Signal classification method and system based on multi-feature fusion and its application

Abstract

The present invention relates to a signal classification method and system for multi-feature fusion and application thereof, comprising a model training stage and a model application stage realized by a deep learning algorithm, characterized in that, in the model training stage, time domain features of an original signal, frequency domain features of an original signal and time domain features of a pre-processed original signal are fused and then sent to a training model for training; the pre-processing comprises the following contents: sending n types of training sample data to a Mel filter group for filtering according to categories, wherein the Mel filter group is provided with N filters; performing feature extraction processing on n types of filtered data respectively output by the N filters in the Mel filter group, and respectively calculating correlation parameters of behavioral features in each type of data after filtering by 1, 2, ... N filters; comparing N groups of correlation parameters, obtaining a filter with the smallest correlation parameter as a filter for model pre-processing; and using the obtained filter to process the signal.

Description

Technical field:

[0001] The present invention relates to the field of data processing technology, and in particular to a signal classification method and system that is particularly suitable for signal sequence classification processing, can fully explore the features of data to be processed, and thus improve the accuracy of data classification, and its application. Background technology:

[0002] With the progress of society and the improvement of science and technology, all walks of life are making full use of large amounts of data and information. Intelligent products are getting closer and closer to daily life. At the same time, people's demand for improving their living standards is becoming increasingly strong. Based on the existing big data, the use of artificial intelligence technology to improve living standards has become an increasingly hot direction.

[0003] Artificial intelligence technology uses existing sensors to collect unknown signals, automatically analyze and classify them. Common methods for classifying unknown signals include deep learning and machine learning, both of which are widely used in current artificial intelligence. Traditional machine learning solutions mostly process signals, extract features, and put them into classifiers for training. Common machine learning algorithms include support vector machines, decision trees, random forests, gradient boosting trees, etc. Deep learning solutions are more flexible and can put raw data or data features into classifiers for model training. Common deep learning methods include convolutional neural networks, recurrent neural networks, deep belief networks, etc.

[0004] Artificial intelligence solutions are increasingly used in various fields for signal classification due to their advantages of speed, accuracy, and cost saving. However, when current artificial intelligence solutions are applied to the field of signal classification, there is often a problem of insufficient comprehensive signal feature extraction. For example, most existing algorithms only extract features from the original signal or the filtered signal, without integrating multiple features. In addition, although filters are also set in the existing technology for data preprocessing, the specific setting of the filters lacks pertinence, and the degree of improvement of the classification performance by the set filters is limited. Summary of the invention:

[0005] In view of the shortcomings and deficiencies in the prior art, the present invention proposes a signal classification method and system and its application that are particularly suitable for signal sequence classification processing, which can fully explore the characteristics of the data to be processed and thus improve the accuracy of data classification.

[0006] The present invention is achieved by the following measures:

[0007] A signal classification method for multi-feature fusion includes a model training stage and a model application stage implemented by a deep learning algorithm, characterized in that in the model training stage, the time domain features of the original signal, the frequency domain features of the original signal, and the time domain features of the pre-processed original signal are fused and sent to a training model for training; the pre-processing includes the following contents:

[0008] Step 1-1: sending n types of training sample data to a Mel filter bank for filtering according to categories, wherein the Mel filter bank is provided with N filters;

[0009] Step 1-2: Perform feature extraction processing on the n types of filtered data output by the N filters in the Mel filter bank, and calculate the correlation parameters of the behavioral features in each type of data after filtering by 1, 2, ... N filters;

[0010] Step 1-3: Compare N groups of correlation parameters and obtain the filter with the smallest correlation parameter as the filter for model preprocessing;

[0011] Step 1-4: Process the signal using the filter obtained in step 1-3.

[0012] The time domain feature extraction of the original signal in the model training stage of the present invention refers to: the mean, variance, maximum, minimum, kurtosis of the signal and the maximum value of every 5 points; the frequency domain feature extraction of the original signal refers to: the maximum, minimum, mean, variance, kurtosis of the signal spectrum amplitude and the amplitude of 0-1kHz on the extracted spectrum; the time domain feature extraction of the original signal after preprocessing refers to: the mean, variance, maximum, minimum, kurtosis of the signal and the maximum value of every 5 points.

