Signal classification method, training method and device based on neural network
By employing parallel feature extraction and classification sub-modules in the neural network, combined with linear and nonlinear layers, features are extracted and classified for each signal one by one, solving the problem of inaccurate signal classification in existing technologies and achieving efficient signal classification and model training.
Patent Information
- Application Number
- CN202411063345.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-02
- Publication Date
- 2026-02-03
AI Technical Summary
Existing technologies struggle to effectively classify mixed signals, especially when multiple types of interference or content categories exist in communication systems, image signals, and audio signals, making the classification methods inaccurate.
A neural network-based signal classification method is adopted, which extracts and classifies signals one by one through parallel feature extraction and classification sub-modules. It utilizes a combination of linear and nonlinear layers for feature extraction and classification, and improves accuracy through a feature pre-identification module.
It achieves effective signal classification, improves the accuracy of signal classification and the efficiency of model training, and is suitable for lightweight neural network models.
Smart Images
Figure CN121456633A_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to the field of signal classification, and more specifically, to a signal classification method and apparatus based on neural networks, a training method and apparatus for a neural network for signal classification, and a neural network system for signal classification. Background Technology
[0002] In signal processing, it is common to encounter situations where a single signal contains multiple signal categories. For example, interference signals in communication systems may include various types of interference (such as interference from smart streetlights or 800MHz telecommunications signals); similarly, an image signal may contain multiple categories of images, such as portraits, animals, or landscapes in terms of content; traffic images may include red, green, and yellow images; and an audio signal may contain multiple categories of audio, such as music or speech. In these cases, signal classification methods are necessary. Summary of the Invention
[0003] This disclosure aims to provide a method and a neural network system for signal classification using artificial intelligence neural networks, which can effectively classify signals.
[0004] According to one aspect of this disclosure, a signal classification method based on a neural network is provided, wherein the neural network includes a feature extraction module and a classification module; the feature extraction module includes multiple parallel feature extraction sub-modules; the classification module includes multiple parallel classification sub-modules corresponding to each of the multiple feature extraction sub-modules; each feature extraction sub-module and each classification sub-module corresponds to a signal category; and the method includes: each extraction sub-module extracting its own feature signal based on an input signal; and each classification sub-module determining whether the input signal contains a signal of a corresponding signal category based on the feature signal extracted by its corresponding feature extraction sub-module.
[0005] In one embodiment, each of the plurality of feature extraction submodules includes a first linear layer and a first nonlinear layer, and each of the plurality of classification submodules includes a second linear layer and a second nonlinear layer.
[0006] In one embodiment, the first nonlinear layer is implemented using the ReLU activation function, and the second nonlinear layer is implemented using the Sigmoid activation function.
[0007] In one embodiment, the neural network is trained via the following steps: each of the plurality of feature extraction submodules extracts its own feature signal based on the sample input signal; each of the plurality of classification submodules determines whether the sample input signal contains a signal of a corresponding signal category based on the feature signal extracted by its corresponding feature extraction submodule; a first loss is determined based on a first loss function between the determination results of the plurality of classification submodules for their respective corresponding signal categories and the labeled signal category of the sample input signal; the feature signals extracted by the plurality of feature extraction submodules based on the sample input signal are summed, and a second loss is determined based on a second loss function between the summed feature signal and the sample input signal; and the model parameters of the neural network are adjusted based on the first loss and the second loss.
[0008] In one embodiment, the feature extraction module further includes a feature pre-identification module, and the method further includes: the feature pre-identification module performing feature pre-identification on the input signal to generate a pre-identification signal, and inputting the pre-identification signal to each of the plurality of feature extraction sub-modules.
[0009] In one embodiment, the input signal includes a communication interference signal, and the signal category includes a communication interference category; the input signal includes an image signal, and the signal category includes an image content category; the input signal includes a video signal, and the signal category includes a video content category; or the input signal includes an audio signal, and the signal category includes an audio content category.
[0010] According to another aspect of this disclosure, a neural network system for signal classification is provided, the method comprising: a feature extraction module including a plurality of parallel feature extraction sub-modules; and a classification module including a plurality of parallel classification sub-modules corresponding to each of the plurality of feature extraction sub-modules, wherein each feature extraction sub-module and each classification sub-module corresponds to a signal category, each feature extraction sub-module is configured to extract its own feature signal based on an input signal, and each classification sub-module is configured to determine whether the input signal contains a signal of a corresponding signal category based on the feature signal extracted by its corresponding feature extraction sub-module.
[0011] In one embodiment, each of the plurality of feature extraction submodules includes a first linear layer and a first nonlinear layer, and each of the plurality of classification submodules includes a second linear layer and a second nonlinear layer.
[0012] In one embodiment, the first nonlinear layer is implemented using the ReLU activation function, and the second nonlinear layer is implemented using the Sigmoid activation function.
[0013] In one embodiment, the neural network system is trained via the following steps: each of the plurality of feature extraction submodules extracts its own feature signal based on the sample input signal; each of the plurality of classification submodules determines whether the sample input signal contains a signal of a corresponding signal category based on the feature signal extracted by its corresponding feature extraction submodule; a first loss is determined based on a first loss function between the determination results of the plurality of classification submodules for their respective corresponding signal categories and the labeled signal category of the sample input signal; the feature signals extracted by the plurality of feature extraction submodules based on the sample input signal are summed, and a second loss is determined based on a second loss function between the summed feature signal and the sample input signal; and the model parameters of the neural network are adjusted based on the first loss and the second loss.
[0014] In one embodiment, the feature extraction module further includes a feature pre-identification module, and the feature pre-identification module is configured to perform feature pre-identification on the input signal to generate a pre-identification signal, and input the pre-identification signal to each of the plurality of feature extraction sub-modules.
[0015] In one embodiment, the input signal includes a communication interference signal, and the signal category includes a communication interference category; the input signal includes an image signal, and the signal category includes an image content category; the input signal includes a video signal, and the signal category includes a video content category; or the input signal includes an audio signal, and the signal category includes an audio content category.
