Deep learning-based electric signal noise reduction method, system, terminal and storage medium

Abstract

The application relates to a deep learning-based electric signal noise reduction method and system, a terminal and a storage medium, relates to the field of track signal processing, and comprises the following steps: acquiring a training circuit signal set of a preset track circuit; the training circuit signal set comprises a noisy signal sample and a corresponding clean signal label; a training task scene is acquired based on the training circuit signal set; a deep neural network model is constructed, and the deep neural network model is trained according to the training circuit signal set and the corresponding training task scene; real-time circuit signals of the track circuit and a corresponding real-time task scene are acquired; and the trained deep neural network model is controlled to output noise reduction circuit signals according to the real-time circuit signals and the real-time task scene. The application has the effect of improving the noise reduction effect of the track circuit signal.

Description

Technical Field

[0001] This application relates to the technical field of track signal processing, and in particular to a method, system, terminal and storage medium for electrical signal noise reduction based on deep learning. Background Technology

[0002] Track circuit signals are an important component of railway signaling systems, used to detect train positions and ensure railway operation safety. By identifying the signal characteristics of track circuit signals and using the correspondence between different signal characteristics and different track states, the specific state of the track at that time can be determined.

[0003] In related technologies, the accuracy of track circuit signals is particularly important due to safety concerns. Currently, noise reduction of acquired track circuit signals is mainly achieved through hardware and software approaches. For example, at the hardware level, the track line is shielded and grounded, filters are added to the transmission path to filter out unwanted high-frequency noise, and components with strong anti-interference capabilities are used to improve system stability. At the software level, algorithms such as Fourier transform and wavelet transform are used to perform frequency domain analysis on the track circuit signals, thereby identifying useful signal components and removing noise.

[0004] Regarding the aforementioned technologies, algorithms such as Fourier transform and wavelet transform are used to denoise track circuit signals. These algorithms perform well for specific types of noise, such as simple and regular Gaussian white noise. However, for unknown noise, such as nonlinear noise like multi-source interference in track signals, Fourier transform and wavelet transform algorithms may fail, resulting in poor noise reduction performance for track circuit signals. There is still room for improvement. Summary of the Invention

[0005] To improve the noise reduction effect of track circuit signals, this application provides a method, system, terminal and storage medium for electrical signal noise reduction based on deep learning.

[0006] Firstly, this application provides a deep learning-based method for denoising electrical signals, employing the following technical solution: Deep learning-based methods for denoising electrical signals include: Obtain a training circuit signal set for a preset track circuit; the training circuit signal set includes noisy signal samples and corresponding clean signal labels; The training task scenario is obtained based on the corresponding signal set of the training circuit. Construct a deep neural network model and train the deep neural network model based on the training circuit signal set and the corresponding training task scenario; Acquire real-time circuit signals of the track circuit and the corresponding real-time task scenario; The noise reduction circuit signal is output by the trained deep neural network model based on the real-time circuit signal and the real-time task scenario.

[0007] By adopting the above technical solution, a deep neural network model is trained based on the training circuit signal set of the track circuit and the corresponding training task scenario. The deep neural network model can handle both linear and simple noise, as well as nonlinear and non-stationary noise. Using the trained deep neural network model to perform noise reduction based on real-time circuit signals and real-time task scenarios can not only improve the noise reduction efficiency of track circuit signals, but also improve the noise reduction effect of track circuit signals.

[0008] Optionally, the steps for training a deep neural network model based on the training circuit signal set and the corresponding training task scenario include: The signal set of the training circuit is analyzed to determine the time-frequency distribution characteristics of the noise; The circuit noise type is determined based on the time-frequency distribution characteristics and the preset relationship between the first time-frequency noise type. The network noise reduction branch is determined based on the circuit noise type and the preset first noise type branch relationship; The scene noise reduction branch is determined based on the training task scenario and the preset first scene branch relationship; The denoising branch of the deep neural network model is switched according to the scene denoising branch and the network denoising branch, and the denoising branch of the deep neural network model is trained according to the signal set of the training circuit.

[0009] By adopting the above technical solution, the scene denoising branch and the network denoising branch are determined according to the training task scenario and the type of circuit noise, respectively. This solves the problem of needing to retrain when the noise distribution drifts in practical applications, and improves the model's ability to quickly adapt to unknown noise. The network denoising branch controls the adaptive switching of the model, which solves the problem of single model performance degradation in mixed noise scenarios, thereby improving the denoising effect on track circuit signals.