[0013] The model implemented by the deep learning algorithm described in the present invention adopts a convolutional neural network model to train the feature data and its corresponding behavior type labels, and optimizes the network structure to obtain the required classification model.

[0014] The model application stage of the present invention refers to using the obtained classification model to classify unknown type data.

[0015] The present invention also proposes a signal classification system with multi-feature fusion, which is characterized in that it is provided with a data preprocessing module, a feature extraction module, a multi-feature fusion module, and a convolutional neural network model, wherein the output end of the data preprocessing module is connected to the feature extraction module; the feature extraction module is provided with an original data time domain feature extraction module, an original data frequency domain feature extraction module and a preprocessed data time domain feature extraction module; the output end of the feature extraction module is connected to the multi-feature fusion module, and the output end of the multi-feature fusion module is connected to the input end of the convolutional neural network model; the data preprocessing module is provided with a Mel filter group for processing n types of behavior data, the N filters provided in the Mel filter group respectively accept the n types of behavior data and respectively output the filtered data to the preprocessing feature extraction module, the back end of the preprocessing feature extraction module is respectively connected to N feature correlation parameter calculation modules, the N feature correlation calculation modules respectively output the correlation parameters of the n types of signal filtered data to the correlation parameter comparison module, and the correlation parameter comparison module determines the filter with the smallest correlation parameter.

[0016] The present invention also proposes an application of a signal classification method of multi-feature fusion, which is characterized by comprising the following steps:

[0017] Step 1: Preprocess the sample data, including:

[0018] Step 1-1: sending n types of training sample data to a Mel filter bank for filtering according to categories, wherein the Mel filter bank is provided with N filters;

[0019] Step 1-2: Perform feature extraction processing on the n types of filtered data output by the N filters in the Mel filter bank, and calculate the correlation parameters of the behavioral features in each type of data after filtering by 1, 2, ... N filters;

[0020] Step 1-3: Compare N groups of correlation parameters and obtain the filter with the smallest correlation parameter as the filter for model preprocessing;

[0021] Step 1-4: Process the signal using the filter obtained in step 1-3;

[0022] Step 2: Extract the time domain features and frequency domain features of the original sample data, and the time domain features of the preprocessed sample data respectively;

[0023] Step 3: Send the data feature data of the behavior fragments obtained in step 2 and their corresponding behavior type labels to the convolutional neural network for training, and optimize the network structure and its parameters to obtain the best classifier and the required classification model;

[0024] Step 4: Use the obtained classification model to make category judgment on unknown category signals.

[0025] The convolutional neural network described in step 3 of the present invention is specifically:

[0026] The data enters a one-dimensional convolutional layer, a normalization layer, and then a pooling layer with a step size of 2 and a pooling size of 5. The above layers are repeated three times to build a three-layer convolutional network, and then pass through a flatten layer, a fully connected layer, a relu activation function, a random dropout layer, a fully connected layer, a random dropout layer, and a softmax layer.

[0027] In step 3 of the present invention, the cross entropy loss function is used to evaluate the deviation between the current training result of the model and the actual data situation. When training the network, the Adam optimization method is used, the learning rate learning_rate = 0.001, beta1 = 0.9, beta2 = 0.999, epochs is set to 20, batch_size is 500, and the network is trained by the above parameters to obtain the required classification model.

