Noise suppression method, training method and device of noise suppression model, and chip

Abstract

The invention discloses a noise suppression method, a training method and device of a noise suppression model and a chip, and belongs to the field of signal processing. The method comprises the following steps: receiving an original signal sent through a power line; the original signal is a noisy signal; performing feature extraction based on the original signal to obtain a signal feature; performing feature extraction on the signal features through a plurality of attention heads to obtain a plurality of groups of attention features; each group of attention features comprises a query vector, a key vector and a value vector, and each group of attention features share the same value vector; performing correlation calculation on the query vector and the value vector based on the multiple groups of attention features to obtain an attention score; and performing linear transformation according to the attention score, and outputting a target signal after noise suppression. According to the invention, noise in power line carrier communication can be adaptively identified and suppressed, and the reliability and stability of communication are improved.

Description

Technical Field

[0001] The present application relates to the field of signal processing, and in particular to a noise suppression method, a training method for a noise suppression model, a device, and a chip. Background Art

[0002] Power Line Communication (PLC) uses power lines as a transmission medium to achieve high-speed information transmission. However, the power line channel itself is subject to strong multipath fading, a large amount of narrowband interference (NBI), and impulse noise, which significantly affect communication performance.

[0003] Related technologies primarily use filters in the time domain to process interference signals. For example, this approach sets a threshold and zeroes or limits signals above the threshold. However, this approach is demanding in selecting the threshold and struggles to suppress the impact of impulse noise tailing on the data.

[0004] Therefore, how to perform accurate noise suppression to improve the reliability and stability of communication is an urgent problem to be solved in this field. Summary of the Invention

[0005] This application aims to solve at least one of the technical problems existing in the related art. To this end, this application proposes a noise suppression method, a noise suppression method, and a noise suppression model training method, device, and chip to adaptively identify and suppress noise in power line carrier communication, thereby improving the reliability and stability of communication.

[0006] In a first aspect, the present application provides a noise suppression method, the method comprising:

[0007] receiving an original signal sent through a power line; the original signal being a noisy signal;

[0008] Perform feature extraction based on the original signal to obtain signal features;

[0009] Extracting features of the signal using multiple attention heads, obtaining multiple sets of attention features; each set of attention features includes a query vector, a key vector, and a value vector, and each set of attention features shares the same value vector;

[0010] performing a correlation calculation on the query vector and the value vector based on the multiple sets of attention features to obtain an attention score;

[0011] A linear transformation is performed according to the attention score, and a noise-suppressed target signal is output.

[0012] In the above technical solution, by extracting features based on the original signal, signal features are obtained, and the original features of the original signal can be learned. Then, feature extraction of the signal features is performed separately by multiple attention heads to obtain multiple groups of attention features, wherein each group of attention features includes a query vector, a key vector and a value vector, and each group of attention features shares the same value vector. This can reduce the number of parameters while introducing multiple attention distributions to capture more dimensional feature retention, and enable each attention head to share unified context information when extracting signal characteristics, which helps to reduce the deviation of feature extraction; the signal features are deeply mined through the multi-head attention mechanism, which can capture key signal information while weakening interference features, further improving the accuracy of noise suppression; then, based on the multiple groups of attention features, the query vector and the value vector are correlated to obtain an attention score, which has strong adaptability and can dynamically adjust the weight distribution according to the signal features under different channel conditions, thereby improving the robustness of communication transmission; finally, a linear transformation is performed according to the attention score, and the feature transformation is mapped to a specific communication signal to obtain the target signal after noise suppression, while maintaining the integrity of the target signal, significantly improving the signal-to-noise ratio, effectively reducing the communication bit error rate, and improving communication performance.

[0013] According to one embodiment of the present application, the performing correlation calculation on the query vector and the value vector based on the multiple sets of attention features to obtain an attention score includes:

[0014] For each set of attention features, performing a dot product calculation based on the query vector and the key vector included therein to obtain a dot product result corresponding to the targeted attention feature;

[0015] Determine an initial relevance weight corresponding to each group of attention features based on the dot product results corresponding to each group of attention features;

[0016] performing differential calculation based on the initial relevance weights corresponding to each set of attention features to obtain a target relevance weight;

[0017] A weighted calculation is performed based on the target relevance weight and the shared value vector to obtain an attention score.

[0018] In the above embodiment, by sharing the value vector, the redundant calculation of independently generating the value vector for each group of attention features is avoided, the computational complexity is reduced, and it is particularly suitable for high-dimensional signal processing scenarios; moreover, the differential operation based on the initial correlation weight can effectively remove low-correlation features and improve the quality of the target weight, thereby enhancing the model's ability to suppress noise.

[0019] In a second aspect, the present application provides a noise suppression device, comprising:

[0020] A signal collection module, configured to receive an original signal transmitted via a power line; the original signal being a noisy signal;

[0021] A feature extraction module, configured to extract features based on the original signal to obtain signal features;

[0022] An attention module is configured to extract features from the signal using multiple attention heads to obtain multiple sets of attention features; each set of attention features includes a query vector, a key vector, and a value vector, and each set of attention features shares the same value vector;

[0023] The attention module is further configured to perform a correlation calculation on the query vector and the value vector based on the multiple sets of attention features to obtain an attention score;

[0024] The output module is used to perform a linear transformation according to the attention score and output a noise-suppressed target signal.

[0025] In the above technical solution, by extracting features based on the original signal, signal features are obtained, and the original features of the original signal can be learned. Then, feature extraction of the signal features is performed separately by multiple attention heads to obtain multiple groups of attention features, wherein each group of attention features includes a query vector, a key vector and a value vector, and each group of attention features shares the same value vector. This can reduce the number of parameters while introducing multiple attention distributions to capture more dimensional feature retention, and enable each attention head to share unified context information when extracting signal characteristics, which helps to reduce the deviation of feature extraction; the signal features are deeply mined through the multi-head attention mechanism, which can capture key signal information while weakening interference features, further improving the accuracy of noise suppression; then, based on the multiple groups of attention features, the query vector and the value vector are correlated to obtain an attention score, which has strong adaptability and can dynamically adjust the weight distribution according to the signal features under different channel conditions, thereby improving the robustness of communication transmission; finally, a linear transformation is performed according to the attention score, and the feature transformation is mapped to a specific communication signal to obtain the target signal after noise suppression, while maintaining the integrity of the target signal, significantly improving the signal-to-noise ratio, effectively reducing the communication bit error rate, and improving communication performance.

[0026] In a third aspect, the present application provides a method for training a noise suppression model, the method comprising:

[0027] Acquire a training set, wherein the training set includes noisy sample signals and corresponding clean sample signals;

[0028] Extracting features from the noisy sample signal using a feature extraction network to obtain features of the sample signal;

[0029] Extracting features of the sample signal using an attention network to obtain multiple sets of attention sample features; the attention network includes multiple attention heads, each set of attention sample features includes a value vector, a key vector, and a query vector, and each set of attention sample features shares the same value vector;

[0030] Performing a correlation calculation on the query vector and the value vector through the attention network to obtain an attention prediction score;

[0031] Performing a linear transformation based on the attention prediction score through the output network, and outputting a noise-suppressed prediction sample signal;

[0032] determining a loss function based on a difference between the predicted sample signal and the clean sample signal;

[0033] The feature extraction network, the attention network and the output network are jointly trained based on the loss function to obtain a trained noise suppression model; the noise suppression model is used to suppress noise on the original signal collected from the power line for data signal transmission.

[0034] In the above technical solution, by utilizing the combination of feature extraction network and attention network, the time domain and frequency domain characteristics of the signal can be deeply explored, and the adaptability of the noise suppression model to complex noise can be improved, especially for difficult-to-handle noise types such as narrowband interference; and, through the attention mechanism of shared value vectors, the model parameters can be reduced, the risk of overfitting can be reduced, and the synergy between different attention heads can be improved, thereby enhancing the adaptability of the model in different power line channel environments; furthermore, based on the difference between the predicted sample signal and the pure sample signal, the loss function is determined, and the feature extraction network, attention network and output network are jointly trained based on the loss function to obtain a trained noise suppression model. The model thus trained can achieve signal denoising with lower latency, meet the real-time requirements of power line carrier communication, and thus ensure the stability and reliability of data signal transmission.

[0035] This end-to-end model training framework avoids the complex manual parameter tuning process used in traditional methods, simplifies the design process of noise suppression systems, and improves development efficiency.

