An environment-adaptive deep learning fusion optical signal noise suppression method and system

Abstract

This invention provides an environment-adaptive deep learning-based fusion method and system for optical signal noise suppression. The invention classifies environment types based on K-means clustering, pre-trains three sub-models (CNN, LSTM, and Transformer), and initializes an environment-adaptive decision-maker based on a BP neural network. The decision-maker performs nonlinear feature extraction and parameter coupling correlation mining on real-time environment parameters, outputting the similarity between the real-time environment and each predefined environment type to identify the current environment type and select the optimal sub-model combination. An attention mechanism is used to fuse the noise reduction results of each sub-model, and the final denoised optical signal is obtained through inverse standardization. Incremental model updates are achieved by combining noise reduction effect evaluation. This invention possesses strong environment adaptability, accurately adapting to the noise characteristics of complex and changing environments, balancing noise reduction accuracy and real-time performance, effectively preserving the essential details of the optical signal, and significantly improving the stability and reliability of optical transmission systems.

Description

Technical Field

[0001] This invention relates to the field of optical signal transmission and processing technology, and in particular to an environment-adaptive deep learning-based optical signal noise suppression method and system. Background Technology

[0002] During transmission, optical signals are susceptible to external environmental factors (such as temperature fluctuations, humidity changes, vibration interference, atmospheric scattering / absorption, and transmission link loss) and internal equipment noise (such as light source noise, detector noise, and circuit thermal noise), leading to optical signal distortion and a decrease in signal-to-noise ratio. This severely affects the communication quality, sensing accuracy, and transmission distance of optical transmission systems. Therefore, optical signal noise suppression is one of the core technologies in the field of optical signal processing.

[0003] Currently, existing optical signal noise suppression methods are mainly divided into two categories: traditional filtering methods and deep learning-based noise suppression methods. Traditional filtering methods (such as Kalman filtering, Wiener filtering, and wavelet filtering) rely on the prior statistical characteristics of noise and achieve noise suppression through fixed filtering rules. However, when faced with complex and changing environments (such as sudden changes in environmental parameters or the superposition of multiple noises), their adaptive adjustment ability is poor, making it difficult to accurately match the dynamic changes in noise characteristics. This results in poor noise suppression performance and the easy loss of effective details in the optical signal, which in turn affects the accuracy of subsequent signal demodulation and recognition.

[0004] With the development of deep learning technology, deep learning-based optical signal noise suppression methods are gradually replacing traditional methods. These methods learn the characteristic differences between noise and effective signals through training on a large number of samples, achieving accurate noise separation. However, existing deep learning-based methods still have significant drawbacks: First, they lack environmental adaptability. Most deep learning models use fixed structures and parameters, which cannot dynamically adjust the model structure or parameters according to changes in real-time environmental parameters (temperature, humidity, vibration frequency, etc.). When the environment changes abruptly, the model's generalization ability is insufficient, and the noise reduction effect drops sharply. Second, the fusion strategy is simplistic. Some multi-model fusion methods only use a simple weighted summation method to fuse the outputs of each sub-model, failing to fully explore the noise reduction advantages of each sub-model in different environments, and thus failing to achieve the optimal match between "environment-model-noise reduction effect". Third, it is difficult to balance model training and inference efficiency. While complex models can improve noise reduction accuracy, their inference speed is slow, which cannot meet the real-time processing requirements of optical signal transmission. On the other hand, lightweight models have fast inference speed, but their noise reduction accuracy is limited.

[0005] Therefore, there is an urgent need for an environment-adaptive deep learning fusion optical signal noise suppression method that can sense environmental changes in real time, dynamically adjust model strategies, and balance noise reduction accuracy and real-time performance, in order to solve technical problems such as poor optical signal noise suppression effect, weak adaptability, and insufficient real-time performance in complex and variable environments. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an environment-adaptive deep learning fusion optical signal noise suppression method and system. This invention achieves accurate and efficient suppression of optical signal noise under different environments through real-time perception of environmental parameters, adaptive model selection and parameter adjustment, and weighted fusion of multiple deep learning sub-models, balancing noise reduction accuracy and real-time processing requirements, and improving the stability and reliability of optical transmission systems.

[0007] In a first aspect, the present invention provides an environment-adaptive deep learning-based optical signal noise suppression method, comprising the following steps:

[0008] S1) Collect real-time environmental parameters and noisy optical signals in the optical transmission link; and preprocess the environmental parameters and noisy optical signals;

[0009] S2) Based on the K-means clustering algorithm, historical environmental parameter samples are clustered to classify different environmental types;

[0010] S3) Pre-train the corresponding deep learning sub-model for each environment type and initialize the environment adaptive decision-maker;

[0011] S4) Based on standardized real-time environmental parameters, an environmental adaptive decision-maker identifies the similarity between standardized real-time environmental parameters and environmental types, and selects the optimal combination of sub-models based on the environmental type similarity.

[0012] S5) Input the preprocessed noisy optical signal into the selected optimal sub-model combination. Each sub-model completes noise suppression and outputs preliminary noise reduction results. The outputs of each sub-model are dynamically fused using an attention mechanism to obtain the final noise-reduced optical signal. Then, the original features of the optical signal are restored through inverse normalization processing.

[0013] S6) Quantitatively evaluate the denoising effect of the final denoised optical signal to determine whether it meets the actual application requirements; if not, update the parameters of the sub-model and the environment adaptive decision-maker through incremental training.

