Neural network-based signal classification method, training method and apparatus

Abstract

To provide a method, a system, a device and a program for effectively performing signal classification using an artificial intelligence neural network.SOLUTION: A signal classification method based on a neural network, wherein the neural network system comprises a feature extraction module and a classification module, the feature extraction module comprises a plurality of parallel feature extraction sub-modules, the classification module comprises a plurality of parallel classification sub-modules respectively corresponding to the plurality of feature extraction sub-modules, each feature extraction sub-module and each classification sub-module correspond to a signal category, each feature extraction sub-module extracts a respective feature signal according to an input signal, and each classification sub-module respectively determines whether the input signal contains a signal of a corresponding signal category according to the feature signal extracted by the feature extraction sub-module corresponding to the classification sub-module.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to the field of signal classification, and more particularly to a method and apparatus for signal classification based on neural networks, a method and apparatus for training a neural network for signal classification, and a neural network system for signal classification. [Background technology]

[0002] In the signal processing process, the same signal often contains multiple signal categories. For example, an interference signal in a communication system may contain multiple types of interference (e.g., intelligent street light interference, telecom 800M interference, etc.). For example, an image signal may contain multiple categories of images. For example, an image signal may contain portraits, animals, landscapes, etc. as content objects. For example, a traffic image signal may contain red images, green images, and yellow images. For example, an audio signal may contain multiple categories of audio, such as music signals and voice signals. In this case, a method for classifying signals is needed. Summary of the Invention [Problem to be solved by the invention]

[0003] The present disclosure aims to provide a method and a neural network system for effectively performing signal classification using an artificial intelligence neural network. [Means for solving the problem]

[0004] According to one aspect of the present disclosure, there is provided a signal classification method based on a neural network, the neural network including a feature extraction module and a classification module, the feature extraction module including a plurality of parallel feature extraction sub-modules, the classification module including a plurality of parallel classification sub-modules respectively corresponding to the plurality of feature extraction sub-modules, each feature extraction sub-module of the plurality of feature extraction sub-modules and each classification sub-module of the plurality of classification sub-modules corresponding to one signal category, the method including: each feature extraction sub-module of the plurality of feature extraction sub-modules extracting a respective feature signal based on an input signal; and each classification sub-module of the plurality of classification sub-modules determining, based on the feature signal extracted by its corresponding feature extraction sub-module, whether the input signal includes a signal of the corresponding signal category.

[0005] In one embodiment, each feature extraction sub-module of the plurality of feature extraction sub-modules includes a first linear layer and a first non-linear layer, and each classification sub-module of the plurality of classification sub-modules includes a second linear layer and a second non-linear layer.

[0006] In one embodiment, the first nonlinear layer is realized by a Relu activation function and the second nonlinear layer is realized by a Sigmoid activation function.

[0007] In one embodiment, the neural network is trained through the following steps: each feature extraction submodule of the plurality of feature extraction submodules extracts a respective feature signal based on a sample input signal; each classification submodule of the plurality of classification submodules determines, based on the feature signal extracted by its corresponding feature extraction submodule, whether a signal of a corresponding signal category is included in the sample input signal; determining, based on a first loss function, a first loss between the determination results of the plurality of classification submodules for the corresponding signal category and a label signal category of the sample input signal; adding up the feature signals extracted by the plurality of feature extraction submodules based on the sample input signal and determining, based on a second loss function, a second loss between the added feature signal and the sample input signal; and adjusting model parameters of the neural network based on the first loss and the second loss.

[0008] In one embodiment, the feature extraction module further includes a feature pre-recognition module, and the method further includes the feature pre-recognition module performing feature pre-recognition on the input signal to generate a pre-recognition signal, and inputting the pre-recognition signal to each feature extraction sub-module of the plurality of feature extraction sub-modules.

[0009] In one embodiment, the input signal comprises a communications interference signal and the signal category comprises a communications interference category, or the input signal comprises an image signal and the signal category comprises an image content category, or the input signal comprises a video signal and the signal category comprises a video content category, or the input signal comprises an audio signal and the signal category comprises an audio content category.

[0010] According to another aspect of the present disclosure, there is provided a neural network system for signal classification, including: a feature extraction module including a plurality of parallel feature extraction submodules; and a classification module including a plurality of parallel classification submodules respectively corresponding to the plurality of feature extraction submodules, wherein each feature extraction submodule of the plurality of feature extraction submodules and each classification submodule of the plurality of classification submodules corresponds to one signal category, each feature extraction submodule of the plurality of feature extraction submodules is configured to extract a respective feature signal based on an input signal, and each classification submodule of the plurality of classification submodules is configured to determine, based on the feature signal extracted by its corresponding feature extraction submodule, whether the input signal contains a signal of a corresponding signal category.

[0011] In one embodiment, each feature extraction sub-module of the plurality of feature extraction sub-modules includes a first linear layer and a first non-linear layer, and each classification sub-module of the plurality of classification sub-modules includes a second linear layer and a second non-linear layer.

[0012] In one embodiment, the first nonlinear layer is realized by a Relu activation function and the second nonlinear layer is realized by a Sigmoid activation function.

[0013] In one embodiment, the neural network system is trained through the following steps: each feature extraction submodule of the plurality of feature extraction submodules extracts a respective feature signal based on a sample input signal; each classification submodule of the plurality of classification submodules determines, based on the feature signal extracted by its corresponding feature extraction submodule, whether the sample input signal contains a signal of a corresponding signal category; determining, based on a first loss function, a first loss between the determination results of the plurality of classification submodules for the corresponding signal category and a label signal category of the sample input signal; adding together the feature signals extracted by the plurality of feature extraction submodules based on the sample input signal and determining, based on a second loss function, a second loss between the added feature signal and the sample input signal; and adjusting model parameters of the neural network based on the first loss and the second loss.

[0014] In one embodiment, the feature extraction module further includes a feature pre-recognition module configured to perform feature pre-recognition on the input signal to generate a pre-recognition signal, and to input the pre-recognition signal to each feature extraction sub-module of the plurality of feature extraction sub-modules.

[0015] In one embodiment, the input signal comprises a communications interference signal and the signal category comprises a communications interference category, or the input signal comprises an image signal and the signal category comprises an image content category, or the input signal comprises a video signal and the signal category comprises a video content category, or the input signal comprises an audio signal and the signal category comprises an audio content category.