[0016] According to another aspect of this disclosure, a training method for a neural network for signal classification is provided. The neural network includes a feature extraction module and a classification module. The feature extraction module includes multiple parallel feature extraction sub-modules, and the classification module includes multiple parallel classification sub-modules corresponding to each of the multiple feature extraction sub-modules. Each feature extraction sub-module and each classification sub-module corresponds to a signal category. The method includes: acquiring a sample input signal and a labeled signal category of the sample input signal; each feature extraction sub-module extracts its own feature signal based on the sample input signal; each classification sub-module determines whether the sample input signal contains a signal of a corresponding signal category based on the feature signal extracted by its corresponding feature extraction sub-module; determining a first loss between the determination result of the multiple classification sub-modules for their respective corresponding signal categories and the labeled signal category of the sample input signal based on a first loss function; summing the feature signals extracted by the multiple feature extraction sub-modules based on the sample input signal and determining a second loss between the summed feature signals and the sample input signal based on a second loss function; and adjusting the model parameters of the neural network based on the first loss and the second loss.
[0017] In one embodiment, the feature extraction module further includes a feature pre-identification module, and the method further includes: the feature pre-identification module is configured to perform feature pre-identification on the sample input signal to generate a sample pre-identification signal, and input the sample pre-identification signal to each of the plurality of feature extraction sub-modules.
[0018] In one embodiment, each of the plurality of feature extraction submodules includes a first linear layer and a first nonlinear layer, and each of the plurality of classification submodules includes a second linear layer and a second nonlinear layer.
[0019] In one embodiment, the first nonlinear layer is implemented using the ReLU activation function, and the second nonlinear layer is implemented using the Sigmoid activation function.
[0020] In one embodiment, the sample input signal includes a communication interference signal, and the signal category includes a communication interference category; the sample input signal includes an image signal, and the signal category includes an image content category; the sample input signal includes a video signal, and the signal category includes a video content category; or the sample input signal includes an audio signal, and the signal category includes an audio content category.
[0021] According to another aspect of this disclosure, an apparatus for signal classification is provided, comprising: a processor; and a memory storing one or more computer programs; wherein, when the one or more computer programs are run by the processor, the processor causes the processor to perform the method described above.
[0022] According to another aspect of this disclosure, a computer program product is provided, which includes a computer program that, when executed by a processor, implements the method as described above.
[0023] According to the signal classification method of this disclosure, by setting corresponding sub-models for a neural network model based on the possible categories of the classified signal, effective signal classification can be achieved. In some embodiments, high-accuracy signal classification can be achieved on a lightweight neural network model. Attached Figure Description
[0024] Figure 1 A neural network system for signal classification according to an embodiment of the present disclosure is schematically illustrated;
[0025] Figure 2 A method for signal classification according to an embodiment of the present disclosure is illustrated schematically;
[0026] Figure 3 The diagram schematically illustrates a training method for a neural network system for signal classification according to embodiments of the present disclosure; and
[0027] Figure 4 A block diagram of a neural network system for classifying communication interference signals according to an embodiment of the present disclosure is shown schematically. Detailed Implementation
[0028] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.
[0029] It should be understood that the terms “comprising” or “may include” as used herein refer to the presence of a corresponding disclosed function, operation, or component that may be used in the various embodiments of this disclosure, rather than limiting the presence of one or more additional functions, operations, or features. Furthermore, the terms “comprising” or “having” may be interpreted as indicating certain characteristics, numbers, steps, operations, constituent elements, components, or combinations thereof, but should not be construed as excluding the possibility of the presence of one or more other characteristics, numbers, steps, operations, constituent elements, components, or combinations thereof.
[0030] The concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules, units, models, data, etc., and are not used to limit the order of the functions performed by these devices, modules, units, models, data, or their interdependencies.
[0031] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".
[0032] As used herein, any reference to “an example” or “example,” “an embodiment” or “an embodiment” means that a particular element, feature, structure, or characteristic described in connection with that embodiment is included in at least one embodiment. The phrases “in one embodiment” or “in one example” appearing in different places in the specification do not necessarily refer to the same embodiment.
[0033] It will be further understood that the terms “including” or “contains” and similar words mean that the element or object preceding the word covers the element or object listed after the word and its equivalents, without excluding other elements or objects.
[0034] The various embodiments discussed below for describing the principles of this disclosure in this patent document are for illustrative purposes only and should not be construed in any way as limiting the scope of this disclosure.
[0035] The following description, with reference to the accompanying drawings, is provided to aid in a thorough understanding of the various embodiments of this disclosure as defined by the claims and their equivalents. This description includes various specific details to aid understanding but should be considered exemplary only. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the various embodiments described herein without departing from the scope and spirit of this disclosure. Furthermore, for clarity and brevity, descriptions of well-known functions and structures may be omitted.
[0036] The terms and wording used in the following description and claims are not limited to their dictionary meanings, but are merely used by the inventors to enable a clear and consistent understanding of this disclosure. Therefore, it will be apparent to those skilled in the art that the following description of various embodiments of this disclosure is for illustrative purposes only and not for limiting the purpose of this disclosure as defined in the appended claims and their equivalents.
[0037] This disclosure provides a neural network system for signal classification, a neural network-based signal classification method and apparatus, and a training method and apparatus for the neural network for signal classification. In the neural network system according to this disclosure, feature extraction and classification are performed sequentially for each of multiple signal categories, rather than providing recognition results for all signal categories at once, thus achieving effective signal recognition. Furthermore, during the training of the neural network system, the signals obtained from feature extraction for each of the multiple signal categories are summed, and the summed signal is compared with the input signal. The comparison result is used to adjust the parameters of the neural network system, which can further improve the efficiency and success rate of model training and the accuracy of signal category recognition. In a further embodiment, using computationally less computationally intensive linear layers in the feature extraction and classification modules can make the neural network system according to this disclosure more lightweight.
[0038] It should be noted that the signals mentioned in this disclosure can be various signals mixed with multiple categories, such as communication interference signals having at least one type of interference, image signals having at least one image content category (e.g., images are divided into multiple image content categories according to target object category, texture category, or color category), video signals having at least one video content category (e.g., classified as still images and moving images), and audio signals having at least one audio content category (e.g., classified as speech signals, music signals, or other noise signals). The method of acquiring the above signals is not specifically limited in this disclosure.
[0039] Figure 1 A neural network system 10 for signal classification according to an embodiment of the present disclosure is shown.
