of channels to ensure that the model has sufficient representation capability to effectively separate and recover spectral features masked by noise.

[0089] Compared with traditional networks with fixed channel configurations, the adaptive channel selection module effectively avoids the decline in feature expression capabilities caused by insufficient number of channels, and avoids feature redundancy and waste of computing resources caused by too many channels. For example, in multi-condition spectral processing experiments, a network configured with 64 channels was able to maintain complete spectral features when processing low SNR spectra and significantly reduce the impact of background noise, thereby achieving better spectral quality and analysis accuracy. In addition, when the number of channels exceeds 64 (for example, 128 channels), the performance improvement tends to stabilize and even slightly decreases, further verifying the scientific nature and rationality of the adaptive module channel selection process.

[0090] The implementation of the adaptive channel selection module provides a flexible approach for optimizing network architecture. In experiments, this module effectively facilitates the model's dynamic adaptability to various spectral conditions, enabling it to adjust its feature representation capabilities according to specific task requirements. This not only provides critical support for spectral noise reduction and analysis in complex environments, but also serves as a reference model for other multi-channel data processing applications. By dynamically adjusting the number of channels, the adaptive module significantly reduces resource waste while improving model performance, laying a solid foundation for intelligent and efficient spectral processing.

[0091] (3)SNR improvement.

[0092] Rigorous evaluation results on the test set show that the improved model architecture achieves significant improvements in signal-to-noise ratio (SNR), with an average improvement of 7.5%. This improvement is particularly evident in the processed spectra, where noise interference is more effectively suppressed and spectral intensity and clarity are significantly enhanced compared to the original model output. In particular, when processing low-SNR, high-fluctuation spectra, the improved model successfully highlights subtle spectral features that were previously masked by noise, providing a more reliable basis for the precise analysis of key spectral properties. The processed spectral peaks are sharper and more complete, and the smoothness of the background noise is significantly improved, making the correlation between spectral features and system dynamics more accurate. This has greatly promoted the further development of spectral processing and analysis research.

[0093] The improved spectral quality is confirmed by experimental SNR data, as shown in Table 1. The table shows the SNR values ​​of the original model and the improved model output under various conditions, indicating that the processed spectra are significantly improved under different settings.

[0094] Table 1 SNR comparison between the improved network and DU-CNN

[0095]

[0096] (4) MAE decreases.

[0097] In the model performance evaluation, the mean absolute error (MAE) metric showed a significant reduction, by 19%. The improved model architecture, optimized through a channel-based attention mechanism, significantly enhanced its ability to accurately capture the complex nonlinear mapping between spectral features and denoised spectra. The denoised spectra output by the model showed significant improvement in fitting the true values, both in terms of low signal-to-noise ratio spectral denoising and feature extraction, with the error distribution more concentrated and approaching zero. For example, in multi-condition spectral processing experiments, the average deviation between the denoised spectra of the improved model and high-precision sensor measurements was significantly reduced under different feature distributions and dynamic conditions. This provides solid data support for the precise adjustment and optimization of spectral processing tasks, significantly reducing analytical errors and the risk of instability caused by noise interference. In addition, the reduction in MAE further demonstrates that the model effectively enhances denoising performance while maintaining spectral integrity and accuracy, injecting strong impetus into the in-depth analysis and intelligent optimization of spectral processing workflows.

[0098] The improvements achieved by the enhanced model are highlighted by the MAE values ​​from the experiments, as shown in Table 2. The table compares the MAE of the outputs of the original and improved models under various conditions, confirming the significant reduction in error and improvement in prediction accuracy.

[0099] Table 2

[0100] MAE comparison between the improved network and DU-CNN

[0101]

[0102] This paper, through an improved CBC architecture, achieves significant performance improvements in denoising short-time spectra in high-noise environments. Experimental results demonstrate that this method outperforms traditional techniques in terms of signal-to-noise ratio (SNR) and mean absolute error (MAE). It successfully addresses the challenge of extracting features from short-time spectra with low SNRs and provides key support for accurate monitoring and parameter optimization of spectral processing.