[0016] According to another aspect of the present disclosure, there is provided a method for training a neural network for signal classification, the neural network including a feature extraction module and a classification module, the feature extraction module including a plurality of parallel feature extraction sub-modules, the classification module including a plurality of parallel classification sub-modules respectively corresponding to the plurality of feature extraction sub-modules, each feature extraction sub-module of the plurality of feature extraction sub-modules and each classification sub-module of the plurality of classification sub-modules corresponding to one signal category, the method including: obtaining sample input signals and label signal categories of the sample input signals; each feature extraction sub-module of the plurality of feature extraction sub-modules extracting a respective feature signal based on the sample input signals; The method includes: each classification sub-module of a class sub-module respectively determining, based on feature signals extracted by a corresponding feature extraction sub-module, whether a signal of a corresponding signal category is included in the sample input signal; determining, based on a first loss function, a first loss between the determination results of the classification sub-modules for the corresponding signal category and a label signal category of the sample input signal; adding together the feature signals extracted by the feature extraction sub-modules based on the sample input signal, and determining, based on a second loss function, a second loss between the added feature signal and the sample input signal; and adjusting 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-recognition module, and the method further includes the feature pre-recognition module being configured to perform feature pre-recognition on the sample input signal to generate a sample pre-recognition signal, and input the sample pre-recognition signal to each feature extraction sub-module of a plurality of feature extraction sub-modules.

[0018] In one embodiment, each feature extraction sub-module of the plurality of feature extraction sub-modules includes a first linear layer and a first non-linear layer, and each classification sub-module of the plurality of classification sub-modules includes a second linear layer and a second non-linear layer.

[0019] In one embodiment, the first nonlinear layer is realized by a Relu activation function and the second nonlinear layer is realized by a Sigmoid activation function.

[0020] In one embodiment, the sample input signal comprises a communications interference signal and the signal category comprises a communications interference category, or the sample input signal comprises an image signal and the signal category comprises an image content category, or the sample input signal comprises a video signal and the signal category comprises a video content category, or the sample input signal comprises an audio signal and the signal category comprises an audio content category.

[0021] According to another aspect of the present disclosure, there is provided an apparatus for signal classification, comprising a processor and a memory having stored thereon one or more computer programs, the one or more computer programs, when executed by the processor, causing the processor to perform the method described above.

[0022] According to another aspect of the present disclosure, there is provided a computer program product comprising a computer program, said computer program being adapted to implement the method described above when executed by a processor. [Effects of the Invention]

[0023] According to the signal classification method of the present disclosure, a neural network model is configured with corresponding sub-models based on possible categories of the signal to be classified, thereby achieving effective classification of the signal. In some embodiments, highly accurate signal classification can be achieved using a lightweight neural network model. [Brief explanation of the drawings]

[0024] [Figure 1] FIG. 1 illustrates an exemplary neural network system for signal classification according to an embodiment of the present disclosure. [Figure 2] FIG. 2 illustrates an exemplary method for signal classification according to an embodiment of the present disclosure. [Figure 3] FIG. 3 illustrates an exemplary method for training a neural network system for signal classification according to an embodiment of the present disclosure. [Figure 4] FIG. 4 illustrates an exemplary block diagram of a neural network system for classifying communication interference signals in accordance with an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0025]

[0023] Hereinafter, embodiments of the present disclosure will be described in more detail with reference to the accompanying drawings. Although several embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be realized in various forms and should not be construed as being limited to the embodiments described herein, but rather that these embodiments are provided to enable a more thorough and complete understanding of the present disclosure. It should be understood that the accompanying drawings and embodiments of the present disclosure are for illustrative purposes only and are not intended to limit the scope of protection sought for the present disclosure.

[0026] The terms "comprises" and "may include" as used herein should be understood to refer to the presence of features, operations, or components that can be used in each embodiment of the present disclosure, and should not be interpreted to limit the presence of one or more additional features, operations, or features. Furthermore, the terms "comprises" and "having" may be interpreted to represent certain features, numbers, steps, operations, elements, components, or combinations thereof, but should not be interpreted to exclude the possible presence of one or more other features, numbers, steps, operations, elements, components, or combinations thereof.

[0027] The concepts of "first," "second," etc. referred to in this disclosure are merely intended to distinguish between different devices, modules, units, models, data, etc., and are not intended to limit the order in which functions are performed, the order in which they are generated, or the dependencies between these devices, modules, units, models, and data.

[0028] It should be noted that the modifications "a" and "a plurality" referred to in this disclosure are intended to be illustrative rather than limiting, and those skilled in the art should understand that they should be understood as "one or more" unless the context clearly dictates otherwise.

[0029] As used herein, any reference to "one example," "example," "one 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 appearances of the phrase "in one embodiment" or "in one example" in different places in the specification do not necessarily all refer to the same embodiment.

[0030] Furthermore, words such as "comprise" or "comprise" are understood to mean the inclusion of the listed elements or objects, and equivalents thereof, and not the exclusion of other elements or objects.

[0031] The various examples discussed below, which are intended to illustrate the principles of the present disclosure in this patent document, are merely illustrative and should not be construed as limiting the scope of the present disclosure in any way.

[0032] The following description of the referenced drawings is provided to facilitate a complete understanding of various embodiments of the present disclosure, as defined by the claims and their equivalents. While the present description includes numerous specific details for ease of understanding, they should be considered merely as examples. Therefore, those skilled in the art will recognize that changes and modifications can be made to the various embodiments described herein without departing from the scope and spirit of the present disclosure. Furthermore, descriptions of well-known functions and configurations may be omitted for clarity and conciseness.

[0033] The terms and expressions used in the following specification and claims are not limited to their dictionary meanings, but are merely used to enable the inventor to clearly and consistently understand the present disclosure. Therefore, it should be apparent to those skilled in the art that the description of various embodiments of the present disclosure is provided for illustrative purposes only, and does not limit the scope of the present disclosure, which is defined by the appended claims and their equivalents.