[0028] Compared with the prior art, the present invention increases the screening of filters in the process of data preprocessing, and pre-screens a filter so that different categories of signals have the highest feature distinction after passing through the filter, which plays a positive role in subsequent signal classification; in addition, multiple types of features of the signal are fused. Compared with the prior art which mostly classifies features extracted from original signals or filtered signals, the present invention fuses the time and frequency domain features of the original signal with the time domain features of the signal after filtering by the screened filter, thereby increasing the number of effective features and achieving full mining and utilization of the original signal features. Description of the drawings:

[0029] Attached Figure 1 Schematic diagram of the principle of preprocessing in the signal classification method of multi-feature fusion of the present invention. Figure 2 It is a schematic diagram of the signal classification method of multi-feature fusion of the present invention. Specific implementation method:

[0030] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0031] As attached Figure 1 As shown, the present invention proposes a signal classification method for multi-feature fusion, including a model training stage and a model application stage implemented by a deep learning algorithm, characterized in that, in the model training stage, the time domain features of the original signal, the frequency domain features of the original signal, and the time domain features of the pre-processed original signal are fused and sent to the training model for training; the pre-processing includes the following contents:

[0032] Step 1-1: sending n types of training sample data to a Mel filter bank for filtering according to categories, wherein the Mel filter bank is provided with N filters;

[0033] Step 1-2: Perform feature extraction processing on the n types of filtered data output by the N filters in the Mel filter bank, and calculate the correlation parameters of the behavioral features in each type of data after filtering by 1, 2, ... N filters;

[0034] Step 1-3: Compare N groups of correlation parameters and obtain the filter with the smallest correlation parameter as the filter for model preprocessing;

[0035] Step 1-4: Process the signal using the filter obtained in step 1-3.

[0036] The time domain feature extraction of the original signal in the model training stage of the present invention refers to: the mean, variance, maximum, minimum, kurtosis of the signal and the maximum value of every 5 points; the frequency domain feature extraction of the original signal refers to: the maximum, minimum, mean, variance, kurtosis of the signal spectrum amplitude and the amplitude of 0-1kHz on the extracted spectrum; the time domain feature extraction of the original signal after preprocessing refers to: the mean, variance, maximum, minimum, kurtosis of the signal and the maximum value of every 5 points. The model implemented by the deep learning algorithm of the present invention adopts a convolutional neural network model to train the feature data and its corresponding behavior type labels, and optimizes the network structure to obtain the required classification model.

[0037] The model application stage of the present invention refers to using the obtained classification model to classify unknown type data.

[0038] The present invention also proposes a signal classification system with multi-feature fusion, which is characterized in that it is provided with a data preprocessing module, a feature extraction module, a multi-feature fusion module, and a convolutional neural network model, wherein the output end of the data preprocessing module is connected to the feature extraction module; the feature extraction module is provided with an original data time domain feature extraction module, an original data frequency domain feature extraction module and a preprocessed data time domain feature extraction module; the output end of the feature extraction module is connected to the multi-feature fusion module, and the output end of the multi-feature fusion module is connected to the input end of the convolutional neural network model; the data preprocessing module is provided with a Mel filter group for processing n types of behavior data, the N filters provided in the Mel filter group respectively accept the n types of behavior data and respectively output the filtered data to the preprocessing feature extraction module, the back end of the preprocessing feature extraction module is respectively connected to N feature correlation parameter calculation modules, the N feature correlation calculation modules respectively output the correlation parameters of the n types of filtered data to the correlation parameter comparison module, and the correlation parameter comparison module determines the filter with the smallest correlation parameter.

[0039] As attached Figure 2 As shown, the present invention also proposes an application of a signal classification method of multi-feature fusion, which is characterized by comprising the following steps:

[0040] Step 1: Preprocess the sample data, including:

[0041] Step 1-1: sending n types of training sample data to a Mel filter bank for filtering according to categories, wherein the Mel filter bank is provided with N filters;

[0042] Step 1-2: Perform feature extraction processing on the n types of filtered data output by the N filters in the Mel filter bank, and calculate the correlation parameters of the behavioral features in each type of data after filtering by 1, 2, ... N filters;

[0043] Step 1-3: Compare N groups of correlation parameters and obtain the filter with the smallest correlation parameter as the filter for model preprocessing;

[0044] Step 1-4: Process the signal using the filter obtained in step 1-3;

[0045] Step 2: Extract the time domain features and frequency domain features of the original sample data, and the time domain features of the preprocessed sample data respectively;

[0046] Step 3: Send the data feature data of the behavior fragments obtained in step 2 and their corresponding behavior type labels to the convolutional neural network for training, and optimize the network structure and its parameters to obtain the best classifier and the required classification model;

[0047] Step 4: Use the obtained classification model to make category judgment on unknown category signals.

