Deep learning signal type recognition method based on indefinite length multi-dimensional input

Abstract

The application belongs to the technical field of signal processing, and provides a deep learning signal type identification method based on indefinite length multi-dimensional input, comprising: performing continuous wavelet transform processing and discrete wavelet transform processing on original data of a signal to be predicted respectively to obtain continuous wavelet transform feature data and discrete wavelet transform feature data; performing normalization processing on the original data, the continuous wavelet transform feature data and the discrete wavelet transform feature data, inputting the original data into a pre-trained deep learning network, and performing signal type identification. The application performs high-dimensional feature extraction and feature compression on original signals and corresponding continuous wavelet transform data and discrete wavelet transform data through a deep learning multi-modal network, completes signal type identification, and can obtain characteristic expressions which are representative in time domain, frequency domain and time-frequency domain, and improve the comprehensiveness and discriminability of features.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of signal processing, and in particular to a deep learning signal type recognition method based on indefinite length multi-dimensional input. BACKGROUND

[0002] Signal type recognition has wide applications and important significance in both civilian and military fields. In the civilian field, signal type recognition technology is widely used in communication, network security, and smart home systems. For example, by recognizing different communication signal types, the efficiency of spectrum utilization can be improved, interference can be reduced, and communication quality can be ensured; in network security, recognizing abnormal signal types can detect and defend against network attacks, protecting user data security; in smart home, recognizing various control signals can achieve automated management and improve the convenience of life. In the military field, signal type recognition technology has more stringent security requirements and strategic significance. It plays a key role in electronic warfare, intelligence collection, and military communication. For example, in electronic warfare, by recognizing enemy signal types, effective electronic countermeasures such as jamming and deception can be taken; in intelligence collection, recognizing and analyzing enemy communication signal types can obtain valuable intelligence to support military decision-making; in military communication, by recognizing friendly and enemy communication signals, the security and confidentiality of military communication can be ensured, and the efficiency of communication on the battlefield can be improved. In summary, signal type recognition technology plays an indispensable role in both civilian and military fields.

[0003] In recent years, with the development of deep learning, deep learning technology has been integrated into various industries and has achieved good results. Similarly, deep learning methods have been applied to signal type recognition by domestic and foreign researchers, and have achieved very good results.

[0004] Because the duration of the signal itself is uncontrollable, the signal sampling rate and bandwidth cannot be determined, but in actual scenarios, there are a large number of signals of different lengths. The network input of the general deep learning method is fixed, and usually needs to be cut or padded when processing signals of different lengths, but because the same length of signal contains a huge difference in feature information, this method has strong limitations. SUMMARY

[0005] In view of the defects in the prior art, the present application provides a deep learning signal type recognition method based on indefinite length multi-dimensional input to solve the problem that the characteristics of the signal itself and the duration are not fixed, resulting in differences in feature extraction of the signal, and poor accuracy of signal type recognition.

[0006] The deep learning signal type recognition method based on indefinite length multi-dimensional input provided by the present application comprises:

[0007] The original data of the signal to be predicted is processed by continuous wavelet transform and discrete wavelet transform respectively to obtain continuous wavelet transform feature data and discrete wavelet transform feature data;

[0008] The original data, the continuous wavelet transform feature data and the discrete wavelet transform feature data are normalized and input into a pre-trained deep learning network for signal type recognition.

[0009] The training sample data of the deep learning network includes original sample features, original sample continuous wavelet transform features and original sample discrete wavelet transform features.

[0010] According to the above technical solution, the deep learning signal type recognition method based on indefinite length multi-dimensional input provided by the application can obtain representative feature expressions in time domain, frequency domain and time-frequency domain, effectively compensate for the deficiency of single features in information dimension, improve the comprehensiveness and discriminability of features, thereby enhancing the recognition ability of the network under different signal types, different lengths and complex noise environments, and finally achieving high-precision and strong-robustness signal type intelligent recognition effect under indefinite length input and complex interference conditions.

[0011] Optionally, the training method of the deep learning network comprises:

[0012] The original real narrowband signal data is extracted in random length to obtain original sample data; the original real narrowband signal data includes data of different signal types, and the extraction length is at least minLen;

[0013] The original sample data is processed by continuous wavelet transform and discrete wavelet transform respectively to obtain continuous wavelet transform output results and discrete wavelet transform output results;

[0014] The original sample data, the continuous wavelet transform output results and the discrete wavelet transform output results are processed by maximum and minimum value normalization;

[0015] The sample label is associated with the original sample data after normalization, and the continuous wavelet transform output results and the discrete wavelet transform output results form a data set;

[0016] The data set is used for model training, and when the loss function converges, a model weight file is obtained.