[0034] The present disclosure provides a neural network system for signal classification, a method and apparatus for neural network-based signal classification, and a method and apparatus for training a neural network for signal classification. In the neural network system of the present disclosure, effective signal recognition is achieved by extracting features and performing classification for each of a plurality of signal categories one by one, rather than providing recognition results for all signal categories at once. Furthermore, in the process of training the neural network system, signals obtained by extracting features for each of a plurality of signal categories are added, the added signal is compared with an input signal, and the comparison results are used to adjust the parameters of the neural network system, thereby improving the efficiency and success rate of model training and the accuracy of signal category recognition. In a further embodiment, linear layers with low computational complexity are used in the feature extraction module and classification module, thereby making the neural network system of the present disclosure more lightweight.

[0035] The signals referred to in the present disclosure may be various signals with a mixture of multiple categories, such as a communication interference signal having at least one interference type, an image signal having at least one image content category (e.g., classifying images into multiple image content categories in a manner such as target object category, texture category, or color category), a video signal having at least one video content category (e.g., classifying into still images and dynamic images), and an audio signal having at least one audio content category (e.g., classifying into voice signals, music signals, or other noise signals). The manner of acquiring the above-mentioned signals is not specifically limited in the present disclosure.

[0036] FIG. 1 illustrates a neural network system 10 for signal classification according to an embodiment of the present disclosure.

[0037] The neural network system 10 may include a feature extraction module 101 for extracting feature signals from an input signal to be recognized, and a classification module 102 for recognizing and classifying the extracted feature signals. Rather than performing a single direct feature recognition to provide a classification result, the neural network system 10 performs feature extraction and classification for each possible type of signal in two steps, feature extraction and feature recognition, thereby expressing the task of signal classification in a Chain-of-Thought (CoT)-like format, thereby improving the accuracy of signal classification.

[0038] The 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 designated with reference numbers 1012-1, 1012-2, and so on up to 1012-N in FIG. 1). The classification module 102 may include N classification sub-modules (N is designated with reference numbers 1021-1, 1021-2, and so on up to 1021-N in FIG. 1), where N is the total number of signal categories that can exist in the input signal to be recognized.

[0039] The N feature extraction sub-modules and the N classification sub-modules correspond to each other in a one-to-one correspondence. Each of the N feature extraction sub-modules and each of the N classification sub-modules corresponds to one signal category. For example, the feature extraction sub-module 1012-1 and the classification sub-module 1021-1 correspond to signal category 1, meaning that the processing of these two sub-modules focuses on signal category 1, and the feature extraction sub-module 1012-2 and the classification sub-module 1021-2 correspond to signal category 2. The same applies to the subsequent feature extraction sub-modules and classification sub-modules.

[0040] Each of the N feature extraction submodules is configured to extract a respective 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 signal extracted by its corresponding feature extraction submodule. The feature signals extracted by each feature extraction submodule may have the same format as the input signal. For example, if the input signal is one-dimensional vector data, the feature signals may also be one-dimensional vector data, with the number of elements in both vectors being the same. In an alternative embodiment, the feature signals extracted by each feature extraction submodule may have a different format from the input signal. In this case, they need to be adjusted to the same signal format as the input signal before addition, so that the sum of the extracted feature signals can be compared with the input signal. For example, if the feature signals have a small number of elements, the number of vector elements in the feature signals may be padded to the same format as the input signal using zero padding.

[0041] Taking communication interference signals as an example, in a communication system, interference categories may include, for example, telecom 800M interference, false base station interference, NR system interference, LTE D band interference, 700M interference, intelligent street light interference, wireless bridge / surveillance camera interference, jammer interference, router interference, clock source failure interference, broadcast signal interference, external interference, and other interference categories. That is, in the above example, communication interference signals may include 13 types of interference categories.

[0042] In view of this, when classifying communication interference signals, the above-mentioned N may be set to 13, that is, the neural network system 10 may include 13 feature extraction sub-modules and 13 classification sub-modules.

[0043] As described above, the 13 feature extraction sub-modules and the 13 classification sub-modules may respectively correspond to the above-mentioned 13 interference categories. For example, the feature extraction sub-module 1012-1 and the classification sub-module 1021-1 may correspond to the intelligent street light interference category among the above-mentioned 13 interference categories, the feature extraction sub-module 1012-1 may extract feature waveforms for the interference category intelligent street light interference, and the classification sub-module 1021-1 may recognize and classify the feature waveforms extracted by the feature extraction sub-module 1012-1 to determine whether the interference category intelligent street light interference exists in the input communication interference signal. In addition, the feature extraction sub-module 1012-2 and the classification sub-module 1021-2 correspond to Telecom 800M interference among the 13 types of interference categories mentioned above, that is, the processing of these two sub-modules focuses on signal category 2, and the classification sub-module 1021-2 may determine whether the interference category of Telecom 800M interference exists in the input communication interference signal.

[0044] In a more specific embodiment, the above-mentioned determination result may be a value of 0 or a value of 1. Here, the value 0 may indicate that a signal of the corresponding signal category does not exist in the input signal, and the value 1 may indicate that a signal of the corresponding signal category exists in the input signal. Taking the classification sub-module 1021-1 for intelligent street light interference as an example, a value of 1 output by the classification sub-module 1021-1 represents that intelligent street light interference exists in the input communication interference signal, and a value of 0 output by the classification sub-module 1021-1 represents that intelligent street light interference does not exist in the input communication interference signal.

[0045] The determination result may also be expressed in other ways, such as in the form of probability. For example, for the classification sub-module 1021-1, a probability greater than 50% represents that intelligent street light interference exists in the input communication interference signal, and a probability less than or equal to 50% represents that intelligent street light interference does not exist in the input communication interference signal. Correspondingly, other classification sub-modules (e.g., classification sub-module 1021-2) may also use a value of 1 or a value of 0, or a probability greater than 50% or less than or equal to 50% to determine whether Telecom 800M interference exists in the input communication interference signal.

[0046] Taking an image signal as an example, various methods for classifying an image may be used and may be configured according to specific needs. For example, the image signal may be classified according to the category of a target object of interest in the image signal, its color category, or its texture category. If the image signal to be classified is to be classified into a person image, an animal image, a landscape image, or other images, the above-mentioned N may be set to 4. That is, the neural network system 10 may include four feature extraction sub-modules and four classification sub-modules.