[0014] Preferably, in step S2), the historical environmental parameter samples are clustered based on the K-means clustering algorithm to classify different environmental types, specifically including the following steps:

[0015] S21) Collect historical environmental parameter samples and perform preprocessing on the historical environmental parameter samples as in step S1) to obtain a standardized historical environmental parameter sample set.

[0016] S22) Based on the distribution characteristics of historical environmental parameter samples, calculate the average profile coefficient corresponding to different cluster numbers K values, and select the average profile coefficient. The largest cluster number K value is taken as the final cluster number;

[0017] S23) Randomly select K samples from the standardized historical environmental parameter sample set as the initial cluster centers;

[0018] S24) Calculate the Euclidean distance between each historical environmental parameter sample and each cluster center, and assign the historical environmental parameter sample to the nearest cluster;

[0019] S25) Calculate the mean of all historical environmental parameter samples within each cluster and use it as the new cluster center;

[0020] S26) Repeat steps S24) and S25) until the change in cluster centers is less than the set threshold; or the number of iterations reaches the maximum number of iterations, then stop the iteration and obtain K environment types;

[0021] S27) Based on the clustering results, analyze the distribution range of environmental parameters for each cluster, define the environmental type, and clarify the parameter threshold range for each environmental type.

[0022] Preferably, in step S3), the deep learning sub-model includes a CNN sub-model, an LSTM sub-model, and a Transformer sub-model; wherein, the CNN sub-model is used to extract local noise features of the noisy light signal; the LSTM sub-model is used to extract temporal features of the noisy light signal; and the Transformer sub-model is used to extract global correlation features of the noisy light signal, including the correlation between noise and the environment.

[0023] Preferably, in step S3), the Transformer sub-model includes a two-layer Transformer encoder; the two-layer Transformer encoder is used to extract global correlation features of the optical signal, including noise and environmental correlation.

[0024] Preferably, in step S3), a BP neural network is used to construct an environment adaptive decision-maker, which is used to select the optimal sub-model combination and fusion weights and dynamically fine-tune the sub-model parameters according to real-time environment parameters; the environment adaptive decision-maker includes an input layer, 2-3 hidden layers, and an output layer.

[0025] The input layer is used to input standardized environmental parameters, and the hidden layer uses the ReLU activation function to perform nonlinear feature extraction on the input environmental parameter vector through the hidden layer, thereby mining the coupling correlation between various environmental parameters and obtaining high-dimensional correlation features.

[0026] The output layer maps the high-dimensional correlation features of the hidden layer to the [0,1] interval through the sigmoid activation function, and outputs the K-dimensional environment type similarity.

[0027] Preferably, step S4) specifically includes the following steps:

[0028] S41) The standardized real-time environmental parameters are input into the environment adaptive decision-maker, which calculates the similarity between the standardized real-time environmental parameters and the cluster centers of each predefined environment type using cosine similarity.

[0029] S42) Select the environment type with the highest similarity as the current environment type;

[0030] S43) Select the optimal combination of sub-models based on the current environment type and the adaptation relationship between each environment type and the sub-model determined in the pre-training stage;

[0031] S44) The environment adaptive decision-maker dynamically fine-tunes the key parameters of the selected sub-model based on the real-time changing trends of environmental parameters. The fine-tuning objective is to maximize the signal-to-noise ratio after noise reduction. The gradient descent method is used for parameter updates.

[0032] Preferably, in step S43), the optimal sub-model combination is selected, specifically as follows:

[0033] If a stable environment is required, choose the CNN sub-model;

[0034] For environments with temperature fluctuations, an LSTM sub-model is selected; for environments with vibration disturbances, a combination of a CNN sub-model and an LSTM sub-model is selected.

[0035] In multi-factor mixed interference environments, a combination of CNN sub-model, LSTM sub-model, and Transformer sub-model is selected.

[0036] Preferably, step S5) specifically includes the following steps:

[0037] S51) Input the segmented noisy optical signal samples into the selected optimal sub-model combination. Each sub-model in the optimal sub-model combination performs noise reduction processing on the segmented noisy optical signal samples and outputs its own preliminary noise reduction results.

[0038] If the optimal combination of sub-models consists of multiple sub-models, then each sub-model outputs a preliminary denoising sample, resulting in a preliminary denoising result set.

[0039] S52) The preliminary denoising results of each sub-model are fused using a multi-head self-attention mechanism to obtain the standardized final denoised optical signal sample;

[0040] S53) Perform denormalization on the final denoised optical signal sample after standardization to restore the amplitude, phase, and frequency characteristics of the original optical signal.

[0041] Preferably, step S6) specifically includes the following steps:

[0042] S61) Calculate the signal-to-noise ratio (SNR), mean square error (MSE), and peak signal-to-noise ratio (PSNR) of the final denoised optical signal; if the signal-to-noise ratio (SNR), mean square error (MSE), and peak signal-to-noise ratio (PSNR) of the final denoised optical signal all meet the threshold requirements, then output the final denoised optical signal.

[0043] If any evaluation indicator fails to meet the threshold requirement, the noise reduction effect is deemed unqualified; proceed to step S62).

[0044] S62) The current standardized environmental parameters, noisy light signal samples and corresponding final denoised light signal samples, and real effective light signal samples are used as new incremental training samples and added to the training sample set of the corresponding environment type.