[0010] Optionally, the steps for training the denoising branch of the deep neural network model based on the signal set of the training circuit include: Define a joint loss function, which is composed of a linearly weighted data fitting term and a physical constraint term; the data fitting term is characterized by the mean square error between the network output and the clean label, and the physical constraint term is characterized by the sum of the loop voltage residual term and the instantaneous power residual term; The parameters of the denoising branch of the trained deep neural network model are optimized based on the preset gradient backpropagation algorithm and the training circuit signal set until the joint loss function is minimized.

[0011] By adopting the above technical solution, a joint loss function is defined using data fitting terms and physical constraint terms, thereby avoiding physically unreliable results caused by purely data-driven approaches, such as signal amplitude abrupt changes violating circuit principles, and thus improving the accuracy of deep neural network models.

[0012] Optionally, the step of controlling the trained deep neural network model to output the noise reduction circuit signal based on the real-time circuit signal and the real-time task scenario includes: Analyze real-time circuit signals to determine the time-frequency characteristics of real-time noise; The real-time noise type is determined based on the real-time noise time-frequency characteristics and the preset second time-frequency noise type relationship; The real-time noise reduction branch of the network is determined based on the real-time noise type and the preset second noise type branch relationship; The scene real-time noise reduction branch is determined based on the real-time task scenario and the preset second scene branch relationship; The denoising branch of the deep neural network model is switched between the scene real-time denoising branch and the network real-time denoising branch, and the real-time circuit signal is input into the denoising branch of the deep neural network model to output the denoising circuit signal.

[0013] By adopting the above technical solution, the noise reduction branch of the deep neural network model is switched according to the real-time noise reduction branch of the scene and the real-time noise reduction branch of the network, and the real-time circuit signal is input into the noise reduction branch of the deep neural network model for noise reduction, so as to remove noise in the circuit signal with the best model, thereby improving the noise reduction effect of the track circuit signal.

[0014] Optionally, the step of inputting the real-time circuit signal into the denoising branch of the deep neural network model to output the denoising circuit signal includes: Analyze real-time circuit signals to determine noise energy indicators; Determine whether the noise energy index meets the preset energy index threshold requirements; If the conditions are met, the denoising branch of the deep neural network model is controlled to denoise the real-time circuit signal in order to generate a denoised circuit signal. If the signal does not meet the requirements, the real-time circuit signal will be smoothed according to a preset lightweight algorithm to generate a noise-reducing circuit signal.

[0015] By adopting the above technical solution, when the noise energy index meets the requirements of the energy index threshold, the noise reduction branch of the deep neural network model is controlled to reduce the noise of the real-time circuit signal. When it does not meet the requirements, the real-time circuit signal is smoothed according to the lightweight algorithm. This not only ensures the noise reduction effect of the track circuit signal, but also reduces the power consumption during noise reduction.

[0016] Optionally, the step of inputting the real-time circuit signal into the denoising branch of the deep neural network model to output the denoising circuit signal further includes: Obtain the trigger signal for the jump in noise energy index; The noise reduction branch of the deep neural network model is controlled by the jump trigger signal to reduce the noise of the real-time circuit signal and generate the first noise-reduced signal. The lightweight algorithm controls the transition trigger signal to smooth the real-time circuit signal in order to generate a second noise-reduced signal. The noise energy index is analyzed to determine the fusion weights; The first and second noise-reduced signals are weighted and fused according to the fusion weights to generate the noise reduction circuit signal.

[0017] By adopting the above technical solution, when a jump trigger signal of the noise energy index is detected, the denoising branch of the deep neural network model is controlled to generate a first denoising signal, and the lightweight algorithm is controlled to generate a second denoising signal. The first denoising signal and the second denoising signal are weighted and fused according to the fusion weight to obtain the denoising circuit signal, ensuring the smoothness of the switching between the deep neural network model and the lightweight algorithm and avoiding amplitude abrupt changes caused by hard switching.

[0018] Optionally, the steps of analyzing noise energy metrics to determine fusion weights include: The noise energy index and the preset activation function are analyzed to determine the first noise weight; The preset total weight and the first noise weight are analyzed to determine the second noise weight; The first noise weight and the second noise weight are correlated to generate the fused weight.

[0019] By adopting the above technical solution, the first noise weight is determined based on the noise energy index and activation function, thereby adjusting the weights of the deep neural network model and the lightweight algorithm according to the different noise energy, realizing a gradual switching and avoiding sudden amplitude changes caused by hard switching.

[0020] Secondly, this application provides a deep learning-based electrical signal noise reduction system, which adopts the following technical solution: A deep learning-based electrical signal noise reduction system includes: The acquisition module is used to acquire the training circuit signal set, training task scenario, real-time circuit signal, and real-time task scenario. A memory for storing programs for deep learning-based electrical signal noise reduction methods as described in any of the preceding claims; The processor and the program in the memory can be loaded and executed by the processor to implement the deep learning-based electrical signal noise reduction method as described in any of the above.