[0047] These four feature extraction sub-modules and four classification sub-modules may correspond to the four image content types described above, respectively. For example, the feature extraction sub-module 1012-1 and the classification sub-module 1021-1 may correspond to the portrait image type of the four image content types described above, where the feature extraction sub-module 1012-1 extracts feature signals for the portrait image content type, and the classification sub-module 1021-1 recognizes and classifies the feature signals extracted by the feature extraction sub-module 1012-1 to determine whether the portrait image type is present in the input image signal. Also, the feature extraction sub-module 1012-2 and the classification sub-module 1021-2 may correspond to the animal image type of the four image content types described above, where the processing of these two sub-modules focuses on the animal image category, and the classification sub-module 1021-2 may determine whether the animal image is present in the input image signal.

[0048] In a more specific embodiment, the above-mentioned determination result may be a value of 0 or a value of 1. Here, the value 0 may indicate that a signal of the corresponding signal category is not present in the input signal, and the value 1 may indicate that a signal of the corresponding signal category is present in the input signal. Taking the classification sub-module 1021-1 for a person image type as an example, a value of 1 output by the classification sub-module 1021-1 indicates that a person image is present in the input image signal, and a value of 0 output by the classification sub-module 1021-1 indicates that a person image is not present in the input image signal. The determination result may also be expressed in other ways, such as in the form of a probability. For example, for the classification sub-module 1021-1 for a person image type, a probability greater than 50% indicates that a person image type is present in the input image signal, and a probability less than 50% indicates that a person image type is not present in the input image signal. Correspondingly, other classification sub-modules (e.g., classification sub-module 1021-2) may also use a value of 1 or a value of 0, or a probability greater than or less than 50%, to determine whether an animal image is present in the input image signal.

[0049] Taking audio signals as an example, there are various ways to classify audio signals, and audio signals may be classified based on the specific content contained in the audio signals or the characteristics of the signals themselves according to specific needs. If the audio signal to be classified is to be classified into a voice signal, a music signal, or other noise signals, the above-mentioned N is set to 3. That is, the neural network system 10 may include three feature extraction sub-modules and three classification sub-modules.

[0050] These three feature extraction sub-modules and three classification sub-modules may correspond to the three audio content types described above, respectively. For example, the feature extraction sub-module 1012-1 and the classification sub-module 1021-1 may correspond to a music signal type among the three audio content types described above, where the feature extraction sub-module 1012-1 extracts feature signals for the music signal content type, and the classification sub-module 1021-1 performs recognition and classification based on the feature signals extracted by the feature extraction sub-module 1012-1 to determine whether a music signal type is present in the input audio signal. Also, the feature extraction sub-module 1012-2 and the classification sub-module 1021-2 may correspond to a voice signal among the three audio content types described above. That is, the processing of these two sub-modules may focus on the voice signal category, and the classification sub-module 1021-2 may determine whether a voice signal is present in the input audio signal.

[0051] In a more specific embodiment, the above-mentioned determination result may be a value of 0 or a value of 1. Here, the value 0 may indicate that a signal of the corresponding signal category is not present in the input signal, and the value 1 may indicate that a signal of the corresponding signal category is present in the input signal. Taking the classification sub-module 1021-1 for a music signal type as an example, a value of 1 output by the classification sub-module 1021-1 indicates that a music signal is present in the input audio signal, and a value of 0 output by the classification sub-module 1021-1 indicates that a music signal is not present in the input audio signal. The determination result may also be expressed in other ways, such as in the form of a probability. For example, for the feature extraction sub-module 1012-1 and the classification sub-module 1021-1 for a music signal type, a probability greater than 50% indicates that a music signal is present in the input audio signal, and a probability less than 50% indicates that a music signal is not present in the input audio signal. Correspondingly, other classification sub-modules (e.g., classification sub-module 1021-2) may also use a value of 1 or a value of 0, or a probability greater than or less than 50%, to determine whether a speech signal is present in the input audio signal.

[0052] The feature extraction module 101 and the classification module 102 are specifically implemented as N feature extraction sub-modules 1012-1 to 1012-N and N classification sub-modules 1021-1 to 1021-N, respectively, and each of the N feature extraction sub-modules 1012-1 to 1012-N and the N classification sub-modules 1021-1 to 1021-N corresponds to one signal category, thereby enabling signals to be analyzed one by one from the perspective of different signal categories. This allows the neural network system to provide signal classification results one by one, rather than directly providing a final result including all classifications, thereby improving the overall accuracy of the neural network system.

[0053] In some embodiments, each of the N feature extraction submodules includes a first linear layer and a first nonlinear layer. That is, each of the N feature extraction submodules includes N first linear layers and N first nonlinear layers. For example, feature extraction submodule 1012-1 includes a first linear layer 1012-11 and a first nonlinear layer 1012-12, feature extraction submodule 1012-2 includes a first linear layer 1012-21 and a first nonlinear layer 1012-22, and subsequent feature extraction submodules also include first linear layers and first nonlinear layers. Furthermore, each of the N classification submodules includes a second linear layer and a second nonlinear layer. That is, each of the N classification submodules includes 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 subsequent classification submodules similarly have second linear layers and second nonlinear layers.

[0054] In one embodiment, the first nonlinear layer may be implemented with a Relu activation function and the second nonlinear layer may be implemented with a Sigmoid activation function, while in other embodiments, the first and second nonlinear layers may be implemented with other activation functions.

[0055] By achieving feature extraction and classification through a combination of linear and nonlinear layers, the neural network system is relatively lightweight, and the configuration of two linear and two nonlinear layers can ensure the recognition performance of the neural network system.

[0056] In a further embodiment, the feature extraction module 101 may further include a feature pre-recognition module 1011. The feature pre-recognition module 1011 is configured to pre-recognize features on the input signal to generate a pre-recognition signal, and input the pre-recognition signal to each of the N feature extraction sub-modules. For example, the feature pre-recognition module 1011 may include a linear layer and a non-linear layer, and the non-linear layer may be realized by, for example, a Relu activation function.

[0057] By installing a feature pre-recognition module before the N feature extraction sub-modules in the feature extraction module, the length of the thought chain can be increased, the accuracy of signal classification can be further improved, and the recognition performance of the neural network system can be improved.

[0058] FIG. 2 illustrates an exemplary method for performing signal classification using a neural network system according to an embodiment of the present disclosure.

