Signal classification method and device

Abstract

The present application provides a signal classification method and device. After obtaining a signal to be classified, the method performs a wavelet transform on the signal to be classified, and generates input data according to the original data of the signal to be classified and the wavelet transform result. The input data is then input into a neural network model to calculate the output classification probability through the neural network model. The input data includes the original data of the signal to be classified, the transformed data obtained after the signal to be classified is subjected to wavelet transform, and the composite data formed by stacking the original data and the transformed data. The time series relationship in the signal to be classified can be retained, and the characteristics of the data input into the neural network model can be increased, thereby improving the accuracy of signal classification.

Description

Technical Field

[0001] The present application relates to the technical field of signal classification, and in particular to a signal classification method and device. Background Art

[0002] Signal classification is a signal processing process that determines the category of a signal by analyzing its characteristics. For example, in a perimeter protection system, by collecting detection signals of targets within the perimeter area and analyzing the characteristics of the detection signals, signal fluctuations caused by human intrusion and natural factors can be distinguished. Since the complexity of signal characteristics varies in different application fields, and the signals in some application scenarios are highly complex, simple signal feature analysis is usually unable to accurately determine the category to which the signal belongs.

[0003] In order to classify complex signals, a neural network model can be used in the classification process. A typical neural network model obtains the model output result by inputting the training signal data into the network model. The model parameters are updated by back propagation based on the model data result and the label information. After multiple trainings, a neural network model that meets the current field is obtained. When classifying signals, the signal to be classified can be input into the neural network model in the form of a vector, and the classification probability of the signal to be classified can be output by the neural network model to finally determine the category of the signal.

[0004] It can be seen that in the above classification process, the specific values ​​in the signal to be classified need to be arranged to form vector data. Commonly used network models, such as convolutional neural networks, treat the signal to be classified as image data and input it into the network model. However, most signals have a temporal relationship, that is, the specific values ​​in the signal are correlated in time. If it is directly used as the input of the network model without processing, this temporal relationship will not be effectively preserved, thereby reducing the accuracy of signal classification. Summary of the invention

[0005] The present application provides a signal classification method and device based on wavelet transform to solve the problem that the traditional signal classification method cannot effectively retain the time series relationship, resulting in reduced classification accuracy.

[0006] On the one hand, the present application provides a signal classification method based on wavelet transform, comprising:

[0007] Acquire a signal to be classified detected by a detection device, wherein the signal to be classified is a time sequence signal;

[0008] Generate input data, the input data including original data of the signal to be classified, transformed data obtained after the signal to be classified is subjected to wavelet transform, and composite data formed by stacking the original data and the transformed data;

[0009] Inputting the input data into a neural network model, wherein the neural network model includes a plurality of LSTM networks, a deep residual network, and a fully connected network;

[0010] An intrusion classification probability corresponding to the signal to be classified is obtained, where the intrusion classification probability is an output result of the neural network model.

[0011] Optionally, the steps of generating input data include:

[0012] Converting the raw data into vector form from the signal to be classified;

[0013] Performing wavelet transform on the signal to be classified to obtain transformed data, wherein the transformed data has the same vector length as the original data;

[0014] The original data and the transformed data are stacked to form the composite data.

[0015] Optionally, the input layer of the neural network model includes a first branch and a second branch, the first branch includes two first LSTM networks for inputting the original data and the transformed data respectively; the second branch includes a second LSTM network, and the number of hidden units of the second LSTM network is the sum of the number of hidden units of the two first LSTM networks in the first branch.

[0016] Optionally, the step of inputting the input data into the neural network model includes:

[0017] Inputting the original data and the transformed data into two first LSTM networks respectively to obtain output results of two first LSTM networks;

[0018] Stacking the output results of the two first LSTM networks to generate first result data;

[0019] Inputting the composite data into the second LSTM network to obtain an output result of the second LSTM network;

[0020] The first result data and the output result of the second LSTM network are stacked to generate second result data.

[0021] Optionally, the step of inputting the input data into the neural network model further includes:

[0022] Inputting the second result data into the deep residual network;

[0023] Extracting features from the second result data using the deep residual network;

[0024] Calculate the classification probability based on the extracted features using the fully connected network and the softmax activation function;

[0025] Output the classification probability.