[0036] In a fourth aspect, the present application provides a training device for a noise suppression model, the device comprising:

[0037] A sample acquisition module is used to acquire a training set, wherein the training set includes noisy sample signals and corresponding clean sample signals;

[0038] A feature extraction network module is used to extract features from the noisy sample signal through a feature extraction network to obtain features of the sample signal;

[0039] An attention network module is configured to extract features of the sample signal through an attention network to obtain multiple sets of attention sample features; the attention network includes multiple attention heads, each set of attention sample features includes a value vector, a key vector, and a query vector, and each set of attention sample features shares the same value vector;

[0040] The attention network module is further configured to perform a correlation calculation on the query vector and the value vector through the attention network to obtain an attention prediction score;

[0041] An output network module, configured to perform a linear transformation based on the attention prediction score through an output network, and output a noise-suppressed prediction sample signal;

[0042] a training module, configured to determine a loss function based on a difference between the predicted sample signal and the clean sample signal;

[0043] The training module is also used to jointly train the feature extraction network, the attention network and the output network based on the loss function to obtain a trained noise suppression model; the noise suppression model is used to suppress noise on the original signal collected from the power line for data signal transmission.

[0044] In a fifth aspect, the present application provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the noise suppression method as described in the first aspect above or the noise suppression model training method as described in the third aspect above is implemented.

[0045] In a sixth aspect, the present application provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the noise suppression method as described in the first aspect above, or the noise suppression model training method as described in the third aspect above.

[0046] In the seventh aspect, the present application provides a chip, which includes a processor and a communication interface, the communication interface and the processor are coupled, and the processor is used to run programs or instructions to implement the noise suppression method described in the first aspect above, or the noise suppression model training method described in the third aspect above.

[0047] In an eighth aspect, the present application provides a computer program product, comprising a computer program, which, when executed by a processor, implements the noise suppression method as described in the first aspect above, or the noise suppression model training method as described in the third aspect above.

[0048] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become obvious from the description below, or will be learned through practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0050] Figure 1 is a schematic diagram of an application scenario of the noise suppression method provided in some embodiments of the present application;

[0051] Figure 2 is a flowchart of a noise suppression method provided in some embodiments of the present application;

[0052] Figure 3 is a signal schematic diagram of a noise signal provided in some embodiments of the present application;

[0053] Figure 4 is a schematic diagram of a signal providing spectrum leakage in some other embodiments of the present application;

[0054] Figure 5 is a signal diagram of narrowband noise provided in some embodiments of the present application;

[0055] Figure 6 is a signal schematic diagram of sending signals provided in some embodiments of the present application;

[0056] Figure 7 is a signal diagram of a received signal provided in some embodiments of the present application;

[0057] Figure 8 is a signal diagram of a target signal obtained after noise suppression provided in some embodiments of the present application;

[0058] Figure 9 is a schematic diagram of the structure of a global network provided in some embodiments of the present application;

[0059] Figure 10A is a schematic diagram of the structure of the network layer of the noise suppression model provided in some embodiments of the present application;

[0060] Figure 10B is a schematic structural diagram of a network layer of a noise suppression model provided in some other embodiments of the present application;

[0061] Figure 10C is a schematic structural diagram of a network layer of a noise suppression model provided in some further embodiments of the present application;

[0062] Figure 11is a schematic structural diagram of a gated recurrent unit provided in some embodiments of the present application;

[0063] Figure 12 is a flowchart of a method for training a noise suppression model provided in some embodiments of the present application;

[0064] Figure 13 is a schematic diagram of the overall flow of the noise suppression process provided in some embodiments of the present application;

[0065] Figure 14 is a schematic structural diagram of a noise suppression device provided in some embodiments of the present application;

[0066] Figure 15 is a schematic structural diagram of a noise suppression model training device provided in some embodiments of the present application;

[0067] Figure 16 It is a schematic diagram of the structure of a computer device provided in some embodiments of the present application. DETAILED DESCRIPTION

[0068] The following will be combined with the accompanying drawings in the embodiments of the present application to clearly describe the technical solutions in the embodiments of the present application. Obviously, the embodiments described are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field are within the scope of protection of this application.

[0069] Unless otherwise defined, all technical and scientific terms used in this application have the same meanings as commonly understood by those skilled in the art to which this application belongs. The terms used in the specification of this application are for the purpose of describing specific embodiments only and are not intended to limit this application. The terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first" and "second" in the specification and claims of this application or the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order or a primary-secondary relationship.

[0070] References to "embodiments" in this application mean that a particular feature, structure, or characteristic described in connection with the embodiment may be included in at least one embodiment of the application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described in this application may be combined with other embodiments.

[0071] In the description of this application, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," "connected," and "attached" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; they may refer to direct connections, indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art will understand the specific meanings of the above terms in this application based on specific circumstances.

[0072] The term "and / or" in this application simply describes an association between related objects, indicating that three possible relationships exist. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " in this application generally indicates that the related objects are in an "or" relationship.

[0073] The term "multiple" in this application refers to more than two (including two). Similarly, "multiple groups" refers to more than two groups (including two groups), and "multiple pieces" refers to more than two pieces (including two pieces).

[0074] Traditional linear filtering methods, such as recursive filters (Infinite Impulse Response, IIR), rely on the assumption of signal linearity and are ineffective in complex noise environments, especially in scenarios where the noise is nonlinear. In power line carrier communication scenarios, the voltage on the power line is very high, and phenomena such as corona and insulator discharge are present, all of which can cause noise interference to communications. Due to the inherent characteristics of the power line channel, power line carrier communication often suffers from strong multipath fading, a large amount of narrowband interference, and pulse interference. Furthermore, the attenuation of the power line carrier channel is large and time-varying. The lower the voltage, the greater the line attenuation and the stronger the time-varying nature. This makes traditional filtering methods difficult to apply.

[0075] In view of this, the embodiment of the present application provides a noise suppression method. By designing a noise suppression model network structure, the recurrent neural network (RNN) and the multi-head-attention mechanism are combined to model complex signal-noise relationships. This method is not limited to linear assumptions and can accurately identify noise interference in a nonlinear noise background. In addition, it can adaptively adjust according to changes in the input signal and noise, memorize signal characteristics over long time spans, and has strong processing capabilities for continuous or intermittent noise, or noise with time dependence and complex noise types, thereby improving the reliability and stability of power line carrier communication.

[0076] It is easy to understand that the noise suppression method provided in the embodiments of the present application can be applied not only to power line carrier communication scenarios, such as industrial / home power grids and smart grid communication systems, but also to other application fields that require signal noise suppression processing, such as noise suppression for data communication between devices in industrial control networks, noise suppression for charging control signals in electric vehicle charging networks, or noise suppression for transmission signals in remote medical equipment or intelligent monitoring systems. The embodiments of the present application are described using the power line carrier communication scenario as an example, but this should not be regarded as a limitation of the present application.

[0077] The noise suppression method and the training method of the noise suppression model provided in the embodiments of the present application are described in detail below with reference to specific embodiments and their application scenarios in conjunction with the accompanying drawings.

[0078] The noise suppression method provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown in FIG. The transmitting end and receiving end refer to communication devices or modules capable of sending and receiving data signals via power lines. For example, the transmitting end may include a PLC modulation chip / module, etc., and the receiving end may include a PLC demodulation chip / module, etc.; alternatively, the receiving end may be a terminal device. The transmitting end sends a signal and transmits it via a power line channel; the receiving end receives the signal and performs noise suppression processing on the received signal to obtain a denoised signal. Thus, the denoised signal can be subsequently processed.

[0079] In some embodiments, the denoised signal may be further processed to restore the original data information. Alternatively, the denoised signal may contain instructions for controlling the operation of the receiving device, so that the denoised signal can be used to control or provide feedback on the operating status of the device.

[0080] Terminal devices include, but are not limited to, one or more of various desktop computers, laptops, smartphones, tablets, in-vehicle terminals, IoT devices, or portable wearable devices. IoT devices may be one or more of smart speakers, smart TVs, smart air conditioners, or smart in-vehicle devices. Portable wearable devices may be one or more of smart watches, smart bracelets, or head-mounted devices. A server may be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, and big data and artificial intelligence platforms.

[0081] The noise suppression method may be applied to a computer device, and may be specifically executed by hardware or software in the computer device.

[0082] The noise suppression method provided in the embodiments of the present application is applicable to a variety of power line carrier communication scenarios, including home power grids, industrial control networks, and smart grid communication systems, and can effectively improve the communication reliability of the system.

[0083] The noise suppression method provided in the embodiments of the present application can be executed by a computer device or a functional module or functional entity in the computer device that can implement the method. The noise suppression method provided in the embodiments of the present application is described below using a computer device as an example of the execution entity.

[0084] like Figure 2 As shown, the noise suppression method includes: steps 210 to 250.

[0085] Step 210: Receive an original signal sent through a power line; the original signal is a noisy signal.