[0059] According to this method, in step S220, each of the N feature extraction sub-modules in the neural network system extracts a respective feature signal based on an input signal.

[0060] 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 signal extracted by the corresponding feature extraction sub-module.

[0061] Optionally, if the feature extraction module of the neural network system further includes a feature pre-recognition module, the method further includes step S210, in which the feature pre-recognition module pre-recognizes features of the input signal to generate pre-recognition signals, and inputs the pre-recognition signals to each of the N feature extraction sub-modules. Note that in step S220, extracting each feature signal based on the input signal may involve extracting features directly based on the input signal, or extracting features based on pre-processed input signals. For example, if step S210 is included, each feature extraction sub-module extracts each feature signal based on the pre-recognition signals, where the pre-recognition signals belong to the pre-processed input signal.

[0062] Taking a communication interference signal as an example, when signal classification is performed using the neural network system 10, each of the 13 feature extraction sub-modules 1012-1 to 1012-13 in the neural network system extracts a feature waveform of the corresponding interference category based on the input communication interference signal. However, the parameters in each feature extraction sub-module are determined through training of the neural network system, and the extracted feature waveform can be determined thereby.

[0063] If the input communication interference signal includes signals on 273 physical resource blocks (PRBs), the amplitude value of the signal corresponding to each PRB may be the average value of the interference noise power within one sampling period. In this case, the input interference signal may be represented as a 1*273 vector.

[0064] Each of the thirteen feature extraction sub-modules 1012-1 to 1012-13 extracts a feature waveform of the corresponding interference category after receiving the input communication interference signal. For example, feature extraction sub-module 1012-1 may focus only on noise power related to the intelligent streetlight interference waveform feature and extract a feature waveform from the input communication interference signal according to trained feature extraction parameters for the intelligent streetlight interference waveform. Feature extraction sub-module 1012-2 may focus only on noise power related to the telecom 800M interference waveform feature and extract a feature waveform from the input communication interference signal according to trained feature extraction parameters for the telecom 800M interference waveform.

[0065] 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 signal extracted by its corresponding feature extraction sub-module. For example, the classification sub-module 1021-1 determines whether the feature waveform extracted by the feature extraction sub-module 1012-1 corresponds to the waveform of intelligent street light interference according to its trained parameters, thereby determining whether intelligent street light interference is present in the input communication interference signal. The classification sub-module 1021-2 determines whether the feature waveform extracted by the feature extraction sub-module 1012-2 corresponds to the waveform of Telecom 800M interference according to its trained parameters, thereby determining whether Telecom 800M interference is present in the input communication interference signal.

[0066] When it is determined how many signal content categories the input signal is classified into, N of the above-mentioned feature extraction sub-module and classification sub-module is set to a corresponding number. For example, if the input signal is an image signal, it is already known that the image signal will be classified into four image categories, so N is set to 4. In this case, the N feature extraction sub-modules mentioned above, i.e., each of the four feature extraction sub-modules, and the N classification sub-modules, i.e., each of the four classification sub-modules, each correspond to one image content category out of these four categories.

[0067] Similarly, if the input signal is a video signal, it is already known that the video signal content can be classified into three categories, so N is set to 3. In this case, the N feature extraction sub-modules, i.e., each of the three feature extraction sub-modules and the N classification sub-modules, i.e., each of the three classification sub-modules, correspond to one video content category among the three categories. If the input signal may include an audio signal, it is already known that the audio signal content can be classified into three categories, so N is also set to 3. In this case, the N feature extraction sub-modules, i.e., each of the three feature extraction sub-modules and the N classification sub-modules, i.e., each of the three classification sub-modules, correspond to one audio content category among the three categories. Based on this, those skilled in the art can easily conceive of adaptive adjustments for other signals (e.g., audio signals) that require classification.

[0068] FIG. 3 illustrates an exemplary method for training the neural network system 10 for signal classification according to an embodiment of the present disclosure.

[0069] The training method of the neural network system 10 for signal classification may include several steps as follows: In step S300, a sample input signal S1 and a labeled signal category of the sample input signal are obtained. The signal categories included in the sample input signal are labeled M1, M2, ..., MX, where X may be a natural number equal to or greater than 1. Taking a communication interference signal as an example, the sample input signal is a sample interference signal corresponding to 273 PRBs, and the amplitude value of the signal corresponding to each PRB is the average value of the interference noise power within one sampling period, and the labeled signal category may include at least one interference category, such as intelligent street light interference, or intelligent street light interference and jammer interference.

[0070] In step S320, each of the N feature extraction sub-modules extracts a respective feature signal based on the sample input signal, which may be labeled sequentially as C1, C2, ..., CN.

[0071] 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 may be sequentially labeled D1, D2, ..., DN. As described above, D1 to DN may be represented by a value of 0 or 1, a probability, or other representation formats.

[0072] In step S340, based on the first loss function, the N classification sub-modules determine first losses between the determination results for the corresponding signal categories (i.e., D1, D2, ..., DN) and the labeled signal categories (i.e., M1, M2, ..., MX) of the sample input signal. Here, D1, D2, ..., DN and M1, M2, ..., MX adopt the same labeling scheme. For example, if D1 to DN are adopted as values ​​0 or 1, when N is 13, the values ​​of D1, D2, ..., D13 output after the first training may be expressed as 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0. Taking the case of classifying communication interference signals as an example, if the sample interference signal is labeled as having intelligent street light interference, the values ​​of M1, M2, ..., MX are 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0. The first loss function, denoted as L1(Dx, Mx), is used to quantify and indicate the degree of deviation between the labeled signal category or true signal category value Mx of the input signal and the neural network system's predicted value for the signal category Dx. During the training process, the smaller the value of L1, the better. During the training process, adjusting the neural network parameters to keep the first loss within a specific first threshold can make the output results of the N classification submodules closer to the true signal category 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 for each interference category by the 13 classification submodules 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 completion conditions.