[0040] The neural network system 10 may include a feature extraction module 101 for extracting feature signals from the input signal to be identified, and a classification module 102 for identifying and classifying the extracted feature signals. By performing feature extraction and feature recognition steps to extract and classify the possible types of the signal separately, rather than directly performing a single feature recognition and providing a classification result, the task of classifying the signal is represented in the form of a thought process chain, thereby improving the accuracy of signal classification.
[0041] Feature extraction module 101 may include N feature extraction sub-modules (N is a natural number greater than or equal to 1, and these feature extraction sub-modules are...) Figure 1 The submodules are labeled 1012-1, 1012-2, and so on, up to 1012-N, while the classification module 102 can include N classification submodules (in...). Figure 1 The input signals are labeled as 1021-1, 1021-2, and so on, up to 1021-N. Here, N represents the total number of possible signal categories in the input signal to be identified.
[0042] There are N feature extraction submodules and N classification submodules, each with a one-to-one correspondence. Furthermore, each feature extraction submodule and each classification submodule corresponds to a specific signal category. For example, feature extraction submodule 1012-1 and classification submodule 1021-1 correspond to signal category 1, meaning their processing focuses on signal category 1. Feature extraction submodule 1012-2 and classification submodule 1021-2 correspond to signal category 2, and so on.
[0043] Each of the N feature extraction submodules is configured to extract its own feature signal based on the input signal, and each of the N classification submodules is configured to determine whether the input signal contains a signal of a corresponding signal category based on the feature signals extracted by its corresponding feature extraction submodule. The feature signals extracted by each feature extraction submodule can have the same format as the input signal. For example, the input signal is a one-dimensional vector data, and the feature signals are also one-dimensional vector data, with the same number of vector elements. In an alternative embodiment, the feature signals extracted by each feature extraction submodule can also have a different format than the input signal. In this case, they need to be adjusted to have the same signal format as the input signal before summing, so that the sum of the extracted feature signals can be compared with the input signal. For example, when the feature signal has a small number of elements, the number of vector elements of the feature signal can be padded with zeros to make it the same as the input signal format.
[0044] Taking communication interference signals as an example, in a communication system, interference types can include the following categories: 800MHz telecommunications interference, fake base station interference, NR system internal interference, LTE D-band interference, 700MHz interference, smart street light interference, wireless bridge / video surveillance interference, jammer interference, router interference, clock source failure interference, broadcast signal interference, external interference, and other interference types. In other words, in the above example, communication interference signals can include a total of 13 types of interference.
[0045] Based on this, when classifying communication interference signals, N can be set to 13, meaning that the neural network system 10 includes 13 feature extraction sub-modules and 13 classification sub-modules.
[0046] As described above, the 13 feature extraction submodules and 13 classification submodules can correspond to the 13 interference types mentioned above. For example, feature extraction submodule 1012-1 and classification submodule 1021-1 can target the smart street light interference type among the 13 interference types. Feature extraction submodule 1012-1 extracts feature waveforms for this type of interference, and classification submodule 1021-1 identifies and classifies the input communication interference signal based on the feature waveforms extracted by feature extraction submodule 1012-1, determining whether the input communication interference signal contains the smart street light interference type. Feature extraction submodule 1012-2 and classification submodule 1021-2 can correspond to the 800M telecommunications interference type among the 13 interference types. That is, the processing of these two submodules focuses on signal category 2, and classification submodule 1021-2 determines whether the input communication interference signal contains the smart street light interference type.
[0047] In a more specific embodiment, the determination result can be a value of 0 or a value of 1, where a value of 0 indicates that there is no signal of the corresponding signal category in the input signal, and a value of 1 indicates that there is a signal of the corresponding signal category in the input signal. Taking the classification submodule 1021-1 for smart street light interference as an example, a value of 1 output by the classification submodule 1021-1 represents that there is smart street light interference in the input communication interference signal, and a value of 0 output by the classification submodule 1021-1 represents that there is no smart street light interference in the input communication interference signal.
[0048] The results can also be represented in other ways, such as using probabilities. For example, for classification submodule 1021-1, a probability greater than 50% indicates that smart street light interference exists in the input communication interference signal, while a probability less than or equal to 50% indicates that smart street light interference does not exist in the input communication interference signal. Correspondingly, other classification submodules (e.g., classification submodule 1021-2) can also use values of 1 or 0, or probabilities greater than 50% or less than or equal to 50%, to determine whether 800M telecommunications interference exists in the input communication interference signal.
[0049] Taking image signals as an example again, there are various ways to classify images, which can be set according to specific needs. For example, the image signal can be classified according to the category of the target object of interest, color category, or texture category in the image signal. Suppose that the current image signal to be classified needs to be classified as a person image, animal image, landscape image, or other image, then the above N can be set to 4, that is, the neural network system 10 includes 4 feature extraction sub-modules and 4 classification sub-modules.
[0050] These four feature extraction submodules and four classification submodules can correspond to the four image content types mentioned above, respectively. For example, feature extraction submodule 1012-1 and classification submodule 1021-1 can target the "human image" type among the four image content types. Feature extraction submodule 1012-1 extracts feature signals for the "human image" content type, and classification submodule 1021-1 performs identification and classification based on the feature signals extracted by feature extraction submodule 1012-1, determining whether the input image signal contains the "human image" type. Similarly, feature extraction submodule 1012-2 and classification submodule 1021-2 can correspond to the "animal image" type among the four image content types. That is, the processing of these two submodules focuses on the "animal image" category, and classification submodule 1021-2 determines whether the input image signal contains an "animal image."
[0051] In a more specific embodiment, the determination result can be a value of 0 or a value of 1, where a value of 0 indicates that the input signal does not contain a signal of the corresponding signal category, and a value of 1 indicates that the input signal contains a signal of the corresponding signal category. Taking the classification submodule 1021-1 for human image types as an example, a value of 1 output by classification submodule 1021-1 represents the presence of a human image in the input image signal, and a value of 0 output by classification submodule 1021-1 represents the absence of a human image in the input image signal. The determination result can also be represented in other ways, such as in the form of a probability. For example, for classification submodule 1021-1 for human image types, a probability greater than 50% indicates the presence of a human image type in the input image signal, while a probability less than or equal to 50% indicates the absence of a human image type in the input image signal. Correspondingly, other classification submodules (e.g., classification submodule 1021-2) can also use values of 1 or 0, or probabilities greater than 50% or less than or equal to 50%, to determine whether an animal image exists in the input image signal.