[0103] In terms of SNR improvement, the improved model dynamically focuses on spectral channels relevant to spectral characteristics, significantly suppressing noise interference. Experimental data show that the improved model achieves a 7.5% improvement in SNR compared to the method proposed by Yoon et al. In particular, when processing spectra in low SNR environments, the model is able to recover key spectral features obscured by noise, significantly enhancing the clarity and stability of spectral curves. Short-term spectra are often limited by the randomness and complexity of noise, and traditional methods have limited performance in such complex environments. However, the improved model significantly improves spectral reliability by precisely selecting characteristic channels. The reduction in Mean Average Error (MAE) also demonstrates that the improved model achieves higher accuracy in spectral denoising. Experimental results show that the improved model reduces MAE by 19%, demonstrating greater robustness and predictive power. In high-noise environments, parameter variations are complex and nonlinear, making it difficult for traditional linear models to effectively capture these changes. In contrast, a deep learning framework incorporating a channel-based attention mechanism can dynamically focus on spectral bands closely related to the target attribute, optimizing the mapping between the spectrum and the target attribute, thereby significantly improving prediction accuracy.

[0104] The adaptive channel selection module plays a key role in performance optimization. Experiments show that different channel configurations have a significant impact on model performance. When the number of channels increases from 16 to 64, both SNR and MAE continue to improve, with the 64-channel configuration performing best. When the number of channels is further increased to 128, the performance improvement levels off and even decreases slightly, indicating that too many channels can lead to computational redundancy and feature dispersion. By dynamically selecting the optimal number of channels, the improved model successfully achieves an effective balance between performance and computing resources.

[0105] The improved CBC architecture demonstrates strong adaptability and processing capabilities for short-term spectral denoising in high-noise environments. The high-quality spectra provide solid data support for real-time monitoring of spectral status, improving system stability, and optimizing efficiency. Future research will continue to explore the applicability of this method to various noise types and complex conditions, while also attempting to expand its application to a wider range of practical scenarios, providing more possibilities for precise control of spectral processing and efficient resource utilization.

[0106] The present invention comprises three parts: a data processing module, a model building module and a result evaluation module.

[0107] In order to reflect the effectiveness and comparability of the improved model, the data processing module of the present invention continues to use the previous research data set, which is a pair of high signal-to-noise ratio (long gate and average) and low signal-to-noise ratio (short gate and instantaneous) spectra collected in the steady-state calibration experiment to train and verify the model. The complete data set is then divided into a training set and a test set at a ratio of approximately 70% and 30%. The training set is used for model core parameter learning and feature extraction capability construction, and the test set is used for the final evaluation of the performance of unknown data after the model training optimization is completed. In the model training stage, the Adam optimizer is selected, the initial learning rate is set and the learning rate is dynamically adjusted in combination with the cosine annealing strategy. The training data is input in batches with a batch size of 128. After 50 training cycles of iterative optimization, the performance indicators of the verification set are stabilized.

[0108] The model building module is the core component of the present invention. It innovates on the traditional CNN model by introducing a channel attention mechanism. After adding the channel attention module to each layer of convolution operation, it can adaptively adjust the importance weights of each channel, enhance the network's ability to focus on key signals, and suppress the influence of noise. Experimental results show that the model proposed in this invention significantly improves the signal-to-noise ratio (SNR) in the denoising task of short-exposure flame emission spectra and achieves higher accuracy in the gas property estimation task. Compared with traditional denoising methods, the method of the present invention not only effectively improves signal quality, but also enhances the real-time monitoring accuracy of the combustion process.

[0109] The result evaluation module provides a scientific standard for measuring model performance. This module uses signal-to-noise ratio and mean absolute error (MAE) as core evaluation indicators. SNR can intuitively quantify the purity and clarity of the spectrum by accurately calculating the ratio of the effective signal power to the noise power in the spectral signal, and is a key indicator for evaluating the denoising effect of the model. MAE is used to evaluate the average deviation between the gas property values ​​predicted by the model and the true values, and can accurately reflect the accuracy and reliability of the model in the gas property prediction task. Through the comprehensive evaluation of these two indicators, the performance of the model in denoising and gas property prediction tasks can be comprehensively and accurately judged, providing a strong basis for model optimization and improvement.

