Method and apparatus for signal processing

Abstract

The application relates to the technical field of artificial intelligence and a signal processing method in power line communication, which comprises the following steps: acquiring a signal segment containing impulse noise, the signal segment comprising at least one sampling point; inputting the signal segment into a neural network model to obtain a noise reduction parameter; and performing noise reduction processing on the at least one sampling point according to the noise reduction parameter. When the neural network model is a local neural network model, the method further comprises the following steps: sending the signal segment to a cloud device; the signal segment is used for training the neural network model by the cloud device, the input of the neural network model is the signal segment, and the output of the neural network model is the noise reduction parameter; receiving the parameters of the neural network model trained by the cloud device; and constructing the local neural network model according to the parameters of the neural network model. Through the combination of a terminal and a cloud, a terminal with limited computing power can quickly construct a neural network model, and the noise reduction parameter required for noise reduction can be obtained according to the model, so that the impulse noise in power line communication can be reduced according to the parameter.

Description

Technical Field

[0001] This invention relates to the field of communication technology, and in particular to methods and apparatuses for signal processing in power line communication, computing devices and computer-readable storage media. Background Technology

[0002] As people's living standards improve, their demands for communication latency, speed, quality, and stability are also increasing. However, due to the limited available spectrum, the capacity of mobile networks cannot be increased indefinitely. Some scholars have proposed using power lines for communication, i.e., Power Line Communication (PLC). Power lines are widely distributed in cities of all sizes, and if fully utilized, PLC has a promising future. Furthermore, PLC operates independently of other communication methods and will not degrade the communication quality of wireless or wired communications.

[0003] Compared to traditional wireless communication, PLC communication environments are extremely harsh, including not only additive white Gaussian noise (AWGN) but also colored background noise, power frequency noise, and periodic impulse noise. These noises severely reduce PLC communication transmission speed, increase bit error rate, and significantly impact communication quality. Among these, impulse noise has the highest power and the greatest impact on communication. Common methods for impulse noise reduction all require setting parameters, such as thresholds for the top-cutting and zeroing methods. However, obtaining the optimal parameters for noise reduction algorithms is often difficult for PLC systems. Due to cost considerations, PLC transmitters and receivers often have limited computing power, and computationally intensive noise reduction algorithms such as brute-force search often require excessive time, failing to meet the requirements of real-time noise reduction.

[0004] In summary, in order to improve the quality and speed of communication, there is a need for a method that can quickly obtain the noise reduction parameters required for noise reduction, and suppress impulse noise in power line communication channels based on these parameters. Summary of the Invention

[0005] In view of the above-mentioned problems of the prior art, this application provides a signal processing method and apparatus, computing device and computer-readable storage medium, which can realize the rapid acquisition of noise reduction parameters required for noise reduction in power line communication equipment, so as to suppress impulse noise of power line communication channel based on these parameters.

[0006] To achieve the above objectives, the first aspect of this application provides a signal processing method, comprising:

[0007] Acquire a signal segment containing impulse noise, the signal segment including at least one sampling point;

[0008] The signal segment is input into the neural network model to obtain the noise reduction parameters;

[0009] Noise reduction is applied to at least one sampling point based on the noise reduction parameters.

[0010] Therefore, the method of this application is applicable to power line communication. In practical applications, even in power line communication devices with limited computing power, the required noise reduction parameters can be quickly solved through the neural network, thereby achieving real-time noise reduction processing for power line communication.

[0011] As one possible implementation of the first aspect, when the neural network model is a local neural network model, the method also includes:

[0012] Send signal segments to cloud devices; the signal segments are used by the cloud devices to train neural network models, and the neural network models take the signal segments as input and output the neural network models with noise reduction parameters as output.

[0013] Receive parameters from the neural network model trained by the cloud device;

[0014] Construct a local neural network model based on the parameters of the neural network model.

[0015] The above framework combines cloud and power line communication (PLC) terminal devices. The cloud device rapidly trains a neural network model, and the trained model parameters are provided to the terminal. The terminal can then quickly reconstruct the neural network model based on these parameters and use it to obtain noise reduction parameters for high-speed PLC noise reduction. This combination of cloud and terminal effectively addresses the limitation of terminal computing power and allows for further improvement in noise reduction performance.

[0016] As one possible implementation of the first aspect, the signal segment is a downsampled signal segment;

[0017] The downsampled signal segment retains the peak points of the impulse noise.

[0018] Optionally, downsampling can be performed on the acquired signal segments by the terminal of the power line communication equipment, or downsampling can be performed on the signal segments transmitted by the terminal by the cloud device. The former reduces the amount of data transmitted between the terminal and the cloud device. Therefore, downsampling can improve the training speed of the neural network model in the cloud and also reduce the network size, making it more suitable for power line communication equipment. Alternatively, it can reduce the amount of data input to the neural network model, allowing for the use of neural network models with more layers, resulting in a better-performing trained neural network model.

[0019] As one possible implementation of the first aspect, downsampling includes using a local maximum downsampling algorithm, including:

[0020] Acquire the sub-segments included in the signal segment;

[0021] Obtain the sampling point with the maximum signal amplitude in each sub-segment, and the sampling points of each sub-segment constitute the downsampled signal segment.

[0022] Therefore, the downsampling method described above is suitable for applications such as... Figure 6B The impulse noise interference shown is a case of point-like impulse noise, meaning each impulse only affects one sampling point of the signal segment. The downsampling process needs to retain all sampling points affected by the point-like impulses. By considering the characteristics of impulse noise and employing a corresponding downsampling method, more of the effective signal for neural network training can be preserved.

[0023] As one possible implementation of the first aspect, downsampling includes using a global maximum windowing downsampling algorithm, including:

[0024] Obtain the sampling point in the signal segment where the signal amplitude is at its maximum value;

[0025] The signal segment contains multiple consecutive sampling points with the maximum signal amplitude. These multiple consecutive sampling points constitute the downsampled signal.

[0026] Therefore, the downsampling method described above is suitable for applications such as... Figure 6C The impulse noise interference shown is clustered impulse noise, meaning that each impulse affects multiple sampling points of the signal segment, and each impulse has a "tail" portion. The goal is to retain as many sampling points affected by the clustered impulses as possible during the downsampling process. By considering the characteristics of impulse noise and employing a corresponding downsampling method, more effective signal for neural network training can be preserved.

[0027] As one possible implementation of the first aspect, the sending includes:

[0028] Buffer each signal segment;

[0029] Once a certain time point is reached, the cached signal fragments are sent to the cloud device.

[0030] Therefore, by setting specific time points, especially time points that do not affect the user's internet access, such as between 2:00 AM and 4:00 AM when the user is asleep, the data transmission of signal segments can minimize interference with the user's internet access.

[0031] As one possible implementation of the first aspect, the noise reduction parameters include a first noise reduction parameter and / or a second noise reduction parameter;

[0032] The first noise reduction parameter is used as the first threshold value to perform noise reduction processing on signals whose amplitude exceeds the first threshold value.

[0033] The second noise reduction parameter is used as a coefficient, and its product with the first noise reduction parameter is used as the second threshold value. This is used to perform noise reduction processing on signals whose amplitude exceeds the second threshold value, setting them to zero. The second noise reduction parameter is greater than 1.

[0034] As shown above, the noise reduction parameters can be quickly obtained through this application, and can be applied to power line communication impulse noise suppression using the three-segment method.

[0035] To achieve the above objectives, a second aspect of this application provides a signal processing method, comprising:

[0036] Receive a signal segment; the signal segment contains impulse noise and includes at least one sampling point;

[0037] A neural network model is trained using signal segments as input and output as noise reduction parameters.

[0038] Send the parameters of the trained neural network model to the terminal device.