[0026] Optionally, the deep residual network includes multiple residual blocks and transition blocks, and the multiple residual blocks are connected through the transition blocks; each of the residual blocks includes multiple convolutional layers and multiple activation function layers;

[0027] Optionally, the method further includes training the neural network model, specifically including:

[0028] Acquire signal sample data, wherein the signal sample data includes a training label;

[0029] Processing the signal sample data by wavelet transform to generate training data;

[0030] Inputting the training data into the initialized neural network model to obtain an output result;

[0031] The model parameters of the neural network model are adjusted according to the output result and the training label.

[0032] Optionally, the method further includes:

[0033] Obtaining judgment data, wherein the judgment data is a combination of one or more of the training set accuracy, training set loss, validation set accuracy, and validation set loss output by the neural network model;

[0034] If the judgment data meets the preset training requirements, the model parameters are output.

[0035] Optionally, the step of adjusting the model parameters of the neural network model according to the output result and the training label includes:

[0036] Calculate the difference between the training labels in the training data and the output results using a loss function;

[0037] With the goal of minimizing the difference, the model parameters of the neural network model are iteratively adjusted through a back-propagation algorithm.

[0038] On the other hand, the present application also provides a signal classification device, including: an acquisition module, a conversion module, a model input module and an output module.

[0039] The acquisition module is used to acquire the signal to be classified detected by the detection device, and the signal to be classified is a time sequence signal;

[0040] The conversion module is used to generate input data, wherein the input data includes original data of the signal to be classified, transformed data obtained after the signal to be classified is subjected to wavelet transformation, and composite data formed by stacking the original data and the transformed data;

[0041] The model input module is used to input the input data into a neural network model, wherein the neural network model includes multiple LSTM networks, a deep residual network, and a fully connected network;

[0042] The output module is used to obtain the intrusion classification probability corresponding to the signal to be classified, and the intrusion classification probability is the output result of the neural network model.

[0043] It can be seen from the above technical solutions that the present application provides a signal classification method and device. After obtaining the signal to be classified, the method performs a wavelet transform on the signal to be classified, and generates input data according to the original data of the signal to be classified and the wavelet transform result. The input data is then input into the neural network model to calculate the output classification probability through the neural network model. Among them, the input data includes the original data of the signal to be classified, the transformed data obtained after the signal to be classified is subjected to the wavelet transform, and the composite data formed by stacking the original data and the transformed data. The time series relationship in the signal to be classified can be retained, and the characteristics of the data input into the neural network model can be increased, thereby improving the accuracy of signal classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solution of the present application, the drawings required for use in the embodiments are briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0045] Figure 1 This is a schematic diagram of an application scenario of a signal classification device in an embodiment of the present application;

[0046] FIG. 2( a ) is a schematic diagram of a signal classification method flow in an embodiment of the present application;

[0047] FIG2( b ) is a schematic diagram of a perimeter security system in an embodiment of the present application;

[0048] FIG2( c ) is a schematic diagram of another perimeter security system in an embodiment of the present application;

[0049] Figure 3 A schematic diagram of a process for generating input data in an embodiment of the present application;

[0050] Figure 4 A schematic diagram of a process of inputting input data into a neural network model in an embodiment of the present application;

[0051] Figure 5 A schematic diagram of a process for outputting classification probabilities in an embodiment of the present application;

[0052] Figure 6 A schematic diagram of the process of training a neural network model in an embodiment of the present application;

[0053] Figure 7 A schematic diagram of a process for outputting model parameters in an embodiment of the present application;

[0054] Figure 8 Schematic diagram of the structure of the signal classification device in the embodiment of the present application. DETAILED DESCRIPTION

[0055] The following embodiments are described in detail, and examples thereof are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following embodiments do not represent all implementations consistent with the present application. They are only examples of systems and methods consistent with some aspects of the present application as detailed in the claims.

[0056] In the embodiment of the present application, signal classification is a signal processing process that determines the category to which the signal belongs by analyzing the features in the signal data. Here, the signal refers to an electrical signal detected by a front-end device such as a sensor in actual engineering. Figure 1 As shown, the signal can be converted into a digital signal that can be read and processed by a control device such as a computer or a controller through A / D conversion. The signal can also be processed by noise reduction and impurity removal before being converted into a digital signal to obtain a signal with better quality.

[0057] It should be noted that in some signal processing processes, the electrical signals detected by the front-end equipment have a time-series relationship. That is, the digital signal can be composed of multiple specific values ​​arranged in time order, that is, the digital signal can be further converted into signal data after encoding, mixing, etc., and the signal data is represented by a vector. Obviously, in each vector, multiple signal original values ​​with a time-series relationship are included to represent the specific data content of the signal at a certain moment.