[0073] In step S350, the feature signals (i.e., C1, C2, ..., C N) extracted by the N feature extraction sub-modules are summed based on the sample input signal S1, and a second loss between the summed feature signal H1 and the sample input signal S1 is determined based on a second loss function. In a more specific embodiment, the second loss function is represented as L2(H1, S1) and is used to quantify the degree of discrepancy between the sum of the feature signals H1 extracted by the N feature extraction sub-modules and the input signal S1. During the training process, the smaller the value of L2, the better. During the training process, the neural network parameters are adjusted to keep the second loss within a specific second threshold, so that the sum of the waveforms extracted by the N feature extraction sub-modules can more faithfully reproduce the input signal, thereby improving the accuracy of signal classification. Taking a communication interference signal as an example, the closer the summed waveform obtained by adding together the feature waveforms extracted for each interference category by the 13 feature extraction sub-modules is to the waveform of the input interference signal, the higher the accuracy of feature extraction for the input signal. Another condition for completing training may be set as follows: a second loss between the summed signal obtained by adding together the feature waveforms extracted for each interference category by the 13 feature extraction sub-modules and the waveform of the input interference signal is smaller than a specific second threshold.

[0074] The loss functions L1 and L2 mentioned above may take the form of a Mean Squared Error (MSE), a Mean Absolute Error (MAE) function, or other loss functions.

[0075] In summary, in an embodiment of the present disclosure, the model parameters of the neural network system are adjusted based on the first loss and the second loss. The above steps may be repeatedly performed to adjust the model parameters of the neural network system until a completion condition is met. For example, the completion condition may be that the first loss is lower than a first threshold and that the second loss is lower than a second threshold. Of course, other conditions may be set as the completion condition as needed.

[0076] If the neural network system includes a feature pre-recognition module, the training method may optionally further include S320, in which the feature pre-recognition module performs feature pre-recognition on the sample input signal S1 to generate a pre-recognition signal P1, and inputs the pre-recognition signal P1 to each of the N feature extraction sub-modules. Correspondingly, during training, each parameter in the feature pre-recognition module is adjusted.

[0077] As mentioned above, prior to use, the neural network system 10 must be trained using a set of sample input signals labeled with signal categories corresponding to the number of signal categories contained in the input signals to be classified.

[0078] When the input signal is an image signal, a sample image signal set, whose image content categories have already been labeled, is used for training. For example, the image signal set may include sample image signals labeled with two categories: human images and animal images. To classify images into four image categories, namely, human images, animal images, landscape images, and other images, the four corresponding feature extraction sub-modules may correspond to human images, animal images, landscape images, and other images, respectively. In this case, the extracted feature signals may be C1, C2, C3, and C4. Specifically, each feature signal may be a pixel value associated with the corresponding category. The feature signals may be set to have the same format as the input image signal. For example, C1 may be a pixel value associated with a human image, and pixel values ​​unrelated to the human feature are set to 0. When identifying the second loss, C1, C2, C3, and C4 are added together to obtain a summed feature signal H1. The extracted feature image signal has the same signal format as the input image signal, or is 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 signal.

[0079] When the input signal is an audio signal, a sample audio signal set already labeled with its audio content category is used for training. The audio signal set may include, for example, a sample audio signal labeled as including a music signal. In this case, the audio signal may be a time-domain audio signal, e.g., a segment of an audio signal in the time dimension, or a frequency-domain audio signal, e.g., an audio signal converted to the frequency domain, and the specific format is not specifically limited herein. To classify audio into three audio categories, i.e., speech signals, music signals, and other noise signals, the three corresponding feature extraction sub-modules may be sequentially applied to speech signals, music signals, and other noise signals. In this case, the extracted feature signals may be C1, C2, and C3. Specifically, each feature signal may be a feature audio signal related to the corresponding category. The feature signals may be set to have the same format as the input audio signal. For example, C1 may be an audio signal related to speech signals, and in this case, audio signal values ​​not related to speech are set to 0. When identifying the second loss, C1, C2, and C3 are added together to obtain a summed feature signal H1. Note that the extracted feature audio signal has the same signal format as the input audio signal or is adjusted to have the same signal format as the input audio signal. This allows the sum of the extracted feature audio signals to be compared with the input signal.

[0080] When the input signal is a video signal, a sample image signal set already labeled with the video content category is used for training. The sample image signal set may include, for example, a sample image signal labeled as including a dynamic image. The video signal may be a video signal within a certain period of time. The video signal may be considered as a three-dimensional matrix, one dimension of which is time and the other two dimensions are the frame height and frame width of the video frame at the corresponding time point, but the specific format is not specifically limited herein. To classify a video into two video categories, namely, still images and dynamic images, the two feature extraction submodules may correspond to still images and dynamic images, respectively. In this case, the extracted feature signals may be C1 and C2. Specifically, each feature signal may be a feature video signal related to the corresponding category. The feature signals may be set to have the same format as the input video signal. For example, C1 may be a video related to a still image signal, and in this case, video signals unrelated to the still image signal are set to 0. When identifying the second loss, C1 and C2 are added together to obtain a summed feature signal H1. The extracted feature video signal has the same signal format as the input video signal, or is adjusted to have the same signal format as the input video signal when the summation is performed. This allows the sum of the extracted feature video signals to be compared with the input signal.

[0081] Based on this, one skilled in the art can easily imagine that when training a neural network system, it can be adaptively adjusted for other signals that require classification.

[0082] 1 to 3 have described the configuration of a neural network system according to an embodiment of the present disclosure, a method for performing classification using the neural network system, and a method for training the neural network system. These are suitable for various signals that require classification. Next, with reference to FIG. 4, an embodiment of classifying communication interference signals using a neural network system according to an embodiment of the present disclosure will be described. In this embodiment, the same content as that described above will not be described again.

[0083] FIG. 4 illustrates an exemplary block diagram of a neural network system for classifying communication interference signals in accordance with an embodiment of the present disclosure.

[0084] As described above, in an example in which the communication interference signal includes 13 types of interference categories, the N feature extraction sub-modules and the N classification sub-modules in the neural network system may be specifically realized as 13 feature extraction sub-modules 1012-1 to 1012-13 and 13 classification sub-modules 1021-1 to 1021-13, respectively. Each of the 13 feature extraction sub-modules 1012-1 to 1012-13 includes a first linear layer (simply indicated as "Linear" in FIG. 4) and a first nonlinear layer (realized by ReLU in FIG. 4, where the official name of ReLU is "Rectified Linear Unit" and is a linear rectification function (normalized linear function)). Each of the 13 classification submodules 1021-1 to 1021-13 includes a second linear layer (simply indicated as "linear" in FIG. 4) and a second nonlinear layer (realized by a sigmoid activation function in FIG. 4). By using at least two linear layers, one for each signal category, the neural network model becomes lighter while ensuring recognition performance.

