Multi-noise spectrum adaptive identification method based on deep learning

Abstract

The invention provides a multi-noise spectrum adaptive identification method based on deep learning, and the method comprises the steps: constructing a spectrum database containing various noises, and dividing the spectrum database into a training set, a verification set and a test set; constructing a noise classification model; constructing a spectrum denoising model; respectively training the noise classification model and the spectrum denoising model through the spectrum database; obtaining noise probability distribution of the noisy spectral data through the trained noise classification model; and inputting the noise probability distribution into the trained spectrum denoising model, and obtaining pure data through probability weighting operation. According to the method, noise classification and de-noising operations are realized through the constructed multiple network models, and an automatic model updating mechanism is introduced, so that the efficiency and precision of multi-noise identification are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of spectrum analysis and signal processing, and in particular to a multi-noise spectrum adaptive recognition method based on deep learning. Background Art

[0002] Spectral analysis, as an important analytical and testing method, is widely used in materials science, chemical analysis, biomedicine, environmental monitoring, and other fields. However, during the actual spectral acquisition process, due to factors such as instrument precision limitations, environmental interference, and random fluctuations, the acquired spectral data often contains various types of noise, which seriously affects the accuracy and reliability of subsequent analysis.

[0003] Traditional spectral denoising methods primarily include wavelet threshold denoising (DW), empirical mode decomposition (EMD), and Savitzky-Golay (SG) filtering. These methods primarily achieve noise reduction through varying degrees of signal smoothing. However, in practical applications with complex noise components and low signal-to-noise ratios, their noise separation capabilities and feature preservation remain to be improved. Furthermore, these methods often require manual configuration of multiple key parameters, and denoising results rely heavily on expert experience.

[0004] In recent years, some studies have attempted to apply convolutional neural networks (CNNs) and recurrent neural networks (RNNs) to spectral denoising. However, these methods are primarily designed for a single noise type and lack the ability to recognize multiple noise types or adaptive processing mechanisms. Furthermore, when new noise types emerge, existing methods often require retraining the entire model, which consumes large amounts of computational resources and can easily lead to "catastrophic forgetting," i.e., a degradation in performance for previously learned noise types. Therefore, it is crucial to design a deep learning-based adaptive multi-noise spectral recognition method. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-noise spectrum adaptive recognition method based on deep learning, which performs noise classification and denoising operations by constructing multiple network models and introduces a model automatic update mechanism to improve the efficiency and accuracy of multi-noise recognition.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A multi-noise spectrum adaptive recognition method based on deep learning includes the following steps:

[0008] Construct a spectral database containing various noises and divide the spectral database into training set, validation set and test set;

[0009] Construct a noise classification model; the noise classification model consists of a connected feature extraction module and a classification decision module;

[0010] Construct a spectral denoising model; the spectral denoising model consists of an encoder and a decoder, and the encoder and decoder are jump-connected;

[0011] The noise classification model and spectral denoising model are trained separately through the spectral database;

[0012] Obtain the noise probability distribution of noisy spectral data through the trained noise classification model;

[0013] The noise probability distribution is input into the trained spectral denoising model, and the clean data is obtained through probability weighting operation.

[0014] Optionally, a spectral database containing various noises is constructed, and the spectral database is divided into a training set, a validation set, and a test set, including:

[0015] Constructing a simulated reference spectrum based on an asymmetric Gaussian function;

[0016] Gaussian white noise, pink noise, impulse noise and mixed noise of Gaussian white noise, pink noise and impulse noise of different intensities are added to the simulated reference spectrum to obtain noisy spectrum samples;

[0017] Based on the preset basic spectrum, the noisy spectrum samples are paired and integrated to obtain a spectrum database;

[0018] The spectral database is divided into training set, validation set and test set.