[0021] By adopting the above technical solution, the control processor loads and executes the program of the deep learning-based electrical signal denoising method stored in the memory, so that the acquisition module acquires a series of data related to the deep learning-based electrical signal denoising. Then, the deep neural network model is trained according to the training circuit signal set of the track circuit and the corresponding training task scenario. The deep neural network model can handle both linear and simple noise and nonlinear and non-stationary noise. Using the trained deep neural network model to denoise according to the real-time circuit signal and the real-time task scenario can not only improve the denoising efficiency of the track circuit signal, but also improve the denoising effect of the track circuit signal.

[0022] Thirdly, this application provides a smart terminal, which adopts the following technical solution: A smart terminal includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described in any of the preceding claims, a deep learning-based electrical signal noise reduction method.

[0023] By adopting the above technical solution, the processor loads and executes the program of the deep learning-based electrical signal denoising method stored in the memory through the operation of the smart terminal. The deep neural network model is trained according to the training circuit signal set of the track circuit and the corresponding training task scenario. The deep neural network model can handle both linear and simple noise as well as nonlinear and non-stationary noise. Using the trained deep neural network model to denoise according to the real-time circuit signal and the real-time task scenario can not only improve the denoising efficiency of the track circuit signal, but also improve the denoising effect of the track circuit signal.

[0024] Fourthly, this application provides a computer storage medium capable of storing corresponding programs, which facilitates the implementation of noise reduction effects to improve track circuit signals, and adopts the following technical solution: A computer-readable storage medium storing a computer program capable of being loaded by a processor and executing any of the deep learning-based electrical signal noise reduction methods described above.

[0025] By adopting the above technical solution, a computer program for electrical signal denoising based on deep learning is stored in a computer-readable storage medium. The processor loads and executes the computer program in the storage medium, thereby training a deep neural network model based on the training circuit signal set of the track circuit and the corresponding training task scenario. The deep neural network model can handle both linear and simple noise as well as nonlinear and non-stationary noise. Using the trained deep neural network model to denoise based on real-time circuit signals and real-time task scenarios can not only improve the denoising efficiency of track circuit signals, but also improve the denoising effect of track circuit signals.

[0026] In summary, this application includes at least one of the following beneficial technical effects: 1. By training a deep neural network model based on the training circuit signal set of the track circuit and the corresponding training task scenario, the deep neural network model can handle both linear and simple noise as well as nonlinear and non-stationary noise. Using the trained deep neural network model to perform noise reduction based on real-time circuit signals and real-time task scenarios can not only improve the noise reduction efficiency of track circuit signals, but also improve the noise reduction effect of track circuit signals. 2. By switching the denoising branch of the deep neural network model according to the real-time denoising branch of the scene and the real-time denoising branch of the network, and inputting the real-time circuit signal into the denoising branch of the deep neural network model for denoising, the noise in the circuit signal is removed with the best model, thereby improving the denoising effect of the track circuit signal. 3. By defining a joint loss function using data fitting terms and physical constraint terms, we can avoid physically unreliable results caused by purely data-driven approaches, such as signal amplitude abrupt changes violating circuit principles, thereby improving the accuracy of deep neural network models. Attached Figure Description

[0027] Figure 1 This is a flowchart of the electrical signal noise reduction method based on deep learning in the embodiments of this application.

[0028] Figure 2 This is a flowchart of the steps for training a deep neural network model based on the training circuit signal set and the corresponding training task scenario in an embodiment of this application.

[0029] Figure 3 This is a flowchart of the steps in this application embodiment for training the noise reduction branch of the deep neural network model according to the signal set of the training circuit.

[0030] Figure 4 This is a flowchart illustrating the steps in this application embodiment of controlling the trained deep neural network model to output noise reduction circuit signals based on real-time circuit signals and real-time task scenarios.

[0031] Figure 5 This is the flowchart of the step in this application embodiment where the real-time circuit signal is input into the denoising branch of the deep neural network model to output the denoising circuit signal. Figure 1 .

[0032] Figure 6 This is the flowchart of the step in this application embodiment where the real-time circuit signal is input into the denoising branch of the deep neural network model to output the denoising circuit signal. Figure 2 .

[0033] Figure 7This is a flowchart of the steps in this application embodiment to analyze noise energy indicators to determine fusion weights. Detailed Implementation

[0034] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figures 1 to 7 The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.