[0039] The above framework combines cloud and power line communication (PLC) terminal devices. The cloud device rapidly trains a neural network model, and the trained model parameters are provided to the terminal. The terminal can then quickly reconstruct the neural network model based on these parameters and use it to obtain noise reduction parameters for high-speed PLC noise reduction. This combination of cloud and terminal effectively addresses the limitation of terminal computing power and allows for further improvement in noise reduction performance.

[0040] This application allows for the rapid acquisition of the aforementioned noise reduction parameters, which can then be applied to power line communication impulse noise suppression using the three-segment method.

[0041] As one possible implementation of the second aspect, the training includes a supervised training method with mean squared error as the loss function, which is: the square of the difference between the optimal first denoising parameter and the predicted first denoising parameter, plus the square of the difference between the optimal second denoising parameter and the predicted second denoising parameter.

[0042] The signal segment has optimal first noise reduction parameters and optimal second noise reduction parameters;

[0043] The first and second noise reduction parameters are predicted by inputting the signal segment into the neural network model.

[0044] As shown above, the supervised training method described above is suitable for training neural networks with known optimal denoising parameters. The training method can be flexibly selected according to the actual situation.

[0045] As one possible implementation of the second aspect, the training includes an unsupervised training method that uses the reciprocal of the signal-to-interference-plus-noise ratio (SINR) as the loss function. The loss function is the reciprocal of the denoised signal-to-interference-plus-noise ratio (SINR) calculated from the denoised signal segment obtained after denoising based on the predicted first denoising parameter and the predicted second denoising parameter.

[0046] The first and second noise reduction parameters are predicted by inputting the signal segment into the neural network model.

[0047] As shown above, the unsupervised training method described above is suitable for training neural networks with unknown optimal denoising parameters. The training method can be flexibly selected according to the actual situation.

[0048] To achieve the above objectives, a third aspect of this application provides a signal processing method, comprising:

[0049] The signal processing method of any of the first aspects mentioned above, and

[0050] The signal processing method of any of the second aspects mentioned above.

[0051] To achieve the above objectives, a fourth aspect of this application provides a signal processing apparatus, comprising:

[0052] A signal segment acquisition unit is used to acquire a signal segment containing impulse noise, wherein the signal segment includes at least one sampling point;

[0053] The noise reduction parameter acquisition unit is used to input signal segments into the neural network model to obtain noise reduction parameters.

[0054] The noise reduction processing unit is used to perform noise reduction processing on at least one sampling point according to the noise reduction parameters.

[0055] As a possible implementation of the fourth aspect, when the neural network model is a local neural network model, the device further includes:

[0056] The transmitting unit is used to send signal segments to the cloud device; the signal segments are used by the cloud device to train the neural network model, and the input of the neural network model is the signal segments, and the output is the neural network model with noise reduction parameters;

[0057] The receiving unit is used to receive the parameters of the neural network model trained by the cloud device;

[0058] The neural network model building unit is used to reconstruct a local neural network model based on the parameters of the neural network model.

[0059] As one possible implementation of the fourth aspect, the signal segment is a downsampled signal segment;

[0060] The downsampled signal segment retains the peak points of the impulse noise.

[0061] As one possible implementation of the fourth aspect, downsampling includes using a local maximum downsampling algorithm, including:

[0062] Acquire the sub-segments included in the signal segment;

[0063] Obtain the sampling point with the maximum signal amplitude in each sub-segment, and the sampling points of each sub-segment constitute the downsampled signal segment.

[0064] As one possible implementation of the fourth aspect, downsampling includes using a global maximum windowing downsampling algorithm, including:

[0065] Obtain the sampling point in the signal segment where the signal amplitude is at its maximum value;

[0066] The signal segment contains multiple consecutive sampling points with the maximum signal amplitude. These multiple consecutive sampling points constitute the downsampled signal.

[0067] As one possible implementation of the fourth aspect, the transmission includes:

[0068] Buffer each signal segment;

[0069] Once a certain time point is reached, the cached signal fragments are sent to the cloud device.

[0070] As one possible implementation of the fourth aspect, the noise reduction parameters include a first noise reduction parameter and a second noise reduction parameter;

[0071] The first noise reduction parameter is used as the first threshold value to perform noise reduction processing on signals whose amplitude exceeds the first threshold value.

[0072] The second noise reduction parameter is used as a coefficient, and its product with the first noise reduction parameter is used as the second threshold value. This is used to perform noise reduction processing on signals whose amplitude exceeds the second threshold value, setting them to zero. The second noise reduction parameter is greater than 1.

[0073] To achieve the above objectives, a fifth aspect of this application provides a signal processing apparatus, comprising:

[0074] A receiving unit is used to receive a signal segment; the signal segment contains impulse noise and includes at least one sampling point;

[0075] The neural network model training unit is used to train a neural network model using signal segments. The input of the neural network model is the signal segment, and the output is the neural network model with noise reduction parameters.

[0076] The transmitting unit is used to send the parameters of the trained neural network model to the terminal device.

[0077] As one possible implementation of the fifth aspect, the training includes a supervised training method with mean squared error as the loss function, which is: the square of the difference between the optimal first denoising parameter and the predicted first denoising parameter, plus the square of the difference between the optimal second denoising parameter and the predicted second denoising parameter.

[0078] The signal segment has optimal first noise reduction parameters and optimal second noise reduction parameters;

[0079] The first and second noise reduction parameters are predicted by inputting the signal segment into the neural network model.

[0080] As one possible implementation of the fifth aspect, the training includes an unsupervised training method that uses the reciprocal of the signal-to-interference-plus-noise ratio (SINR) as the loss function. The loss function is the reciprocal of the denoised signal-to-interference-plus-noise ratio (SINR) calculated from the denoised signal segment obtained after denoising based on the predicted first denoising parameter and the predicted second denoising parameter.

[0081] The first and second noise reduction parameters are predicted by inputting the signal segment into the neural network model.

[0082] To achieve the above objectives, a sixth aspect of this application provides a signal processing apparatus, comprising:

[0083] The signal processing apparatus of any of the fourth aspects mentioned above, and the signal processing apparatus of any of the fifth aspects mentioned above.

[0084] To achieve the above objectives, a seventh aspect of this application provides a computing device, including...

[0085] Communication interface;

[0086] At least one processor connected to a communication interface; and

[0087] At least one memory, connected to a processor and storing program instructions, which, when executed by at least one processor, cause at least one processor to perform any of the signal processing methods described in the first aspect, or the program instructions, when executed by at least one processor, cause at least one processor to perform any of the signal processing methods described in the second aspect.

[0088] To achieve the above objectives, the eighth aspect of this application provides a computer-readable storage medium having program instructions stored thereon, wherein when executed by a computer, the program instructions cause the computer to perform any of the signal processing methods described in the first aspect, or when executed by a computer, the program instructions cause the computer to perform any of the signal processing methods described in the second aspect.

[0089] These and other aspects of the invention will become more apparent from the following description of several embodiments. Attached Figure Description

[0090] The various features of the present invention and the relationships between them are further explained below with reference to the accompanying drawings. The drawings are exemplary; some features are not shown to scale, and some drawings may omit conventional features in the field of this application that are not essential to this application, or additional features that are not essential to this application may be shown. The combination of features shown in the drawings is not intended to limit the present application. Furthermore, throughout this specification, the same reference numerals refer to the same things. Specific descriptions of the drawings are as follows:

[0091] Figure 1 A schematic diagram of noise suppression using the three-stage method;

[0092] Figure 2A This is a schematic diagram illustrating the principle of the noise suppression method for power line communication in this application;

[0093] Figure 2B This is a schematic diagram of an embodiment of a terminal of the power line communication equipment of this application;

[0094] Figure 3A This is a flowchart of a first embodiment of the signal processing method of this application;

[0095] Figure 3B A flowchart of another embodiment of the signal processing method of this application;

[0096] Figure 4 This is a flowchart of a second embodiment of the signal processing method of this application;

[0097] Figure 5 This is a flowchart of a first specific embodiment of the signal processing method of this application;

[0098] Figure 6A This is a schematic diagram of an impulse noise signal;

[0099] Figure 6B This is a schematic diagram showing that the impulse noise interference is a point-like impulse noise.