[0017] According to the technical scheme, the multi-dimensional input network structure is used, the original signal itself is included in the input, the original signal is processed through continuous wavelet transform and discrete wavelet transform, the obtained feature data and the original data are input into the deep learning network, high-order semantic features are compressed and fused, and type recognition of the signal is completed; the wavelet transform can simultaneously observe the frequency and time axis, has good time resolution when the frequency is high, has good frequency resolution when the frequency is low, is suitable for analyzing mutation signals and singular signals, and can analyze different sizes of signal scales, and has high recognition accuracy for various types of signals; and the method can adapt to various complex signal scenes, and has strong adaptability to burst signals and variable modulation signals.

[0018] Optionally, the original sample data is processed through continuous wavelet transform, and the processing includes:

[0019] The Mexican hat wavelet is selected as the mother wavelet.

[0020] The continuous wavelet transform processing outputs a two-dimensional matrix of ; wherein is the input signal, is the mother wavelet function, a is the scale parameter, b is the translation parameter, m is the length of the scale sequence, is the length of the original sample data.

[0021] Optionally, the original sample data is processed through discrete wavelet transform, and the processing includes:

[0022] The Mexican hat wavelet is selected as the mother wavelet.

[0023] The signal discrete wavelet transform result is obtained through the original sample data and the mother wavelet, and the output data is the approximation coefficient and the detail coefficient; the approximation coefficient , and the detail coefficient , and are low-pass and high-pass filter coefficients of the wavelet analysis filter.

[0024] Optionally, the deep learning network includes a convolution module I and a convolution module II,

[0025] The signal original data is sequentially processed through the convolution module I with 128 and 256 convolution kernels, and then is processed through global maximum pooling;

[0026] The discrete wavelet transform feature data is sequentially processed through the convolution module II with 128 and 256 convolution kernels, the convolution kernel sizes are 2*3 and 1*3 respectively, the pooling methods are 2*2 and 1*2 respectively, and then is processed through global maximum pooling.

[0027] The continuous wavelet transform feature data is sequentially processed by the convolution module II with 128, 256, 256, and 256 convolution kernels, the convolution kernel size is 3*3, the pooling method is 2*2, and then global maximum pooling processing is performed;

[0028] The processed signal original data, the discrete wavelet transform feature data, and the continuous wavelet transform feature data are subjected to feature splicing processing and dimension expansion, then the convolution module I with 64, 32, and 5 convolution kernels is used for feature fusion and feature compression, and finally global maximum pooling is performed.

[0029] Optionally, the signal type recognition is based on the following formula:

[0030] ,

[0031] Among them, represents an indefinite length one-dimensional input signal;

[0032] DWT represents the discrete wavelet transform;

[0033] CWT represents the continuous wavelet transform;

[0034] , , respectively represent global maximum pooling in time, scale-time, and scale-translation dimensions;

[0035] represents vector channel splicing; is a fusion / compression linear mapping, BN represents batch normalization, σ is a nonlinear operation;

[0036] represents the weight parameter in the classifier parameter;

[0037] represents the offset parameter in the classifier parameter;

[0038] represents the probability of the cth class;

[0039] represents the output sample label.

[0040] By adopting the above technical solutions, the present application has the following beneficial effects:

[0041] The application adopts a multi-modal input method of processing original data, discrete wavelet features and continuous wavelet features, aims to simultaneously retain time domain information, multi-scale frequency domain details and continuous time-frequency change features, can obtain representative feature expression in time domain, frequency domain and time-frequency domain, effectively makes up for the deficiency of single feature in information dimension, improves the comprehensiveness and discriminability of the feature, thereby enhances the recognition ability of the network under different signal types, different lengths and complex noise environment, and finally realizes the effect of high-precision and strong-robustness intelligent signal type recognition under the condition of indefinite length input and complex interference;

[0042] The application adopts a multi-input fusion network method, aims to fully utilize complementary information from different feature domains, cooperatively models original time domain features and multi-scale time-frequency features in a unified model, can obtain balanced and comprehensive feature expression in time domain integrity, frequency domain resolution and time-frequency local details, and effectively eliminates the structural limitation caused by the input length difference through a global pooling and adaptive fusion mechanism, thereby supports indefinite length input while maintaining the stability and discriminability of the feature expression, and finally realizes the effect of high-precision and strong-robustness signal type recognition under various signal duration and complex interference conditions. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the drawings needed in the specific embodiments or prior art description will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, each element or part is not necessarily drawn according to the actual scale.