[0052] Taking audio signals as an example again, there are various ways to classify audio signals. Audio can be classified according to specific needs, based on the specific content included in the audio signal, or the characteristics of the signal itself. Suppose we need to classify the audio signal to be classified as a speech signal, a music signal, or other noise signal. Then, we can set N to 3, meaning that the neural network system 10 includes 3 feature extraction submodules and 3 classification submodules.
[0053] These three feature extraction submodules and three classification submodules can correspond to the three audio content types mentioned above, respectively. For example, feature extraction submodule 1012-1 and classification submodule 1021-1 can target the music signal type among the three audio content types. Feature extraction submodule 1012-1 extracts feature signals for the music signal content type, and classification submodule 1021-1 identifies and classifies the input audio signal based on the feature signals extracted by feature extraction submodule 1012-1, determining whether the input audio signal contains a music signal. Similarly, feature extraction submodule 1012-2 and classification submodule 1021-2 can correspond to the speech signal type among the three audio content types. That is, the processing of these two submodules focuses on the speech signal category, and classification submodule 1021-2 determines whether the input audio signal contains a speech signal.
[0054] In a more specific embodiment, the determination result can be a value of 0 or a value of 1, where a value of 0 indicates that the input signal does not contain a signal of the corresponding signal category, and a value of 1 indicates that the input signal contains a signal of the corresponding signal category. For example, taking the classification submodule 1021-1 for music signal types as an example, a value of 1 output by the classification submodule 1021-1 represents that a music signal exists in the input image signal, and a value of 0 output by the classification submodule 1021-1 represents that a music signal does not exist in the input image signal. The determination result can also be represented in other ways, such as in the form of probability. For example, for the feature extraction submodule 1012-1 and the classification submodule 1021-1 for music signal types, a probability greater than 50% indicates that a music signal exists in the input audio signal, while a probability less than or equal to 50% indicates that a music signal does not exist in the input audio signal. Correspondingly, other classification submodules (e.g., classification submodule 1021-2) can also use a value of 1 or 0, or a probability greater than 50% or less than or equal to 50%, to determine whether a speech signal exists in the input audio signal.
[0055] By implementing the feature extraction module 101 and the classification module 102 into N feature extraction sub-modules 1012-1 to 1012-N and N classification sub-modules 1021-1 to 1021-N respectively, and making each of the N feature extraction sub-modules 1012-1 to 1012-N and the N classification sub-modules 1021-1 to 1021-N correspond to a signal category, the signals can be analyzed one by one from the perspective of different signal categories. This allows the neural network system to give the result of classifying the signal one by one, rather than directly giving the final result containing all categories, thereby improving the overall accuracy of the neural network system.
[0056] In some embodiments, each of the N feature extraction submodules includes a first linear layer and a first nonlinear layer, meaning that the N feature extraction submodules together include N first linear layers and N first nonlinear layers. For example, feature extraction submodule 1012-1 has a first linear layer 1012-11 and a first nonlinear layer 1012-12, feature extraction submodule 1012-2 has a first linear layer 1012-21 and a first nonlinear layer 1012-22, and so on. Furthermore, each of the N classification submodules includes a second linear layer and a second nonlinear layer, meaning that the N classification submodules together include N second linear layers and N second nonlinear layers. For example, classification submodule 1021-1 has a second linear layer 1021-11 and a second nonlinear layer 1021-12, classification submodule 1021-2 has a second linear layer 1021-21 and a second nonlinear layer 1021-22, and so on.
[0057] In one embodiment, the first nonlinear layer can be implemented using the ReLU activation function, and the second nonlinear layer can be implemented using the Sigmoid activation function. In other embodiments, the first and second nonlinear layers can also be implemented using other activation functions.
[0058] By using a combination of linear and nonlinear layers to achieve feature extraction and classification, the neural network system becomes relatively lightweight, while the construction of two consecutive linear and nonlinear layers ensures the recognition performance of the neural network system.
[0059] In a further embodiment, the feature extraction module 101 may further include a feature pre-identification module 1011, which is configured to perform feature pre-identification on the input signal to generate a pre-identification signal, and input the pre-identification signal to each of the N feature extraction sub-modules. For example, the feature pre-identification module 1011 may include linear layers and nonlinear layers, and the nonlinear layers may be implemented, for example, using a ReLU activation function.
[0060] By setting a feature pre-identification module before the N feature extraction sub-modules in the feature extraction module, the length of the thought chain can be increased, further improving the accuracy of signal classification and the recognition performance of the neural network system.
[0061] Figure 2 A method for signal classification using a neural network system according to an embodiment of the present disclosure is illustrated schematically.
[0062] The method includes: in step S220, each of the N feature extraction submodules in the neural network system extracts its own feature signal based on the input signal.
[0063] In step S230, each of the N classification sub-modules in the neural network system determines whether the input signal contains a signal of a corresponding signal category based on the feature signals extracted by its respective feature extraction sub-module.
[0064] Optionally, if the feature extraction module of the neural network system further includes a feature pre-identification module, the method further includes: step S210, whereby the feature pre-identification module performs feature pre-identification on the input signal to generate a pre-identification signal, and inputs the pre-identification signal to each of the N feature extraction sub-modules. It should be noted that the feature signal extraction based on the input signal in step S220 can be performed directly based on the input signal, or it can be performed based on a pre-processed input signal. For example, if step S210 is included, each feature extraction sub-module will extract its own feature signal based on the pre-identification signal, where the pre-identification signal is a pre-processed input signal.
[0065] Taking communication interference signals as an example, when using the neural network system 10 for signal classification, each of the 13 feature extraction sub-modules 1012-1 to 1012-13 in the neural network system extracts the feature waveform of its target interference category based on the input communication interference signal. The parameters within each feature extraction sub-module are determined through training of the neural network system, thereby determining the extracted feature waveform.