[0048] The convolutional neural network described in step 3 of the present invention is specifically as follows: the data enters a one-dimensional convolution layer, a normalization layer, followed by a pooling layer, the pooling layer step size is 2, the pooling size is 5, the above layers are repeated three times, a three-layer convolutional network is constructed, and then passes through a flatten layer, a fully connected layer, a relu activation function, a random dropout layer, a fully connected layer, a random dropout layer, and a softmax layer.

[0049] In step 3 of the present invention, the cross entropy loss function is used to evaluate the deviation between the current training result of the model and the actual data situation. When training the network, the Adam optimization method is used, the learning rate learning_rate = 0.001, beta1 = 0.9, beta2 = 0.999, epochs is set to 20, batch_size is 500, and the network is trained by the above parameters to obtain the required classification model.

[0050] Embodiment 1:

[0051] This example proposes an application of a multi-feature fusion signal classification method in the classification and processing of optical fiber vibration signals, specifically:

[0052] Step 1: Collect signal sequences of different categories, record corresponding behavior types, and build a training data sample library;

[0053] Step 2: From the training data sample library in step 1, randomly select n samples of each type of signal and pass them through a preset Mel filter;

[0054] Since the human ear has different sensitivities to different frequencies and the relationship is nonlinear, we divide the spectrum into multiple Mel filter groups according to the sensitivity of the human ear. Within the Mel scale range, the center frequencies of each filter are linearly distributed with equal intervals, but they are not equally spaced in the FFT frequency range. The relationship between Mel frequency and FFT frequency conversion is as follows:

[0055] F mel (f) = 1125ln (1 + f / 700)

[0056] The method for calculating the upper and lower frequency limits of the Mel filter is as follows:

[0057] Step 2-1: Determine the lowest (generally 0 Hz) and highest (generally half of the sampling rate) frequencies of the signal and the number of Mel filters. In this embodiment, the sampling rate of the optical fiber vibration data is 16 kHz, and the number of Mel filters is 24;

[0058] Step 2-2: Calculate the Mel frequency corresponding to the lowest and highest frequencies. According to the form and parameters of the signal in this embodiment, the upper limit of the lowest Mel frequency is 0 and the lower limit is 2835;

[0059] Step 2-3: Calculate the distance between the center frequencies of two adjacent Mel filters - (highest Mel frequency - lowest Mel frequency) / (number of filters + 1). According to the form of the signal in this embodiment, the distance between the center frequencies of two adjacent Mel filters is 113.4, so the center Mel frequencies of the Mel filter group obtained are 0, 113.4, 226.8, 340.2, 453.6, 567, 680.4, 793.8, 907.2, 1020.6, 1134, 1247.4, 1360.8, 1474.2, 1587.6, 1701, 1814.4, 1927.8, 2041.2, 2154.6, 2268, 2381.4, 2494.8, 2608.2, 2721.6, 2835.

[0060] Step 2-4: Convert each center Mel frequency into frequency - According to the conversion relationship between Mel frequency and FFT frequency,

[0061] In this embodiment, according to the Mel frequencies obtained above, the corresponding FFT frequencies are 0, 74.24, 156.35, 247.17, 347.62, 458.73, 581.62, 717.54, 867.88, 1034.16, 1218.08, 1421.5, 1646.5, 1895.36, 2170.61, 2475.05, 2811.78, 3184.23, 3596.17, 4051.8, 4555.75, 5113.15, 5729.67, 6411.57, 7165.79, 8000;

[0062] Step 2-5: Calculate the subscript of the frequency corresponding to the FFT midpoint - According to the above steps, the upper and lower limit frequencies of the 24 filters can be obtained respectively:

[0063] [0,156.35],[74.24,247.17],[156.35,347.62],[247.17,458.73],[347.62,581.62],[458.73,717.54],[581.62,867.88],[717.54,1034.16],[867.88,1218.08],[1034.16,1421.5],[1218.08,1646.5],[1421.5,1895.36],[1646.5,2 170.61],[1895.36,2475.05],[2170.61,2811.78],[2475.05,3184.23],[2811.78,3596.17],[3184.23,4051.8],[3596.17,4555.75],[4051.8,5113.15],[4555.75,5729.67][5113.15,6411.57],[5729.67,7165.79],[6411.57,8000].