[0058] In some application scenarios, the signal can be generated by multiple front-end devices, or generated by one front-end device but includes multiple specific meanings, so this signal data can also include multiple vectors. For example, for a perimeter security system, sensor devices based on different principles such as laser radar, infrared radar, and acoustic radar can be used to detect the perimeter of the defense zone to obtain a detection signal. In the process of using the detection signal to detect objects that invade the perimeter of the defense zone, different invading objects will generate different signal change characteristics.

[0059] In order to eliminate the influence of irrelevant interference on the system intrusion judgment, the perimeter security system needs to distinguish between "human intrusion" and "non-human interference". Among them, "human intrusion" signals usually include human vertical intrusion signals, human line intrusion signals, etc., and "non-human interference" signals usually include weather change signals, animal intrusion signals, vehicle intrusion signals, etc. Obviously, the signal characteristics detected by the above different types of intrusion methods are different, and because the intrusion process is a continuous process, the detected signal will continue to change with the intrusion process, that is, it has a time sequence relationship.

[0060] The classification to which the signal belongs can be determined by analyzing the features in the signal data. For example, the signal data may include features such as the moving speed and path of the intruding object that can be determined by analyzing the signal data to determine the type of the intruding object. It is also possible to obtain a more accurate classification result by comprehensively analyzing multiple features in the signal data. For example, the signal data can be input into a pre-trained classification model, and the classification model can be used to calculate the output classification probability, thereby determining which intrusion type the detection signal corresponds to.

[0061] The classification model used can be a convolutional neural network model obtained through deep training of a large amount of sample data. The neural network model can comprehensively analyze multiple features in the signal data for calculation, so that the output classification result has a greater advantage in accuracy.

[0062] It should be noted that the time series relationship refers to the time series order between the specific data values ​​in the vector in the signal data, and / or the correlation between the specific data values ​​in the time dimension, that is, the specific value of the signal data collected at the previous moment is correlated with the specific value of the signal data collected at the next moment. For example, the target area of ​​personnel in the intrusion defense zone will change continuously as the personnel gradually enter the defense zone from the outside. And according to the movement direction and speed of the personnel, the detected target area has a correlation in the time dimension.

[0063] In order to input the signal data into the classification model, the vectors in the signal data can be converted into an image-like form and subsequently processed as an image. However, this processing will cause the temporal relationship within the vector to be lost, so that the temporal relationship between signals is not utilized in the features analyzed by the neural network model.

[0064] To this end, as shown in FIG2(a), some embodiments of the present application provide a signal classification method based on wavelet transform, including:

[0065] S100: Acquire a signal to be classified detected by a detection device.

[0066] The signal classification method provided in the embodiment of the present application can be configured in a signal classification device in a specific application system. The signal classification device can be connected to a front-end device to obtain signal data to be classified through the front-end device. The signal to be classified is a time sequence signal. For example, in a perimeter security system, the front-end device is a detection device capable of acquiring wireless signals (microwave signals), and the detection device can perform detection within the defense zone to obtain a detection signal.

[0067] The acquired detection signal can be sent by the front-end device to the signal classification device, during which signal processing processes such as noise reduction, analog-to-digital conversion, modulation / demodulation, etc. are performed to form the signal data to be classified. The signal data to be classified includes one or more data vectors, each of which is a sequence of multiple numerical values, and the data in the vector has a time series relationship.

[0068] It should be noted that in order to obtain the detection signal, the detection equipment can use different detection principles to obtain the signal. For example, the perimeter security system is composed of a transmitting sensor and a receiving sensor of a microwave radio frequency signal, wherein each transmitting sensor can transmit a wireless signal to s receiving sensors, and correspondingly, each receiving sensor can receive the wireless signals transmitted by t transmitting sensors, thereby forming a wireless network, wherein s and t are positive integers. When an intruder or an interfering object enters the wireless network, the wireless network will experience network fluctuations.

[0069] As shown in Figure 2(b) and Figure 2(c), each transmitting sensor can transmit a wireless signal to three receiving sensors (i.e., s is 3), and accordingly, each receiving sensor can receive wireless signals transmitted by three transmitting sensors (i.e., t is 3). In addition, in order to improve the intrusion detection accuracy of the perimeter security system, the density of the wireless network formed between the transmitting sensor and the receiving sensor can also be increased. Therefore, in Figure 2(c), each transmitting sensor can also transmit a wireless signal to five receiving sensors (i.e., s is 5), and accordingly, each receiving sensor can receive wireless signals transmitted by five transmitting sensors (i.e., t is 5). Of course, according to different detection requirements, the transmitting sensor can also be set to transmit wireless signals to a larger number of receiving sensors, and the embodiments of the present application are not limited to this.