[0066] Assume the input communication interference signal comprises signals from 273 Physical Resource Blocks (PRBs), and the signal amplitude corresponding to each PRB can be the average value of the interference noise power within one sampling period. In this case, the input interference signal can be represented as a 1*273 vector.
[0067] After each of the 13 feature extraction submodules 1012-1 to 1012-13 acquires the input communication interference signal, it extracts the feature waveform of the interference category targeted by that submodule. For example, feature extraction submodule 1012-1 only focuses on the noise power related to the waveform characteristics of smart street light interference, and it can extract the feature waveform of the input communication interference signal according to the trained feature extraction parameters for smart street light interference waveforms. Meanwhile, feature extraction submodule 1012-2 only focuses on the noise power related to the waveform characteristics of 800M telecommunications interference, and it can extract the feature waveform of the input communication interference signal according to the trained feature extraction parameters for 800M telecommunications interference.
[0068] In step S230, each of the 13 classification sub-modules 1021-1 to 1021-N in the neural network system determines whether the input communication interference signal contains a signal of a corresponding signal category based on the feature signals extracted by its corresponding feature extraction sub-module. For example, classification sub-module 1021-1 determines whether the feature waveform extracted by feature extraction sub-module 1012-1 corresponds to the waveform of smart street light interference based on the parameters trained in this sub-module to determine whether smart street light interference exists in the input communication interference signal. Classification sub-module 1021-2 determines whether the feature waveform extracted by feature extraction sub-module 1012-2 corresponds to the waveform of 800M telecommunications interference based on the parameters trained in this sub-module to determine whether 800M telecommunications interference exists in the input communication interference signal.
[0069] Once it is determined how many signal content categories the input signal needs to be divided into, the N values of the aforementioned feature extraction submodule and classification submodule are set to corresponding numbers. For example, if the input signal is an image signal and it is known that the image signal needs to be divided into 4 image categories, then N is set to 4. In this case, each of the N feature extraction submodules (i.e., the 4 feature extraction submodules) and each of the N classification submodules (i.e., the 4 classification submodules) mentioned above corresponds to one of these four image content categories.
[0070] Similarly, when the input signal is a video signal, and it is known that the video signal content needs to be divided into 3 categories, N is set to 3. In this case, each of the N feature extraction submodules (i.e., each of the 3 feature extraction submodules) and the N classification submodules (i.e., each of the 3 classification submodules) corresponds to one of the three video content categories. When the input signal may include an audio signal, and it is known that the video signal content needs to be divided into 3 categories, N is also set to 3. In this case, each of the N feature extraction submodules (i.e., each of the 3 feature extraction submodules) and the N classification submodules (i.e., each of the 3 classification submodules) corresponds to one of the three audio content categories. Based on this, those skilled in the art will readily conceive of adaptive adjustments for other signals that need to be classified (e.g., audio signals).
[0071] Figure 3 The training method of a neural network system 10 for signal classification according to an embodiment of the present disclosure is illustrated schematically.
[0072] The training method for the neural network system 10 used for signal classification may include the following steps. In step S300, the sample input signal S1 and the labeled signal category of the sample input signal are obtained. The signal categories included in the sample input signal can be labeled as: M1, M2, ..., MX, where X can be a natural number greater than or equal to 1. Taking communication interference signals as an example, the sample input signal can be the sample interference signal corresponding to 273 PRBs. The signal amplitude corresponding to each PRB is the average value of the interference noise power within one sampling period. The labeled signal category will include at least one type of interference, such as smart street light interference, or, for example, smart street light interference and jammer interference.
[0073] In step S320, each of the N feature extraction sub-modules extracts its own feature signal based on the sample input signal, and these feature signals can be labeled as C1, C2, ..., CN respectively.
[0074] In step S330, each of the N classification sub-modules determines whether the sample input signal contains a signal of a corresponding signal category based on the feature signal extracted by its corresponding feature extraction sub-module. The determination results of the N classification sub-modules can be sequentially labeled as D1, D2, ..., DN. As mentioned above, D1 to DN can be represented by values 0 or 1, probabilities, or other representations.
[0075] In step S340, a first loss is determined based on the first loss function between the determination results (i.e., D1, D2, ..., DN) of the signal categories of the N classification sub-modules for their respective corresponding signal categories and the labeled signal categories (i.e., M1, M2, ..., MX) of the sample input signal. D1, D2, ..., DN and M1, M2, ..., MX use the same labeling method. For example, assuming that D1 to DN can use values 0 or 1, when N is 13, the values of D1, D2, ..., D13 output in the first training can be represented as 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0. Taking the classification of communication interference signals as an example, for the case where smart street light interference is marked in the sample interference signal, the values of M1, M2, ..., MX should be 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0. The first loss function can be represented as L1(Dx,Mx), which measures the degree of inconsistency between the value Mx indicating the labeled or true signal category of the input signal and the neural network system's prediction value Dx for the signal category. During training, the value of L1 should be as small as possible. By adjusting the parameters of the neural network during training to keep this first loss within a specific first threshold, the output results of the N classification sub-modules can be made closer to the true or labeled signal category, thereby improving the accuracy of signal classification. Taking communication interference signals as an example, during training, the parameters of the neural network system need to be adjusted based on the labeled interference category so that the first loss between the results determined by the 13 classification sub-modules for each interference category and the labeled interference category of the input sample interference signal is less than a specific first threshold. In one embodiment, keeping the first loss within a specific first threshold can be set as one of the training termination conditions.
[0076] In step S350, the feature signals (i.e., C1, C2, ..., CN) extracted by the N feature extraction sub-modules based on the sample input signal S1 are summed, and a second loss is determined between the summed feature signal H1 and the sample input signal S1 based on a second loss function. In a more specific embodiment, the second loss function can be expressed as L2(H1,S1), which measures the degree of inconsistency between the sum of the feature signals H1 extracted by the N feature extraction sub-modules and the input signal S1. During training, the value of L2 should be as small as possible. By adjusting the parameters of the neural network during training to keep the second loss within a specific second threshold, the summation of the waveforms extracted by the N feature extraction sub-modules can more realistically reproduce the input signal, thereby improving the accuracy of signal classification. Taking communication interference signals as an example, the closer the sum of the feature waveforms extracted by the 13 feature extraction submodules for each interference category is to the waveform of the input interference signal, the higher the accuracy of feature extraction for the input signal. The second loss between the sum of the feature waveforms extracted by the 13 feature extraction submodules for each interference category and the waveform of the input interference signal is less than a specific first threshold can be set as another condition for the end of training.