[0100] Figure 6C This is a schematic diagram illustrating that impulse noise interference is a cluster of impulse noise.

[0101] Figure 7A This is a schematic diagram of a convolutional neural network model structure that can be used in this application;

[0102] Figure 7B This is a schematic diagram of another convolutional neural network model structure that can be used in this application;

[0103] Figure 8A This is a schematic diagram illustrating the supervised training method for the neural network model in this application;

[0104] Figure 8B This is a schematic diagram illustrating the unsupervised training method for the neural network model in this application;

[0105] Figure 9A This is a schematic diagram of a first embodiment of the signal processing apparatus of this application;

[0106] Figure 9B This is a schematic diagram of a second embodiment of the signal processing apparatus of this application;

[0107] Figure 10 This is a schematic diagram of the computing device used in this application. Detailed Implementation

[0108] The terms "first, second, third, etc." or similar terms such as module A, module B, module C, etc., used in the specification and claims are only used to distinguish similar objects and do not represent a specific ordering of objects. It is understood that a specific order or sequence may be interchanged where permitted so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0109] In the following description, the labels of the steps, such as S110, S120, etc., do not necessarily mean that the steps will be executed in this way. The order of the steps can be interchanged or executed simultaneously if permitted.

[0110] The term "comprising" as used in the specification and claims should not be construed as limiting itself to what follows; it does not exclude other elements or steps. Therefore, it should be interpreted as specifying the presence of the mentioned feature, integral, step, or component, but does not exclude the presence or addition of one or more other features, integrals, steps, or components, or groups thereof. Thus, the statement "device comprising means A and B" should not be limited to a device consisting solely of components A and B.

[0111] The terms "some embodiments" or "embodiments" used in this specification mean that a particular feature, structure, or characteristic described in conjunction with that embodiment is included in at least some embodiments of the invention. Therefore, the terms "in some embodiments" or "in an embodiment" appearing throughout this specification do not necessarily refer to the same embodiment, but may refer to the same embodiment. Furthermore, in one or more embodiments, particular features, structures, or characteristics can be combined in any suitable manner, as will be apparent to those skilled in the art from this disclosure.

[0112] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. In case of any inconsistency, the meaning set forth in this specification or derived from the content described herein shall prevail. Furthermore, the terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application. To accurately describe the technical content of this application and to accurately understand the invention, the following explanations or definitions of the terms used in this specification are provided before describing specific embodiments:

[0113] 1) Power Line Communication (PLC): This refers to a communication method that uses power lines to transmit data and media signals. The principle is that the transmitting end loads a high-frequency signal carrying information onto an electric current and then transmits it through the power line; the receiving end then separates the high-frequency signal from the current to achieve information transmission.

[0114] 2) Orthogonal Frequency Division Multiplexing (OFDM): A multi-carrier modulation method. Its main idea is to divide the channel into several orthogonal sub-channels, converting high-speed data signals into parallel low-speed sub-data streams, and modulating them onto each sub-channel for transmission. For example, 2048 serially transmitted signals can be converted into 2048 parallel signals through OFDM modulation and transmitted in parallel on their respective sub-channels.

[0115] 3) Supervised learning: One of the training methods for neural network models. In supervised learning, the sample data includes labels. Through supervised learning, a function or model that can map the sample data to the labels can be learned.

[0116] 4) Unsupervised learning: One of the training methods for neural network models. In unsupervised learning, the sample data does not include any labels. Through unsupervised learning, effective features, categories, structures, or probability distributions hidden in the data can be discovered.

[0117] For the noise suppression scheme of power line communication, one technique is to suppress the impulse noise of power line communication by the three-segment method. Specifically: two thresholds are set for the time-domain signal: the first threshold T, and the second threshold aT formed by the product of the coefficient a and the first threshold, where the coefficient a > 1. See Figure 1 the schematic diagram of impulse noise suppression by the three-segment method shown in. When the signal amplitude exceeds the first threshold T, the signal is clipped; when the signal exceeds the second threshold aT, the signal is directly set to zero. By these two thresholds, the signals with amplitudes exceeding the thresholds are clipped or set to zero, so as to achieve the effect of suppressing impulse noise. Described by the formula as follows:

[0118]

[0119] where x(t) represents the signal of power line communication, sgn is the step function, and when T ≤ |x(t)| < aT, the value of sgn(x(t))T is T. The key to noise suppression by this three-segment method is the optimal design of two noise reduction parameters, such as the first threshold T and the second threshold aT (or rather the first threshold T and the coefficient a). If the signal distribution or noise distribution is unknown, these two thresholds need to be determined by experimental or simulation methods, usually using the exhaustive search method to determine the thresholds. Given the limited computing power of power line communication devices, the time required for exhaustive search is too long, and it often fails to meet the requirements of real-time noise reduction.

[0120] This application provides another signal processing method applicable to noise suppression in power line communication, which uses a neural network model to quickly obtain the noise reduction parameters of the three-segment method, and suppresses the power line communication noise through these noise reduction parameters, especially the suppression of impulse noise. As Figure 2A shown, the basic principle of this method is: the terminal of the power line communication device receives the power line communication signal, performs signal sampling, and sends it to the cloud device. The cloud device trains a neural network model according to the sampling data, and sends the parameters of the trained neural network model to the terminal. The terminal reconstructs or updates the local neural network model according to these parameters, and then predicts the noise reduction parameters for the received power line communication signal according to the local neural network model, and uses the predicted noise reduction parameters for noise reduction processing. Since the training of the neural network model is carried out on the cloud device with stronger computing power, and the power line communication terminal receives the parameters of the trained neural network model, the problem of training the neural network model in the case of limited computing power of the power line communication terminal is solved.

[0121] The application scenario of this application can be applied to power line communication devices, for example, routers based on power line communication, home gateways, modems (or power line network adapters), switches, etc. As Figure 2BThe diagram illustrates a terminal embodiment of a power line communication (PLC) device 100. This PLC device 100 may include at least one or more transceivers 101, one or more processors 102, and one or more memories 103. The memories 103 are used to store instructions. The processors 102 can call the instructions in the memories 103 to enable the PLC device to perform noise reduction methods, etc. The processors 102, transceivers 101, and memories 103 are electrically connected, for example, via a bus, to facilitate data exchange. The transceivers 101, under the control of the processors 102, implement PLC communication, which includes communication with other devices and communication with cloud devices via a PLC device (such as a modem) that can access the Internet.

[0122] [First Implementation of a Signal Processing Method]

[0123] The first embodiment of the signal processing method of this application is described below. Specifically, the signal processing method in this first embodiment is used to suppress impulse noise in power line communication, and this embodiment is applied to the terminal side of a power line communication device. (Refer to...) Figure 3A The flowchart shown includes the following steps:

[0124] S110: The terminal of the power line communication device acquires a signal segment containing impulse noise, the signal segment including at least one sampling point, the number of sampling points being denoted as N0, where N0 is an integer greater than or equal to 1.

[0125] In this embodiment, the terminal of the power line communication device is the signal receiver. Another power line communication device, acting as the signal transmitter, sends a pulse signal containing a known sequence (i.e., the encoded sequence to be transmitted). This signal is irradiated after transmission through the channel. The power line communication device at the receiver acquires this signal and treats it as a signal to be processed. Figure 6A This is a schematic diagram of an impulse noise signal, such as... Figure 6A As shown, when a pulse arrives, the signal will be disturbed for a short period of time. The additional signal corresponding to this part is the pulse noise. When there is no additional pulse noise, the signal is relatively stable, that is, when there is no pulse noise, noise reduction is not necessary.