[0070] S200: Generate input data.

[0071] After obtaining the signal to be classified, input data can be generated according to the signal to be classified, that is, the input signal to be classified is processed using wavelet transform to form new input data. Among them, wavelet transform is a signal transformation analysis method, which extracts information from the signal by local transformation of space (time) and frequency. Wavelet transform can fully highlight the characteristics of certain aspects of the signal through transformation, and can perform local analysis of time (space) frequency, and gradually refine the signal (function) at multiple scales through telescoping and translation operations, and finally achieve time subdivision at high frequencies and frequency subdivision at low frequencies, automatically adapting to the requirements of time-frequency signal analysis, so as to focus on any details of the signal.

[0072] The transformed data obtained by wavelet transform can retain the time series relationship in the signal to be classified, so as to facilitate the subsequent analysis of the characteristics of the time series relationship. In order to obtain a more accurate classification result, while performing the wavelet transform, the original data of the signal to be classified can also be retained, and the original data and the transformed data can be recombined to generate input data. That is, the input data includes the original data of the signal to be classified, the transformed data obtained after the signal to be classified is transformed by wavelet transform, and the composite data formed by stacking the original data and the transformed data.

[0073] In this embodiment, input data can be generated according to the signal to be classified by wavelet transform, wherein the input data includes original data, transformed data and composite data. The time series relationship features in the signal to be classified are retained by the transformed data, other features in the signal to be classified except the time series relationship are retained by the original data, and the connection between the time series relationship features and other features is retained by the composite data. Therefore, compared with the traditional signal classification method, the number of features in the classification process can be increased to obtain a more accurate classification result.

[0074] S300: Input the input data into a neural network model.

[0075] After the input data is generated, the input data can be input into the classification model to obtain the classification probability through the classification model. Among them, the classification model can be a pre-trained neural network model. The neural network model includes multiple LSTM networks, deep residual networks and fully connected networks. After the input data enters the neural network model through the input layer, it can be subjected to multiple operations of multiple LSTM networks, deep residual networks and fully connected networks to analyze the signal characteristics, obtain the classification probability, and finally output the classification result through the output layer.

[0076] S400: Obtaining an intrusion classification probability corresponding to the signal to be classified.

[0077] After the input data is input into the neural network model, the result can be output by the neural network model. Since the neural network model in this application is used for signal classification, its output result is the classification probability of the signal to be classified corresponding to each category.

[0078] Obviously, the category with the highest classification probability can be used as the classification result of the signal data to be classified. For example, in the perimeter security system, the probability that a signal data corresponds to the classification category of "personnel intrusion" is 85%, and the probability that the corresponding classification category is "non-human interference" is 15%, so the classification category of the current signal data is determined to be "personnel intrusion". After being processed by the neural network, the probability of the input classification probability being "personnel intrusion" is greater, so it is determined that the current signal data is triggered by human intrusion, so that further measures can be taken to handle related events, such as generating alarm signals.

[0079] It should be noted that the input data into the neural network model can contain data corresponding to multiple vectors. Multiple vectors are selected to distinguish the classification results more effectively, such as distinguishing between vector fluctuations caused by "human intrusion" and vector fluctuations caused by natural factors. The specific number of vectors supported can be set according to the form of the model training data set.

[0080] It can be seen from the above technical solutions that the signal classification method based on wavelet transform provided in the above embodiments can generate input data through wavelet transform after obtaining the signal to be classified, and perform feature analysis on the input data through the neural network model to obtain the classification probability. The method extracts the temporal relationship characteristics of the input samples through wavelet transform, and reconstructs a new input in combination with the original input, so that the data input into the neural network model can retain a more significant temporal relationship, increase data features, and output a more accurate classification probability.

[0081] In order to generate data, such as Figure 3 As shown, in some embodiments of the present application, the step of generating input data further includes:

[0082] S201: Converting the signal to be classified into the original data in vector form;

[0083] S202: Performing wavelet transform on the signal to be classified to obtain transformed data;

[0084] S203: stacking the original data and the transformed data to form the composite data.

[0085] After obtaining the signal to be classified, the classified signal can be copied first, and then one of the signals to be classified can be represented in vector form to convert it into original data. The other signal to be classified is subjected to wavelet transform to obtain transformed data. The wavelet basis used in the wavelet transform can be set according to the features to be retained in the signal to be classified. For example, if the db1 wavelet basis is used, the result returned is a single vector with the same length as the input vector, that is, the transformed data has the same vector length as the original data. If you want to obtain more features, you can choose a different wavelet basis.