[0086] The original signal refers to the signal transmitted through the power line, unprocessed. The original signal typically carries various types of noise interference, such as additive white Gaussian noise (AWGN), narrowband noise, periodic impulsive noise, and asynchronous impulsive noise. In other words, the original signal is the superposition of the target signal and the interfering signal. Figure 3 The following diagram shows various noise signal waveforms: the left side shows the time domain waveform, and the right side shows the power spectral density (PSD). From top to bottom, they represent AWGN noise, narrowband noise, periodic impulse noise, asynchronous impulse noise, and the superposition of the four noise types. In the time domain waveform, the horizontal axis represents time (in microseconds (μs)) and the vertical axis represents amplitude. In the power spectral density diagram, the horizontal axis represents frequency (in megahertz (MHz)) and the vertical axis represents power spectral density (in decibel milliwatts per hertz (dBm / Hz).

[0087] In power line carrier communication, frequency offset will aggravate spectrum leakage due to the multipath effect of the channel. Figure 4 The effect of spectrum leakage on the signal is shown. Figure 4As can be seen in the figure, in single-tone narrowband noise, a represents the frequency deviation factor. A larger a value indicates more severe spectral leakage. Spectral leakage causes signal energy to spread beyond the target frequency band, reducing communication efficiency and potentially interfering with other signals within the band. Power lines inherently contain narrowband interference and impulse noise, and spectral leakage can make signals more susceptible to interference, especially in high-noise environments.

[0088] In power line communications, narrowband noise often comes from switching devices, transformers or motors in the power grid and may exist in the form of multiple tones, that is, interference signals at multiple discrete frequencies. Figure 5 Figure 2 shows a schematic diagram of a multi-tone narrowband noise signal, where d is the normalized frequency spacing. The impact of narrowband noise on communication performance is closely related to its spectrum distribution.

[0089] If multiple narrowband interference frequencies are too close together, they may overwhelm multiple subcarriers of the communication signal, significantly increasing the system's bit error rate. Depending on the frequency spacing d, the noise impact range may be relatively concentrated or relatively dispersed, necessitating adaptive noise identification and suppression. Consequently, the traditional filtering method of presetting fixed thresholds cannot accurately identify noise.

[0090] The power line carrier communication scenario used in the embodiments of the present application can be, for example, a communication transmission scenario for typical long-distance, multi-branch medium and low voltage power lines such as underground and overhead lines. For example, the number of branches of the medium and low voltage power lines is not less than 16, and the frequency band range is not less than 100 to 500 kHz, and the power line signal transmission scenario is 0.7 MHz to 30 MHz. The computer device can collect the original signal in the power line through itself or the power line carrier communication device, and the original signal is a signal containing noise.

[0091] In some embodiments, after receiving the raw signal, the computer device may perform preprocessing on it to improve the effectiveness of subsequent processing. Preprocessing steps include, but are not limited to, removing the DC component (i.e., DC offset) and signal normalization (i.e., adjusting the signal value to a specific range, such as between 0 and 1) to eliminate redundant and interfering information in the data. Preprocessing steps may also include signal segmentation or windowing to better capture local features of the signal during subsequent processing.

[0092] Step 220: Perform feature extraction based on the original signal to obtain signal features.

[0093] Computer devices can process the original signal using a pre-trained noise suppression model to obtain a noise-suppressed target signal. This noise suppression model consists of an input layer, a hidden layer, and an output layer. The input layer receives the pre-processed signal data, i.e., the original signal. The hidden layer combines a recurrent neural network with an attention mechanism to capture the temporal features in the original signal. The attention mechanism adaptively adjusts the weights of the signal features, thereby highlighting key signal information while weakening interference features. This allows adaptive identification and suppression of different types of noise, improving the quality of the target signal and enhancing communication performance. The output layer outputs the noise-suppressed result, i.e., the target signal.

[0094] Specifically, the computer device first extracts features from the original signal through a feature extraction network, captures the global or local features of the signal, and generates a signal representation of the original signal, namely, the signal feature.

[0095] In some embodiments, the feature extraction network includes one or more of a recurrent neural network and its variants, a convolutional neural network (CNN), a gated recurrent unit (GRU), a long short-term memory network (LSTM), etc.

[0096] Step 230: extract the signal features through multiple attention heads to obtain multiple groups of attention features; each group of attention features includes a query vector, a key vector and a value vector, and each group of attention features shares the same value vector.

[0097] Specifically, the computer device performs deep feature extraction on signal features through an attention network. The attention network can include multiple attention heads, each of which has an independent linear unit for generating a set of attention features. The linear unit can include one or more linear layers. As a result, each attention head can capture different subspace information of the input original signal, thereby achieving multi-angle feature extraction.

[0098] Among them, a set of attention features includes one or more query vectors Q, one or more key vectors K and a value vector V, where multiple attention heads share the same set of value vectors to capture the correlation of signals between different feature dimensions. During the entire calculation process, the value vector V is shared, that is, all attention calculations are based on the same value vector V. This can reduce the number of parameters while introducing multiple attention distributions to capture feature retention in more dimensions. In addition, the shared value vector enables each attention head to share unified contextual information when extracting signal characteristics, which helps to reduce the deviation of feature extraction.

[0099] Step 240: perform correlation calculation on the query vector and the value vector based on multiple sets of attention features to obtain an attention score.

[0100] Based on the attention mechanism, the computer device performs a dot product operation on the query vector and the key vector, calculates their correlation, and then calculates an attention score. The computer device then uses the attention score as a weight to perform a weighted calculation on the value vector to generate a noise-suppressed signal representation.

[0101] Step 250: Perform a linear transformation based on the attention score and output the target signal after noise suppression.

[0102] The computer device can perform linear transformation on the above-mentioned noise-suppressed signal representation through a linear transformation module to generate a final target signal, which has significantly reduced noise interference and has higher communication quality.

[0103] For example, in Figure 1 Based on Figures 6 to 8 , Figure 6 shows the waveform of the transmission signal sent by the transmitting end, Figure 7 The received signal is the received signal received by the receiving end, which is superimposed with noise interference compared to the transmitted signal. After the computer device suppresses the noise of the received signal through the above steps, the output target signal is as follows Figure 8 As shown, the target signal significantly reduces noise interference compared to the received signal.

[0104] The target signal can be used for subsequent communication processing, such as demodulation, decoding, etc.

[0105] The noise suppression method provided in the embodiment of the present application obtains signal features by extracting features based on the original signal, and can learn the original features of the original signal. Then, the signal features are extracted by multiple attention heads to obtain multiple groups of attention features, wherein each group of attention features includes a query vector, a key vector and a value vector, and each group of attention features shares the same value vector. This can reduce the number of parameters while introducing multiple attention distributions to capture more dimensional feature retention, and enable each attention head to share unified context information when extracting signal characteristics, which helps to reduce the deviation of feature extraction; the signal features are deeply mined through the multi-head attention mechanism, which can capture key signal information while weakening interference features, further improving the accuracy of noise suppression; then, based on the multiple groups of attention features, the query vector and the value vector are correlated to obtain an attention score, which has strong adaptability and can dynamically adjust the weight distribution according to the signal features under different channel conditions, thereby improving the robustness of communication transmission; finally, a linear transformation is performed according to the attention score, and the feature transformation is mapped to a specific communication signal to obtain the target signal after noise suppression, while maintaining the integrity of the target signal, significantly improving the signal-to-noise ratio, effectively reducing the communication bit error rate, and improving communication performance.

[0106] In the noise suppression task of power line carrier communication, signal processing needs to be effectively modeled and optimized for complex noise environments. Traditional methods rely on linear filtering technology, but their effectiveness is limited in the coexistence of multiple interference types and in nonlinear backgrounds. The attention mechanism in deep learning technology has gradually become a solution due to its efficient modeling ability of feature relationships. By modeling the key characteristics of the input signal through multiple sets of attention features, the correlation of the signal in the time and frequency domain can be extracted, thereby suppressing the interference signal. However, calculating independent weights for different attention heads may lead to redundant calculations and increase the complexity of the model. Therefore, this application proposes a method for calculating the correlation of multiple sets of attention features through shared value vectors, which effectively reduces the amount of calculation while ensuring the adequacy and accuracy of feature extraction.

[0107] Based on this, in some embodiments, the query vector and the value vector are correlated based on multiple groups of attention features to obtain an attention score, including: for each group of attention features, a dot product calculation is performed based on the included query vector and key vector to obtain a dot product result corresponding to the targeted attention feature; based on the dot product results corresponding to each group of attention features, the initial correlation weight corresponding to each group of attention features is determined; based on the initial correlation weight corresponding to each group of attention features, a differential calculation is performed to obtain a target correlation weight; and a weighted calculation is performed based on the target correlation weight and the shared value vector to obtain an attention score.