[0126] In some embodiments, the signal to be processed may include n sampling points, where n is an integer greater than or equal to N0. The terminal samples the signal to be processed at a fixed sampling frequency to obtain a signal segment containing N0 sampling points.

[0127] There are several ways to select N0, which can be related to the length of the signal that needs to be processed in a single time slot of the communication system. For example, N0 can be selected as 2048, which is the length of a single Orthogonal Frequency Division Multiplexing (OFDM) time-domain signal. It is easy to understand that other values ​​can also be selected.

[0128] S120: The terminal of the power line communication equipment inputs the signal segment into its local neural network model to obtain noise reduction parameters.

[0129] The neural network model is a pre-trained neural network model. The training process of the neural network model can be carried out by a third device with strong computing power, such as a cloud device. The parameters of the trained neural network model are then provided to the terminal of the power line communication device, so that the terminal can use these network parameters to reconstruct the neural network model.

[0130] In some embodiments, the neural network model may be a multi-layer fully connected neural network model, a convolutional neural network model, a recurrent neural network model, etc.

[0131] S130: The terminal of the power line communication equipment performs noise reduction processing on the at least one sampling point according to the noise reduction parameters.

[0132] In some embodiments, such as Figure 3B As shown, when the neural network model is a local neural network model on the terminal side, before step S120, the method further includes:

[0133] S112: The terminal of the power line communication device sends the signal segment to the cloud device; the signal segment is used by the cloud device to train a neural network model, and the input of the neural network model is the signal segment, and the output is the neural network model with the noise reduction parameters.

[0134] S114: The terminal of the power line communication equipment receives the parameters of the neural network model trained by the cloud device.

[0135] S116: The terminal of the power line communication device reconstructs a local neural network model based on the parameters of the neural network model.

[0136] The above is an example of training the neural network model using a cloud-based device with strong computing power. In this way, for power line communication devices with limited computing power in practical applications, the cloud-based device can quickly train the neural network model and provide the trained neural network model parameters to the terminal. The terminal can then quickly reconstruct the neural network model based on these parameters and use it to obtain noise reduction parameters for high-speed power line communication. By combining the cloud and the terminal, the problem of limited terminal computing power can be effectively solved, and there is room for improvement in noise reduction performance.

[0137] In some embodiments, the signal segment may be a downsampled signal segment; the downsampled signal segment retains the peak points of the impulse noise. Here, the number of sampling points contained in the downsampled signal segment can be denoted as N1, where N1 is an integer greater than 1 and less than N0. The value of N1 can be matched to the number of inputs to the neural network model; for example, when N0 is selected as 2048, N1 can be 256, where 256 is the number of input data points to the neural network model.

[0138] By downsampling the signal, the length of each signal segment is reduced. Therefore, given the limited size and computing power of the neural network model, the amount of data input to the model can be reduced. This allows for the use of neural network models with more layers, resulting in better performance after training. Downsampling can be performed by the terminal of the power line communication equipment or by the cloud device on the signal segments transmitted from the terminal. The former reduces the amount of data transmitted between the terminal and the cloud device.

[0139] In some embodiments, the downsampling includes downsampling using a local maximum downsampling algorithm, which is suitable for applications such as... Figure 6B The pulse noise interference shown is a case of point-like pulse noise, meaning each pulse only affects one sampling point of the signal segment. The downsampling process needs to retain all sampling points affected by the point-like pulses. This downsampling step includes:

[0140] First, the sub-segments comprising the signal segment are obtained; then, a sampling point with the maximum signal amplitude in each sub-segment is obtained, and the sampling points of each sub-segment constitute the downsampled signal segment. This process will be described in more detail in the following specific embodiments.

[0141] In some embodiments, the downsampling includes downsampling using a global maximum windowing downsampling algorithm, which is suitable for applications such as... Figure 6CThe impulse noise interference shown is clustered impulse noise, meaning each pulse affects multiple sampling points of the signal segment, and each pulse has a "tail" portion. The goal is to preserve as many sampling points affected by the clustered impulses as possible during the downsampling process. This downsampling step includes:

[0142] First, the sampling point with the maximum signal amplitude in the signal segment is obtained; then, multiple consecutive sampling points containing the sampling point with the maximum signal amplitude in the signal segment are obtained, and these multiple consecutive sampling points constitute the downsampled signal. This process will be described in more detail in the following specific embodiments.

[0143] In some embodiments, the sending in step S112 above includes:

[0144] First, during operation, the power line communication equipment terminal buffers various signal segments, including sampled signals over a period of time, such as several signal segments containing impulse noise or downsampled signal segments. Then, at a certain time point, or after the buffer exceeds a certain capacity, the power line terminal equipment packages these segments into data packets and sends them to the cloud device. The size of the buffered data packets varies depending on the terminal's storage capacity.

[0145] The time point can be a time that does not affect the user's internet access, such as the time between 2:00 AM and 4:00 AM when the user is asleep.

[0146] In some embodiments, when using the three-segment method for power line communication impulse noise suppression, the noise reduction parameters include a first noise reduction parameter T and a second noise reduction parameter a; the first noise reduction parameter T serves as a first threshold value, used for clipping noise reduction processing of signals whose signal amplitude exceeds the first threshold; the second noise reduction parameter a serves as a coefficient, and its product with the first noise reduction parameter serves as a second threshold value, used for zeroing noise reduction processing of signals whose signal amplitude exceeds the second threshold, wherein the second noise reduction parameter a is greater than 1.

[0147] In other words, the above-mentioned noise reduction parameters can be obtained quickly through this application, and can be applied to power line communication impulse noise suppression using the three-segment method.

[0148] [Second Embodiment of Signal Processing Method]

[0149] The second embodiment of the signal processing method of this application is described below. Specifically, the signal processing method in this second embodiment is used to suppress impulse noise in power line communication, and this embodiment is applied to the cloud device side of the power line communication equipment. (Refer to...) Figure 4 The flowchart shown includes the following steps:

[0150] S210: The cloud device receives a signal segment transmitted by the terminal of the power line communication device; the signal segment contains impulse noise and includes at least one sampling point.

[0151] Cloud devices can receive sampled signal data packets from multiple terminals, forming a large database for use in training the neural network described later.

[0152] S220: The cloud device uses the downsampled signal segment as a dataset to train a neural network model, wherein the input of the neural network model is the signal segment and the output is a neural network model with noise reduction parameters.

[0153] In some embodiments, the noise reduction parameters include the first noise reduction parameter T and the second noise reduction parameter a.

[0154] In some embodiments, after a certain period of time or after collecting a certain amount of data packets, the cloud device can use new data to further train the neural network model in order to update the neural network model.

[0155] S230: The cloud device sends the parameters of the trained neural network model to the terminal device.

[0156] In some embodiments, the neural network model includes the number of layers, structural information of each layer, and weight information of each layer. The weight information of each layer includes, for example, the weights of each layer in a fully connected neural network model, and the convolution kernels of a convolutional network.

[0157] In some embodiments, the training includes a supervised training method using mean squared error as a loss function, one embodiment of which includes:

[0158] If the signal segment has the optimal first noise reduction parameter and the optimal second noise reduction parameter;

[0159] If the signal segment is input into the neural network model, the model outputs a first predicted noise reduction parameter and a second predicted noise reduction parameter.

[0160] The loss function is then calculated as the square of the difference between the optimal first denoising parameter and the predicted first denoising parameter, plus the square of the difference between the optimal second denoising parameter and the predicted second denoising parameter.

[0161] In some embodiments, the training includes an unsupervised training method using the inverse of the signal-to-interference-plus-noise ratio (SINR) as a loss function, one embodiment of which includes:

[0162] If the signal segment is input into the neural network model, the model outputs a first predicted noise reduction parameter and a second predicted noise reduction parameter.