[0110] This paper utilizes a convolutional neural network and a channel attention mechanism to adaptively adjust the importance weights of each spectral channel, enabling the network to focus more on key features while suppressing the influence of irrelevant noise, significantly improving denoising effectiveness. This approach has significant research value for applications such as denoising short-exposure flame emission spectra and high-precision flame emission spectrum analysis in combustion environments.

Claims

1. An improved CNN combined with channel attention mechanism for flame emission spectrum denoising method, characterized by: The following steps are involved: Obtain a spectral dataset and divide it into training set, validation set and test set; A flame emission spectrum denoising model was constructed. The model incorporates a channel attention mechanism after each convolution operation of a convolutional neural network and adaptively adjusts the weights of each channel. The input of the model is a plurality of downsampled spectral data, which are input into the model after preprocessing; the output of the model is the generated denoised spectrum; Use the training set to train the model, use the validation set to evaluate the model and tune hyperparameters, and then use the test set to evaluate the model's performance and measure the model's applicability and robustness in the application scenario.

2. The method according to claim 1, characterized in that The model includes several convolution blocks, batch normalization, activation functions and adaptive channel number selection modules; Convolution outputs multi-channel feature map F∈R C×H×W , where C is the number of channels, H×W is the spatial size, H is the height of the feature map, and W is the width of the feature map; The multi-channel feature map F output by the convolution is batch normalized. The model processes the multi-channel feature map F, and each channel undergoes a global average pooling operation to compress the spatial dimension into a feature vector z c ∈R C : Among them, F(c,i,j) represents the value of the c-th channel at the spatial position (i,j), z c Encodes the global semantic information of the c-th channel; Eigenvector z c The input is fed into a multilayer perceptron, a two-layer fully connected network, for nonlinear mapping to generate channel importance weights. The mapping process is as follows: y=ReLU(W1 z c +b1) Among them, y is the output result after the first layer operation, W1∈R d×C and b1∈R d are the weight and bias of the first layer, d is the intermediate dimension, and ReLU is the activation function; w c =σ(W2·y+b2) Among them, w c is the output result after the second layer operation, W2∈R C×d and b2∈R C are the weights and biases of the second layer respectively. The Sigmoid function σ maps the output to the range [0,1], and the generated weight vector w c Indicates the importance of each channel; Using the generated channel weight w c , the model applies dynamic weighting to each channel of the feature map F: in, Represents the weighted value of the c-th channel at the spatial position (i, j), and the weighted feature map is The adaptive channel number selection module dynamically adjusts the number of channels according to input features, where the input features are feature maps output by the convolutional layer; After multiple layers of convolution, adaptive weighting, and channel number selection, the model generates optimized feature maps. The final output is the spectral characteristics.

3. The method according to claim 1 or 2, characterized in that The adaptive channel number selection module dynamically adjusts the number of output channels C of each convolutional layer ′ , specifically: The adaptive channel number selection module adopts a strategy of gradually increasing the number of channels, with each increase in the number of channels as a stage. After each stage of channel number increase, the model evaluates the performance indicators; If the performance reaches its peak, stop adjusting the number of channels and select the number of channels C that provided the best performance in the previous round. opt As the final configuration.

4. An improved CNN combined with channel attention mechanism is used in flame emission spectrum denoising system, characterized by: It includes three parts: data processing module, model building module and result evaluation module; The data processing module is used to obtain a data set, wherein the data set is a pair of high signal-to-noise ratio and low signal-to-noise ratio spectra collected in a steady-state calibration experiment, which is used to train and verify the model; The data processing module is used to divide the complete data set into a training set and a test set in proportion; the training set is used for learning the core parameters of the model and building feature extraction capabilities, and the test set is used for the final evaluation of the unknown data performance after the model training optimization is completed; The model building module introduces a channel attention mechanism based on the traditional CNN model. The channel attention module is added after each convolution operation to adaptively adjust the importance weight of each channel. The result evaluation module is used to measure the model performance. The module uses signal-to-noise ratio and mean absolute error as core evaluation indicators; The result evaluation module is used to quantify the purity and clarity of the spectrum by calculating the ratio of the effective signal power to the noise power in the spectral signal; and is used to evaluate the average deviation between the gas property values ​​predicted by the model and the true values ​​through the mean absolute error, reflecting the accuracy and reliability of the model in the gas property prediction task.

Images