[0086] After obtaining the original data and the transformed data respectively, the obtained original data and the transformed data can also be stacked to generate composite data. Since the composite data includes the contents of the original data and the transformed data, the vector length of the composite data is twice that of the original data.

[0087] For example, the wavelet transform result and the original data have the same length, and the two vectors can be directly combined to form a new vector. After wavelet transform, the original data X of size n×1 can be obtained from the signal to be classified. ORI , size is n×1 transformation data X trans The original data and the wavelet transform results are then stacked to form a composite data X of size n×2. cat .

[0088] Similarly, when classifying a multi-vector signal, if each sample has m vectors, then the original data X ORI The size is n×m, the transformed data X trans The size is n×m, and the composite data X after stacking cat The size is n×2m.

[0089] It can be seen from the above technical solution that in the above embodiment, wavelet transform is used to process the input signal to be classified to form new input data to retain the time series relationship in the input data. Since the input data includes original data, converted data and composite data, these data can retain different features in the signal to be classified respectively. Therefore, the original data, converted data and composite data need to be input into the neural network model respectively. In order to adapt to the form of the input data, the input layer of the neural network model includes a first branch and a second branch, and the first branch includes two first LSTM networks for inputting original data and transformed data respectively; the second branch includes a second LSTM network, and the second LSTM network is used to input composite data. Since the composite data is formed by stacking the original data and the transformed data, the size of the composite data is the sum of the size of the original data and the transformed data, so the number of hidden units of the second LSTM network is the sum of the number of hidden units of the two first LSTM networks in the first branch.

[0090] That is Figure 4 As shown, in some embodiments of the present application, the step of inputting the input data into the neural network model includes:

[0091] S301: inputting the original data and the transformed data into two first LSTM networks respectively to obtain output results of two first LSTM networks;

[0092] S302: stacking the output results of the two first LSTM networks to generate first result data;

[0093] S303: Inputting the composite data into the second LSTM network to obtain an output result of the second LSTM network;

[0094] S304: Stack the first result data and the output result of the second LSTM network to generate second result data.

[0095] Among them, the LSTM network (Long Short-Term Memory) is a time recurrent neural network that can be used to process and predict important events with very long intervals and delays in time series. In this embodiment, the LSTM layer in the recurrent neural network can be used to process the input data to extract the features sent to the subsequent convolutional layer. In order to make full use of the three types of input data, the input layer of the neural network model can include multiple branches, each of which uses these inputs as inputs for different branches.

[0096] For example, the first branch of the neural network is divided into two small branches, both of which contain two LSTM layers. ORI As input, the second branch takes the transformed data X after wavelet transformation trans As input. The number of hidden units in the two LSTM networks is the same, for example, both are set to 64. Then after two small branches, for an input vector of length n, we can get two output results of size n×64, which are directly stacked to form the output result of the first branch, with a size of n×128.

[0097] The second branch of the neural network stacks the results, i.e. the composite data X cat As input. The number of hidden units in the LSTM layer in this branch is set to the sum of the number of hidden units in the two LSTM layers in the first branch, that is, 128. After two LSTM layers, the second result data can be obtained, that is, an output of size n×128. The output results of the two branches are stacked up to form data in n×256 format as the input of the convolution layer. Obviously, in order to standardize the final stacked result, the setting of hidden units can be adjusted with the size of n.

[0098] It can be seen from the above technical solution that in this embodiment, multiple branches are set in the input layer of the neural network model to be used for the input of original data, converted data and composite data respectively, and then the processing results of multiple branches are stacked to form data that can adapt to the convolutional network, thereby completing the data input while retaining the timing relationship.

[0099] Based on the above input data input method, such as Figure 5 As shown, in order to obtain the classification probability, the method further includes:

[0100] S401: Inputting the second result data into the deep residual network;

[0101] S402: Extracting features from the second result data using the deep residual network;

[0102] S403: Calculate the classification probability according to the extracted features using the fully connected network and the softmax activation function;

[0103] S404: Output the classification probability.

[0104] After stacking the first result data and the output result of the second LSTM network to generate the second result data, the neural network model can also input the second result data into the deep residual network to extract features from the second result data using the deep residual network. Since the temporal relationship characteristics of the signal are retained in the input data, the deep residual network can extract features related to the temporal relationship from the second result data. The fully connected network and the softmax activation function are then used to calculate the classification probability based on the extracted features, thereby outputting the classification probability.