[0019] Optionally, the feature extraction module consists of three convolutional layers and two maximum pooling layers connected in sequence; the convolution kernel and padding of the convolution layer are both decreasing strategies, the decreasing order of the convolution kernel is 9×1, 5×1 and 3×1, and the decreasing order of padding is 4, 2 and 1; the step size of the convolution layer is 1; the size and step size of the maximum pooling layer are both 2; the convolution layer has a built-in batch normalization function and ReLU activation function.

[0020] Optionally, the classification decision module consists of a flattening layer, three fully connected layers and an output layer connected in sequence; the flattening layer is used to quantize the output of the feature extraction module into 64,000-dimensional features; the number of nodes in the fully connected layers is 512, 128 and 4 respectively, and the Dropout strategy is used between the fully connected layers; the output layer has a built-in linear activation function and a cross-entropy loss function, and the output layer is used to output unnormalized probability values.

[0021] Optionally, the encoder consists of three cascaded encoding units, each consisting of a connected convolutional layer and a maximum pooling layer; the convolutional layer has a convolution kernel of 3×1, a stride of 1, and a padding of 1; the maximum pooling layer has a receptive field size of 2×1 and a stride of 2;

[0022] The decoder and encoder have a mirror-symmetrical structure; the decoding unit of the decoder consists of a connected upsampling layer and a transposed convolution layer; the amplification factor of the upsampling layer is 2; the output ends of the convolution layer and the transposed convolution layer are both set with batch normalization function and PReLU activation function.

[0023] Optionally, the step of training the noise classification model using the spectral database includes:

[0024] Based on the training set, the noise classification model is trained through the cross entropy loss function to obtain the loss value;

[0025] The gradient of the loss value is calculated through back propagation, and the parameters of the noise classification model are updated using the Adam optimizer to obtain the candidate model;

[0026] The candidate model is verified through the validation set to obtain the validation accuracy. When the validation accuracy does not exceed the optimal candidate model for 5 consecutive cycles, the training is stopped and the parameters of the optimal candidate model are retained.

[0027] The confusion matrix of the optimal candidate model is drawn through the test set for model testing to obtain the trained noise classification model.

[0028] Optionally, the step of training the spectral denoising model using the spectral database includes:

[0029] Based on the training set, the coding unit is trained through the MSE loss function to obtain the loss value;

[0030] Based on the loss value, the parameters of the coding unit are updated through back propagation to obtain the candidate coding unit;

[0031] The validation loss of the candidate coding unit is calculated using the validation set. When the validation loss does not exceed the optimal candidate coding unit for 10 consecutive cycles, the training is stopped and the parameters of the optimal candidate coding unit are retained.

[0032] After all encoding units are trained, use the Adam optimizer and learning rate scheduler to stack the trained encoding units to obtain the trained encoder;

[0033] The signal-to-noise ratio and percentage root mean square error of the trained encoder are obtained through the test set for model testing, and the trained spectral denoising model is obtained.

[0034] Optionally, the noise probability distribution is input into the trained spectral denoising model to obtain clean data through probability weighting operations, including:

[0035] Input the noisy spectral data into the trained spectral denoising model to obtain multiple denoising results;

[0036] The noise probability distribution is used as the weight to perform weighted combination on the denoising results to obtain pure data.

[0037] Optionally, the method further includes: when the noise classification model cannot identify the noise type, constructing a new denoising branch and adding it to the noise classification model to obtain an updated noise classification model, specifically including:

[0038] When the highest classification probability value of the noise classification model is lower than the preset threshold, the noisy spectral data is marked as a new noise type;

[0039] Build a new denoising branch and incrementally train it with the new noise type;

[0040] The trained denoising branch is added to the noise classification model, and the parameters of the noise classification model are kept unchanged to obtain an updated noise classification model.

[0041] Optionally, construct a new denoising branch and incrementally train it with the new noise type, including:

[0042] Based on the principle of transfer learning, the output layer of the noise classification model is initialized and expanded using the Xavier method;

[0043] The expanded output layer is trained through the regularization loss function to obtain the trained denoising branch; the regularization loss function is obtained by weighting the cross entropy loss function and the knowledge distillation loss function.