[0077] The loss functions L1 and L2 mentioned above can take the form of mean squared error, mean absolute error, or other loss functions.
[0078] In summary, in the embodiments of this disclosure, the model parameters of the neural network system are adjusted based on a first loss and a second loss. The above steps can be repeated to adjust the model parameters of the neural network system until a termination condition is met. For example, the termination condition could be that the first loss is lower than a first threshold and the second loss is lower than a second threshold. Of course, the termination condition can also be set to other conditions as needed.
[0079] When the neural network system includes a feature pre-identification module, the above training method may optionally include S320, which involves generating a pre-identification signal P1 by performing feature pre-identification on the sample input signal S1 through the feature pre-identification module, and inputting the pre-identification signal P1 into each of the N feature extraction sub-modules. Correspondingly, the parameters in the feature pre-identification module will also be adjusted during training.
[0080] As mentioned earlier, in accordance with the number of signal categories included in the input signal to be classified, the neural network system 10 needs to be trained using a set of sample input signals labeled with signal categories before it can be used.
[0081] When the input signal is an image signal, a set of sample image signals labeled with their content categories is used during training. For example, this set could include sample image signals labeled with two categories: people images and animal images. When it is necessary to classify images into four categories—people images, animal images, landscape images, and other images—the four corresponding feature extraction submodules can be applied sequentially to people images, animal images, landscape images, or other images. In this case, the extracted feature signals are C1, C2, C3, and C4. Specifically, each feature signal can be a pixel value associated with the corresponding category. The feature signals can be set to have the same format as the input image signal. For example, C1 can be a pixel value associated with people images, in which case pixel values unrelated to people features will be set to 0. When determining the second loss, C1, C2, C3, and C4 are summed to obtain the summed feature signal H1. It should be noted that the extracted feature image signals have the same signal format as the input image signal or are adjusted to have the same signal format as the input image signal, so that the sum of the extracted feature image signals can be compared with the input image signal.
[0082] When the input signal is an audio signal, a set of sample audio signals labeled with their audio content categories is used during training. For example, this set may include sample audio signals labeled as including music signals. The audio signal can be a time-domain signal, such as a segment of audio signal in the time dimension, or a frequency-domain signal, such as an audio signal transformed from the frequency domain; the specific form is not specifically limited here. When it is necessary to classify audio into three categories—speech signal, music signal, and other noise signal—the corresponding three feature extraction submodules can target speech signal, music signal, and other noise signal sequentially. The extracted feature signals are C1, C2, and C3. Specifically, each feature signal can be a feature audio signal associated with its corresponding category. The feature signals can be set to have the same format as the input audio signal; for example, C1 can be an audio signal associated with the speech signal, in which case the values of audio signals unrelated to speech will be set to 0. When determining the second loss, C1, C2, and C3 are summed to obtain the summed feature signal H1. It should be mentioned that the extracted feature audio signals have the same signal format as the input audio signal or are adjusted to have the same signal format as the input audio signal, so that the sum of the extracted feature audio signals can be compared with the input audio signal.
[0083] When the input signal is a video signal, a set of sample image signals labeled with their video content categories is used during training. For example, this may include sample image signals labeled as including moving images. The video signal can be a video signal over a period of time, which can be considered as a three-dimensional matrix, where one dimension can be time, and the other two dimensions can be the frame height and frame width of the video frame at the corresponding time point. However, the specific form of the video signal is not specifically limited here. When it is necessary to classify the video into two video categories, still images and moving images, the two feature extraction submodules can target the still image signal and the moving image signal respectively. At this time, the extracted feature signals are C1 and C2. Specifically, each feature signal can be a feature video signal related to the corresponding category. The feature signals can be set to have the same format as the input video signal. For example, C1 can be a video signal related to still image signals, and video signals unrelated to still image signals will be set to 0. When determining the second loss, C1 and C2 are summed to obtain the summed feature signal H1. It should be mentioned that the extracted feature video signals have the same signal format as the input video signal, or the summation will be adjusted to have the same signal format as the input video signal, so that the sum of the extracted feature video signals can be compared with the input video signal.
[0084] Based on this, those skilled in the art will readily conceive of making adaptive adjustments to other signals that need to be classified when training a neural network system.
[0085] Figures 1 to 3 The following describes the structure of a neural network system according to embodiments of the present disclosure, a method for performing classification using the neural network system, and a method for training the neural network system, which is applicable to various signals requiring classification. The following will combine... Figure 4 This section describes an embodiment of classifying communication interference signals using a neural network system according to embodiments of the present disclosure. In this embodiment, the same content as described above will not be repeated.
[0086] Figure 4 A block diagram of a neural network system for classifying communication interference signals according to an embodiment of the present disclosure is shown schematically.
[0087] As mentioned earlier, in the example case where the communication interference signal includes 13 types of interference, the N feature extraction submodules and N classification submodules in the neural network system can be specifically implemented as 13 feature extraction submodules 1012-1 to 1012-13 and 13 classification submodules 1021-1 to 1021-13, respectively. Furthermore, each of the 13 feature extraction submodules 1012-1 to 1012-13 includes a first linear layer (…). Figure 4 The text is simply shown as "linear" and the first nonlinear layer (…). Figure 4This is implemented using ReLU, where ReLU stands for "Rectified Linear Unit". Furthermore, each of the 13 classification sub-modules 1021-1 to 1021-13 includes a second linear layer. Figure 4 It is also simply shown as "linear" in the middle) and the second nonlinear layer (in Figure 4 (This is achieved using the Sigmoid activation function). Using at least two linear layers and classifying each signal category individually will make the neural network model more lightweight while maintaining recognition performance.