[0108] Specifically, the computer device outputs multiple sets of attention features through multiple attention heads, such as the first set of attention features: query vector Q1, key vector K1, and value vector V; the second set of attention features: query vector Q2, key vector K2, and value vector V; etc. Each set of attention features shares the same value vector V.

[0109] For each set of attention features, the computer device performs a dot product calculation based on the query vector Q and the key vector K to obtain a dot product result for each set of attention features. For example, for the first set of attention features, the computer performs a dot product calculation based on the query vector Q1 and the key vector K1 to obtain a dot product result Q1K1.

[0110] Then, for each set of attention features, the computer device performs a normalization operation on the corresponding dot product results to obtain the importance score of each key vector to the query vector to generate an initial relevance weight. For example, the computer device can perform normalization processing through the Softmax function, which can be expressed as: Softmax(Q i K i T ), i is a positive integer. The output of Softmax is a probability distribution, which represents the correlation weight between the query vector and the key vector.

[0111] Furthermore, the computer device performs differential calculation based on the initial correlation weight to reduce the influence of redundant or low-correlation features, eliminate attention noise, and generate target correlation weights, thereby enabling the attention mechanism to have a better effect. In some embodiments, the computer device can perform differential calculations on the initial correlation weights of a certain group and the initial correlation weights of other groups to obtain target correlation weights. The initial correlation weights of other groups can be represented by cumulative average, weighted sum, etc. Furthermore, the computer device uses the target correlation weight to perform weighted calculations on the shared value vector to obtain the final attention score. For example, the computer device can multiply the target correlation weight by the value vector to obtain an attention score.

[0112] Exemplarily, the step of calculating the attention score P by the computer device can be expressed by the following formula:

[0113]

[0114] Wherein, λ is a preset weight coefficient, and N is the number of groups of attention features. In some embodiments, each group of attention features may correspond to its own weight coefficient, i.e., λ = λ1, λ2, ..., λ N , N is a positive integer.

[0115] In the above embodiment, by sharing the value vector, the redundant calculation of independently generating the value vector for each group of attention features is avoided, the computational complexity is reduced, and it is particularly suitable for high-dimensional signal processing scenarios; moreover, the differential operation based on the initial correlation weight can effectively remove low-correlation features and improve the quality of the target weight, thereby enhancing the model's ability to suppress noise.

[0116] In some embodiments, feature extraction is performed based on the original signal to obtain signal features, including: extracting features from the original signal through a global network to obtain global features of the original signal; inputting the global features and the original signal into a hierarchical network for feature extraction to obtain hierarchical features of the original signal; and using the hierarchical features as signal features of the original signal.

[0117] Feature extraction networks include global networks and hierarchical networks. Global networks excel at extracting the overall distribution characteristics of the signal, while hierarchical networks can deeply explore local details and hierarchical structures. By combining global and local information, the noise suppression effect is improved.

[0118] Specifically, the computer device receives the original signal, inputs it into the global network, and extracts the global features of the original signal. The global network can be one or more of a convolutional neural network, a recurrent neural network, a gated recurrent unit, or a long short-term memory network. For example, Figure 9 As shown, the global network, for example, consists of a dense layer, a nonlinear activation layer (Tanh layer), a gated recurrent unit (GRU), and a linear activation layer (Relu layer). The dense layer transforms the input signal's dimensionality, mapping the signal features into a high-dimensional space. The nonlinear activation layer, through the tanh activation function, enhances the nonlinear representation of the signal features, facilitating the capture of complex signal characteristics. The GRU captures the signal's time series characteristics. The linear activation layer optimizes the output range of the signal features, generating a global feature representation for subsequent hierarchical network processing. Thus, the combination of the dense and nonlinear activation layers maps the original signal to a higher dimension and enhances nonlinear capabilities, laying the foundation for subsequent feature extraction. The GRU, through its recurrent structure, memorizes historical information and effectively extracts temporal characteristics from the signal. It is particularly suitable for processing highly continuous signals, particularly in power line communication transmission scenarios. The linear activation layer further adjusts the range of feature output, improving the discriminability and stability of signal features. The multi-layer design of the global network allows for flexible adjustment, allowing the appropriate network structure to be selected based on specific application scenarios, resulting in excellent scalability and adaptability. Moreover, the output of global feature extraction can significantly improve the feature extraction effect of subsequent hierarchical networks, thereby improving the overall noise suppression performance.

[0119] Then, the computer device takes the global features and the original signal as joint input and passes them to the hierarchical network to extract the hierarchical features of the signal. The hierarchical network can be one or more of a stacked multi-layer perceptron, a convolutional neural network, a recurrent neural network, a gated recurrent unit, or a long short-term memory network. For example, as Figure 9 As shown, the hierarchical network includes, for example, a gated recurrent unit layer and a linear activation layer to further process global features and capture more subtle signal and local characteristics. Ultimately, the computer device uses the output of the hierarchical network as a signal feature for subsequent noise suppression.

[0120] In the above embodiment, through the combination of the global network and the hierarchical network, the overall distribution characteristics and local detail characteristics of the signal can be captured at the same time, making the signal characteristics more comprehensive; and, by jointly inputting the global features and the original signal into the hierarchical network, multiple interactions of feature information can be realized in the network hierarchy, thereby improving the accuracy of feature extraction, and then the precise signal characteristics help the subsequent noise suppression module to better separate the target signal from the noise, further improving the communication performance.

[0121] In order to generate a more accurate signal feature representation, in some embodiments, after obtaining the hierarchical features of the original signal, the method further includes: fusing the global features and the hierarchical features through a fusion network to obtain fused features of the original signal; and using the fused features as the signal features of the original signal.

[0122] That is, the feature extraction network also includes a fusion network for further fusing global features and hierarchical features, optimizing the expression of features, so that the fused features finally obtained can more fully reflect the key information of the original signal.

[0123] In the above embodiment, by fusing global features and hierarchical features, the accuracy of signal feature expression is enhanced, so that the noise in the original signal can be more accurately identified and separated and suppressed; and it can show better robustness in complex scenarios, reducing the performance degradation caused by single local features or insufficient global features.

[0124] In a specific example, Figure 10A and Figure 10B As shown in , the noise suppression model includes an input layer, a hidden layer, and an output layer. The input layer receives the preprocessed original signal. The input of each time step can be a sampling point of the signal or a segmented signal in a sliding window, and the segmented signal contains multiple sampling points. The hidden layer can include a feature extraction network and an attention network. Among them, the feature extraction network can include a global network and a hierarchical network, such as Figure 10A Alternatively, Figure 10BAs shown in Figure 2, the feature extraction network can also include a global network, a hierarchical network, and a fusion network. The attention network includes multiple attention heads. Figure 10C , each attention head includes a linear unit, and a linear unit can be one or more linear layers (Linear). The attention network obtains multiple sets of attention features through multiple attention heads, and calculates attention scores based on the multiple sets of attention features. Furthermore, the attention network can output through a normalization layer (LayerNorm), a linear layer, and one or more multilayer perceptrons (Multilayer Perceptron, MLP). The output layer is connected to the last unit of the hidden layer and is used to output the signal after noise suppression. The output can be the predicted value of each time step or the predicted value of the entire signal sequence. The mathematical expression of the entire noise suppression model is as follows:

[0125]

[0126] in, is the model output, R is the model input, are the parameter sets of the global network f1, the hierarchical network f2, and the attention network f3. If the feature extraction network also includes a fusion network, the above formula can also introduce As the parameter set of the fusion network f4.

[0127] For example, the structure of the gated recurrent unit can be as follows Figure 11 As shown. Among them, r is the reset gate, z is the update gate, state is the internal state, and h is the hidden state. In the noise suppression task, the gated recurrent unit learns the timing characteristics of the signal through the recurrent structure, extracts the features of the input signal and dynamically filters the noise signal. For example, receiving the input signal h at the current moment t and the state h at the previous moment t-1 , as the basis for subsequent calculations. The gated recurrent unit first calculates the reset gate r, which determines how much information in the previous state needs to be forgotten to suppress historical noise that is irrelevant to the current input; then based on the current input signal h tThe gated recurrent unit uses the state at the previous moment after the reset to determine a candidate state. This candidate state captures the valid features of the current input while eliminating short-term noise interference. The gated recurrent unit then calculates the update gate z to dynamically adjust the weight of the current signal features and historical information. A weighted calculation is then performed based on the update gate z and the candidate states to obtain the final state, which integrates the valid features of the current signal and historical information while effectively suppressing noise. Thus, the gated recurrent unit fully utilizes the correlation between the input signal and historical information while dynamically adjusting signal features. The reset gate suppresses irrelevant noise, and the update gate balances the temporal continuity of the signal with the current valid features, thereby achieving efficient noise suppression and signal feature extraction.