[0105] The deep residual network includes multiple residual blocks and transition blocks, and the multiple residual blocks are connected by the transition blocks; each residual block includes multiple convolutional layers and multiple activation function layers. Each layer of the deep residual network is used to extract features from the stacked results of the LSTM output. The residual block of the deep residual network consists of an input layer, a convolutional layer, an activation function layer, a convolutional layer, an activation function layer, and a convolutional layer. The final output is the sum of the input and the calculated result. Different residual blocks are connected using transition blocks, which are composed of 1×1 convolutions. Finally, the output is output through the fully connected layer and the softmax layer, and the output result is the probability that the signal belongs to a certain category.

[0106] From the above technical solution, it can be seen that the signal classification method provided in the above embodiment can use the LSTM part of the neural network model to construct two branch networks to extract the input time series features, and after stacking the output results, connect them with the convolutional neural network to further extract features. The neural network model is implemented to extract and analyze features including time series relationships, and obtain the signal classification probability by comprehensively considering the feature content.

[0107] In the above embodiment, the classification probability can be obtained by processing the signal data to be classified using the neural network model, and the adaptability of the neural network model will directly affect the accuracy of the classification result. In order to obtain a neural network model that is more suitable for the current application scenario, such as Figure 6 As shown, before classifying the signal, the neural network model can also be trained, including:

[0108] S501: Acquire signal sample data, where the signal sample data includes training labels;

[0109] S502: Processing the signal sample data by wavelet transform to generate training data;

[0110] S503: Inputting the training data into the initialized neural network model to obtain an output result;

[0111] S504: Adjust model parameters of the neural network model according to the output result and the training label.

[0112] The training process of the neural network model is to construct the initial model, obtain the model output results by inputting sample data, and backpropagate according to the difference between the model output results and the training labels, so as to optimize the model parameters in the neural network model.

[0113] Therefore, when conducting model training, it is necessary to first obtain signal sample data. The signal sample data includes training labels. Signal sample data is signal data with training labels that are obtained in advance based on statistics, data mining, etc., and its training labels are used to indicate the category to which the corresponding signal data belongs. Obviously, in order to adapt to the current application scenario, the source of the signal sample data should also be in the current application scenario. For example, in a perimeter security system, the signal sample data is composed of signal data collected by each front-end device under different target intrusion methods through experiments and statistics. The training label in the signal sample data corresponds to the target intrusion method adopted under the corresponding signal data.

[0114] After acquiring the signal sample data, the signal sample data also needs to be processed by wavelet transform to obtain training data with time series relationship. The specific processing method of processing the signal sample data by wavelet transform is the same as that provided in the above embodiment, which will not be repeated here.

[0115] By performing wavelet transform on the signal sample data, the temporal relationship in the training data can be preserved and more in-depth information can be discovered. Therefore, after the training data is input into the initialized neural network model, the output result with reference to the temporal relationship characteristics can be obtained. Then, based on the difference between the output result and the training label, the difference is back-propagated, and the model parameters of the neural network model are adjusted based on the minimum difference.

[0116] Among them, the loss function can be used to calculate the difference between the training label and the output result in the training data, and the model parameters of the neural network model can be iteratively adjusted through the back propagation algorithm with the goal of minimizing the difference. For example, the input signal sample data is processed using wavelet transform. The result after wavelet transform and the original input are stacked to form composite data for input. The loss function uses the cross entropy function to calculate the difference between the input label and the output result of the neural network. The processed training samples and training labels are input into the initialized neural network model. The original data and the transformed data after wavelet transform are respectively sent to the two small branches of the first branch of the neural network model, and the composite data is input to the second branch, so as to obtain the output result through the neural network model. The network parameters are then iteratively adjusted on the data set through the back propagation algorithm. That is, in the perimeter security system, the signal sample data with the training label of "personnel intrusion" is a positive sample, and the signal sample data with the training label of "non-human intrusion" is a negative sample.

[0117] After inputting signal sample data multiple times and repeatedly adjusting the model parameters, the output results of the neural network model can be made closer and closer to the training labels, that is, the adaptability of the neural network model to the current scenario is getting higher and higher. As the difference between the output results and the training labels becomes smaller and smaller, the model parameters at this time can be output after the difference reaches a certain set threshold, thereby training the initialized neural network model into a neural network model suitable for the current application scenario.

[0118] In order to output the model parameters of the neural network model, such as Figure 7 As shown, in some embodiments of the present application, the method further includes:

[0119] S505: Obtaining judgment data;

[0120] S506: If the judgment data meets the preset training requirements, the model parameters are output.

