CN111652177A - Signal feature extraction method based on deep learning - Google Patents

Abstract

The invention discloses a signal feature extraction method based on deep learning, comprising the following steps: 1) compressing an original sparse spectrum signal to make it a time-series spectrum signal, and completing denoising processing after compression to obtain a reconstructed time-series signal Spectral signals, as training data; 2) Mark the time-series spectrum signals that are not marked in the training data with semi-supervised learning method; simulate each group of time-series spectrum signals, and use genetic algorithm to analyze the correlation between time-series spectrum signals Measure; 3) Use the marked training data to classify and predict each group of time-series spectrum signals by using the deep learning method. The present invention improves model training efficiency through multi-method fusion, realizes automatic labeling of unlabeled labels, and results in more accurate prediction results.

Description

Signal feature extraction method based on deep learning

technical field

The invention belongs to the field of information technology, and in particular relates to a signal feature extraction method based on deep learning.

Background technique

There are two common deep learning model training methods: deep learning models with feedback regulation and cascaded deep models without feedback. Deep models including feedback regulation such as convolutional neural networks, deep belief networks, and auto-encoding representations, etc., and deep models without feedback include DeepBoosting, DeepFisher, and DeepPCA. Both of these deep learning models are regular combinations of feature extraction and feature selection. Therefore, how to set up feature transformation and feature selection methods and their combination methods is an important part of research on the basis of predecessors.

(1) Feature extraction

Feature extraction is the process of obtaining information from data features. The information of the data is obtained by transforming the features. The simplest feature transformation methods include PCA transformation, linear discriminant analysis, etc. PCA transformation is to find the main axis direction of the data, so that the data can be projected with as few features as possible to represent as much information as possible; linear discriminant analysis is through linear projection. The data in space has larger inter-class distance and smaller intra-class distance; Gabor filter transformation is an important transformation method in the image field. By setting different parameters of the Gaussian kernel function in the filter, it can be effectively extracted. The edge information in the image; by using the DOG operator to detect key points in the multi-scale space, the SIFT feature extracts the local information of the image, so that it maintains good invariance to rotation, scale scaling, and brightness transformation; Divided into several connected regions and counted the histogram of each region, the HOG feature can effectively deal with the rotation and scale invariance of the image.

(2) Feature selection

Feature selection is a method of picking out a subset of features from the data features that are helpful for the task. Common feature selection methods can be divided into three categories: Filter, Wrapper and Embedded methods. The Filer method selects features one by one according to the size of the corresponding feature evaluation indicators for each feature, which are often some statistical characteristics of the data. The Wrapper method selects features according to the classification accuracy of the selected feature subspace for classification. Since the feature subspace is uncertain, the Wrapper method requires multiple trainings to give the selected feature space. Embedded methods first identify the model using features, and then search the feature space for feature subspaces that can improve model performance. From whether the sample contains category information, feature selection can be divided into supervised learning and unsupervised learning. Early unsupervised learning first defined some metric characteristics, then calculated the metric characteristics of the features one by one, and selected them in order. These metric characteristics may be clustering effect, information redundancy, etc. Some representative measures include Laplacian Score, Traceratio Wait. However, this feature extraction method relying on search requires a huge amount of computation, so researchers began to consider clustering algorithms that no longer require a search space. Based on the similar characteristics of features, a series of spectral clustering algorithms were proposed.

Through a reasonable combination of appropriate feature extraction and feature selection methods, deep learning has achieved a series of results in various fields. The successful use of convolutional neural networks has made breakthroughs in the field of speech recognition, and convolutional neural networks have also achieved the best results in the field of target recognition.

Besides convolutional neural networks, deep belief networks are also an important part of deep learning. The idea of deep belief network is to learn shallow features through a greedy strategy, and then obtain a more abstract description through the combination of shallow features. With the development of deep-structured neural networks, many improved algorithms for deep learning have emerged, and deep learning has been widely used. Deep learning has been successfully applied in speech recognition, image processing, etc. Hinton pointed out that in order to train a DBN, the restricted Boltzmann machine of each layer is firstly trained through unsupervised greed, and a deep belief network DBN is constructed by a combination of a set of restricted Boltzmann machines, and then through traditional global learning The algorithm fine-tunes the constructed DBN to make the network optimal. The core idea of deep learning is as follows: unsupervised learning is used for the pre-training of each layer of the network; only one layer is trained each time with unsupervised learning, and its training result is used as the input of its higher layer; supervised learning is used to adjust all the Floor.