[0035] This application discloses a method for denoising electrical signals based on deep learning. An operator collects a training circuit signal set of the track circuit and the corresponding training task scenario. After constructing a deep neural network model, the model is trained based on the training circuit signal set and the training task scenario. After training, real-time circuit signals of the track circuit and the corresponding real-time task scenario are collected. The trained deep neural network model is then controlled to denoise the real-time circuit signals based on the real-time circuit signals and the real-time task scenario, and then outputs a denoised circuit signal, thereby improving the denoising effect of the track circuit signals.

[0036] Reference Figure 1 This application discloses a method for denoising electrical signals based on deep learning, including the following steps: Step S100: Obtain the training circuit signal set of the preset track circuit; the training circuit signal set includes noisy signal samples and corresponding clean signal labels.

[0037] The track circuit refers to the circuit in the track used to monitor the track status, including the power supply end, track line and power receiving end. The track status is determined by analyzing the circuit signals received at the power receiving end.

[0038] The training circuit signal set refers to the data samples used to train the deep neural network model. It includes noisy signal samples, which are obtained by the sensor sampling the track circuit signals at the power receiving end of the track circuit; it also includes clean signal labels corresponding to the noisy signal samples, i.e. noise-free circuit signals, which can be generated by shielding the environment or through simulation.

[0039] Step S101: Obtain the training task scenario based on the signal set of the training circuit.

[0040] The training task scenario refers to the actual scenario when noisy signals occur, including parameters such as temperature, humidity, vibration, and electromagnetic environment. These parameters are detected by the power receiving end of the synchronous track circuit of temperature, humidity, vibration, and electromagnetic sensors installed along the track and the timestamps are aligned to eliminate sensor errors.

[0041] Step S102: Construct a deep neural network model and train the deep neural network model based on the training circuit signal set and the corresponding training task scenario.

[0042] The process involves operators building a deep neural network model on the processing terminal and training the model based on the training circuit signal set and the corresponding training task scenario. Specific methods are detailed below. Figure 2 This process provides a foundation for subsequent noise reduction of track circuit signals.

[0043] A deep neural network model is a model used to denoise track circuit signals. It includes a feature extraction module and a signal reconstruction module. The feature extraction module extracts the time and frequency domain features of the noisy signal through multi-layer convolution operations. The signal reconstruction module generates the denoised signal based on the features. The deep neural network model is an encoder-decoder structure. The encoder uses a one-dimensional dilated convolutional layer to gradually expand the receptive field to capture multi-scale noise features. The decoder uses a transposed convolutional layer to gradually restore the signal resolution.

[0044] Step S103: Obtain the real-time circuit signals of the track circuit and the corresponding real-time task scenario.

[0045] Among them, real-time circuit signals refer to the signals of the track circuit collected in real time by the power receiving end, which are acquired by sensors at the power receiving end and sent to the processing terminal. Real-time task scenarios refer to the scenarios corresponding to the real-time circuit signals, which are detected by temperature, humidity, vibration, and electromagnetic sensors installed along the track and sent to the processing terminal.

[0046] Step S104: Control the trained deep neural network model to output the noise reduction circuit signal based on the real-time circuit signal and the real-time task scenario.

[0047] Specifically, after receiving the real-time circuit signal and the real-time task scenario, the processing terminal controls a trained deep neural network model to denoise the real-time circuit signal based on the real-time circuit signal and the real-time task scenario, thereby obtaining the denoised circuit signal. The specific method is described in [reference needed]. Figure 4 The steps.

[0048] Noise-reduced circuit signals refer to real-time circuit signals that are free of noise, obtained by noise reduction of real-time circuit signals using a deep neural network model.

[0049] Reference Figure 2 The steps for training a deep neural network model based on the training circuit signal set and the corresponding training task scenario include: Step S200: Analyze the signal set of the training circuit to determine the time-frequency distribution characteristics of noise.

[0050] Among them, the noise time-frequency distribution characteristics refer to the characteristics of the noisy signal in the training circuit signal set in the time-frequency plane. The processing terminal selects the wavelet basis function to perform wavelet transform on the noisy signal in the training circuit signal set to obtain the energy distribution in the time-frequency plane. Then, the wavelet coefficients are analyzed, and the time-frequency distribution characteristics of the noisy signal are extracted from the wavelet coefficients by setting a threshold.

[0051] Step S201: Determine the circuit noise type based on the time-frequency distribution characteristics and the preset first time-frequency noise type relationship.