[0045] S63) Use a mini-batch training method to incrementally train the selected sub-model;

[0046] S64) Use the current environmental parameters and the corresponding optimal sub-model combination and parameter adjustment coefficients as training samples to incrementally train the environmental adaptive decision-maker;

[0047] (S65) After incremental training is completed, save the updated sub-model parameters and decision-maker parameters for subsequent noise suppression processing; at the same time, add the current environmental parameter samples to the historical environmental parameter sample set. If the addition of samples causes a change in the distribution of environmental types, re-execute the K-means clustering in step S2) to update the environmental type classification results.

[0048] Secondly, the present invention provides an environment-adaptive deep learning-based optical signal noise suppression system, comprising:

[0049] The data acquisition module is used to acquire real-time environmental parameters and noisy optical signals in the optical transmission link; and to preprocess the environmental parameters and noisy optical signals.

[0050] The environment type classification module uses the K-means clustering algorithm to cluster historical environment parameter samples and classify them into different environment types.

[0051] The training module is used to pre-train the corresponding deep learning sub-model for each environment type and initialize the environment adaptive decision-maker;

[0052] The model selection module, based on standardized real-time environmental parameters, identifies the similarity between standardized real-time environmental parameters and environmental types through an environmental adaptive decision-maker, and selects the optimal combination of sub-models based on the environmental type similarity.

[0053] The noise reduction module uses the selected optimal sub-model combination to suppress noise in the preprocessed noisy optical signal and outputs the preliminary noise reduction result. It also uses an attention mechanism to dynamically fuse the outputs of each sub-model to obtain the final noise-reduced optical signal. Finally, it restores the original features of the optical signal through denormalization.

[0054] The evaluation module quantitatively evaluates the denoising effect of the final denoised optical signal to determine whether it meets the actual application requirements; if not, it updates the parameters of the sub-model and the environment adaptive decision-maker through incremental training.

[0055] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the optical signal noise suppression method.

[0056] The beneficial effects of this invention are as follows:

[0057] 1. This invention synchronously collects real-time environmental parameters, combines K-means clustering to classify environmental types, and uses an environmental adaptive decision-maker to dynamically select the optimal sub-model combination and fine-tune parameters. It can accurately adapt to the noise characteristics under different environments and solve the technical problems of poor environmental adaptability and sharp decline in noise reduction effect when the environment changes suddenly in existing methods.

[0058] 2. This invention employs a fusion strategy of three sub-models: CNN, LSTM, and Transformer, to extract local, temporal, and global features of the optical signal respectively. Combined with an attention mechanism, the fusion weights are dynamically allocated to fully leverage the noise reduction advantages of each sub-model. This enables accurate separation of various types of noise while preserving the effective details in the optical signal to the maximum extent.

[0059] 3. This invention significantly reduces the computational complexity of model inference and updates through sub-model pre-training, parameter fine-tuning optimization, and small-batch incremental training, keeping the inference latency within 10ms, thus meeting the real-time processing requirements of optical signal transmission. Attached Figure Description

[0060] Figure 1 This is a flowchart illustrating the method of Embodiment 1 of the present invention;

[0061] Figure 2 This is a schematic diagram of the process for selecting the optimal combination of sub-models in Embodiment 1 of the present invention;

[0062] Figure 3This is a schematic diagram of the process of denoising and restoring a noisy optical signal according to Embodiment 1 of the present invention. Detailed Implementation

[0063] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings:

[0064] Example 1

[0065] like Figure 1 As shown, this embodiment provides an environment-adaptive deep learning-based optical signal noise suppression method, including the following steps:

[0066] S1) Collect real-time environmental parameters and noisy optical signals in the optical transmission link; and preprocess the environmental parameters and noisy optical signals;

[0067] In this embodiment, the environmental parameters include temperature T, humidity H, vibration frequency f, and optical power attenuation A;

[0068] The noisy optical signal includes the optical signal amplitude. Phase ,frequency The noisy optical signal mentioned above Represented as:

[0069] ;

[0070] In the formula, express The amplitude of the optical signal at a given moment; This indicates a noise signal.

[0071] In this embodiment, the preprocessing includes:

[0072] Outlier removal is performed on the collected raw data using the following methods: The criteria include removing extreme anomalies caused by sensor malfunctions; for data gaps after outlier removal, linear interpolation is used to fill in the gaps, ensuring data continuity. The linear interpolation calculation formula is as follows:

[0073]

[0074] In the formula, For the first Fill in the missing data in each position; For the first Sampling time at each vacancy location; The first Valid data before and after each vacancy position; Data respectively Sampling time;

[0075] The completed environmental parameters and noisy optical signals are standardized to eliminate interference from differences in parameter magnitudes and signal amplitudes. Simultaneously, the noisy optical signals are segmented. Specifically:

[0076] The environmental parameters are standardized using the min-max method, mapping each parameter to the [0,1] interval to eliminate the influence of environmental parameters of different magnitudes; thus obtaining standardized environmental parameters. ;

[0077]

[0078] In the formula, , These are the maximum and minimum values ​​of the environmental parameters, respectively. These are the original environmental parameters;

[0079] Finally, the standardized environmental parameter set Represented as:

[0080] ;

[0081] ;

[0082] In the formula, The first The standardized real-time temperature, humidity, vibration frequency, and optical power attenuation; This represents the total number of real-time environmental parameters.

[0083] For noisy optical signals, the Z-score normalization method is used to normalize the signal amplitude, phase, and frequency features to a standard normal distribution with a mean of 0 and a variance of 1, reducing the interference of signal amplitude differences on subsequent model training.

[0084]

[0085] In the formula, The standardized optical signal containing noise; The original, noisy optical signal; The mean value of the noisy optical signal; This represents the standard deviation of the noisy optical signal.