[0121] The judgment data is a combination of one or more of the training set accuracy, training set loss, validation set accuracy, and validation set loss output by the neural network model. In practical applications, different parameters can be selected as judgment data according to different application scenario requirements, and different training requirements can be set.

[0122] After obtaining the judgment data, the judgment data can be compared with the preset training requirements to determine whether the judgment data meets the preset training requirements. When the judgment data meets the preset training requirements, that is, the difference between the output result under the current model and the sample label is within an acceptable range, the model parameters can be output to complete the model training.

[0123] In addition, the training process will simultaneously output the training set accuracy, training set loss, validation set accuracy, and validation set loss to adjust the model parameters more effectively. The results on the validation set are used as the standard for when to stop training the model. After the training is completed, the final model parameters will be saved or the model will be directly output. For example, if the result of a certain iteration exceeds the previous result, the current model will be automatically saved as the best model at the moment.

[0124] After obtaining the trained neural network model, the neural network model can also be tested. The original data of the test sample, the transformed data after wavelet transformation, and the stacked composite data can be sent to the trained neural network model. The neural network model finally outputs the probability value of the signal to be classified belonging to the category of "human intrusion" and "non-human intrusion" after calculation.

[0125] It can be seen from the above technical solutions that the signal classification method based on wavelet transform provided in the above embodiments can use wavelet transform to retain the temporal relationship characteristics in the input data and explore deep information, and reconstruct new inputs in combination with the original data. Two branch networks are formed through the LSTM network to extract the temporal characteristics in the input data, and the output is stacked and connected to the convolutional neural network to complete the feature analysis in the data. The method supports single vector input and multi-vector input, and uses different wavelet bases and wavelet coefficients to form features.

[0126] Based on the above signal classification method, such as Figure 8 As shown, some embodiments of the present application also provide a signal classification device based on wavelet transform, including: an acquisition module, a conversion module, a model input module and an output module.

[0127] The acquisition module is used to acquire the signal to be classified detected by the detection device, and the signal to be classified is a time sequence signal;

[0128] The conversion module is used to generate input data, wherein the input data includes original data of the signal to be classified, transformed data obtained after the signal to be classified is subjected to wavelet transformation, and composite data formed by stacking the original data and the transformed data;

[0129] The model input module is used to input the input data into a neural network model, wherein the neural network model includes multiple LSTM networks, a deep residual network, and a fully connected network;

[0130] The output module is used to obtain the intrusion classification probability corresponding to the signal to be classified, and the intrusion classification probability is the output result of the neural network model.

[0131] In practical applications, the signal classification device can be arranged in a terminal device with data processing function such as a computer or a server. The acquisition module of the signal classification device is connected to the front-end device to obtain the signal to be classified through the front-end device. After obtaining the signal to be classified, the acquisition module can send the signal data to the conversion module, so as to perform a wavelet transform on the signal to be classified to generate input data with a time series relationship preserved. The conversion module then sends the generated input data to the model input module. The model input module has a built-in neural network model, which can calculate the classification probability based on the input data, and output the classification probability through the output module to complete the signal classification.

[0132] It should be noted that, for other implementations provided in the above embodiments, additional functional modules, such as a training module, etc., may be configured in the signal classification device; or a functional processing unit may be configured in any module to respectively execute the relevant program steps in the above embodiments. For example, the conversion module may also include a raw data conversion unit, a wavelet transformation unit, a stacking unit, etc.

[0133] Among them, the original data conversion unit is used to convert the signal to be classified into the original data in vector form; the wavelet transformation unit is used to perform wavelet transformation on the signal to be classified to obtain transformed data; and the stacking unit is used to stack the original data and the transformed data to form the composite data.

[0134] Similarly, for the other implementations mentioned above, corresponding functional modules or functional units may also be configured to complete various processing of signal data, which will not be listed one by one here.

[0135] It can be seen from the above technical solutions that the present application provides a signal classification method and device based on wavelet transform. After obtaining the signal to be classified, the method performs wavelet transform on the signal to be classified, and generates input data according to the original data of the signal to be classified and the wavelet transform result. The input data is then input into the neural network model to calculate the output classification probability through the neural network model. Among them, the input data includes the original data of the signal to be classified, the transformed data obtained after the signal to be classified is subjected to wavelet transform, and the composite data formed by stacking the original data and the transformed data. The time series relationship in the signal to be classified can be retained, and the characteristics of the data input into the neural network model can be increased, thereby improving the accuracy of signal classification.