[0044] Figure 1 A flowchart of a deep learning signal type recognition method based on indefinite length multi-dimensional input provided by an embodiment of the present application is shown;

[0045] Figure 2 A flowchart of a training method of a deep learning network provided by an embodiment of the present application is shown;

[0046] Figure 3 A flowchart of a deep learning signal type recognition method based on indefinite length multi-dimensional input provided by an embodiment of the present application is shown;

[0047] Figure 4 A schematic diagram of a deep learning network provided by an embodiment of the present application is shown;

[0048] Figure 5 A schematic diagram of a Mexican hat wavelet provided by an embodiment of the present application is shown;

[0049] Figure 6The feature effect diagram after continuous wavelet transform processing provided by the embodiment of the application is shown. DETAILED DESCRIPTION

[0050] The embodiments of the technical solutions of the application will be described in detail below with reference to the drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the application, and therefore only serve as examples, and cannot be used to limit the protection scope of the application.

[0051] It should be noted that, unless otherwise specified, the technical terms or scientific terms used in the present application should be understood as the general meanings understood by the skilled in the art to which the application belongs.

[0052] In one embodiment, as shown in Figure 1 a deep learning signal type recognition method based on indefinite-length multi-dimensional input is provided, comprising:

[0053] S110. The original data of the signal to be predicted is respectively subjected to continuous wavelet transform processing and discrete wavelet transform processing to obtain continuous wavelet transform feature data and discrete wavelet transform feature data;

[0054] S120. The original data, the continuous wavelet transform feature data and the discrete wavelet transform feature data are subjected to normalization processing and input into a pre-trained deep learning network for signal type recognition.

[0055] The training sample data of the deep learning network includes original sample features, original sample continuous wavelet transform features and original sample discrete wavelet transform features.

[0056] The original signal is respectively subjected to continuous wavelet processing and discrete wavelet processing to obtain two groups of signal processing features. High-dimensional feature extraction and feature compression are performed on the original signal and the corresponding continuous wavelet transform data and discrete wavelet transform data by a deep learning multi-modal network to complete signal type recognition. Representative feature expressions in the time domain, the frequency domain and the time-frequency domain can be obtained, the deficiency of a single feature in the information dimension is effectively made up, the comprehensiveness and the discriminability of the features are improved, and thus the recognition ability of the network in different signal types, different lengths and complex noise environments is enhanced, and finally the intelligent signal type recognition effect with high precision and strong robustness is still maintained under the conditions of indefinite-length input and complex interference.

[0057] In one embodiment, as shown in Figure 2 a deep learning network training method, comprising:

[0058] S210. Original sample data is obtained by performing random length extraction on original real narrowband signal data; the original real narrowband signal data includes data of different signal types, and the extraction length is at least minLen; minLen = 2048.

[0059] S220. The original sample data is respectively processed by continuous wavelet transform and discrete wavelet transform to obtain continuous wavelet transform feature output results and discrete wavelet transform output results;

[0060] S230. The original sample data, the continuous wavelet transform output results and the discrete wavelet transform output results are processed by maximum and minimum value normalization;

[0061] S240. The sample label is associated with the original sample data after normalization, and a data set is formed with the continuous wavelet transform output results and the discrete wavelet transform output results.

[0062] According to different types of original signals, sample labels are made, and the label form is one-hot type. The sample label is combined with the normalized original sample data to form a data set. In a specific embodiment, the sample type includes five types, and the sample labels are [1, 0, 0, 0, 0], [0, 1, 0, 0, 0], [0, 0, 1, 0, 0], [0, 0, 0, 1, 0], and [0, 0, 0, 0, 1] respectively.

[0063] S250. The data set is used for model training, and when the loss function converges, a model weight file is obtained. The loss function can use a multi-class cross-entropy loss function.