[0064] Step 3: Extract features of various types of signals that have passed through the Mel filter group in step 2, and use the correlation coefficient as a measurement indicator to compare the size of the correlation coefficients of the features of different types of signals after passing through the filter group, and select a bandpass filter (156Hz-347Hz) with the smallest correlation parameter of different types of signal features (the higher the discrimination of different types of signals, the smaller the correlation parameter) as the filter used this time, and process the signal as part of the feature extraction module;

[0065] Step 3-1: In this embodiment, the mean, variance, maximum, minimum, kurtosis and the maximum value of every 5 points of the signal are taken as the features of the signal;

[0066] Step 3-2: Pass the original signal through multiple preset Mel filters respectively;

[0067] Step 3-3: Extract features of the signals after filtering;

[0068] Step 3-4: Calculate the correlation coefficients of various signal features after passing through each filter, and compare the correlation coefficients after passing through different filters; select a filter with the smallest correlation coefficient as part of the feature extraction module for signal processing.

[0069] Step 4: Send the data of the training data sample library in step 1 to the feature extraction module. In the feature extraction module, the time domain features, frequency domain features and the features of the signal after filtering by the selected filter are integrated to obtain more comprehensive features;

[0070] In this embodiment, the original data is used in the feature extraction module to extract the time domain features (mean, variance, maximum, minimum, kurtosis and the maximum value of every 5 points of the signal), frequency domain features (maximum value, minimum value, mean, variance, kurtosis of the signal spectrum amplitude and the amplitude of 0-1kHz on the extracted spectrum) and the time domain features of the signal after filtering by the screened filter, and the above three types of features are fused to make full use of the original signal.

[0071] Step 5: Send the feature data of the behavior fragments obtained in step 4 and their corresponding behavior type labels into the convolutional neural network for training, optimize the network structure and its parameters, select the best classifier, and obtain the required classification model.

[0072] In this embodiment, a convolutional neural network is used as a classifier to classify the signal. During model training, feature extraction is first performed on data of different categories in the manner of step four, and then the features of the data and the labels corresponding to their behavior categories are placed in the convolutional neural network for model training: first, the data enters a one-dimensional convolutional layer with 16 filters, a step size of 2, a convolution kernel size of 5, followed by a relu activation function, and then enters a normalization layer, followed by a pooling layer with a step size of 2 and a pooling size of 5. The above layers are repeated three times to construct a three-layer convolutional network, and then pass through a flatten layer, a fully connected layer (with 128 neurons), a relu activation function, a random dropout layer (with a threshold set to 0.5), a fully connected layer (with 64 neurons), a random dropout layer (with a threshold set to 0.5), and a softmax layer (with 2 neurons). The above is the network structure used for this data training.

[0073] During the model training process, the cross entropy loss function (categorical_cross entropy) is used to evaluate the deviation between the current training results of the model and the actual data situation. When training the network, the Adam optimization method is used, the learning rate learning_rate=0.001, beta1=0.9, beta2=0.999, epochs is set to 20, batch_size is set to 500, and the network is trained with the above parameters to obtain the required classification model.

[0074] Compared with the prior art, the present invention increases the screening of filters in the process of data preprocessing, and pre-screens a filter so that different categories of signals have the highest feature distinction after passing through the filter, which plays a positive role in subsequent signal classification; in addition, multiple types of features of the signal are fused. Compared with the prior art which mostly classifies features extracted from original signals or filtered signals, the present invention fuses the time and frequency domain features of the original signal with the time domain features of the signal after filtering by the screened filter, thereby increasing the number of effective features and achieving full mining and utilization of the original signal features.