[0052] The first time-frequency noise type refers to the correspondence between time-frequency characteristics and noise types. For example, the time-frequency distribution characteristics of white noise are that the energy is evenly distributed across all frequency ranges, the time-frequency distribution characteristics of Gaussian noise are that the statistical characteristics conform to a normal distribution and the energy distribution is relatively uniform, and the time-frequency distribution characteristics of impulse noise are that the energy is concentrated at a specific time point and the frequency range is relatively wide. The operator forms a mapping table by matching the video distribution characteristics with the noise types one by one.

[0053] Circuit noise type refers to the noise type of the noisy signal, which is obtained by the processing terminal by looking up the mapping table corresponding to the first time-frequency noise type relationship based on the time-frequency distribution characteristics.

[0054] Step S202: Determine the network noise reduction branch based on the circuit noise type and the preset first noise type branch relationship.

[0055] Among them, the first noise type branch relationship refers to the correspondence between noise type and noise reduction branch. The operator constructs the noise reduction branch in the deep neural network model and forms a mapping table by matching the noise reduction branch with different noises one by one.

[0056] The network noise reduction branch refers to the specific branch in the deep neural network model that performs noise reduction on noisy signals. It is obtained by the processing terminal by looking up the corresponding mapping table of the first noise type branch relationship based on the circuit noise type.

[0057] Step S203: Determine the scene noise reduction branch based on the training task scenario and the preset first scene branch relationship.

[0058] The first scenario branch relationship refers to the correspondence between the task scenario and the scenario denoising branch. The operator constructs the denoising branch in the deep neural network model and forms a mapping table by mapping the denoising branch to different task scenarios.

[0059] The scene denoising branch refers to the branch in the deep neural network model that performs denoising based on the scene. The scene denoising branch includes multiple network denoising branches, which are obtained by the processing terminal by looking up the mapping table corresponding to the first scene branch relationship based on the training task scene.

[0060] Step S204: Switch the denoising branch of the deep neural network model according to the scene denoising branch and the network denoising branch, and train the deep neural network model according to the signal set of the training circuit.

[0061] After determining the scene denoising branch and the network denoising branch, the deep neural network model automatically switches to the network denoising branch within the scene denoising branch, and trains the deep neural network model according to the denoising branch corresponding to the training circuit signal set. For specific methods, refer to... Figure 3 The steps.

[0062] Reference Figure 3 The steps for determining the denoising branch of the deep neural network model based on the training circuit signal set include: Step S300: Define the joint loss function, which is composed of a linear weighted sum of the data fitting term and the physical constraint term; the data fitting term is represented by the mean square error between the network output and the clean label, and the physical constraint term is represented by the sum of the loop voltage residual term and the instantaneous power residual term.

[0063] The joint loss function refers to the loss function of the deep neural network model, which is composed of a weighted average of data fitting terms and physical constraint terms. The data fitting term refers to the mean square error between the denoised signal and the clean signal. The physical constraint terms are constructed based on the physical laws of the circuit system, including a loop voltage residual term constructed according to Kirchhoff's laws and an instantaneous power residual term constructed according to the energy conservation equation. The weight coefficients of the data fitting term and the physical constraint term adopt a dynamic adjustment strategy, using adaptive weighting based on Bayesian uncertainty estimation, so that the weight coefficients are inversely proportional to the variance of each loss term. By embedding the physical equations into the loss function, it is ensured that the denoised signal output by the network conforms to physical laws.

[0064] Step S301: Optimize the parameters of the denoising branch of the trained deep neural network model according to the preset gradient backpropagation algorithm and the training circuit signal set until the joint loss function is minimized.

[0065] Specifically, the parameters of the denoising branch of the deep neural network model are optimized by using the gradient backpropagation algorithm and the labels of noisy and clean signals in the training circuit signal set, so as to minimize the joint loss function. The gradient of the physical constraint term is calculated by automatic differentiation and fused with the gradient of the data fitting term in a weighted manner. Finally, the parameters of the deep neural network model are optimized to the optimal state, and the training of the deep neural network model is completed.

[0066] The gradient directional propagation algorithm is a core algorithm for training deep neural network models. It calculates the gradient of the loss function with respect to the network parameters and uses gradient descent to optimize these parameters, thereby minimizing the loss function.

[0067] Reference Figure 4 The steps for controlling the trained deep neural network model to output noise reduction circuit signals based on real-time circuit signals and real-time task scenarios include: Step S400: Analyze the real-time circuit signal to determine the real-time noise frequency characteristics.

[0068] Among them, the real-time noise time-frequency characteristics refer to the characteristics of noise in the real-time circuit signal in the time-frequency plane. The processing terminal selects wavelet basis functions to perform wavelet transform on the real-time circuit signal to obtain the energy distribution in the time-frequency plane. Then, the wavelet coefficients are analyzed, and the real-time noise time-frequency characteristics of the real-time circuit signal are extracted from the wavelet coefficients by setting a threshold.