[0086] The standardized noisy optical signal is segmented using a fixed-length segmentation method. Each segment length L is set to 64-256 sampling points, meaning each segment contains L consecutive sample values. After segmentation, several segments of noisy optical signal samples are obtained, namely:

[0087] ;

[0088] ;

[0089] In the formula, For the first A segment of noisy optical signal sample; Indicates the first The first segment of the noisy optical signal sample A noisy optical signal.

[0090] S2) Based on the K-means clustering algorithm, historical environmental parameter samples are clustered to classify different environmental types and clarify the differences in noise characteristics under each environment; specifically as follows:

[0091] S21) Collect historical environmental parameter samples and preprocess them as in step S1) to obtain a standardized historical environmental parameter sample set. ,Right now:

[0092] ;

[0093] ;

[0094] In the formula, The first The standardized temperature, humidity, vibration frequency, and optical power attenuation; This represents the total number of historical environmental parameter samples.

[0095] S22) Based on the distribution characteristics of historical environmental parameter samples, combined with the profile coefficient Verification method to determine the number of clusters The aforementioned contour coefficient Represented as:

[0096]

[0097] In the formula, For the first The average distance between a historical environmental parameter and all other historical environmental parameters in the same cluster; For the first The average distance between each historical environmental parameter and all samples within the nearest heterocluster; Indicates taking The maximum value in.

[0098] Profile coefficient The value range is [-1, 1], and the profile coefficient is... The closer the value is to 1, the better the sample clustering effect; calculate the average silhouette coefficient corresponding to different cluster numbers K values ​​(4~6), and select the average silhouette coefficient. The largest cluster number K value is taken as the final cluster number;

[0099] S23), from the standardized historical environmental parameter sample set K samples were randomly selected as the initial cluster centers. ,Right now:

[0100] ;

[0101] In the formula, For the first An environmental parameter vector for each cluster center.

[0102] S24) Calculate each historical environmental parameter sample with each cluster center Euclidean distance and samples of historical environmental parameters Assign to the nearest cluster; the Euclidean distance Represented as:

[0103]

[0104] In the formula, For the first The first historical environmental parameter sample One environmental parameter, At that time, the corresponding values ​​are temperature, humidity, vibration frequency, and power attenuation, respectively. Indicates the first The first cluster center One environmental parameter;

[0105] S25) Calculate the mean of all historical environmental parameter samples within each cluster, and use this mean as the new cluster center, i.e.:

[0106]

[0107] In the formula, For the first k The updated cluster centers of each cluster; For the first The number of historical environmental parameter samples within each cluster; For the first Each cluster contains a set of historical environmental parameter samples;

[0108] S26) Repeat steps S24) and S25) until the change in cluster centers is less than a set threshold; or until the maximum number of iterations is reached, stop the iteration and obtain the result. K Environment types.

[0109] S27) Based on the clustering results, analyze the distribution range of environmental parameters for each cluster, define the environmental type, and clarify the parameter threshold range for each environmental type;

[0110] In this embodiment, the environment types include:

[0111] Stable environment, temperature 20℃~30℃, humidity 40%RH~60%, vibration frequency ≤50Hz, optical power attenuation ≤-20dBm;

[0112] Temperature fluctuation environment: temperature -10℃~50℃, humidity 40%RH~60%, vibration frequency ≤50Hz, optical power attenuation ≤-20dBm;

[0113] Vibration interference environment, temperature 20℃~30℃, humidity 40%RH~60%, vibration frequency 50~500Hz, optical power attenuation ≤-20dBm;

[0114] Multi-factor mixed interference environment, temperature -10℃~50℃, humidity 20%RH~80%, vibration frequency 50~500Hz, optical power attenuation >-20dBm.

[0115] S3) Pre-train a corresponding deep learning sub-model for each environment type and initialize the environment adaptive decision-maker; specifically as follows:

[0116] In this embodiment, the deep learning sub-model includes a CNN sub-model, an LSTM sub-model, and a Transformer sub-model; wherein, the CNN sub-model is used to extract local noise features of the noisy light signal; the LSTM sub-model is used to extract temporal features of the noisy light signal; and the Transformer sub-model is used to extract global features of the noisy light signal, including the correlation features between noise and the environment.

[0117] In this embodiment, the CNN sub-model includes multiple convolutional layers, multiple max-pooling layers, and one fully connected layer. The convolutional layers use the ReLU activation function. First, shallow convolutional layers capture the noise fluctuation features between adjacent sampling points of the noisy optical signal. Then, max-pooling layers compress the dimension of the noise fluctuation features and retain the core noise features. Subsequently, deep convolutional layers expand the local feature capture range and extract the noise correlation features between short-distance sampling points. Max-pooling layers further enhance the noise feature discriminativeness. Finally, a fully connected layer flattens and fuses the features extracted by the multiple convolutional-pooling layers to output local noise features.

[0118] In this embodiment, the LSTM sub-model includes multiple LSTM hidden layers and one fully connected layer. The LSTM hidden layers use the tanh activation function, and the fully connected layer uses a linear activation function. The LSTM hidden layers capture the temporal dependence of noise between different sampling times of the optical signal through memory units, forget gates, input gates, and output gates, thereby achieving the extraction and separation of temporal noise features.