[0163] If the noise reduction process is performed according to the predicted first noise reduction parameter and the predicted second noise reduction parameter, and a noise-reduced signal segment is obtained, the noise reduction signal segment and the original signal segment are used to calculate the noise reduction signal-to-interference-plus-noise ratio.

[0164] The loss function is then the reciprocal of the signal-to-interference-plus-noise ratio after denoising.

[0165] [First Specific Implementation of the Signal Processing Method]

[0166] The following description, in order to better understand this application, will use a specific embodiment of the signal noise reduction processing method applied to power line communication as an example. (Refer to...) Figure 5 The flowchart of this specific embodiment shown includes the following steps:

[0167] S310: The terminal of the power line communication device acquires a signal segment containing impulse noise, the signal segment comprising N0 sampling points, where N0 is an integer greater than or equal to 1.

[0168] In this embodiment, the power line communication device terminal is the signal receiving end, which samples the received signal to be processed at a fixed sampling rate to obtain a signal segment containing N0 sampling points. The signal to be processed includes n sampling points, where n is an integer greater than or equal to N0. In this specific embodiment, N0 is selected as 2048, which is the length of a single orthogonal frequency division multiplexing time-domain signal.

[0169] S320: The terminal of the power line communication equipment processes the signal segment using a downsampling algorithm to obtain a downsampled signal segment containing impulse noise. The downsampled signal segment includes N1 sampling points, where N1 is an integer greater than 1 and less than N0.

[0170] Let's continue using OFDM time-domain signals as an example. Each OFDM symbol has a length of 2048 sampling points, which is reduced to 256 sampling points after downsampling. The reason for using downsampling is that when using the received signal as input to a neural network model, considering the long OFDM data (2048 data points), using all 2048 data points as input might result in a shallow neural network model with poor performance. By using only a portion of the data, i.e., the downsampled data, the input data to the neural network model is reduced to 256 points. This allows for an increase in the number of layers in the neural network model, enabling the use of a more layered model and thus achieving better performance.

[0171] In this embodiment, the power line communication device downsamples the signal segment containing impulse noise. Optionally, the downsampling algorithm includes a local maximum downsampling algorithm and a global maximum windowing downsampling algorithm. The downsampling algorithm can be programmed and embedded into the terminal of the power line communication device.

[0172] In some embodiments, after downsampling, in order to better preserve the information of the original signal, a downsampling algorithm can be selected based on the type of impulse noise. For example, when the impulse noise interference is point-like impulse noise, a local maximum downsampling algorithm can be used; when the impulse noise interference is cluster-like impulse noise, a global maximum windowing downsampling algorithm can be used. See below. Figures 6A-6C Furthermore, the downsampling algorithm described above is introduced in conjunction with the analysis of noise.

[0173] like Figure 6A This is a schematic diagram of an impulse noise signal, such as... Figure 6A As shown, when a pulse arrives, the signal is disturbed for a short duration, while the signal is relatively stable without additional pulse noise; that is, noise reduction is not necessary when there is no pulse noise. However, for additional pulse noise, including... Figure 6B The pulse noise interference shown is point-like pulse noise, or Figure 6C The impulse noise interference shown is clustered impulse noise. The specific noise reduction methods for these two types of noise are as follows:

[0174] 1) When the impulse noise interference is as follows Figure 6B The illustrated point-like pulse noise, meaning each pulse only affects one sampling point of the signal segment, necessitates that the downsampling process retain all sampling points affected by the point-like pulses. For example... Figure 6B As shown in the figure, in the signal segment interfered with by impulse noise, the positions where the point pulses appear are the sampling points circled in the figure. Therefore, during the downsampling process, it is necessary to retain these sampling points circled in the figure.

[0175] For point-like impulse noise, the downsampling algorithm used can be the local maximum downsampling algorithm. Specifically, if the downsampling factor is k = N0 / N1, then sub-segments of length k are extracted sequentially from the signal segment, containing a total of N1 sub-segments. For each sub-segment of length k, only one sampling point with the maximum amplitude among the k sampling points is retained, resulting in a total of N1 local maximum sampling points. The N1 local maximum sampling points constitute the downsampled signal of the original signal segment.

[0176] For example, when N0 is 2048 and N1 is 256, the calculated k is 8 (i.e., each sub-segment contains 8 sampling points). Then, for each of the 256 sub-segments, for every sub-segment of length 8, the sampling point with the largest amplitude among these 8 sampling points is retained, thus ultimately obtaining 256 sampling points. This example is described using the following logical algorithm:

[0177] 1: Input s2(t), t=1,2,…,2048

[0178] 2:n=0,x(t)=0,t=1,2,…,256

[0179] 3: Start the loop, termination condition: n = 256

[0180] 4: S = {y: 8n ≤ y < 8(n+1), y ∈ N}

[0181] 5: i = argmax t∈S |s2(t)|

[0182] 6: x(n) = s2(i)

[0183] 7: n = n + 1

[0184] 8: Output x(t)

[0185] 2) When the impulse noise interference is as follows Figure 6C The illustrated clustered pulse noise, where each pulse affects multiple sampling points of a signal segment and each pulse has a "tail," necessitates that the downsampling process retain as many sampling points affected by the clustered pulses as possible. For example... Figure 6C As shown, in a signal segment interfered with by impulse noise, the cluster pulses mainly appear at the multiple sampling points circled in the figure. Therefore, during the downsampling process, it is necessary to retain the multiple sampling points shown in the circle.

[0186] For clustered impulse noise, the downsampling algorithm used can be a global maximum windowing downsampling algorithm, specifically: obtaining the sampling point with the maximum amplitude of the signal segment, denoted as M; if the length of the downsampled signal segment is N1, then retaining N1 consecutive sampling points in the signal segment that contain the sampling point M; the method for obtaining these N1 consecutive sampling points is to take a sampling points before sampling point M and b sampling points after sampling point M based on empirical knowledge of clustered impulse noise, satisfying a+b+1=N1, and these N1 consecutive sampling points should include the sampling points mainly affected by the impulse noise; these N1 consecutive sampling points constitute the downsampled signal.

[0187] For example, when N0 is 2048 and N1 is 256, assuming the maximum amplitude sampling point M is the k-th sampling point, and a is 10 and b is 254, then for these 2048 sampling points, a sub-segment of length 256 can be obtained, and this sub-segment contains the maximum sampling point, as well as its preceding 10 and following 254 sampling points. This example is described by the following logical algorithm:

[0188] 1: Input s2(t), t=1,2,…,2048

[0189] 2:n=0,x(t)=0,t=1,2,…,256

[0190] 3:k = argmax t |s2(t)|

[0191] 4:t min =max{1,k-10}

[0192] 5:t max =min{k+245,2048}

[0193] 6: x(t) = s2(t) min :t max )

[0194] 7: Output x(t)

[0195] S330: The terminal of the power line device sends the downsampled signal segment to the cloud device.

[0196] During operation, the terminal device buffers sampled signals for a period of time, including several downsampled signal segments containing impulse noise. When a certain time point is reached, or when the buffer exceeds a certain capacity, the power line terminal device packages these segments into data packets and sends them to the cloud device. The size of the buffered data packets varies depending on the terminal's storage capacity.

[0197] S340: The cloud device receives the plurality of signal segments, and the cloud device uses the downsampled signal segments as a dataset to train a neural network model, wherein the output of the neural network model is the first noise reduction parameter and the second noise reduction parameter of the downsampled signal segments.

[0198] If the cloud device has already trained the neural network model, it can be further trained using a certain amount of new data collected at regular intervals to update the neural network model.