Deep Belief Network (DBN) is a typical generative deep structure. Deep learning has been successfully applied to a variety of pattern classification problems, bringing great convenience to human beings, but it is still only in the development stage, and there are still many problems worthy of further in-depth solutions, such as:

(1) The model training time is generally too long, how to improve the training speed and improve the practicability of deep learning, etc.

(2) The feature learning of labeled data is still dominant, and the data in real life are all unlabeled data. It is obviously unrealistic to add manual labels to these data one by one. How can we study the automatic labeling of unlabeled labels? Technology.

(3) A single deep learning method cannot bring good results. How can the deep learning method be integrated with other methods or other methods to be perfectly applied to practical applications.

SUMMARY OF THE INVENTION

In view of this, in order to solve the above-mentioned problems in the prior art, the present invention proposes a deep learning-based signal feature extraction method that improves model training efficiency and realizes multi-method fusion of automatic labeling without labeling.

The technical solution of the present invention is to provide a signal feature extraction method based on deep learning, comprising the following steps:

1) Compress the original sparse spectrum signal to make it a time-series spectrum signal, complete the denoising process after compression, and obtain the reconstructed time-series spectrum signal as training data;

2) Mark the time-series spectrum signals that are not marked in the training data with a semi-supervised learning method; simulate each group of time-series spectrum signals, and use genetic algorithms to measure the correlation between the time-series spectrum signals;

3) Using the marked training data, the deep learning method is used to classify and predict each group of time-series spectrum signals.

Optionally, in step 1), the dynamic subspace tracking algorithm in the greedy tracking algorithm is used to compress the original sparse spectral signal, and the unscented Kalman filtering algorithm is used to denoise the compressed sparse spectral signal, or reconstruct the original sparse spectral signal. The obtained time series spectrum signal is directly predicted by using the unscented Kalman filtering algorithm to obtain the first prediction result.

Optionally, the semi-supervised learning method includes an active learning method and a deep belief net in deep learning, the active learning method is used to mark incompletely marked training data, and the deep belief net is a probability generation model, so The above probability generation model includes multiple restricted Boltzmann machines to establish a joint distribution between observation data and labels. The restricted Boltzmann machine consists of a visible layer and a hidden layer, and there are connections between the layers. There is no connection between the units in the layer, and the deep belief network classifies and predicts the training data, and obtains a classification result and a second prediction result.

Optionally, the active learning method is incorporated into the deep belief net as the first hidden layer.

Optionally, first use the training-restricted Boltzmann machine to set the energy of each layer to the extreme value. After all the restricted Boltzmann machines have been trained, use the label results of active learning to set the weights of each layer. Make adjustments to finally obtain a classification surface of a deep belief network; when the deep belief network is used for prediction, all training data will be replaced by the residuals in the unscented Kalman filter algorithm model, and the deep belief network will be used as a The machine learning model is used to predict the change curve of the residual in the unscented Kalman filtering algorithm model. Finally, the prediction result of the deep belief network will be combined with the prediction result of the unscented Kalman filtering algorithm to obtain the final result.

Compared with the prior art, the present invention has the following advantages:

In the present invention, the classification and prediction results obtained by the extended deep belief network (combined with active learning and unscented Kalman filtering algorithm) will be superior to the deep belief network of the prior art, and have great advantages in the field of signal processing and machine learning. The application value of the invention improves the model training efficiency through multi-method fusion, and realizes the automatic labeling of unlabeled labels, and the prediction result is more accurate.

Description of drawings

Figure 1 is a flow chart of the present invention.

Detailed ways

The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention is not limited to these embodiments. The present invention covers any alternatives, modifications, equivalent methods and arrangements made within the spirit and scope of the present invention.

In order to give the public a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention, and those skilled in the art can fully understand the present invention without the description of these details.

1, the specific implementation process of the present invention is as follows:

The signal feature extraction method based on deep learning includes the following steps:

1) Compress the original sparse spectrum signal to make it a time-series spectrum signal, complete the denoising process after compression, and obtain the reconstructed time-series spectrum signal as training data;

2) Mark the time-series spectrum signals that are not marked in the training data with a semi-supervised learning method; simulate each group of time-series spectrum signals, and use genetic algorithms to measure the correlation between the time-series spectrum signals;

3) Using the marked training data, the deep learning method is used to classify and predict each group of time-series spectrum signals.