[0044] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: the multi-noise spectral adaptive recognition method based on deep learning provided by the present invention comprises: constructing a spectral database containing multiple noises, and dividing the spectral database into a training set, a validation set, and a test set; constructing a noise classification model; constructing a spectral denoising model; training the noise classification model and the spectral denoising model respectively through the spectral database; obtaining the noise probability distribution of the noisy spectral data through the trained noise classification model; inputting the noise probability distribution into the trained spectral denoising model, and obtaining pure data through probability weighting operation. The method realizes noise classification and denoising operations through the construction of multiple network models, and introduces a model automatic update mechanism, thereby improving the efficiency and accuracy of multi-noise recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1 This is a flow chart of the multi-noise spectrum adaptive recognition method of the present invention;

[0047] Figure 2 Schematic diagram of a spectral denoising model according to an embodiment of the present invention;

[0048] Figure 3 This is a confusion matrix diagram of the spectrum after being classified by the noise classification model according to an embodiment of the present invention;

[0049] Figure 4 This is a comparison diagram of a spectral signal with a signal-to-noise ratio of 5dB before and after denoising according to an embodiment of the present invention;

[0050] Figure 5 The SNR and RMSE comparison chart of the method of the present invention and other denoising methods;

[0051] Figure 6 Construct a flow chart for the new denoising branch of the present invention. DETAILED DESCRIPTION

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0053] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] like Figure 1 As shown, the present invention provides a multi-noise spectrum adaptive recognition method based on deep learning, comprising the following steps:

[0055] Step 100: constructing a spectral database containing various noises, and dividing the spectral database into a training set, a validation set, and a test set;

[0056] Specifically, a simulated baseline spectrum is first constructed based on an asymmetric Gaussian function. To this, various intensities of Gaussian white noise, pink noise, impulse noise, and a mixture of the three are added to generate a variety of noisy spectral samples. These noisy spectra are then paired and integrated with the corresponding baseline spectra to construct a complete spectral database. Finally, all simulated spectra in the spectral database are divided into training, validation, and test sets.

[0057] Furthermore, the reference spectrum signal is generated by the following formula:

[0058]

[0059] Where I(λ) represents the spectral intensity at wavelength λ, I0 is the peak intensity, λ is the wavelength of the spectrum, and λ B is the central wavelength, W is the spectrum width parameter, v is the shape parameter, which is used to control the steepness of the spectrum peak, χ is the asymmetry coefficient, which is used to control the asymmetric distortion of the spectrum, the value range of I0 is [0.8, 0.85], λ B The value range of is [1545.5, 1546.5], the value range of W is [0.2, 0.22], the value of v is 2, and the value range of χ is [1.5, 1.7]. The expression of the spectral signal containing Gaussian white noise is:

[0060] y(t)=s(t)+n w (t);

[0061] Among them, s(t) is the original signal, n w (t) is Gaussian white noise, and the standard deviation of Gaussian white noise is calculated as: The spectral signal expression of pink noise is:

[0062]

[0063] Among them, P s is the average power of the signal, which is calculated as n′ p (t) is noise with spectral characteristics. The spectral signal expression of impulse noise is:

[0064]

[0065] The standard deviation of the impulse noise amplitude is The expressions of the three mixed noises are: Each of the four noise types was power-adjusted to ensure the final noisy signal achieved the desired SNR. This resulted in 60,000 noise signals of each type, with SNRs ranging from 5 to 30 dB. 120,000 of these data pairs were used for the noise classification model, while the spectral denoising model used 60,000 pairs.