[0128] In power line carrier communications, signal transmission is severely affected by multipath fading, narrowband interference, and impulse noise, making it difficult for the receiving end to directly and effectively process the original signal. To enhance the comprehensive analysis of the original signal's time and frequency domain characteristics, it is usually necessary to preprocess the received original signal, dividing the signal into multiple segments to capture its local characteristics. This is combined with the splicing operation of frequency and time domain features to provide higher-quality data input for subsequent feature extraction. Sliding window technology, as an effective signal segmentation method, can flexibly adjust the window length according to the frequency characteristics of the power line carrier signal, thereby maximizing the continuity and integrity of the signal characteristics.

[0129] Based on this, in some embodiments, after receiving the original signal sent through the power line, the above method also includes: performing time series segmentation on the original signal to obtain multiple segments of the original signal; wherein the time series segmentation adopts a sliding window technology, and the length of each sliding window is determined based on the frequency characteristics of the power line carrier signal; performing fast Fourier transform on each segment of the original signal to obtain the frequency domain characteristics of each segment of the original signal, and retaining the time domain characteristics of each segment of the original signal; splicing the frequency domain characteristics and time domain characteristics of each segment of the original signal, and each segment of the spliced original signal is used for feature extraction.

[0130] After receiving the original signal transmitted via the power line, the computer device performs time series segmentation using a sliding window technique. The length of each sliding window is dynamically adjusted based on the frequency characteristics of the power line carrier signal to ensure that the signal characteristics are fully captured. Adjacent sliding windows can have a certain overlap area to preserve signal continuity across windows. This overlapping sliding window design reduces the boundary effects caused by time series segmentation and preserves the overall timing information of the signal.

[0131] The computer can perform a Fast Fourier Transform (FFT) on each segment of the original signal to extract its frequency domain features. At the same time, the original time domain feature information of each signal segment is retained to avoid losing important details related to the timing.

[0132] The computer then concatenates the frequency and time domain features of each signal segment to create a signal representation that combines time and frequency characteristics. This concatenated signal not only contains the signal's frequency distribution but also preserves its evolution over time, providing a comprehensive data foundation for subsequent deep feature extraction.

[0133] In the above embodiment, sliding window segmentation and time-frequency feature splicing help to accurately capture the local features of the signal, and the local features are effectively combined with the global features, thereby enhancing the ability to describe the original signal and improving the noise suppression effect.

[0134] Correspondingly, after outputting the target signal after noise suppression, the above method further includes: reconstructing the target signal to obtain a reconstructed target signal; the reconstructed target signal is used for data signal transmission of power line carrier communication. Specifically, the computer device performs a signal reconstruction operation on the output noise-suppressed target signal. During the reconstruction process, based on the spectrum distribution characteristics of the target signal, a filter or transformation algorithm is used to optimize and adjust the signal to ensure that the signal characteristics meet the requirements of power line carrier communication. Furthermore, the reconstructed target signal is used for data signal transmission in power line carrier communication, and the signal is transmitted over long distances using the power line as a medium. The reconstructed signal can significantly reduce the impact of multipath interference, narrowband noise and other external noise during the transmission process.

[0135] For example, the computer device can use an adaptive filter to smooth the spectrum of the target signal to eliminate spectral artifacts that may be introduced during the suppression process. In another example, the computer device can use a bandpass filter to limit the bandwidth of the target signal to ensure that it conforms to the frequency range of power line carrier communication (e.g., 2-30 MHz).

[0136] Subsequently, the computer can load the reconstructed signal into the physical layer modulation scheme of power line carrier communication, such as orthogonal frequency division multiplexing (OFDM) or single carrier modulation, to further improve the signal's transmission stability and anti-interference capabilities. The computer can also verify the reliability and consistency of the reconstructed target signal during transmission by demodulating and restoring the data content.

[0137] In the above embodiments, the signal reconstruction process is used to effectively repair signal distortion or spectrum unevenness that may be caused by noise suppression, thereby ensuring signal integrity and consistency; and by optimizing the frequency distribution and bandwidth usage of the target signal, signal reconstruction can improve the spectrum efficiency of the power line carrier communication system.

[0138] In power line carrier communications, the target signal after noise suppression may suffer from missing or distorted frequency components due to filtering, feature extraction, or transformation steps in the suppression process, which can affect signal transmission performance and demodulation efficiency. Therefore, after noise suppression, the signal must undergo a reconstruction process to restore its spectral integrity and time domain characteristics. Extracting the signal's frequency components through a fast Fourier transform (FFT), combined with filtering compensation and inverse Fourier transform (IFFT) reconstruction, effectively improves signal quality, providing greater signal reliability and consistency for subsequent communication processes.

[0139] Based on this, in some embodiments, the target signal is reconstructed to obtain a reconstructed target signal, including: performing a fast Fourier transform on the target signal to obtain the frequency component of the target signal; based on the frequency component, filtering the target signal through a preset filter to obtain a compensated target signal; performing an inverse fast Fourier transform on the compensated target signal to obtain a reconstructed target signal.

[0140] The computer equipment first performs a fast Fourier transform on the target signal after noise suppression to obtain the frequency components of the target signal and analyzes the spectral distribution characteristics of the signal, including frequency amplitude and phase information.

[0141] Computer equipment can use preset filters, such as bandpass filters or adaptive filters, to filter the frequency components of the target signal, suppressing noise or interfering frequencies and highlighting the retained frequency components, thereby making the effective frequency components of the target signal more prominent. For missing or weakened frequency components, gain compensation technology is applied to restore the complete spectral distribution of the target signal to compensate for signal distortion or frequency component weakening caused by noise suppression in the spectrum. The filter parameters are pre-designed based on the frequency band characteristics of power line carrier communication.

[0142] The computer then performs an inverse fast Fourier transform on the filtered and compensated frequency components, restoring the signal from the frequency domain to the time domain to obtain the reconstructed target signal. This reconstructed signal exhibits improved interference immunity and spectral uniformity, significantly reducing the complexity of signal recovery and demodulation at the receiving end and improving the overall efficiency and stability of the communication link.

[0143] In the above-described embodiments, the use of fast Fourier transforms and filtering compensation can repair spectrum attenuation or distortion caused by the noise suppression process, bringing the signal's frequency components closer to the original signal. Furthermore, the filtering compensation process effectively suppresses residual noise while optimizing the frequency amplitude and phase characteristics of the target signal, thereby improving signal clarity and stability. Restoring the signal to the time domain through an inverse fast Fourier transform ensures that the signal's time domain performance meets the requirements of power line carrier communication, based on frequency domain optimization.

[0144] The noise suppression model in the embodiments of the present application is pre-trained and deployed at the receiving end. To this end, the present application also provides a method for training the noise suppression model. The method can be performed by a computer device or a functional module or functional entity within the computer device that can implement the method. The following describes the method for training the noise suppression model provided in the embodiments of the present application, using a computer device as an example.

[0145] like Figure 12 As shown, the method includes steps 1210 to 1270:

[0146] Step 1210: Obtain a training set, where the training set includes noisy sample signals and corresponding clean sample signals.

[0147] The computer device can obtain training data including noisy sample signals and clean sample signals. The noisy sample signals represent signals interfered with by noise in power line carrier communication, and the clean sample signals are corresponding noise-free signals.

[0148] In some embodiments, the power line carrier communication scenario can be, for example, a communication transmission scenario for typical long-distance, multi-branch medium and low voltage power lines, such as those located underground or overhead. For example, the number of branches of the medium and low voltage power lines is no less than 16, and the frequency band range is no less than 100 to 500 kHz, and the power line signal transmission scenario is 0.7 MHz to 30 MHz. The computer device can collect the original signal and the received signal from the power line through itself or a power line carrier communication device. The original signal is a signal containing noise.

[0149] In some embodiments, after receiving the raw signal, the computer device may perform preprocessing on it to improve the effectiveness of subsequent processing. Preprocessing steps include, but are not limited to, removing the DC component (i.e., DC offset) and signal normalization (i.e., adjusting the signal value to a specific range, such as between 0 and 1) to eliminate redundant and interfering information in the data. Preprocessing steps may also include signal segmentation or windowing to better capture local features of the signal during subsequent processing.

[0150] The computer device can divide the preprocessed signal data into a training set and a test set. The training set is used to train the model, and the test set is used to evaluate the performance of the model.