[0069] Step S401: Determine the real-time noise type based on the real-time noise time-frequency characteristics and the preset second time-frequency noise type relationship.

[0070] The second time-frequency noise type relationship in this step is consistent with the first time-frequency noise type relationship in step S201, and will not be elaborated here.

[0071] Real-time noise type refers to the type of noise contained in the real-time circuit signal, which is obtained by the processing terminal by looking up the mapping table corresponding to the second time-frequency noise type relationship based on the time-frequency characteristics of the real-time noise.

[0072] Step S402: Determine the real-time noise reduction branch of the network based on the real-time noise type and the preset second noise type branch relationship.

[0073] The second noise type branch relationship in this step is consistent with the first noise type branch relationship in step S202, and will not be elaborated here.

[0074] The real-time noise reduction branch in a network refers to the network branch in a deep neural network model that performs noise reduction on real-time circuit signals. It is obtained by the processing terminal by looking up the corresponding mapping table in the second noise type branch relationship based on the real-time noise type.

[0075] Step S403: Determine the scene real-time noise reduction branch based on the real-time task scene and the preset second scene branch relationship.

[0076] The second scene branch relationship in this step is consistent with the first scene branch relationship in step S203, and will not be elaborated here.

[0077] The real-time scene denoising branch refers to the branch in the deep neural network model that performs denoising based on the real-time task scene. It is obtained by the processing terminal by looking up the corresponding mapping table in the second scene branch relationship based on the real-time task scene.

[0078] Step S404: Switch the denoising branch of the deep neural network model according to the scene real-time denoising branch and the network real-time denoising branch, and input the real-time circuit signal into the denoising branch of the deep neural network model to output the denoising circuit signal.

[0079] After determining the scene real-time denoising branch and the network real-time denoising branch, the deep neural network model automatically switches to the network real-time denoising branch within the scene real-time denoising branch. The real-time circuit signal is then input into the denoising branch of the deep neural network model for denoising, and finally, the denoising circuit signal is output. The specific method is described in [reference needed]. Figure 5 and Figure 6 The steps.

[0080] Reference Figure 5 The steps of inputting real-time circuit signals into the denoising branch of a deep neural network model to output denoising circuit signals include: Step S500: Analyze the real-time circuit signal to determine the noise energy index.

[0081] Among them, the noise energy index refers to the energy of the noise contained in the real-time circuit signal, which is obtained by the processing terminal based on the noise energy of the real-time circuit signal detected by the edge computing node.

[0082] Step S501: Determine whether the noise energy index meets the preset energy index threshold requirements.

[0083] Among them, the energy index threshold refers to the range of noise reduction of real-time circuit signals when switching between deep neural network models and lightweight algorithms. The specific value is determined by the operator according to the actual situation. The requirement for the energy index threshold is that it is not less than the energy index threshold.

[0084] By processing the terminal to determine whether the noise energy index is not less than the energy index threshold, the system can determine whether to use a deep neural network model or a lightweight algorithm to reduce noise in the real-time circuit signal.

[0085] Step S5011: If the condition is met, control the denoising branch of the deep neural network model to denoise the real-time circuit signal to generate a denoised circuit signal.

[0086] If the processing terminal determines that the noise energy index is not less than the energy index threshold, it indicates that the noise energy in the real-time circuit signal is high. Therefore, only the noise reduction branch of the deep neural network model is controlled to extract the local temporal features of the real-time circuit signal using a one-dimensional causal convolutional layer. Then, the signal is converted to the frequency domain by real-time fast Fourier transform, and the spectral features are extracted by the frequency domain convolutional layer. Finally, the time domain and frequency domain features are fused through the attention mechanism to generate the noise reduction circuit signal, thereby ensuring the noise reduction effect.

[0087] Step S5012: If it does not meet the requirements, the real-time circuit signal is smoothed according to the preset lightweight algorithm to generate a noise reduction circuit signal.

[0088] If the processing terminal determines that the noise energy index is less than the energy index threshold, it indicates that the noise energy in the real-time circuit signal is low. At this time, the processing terminal performs smoothing processing on the real-time circuit signal according to a lightweight algorithm to generate a noise-reducing circuit signal, thereby reducing the power consumption of the noise reduction process.

[0089] Lightweight algorithms refer to algorithms that smooth real-time circuit signals with low noise energy. In this embodiment, a lightweight Kalman filter is used. A simplified Kalman filter state equation is first constructed, which only contains the first derivative of the circuit signal as the state variable. Its observation equation is a linear mapping of the amplitude of the noisy signal and adopts a fixed-point operation mode. The update of the covariance matrix is ​​performed using 8-bit integer operations.