[0066] Step 200: Construct a noise classification model; the noise classification model consists of a connected feature extraction module and a classification decision module;

[0067] Specifically, the feature extraction module consists of three sequentially connected convolutional layers and two maximum pooling layers, capable of adaptively extracting deep features from spectral data containing different types of noise. The three convolutional layers use a decreasing convolution kernel strategy, using 9×1, 5×1, and 3×1 convolutions, respectively, with corresponding padding of 4, 2, and 1, and a stride of 1 for all convolutional layers. The window size and stride of the two maximum pooling layers are both 2, reducing the feature dimensionality. Batch normalization and ReLU activation functions are introduced after each convolutional layer.

[0068] Specifically, the classification decision module consists of a flattening layer, three fully connected layers and an output layer connected in sequence, which is used to accurately identify the noise type based on the extracted features; the flattening layer is used to quantize the output feature vector of the feature extraction module into 64,000-dimensional features; the number of nodes in the subsequent three fully connected layers are 512, 128 and 4 respectively, and a hierarchically differentiated Dropout strategy is adopted between the fully connected layers; the output layer has a built-in linear activation function and a cross-entropy loss function, which directly outputs the unnormalized probability values of the four types of noise, and end-to-end model optimization is achieved through the cross-entropy loss function.

[0069] It should be noted that the input data dimension of the noise classification model is 1000×1, the batch size is set to 64, and the feature extraction module realizes local adjustment and dimensionality reduction of the captured signal through a combination of convolutional layers and pooling layers. The dimension of the output data becomes 256×250, and the number of channels is 256. It then enters the classification module, gradually reduces the dimension through the fully connected layer, and finally outputs a 4-dimensional vector, which is the probability score of the corresponding signal belonging to the four noise types. In this embodiment, the inactivation rate of the first layer after full connection is 50%, and the inactivation rate of the second layer is reduced to 30%. This dynamic sparsification mechanism effectively suppresses overfitting and achieves the optimal balance between model capacity and generalization ability.

[0070] Step 300: Construct a spectral denoising model; the spectral denoising model consists of an encoder and a decoder, and the encoder and the decoder are jump-connected;

[0071] like Figure 2 As shown in Figure 2, the spectral denoising model employs a symmetric encoder-decoder architecture and introduces a skip connection mechanism. This model uses a noise-contaminated spectral signal as input and achieves signal reconstruction and denoising through deep learning. Its encoder consists of three cascaded encoding units, each of which is composed of a connected convolutional layer and a maximum pooling layer. All convolutional layers use a 3×1 convolution kernel with a stride of 1 and padding of 1. The maximum pooling layer has a receptive field size of 2×1 and a stride of 2. The encoder gradually maps the noisy spectrum into a low-dimensional feature space, thereby extracting the essential characteristic representation of the signal.

[0072] The decoder and encoder have mirror-symmetric structures. Each decoding unit in the decoder consists of a connected upsampling layer and a transposed convolutional layer. The upsampling layer has a magnification factor of 2, and the parameters of the transposed convolutional layer are consistent with those of the corresponding encoding layer. Batch normalization and the Pre-Reduced Unit (PReLU) activation function are introduced after each convolutional layer and transposed convolutional layer.

[0073] Step 400: training the noise classification model and the spectrum denoising model respectively through the spectrum database;

[0074] Specifically, the steps for training the noise classification model using the spectral database are as follows: first, the cross entropy loss function is used as the basic loss function of the noise classification model for training, and the loss value is calculated using the following formula:

[0075]

[0076] where p(x i ) is the true category of the sample, q(x i ) is the probability that the model predicts this category. The gradient of the loss value is then calculated by backpropagation, and the parameters of the noise classification model are updated using the Adam optimizer. The learning rate is initially set to 0.0005, and the batch size is 32. When the verification loss does not improve for three consecutive cycles, the learning rate is automatically halved to improve the optimization accuracy in the later stages of training. The candidate model is then verified using the verification set to obtain the verification accuracy. When the verification accuracy does not exceed the current best model for five consecutive cycles, the accuracy threshold of the best model in this embodiment is 0.0001, the training process is terminated early, and the best model parameters are saved. Finally, the confusion matrix of the best model is drawn out through the test set for model testing, thereby obtaining the trained noise classification model. In this embodiment, the confusion matrix of the spectrum after classification by the noise classification model is as follows: Figure 3 shown.