Claims

1. A signal classification method for multi-feature fusion, comprising a model training phase and a model application phase implemented by a deep learning algorithm, characterized in that: In the model training stage, the time domain features of the original signal, the frequency domain features of the original signal, and the time domain features of the pre-processed original signal are fused and sent to the training model for training; the pre-processing includes the following contents: Step 1-1: sending n types of training sample data to a Mel filter bank for filtering according to the categories, wherein the Mel filter bank is provided with N filters; Step 1-2: Perform feature extraction processing on the n types of filtered data output by the N filters in the Mel filter bank, and calculate the correlation coefficients of the behavioral features in each type of data after filtering by 1, 2, ... N filters; Step 1-3: Compare N groups of correlation coefficients and obtain the filter with the smallest correlation coefficient as the preprocessing filter. The higher the discrimination of different types of signals, the smaller the correlation coefficient. Step 1-4: Use the filter obtained in step 1-3 to process other signals.

2. The signal classification method of multi-feature fusion according to claim 1 is characterized in that: The time domain feature extraction, frequency domain feature extraction and filtered time domain feature extraction of the original signal in the model training stage refer to the extraction of the statistical characteristics of the signal and the features after downsampling, and the fusion of multiple types of features.

3. The signal classification method of multi-feature fusion according to claim 1 is characterized in that: The model implemented by the deep learning algorithm adopts a convolutional neural network model to train the feature data and its corresponding behavior type labels, and optimizes the network structure to obtain the required classification model.

4. The signal classification method of multi-feature fusion according to claim 1 is characterized in that: The model application stage refers to using the obtained classification model to classify unknown type data.

5. A system for executing the signal classification method of multi-feature fusion according to any one of claims 1 to 4, characterized in that: A data preprocessing module, a feature extraction module, a multi-feature fusion module, and a convolutional neural network model are provided, wherein the output end of the data preprocessing module is connected to the feature extraction module; the feature extraction module is provided with an original data time domain feature extraction module, an original data frequency domain feature extraction module, and a preprocessed data time domain feature extraction module; the output end of the feature extraction module is connected to the multi-feature fusion module, and the output end of the multi-feature fusion module is connected to the input end of the convolutional neural network model; the data preprocessing module is provided with a Mel filter group for processing n types of behavior data, the N filters provided in the Mel filter group respectively accept the n types of behavior data and respectively output the filtered data to the preprocessing feature extraction module, the back end of the preprocessing feature extraction module is respectively connected to N feature correlation coefficient calculation modules, the N feature correlation calculation modules respectively output the correlation coefficients of the n types of filtered data to the correlation coefficient comparison module, and the correlation coefficient comparison module determines the filter with the smallest correlation coefficient.

6. An application of the signal classification method of multi-feature fusion according to any one of claims 1 to 4, characterized in that: The following steps are involved: Step 1: Collect signal sequences of different categories, record corresponding behavior types, build a training data sample library, and preprocess the sample data, including: Step 1-1: sending n types of training sample data to a Mel filter bank for filtering according to categories, wherein the Mel filter bank is provided with N filters; Step 1-2: Perform feature extraction processing on the n types of filtered data output by the N filters in the Mel filter bank, and calculate the correlation coefficients of the behavioral features in each type of data after filtering by 1, 2, ... N filters; Step 1-3: Compare N groups of correlation coefficients and obtain the filter with the smallest correlation coefficient as the filter for model preprocessing; Step 1-4: Process the signal using the filter obtained in step 1-3; Step 2: Extract the time domain features and frequency domain features of the original sample data, and the time domain features of the preprocessed sample data respectively; Step 3: Send the data feature data of the behavior fragments obtained in step 2 and their corresponding behavior type labels to the convolutional neural network for training, and optimize the network structure and its parameters to obtain the best classifier and the required classification model; Step 4: Use the obtained classification model to make category judgment on unknown category signals.

Images