[0090] Reference Figure 6 The step of inputting the real-time circuit signal into the denoising branch of the deep neural network model to output the denoising circuit signal further includes: Step S600: Obtain the jump trigger signal of the noise energy index.

[0091] Among them, the jump trigger signal refers to the trigger signal when the noise energy index is within the energy index threshold. It is output by the processing terminal after comparing the noise energy index with the energy index threshold in real time.

[0092] Step S601: Based on the jump trigger signal, control the denoising branch of the deep neural network model to denoise the real-time circuit signal to generate the first denoised signal.

[0093] In this process, after the processing terminal receives the transition trigger signal, the processing terminal controls the denoising branch of the deep neural network model to denoise the real-time circuit signal, thereby obtaining the first denoised signal. The specific process is the same as that in step S5011.

[0094] The first denoised signal refers to the signal obtained after the deep neural network model denoises the real-time circuit signal.

[0095] Step S602: Based on the transition trigger signal, control the lightweight algorithm to smooth the real-time circuit signal to generate the second noise-reduced signal.

[0096] In this process, after the processing terminal receives the transition trigger signal, the processing terminal controls a lightweight algorithm to smooth the real-time circuit signal, thereby obtaining the second noise-reduced signal. The specific process is the same as that in step S5012.

[0097] The second noise reduction signal refers to the signal obtained after the real-time circuit signal is smoothed by a lightweight algorithm.

[0098] Step S603: Analyze the noise energy index to determine the fusion weight.

[0099] The fusion weight refers to the weight used to fuse the first and second denoised signals, which is determined by the processing terminal based on noise energy index analysis. The specific method is described in [reference needed]. Figure 7 The steps.

[0100] Step S604: The first noise reduction signal and the second noise reduction signal are weighted and fused according to the fusion weight to generate the noise reduction circuit signal.

[0101] In this step, the noise reduction circuit signal is the same as the noise reduction circuit signal in step S104. It is obtained by the processing terminal after weighted fusion of the first noise reduction signal and the second noise reduction signal according to the fusion weight.

[0102] Reference Figure 7 The steps for analyzing noise energy metrics to determine fusion weights include: Step S700: Analyze the noise energy index and the preset activation function to determine the first noise weight.

[0103] The activation function is the function that determines the weight of the first noise-reduced signal. In this embodiment, the Sigmoid function is used.

[0104] The first noise weight refers to the weight of the first noise-reduced signal. It is calculated by the processing terminal based on the difference between the noise energy index and the energy index threshold. The difference is then used as an unknown in the activation function to calculate the first noise weight. The larger the difference, the greater the weight.

[0105] Step S701: Analyze the preset total weight and the first noise weight to determine the second noise weight.

[0106] The total weight refers to the sum of the weights, which is 1. The second noise weight refers to the weight of the second noise-reduced signal, which is obtained by the processing terminal calculating the difference between the total weight and the first noise weight.

[0107] Step S702: Associate the first noise weight and the second noise weight to generate the fused weight.

[0108] The fusion weight in this step is the same as the fusion weight in step S603, and is obtained by associating the first noise weight and the second noise weight by the processing terminal.

[0109] Based on the same inventive concept, embodiments of this application provide a deep learning-based electrical signal noise reduction system, comprising: The acquisition module is used to acquire the training circuit signal set, training task scenario, real-time circuit signal, real-time task scenario, and transition trigger signal. Memory used to store programs for deep learning-based electrical signal noise reduction methods; The processor can load and execute programs in memory to implement deep learning-based electrical signal noise reduction methods.

[0110] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device, and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0111] This application provides a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as a deep learning-based electrical signal noise reduction method.

[0112] Computer storage media include, for example, USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media that can store program code.

[0113] Based on the same inventive concept, embodiments of this application provide a smart terminal, including a memory and a processor, wherein the memory stores a computer program that can be loaded and executed by the processor to perform an electrical signal noise reduction method based on deep learning.