[0064] Optionally, the original sample data is respectively processed by continuous wavelet transform, including:

[0065] The Mexican hat wavelet is selected as the mother wavelet. The formula of the Mexican hat wavelet is as follows:

[0066] , m(x) is the result of wavelet transform, x is the number of time points, is a scale function for determining the width of the wavelet function;

[0067] The wavelet shape is as shown in Figure 5 .

[0068] Continuous wavelet transform processing , the output data is a two-dimensional matrix of ; wherein is the input signal, is the mother wavelet function, a is the scale parameter, b is the translation parameter, m is the length of the scale sequence, is the length of the original sample data. The length of the scale sequence m = 512, and the characteristic effect after continuous wavelet transform processing is as shown in Figure 6 .

[0069] Optionally, the original sample data is respectively processed by discrete wavelet transform, including:

[0070] The Mexican hat wavelet is selected as the mother wavelet;

[0071] The signal discrete wavelet transform result is obtained through the original sample data and the mother wavelet, and the output data is the approximation coefficient and the detail coefficient; the approximation coefficient , the detail coefficient , and are the low-pass and high-pass filter coefficients of the wavelet analysis filter respectively.

[0072] It should be noted that, in specific implementation, the deep learning signal type recognition method based on indefinite length multi-dimensional input is as shown in Figure 3 , first, the original real signal data containing different signal types are extracted in random length to obtain original signal sample data; then, the original signal sample data are subjected to continuous wavelet transform and discrete wavelet transform; the original signal sample data and the two kinds of wavelet transform results are subjected to normalization processing, and a training data set is constructed in combination with signal type label data; the data set is input into the deep learning network for model training, and the weight file is obtained after the model converges; after the prediction signal data are subjected to wavelet transform and normalization processing, they are input into the trained model for prediction, and the signal type recognition result can be obtained. The network input data contains the original signal and the features of continuous wavelet transform and discrete wavelet transform, which can separate the fine or coarse components of the signal and has strong anti-noise ability; different input lengths of the same signal type increase the robustness and feature generalization ability of signal recognition, and the present application still has strong recognition accuracy in a complex noise environment.

[0073] In one embodiment, as shown in Figure 4 , the deep learning network includes convolution module I and convolution module II,

[0074] The signal original data are subjected to feature processing using the convolution module I with 128 and 256 convolution kernels in sequence, and are subjected to global maximum value pooling processing after the processing;

[0075] The discrete wavelet transform feature data are subjected to feature processing using the convolution module II with 128 and 256 convolution kernels in sequence, the convolution kernel size is 2*3 and 1*3 respectively, the pooling method is 2*2 and 1*2 respectively, and the processing is subjected to global maximum value pooling processing after the processing;

[0076] The continuous wavelet transform feature data are subjected to feature processing using the convolution module II with 128, 256, 256 and 256 convolution kernels in sequence, the convolution kernel size is 3*3, and the pooling method is 2*2, and the processing is subjected to global maximum value pooling processing after the processing;

[0077] The processed signal raw data, discrete wavelet transform feature data and continuous wavelet transform feature data are subjected to feature splicing processing and dimension expansion; then a convolution module I with 64, 32 and 5 convolution kernels is used for feature fusion and feature compression; finally, global maximum pooling is performed.

[0078] The purpose of using the multi-input fusion network method is to make full use of complementary information from different feature domains, model the original time domain features and multi-scale time-frequency features in a unified model, obtain balanced and comprehensive feature expression in terms of time domain integrity, frequency domain resolution and time-frequency local details, and effectively eliminate the structural limitations caused by the length difference of the input through the global pooling and adaptive fusion mechanism, so as to support indefinite length input while maintaining the stability and discriminability of the feature expression, and finally achieve high-precision and strong-robust signal type recognition effect under various signal duration and complex interference conditions.

[0079] In one embodiment, the signal type recognition in step S2 is determined according to the following formula:

[0080] ,

[0081] wherein, represents an indefinite length one-dimensional input signal;

[0082] DWT represents discrete wavelet transform; CWT represents continuous wavelet transform;

[0083] , , respectively represent global maximum pooling in time, scale-time and scale-translation dimensions; the variable length feature is mapped to a fixed length vector to ensure support for indefinite length input.