[0199] In some embodiments, when using a convolutional neural network model, networks with different parameters can be employed. For example... Figure 7AOne possible convolutional neural network model structure is shown as follows: the neural network model accepts an input vector of length 2048 (for example, 2048 sampled signals can be used as input), and after passing through multiple convolutional layers, pooling layers, and fully connected layers, the output vector has a length of 2 (i.e., the output consists of two of the aforementioned noise reduction parameters). For example... Figure 7B Another possible convolutional neural network model structure is shown below: the neural network model accepts an input vector of length 256 (for example, 256 sampled signals can be used as input), and after passing through multiple convolutional layers, pooling layers, and fully connected layers, the output vector has a length of 2 (i.e., the output consists of two of the aforementioned noise reduction parameters). It is easy to understand that the structure of the neural network model, such as the number of convolutional and pooling layers, the size of the convolutional kernels and the stride, the sampling method (mean or maximum sampling) and stride of the pooling layers, and the number of nodes in each layer, can all be designed according to computing power.

[0200] The training of the corresponding neural network model will be described later and will not be elaborated here.

[0201] S350: The cloud device sends the trained neural network model to the terminal of the power line device; that is, the network parameters of the neural network model are sent to the terminal. In this embodiment, the neural network model includes information about the structure of the neural network model described in step S340, such as the weight information of each layer. The weight information of each layer includes, for example, the weights of each layer in a fully connected neural network model, and the convolution kernels of a convolutional network.

[0202] S360: The terminal of the power line communication device generates or updates a local neural network model based on the parameters of the received neural network model.

[0203] Therefore, the neural network model reconstructed by the terminal of the power line communication equipment should be consistent with the neural network model trained by the cloud device, so as to ensure that the updated neural network model of the power line equipment terminal is the latest trained model.

[0204] S370: The terminal of the power line communication equipment uses the local neural network model to calculate the first noise reduction parameter and the second noise reduction parameter for the signal segment or the downsampled signal segment, and performs noise reduction processing on at least one of the sampling points to achieve noise reduction of the signal.

[0205] In some embodiments, the noise reduction process includes a zeroing method. For example, a power line communication device determines at least one sampling point from the sampling points of signal segment N0 or downsampled signal segment N1 whose amplitude is greater than the product of a first noise reduction parameter and a second noise reduction parameter, and zeroes the amplitude of the at least one sampling point. Figure 1 As shown, Figure 1The diagram illustrates four consecutive discrete-time sampling points. From these, we can determine that the sampling point at position n-2, whose amplitude is greater than the product of the first and second noise reduction parameters, will be zeroed out; that is, its amplitude will be set to 0. Sampling points that are zeroed out are considered to be affected by high-amplitude pulses and are directly zeroed out to eliminate impulse noise.

[0206] In some embodiments, the noise reduction process includes a top-cutting method. For example, from the sampling points of signal segment N0 or downsampled signal segment N1, at least one sampling point is determined whose amplitude is greater than or equal to a first noise reduction parameter and less than or equal to the product of the first noise reduction parameter and a second noise reduction parameter, and the amplitude of the at least one sampling point is set to the first noise reduction parameter. Figure 1 As shown, Figure 1 The diagram illustrates four consecutive discrete-time sampling points. From these, we can determine that the sampling point at position n-1 satisfies the condition that its amplitude is greater than or equal to the first noise reduction parameter and less than or equal to the product of the first and second noise reduction parameters. This point will undergo peak clipping, meaning its amplitude will be set to the first noise reduction parameter. Sampling points undergoing peak clipping are considered to be affected by pulses with small amplitudes; this is equivalent to limiting their amplitude and restoring their original signal value.

[0207] In this embodiment, a noise reduction method combining the zeroing method and the top-cutting method based on two noise reduction parameters is employed, namely, a three-segment method for impulse noise suppression in power line communication. Because this application comprehensively considers the varying degrees of impact from impulse noise, it can retain more effective signals and improve noise reduction performance.

[0208] In some other embodiments, when not limited to the size of the neural network model or the computing power of the device, the downsampling step described in step S320 may be omitted, and the signal segment described in step S310 may be processed directly. That is, the subsequent steps S330, S340 and S370 process the signal segment, rather than the downsampled signal judgment. This will not be described in detail again.

[0209] [A specific implementation method for the neural network model training method in step S340]

[0210] Reference Figure 8A The figure shows a flowchart of a neural network model training method provided in an embodiment of this application. This training method is based on a supervised method using mean squared error as the loss function. As shown in the figure, the method includes the following steps:

[0211] Step 1: Input the dataset consisting of signal segments sent by the terminal, and predict the first and second noise reduction parameters corresponding to the signal segments;

[0212] Step 2: Calculate the squared error of the noise reduction parameters using the first noise reduction parameter, the second noise reduction parameter, and the optimal first noise reduction parameter and the second noise reduction parameter of the signal segment. Feed this error backpropagation into the neural network model as the value of the loss function and update the neural network model.

[0213] In this embodiment of the application, the loss function can be expressed as: in, These are the predicted first and second denoising parameters calculated by the neural network model, while T and a are the calculated optimal first and optimal denoising parameters.

[0214] [Another specific implementation of the neural network model training method in step S340]

[0215] Reference Figure 8B Figure 1 shows a flowchart of another neural network model training method provided in this application embodiment. This training method is based on an unsupervised method using the inverse of the signal-to-interference-plus-noise ratio as the loss function. As shown in the figure, the method includes the following steps:

[0216] Step 1: Input the dataset consisting of signal segments sent by the terminal, and predict the first and second noise reduction parameters corresponding to the signal segments;

[0217] Step 2: Perform noise reduction on the signal segment according to the first noise reduction parameter and the second noise reduction parameter to obtain the noise-reduced signal segment;

[0218] The third step is to calculate the signal-to-interference-plus-noise ratio (SINR) of the denoised signal segment based on the signal segments before and after denoising, and then use the reciprocal of the SINR as the value of the loss function to feed back into the neural network model for error backpropagation and update the neural network model.

[0219] In this embodiment of the application, the loss function can be expressed as: Among them, P n P is the average power of the difference between the denoised signal and the original signal. x It is the average power of the original signal.

[0220] like Figure 9A , Figure 9B As shown, this application also provides embodiments of corresponding signal processing apparatuses. For the beneficial effects of the apparatuses or the technical problems they solve, please refer to the descriptions in the methods corresponding to each apparatus, or to the descriptions in the invention summary, which will not be repeated here.

[0221] [First Embodiment of a Signal Processing Apparatus]

[0222] like Figure 9AIn the first embodiment of the signal processing apparatus shown, which is applied to the terminal side of a power line communication device, the signal processing apparatus includes:

[0223] The signal segment acquisition unit 210 is used to acquire a signal segment containing impulse noise, the signal segment including at least one sampling point;

[0224] The noise reduction parameter acquisition unit 220 is used to input the signal segment into a neural network model to obtain noise reduction parameters;

[0225] The noise reduction processing unit 230 is used to perform noise reduction processing on the at least one sampling point according to the noise reduction parameters.

[0226] In some embodiments, when the neural network model is a local neural network model, the apparatus further includes:

[0227] The transmitting unit 240 is used to transmit the signal segment to the cloud device; the signal segment is used by the cloud device to train a neural network model, and the neural network model takes the signal segment as input and outputs the noise reduction parameters as output.

[0228] The receiving unit 250 is used to receive parameters of the neural network model trained by the cloud device;

[0229] The neural network model building unit 260 is used to reconstruct a local neural network model based on the parameters of the neural network model.

[0230] In some embodiments, the signal segment is a downsampled signal segment; the downsampled signal segment retains the peak points of the impulse noise.

[0231] In some embodiments, the downsampling includes downsampling using a local maximum downsampling algorithm, including:

[0232] Obtain the sub-segments included in the signal segment;

[0233] Obtain a sampling point with the maximum signal amplitude in each of the sub-segments, and the sampling points of each of the sub-segments constitute the downsampled signal segment.