[0136] Similar parts between the embodiments provided in this application can be referenced to each other. The specific implementation methods provided above are only a few examples under the general concept of this application and do not constitute a limitation on the protection scope of this application. For those skilled in the art, any other implementation methods expanded based on the scheme of this application without creative work belong to the protection scope of this application.

Claims

1. A signal classification method, characterized in that: include: Acquire a signal to be classified detected by a detection device, wherein the signal to be classified is a time sequence signal; Generate input data, the input data including original data of the signal to be classified, transformed data obtained after the signal to be classified is subjected to wavelet transform, and composite data formed by stacking the original data and the transformed data; Inputting the input data into a neural network model, wherein the neural network model includes a plurality of LSTM networks, a deep residual network, and a fully connected network; The input layer of the neural network model includes a first branch and a second branch, the first branch includes two first LSTM networks for inputting the original data and the transformed data respectively; the second branch includes a second LSTM network, and the number of hidden units of the second LSTM network is the sum of the number of hidden units of the two first LSTM networks in the first branch; Inputting the original data and the transformed data into two first LSTM networks respectively to obtain output results of two first LSTM networks; Stacking the output results of the two first LSTM networks to generate first result data; Inputting the composite data into the second LSTM network to obtain an output result of the second LSTM network; Stacking the first result data and the output result of the second LSTM network to generate second result data; An intrusion classification probability corresponding to the signal to be classified is obtained, where the intrusion classification probability is an output result of the neural network model.

2. The signal classification method according to claim 1, characterized in that: The steps to generate input data include: Converting the raw data into vector form from the signal to be classified; Performing wavelet transform on the signal to be classified to obtain transformed data, wherein the transformed data has the same vector length as the original data; The original data and the transformed data are stacked to form the composite data.

3. The signal classification method according to claim 1, characterized in that: The step of inputting the input data into the neural network model also includes: Inputting the second result data into the deep residual network; Extracting features from the second result data using the deep residual network; Calculate the classification probability based on the extracted features using the fully connected network and the softmax activation function; Output the classification probability.

4. The signal classification method according to claim 3, characterized in that: The deep residual network includes multiple residual blocks and transition blocks, and the multiple residual blocks are connected through the transition blocks; each of the residual blocks includes multiple convolutional layers and multiple activation function layers.

5. The signal classification method according to claim 1, characterized in that: The method further includes training the neural network model, specifically comprising: Acquire signal sample data, wherein the signal sample data includes a training label; Processing the signal sample data by wavelet transform to generate training data; Inputting the training data into the initialized neural network model to obtain an output result; The model parameters of the neural network model are adjusted according to the output result and the training label.

6. The signal classification method according to claim 5, characterized in that: The method further comprises: Obtaining judgment data, wherein the judgment data is a combination of one or more of the training set accuracy, training set loss, validation set accuracy, and validation set loss output by the neural network model; If the judgment data meets the preset training requirements, the model parameters are output.

7. The signal classification method according to claim 5, characterized in that: The step of adjusting the model parameters of the neural network model according to the output result and the training label includes: Calculate the difference between the training labels in the training data and the output results using a loss function; With the goal of minimizing the difference, the model parameters of the neural network model are iteratively adjusted through a back-propagation algorithm.

8. A signal classification device, characterized in that: include: An acquisition module, used for acquiring a signal to be classified detected by a detection device, wherein the signal to be classified is a time sequence signal; A conversion module, used to generate input data, wherein the input data includes original data of the signal to be classified, transformed data obtained after the signal to be classified is subjected to wavelet transformation, and composite data formed by stacking the original data and the transformed data; A model input module, used to input the input data into a neural network model, wherein the neural network model includes multiple LSTM networks, a deep residual network, and a fully connected network; The input layer of the neural network model includes a first branch and a second branch, the first branch includes two first LSTM networks for inputting the original data and the transformed data respectively; the second branch includes a second LSTM network, and the number of hidden units of the second LSTM network is the sum of the number of hidden units of the two first LSTM networks in the first branch; Inputting the original data and the transformed data into two first LSTM networks respectively to obtain output results of two first LSTM networks; Stacking the output results of the two first LSTM networks to generate first result data; Inputting the composite data into the second LSTM network to obtain an output result of the second LSTM network; Stacking the first result data and the output result of the second LSTM network to generate second result data; The output module is used to obtain the intrusion classification probability corresponding to the signal to be classified, and the intrusion classification probability is the output result of the neural network model.

Images