[0119] ;

[0120] ;

[0121] ;

[0122] ;

[0123] ;

[0124] ;

[0125] In the formula, This is the final hidden state; This is the output of the output gate; This is the state vector of the memory unit at the current moment; Candidate cell state; For input gate output; Output for the forget gate; The sample of the noisy optical signal at time t; , , , These are the weights of the forget gate, input gate, memory bypass unit, and output gate, respectively. , , , These are the bias terms for the forget gate, input gate, memory bypass unit, and output gate, respectively. It is the sigmoid activation function; This indicates element-wise multiplication;

[0126] Finally, the final hidden states of all LSTM hidden layers are fused through a fully connected layer to obtain the temporal noise features.

[0127] In this embodiment, the Transformer sub-model includes a two-layer Transformer encoder; the two-layer Transformer encoder extracts global correlation features and noise-environment coupling features of the optical signal; the first-layer Transformer encoder is as follows:

[0128] The Transformer encoder described above injects position information by introducing a sinusoidal position-coded, segmented, normalized, noisy optical signal sample into the input, i.e.:

[0129] ;

[0130] ;

[0131] In the formula, This is the location index at the sampling time. For feature index, For the Transformer sub-model dimension; For position encoding;

[0132] The position code is added element-by-element to the input sequence to obtain the input feature sequence infused with position information. ;Right now:

[0133] ;

[0134] Input feature sequence with injected location information The query matrix Q, key matrix K, and value matrix V are generated separately through a linear mapping layer; the self-attention outputs are independently computed by multiple parallel self-attention heads; then the self-attention outputs of the multiple self-attention heads are concatenated, and a dimensionality transformation is performed through a linear mapping weight matrix to obtain the final multi-head self-attention output, i.e.:

[0135] ;

[0136] This is a characteristic of multi-head attention; For the first Calculate attention score based on size; , These are the linear mapping weights and bias terms, respectively.

[0137] Multi-head attention features output by multi-head self-attention are obtained through residual connections. With the input feature sequence of this layer The components are added together, and then normalized using layer normalization to eliminate the influence of feature distribution differences on subsequent calculations.

[0138] ;

[0139] In the formula, The normalized multi-head attention features; Presentation layer normalization operation;

[0140] Multi-head attention features after layer normalization are processed using a feedforward fully connected network. Nonlinear transformation and deep fusion are performed to enhance the global feature differentiation between noise and effective signals; then, residual connections and layer normalization are applied again to obtain the final output of the single-layer Transformer encoder, namely:

[0141] ;

[0142] ;

[0143] In the formula, This is a primary global feature; The fusion features are the result of nonlinear transformation and deep fusion. , These are the weights of the two linear mapping layers in the feedforward fully connected network; , These are the biases of the two linear mapping layers of the feedforward fully connected network.

[0144] The primary global features output from the first-layer Transformer encoder are used as input to the second-layer Transformer encoder, and the operation of the first-layer Transformer encoder is repeated. The extraction and fusion of global features are further deepened through the two-layer Transformer encoder to capture more complex noise-environment coupling features. By performing global average pooling on the secondary global features output from the second-layer Transformer encoder, the temporal dimension is compressed to obtain a 128-dimensional global feature vector. Then, through the weight matrix of the fully connected layer, a 128-dimensional global noise feature vector is finally output.

[0145] In this embodiment, matching training sample sets are constructed for the above four environment types: no less than 6,000 sets of noisy light signal samples and corresponding real effective light signal samples are collected for each environment type. Each set of samples includes standardized 128×3-dimensional noisy light signal segment samples and corresponding standardized 128×3-dimensional real effective light signal segment samples. The samples cover the full parameter fluctuation range of the corresponding environment type to ensure the comprehensiveness and representativeness of the training samples and avoid model overfitting.

[0146] The Adam optimizer was used to pre-train the three sub-models separately for different environments. The optimization objective was to minimize the mean squared error (MSE) between the model output and the label. The core parameters for pre-training were set as follows: initial learning rate... Training rounds E = 50-80 rounds, batch size =32-64; During training, a learning rate decay strategy is adopted, with the learning rate decaying to 0.9 times the original value every 10 rounds. At the same time, an early stopping strategy is adopted, and training is stopped when the validation set loss does not decrease for 10 consecutive rounds. The optimal weight parameters of each sub-model under the corresponding environment type are saved to ensure that the pre-trained model has the best noise reduction performance in the corresponding environment.

[0147] In this embodiment, a BP neural network is used to construct an environment adaptive decision-maker, which is used to select the optimal combination of sub-models and fusion weights and dynamically fine-tune the sub-model parameters based on real-time environmental parameters; the environment adaptive decision-maker includes an input layer, 2-3 hidden layers, and an output layer;

[0148] The input layer is used to input standardized environmental parameters, and the hidden layer uses the ReLU activation function to perform nonlinear feature extraction on the input environmental parameter vector and explore the coupling relationship between various environmental parameters.

[0149] The output layer maps the high-dimensional correlation features of the hidden layer to the [0,1] interval through the sigmoid activation function, and outputs the K-dimensional environment type similarity.

[0150] The network weights are initialized using a random normal distribution, and the bias term is initialized to 0. Historical environmental parameter samples and the optimal sub-model combination and parameter adjustment coefficients under corresponding environments are collected as training samples. The environment adaptive decision-maker is pre-trained using the cross-entropy loss function, with 50 training rounds and a batch size of [missing information]. =64, trained until the environment adaptive decision-maker converges. The convergence criterion is that the accuracy of the environment type classification on the validation set is ≥95%, ensuring the accuracy of the decision-maker's environment recognition and parameter adjustment.