[0077] Specifically, the spectral denoising model is divided into two training stages. First, the stacked coding units are greedily trained layer by layer. The initial learning rate is 0.001, which decreases as the number of layers increases. Based on the training set, the MSE loss of the spectrum before and after denoising the coding unit is calculated using the MSE loss function to obtain the loss value. Then, the parameters of the coding unit are updated through backpropagation. The calculation formula of the MSE loss is:

[0078]

[0079] Where n is the number of spectral sampling points, which is 1000 in this embodiment, and y i is the spectral intensity at sampling point i, is the spectral intensity predicted by the model at sampling point i. The validation loss is calculated using the validation set for each training cycle. When the validation loss does not exceed the current best coding unit for 10 consecutive cycles, the training is stopped and the parameters of the best coding unit are retained. After completing the training of all coding units, the trained coding units are stacked using the Adam optimizer and the learning rate scheduler to obtain the trained encoder. The initial learning rate is 0.001, which decreases as the number increases, that is, 1 / 2 is used for the kth coding unit. k The learning rate scaling factor is . Finally, the trained spectrum denoising model is tested using the test set. The signal-to-noise ratio (SNR) and percentage root mean square error (RMSE) of the spectrum denoising with different signal-to-noise ratios are tested. In this embodiment, the comparison of the spectrum signal with a signal-to-noise ratio of 5dB before and after denoising is as follows: Figure 4 As shown in the figure, the SNR and RMSE comparison of this embodiment and other denoising methods are shown in Figure 5 As shown, CSDAE is the spectral denoising model of this embodiment.

[0080] It's important to note that SNR is a key indicator of signal quality; a larger value indicates better signal quality. RMSE, on the other hand, directly measures the quality of signal reconstruction, maintaining the same dimension as the original signal. A smaller value indicates better reconstruction.

[0081] Step 500: Obtaining noise probability distribution of noisy spectral data through the trained noise classification model;

[0082] Specifically, the noisy spectrum to be processed is input into the trained noise classification model, and then the probability distribution vector P = [P1, P2, ..., P n ], where P i represents the probability of the i-th type of noise, and ∑P i =1.

[0083] Step 600: Input the noise probability distribution into the trained spectral denoising model, and obtain clean data through probability weighting operation.

[0084] Specifically, the noisy spectra to be processed are simultaneously input into the trained spectrum denoising model to obtain n denoising results S1, S2, ... S n ; Using the probability distribution vector as the weight, perform weighted combination on the n denoising results to calculate the final denoising spectrum S, i.e., the pure data. The calculation formula is: S = ∑(P i ×S i ).

[0085] Furthermore, when the noise classification model cannot identify the noise type, a new denoising branch is constructed and added to the noise classification model to obtain an updated noise classification model. The specific steps are as follows: Figure 6 As shown, including:

[0086] Step 601: When the highest classification probability value of the noise classification model is lower than a preset threshold, the noisy spectral data is marked as a new noise type;

[0087] Specifically, the preset threshold τ in this embodiment is 0.3, and such sample data is systematically collected and labeled while being marked.

[0088] Step 602: construct a new denoising branch and perform incremental training on the denoising branch using the new noise type;

[0089] Specifically, based on the principle of transfer learning, the original n-class output is expanded to n+1 classes, the weight parameters of the original n classes are retained, and only the connection weights related to the newly added classes are randomly initialized. The expanded expression is:

[0090] W n+1 =[W n |W new ];

[0091] Among them, W n is the weight matrix corresponding to the original n types of noise, W new The weight vector corresponding to the newly added noise category is initialized using the Xavier method. Then, the expanded noise classification model is trained by combining knowledge distillation with incremental learning, and a regularization loss term is introduced to prevent catastrophic forgetting. The expression of the regularization loss term is:

[0092] L total =a·L CE +(1-a)·L KD ;

[0093] Among them, L CE is the cross entropy loss function, L KD is the knowledge distillation loss function, α is the balancing factor. During training, only the parameters related to the new noise type are updated, and the learning rate is set to 50% of the base learning rate.