[0085] The thirteen feature extraction sub-modules 1012-1 to 1012-13 and the thirteen classification sub-modules 1021-1 to 1021-13 each correspond to an interference category. For example, the first feature extraction sub-module 1012-1 and the first classification sub-module 1021-1 may correspond to an intelligent street light interference category, where the first feature extraction sub-module extracts feature waveforms for the interference category of intelligent street light interference, and the first classification sub-module performs recognition and classification based on the feature waveforms extracted by the first feature extraction sub-module to determine whether the interference category of intelligent street light interference exists in the input signal. The second feature extraction submodule 1012-2 and the second classification submodule 1021-2 may correspond to a telecom 800M interference category. The second feature extraction submodule extracts a feature waveform for the interference category "telecom 800M interference." The second classification submodule 1021-2 performs recognition and classification based on the feature waveform extracted by the second feature extraction submodule 1012-2 to determine whether the interference category "telecom 800M interference" exists in the input signal. More specifically, since the frequency range of telecom 800M interference is 824 to 849 / 869 to 894 MHz, when extracting the feature waveform using the second feature extraction submodule 1012-2, waveforms within this frequency range are extracted. Then, when recognizing the extracted waveform using the second classification submodule 1012-2, it determines whether the waveform exists within the 824 to 849 / 869 to 894 MHz frequency range. If it exists, it is determined that Telecom 800M interference exists; if not, it is determined that Telecom 800M interference does not exist.

[0086] As previously mentioned, the neural network system must be trained before it can be used to perform interference classification.

[0087] In the process of training the neural network system, first, a set of sample interference signals and their corresponding labeled interference categories are generated (for example, as described above, the sample interference signals are sample interference signals corresponding to 273 PRBs, and the amplitude value of the signal corresponding to each PRB is the average value of the interference noise power within one sampling period, and the labeled signal category may include at least one interference category, such as intelligent street light interference). Specifically, the sample interference signals on the 0th to 272nd PRBs are input to the neural network system. Next, the first linear layer of each feature extraction sub-module is used to extract feature waveforms of each interference category from the input sample interference signals. Then, the second linear layer of each classification sub-module is used to classify each extracted waveform. More specifically, for telecom 800M interference, when extracting feature waveforms using the second feature extraction sub-module 1012-2, the waveforms of PRBs within the 824-849 / 869-894 MHz frequency range are extracted. In this case, waveform or signal amplitude values ​​on other PRBs may be set to 0 so that the feature signal extracted by the feature extraction sub-module has the same format as the input signal. For example, if the input signal is one-dimensional vector data, the feature signal is also one-dimensional vector data, and both vectors have the same number of elements. When recognizing the waveform extracted using the second classification sub-module 1012-2, whether the waveform exists within the 824-849 / 869-894 MHz frequency range is determined to determine whether Telecom 800M interference is present in the sample interference signal. This is merely a simple example; in practice, the model parameters used to extract and classify features using each feature extraction sub-module and each classification sub-module are continuously optimized during the training process.

[0088] Next, based on the first loss function L1(Dx, Mx), a first loss is calculated between the determined results for the interference categories (labeled D1, D2, ..., D13) of all classification sub-modules and the labeled interference category of the sample interference signal (labeled M1, M2, ..., M13). Then, based on the second loss function L2(H1, S1), a second loss is calculated between the sum H1 of the feature waveforms extracted by all feature extraction sub-modules and the waveform S1 of the sample interference signal. Based on the first loss and the second loss, the parameters of each term in the neural network system are adjusted. In a more specific embodiment, the completion condition for training may be based on the first loss being smaller than a first threshold and the second loss being smaller than a second threshold. For a specific training method, please refer to the description of FIG. 3 in this specification, and a repeated description will be omitted here.

[0089] Now that a trained neural network system has been obtained, the trained neural network system may be utilized to perform interference classification on communication interference signals.

[0090] In one embodiment, first, a communication interference signal to be classified (e.g., interference noise on PRBs 0 to 272 that needs to be classified) is input to a neural network system. Next, each of the 13 feature extraction submodules is used to extract a feature waveform of each interference category from the input communication interference signal. Then, each of the 13 classification submodules is used to classify each extracted feature waveform. More specifically, each classification submodule determines whether a signal of the corresponding interference signal category is included in the communication interference signal based on the feature waveform extracted by the corresponding feature extraction submodule. For example, when the representation format of the output result by the classification sub-module is the value 1 or the value 0, if the determination result by the first classification sub-module 1021-1 (for intelligent streetlight interference) and the second classification sub-module 1021-2 (for telecom 800M interference) among the 13 classification sub-modules is 1, and the determination result by the remaining classification sub-modules is 0, then, as can be seen from the above-mentioned embodiment, the communication interference signal contains intelligent streetlight interference and telecom 800M interference, and no other interference signals exist.

[0091] According to an embodiment of the present disclosure, there is provided an apparatus for signal classification, including a processor and a memory having one or more computer programs stored therein, the one or more computer programs, when executed by the processor, causing the processor to perform any one of the methods for signal classification described in the various embodiments above. According to an embodiment of the present disclosure, there is also provided a computer-readable storage medium or a computer program product having a computer program stored therein, the computer program, when executed by at least one processor, causing the at least one processor to perform any one of the methods described above in the exemplary embodiments of the present disclosure. Here, examples of computer-readable storage media or computer program products 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 memory, hard disk drive (HDD), solid-state drive (SSD), card memory (e.g., multimedia card, Secure Digital (SD) card or Extreme Digital Card (XD) card), magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, and any other device. Any of the other devices are configured to non-transiently store computer programs and any associated data, data files, and data structures, and to provide the computer programs and any associated data, data files, and data structures to a processor or computer so that the processor or computer can execute the programs.The instructions or computer programs contained in the computer-readable storage medium or computer program product described above may be executed in an environment deployed on computing devices, such as clients, hosts, proxy devices, servers, etc. Also, in one example, the computer program and any associated data, data files, and data structures may be distributed across a networked computer system and stored, accessed, and executed in a distributed manner by one or more processors or computers.