In step 1), the dynamic subspace tracking algorithm in the greedy tracking algorithm is used to compress the original sparse spectrum signal, and the unscented Kalman filter algorithm is used to denoise the compressed sparse spectrum signal, or the reconstructed time series spectrum. signal, and at the same time use the unscented Kalman filter algorithm to directly predict to obtain the first prediction result.

The semi-supervised learning method includes an active learning method and a deep belief network in deep learning. The active learning method is used to mark incompletely marked training data, and the deep belief network is a probability generation model. The probability generation model Including multiple restricted Boltzmann machines to establish a joint distribution between observation data and labels Without connections, the deep belief network classifies and predicts the training data, obtaining a classification result and a second prediction result.

The idea of deep belief network is to learn shallow features through a greedy strategy, and then obtain a more abstract description through the combination of shallow features. With the development of deep-structured neural networks, many improved algorithms for deep learning have emerged, and deep learning has been widely used. Deep learning has been successfully applied in speech recognition, image processing, etc. Hinton pointed out that in order to train a DBN, the restricted Boltzmann machine of each layer is firstly trained through unsupervised greed, and a deep belief network DBN is constructed by a combination of a set of restricted Boltzmann machines, and then through traditional global learning The algorithm fine-tunes the constructed DBN to make the network optimal. The core idea of deep learning is as follows: unsupervised learning is used for the pre-training of each layer of the network; only one layer is trained each time with unsupervised learning, and its training result is used as the input of its higher layer; supervised learning is used to adjust all the Floor.

Deep Belief Network (DBN) is a typical generative deep structure. Assuming that a DBN has n hidden layers, let gi represent the vector of the ith hidden layer, then the model of the DBN can be expressed as:

p(x,g ¹ ,g ² ,...,g ⁿ )=p(x|g ¹ )p(g ¹ |g ² )...p(g ^n-1 ,g ⁿ ) (1)

Among them, the conditional probability p(g ⁱ |g ⁱ⁺¹ ) is:

——第i层的第j个节点，Wi第i层的权重矩阵，σ为where:

——The jth node of the i-th layer, the weight matrix of the i-th layer of Wi, σ is

The active learning method is incorporated in the deep belief net as the first hidden layer.

First, the training-restricted Boltzmann machine is used to set the energy of each layer to the extreme value. After all the restricted Boltzmann machines are trained, the weights of each layer are adjusted using the label results of active learning. Finally, Obtain a classification surface of the deep belief net; when the deep belief net is used for prediction, all the training data will be replaced by the residuals in the unscented Kalman filter algorithm model, and the deep belief net will be used as a machine learning model to Predict the change curve of the residual in the unscented Kalman filter algorithm model, and finally, the prediction result of the deep belief network will be combined with the prediction result of the unscented Kalman filter algorithm to get the final result.

A more specific embodiment of the present invention is: using the Google spectrum database to implement and verify the method of the present invention. First, the resulting sparse data will be used for compression processing. During the compression process, the compression results between various signal reconstruction techniques are compared. At the same time, it can be concluded that the accuracy improvement rate of Dynamic SP to the traditional SP algorithm has a profound impact on the entire spectrum signal processing field.

Next, the compressed spectral signal is filtered using UKF. UKF has a good fit for nonlinear time series, especially time series with Gaussian distribution. In this step, other mathematical models can be compared, such as the fitting results of autoregressive ARX, ARMA, ARMAX and ARIMA models, etc. For spectral signals, the fit of the UKF model can be predicted to be optimal. UKF can be directly used for prediction, and the filtered waveform after UKF filtering can be used for prediction through DBNs, and the effect is better.

As a next step, in the labeled training dataset, we can compare the labeling results of two separate machine learning methods: the active learning method and the deep belief net method, and combine the two methods for labeling.

Finally the deep belief net will be used for classification and prediction. The classification and prediction results obtained by the extended deep belief net in the present invention (combining active learning and UKF) will be superior to the traditional deep belief net. It has considerable research value in the field of signal processing and machine learning.

The used main method of the present invention is as follows:

1) Signal reconstruction for sparse or compressible signals.

The NyKuist sampling theorem determines that the sampling rate of the signal in the traditional signal processing method is at least twice the signal bandwidth. However, with the rapid development of telecommunication technology, the bandwidth required for signal transmission has become wider and wider, and the existing hardware facilities have been difficult to achieve. To meet the requirements, the resulting signal data that is lower than the standard adopted frequency is called a sparse signal. Sparse signals require reconstruction algorithms to reconstruct densely sampled signals to form ordered, accurate time series data that can be used for machine learning.