[0094] Step 603: Add the trained denoising branch to the noise classification model, and keep the parameters of the noise classification model unchanged to obtain an updated noise classification model.

[0095] The beneficial effects of the present invention are as follows:

[0096] 1) Through noise type identification and probability weighted combination strategy, accurate removal of multiple noise types is achieved, improving denoising accuracy;

[0097] 2) Through the adaptive expansion mechanism, when the system detects a new noise type, it only needs to build and train a new denoising branch while freezing the parameters of the original branch. This avoids the resource consumption and "catastrophic forgetting" problem caused by retraining the entire model, achieves rapid adaptation to new noise types, and enhances generalization and adaptability.

[0098] 3) Deep learning models are used to automatically identify noise types and select denoising strategies, eliminating the need for manual parameter setting. This reduces reliance on expert experience and improves the automation level of the denoising process.

[0099] 4) A spectral denoising model based on a symmetric codec architecture and a skip connection mechanism, combined with deep feature extraction technology, improves the degree of detail preservation of the original spectrum;

[0100] 5) A phased training strategy of layer-by-layer greedy training and end-to-end fine-tuning, combined with dynamic learning rate adjustment and early stopping mechanism, shortens model training time, reduces computing resource consumption, and improves training efficiency.

[0101] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0102] The present invention uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A multi-noise spectrum adaptive recognition method based on deep learning, characterized in that: The steps include: Constructing a spectral database containing various noises, and dividing the spectral database into a training set, a validation set, and a test set; Constructing a noise classification model; the noise classification model consists of a connected feature extraction module and a classification decision module; Construct a spectral denoising model; The spectral denoising model consists of an encoder and a decoder, and the encoder and the decoder are jump-connected; Training the noise classification model and the spectrum denoising model respectively through the spectrum database; Obtaining the noise probability distribution of the noisy spectral data through the trained noise classification model; The noise probability distribution is input into the trained spectrum denoising model, and clean data is obtained through probability weighting operation.

2. The multi-noise spectrum adaptive recognition method based on deep learning according to claim 1 is characterized in that A spectral database containing various noises is constructed, and the spectral database is divided into a training set, a validation set, and a test set, including: Constructing a simulated reference spectrum based on an asymmetric Gaussian function; Adding Gaussian white noise, pink noise, impulse noise and a mixed noise of the Gaussian white noise, the pink noise and the impulse noise of different intensities to the simulated reference spectrum to obtain a noisy spectrum sample; Based on a preset basic spectrum, the noisy spectrum samples are paired and integrated to obtain the spectrum database; The spectral database is divided into the training set, the validation set and the test set.

3. The multi-noise spectrum adaptive recognition method based on deep learning according to claim 1 is characterized in that: The feature extraction module consists of three convolutional layers and two maximum pooling layers connected in sequence; the convolution kernel and padding of the convolution layer are both decreasing strategies, the decreasing order of the convolution kernel is 9×1, 5×1 and 3×1, and the decreasing order of the padding is 4, 2 and 1; the step size of the convolution layer is 1; the size and step size of the maximum pooling layer are both 2; the convolution layers all have built-in batch normalization function and ReLU activation function.

4. The multi-noise spectrum adaptive recognition method based on deep learning according to claim 1 is characterized in that The classification decision module consists of a flattening layer, three fully connected layers and an output layer connected in sequence; the flattening layer is used to quantize the output of the feature extraction module into 64,000-dimensional features; the number of nodes in the fully connected layers is 512, 128 and 4 respectively, and the Dropout strategy is used between the fully connected layers; the output layer has a built-in linear activation function and a cross-entropy loss function, and the output layer is used to output unnormalized probability values.