[0151] S4) Based on standardized real-time environmental parameters Standardized real-time environmental parameters are identified through an environment adaptive decision-maker. The similarity to the environment type is used to select the optimal combination of sub-models, and the key parameters of the sub-models are dynamically fine-tuned according to the trend of environmental change; such as... Figure 2 As shown, the details are as follows:

[0152] S41) Standardize real-time environmental parameters An environment adaptive decision-maker is input, which calculates standardized real-time environmental parameters using a cosine similarity formula. Clustering centers with each predefined environment type similarity ,Right now:

[0153]

[0154] In the formula, It is an L2 norm; Indicates the transpose operation;

[0155] The larger the value, the higher the degree of matching between the real-time environmental parameters and the environment type.

[0156] S42) Select the environment type with the highest similarity as the current environment type, i.e., the current environment type. Represented as: ;

[0157] If the maximum similarity is less than 0.6, the current environment is determined to be an unknown environment. All sub-models are used for fusion and noise reduction, and the environmental parameters and the corresponding noise reduction data are included in the subsequent incremental training samples.

[0158] S43) Based on the current environment type Based on the adaptation relationship between each environment type and the sub-model determined during the pre-training phase, the optimal combination of sub-models is selected; that is:

[0159] If a stable environment is required, choose the CNN sub-model;

[0160] For environments with temperature fluctuations, an LSTM sub-model is selected; for environments with vibration disturbances, a combination of a CNN sub-model and an LSTM sub-model is selected.

[0161] In multi-factor mixed interference environments, a combination of CNN sub-model, LSTM sub-model, and Transformer sub-model is selected;

[0162] S44) The environment adaptive decision-maker dynamically fine-tunes the key parameters of the selected sub-model based on the real-time changing trends of environmental parameters. The fine-tuning objective is to maximize the signal-to-noise ratio after noise reduction. The gradient descent method is used for parameter updates; that is:

[0163]

[0164] In the formula, These are the fine-tuned parameters; These are the parameters before fine-tuning; The learning rate; For the objective function with respect to parameters The partial derivatives of .

[0165] In this embodiment, the environmental adaptive decision-maker calculates the rate of change of real-time environmental parameters over the most recent M sampling periods. To determine the changing trend of environmental parameters, that is:

[0166]

[0167] In the formula, These are the environmental parameters at the current moment; These are the environmental parameters before a sampling period; The sampling period;

[0168] When | When environmental parameters are stable, it is determined that they are stable; when When environmental parameters rise, it is determined that they are increasing.

[0169] when When environmental parameters decrease, it is determined that the environmental parameters have decreased.

[0170] S5) The preprocessed noisy optical signal is input into the selected optimal sub-model combination. Each sub-model performs noise suppression and outputs preliminary denoising results. An attention mechanism is used to dynamically fuse the outputs of each sub-model to obtain the final denoised optical signal. Then, the original features of the optical signal are restored through denormalization. Figure 3 As shown, the specific steps include the following:

[0171] S51) Segment the noisy optical signal samples In the optimal combination of input sub-models, each sub-model in the optimal combination is used to process segmented noisy optical signal samples. Denoising is performed, and each sub-model outputs its preliminary denoising result. If the optimal sub-model combination consists of multiple sub-models, then each sub-model outputs a preliminary denoising sample, resulting in a set of preliminary denoising results. , This is the initial noise reduction result output by the k-th sub-model;

[0172] S52), the preliminary denoising results of each sub-model are fused using a multi-head self-attention mechanism, as detailed below:

[0173] Calculate the initial denoising results of each sub-model in the optimal sub-model combination. signal-to-noise ratio and mean square error As the basis for weight allocation;

[0174] Based on signal-to-noise ratio and mean square error Calculate the fusion weights of each sub-model ,Right now:

[0175]

[0176]

[0177] In the formula, These are the weighting coefficients. The fusion weights before normalization; This represents the number of sub-models in the optimal sub-model combination.

[0178] The preliminary denoising results of each sub-model are fused using a weighted summation method to obtain the standardized final denoised optical signal sample, namely:

[0179]

[0180] In the formula, For the final noise reduction sample, the first The sampling point Standardized values ​​of each feature; For the first The first of the preliminary noise reduction results of the sub-model The sampling point The standardized value of each feature.

[0181] S53) The final denoised optical signal sample after standardization After denormalization, the amplitude, phase, and frequency characteristics of the original optical signal are restored, i.e.:

[0182]

[0183] In the formula, The final denoised optical signal after standardization The sampling point One feature; The mean value of the noisy optical signal; This represents the standard deviation of the noisy optical signal.

[0184] S6) Quantitatively evaluate the denoising effect of the final denoised optical signal to determine whether it meets the requirements of practical applications; if not, update the parameters of the sub-model and the environment adaptive decision-maker through incremental training; specifically including the following steps:

[0185] S61) Calculate the signal-to-noise ratio (SNR), mean square error (MSE), and peak signal-to-noise ratio (PSNR) of the final denoised optical signal; if the signal-to-noise ratio (SNR), mean square error (MSE), and peak signal-to-noise ratio (PSNR) of the final denoised optical signal all meet the threshold requirements, then output the final denoised optical signal.

[0186] If any evaluation indicator fails to meet the threshold requirement, the noise reduction effect is deemed unqualified; proceed to step S62).

[0187] S62) The current standardized environmental parameters, noisy light signal samples and corresponding final denoised light signal samples, and real effective light signal samples are used as new incremental training samples and added to the training sample set of the corresponding environment type.