[0088] Each of the 13 feature extraction submodules 1012-1 to 1012-13 and the 13 classification submodules 1021-1 to 1021-13 corresponds to a specific interference category. For example, the first feature extraction submodule 1012-1 and the first classification submodule 1021-1 can target the smart street light interference type. The first feature extraction submodule extracts feature waveforms for this type of interference, and the first classification submodule identifies and classifies the input signal based on the feature waveforms extracted by the first feature extraction submodule, determining whether the input signal exhibits this type of interference. Similarly, the second feature extraction submodule 1012-2 and the second classification submodule 1021-2 can target the 800M telecommunications interference type. The second feature extraction submodule extracts feature waveforms for this type of interference, and the second classification submodule 1021-2 identifies and classifies the input signal based on the feature waveforms extracted by the second feature extraction submodule 1012-2, determining whether the input signal exhibits this type of interference. More specifically, for 800MHz telecommunications interference, the frequency range is 824-849 / 869-894MHz. Therefore, when using the second feature extraction submodule 1012-2 to extract feature waveforms, waveforms within this frequency range will be extracted. When using the second classification submodule 1012-2 to identify the extracted waveforms, it will determine whether a waveform exists within the 824-849 / 869-894MHz frequency range. If it exists, it will be determined that 800MHz telecommunications interference exists; otherwise, it will be determined that 800MHz telecommunications interference does not exist.
[0089] As mentioned earlier, this neural network system needs to be trained before it can be used for interference classification.
[0090] During the training of the neural network system, it is first necessary to obtain a set of sample interference signals and their corresponding labeled interference categories (for example, as mentioned above, the sample interference signals can be the sample interference signals corresponding to 273 PRBs, the signal amplitude of each PRB is the average value of the interference noise power within one sampling period, and the labeled signal category includes at least one type of interference, such as smart street light interference). Specifically, the sample interference signals on PRBs 0 to 272 need to be input into the neural network system. Then, the first linear layer of each feature extraction submodule is used to extract the feature waveforms of each interference category in the input sample interference signals. Next, the second linear layer of each classification submodule is used to classify each extracted waveform. More specifically, for 800MHz telecommunications interference, when using the second feature extraction submodule 1012-2 to extract feature waveforms, the waveforms of PRBs in the frequency range of 824-849 / 869-894MHz will be extracted. At this time, the waveforms or signal amplitudes on other PRBs can be set to 0 so that the feature signals extracted by the feature extraction submodule have the same format as the input signal. For example, the input signal is a one-dimensional vector data, and the feature signal is also a one-dimensional vector data, with both having the same number of vector elements. When using the second classification submodule 1012-2 to identify the extracted waveform, it will determine whether a waveform exists in the frequency range of 824-849 / 869-894MHz, thereby determining whether there is 800MHz telecommunications interference in the sample interference signal. This is just a simple example; in reality, the model parameters used for feature extraction and classification by each feature extraction submodule and each classification submodule will be continuously optimized during training.
[0091] Next, based on the first loss function L1(Dx,Mx), a first loss is calculated between the determination results of all classification sub-templates for interference types (labeled as D1, D2, ..., D13) and the labeled interference categories of the sample interference signals (labeled as M1, M2, ..., M13). A second loss is calculated based on the second loss function L2(H1,S1) between the sum of the feature waveforms extracted by all feature extraction sub-modules (H1) and the waveform S1 of the sample interference signals. The parameters of the neural network system are then adjusted based on the first and second losses. In a more specific embodiment, the training termination condition can be based on the first loss being less than a first threshold and the second loss being less than a second threshold. Specific training methods can be found in the section on... Figure 3 The description of that will not be repeated here.
[0092] Thus, a trained neural network system can be obtained, which can be used to classify communication interference signals.
[0093] In one embodiment, the communication interference signal to be classified (e.g., interference noise on PRBs 0 to 272 to be classified) is first input into the neural network system. Then, each of the 13 feature extraction sub-modules extracts the feature waveforms of each interference category in the input communication interference signal. Next, each of the 13 classification sub-modules classifies each extracted feature waveform. More specifically, each classification sub-module determines whether the sample interference signal contains a signal of its corresponding interference signal category based on the feature waveforms extracted by its corresponding feature extraction sub-module. For example, when the output of the classification sub-module is represented as a value of 1 or 0, if the determination result of the first classification sub-module 1021-1 (for smart street light interference) and the second classification sub-module 1021-2 (for 800M telecommunications interference) is 1 while the determination result of the remaining classification sub-modules is 0, it can be seen from the above embodiment that the communication interference signal contains smart street light interference and 800M telecommunications interference, but no other interference signals.
[0094] According to embodiments of the present disclosure, an apparatus for signal classification is also provided, comprising: a processor; and a memory storing one or more computer programs; wherein, when the one or more computer programs are executed by the processor, the processor causes the processor to perform any one of the methods for signal classification as described in the various embodiments above. According to embodiments of the present disclosure, a computer-readable storage medium or computer program product storing a computer program may also be provided, wherein, when the computer program is executed by at least one processor, it causes the at least one processor to perform any one of the methods described above according to exemplary embodiments of the present disclosure. Examples of computer-readable storage media or computer program products herein include: read-only memory (ROM), random access programmable read-only memory (PROM), electrically erasable programmable read-only memory (EEPROM), random access memory (RAM), dynamic random access memory (DRAM), static random access memory (SRAM), flash memory, non-volatile memory, CD-ROM, CD-R, CD+R, CD-RW, CD+RW, DVD-ROM, DVD-R, DVD+R, DVD-RW, DVD+RW, DVD-RAM, BD-ROM, BD-R, BD-R LTH, BD-RE, Blu-ray or optical disc storage, hard disk drive (HDD), solid-state drive (SSD), card storage (such as multimedia cards, secure digital (SD) cards, or ultra-fast digital (XD) cards), magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, and any other device configured to store a computer program and any associated data, data files, and data structures in a non-transitory manner and to provide the computer program and any associated data, data files, and data structures to a processor or computer so that the processor or computer can execute the computer program. The instructions or computer program in the aforementioned computer-readable storage medium or computer program product can run in an environment deployed in computer devices such as clients, hosts, agent devices, servers, etc. Furthermore, in one example, the computer program and any associated data, data files, and data structures are distributed across a networked computer system, such that the computer program and any associated data, data files, and data structures are stored, accessed, and executed in a distributed manner through one or more processors or computers.