[0151] In an embodiment of the present application, the noise suppression model is obtained by combining a recurrent neural network and an attention mechanism. The recurrent neural network extracts the timing characteristics of the power line carrier communication signal, and the attention mechanism adaptively adjusts the weights of the signal features, thereby highlighting the key signal information in the signal and weakening the interference features, thereby adaptively identifying and suppressing different types of noise, improving the quality of the target signal, and improving communication performance; the output layer is used to output the result after noise suppression, that is, the target signal. The input layer of the noise suppression model receives the preprocessed signal data, and the input of each time step can be a sampling point of the signal or a window containing multiple sampling points. The middle layer consists of three parts, including a feature extraction network (including a global network f1 and a hierarchical network f2) and an attention network f3. Exemplarily, the global network f1 includes a Dense layer, 24 GRUs, a Tanh layer, and a Relu layer, and the hierarchical network f2 includes 48 GRUs and Relu layers. The attention network f3 can be, for example, a differential Transformer model, which first performs global attention extraction, obtains multiple sets of attention features through several Linear layers, and then calculates the attention score through the multiple sets of attention features to obtain an attention prediction score. When calculating attention scores, a differential attention mechanism is used to further eliminate attention noise, enhancing the effectiveness of the attention mechanism. For example, the attention network can produce outputs through LayerNorm layers, Linear layers, and multilayer perceptrons. The output layer is connected to the last unit in the hidden layer and outputs the noise-suppressed signal. The output can be a predicted value for each time step or the predicted value for the entire signal sequence.

[0152] The specific processing steps of the noise suppression model are described below.

[0153] Step 1220: extract features from the noisy sample signal through a feature extraction network to obtain sample signal features.

[0154] Step 1230: extract the sample signal features through the attention network to obtain multiple groups of attention sample features; the attention network includes multiple attention heads, each group of attention sample features includes a value vector, a key vector and a query vector, and each group of attention sample features shares the same value vector.

[0155] Step 1240: Calculate the correlation between the query vector and the value vector through the attention network to obtain an attention prediction score.

[0156] Step 1250: Perform a linear transformation based on the attention prediction score through the output network, and output the noise-suppressed prediction sample signal.

[0157] The computer device can use the feature extraction network to process the noisy sample signal and extract the key features of the signal. Afterwards, the computer device can input the extracted sample signal features into the attention network. The attention network contains multiple attention heads, each of which independently generates a query vector (Query), a key vector (Key), and a value vector (Value). All attention heads share the same value vector. Furthermore, the computer device calculates an attention score based on the query vector and the value vector, indicating the importance of different features in the noise suppression task. When calculating the attention score, the differential attention mechanism is used to further eliminate attention noise and make the attention mechanism more effective. Finally, the computer device uses the output network to perform a linear transformation on the attention score to generate a denoised predicted sample signal. The output network is usually a set of fully connected layers.

[0158] Step 1260: Determine a loss function based on the difference between the predicted sample signal and the clean sample signal.

[0159] The computer device compares the predicted sample signal with the pure sample signal, calculates the difference, and obtains a loss function to measure the difference between the predicted signal and the true signal or the pure signal.

[0160] The loss function may be one or more of the mean absolute error, logarithmic mean square error, smooth L1 loss, frequency domain difference loss, etc. between the predicted sample signal and the clean sample signal. For example, the loss function may be the mean square error (MSE). A smaller MSE indicates better model performance.

[0161]

[0162] Among them, L is the loss, R is the received signal at the receiving end, and R p is a pure signal, and F represents the F-norm.

[0163] Step 1270: Jointly train the feature extraction network, the attention network, and the output network based on the loss function to obtain a trained noise suppression model; the noise suppression model is used to suppress noise on the original signal collected from the power line for data signal transmission.

[0164] Based on the loss function, the computer device jointly trains the feature extraction network, attention network, and output network. The fully trained noise suppression model can be used to process the original signal in power line carrier communication in real time, generating a noise-suppressed signal to support data transmission. For example, the computer device can use the Adam optimization algorithm to train the model. The Adam algorithm combines the advantages of the momentum method and the RMSprop algorithm, capable of adaptively adjusting the learning rate to accelerate model convergence.

[0165] During training, the computer inputs the training set data into the model, calculates the loss function, and updates the model parameters through the backpropagation algorithm. The loss of the validation set can be monitored during training to prevent overfitting of the model.

[0166] For example, the model parameters and corresponding parameter configurations can be referred to as shown in Table 1 below:

[0167] Table 1

[0168] parameter Parameter configuration training set 8000 Test set 2000 Number of iterations 100 Learning rate 0.01 Optimizer Adam Loss Function MSE

[0169] In practical applications, a computer device can deploy the trained model to the receiving end of a power line system and input the received power line carrier communication signal into the noise suppression model. The noise suppression model adaptively identifies and suppresses noise based on the timing characteristics of the input signal, and outputs a noise-suppressed target signal. The target signal can then be used for subsequent communication processing, such as demodulation and decoding.

[0170] The training method of the noise suppression model provided in the embodiment of the present application can deeply explore the time domain and frequency domain characteristics of the signal by utilizing the combination of the feature extraction network and the attention network, and improve the adaptability of the noise suppression model to complex noise, especially to difficult-to-handle noise types such as narrowband interference; and, through the attention mechanism of the shared value vector, the model parameters can be reduced, the risk of overfitting can be reduced, and the synergy between different attention heads can be improved, thereby enhancing the adaptability of the model in different power line channel environments; furthermore, based on the difference between the predicted sample signal and the pure sample signal, the loss function is determined, and the feature extraction network, the attention network and the output network are jointly trained based on the loss function to obtain a trained noise suppression model. The model thus trained can achieve signal denoising with lower latency, meet the real-time requirements of power line carrier communication, and thus ensure the stability and reliability of data signal transmission.

[0171] This end-to-end model training framework avoids the complex manual parameter tuning process used in traditional methods, simplifies the design process of noise suppression systems, and improves development efficiency.

[0172] Based on the above noise suppression method and noise suppression model training method, the overall process of noise suppression in this application is as follows: Figure 13 As shown, raw signal data from power line carrier communication (PLC) is first collected, including noisy received and transmitted signals. Computer equipment preprocesses the raw signal data, such as removing DC components and normalizing it, to eliminate redundancy and interference. A noise suppression model based on a recurrent neural network is constructed, consisting of an input layer, hidden layers, and an output layer. The input layer receives the preprocessed signal data. The hidden layer uses a network structure that combines a recurrent neural network with an attention mechanism to capture signal characteristics, including but not limited to one or more of frequency characteristics, delay characteristics, multipath effects, attenuation characteristics, interference characteristics, and noise characteristics. Frequency characteristics refer to the fact that PLC signals are typically transmitted at frequencies higher than the power transmission frequency (e.g., 50 / 60 Hz). Delay characteristics refer to the delay in signal transmission caused by the length and topology of the power line. This delay can be deterministic (e.g., determined by line length) or random (e.g., affected by load variations). Multipath effects refer to the fact that in complex power line environments, signals may reach the receiver via multiple paths, resulting in multipath interference and, in turn, random variations in the signal's phase and amplitude. Attenuation refers to the frequency-dependent attenuation of signals transmitted over power lines, especially at high frequencies. Since signals of different frequency components may arrive at the receiver at different times, this can affect signal timing. Interference refers to the presence of a large number of electrical devices on power lines, which can generate various interference signals that affect PLC signal timing, particularly transient interference and noise. Noise refers to the non-Gaussian and time-varying nature of power line noise, including background noise, narrowband interference, and impulse noise, which can lead to unstable signal timing.

[0173] After acquiring the captured signal and preprocessing the data, the computer first completes model training, using the captured received signal as the input for the noise suppression model. Training is performed using the transmitted signal as the label and the mean square error (MSE) as the loss function. After model training is complete, the computer deploys the noise suppression model at the receiving end of the PLC communication link to implement noise suppression on the received signal.

[0174] Therefore, by using a noise suppression model to suppress noise in power line carrier communications, communication reliability can be effectively improved and communication failures caused by noise interference can be reduced. Furthermore, the noise suppression model adaptively identifies and suppresses different types of noise, allowing communication signals to maintain good stability in different environments. Compared with traditional noise suppression methods, this application eliminates the need for additional hardware at the receiving end and implements noise suppression through software algorithms, reducing implementation costs and having industrial application value.

[0175] The noise suppression method provided in the embodiment of the present application can be executed by a noise suppression device. In the embodiment of the present application, the noise suppression device provided in the embodiment of the present application is described by taking the noise suppression method executed by the noise suppression device as an example.

[0176] The embodiment of the present application also provides a noise suppression device, which is applied to computer equipment. Figure 14 As shown, the noise suppression device includes a signal collection module 1401, a feature extraction module 1402, an attention module 1403 and an output module 1404.