[0084] represents vector concatenation by channel; is a fusion / compression linear mapping, BN represents batch normalization, σ is a nonlinear operation;

[0085] represents the weight parameter in the classifier parameter;

[0086] represents the offset parameter in the classifier parameter;

[0087] represents the probability of the cth class;

[0088] represents the output sample label.

[0089] The above embodiments only express several implementation ways of the present application, and the description is specific and detailed, but it should not be understood as a limitation to the patent scope of the application. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, several modifications and improvements can be made, which are all within the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A deep learning signal type recognition method based on indefinite length multi-dimensional input, characterized in that, The application relates to a signal type recognition method based on deep learning, and belongs to the technical field of signal processing. The original data of a signal to be predicted is subjected to continuous wavelet transform processing and discrete wavelet transform processing respectively to obtain continuous wavelet transform feature data and discrete wavelet transform feature data; The original data of the signal to be predicted is a real narrowband signal; The original data, the continuous wavelet transform feature data and the discrete wavelet transform feature data are subjected to normalization processing and are input into a pre-trained deep learning network for signal type recognition; The training sample data of the deep learning network comprises original sample features, original sample continuous wavelet transform features and original sample discrete wavelet transform features; The deep learning network comprises a convolution module I and a convolution module II, The original data of the signal is subjected to feature processing by using the convolution module I with 128 and 256 convolution kernels in sequence, and is subjected to global maximum value pooling processing after the processing; The discrete wavelet transform feature data is subjected to feature processing by using the convolution module II with 128 and 256 convolution kernels in sequence, the convolution kernel sizes are 2*3 and 1*3 respectively, the pooling methods are 2*2 and 1*2 respectively, and the discrete wavelet transform feature data is subjected to global maximum value pooling processing after the processing; The continuous wavelet transform feature data is subjected to feature processing by using the convolution module II with 128, 256, 256 and 256 convolution kernels in sequence, the convolution kernel size is 3*3, the pooling method is 2*2, and the continuous wavelet transform feature data is subjected to global maximum value pooling processing after the processing; The processed original data of the signal, the discrete wavelet transform feature data and the continuous wavelet transform feature data are subjected to feature splicing processing and dimension expansion, are subjected to feature fusion and feature compression by using the convolution module I with 64, 32 and 5 convolution kernels respectively, and finally are subjected to global maximum value pooling.

2. The method of claim 1, wherein, The training method of the deep learning network comprises the following steps: Original real narrowband signal data is subjected to random length extraction to obtain original sample data; the original real narrowband signal data comprises data of different signal types, and the extraction length is at least minLen; The original sample data is subjected to continuous wavelet transform processing and discrete wavelet transform processing respectively to obtain continuous wavelet transform output results and discrete wavelet transform output results; The original sample data, the continuous wavelet transform output results and the discrete wavelet transform output results are subjected to maximum and minimum value normalization processing; Sample labels are associated with the original sample data subjected to the normalization processing, the continuous wavelet transform output results and the discrete wavelet transform output results to form a data set; A model is trained by using the data set, and a model weight file is obtained after the loss function converges.

3. The method of claim 2, wherein, The continuous wavelet transform processing of the original sample data comprises the following steps: A Mexican hat wavelet is selected as a mother wavelet; Continuous wavelet transform processing The output data is A two-dimensional matrix; where For input signal, Let be the mother wavelet function, a be the scaling parameter, b be the translation parameter, and m be the length of the scaling sequence. The length of the original sample data.

4. The method of claim 3, wherein, The discrete wavelet transform processing of the original sample data comprises the following steps: A Mexican hat wavelet is selected as a mother wavelet; The signal discrete wavelet transform result is obtained by the original sample data and the mother wavelet, and the output data is the approximation coefficient and the detail coefficient , the detail coefficient , and are low-pass and high-pass filter coefficients of the wavelet analysis filter respectively.

5. The method of claim 2, wherein, The signal type recognition is based on the following formula: ， wherein represents an indefinite-length one-dimensional input signal; DWT represents the discrete wavelet transform; CWT represents the continuous wavelet transform; , , respectively denote the global maximum optimization over time, scale-time, scale-translation dimension. denotes vector concatenation by channel; is a fusion / compression linear map, BN denotes batch normalization, ​ is a nonlinear operation; denotes a weight parameter in the classifier parameters; denotes an offset parameter in the classifier parameters; denotes the c-th class probability; Sample label representing the output.

Images