[0234] In some embodiments, the downsampling includes downsampling using a global maximum windowing downsampling algorithm, including:

[0235] Obtain the sampling point in the signal segment where the signal amplitude is at its maximum value;

[0236] Obtain a series of consecutive sampling points in the signal segment that contain the maximum value of the signal amplitude, and the series of consecutive sampling points constitute the downsampled signal.

[0237] In some embodiments, the sending includes:

[0238] Each of the aforementioned signal segments is buffered;

[0239] Upon reaching a certain time point, the cached signal fragments are sent to the cloud device.

[0240] In some embodiments, the noise reduction parameters include a first noise reduction parameter and a second noise reduction parameter;

[0241] The first noise reduction parameter serves as a first threshold value, used to perform top-trimming noise reduction processing on signals whose amplitude exceeds the first threshold value;

[0242] The second noise reduction parameter is used as a coefficient, and its product with the first noise reduction parameter is used as a second threshold value. This is used to perform noise reduction processing on signals whose amplitude exceeds the second threshold value, and the second noise reduction parameter is greater than 1.

[0243] [Second Embodiment of the Signal Processing Apparatus]

[0244] like Figure 9B In a second embodiment of the signal processing apparatus shown, which is applied to a cloud device, the signal processing apparatus includes:

[0245] The receiving unit 310 is configured to receive a signal segment; the signal segment contains impulse noise and includes at least one sampling point;

[0246] The neural network model training unit 320 is used to train a neural network model using the signal segment, wherein the input of the neural network model is the signal segment and the output is a neural network model with noise reduction parameters;

[0247] The transmitting unit 330 is used to send the parameters of the trained neural network model to the terminal device.

[0248] In some embodiments, the training includes a supervised training method using mean squared error as a loss function, wherein the loss function is: the square of the difference between the optimal first denoising parameter and the predicted first denoising parameter, plus the square of the difference between the optimal second denoising parameter and the predicted second denoising parameter.

[0249] The signal segment has the optimal first noise reduction parameter and the optimal second noise reduction parameter; the predicted first noise reduction parameter and the predicted second noise reduction parameter are obtained by inputting the signal segment into the neural network model.

[0250] In some embodiments, the training includes an unsupervised training method using the reciprocal of the signal-to-interference-plus-noise ratio (SINR) as a loss function. The loss function includes the reciprocal of the denoised SINR calculated from the denoised signal segment obtained after the denoising process based on the predicted first denoising parameter and the predicted second denoising parameter. The predicted first denoising parameter and the predicted second denoising parameter are obtained by inputting the signal segment into the neural network model.

[0251] [Analysis of the effectiveness of this application]

[0252] The effect of noise reduction processing under different signal-to-interference ratio (SNR) conditions was analyzed. In this experiment, the SNR corresponding to white noise was set to 15dB; the impulse noise was a simulated home environment consisting of a set-top box, a humidifier, and an LED spotlight; the multipath channel was set to 3 paths with a path interval of 100 sampling points and gain coefficients of [1, -0.3, 0.2] respectively; and the transmission delay was 1000 sampling points.

[0253] Table 1 below shows the training results of a network with an input signal length of 256:

[0254]

[0255] Table 1

[0256] Table 2 below shows the training results of the network with an input signal length of 2048:

[0257]

[0258]

[0259] Table 2

[0260] Table 3 below shows the simulation noise reduction results:

[0261]

[0262] Table 3 shows the "Interference Reduction" column, which is calculated using the following formula and process:

[0263]

[0264] Among them, "Pn before" and "Pn after" are "noise power before noise reduction" and "noise power after noise reduction", respectively, and "Ps" is "signal power".

[0265] The SINR before noise reduction is -1.193dB, which was obtained under experimental conditions (white noise power -20dB, impulse noise power 0dB, signal power 0dB) when the signal was applied. It can be obtained using the following formula:

[0266] Before noise cancellation

[0267] The "SINR after noise reduction" in Table 3 above is calculated by averaging the SINR of the training set and the SINR of the test set in Table 1 or Table 2. Specifically, it can be calculated using the following formula:

[0268] After noise reduction, SINR = (training set SINR + test set SINR) / 2

[0269] As shown in Table 3 above, the various combinations of neural networks and loss functions used in the experiments all reduced interference by more than 10 dB during the noise reduction process. This result demonstrates that the method described in this application can select different combinations to generate and train the neural network model based on the actual noise data conditions (data size, amplitude length of clustered trails, etc.), thereby achieving better noise reduction results.

[0270] [Examples of the computing device in this application]

[0271] Figure 10 This is a schematic structural diagram of a computing device 900 provided in an embodiment of this application. The computing device 900 includes: a processor 910, a memory 920, and a communication interface 930.

[0272] It should be understood that Figure 10 The communication interface 930 in the computing device 900 shown can be used to communicate with other devices.

[0273] The processor 910 can be connected to the memory 920. The memory 920 can be used to store the program code and data. Therefore, the memory 920 can be a storage unit inside the processor 910, an external storage unit independent of the processor 910, or a component that includes both the storage unit inside the processor 910 and the external storage unit independent of the processor 910.

[0274] Optionally, the computing device 900 may also include a bus. The memory 920 and communication interface 930 can be connected to the processor 910 via the bus. The bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc.

[0275] It should be understood that in the embodiments of this application, the processor 910 may be a central processing unit (CPU). The processor may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor. Alternatively, the processor 910 may employ one or more integrated circuits to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0276] The memory 920 may include read-only memory and random access memory, and provides instructions and data to the processor 910. A portion of the processor 910 may also include non-volatile random access memory. For example, the processor 910 may also store device type information.

[0277] When the computing device 900 is running, the processor 910 executes the computer execution instructions in the memory 920 to perform the operation steps of the above method.

[0278] It should be understood that the computing device 900 according to the embodiments of this application can correspond to the corresponding subject in executing the methods according to the various embodiments of this application, and the above and other operations and / or functions of each module in the computing device 900 are respectively for implementing the corresponding processes of the methods of this embodiment. For the sake of brevity, they will not be described in detail here.

[0279] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0280] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0281] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0282] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0283] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0284] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0285] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, performs a diversified problem generation method, including at least one of the schemes described in the above embodiments.

[0286] The computer storage medium in this application embodiment can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. For example, a computer-readable storage medium can be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0287] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0288] The program code contained on a computer-readable medium may be transmitted using any suitable medium, including, but not limited to, wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0289] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0290] Note that the above are merely preferred embodiments and the technical principles employed in this application. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present application has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, all of which fall within the scope of protection of the present invention.

Claims

1. A signal processing method, characterized in that, include: A signal segment containing impulse noise is acquired, the signal segment including at least one sampling point; the signal segment is a downsampled signal segment, the downsampled signal segment retaining the peak points of the impulse noise; wherein, the downsampling algorithm is selected according to the type of impulse noise: when the impulse noise is point-like impulse noise, a local maximum downsampling algorithm is used for downsampling; when the impulse noise interference is cluster-like impulse noise, a global maximum windowing downsampling algorithm is used for downsampling. The signal segment is input into a neural network model to obtain noise reduction parameters; The noise reduction process is performed on the at least one sampling point according to the noise reduction parameters.

2. The method according to claim 1, characterized in that, When the neural network model is a local neural network model, the method further includes: The signal segment is sent to a cloud device; the signal segment is used by the cloud device to train a neural network model, and the neural network model takes the signal segment as input and outputs the neural network model with the noise reduction parameters. Receive parameters from the neural network model trained by the cloud device; The local neural network model is constructed based on the parameters of the neural network model.

3. The method according to claim 1, characterized in that, The downsampling includes downsampling using a local maximum downsampling algorithm, including: Obtain the sub-segments included in the signal segment; Obtain a sampling point with the maximum signal amplitude in each of the sub-segments, and the sampling points of each of the sub-segments constitute the downsampled signal segment.