[0177] in:

[0178] The signal collection module 1401 is used to receive the original signal sent through the power line; the original signal is a noisy signal.

[0179] The feature extraction module 1402 is used to extract features based on the original signal to obtain signal features.

[0180] The attention module 1403 is used to extract signal features through multiple attention heads to obtain multiple groups of attention features; each group of attention features includes a query vector, a key vector and a value vector, and each group of attention features shares the same value vector.

[0181] The attention module 1403 is further configured to perform correlation calculation on the query vector and the value vector based on multiple sets of attention features to obtain an attention score.

[0182] The output module 1404 is used to perform linear transformation according to the attention score and output the target signal after noise suppression.

[0183] According to the noise suppression device provided by the embodiment of the present application, by extracting features based on the original signal, signal features are obtained, and the original features of the original signal can be learned. Then, feature extraction is performed on the signal features through multiple attention heads to obtain multiple groups of attention features, wherein each group of attention features includes a query vector, a key vector and a value vector, and each group of attention features shares the same value vector. This can reduce the number of parameters while introducing multiple attention distributions to capture more dimensional feature retention, and enable each attention head to share unified context information when extracting signal characteristics, which helps to reduce the deviation of feature extraction; through the multi-head attention mechanism, the signal features are deeply mined, which can capture key signal information while weakening interference features, further improving the accuracy of noise suppression; then, based on the multiple groups of attention features, the query vector and the value vector are correlated to obtain an attention score, which has strong adaptability and can dynamically adjust the weight distribution according to the signal features under different channel conditions, thereby improving the robustness of communication transmission; finally, a linear transformation is performed according to the attention score, and the feature transformation is mapped to a specific communication signal to obtain the target signal after noise suppression, while maintaining the integrity of the target signal, significantly improving the signal-to-noise ratio, effectively reducing the communication bit error rate, and improving communication performance.

[0184] In some embodiments, the attention module is also used to perform dot product calculations on each group of attention features based on the included query vector and key vector to obtain dot product results corresponding to the targeted attention features; determine the initial correlation weights corresponding to each group of attention features based on the dot product results corresponding to each group of attention features; perform differential calculations based on the initial correlation weights corresponding to each group of attention features to obtain target correlation weights; and perform weighted calculations based on the target correlation weights and the shared value vector to obtain an attention score.

[0185] In some embodiments, the feature extraction module is also used to extract features from the original signal through a global network to obtain global features of the original signal; input the global features and the original signal into a hierarchical network for feature extraction to obtain hierarchical features of the original signal; and use the hierarchical features as signal features of the original signal.

[0186] In some embodiments, the feature extraction module is further configured to fuse the global features and the hierarchical features through a fusion network to obtain fused features of the original signal; and use the fused features as signal features of the original signal.

[0187] In some embodiments, the above-mentioned device also includes a preprocessing module for performing time series segmentation on the original signal to obtain multiple segments of the original signal; wherein the time series segmentation adopts a sliding window technology, and the length of each sliding window is determined based on the frequency characteristics of the power line carrier signal; each segment of the original signal is subjected to a fast Fourier transform to obtain the frequency domain characteristics of each segment of the original signal, and retain the time domain characteristics of each segment of the original signal; the frequency domain characteristics and time domain characteristics of each segment of the original signal are spliced, and each segment of the spliced original signal is used for feature extraction.

[0188] In some embodiments, the above-mentioned device further includes a post-processing module for performing signal reconstruction on the target signal to obtain a reconstructed target signal; the reconstructed target signal is used for data signal transmission of power line carrier communication.

[0189] In some embodiments, the post-processing module is also used to perform a fast Fourier transform on the target signal to obtain the frequency component of the target signal; based on the frequency component, the target signal is filtered through a preset filter to obtain a compensated target signal; and the compensated target signal is inversely fast Fourier transformed to obtain a reconstructed target signal.

[0190] The noise suppression method provided in the embodiment of the present application can be executed by a noise suppression device. In the embodiment of the present application, the noise suppression device provided in the embodiment of the present application is described by taking the noise suppression method executed by the noise suppression device as an example.

[0191] The present application also provides a noise suppression model training device for use in computer equipment. Figure 15 As shown, the training device of the noise suppression model includes a sample acquisition module 1501, a feature extraction network module 1502, an attention network module 1503, an output network module 1504 and a training module 1505.

[0192] The sample acquisition module 1501 is used to acquire a training set, which includes noisy sample signals and corresponding clean sample signals.

[0193] The feature extraction network module 1502 is used to extract features from the noisy sample signal through a feature extraction network to obtain sample signal features.

[0194] The attention network module 1503 is used to extract features of sample signals through the attention network to obtain multiple groups of attention sample features; the attention network includes multiple attention heads, each group of attention sample features includes a value vector, a key vector and a query vector, and each group of attention sample features shares the same value vector.

[0195] The attention network module 1503 is further used to perform correlation calculation on the query vector and the value vector through the attention network to obtain an attention prediction score.

[0196] The output network module 1504 is used to perform linear transformation based on the attention prediction score through the output network and output the noise-suppressed prediction sample signal.

[0197] The training module 1505 is configured to determine a loss function based on the difference between the predicted sample signal and the clean sample signal.

[0198] The training module 1505 is also used to jointly train the feature extraction network, the attention network and the output network based on the loss function to obtain a trained noise suppression model; the noise suppression model is used to suppress the noise of the original signal collected from the power line for data signal transmission.

[0199] The training device for the noise suppression model provided in the embodiment of the present application can deeply explore the time domain and frequency domain characteristics of the signal by utilizing the combination of the feature extraction network and the attention network, and improve the adaptability of the noise suppression model to complex noise, especially to difficult-to-handle noise types such as narrowband interference; and, through the attention mechanism of the shared value vector, it can reduce model parameters, reduce the risk of overfitting, and improve the synergy between different attention heads, thereby enhancing the adaptability of the model in different power line channel environments; furthermore, based on the difference between the predicted sample signal and the pure sample signal, the loss function is determined, and the feature extraction network, the attention network and the output network are jointly trained based on the loss function to obtain a trained noise suppression model. The model thus trained can achieve signal denoising with lower latency, meet the real-time requirements of power line carrier communication, and thus ensure the stability and reliability of data signal transmission.

[0200] The noise suppression device in the embodiment of the present application can be a computer device, or a component in the computer device, such as an integrated circuit or a chip. The computer device can be a communication device, a terminal device, or a server, etc. Exemplarily, the computer device can be a mobile phone, a tablet computer, a laptop computer, a PDA, a car computer device, a mobile Internet device (Mobile Internet Device, MID), an augmented reality (Augmented Reality, AR) / virtual reality (Virtual Reality, VR) device, a robot, a wearable device, an ultra-mobile personal computer (Ultra-mobile Personal Computer, UMPC), a netbook or a personal digital assistant (Personal Digital Assistant, PDA), etc. It can also be a server, a network attached storage (Network Attached Storage, NAS), a personal computer (Personal Computer, PC), a television (Television, TV), a teller machine or a self-service machine, etc., and the embodiment of the present application does not specifically limit it.

[0201] The noise suppression device / noise suppression model training device in the embodiments of the present application can be a device having an operating system. The operating system can be a Microsoft (Windows) operating system, an Android operating system, an iOS operating system, or other possible operating systems, which are not specifically limited in the embodiments of the present application.

[0202] The noise suppression device provided in the embodiment of the present application can achieve Figure 2 To avoid repetition, the various processes implemented in the method embodiment are not described here.

[0203] The training device of the noise suppression model provided in the embodiment of the present application can achieve Figure 12 To avoid repetition, the various processes implemented in the method embodiment are not described here.

[0204] In some embodiments, as Figure 16 As shown, an embodiment of the present application also provides a computer device 1600, including a processor 1601, a memory 1602, and a computer program stored in the memory 1602 and executable on the processor 1601. When the program is executed by the processor 1601, each process of the above-mentioned method embodiments is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be described here.

[0205] It should be noted that the computer devices in the embodiments of the present application include the mobile computer devices and non-mobile computer devices mentioned above.

[0206] An embodiment of the present application also provides a non-transitory computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the various processes of the above-mentioned noise suppression method or the noise suppression model training method embodiment, and can achieve the same technical effect. To avoid repetition, it will not be repeated here.

[0207] The processor is the processor in the computer device described in the above embodiment. The readable storage medium includes a computer readable storage medium, such as a computer read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0208] An embodiment of the present application also provides a computer program product, including a computer program, which, when executed by a processor, implements the above-mentioned noise suppression method or the above-mentioned noise suppression model training method.