[0095] Those skilled in the art will understand that the illustrative embodiments described above are not intended to be limiting. It should be understood that any two or more of the embodiments disclosed herein can be combined in any combination. Furthermore, other embodiments may be utilized and other changes may be made without departing from the spirit and scope of the subject matter presented herein. It will be readily understood that aspects of the invention disclosed herein, as generally described herein and illustrated in the accompanying drawings, can be arranged, substituted, combined, separated, and designed in a variety of different configurations, all of which are contemplated herein.
[0096] Those skilled in the art will understand that the various illustrative logic blocks, modules, circuits, and steps described herein can be implemented in hardware, software, or a combination of both. To clearly illustrate this interchangeability between hardware and software, the various illustrative components, blocks, modules, circuits, and steps are described above in the form of sets of functions. Whether such sets of functions are implemented in hardware or software depends on the specific application and the design constraints imposed on the overall system. Those skilled in the art may implement the described sets of functions in different ways for each specific application, but such design decisions should not be construed as departing from the scope of this application.
[0097] The various illustrative logic blocks, modules, and circuits described in this application may be implemented or performed using a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. The general-purpose processor may be a microprocessor, but in alternatives, the processor may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors cooperating with a DSP core, or any other such configuration.
[0098] The steps of the methods or algorithms described in this application may be embodied directly in hardware, in a software module executed by a processor, or in a combination of both. The software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, removable disk, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor so that the processor can read and write information to / from the storage medium. In an alternative, the storage medium may be integrated into the processor. The processor and storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In an alternative, the processor and storage medium may reside as discrete components in the user terminal.
[0099] In one or more exemplary designs, the functionality may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software, the functionality may be stored or transmitted as one or more instructions or code on a computer-readable medium. Computer-readable media includes both computer storage media and communication media, the latter including any medium that facilitates the transfer of a computer program from one location to another. Storage media may be any available medium that can be accessed by a general-purpose or special-purpose computer.
[0100] The above description is merely an exemplary embodiment of this application and is not intended to limit the scope of protection of this application. The scope of protection of this application is determined by the appended claims.
Claims
1. A signal classification method based on neural networks, wherein... The neural network includes a feature extraction module and a classification module; The feature extraction module includes multiple parallel feature extraction sub-modules; The classification module includes multiple parallel classification sub-modules corresponding to each of the multiple feature extraction sub-modules; Each of the multiple feature extraction submodules and each of the multiple classification submodules corresponds to a signal category; and The method includes: Each of the multiple feature extraction submodules extracts its own feature signal based on the input signal; as well as Each of the multiple classification sub-modules determines whether the input signal contains a signal of a corresponding signal category based on the feature signals extracted by its respective feature extraction sub-module.
2. The method as described in claim 1, wherein, Each of the plurality of feature extraction sub-modules includes a first linear layer and a first nonlinear layer, and Each of the plurality of classification sub-modules includes a second linear layer and a second nonlinear layer.
3. The method as described in claim 2, wherein, The first nonlinear layer is implemented using the ReLU activation function, and the second nonlinear layer is implemented using the Sigmoid activation function.
4. The method as described in any one of claims 1 to 3, wherein, The neural network is trained via the following steps: Each of the multiple feature extraction submodules extracts its own feature signal based on the sample input signal; Each of the multiple classification sub-modules determines whether the sample input signal contains a signal of a corresponding signal category based on the feature signal extracted by its respective feature extraction sub-module; The first loss is determined based on the first loss function between the determination results of the multiple classification sub-modules for their respective corresponding signal categories and the labeled signal category of the sample input signal. The feature signals extracted by the multiple feature extraction submodules based on the sample input signal are summed, and a second loss is determined between the summed feature signals and the sample input signal based on a second loss function; and The model parameters of the neural network are adjusted based on the first loss and the second loss.
5. The method according to any one of claims 1 to 3, wherein the feature extraction module further comprises a feature pre-identification module, and the method further comprises: The feature pre-identification module performs feature pre-identification on the input signal to generate a pre-identification signal, and inputs the pre-identification signal to each of the plurality of feature extraction sub-modules.
6. The method as described in any one of claims 1 to 3, wherein, The input signal includes a communication interference signal, and the signal category includes a communication interference category; The input signal includes an image signal, and the signal category includes an image content category; The input signal includes a video signal, and the signal category includes a video content category; or The input signal includes an audio signal, and the signal category includes an audio content category.
7. A neural network system for signal classification, comprising: The feature extraction module includes multiple parallel feature extraction sub-modules; as well as The classification module includes multiple parallel classification sub-modules corresponding to each of the multiple feature extraction sub-modules. Each feature extraction submodule and each classification submodule in the plurality of feature extraction submodules corresponds to a signal category. Each of the plurality of feature extraction submodules is configured to extract its respective feature signal based on the input signal, and Each of the multiple classification sub-modules is configured to determine whether the input signal contains a signal of a corresponding signal category based on the feature signals extracted by its respective feature extraction sub-module.
8. A training method for a neural network for signal classification, the neural network comprising a feature extraction module and a classification module, the feature extraction module comprising multiple parallel feature extraction sub-modules, the classification module comprising multiple parallel classification sub-modules corresponding to each of the multiple feature extraction sub-modules, each feature extraction sub-module and each classification sub-module corresponding to a signal category, the method comprising: Acquire the sample input signal and the label signal category of the sample input signal; Each of the multiple feature extraction submodules extracts its own feature signal based on the sample input signal; Each of the multiple classification sub-modules determines whether the sample input signal contains a signal of a corresponding signal category based on the feature signal extracted by its respective feature extraction sub-module; The first loss is determined based on the first loss function between the determination results of the multiple classification sub-modules for their respective corresponding signal categories and the labeled signal category of the sample input signal. The feature signals extracted by the multiple feature extraction submodules based on the sample input signal are summed, and a second loss is determined between the summed feature signals and the sample input signal based on a second loss function; and The model parameters of the neural network are adjusted based on the first loss and the second loss.
9. An apparatus for a neural network, comprising: processor; and Memory, which stores one or more computer programs; Wherein, when the one or more computer programs are run by the processor, the processor causes the processor to perform the method according to any one of claims 1 to 6 and 8.
10. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1 to 6 and 8.