4. The method according to claim 1, characterized in that, The downsampling includes downsampling using a global maximum windowing downsampling algorithm, including: Obtain the sampling point in the signal segment where the signal amplitude is at its maximum value; Obtain a series of consecutive sampling points in the signal segment that contain the maximum value of the signal amplitude, and the series of consecutive sampling points constitute the downsampled signal.

5. The method according to any one of claims 2-4, characterized in that, The sending includes: Each of the aforementioned signal segments is buffered; Upon reaching a certain time point, the cached signal fragments are sent to the cloud device.

6. The method according to claim 1 or 2, characterized in that: The noise reduction parameters include a first noise reduction parameter and a second noise reduction parameter; The first noise reduction parameter serves as a first threshold value, used to perform top-trimming noise reduction processing on signals whose amplitude exceeds the first threshold value; The second noise reduction parameter is used as a coefficient, and its product with the first noise reduction parameter is used as a second threshold value. This is used to perform noise reduction processing on signals whose amplitude exceeds the second threshold value, and the second noise reduction parameter is greater than 1.

7. A signal processing method, characterized in that, include: Received signal segments; The signal segment contains impulse noise, and the signal segment includes at least one sampling point; the signal segment is a downsampled signal segment, and the downsampled signal segment retains the peak points of the impulse noise; wherein, the downsampling algorithm is selected according to the type of impulse noise: when the impulse noise is point-like impulse noise, a local maximum downsampling algorithm is used for downsampling; when the impulse noise interference is clustered impulse noise, a global maximum windowing downsampling algorithm is used for downsampling. The signal segment is used to train a neural network model, wherein the input of the neural network model is the signal segment and the output is a neural network model with noise reduction parameters; Send the parameters of the trained neural network model to the terminal device.

8. The method according to claim 7, characterized in that, The training includes a supervised training method using mean squared error as the loss function, wherein the loss function is: the square of the difference between the optimal first denoising parameter and the predicted first denoising parameter, plus the square of the difference between the optimal second denoising parameter and the predicted second denoising parameter. The signal segment has the optimal first noise reduction parameter and the optimal second noise reduction parameter; The predicted first noise reduction parameter and the predicted second noise reduction parameter are obtained by inputting the signal segment into the neural network model.

9. The method according to claim 7, characterized in that, The training includes an unsupervised training method that uses the reciprocal of the signal-to-interference-plus-noise ratio (SINR) as a loss function. The loss function includes the reciprocal of the denoised signal-to-interference-plus-noise ratio (SINR) calculated from the denoised signal segment obtained after the denoising process based on the predicted first denoising parameter and the predicted second denoising parameter. The predicted first noise reduction parameter and the predicted second noise reduction parameter are obtained by inputting the signal segment into the neural network model.

10. A signal processing method, characterized in that, include: The signal processing method according to any one of claims 1-6, and The signal processing method according to any one of claims 7-9.

11. A signal processing apparatus, characterized in that, include: A signal segment acquisition unit is used to acquire a signal segment containing impulse noise, the signal segment including at least one sampling point; the signal segment is a downsampled signal segment, the downsampled signal segment retains the peak points of the impulse noise; wherein, the downsampling algorithm is selected according to the type of impulse noise: when the impulse noise is point-like impulse noise, a local maximum downsampling algorithm is used for downsampling; when the impulse noise interference is clustered impulse noise, a global maximum windowing downsampling algorithm is used for downsampling. A noise reduction parameter acquisition unit is used to input the signal segment into a neural network model to obtain noise reduction parameters; A noise reduction processing unit is used to perform noise reduction processing on the at least one sampling point according to the noise reduction parameters.

12. The apparatus according to claim 11, characterized in that: When the neural network model is a local neural network model, the device further includes: A transmitting unit is used to transmit the signal segment to a cloud device; the signal segment is used by the cloud device to train a neural network model, and the neural network model takes the signal segment as input and outputs the noise reduction parameters as output. The receiving unit is used to receive the parameters of the neural network model trained by the cloud device; A neural network model building unit is used to reconstruct a local neural network model based on the parameters of the neural network model.

13. The apparatus according to claim 11, characterized in that, The downsampling includes downsampling using a local maximum downsampling algorithm, including: Obtain the sub-segments included in the signal segment; Obtain a sampling point with the maximum signal amplitude in each of the sub-segments, and the sampling points of each of the sub-segments constitute the downsampled signal segment.

14. The apparatus according to claim 11, characterized in that, The downsampling includes downsampling using a global maximum windowing downsampling algorithm, including: Obtain the sampling point in the signal segment where the signal amplitude is at its maximum value; Obtain a series of consecutive sampling points in the signal segment that contain the maximum value of the signal amplitude, and the series of consecutive sampling points constitute the downsampled signal.

15. The apparatus according to any one of claims 11-14, characterized in that, The sending includes: Each of the aforementioned signal segments is buffered; Upon reaching a certain time point, the cached signal fragments are sent to the cloud device.

16. The apparatus according to claim 11, characterized in that: The noise reduction parameters include a first noise reduction parameter and a second noise reduction parameter; The first noise reduction parameter serves as a first threshold value, used to perform top-trimming noise reduction processing on signals whose amplitude exceeds the first threshold value; The second noise reduction parameter is used as a coefficient, and its product with the first noise reduction parameter is used as a second threshold value. This is used to perform noise reduction processing on signals whose amplitude exceeds the second threshold value, and the second noise reduction parameter is greater than 1.

17. A signal processing apparatus, characterized in that, include: The receiving unit is used to receive signal segments; The signal segment contains impulse noise, and the signal segment includes at least one sampling point; the signal segment is a downsampled signal segment, and the downsampled signal segment retains the peak points of the impulse noise; wherein, the downsampling algorithm is selected according to the type of impulse noise: when the impulse noise is point-like impulse noise, a local maximum downsampling algorithm is used for downsampling; when the impulse noise interference is clustered impulse noise, a global maximum windowing downsampling algorithm is used for downsampling. A neural network model training unit is used to train a neural network model using the signal segment, wherein the input of the neural network model is the signal segment and the output is a neural network model with noise reduction parameters; The transmitting unit is used to send the parameters of the trained neural network model to the terminal device.

18. The apparatus according to claim 17, characterized in that, The training includes a supervised training method using mean squared error as the loss function, wherein the loss function is: the square of the difference between the optimal first denoising parameter and the predicted first denoising parameter, plus the square of the difference between the optimal second denoising parameter and the predicted second denoising parameter. The signal segment has the optimal first noise reduction parameter and the optimal second noise reduction parameter; The predicted first noise reduction parameter and the predicted second noise reduction parameter are obtained by inputting the signal segment into the neural network model.

19. The apparatus according to claim 17, characterized in that, The training includes an unsupervised training method that uses the reciprocal of the signal-to-interference-plus-noise ratio (SINR) as a loss function. The loss function includes the reciprocal of the denoised signal-to-interference-plus-noise ratio (SINR) calculated from the denoised signal segment obtained after the denoising process based on the predicted first denoising parameter and the predicted second denoising parameter. The predicted first noise reduction parameter and the predicted second noise reduction parameter are obtained by inputting the signal segment into the neural network model.

20. A signal processing apparatus, characterized in that, include: The signal processing apparatus according to any one of claims 11-16, and The signal processing apparatus according to any one of claims 17-19.

21. A computing device, characterized in that, include: Communication interface; At least one processor connected to the communication interface; as well as At least one memory connected to the processor and storing program instructions that, when executed by the at least one processor, cause the at least one processor to perform the signal processing method of any one of claims 1-6, or the program instructions that, when executed by the at least one processor, cause the at least one processor to perform the signal processing method of any one of claims 7-9.

22. A computer-readable storage medium having program instructions stored thereon, characterized in that, When the program instructions are executed by a computer, the computer performs the signal processing method according to any one of claims 1-6, or the program instructions, when executed by a computer, perform the signal processing method according to any one of claims 7-9.

Images