[0209] The processor is the processor in the computer device described in the above embodiment. The readable storage medium includes a computer readable storage medium, such as a computer read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0210] An embodiment of the present application further provides a chip, which includes a processor and a communication interface, wherein the communication interface is coupled to the processor, and the processor is used to run programs or instructions to implement the various processes of the above-mentioned method embodiments and achieve the same technical effects. To avoid repetition, they will not be described here.

[0211] It should be understood that the chip mentioned in the embodiments of the present application can also be called a system-level chip, a system chip, a chip system or a system-on-chip chip, etc.

[0212] It should be noted that, in this article, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the statement "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device comprising the element. In addition, it should be noted that the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in the opposite order according to the functions involved. For example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0213] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application, or the part that contributes to the relevant technology, can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for enabling a terminal (which can be a mobile phone, computer, server, or network device, etc.) to execute the methods described in each embodiment of the present application.

[0214] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of this application, ordinary technicians in this field can also make many forms without departing from the purpose of this application and the scope of protection of the claims, all of which are within the protection of this application.

[0215] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples" means that the specific features, structures, materials, or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0216] Unless otherwise specified, all embodiments and optional embodiments of the present application can be combined with each other to form a new technical solution.

[0217] Unless otherwise specified, all technical features and optional technical features of this application can be combined with each other to form a new technical solution.

[0218] Unless otherwise specified, all steps of the present application may be performed sequentially or randomly, preferably sequentially. For example, the method includes steps (a) and (b), which means that the method may include steps (a) and (b) performed sequentially, or may include steps (b) and (a) performed sequentially. For example, it is mentioned that the method may also include step (c), which means that step (c) may be added to the method in any order, for example, the method may include steps (a), (b) and (c), or may include steps (a), (c) and (b), or may include steps (c), (a) and (b), etc.

[0219] The above are merely preferred embodiments of the present application and are not intended to limit the present application. Those skilled in the art will readily appreciate that various modifications and variations are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present application shall be included within the scope of protection of the present application.

Claims

1. A noise suppression method, characterized in that: The method comprises: receiving an original signal sent through a power line; the original signal being a noisy signal; Perform feature extraction based on the original signal to obtain signal features; Extracting features of the signal using multiple attention heads, obtaining multiple sets of attention features; each set of attention features includes a query vector, a key vector, and a value vector, and each set of attention features shares the same value vector; Performing a correlation calculation on the query vector and the value vector based on the multiple sets of attention features to obtain an attention score; A linear transformation is performed according to the attention score, and a noise-suppressed target signal is output.

2. The method according to claim 1, characterized in that The performing correlation calculation on the query vector and the value vector based on the multiple sets of attention features to obtain an attention score includes: For each set of attention features, performing a dot product calculation based on the query vector and the key vector included therein to obtain a dot product result corresponding to the targeted attention feature; Determine an initial relevance weight corresponding to each group of attention features based on the dot product results corresponding to each group of attention features; performing differential calculation based on the initial relevance weights corresponding to each set of attention features to obtain a target relevance weight; A weighted calculation is performed based on the target relevance weight and the shared value vector to obtain an attention score.

3. The method according to claim 1 or 2, characterized in that The extracting features based on the original signal to obtain signal features includes: Extracting features from the original signal using a global network to obtain global features of the original signal; Inputting the global features and the original signal into a hierarchical network for feature extraction to obtain hierarchical features of the original signal; The hierarchical features are used as signal features of the original signal.

4. The method according to claim 3, characterized in that After obtaining the hierarchical features of the original signal, the method further includes: fusing the global features and the hierarchical features through a fusion network to obtain a fusion feature of the original signal; The fused feature is used as the signal feature of the original signal.

5. The method according to claim 1, wherein After receiving the original signal transmitted through the power line, the method further includes: Performing time series segmentation on the original signal to obtain multiple segments of the original signal; wherein the time series segmentation adopts a sliding window technology, and the length of each sliding window is determined based on the frequency characteristics of the power line carrier signal; Perform fast Fourier transform on each segment of the original signal to obtain the frequency domain characteristics of each segment of the original signal and retain the time domain characteristics of each segment of the original signal; The frequency domain features and time domain features of each segment of the original signal are spliced together, and each segment of the spliced original signal is used for feature extraction.

6. The method according to claim 5, characterized in that After outputting the noise-suppressed target signal, the method further includes: Signal reconstruction is performed on the target signal to obtain a reconstructed target signal; the reconstructed target signal is used for data signal transmission of power line carrier communication.

7. The method according to claim 6, characterized in that The reconstructing the target signal to obtain a reconstructed target signal includes: Performing a fast Fourier transform on the target signal to obtain a frequency component of the target signal; Based on the frequency component, filtering the target signal through a preset filter to obtain a compensated target signal; Performing an inverse fast Fourier transform on the compensated target signal to obtain a reconstructed target signal.

8. A method for training a noise suppression model, characterized in that: The method comprises: Acquire a training set, wherein the training set includes noisy sample signals and corresponding clean sample signals; Extracting features from the noisy sample signal using a feature extraction network to obtain features of the sample signal; Extracting features of the sample signal using an attention network to obtain multiple sets of attention sample features; the attention network includes multiple attention heads, each set of attention sample features includes a value vector, a key vector, and a query vector, and each set of attention sample features shares the same value vector; Performing a correlation calculation on the query vector and the value vector through the attention network to obtain an attention prediction score; Performing a linear transformation based on the attention prediction score through the output network, and outputting a noise-suppressed prediction sample signal; determining a loss function based on a difference between the predicted sample signal and the clean sample signal; The feature extraction network, the attention network and the output network are jointly trained based on the loss function to obtain a trained noise suppression model; the noise suppression model is used to suppress noise on the original signal collected from the power line for data signal transmission.

9. A noise suppression device, characterized in that: The device comprises: A signal collection module, configured to receive an original signal transmitted via a power line; the original signal being a noisy signal; A feature extraction module, configured to extract features based on the original signal to obtain signal features; An attention module is configured to extract features from the signal using multiple attention heads to obtain multiple sets of attention features; each set of attention features includes a query vector, a key vector, and a value vector, and each set of attention features shares the same value vector; The attention module is further configured to perform a correlation calculation on the query vector and the value vector based on the multiple sets of attention features to obtain an attention score; The output module is used to perform a linear transformation according to the attention score and output a noise-suppressed target signal.

10. The device according to claim 9, characterized in that The attention module is also used to: For each set of attention features, performing a dot product calculation based on the query vector and the key vector included therein to obtain a dot product result corresponding to the targeted attention feature; Determine an initial relevance weight corresponding to each group of attention features based on the dot product results corresponding to each group of attention features; performing differential calculation based on the initial relevance weights corresponding to each set of attention features to obtain a target relevance weight; A weighted calculation is performed based on the target relevance weight and the shared value vector to obtain an attention score.

11. The device according to claim 9 or 10, characterized in that The attention module is also used to: Extracting features from the original signal using a global network to obtain global features of the original signal; Inputting the global features and the original signal into a hierarchical network for feature extraction to obtain hierarchical features of the original signal; The hierarchical features are used as signal features of the original signal.

12. A training device for a noise suppression model, characterized in that: The device comprises: A sample acquisition module is used to acquire a training set, wherein the training set includes noisy sample signals and corresponding clean sample signals; A feature extraction network module is used to extract features from the noisy sample signal through a feature extraction network to obtain features of the sample signal; An attention network module is configured to extract features of the sample signal through an attention network to obtain multiple sets of attention sample features; the attention network includes multiple attention heads, each set of attention sample features includes a value vector, a key vector, and a query vector, and each set of attention sample features shares the same value vector; The attention network module is further configured to perform a correlation calculation on the query vector and the value vector through the attention network to obtain an attention prediction score; An output network module, configured to perform a linear transformation based on the attention prediction score through an output network, and output a noise-suppressed prediction sample signal; a training module, configured to determine a loss function based on a difference between the predicted sample signal and the clean sample signal; The training module is also used to jointly train the feature extraction network, the attention network and the output network based on the loss function to obtain a trained noise suppression model; the noise suppression model is used to suppress noise on the original signal collected from the power line for data signal transmission.

13. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the noise suppression method according to any one of claims 1 to 7 or the noise suppression model training method according to claim 8 is implemented.

14. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the noise suppression method according to any one of claims 1 to 7 or the noise suppression model training method according to claim 8 is implemented.

15. A chip, characterized in that: The chip includes a processor and a communication interface, the communication interface is coupled to the processor, and the processor is used to run a program or instruction to implement the noise suppression method according to any one of claims 1 to 7, or the noise suppression model training method according to claim 8.

Images