[0114] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device, and unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0115] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A method for denoising electrical signals based on deep learning, characterized in that, include: Obtain the training circuit signal set of the preset track circuit; The training circuit signal set includes noisy signal samples and corresponding clean signal labels; The training task scenario is obtained based on the signal set of the training circuit; the training task scenario refers to the actual scenario when noisy signals occur, including temperature, humidity, vibration and electromagnetic environment scenarios. Construct a deep neural network model and train the deep neural network model based on the training circuit signal set and the corresponding training task scenario; Acquire real-time circuit signals of the track circuit and the corresponding real-time task scenario; The training deep neural network model is controlled to output noise reduction circuit signals based on real-time circuit signals and real-time task scenarios. The steps for training a deep neural network model based on the training circuit signal set and the corresponding training task scenario include: The signal set of the training circuit is analyzed to determine the time-frequency distribution characteristics of the noise; The circuit noise type is determined based on the time-frequency distribution characteristics and the preset relationship between the first time-frequency noise type. The network noise reduction branch is determined based on the circuit noise type and the preset first noise type branch relationship; The scene denoising branch is determined based on the training task scenario and the preset first scene branch relationship; the scene denoising branch refers to the branch in the deep neural network model that performs denoising according to the scene, and the scene denoising branch includes multiple network denoising branches; The noise reduction branch of the deep neural network model is switched according to the scene noise reduction branch and the network noise reduction branch, and the noise reduction branch of the deep neural network model is trained according to the signal set of the training circuit. The steps for the noise reduction branch of the deep neural network model corresponding to the training circuit signal set include: Define a joint loss function, which is composed of a linearly weighted data fitting term and a physical constraint term; the data fitting term is characterized by the mean square error between the network output and the clean label, and the physical constraint term is characterized by the sum of the loop voltage residual term and the instantaneous power residual term; The parameters of the denoising branch of the trained deep neural network model are optimized based on the preset gradient backpropagation algorithm and the training circuit signal set until the joint loss function is minimized.

2. The electrical signal denoising method based on deep learning according to claim 1, characterized in that, The steps involved in controlling the trained deep neural network model to output noise reduction circuit signals based on real-time circuit signals and real-time task scenarios include: Analyze real-time circuit signals to determine the time-frequency characteristics of real-time noise; The real-time noise type is determined based on the real-time noise time-frequency characteristics and the preset second time-frequency noise type relationship; The network real-time noise reduction branch is determined based on the real-time noise type and the preset second noise type branch relationship; The scene real-time noise reduction branch is determined based on the real-time task scenario and the preset second scene branch relationship; The denoising branch of the deep neural network model is switched between the scene real-time denoising branch and the network real-time denoising branch, and the real-time circuit signal is input into the denoising branch of the deep neural network model to output the denoising circuit signal.

3. The electrical signal denoising method based on deep learning according to claim 2, characterized in that, The steps of inputting real-time circuit signals into the denoising branch of a deep neural network model to output denoising circuit signals include: Analyze real-time circuit signals to determine noise energy indicators; Determine whether the noise energy index meets the preset energy index threshold requirements; If the conditions are met, the denoising branch of the deep neural network model is controlled to denoise the real-time circuit signal in order to generate a denoised circuit signal. If the signal does not meet the requirements, the real-time circuit signal will be smoothed according to a preset lightweight algorithm to generate a noise-reducing circuit signal.

4. The deep learning-based electrical signal noise reduction method according to claim 3, characterized in that, The step of inputting real-time circuit signals into the denoising branch of a deep neural network model to output denoising circuit signals also includes: Obtain the trigger signal for the jump in noise energy index; The noise reduction branch of the deep neural network model is controlled by the jump trigger signal to reduce the noise of the real-time circuit signal and generate the first noise-reduced signal. The lightweight algorithm controls the transition trigger signal to smooth the real-time circuit signal in order to generate a second noise-reduced signal. The noise energy index is analyzed to determine the fusion weights; The first and second noise-reduced signals are weighted and fused according to the fusion weights to generate the noise reduction circuit signal.

5. The electrical signal denoising method based on deep learning according to claim 4, characterized in that, The steps for analyzing noise energy metrics to determine fusion weights include: The noise energy index and the preset activation function are analyzed to determine the first noise weight; The preset total weight and the first noise weight are analyzed to determine the second noise weight; The first noise weight and the second noise weight are correlated to generate the fused weight.

6. A deep learning-based electrical signal noise reduction system, characterized in that, include: The acquisition module is used to acquire the training circuit signal set, training task scenario, real-time circuit signal, and real-time task scenario. A memory for storing the program of the deep learning-based electrical signal noise reduction method as described in any one of claims 1 to 5; The processor and the program in the memory can be loaded and executed by the processor to implement the deep learning-based electrical signal noise reduction method as described in any one of claims 1 to 5.

7. A smart terminal, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as described in any one of claims 1 to 5, which is based on deep learning for electrical signal noise reduction.

8. A computer-readable storage medium, characterized in that, The computer program is stored that can be loaded by a processor and execute the deep learning-based electrical signal noise reduction method as described in any one of claims 1 to 5.

Images