[0188] S63) Employ a mini-batch training method to incrementally train the selected sub-model; during training, only some parameters of the sub-model are updated, while the low-level feature extraction parameters trained in the pre-training stage are fixed; the mean squared error loss function is used for training, and the Adam optimizer is used;

[0189] S64) Use the current environmental parameters and the corresponding optimal sub-model combination and parameter adjustment coefficients as training samples to incrementally train the environmental adaptive decision-maker, update the weights and biases of the environmental adaptive decision-maker, and adjust the similarity calculation coefficients.

[0190] (S65) After incremental training is completed, save the updated sub-model parameters and decision-maker parameters for subsequent noise suppression processing; at the same time, add the current environmental parameter samples to the historical environmental parameter sample set. If the addition of samples causes a change in the distribution of environmental types, re-execute the K-means clustering in step S2) to update the environmental type classification results.

[0191] Example 2

[0192] This embodiment provides an environment-adaptive deep learning-based optical signal noise suppression system, comprising:

[0193] The data acquisition module is used to acquire real-time environmental parameters and noisy optical signals in the optical transmission link; and to preprocess the environmental parameters and noisy optical signals.

[0194] The environment type classification module uses the K-means clustering algorithm to cluster historical environment parameter samples and classify them into different environment types.

[0195] The training module is used to pre-train the corresponding deep learning sub-model for each environment type and initialize the environment adaptive decision-maker;

[0196] The model selection module, based on standardized real-time environmental parameters, identifies the similarity between standardized real-time environmental parameters and environmental types through an environmental adaptive decision-maker, and selects the optimal combination of sub-models based on the environmental type similarity.

[0197] The noise reduction module uses the selected optimal sub-model combination to suppress noise in the preprocessed noisy optical signal and outputs the preliminary noise reduction result. It also uses an attention mechanism to dynamically fuse the outputs of each sub-model to obtain the final noise-reduced optical signal. Finally, it restores the original features of the optical signal through denormalization.

[0198] The evaluation module quantitatively evaluates the denoising effect of the final denoised optical signal to determine whether it meets the actual application requirements; if not, it updates the parameters of the sub-model and the environment adaptive decision-maker through incremental training.

[0199] Example 3

[0200] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the optical signal noise suppression method described in Embodiment 1.

[0201] In this embodiment, the memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. A processor, coupled to the memory, is used to execute computer programs stored in the memory.

[0202] The computer program includes computer program code, which may be in the form of source code, object code, executable file, or some intermediate form.

[0203] The embodiments and descriptions above are merely illustrative of the principles and preferred embodiments of the present invention. Various changes and modifications may be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed.

Claims

1. An environment-adaptive deep learning fusion optical signal noise suppression method, characterized in that, Includes the following steps: S1) Collect real-time environmental parameters and noisy optical signals in the optical transmission link; and preprocess the environmental parameters and noisy optical signals; S2) Based on the K-means clustering algorithm, historical environmental parameter samples are clustered to classify different environmental types; S3) Pre-train the corresponding deep learning sub-model for each environment type and initialize the environment adaptive decision-maker; where, The deep learning sub-models include a CNN sub-model, an LSTM sub-model, and a Transformer sub-model; wherein, the CNN sub-model is used to extract local noise features of noisy optical signals; the LSTM sub-model is used to extract temporal features of noisy optical signals; and the Transformer sub-model is used to extract global correlation features of noisy optical signals, including noise and environmental correlation. S4) Based on standardized real-time environmental parameters, an environmental adaptive decision-maker identifies the similarity between standardized real-time environmental parameters and environmental types, and selects the optimal combination of sub-models based on the similarity of environmental types. S5) Input the preprocessed noisy optical signal into the selected optimal sub-model combination. Each sub-model in the optimal sub-model combination completes noise suppression and outputs preliminary noise reduction results. The outputs of each sub-model are dynamically fused using an attention mechanism to obtain the final noise-reduced optical signal. Then, the original features of the optical signal are restored through inverse normalization processing. S6) Quantitatively evaluate the denoising effect of the final denoised optical signal to determine whether it meets the actual application requirements; if not, update the parameters of the sub-model and the environment adaptive decision-maker through incremental training. 2.The environment-adaptive deep learning fusion optical signal noise suppression method according to claim 1, characterized in that: In step S2), the historical environmental parameter samples are clustered based on the K-means clustering algorithm to classify different environmental types. This specifically includes the following steps: S21) Collect historical environmental parameter samples and perform preprocessing on the historical environmental parameter samples as in step S1) to obtain a standardized historical environmental parameter sample set. S22) Based on the distribution characteristics of historical environmental parameter samples, calculate the average profile coefficient corresponding to different cluster numbers K values, and select the cluster number K value with an average profile coefficient greater than or equal to the profile threshold as the final cluster number. S23) Randomly select K samples from the standardized historical environmental parameter sample set as the initial cluster centers; S24) Calculate the Euclidean distance between each historical environmental parameter sample and each cluster center, and assign the historical environmental parameter sample to the nearest cluster; S25) Calculate the mean of all historical environmental parameter samples within each cluster and use it as the new cluster center; S26) Repeat steps S24) and S25) until the change in cluster centers is less than the set threshold; or the number of iterations reaches the maximum number of iterations, then stop the iteration and obtain K environment types; S27) Based on the clustering results, analyze the distribution range of environmental parameters for each cluster, define the environmental type, and clarify the parameter threshold range for each environmental type.