Signal reconstruction techniques can be divided into three types: combinatorial optimization algorithms, convex optimization algorithms and greedy pursuit algorithms. Among them, the greedy tracking method is a relatively new algorithm, and takes into account the complexity and reconstruction accuracy, and is the focus of this project. Greedy pursuit algorithm can be divided into non-backtracking algorithm represented by Orthogonal Matching Pursuit (OMP) algorithm and backtracking algorithm represented by Subspace Pursuit (SP) algorithm. Compared with the non-backtracking algorithm, the backtracking algorithm has higher reconstruction accuracy. This project will use the backtracking algorithm to reconstruct sparse radio signals.

2) Denoising the reconstructed signal data.

Denoising using mathematical models is a recent research achievement by members of the project team. For nonlinear time series such as radio signal data, we can use the auto-regressive model with exogenous variables (ARX) model or the unscented Kalman Filter (UKF) model for denoising. The mathematical models produced during the denoising process can also be used for classification and prediction.

3) Label and classify incompletely labeled signals.

The radio signal datasets obtained from the original database are often incompletely labeled, and some datasets even have only a few signals labeled. At this point, the training dataset needs to be labeled and classified. Traditional semi-supervised classification methods include active learning and deep belief nets in deep learning. In this project, we will combine the above two methods, and use genetic algorithm to evaluate the correlation between spectral signals to label and classify the training data, in order to obtain a more complete classifier.

4) Classify and predict the spectrum signal.

Using the classifier formed in the previous step, the unknown spectral signal can be classified and predicted. In the prediction process, the bad influence of the mathematical model generated in the denoising process can also be added, so as to make the prediction result more accurate.

Although the embodiments are described and described separately above, some common technologies are involved, and in the opinion of those of ordinary skill in the art, they can be replaced and integrated between the embodiments. Reference is made to another example described.

The above-mentioned embodiments do not constitute a limitation on the protection scope of the technical solution. Any modifications, equivalent replacements and improvements made within the spirit and principles of the above-mentioned embodiments shall be included within the protection scope of this technical solution.

Claims (5)

1. A signal feature extraction method based on deep learning is characterized in that: the method comprises the following steps:

1) compressing the original sparse spectrum signal to form a time sequence spectrum signal, and completing denoising processing after compression to obtain a reconstructed time sequence spectrum signal serving as training data;

2) marking the unmarked time sequence spectrum signals in the training data by using a semi-supervised learning method; simulating each group of time sequence spectrum signals, and measuring the correlation among the time sequence spectrum signals by using a genetic algorithm;

3) and classifying and predicting each group of time sequence spectrum signals by using the marked training data and a deep learning method.

2. The signal feature extraction method based on deep learning of claim 1, characterized in that: in the step 1), an original sparse spectrum signal is compressed by adopting a dynamic subspace tracking algorithm in a greedy tracking algorithm, the compressed sparse spectrum signal is subjected to denoising processing by adopting an unscented kalman filtering algorithm, or a reconstructed time sequence spectrum signal is directly predicted by utilizing the unscented kalman filtering algorithm, and a first prediction result is obtained.

3. The signal feature extraction method based on deep learning according to claim 2, characterized in that: the semi-supervised learning method comprises an active learning method and a deep belief network in deep learning, wherein the active learning method is used for marking incompletely marked training data, the deep belief network is a probability generation model, the probability generation model comprises a plurality of limiting type Boltzmann machines, joint distribution between observation data and labels is established, the limiting type Boltzmann machines are composed of a visible layer and a hidden layer, the layers are connected, the units in the layers are not connected, and the deep belief network classifies and predicts the training data to obtain a classification result and a second prediction result.

4. The signal feature extraction method based on deep learning according to claim 3, characterized in that: the active learning method is combined in the deep belief network as a first hidden layer.

5. The signal feature extraction method based on deep learning of claim 4, characterized in that: firstly, setting the energy of each layer to an extreme value by utilizing a training restricted Boltzmann machine, and adjusting the weight of each layer by utilizing an actively learned label result after all the restricted Boltzmann machines are trained to finally obtain a classification surface of a deep belief network; when the depth belief network is used for prediction, all training data are converted into residual errors in the unscented Kalman filtering algorithm model, the depth belief network is used as a machine learning model to predict the change curve of the residual errors in the unscented Kalman filtering algorithm model, and finally, the prediction result of the depth belief network is combined with the prediction result of the unscented Kalman filtering algorithm to obtain a final result.

Images