5. The multi-noise spectrum adaptive recognition method based on deep learning according to claim 1 is characterized in that: The encoder consists of three cascaded encoding units, each consisting of a connected convolutional layer and a maximum pooling layer; the convolution kernel of the convolution layer is 3×1, the stride is 1, and the padding is 1; the receptive field size of the maximum pooling layer is 2×1, and the stride is 2; The decoder and the encoder have a mirror-symmetrical structure; the decoding unit of the decoder consists of a connected upsampling layer and a transposed convolution layer; the amplification factor of the upsampling layer is 2; The output ends of the convolutional layer and the transposed convolutional layer are both provided with a batch normalization function and a PReLU activation function.

6. The multi-noise spectrum adaptive recognition method based on deep learning according to claim 1 is characterized in that: The step of training the noise classification model using the spectral database includes: Based on the training set, performing loss training on the noise classification model using a cross entropy loss function to obtain a loss value; Calculate the gradient of the loss value by backpropagation, and use the Adam optimizer to update the parameters of the noise classification model to obtain a candidate model; Verify the candidate model using the verification set to obtain a verification accuracy rate, and stop training when the verification accuracy rate does not exceed the optimal candidate model for five consecutive cycles and retain the parameters of the optimal candidate model; The optimal confusion matrix of the candidate model is drawn out through the test set to perform model testing, thereby obtaining the trained noise classification model.

7. The multi-noise spectrum adaptive recognition method based on deep learning according to claim 5 is characterized in that: The step of training the spectral denoising model using the spectral database includes: Based on the training set, performing loss training on the coding unit by using an MSE loss function to obtain a loss value; Based on the loss value, updating the parameters of the coding unit by backpropagation to obtain a candidate coding unit; Calculating the verification loss of the candidate coding unit using the verification set, and stopping training when the verification loss does not exceed the optimal candidate coding unit for 10 consecutive cycles and retaining the parameters of the optimal candidate coding unit; After all the encoding units are trained, the trained encoding units are stacked using the Adam optimizer and the learning rate scheduler to obtain a trained encoder; The signal-to-noise ratio and percentage root mean square difference of the trained encoder are obtained through the test set to perform model testing, thereby obtaining the trained spectral denoising model.

8. The multi-noise spectrum adaptive recognition method based on deep learning according to claim 1 is characterized in that: The noise probability distribution is input into the trained spectral denoising model, and clean data is obtained through probability weighting operations, including: Inputting the noisy spectral data into the trained spectral denoising model to obtain multiple denoising results; The noise probability distribution is used as a weight to perform weighted combination on the denoising results to obtain the clean data.

9. The multi-noise spectrum adaptive recognition method based on deep learning according to claim 1 is characterized in that: Also includes: When the noise classification model cannot identify the noise type, a new denoising branch is constructed and added to the noise classification model to obtain an updated noise classification model, specifically including: When the highest classification probability value of the noise classification model is lower than a preset threshold, marking the noisy spectral data as a new noise type; Constructing a new denoising branch and performing incremental training on the denoising branch using the new noise type; The trained denoising branch is added to the noise classification model, and the parameters of the noise classification model are kept unchanged to obtain an updated noise classification model.

10. The multi-noise spectrum adaptive recognition method based on deep learning according to claim 9 is characterized in that: Constructing a new denoising branch and performing incremental training on the denoising branch using the new noise type, comprising: Based on the principle of transfer learning, the output layer of the noise classification model is initialized by the Xavier method, and the output layer is expanded; The expanded output layer is trained by a regularization loss function to obtain the trained denoising branch; the regularization loss function is obtained by weighting the cross entropy loss function and the knowledge distillation loss function.

Images