3. The method of claim 2, wherein the method is characterized by: In step S2), the environment type includes: Stable environment, temperature 20℃~30℃, humidity 40%RH~60%, vibration frequency ≤50Hz, optical power attenuation ≤-20dBm; Temperature fluctuation environment: temperature -10℃~50℃, humidity 40%~60% RH, vibration frequency ≤50Hz, optical power attenuation ≤-20dBm; Vibration interference environment, temperature 20℃~30℃, humidity 40%RH~60%, vibration frequency 50~500Hz, optical power attenuation ≤-20dBm; Multi-factor mixed interference environment, temperature -10℃~50℃, humidity 20%RH~80%, vibration frequency 50~500Hz, optical power attenuation >-20dBm.

4. The environmentally adaptive deep learning fusion optical signal noise suppression method according to claim 1, characterized in that: In step S3), the Transformer sub-model includes a two-layer Transformer encoder; the two-layer Transformer encoder is used to extract global correlation features of the noisy optical signal, including the correlation between noise and the environment.

5. The method of claim 1, wherein the method is characterized by: In step S3), a BP neural network is used to construct an environment adaptive decision-maker, which is used to select the optimal combination of sub-models and fusion weights and dynamically fine-tune the sub-model parameters based on real-time environment parameters; the environment adaptive decision-maker includes an input layer, 2-3 hidden layers, and an output layer. The input layer is used to input standardized environmental parameters, and the hidden layer uses the ReLU activation function to perform nonlinear feature extraction on the input environmental parameter vector through the hidden layer, thereby mining the coupling correlation between various environmental parameters and obtaining high-dimensional correlation features. The output layer maps the high-dimensional correlation features of the hidden layer to the [0,1] interval through the sigmoid activation function, and outputs the K-dimensional environment type similarity.

6. The method of claim 5, wherein the method is characterized by: Step S4 specifically includes the following steps: S41) The standardized real-time environmental parameters are input into the environment adaptive decision-maker, which calculates the similarity between the standardized real-time environmental parameters and the cluster centers of each predefined environment type using cosine similarity. S42) Select the environment type with the highest similarity as the current environment type; S43) Based on the current environment type and the adaptation relationship between each environment type and the sub-model determined in the pre-training stage, select the optimal combination of sub-models. S44) The environment adaptive decision-maker dynamically fine-tunes the key parameters of the selected sub-model based on the changing trend of real-time environmental parameters. The fine-tuning objective is to maximize the signal-to-noise ratio (SNR) of the denoised signal, and the gradient descent method is used for parameter updating.

7. The method of claim 6, wherein the method is characterized by: Step S5 specifically includes the following steps: S51) Input the segmented noisy optical signal samples into the selected optimal sub-model combination. Each sub-model in the optimal sub-model combination performs noise reduction processing on the segmented noisy optical signal samples and outputs its own preliminary noise reduction results. If the optimal combination of sub-models consists of multiple sub-models, then each sub-model outputs a preliminary denoising sample, resulting in a preliminary denoising result set. S52) The preliminary denoising results of each sub-model are fused using a multi-head self-attention mechanism to obtain the standardized final denoised optical signal sample; S53) Perform denormalization on the final denoised optical signal sample after standardization to restore the amplitude, phase, and frequency characteristics of the original optical signal.

8. The method of claim 7, wherein the method is characterized by: Step S6) specifically includes the following steps: S61) Calculate the three evaluation indicators of the final denoised optical signal: signal-to-noise ratio (SNR), mean square error (MSE), and peak signal-to-noise ratio (PSNR). If all three evaluation indicators meet the threshold requirements, output the final denoised optical signal. If any one of the evaluation indicators fails to meet the threshold requirements, the denoising effect is deemed unqualified, and proceed to step S62). S62) The current standardized environmental parameters, noisy light signal samples and corresponding final denoised light signal samples, and real effective light signal samples are used as new incremental training samples and added to the training sample set of the corresponding environment type. S63) Use a mini-batch training method to incrementally train the selected sub-model; S64) The current environmental parameters and the corresponding optimal sub-model combination and parameter adjustment coefficients are used as training samples to incrementally train the environmental adaptive decision-maker.

9. An environment-adaptive deep learning fusion optical signal noise suppression system, characterized in that, include: The data acquisition module is used to collect real-time environmental parameters and noisy optical signals in the optical transmission link; It also preprocesses environmental parameters and noisy optical signals; The environment type classification module uses the K-means clustering algorithm to cluster historical environment parameter samples and classify them into different environment types. The training module is used to pre-train the corresponding deep learning sub-model for each environment type and initialize the environment adaptive decision-maker; The deep learning sub-models include a CNN sub-model, an LSTM sub-model, and a Transformer sub-model. The CNN sub-model is used to extract local noise features of noisy light signals. The LSTM sub-model is used to extract temporal features of noisy light signals. The Transformer sub-model is used to extract global correlation features of noisy light signals, including noise and environmental correlation. The model selection module, based on standardized real-time environmental parameters, identifies the similarity between standardized real-time environmental parameters and environmental types through an environmental adaptive decision-maker, and selects the optimal combination of sub-models based on the environmental type similarity. The noise reduction module uses the selected optimal sub-model combination to suppress noise in the preprocessed noisy optical signal and outputs the preliminary noise reduction result. It also uses an attention mechanism to dynamically fuse the outputs of each sub-model to obtain the final noise-reduced optical signal. Finally, it restores the original features of the optical signal through denormalization. The evaluation module quantitatively evaluates the denoising effect of the final denoised optical signal to determine whether it meets the actual application requirements; if not, it updates the parameters of the sub-model and the environment adaptive decision-maker through incremental training.

Images