[0092] Those skilled in the art will appreciate that the exemplary embodiments described above are not limiting and are not intended to be limiting. It will be understood that any two or more embodiments disclosed herein may be used together in any combination. Other modifications may also be made using other embodiments without departing from the spirit and scope of the subject matter disclosed herein. The various aspects of the disclosed invention, as generally described herein and illustrated in the drawings, can be arranged, substituted, combined, separated, and designed in a variety of different configurations, all of which are contemplated herein.

[0093] Those skilled in the art will understand that the various illustrative logical blocks, modules, circuits, and steps described herein can be implemented as hardware, software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the various illustrative components, blocks, modules, circuits, and steps have been generally described above in terms of their sets of functions. Whether such sets of functions are implemented in hardware or software depends on the actual application and 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 interpreted as departing from the scope of the present application.

[0094] Each illustrative logical block, module, and circuit described herein may be implemented or performed by 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 designed to perform the functions described herein. A general-purpose processor may be a microprocessor, but alternatively, the processor may be any conventional processor, controller, microcontroller, or state machine. A 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 in conjunction with a DSP core, or any other configuration.

[0095] The steps of a method or algorithm described herein may be implemented 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, a hard disk, a removable disk, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor such that the processor can read information from and write information to the storage medium. Alternatively, the storage medium may be integral to the processor. The processor and the storage medium may reside in an ASIC. The ASIC may reside in a user terminal. Alternatively, the processor and the storage medium may reside as discrete components in a user terminal.

[0096] In one or more exemplary designs, the functions may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software, the functions may be stored on or transmitted over as one or more instructions or code on a computer-readable medium. Computer-readable media includes both computer storage media and communication media, with communication media including any medium that facilitates transfer of a computer program from one place to another. Storage media may be any available medium accessible by a general-purpose or special-purpose computer.

[0097] The above-described contents are merely exemplary embodiments of the present application and do not limit the scope of protection sought by the present application, which is defined by the appended claims.

Claims

1. 1. A neural network based signal classification method, comprising: the neural network includes a feature extraction module and a classification module; the feature extraction module includes a plurality of feature extraction sub-modules in parallel; the classification module includes a plurality of parallel classification sub-modules corresponding to the plurality of feature extraction sub-modules, respectively; each feature extraction sub-module of the plurality of feature extraction sub-modules and each classification sub-module of the plurality of classification sub-modules corresponds to one signal category; each of the plurality of feature extraction sub-modules extracting a respective feature signal based on an input signal; and each classification sub-module of the plurality of classification sub-modules determining, based on the feature signal extracted by its corresponding feature extraction sub-module, whether the input signal includes a signal of a corresponding signal category.

2. each of the plurality of feature extraction sub-modules includes a first linear layer and a first non-linear layer; The method of claim 1 , wherein each classification sub-module of the plurality of classification sub-modules includes a second linear layer and a second non-linear layer.

3. the first nonlinear layer is realized by a Relu activation function; The method of claim 2 , wherein the second nonlinear layer is realized by a Sigmoid activation function.

4. The neural network each of the plurality of feature extraction sub-modules extracting a respective feature signal based on a sample input signal; each classification sub-module of the plurality of classification sub-modules respectively determining whether the sample input signal includes a signal of a corresponding signal category based on the feature signal extracted by the corresponding feature extraction sub-module; determining a first loss between the determination results for the signal categories respectively corresponding to the plurality of classification sub-modules and the label signal category of the sample input signal based on a first loss function; adding up the feature signals extracted by the plurality of feature extraction sub-modules based on the sample input signal, and determining a second loss between the added feature signal and the sample input signal based on a second loss function; and adjusting model parameters of the neural network based on the first loss and the second loss.

5. The feature extraction module further includes a feature pre-recognition module; The signal classification method includes: The method according to claim 1 , further comprising: the feature pre-recognition module pre-recognizing features of the input signal to generate a pre-recognition signal; and inputting the pre-recognition signal to each of the plurality of feature extraction sub-modules.

6. the input signal includes a communication interference signal and the signal category includes a communication interference category; or the input signal comprises an image signal and the signal category comprises an image content category; or the input signal comprises a video signal and the signal category comprises a video content category; or The method of claim 1 , wherein the input signal comprises an audio signal and the signal categories comprise audio content categories.

7. a feature extraction module including a plurality of feature extraction sub-modules in parallel; a classification module including a plurality of parallel classification sub-modules respectively corresponding to the plurality of feature extraction sub-modules; each feature extraction sub-module of the plurality of feature extraction sub-modules and each classification sub-module of the plurality of classification sub-modules corresponds to one signal category; each of the plurality of feature extraction sub-modules is configured to extract a respective feature signal based on an input signal; a neural network system for signal classification, wherein each classification sub-module of the plurality of classification sub-modules is configured to determine whether the input signal includes a signal of a corresponding signal category based on a feature signal extracted by a corresponding feature extraction sub-module.

8. 1. A method for training a neural network for signal classification, comprising: the neural network includes a feature extraction module and a classification module; the feature extraction module includes a plurality of feature extraction sub-modules in parallel; the classification module includes a plurality of parallel classification sub-modules corresponding to the plurality of feature extraction sub-modules, respectively; each feature extraction sub-module of the plurality of feature extraction sub-modules and each classification sub-module of the plurality of classification sub-modules corresponds to one signal category; obtaining a sample input signal and labeling signal categories of the sample input signal; each of the plurality of feature extraction sub-modules extracting a respective feature signal based on a sample input signal; Each classification sub-module of the plurality of classification sub-modules determines whether the sample input signal includes a signal of a corresponding signal category based on the feature signal extracted by the corresponding feature extraction sub-module; determining a first loss between the determination results of the plurality of classification submodules for the corresponding signal categories and the label signal category of the sample input signal based on a first loss function; summing each feature signal extracted by the plurality of feature extraction sub-modules based on the sample input signal, and determining a second loss between the summed feature signal and the sample input signal based on a second loss function; and adjusting model parameters of the neural network based on the first loss and the second loss.

9. a processor; a memory having one or more computer programs stored therein; 10. An apparatus for a neural network, wherein the one or more computer programs, when executed by the processor, cause the processor to perform the method of claim 1 or 8.

10. including computer programs, A computer program product, which implements the method of claim 1 or 8 when said computer program is executed by a processor.

Images