An abnormal broadcast identification model and its construction method and use method
Through the neural network identification model, the problem of inefficient observation is solved, automatic identification of abnormal broadcast signals and accurate analysis of transmission time distribution is realized, and recognition efficiency and accuracy are improved.
Patent Information
- Application Number
- CN202110670190.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-06-17
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2041-06-17
AI Technical Summary
The prior art relies on manual observation when identifying abnormal broadcast signals, which is inefficient and requires the construction of a complete spectrum license database and understanding the prior probability distribution of many frequency-using parameters. It is difficult to obtain and resolve information, resulting in frequent error detection and miss detection and cost. A lot of manpower and material resources.
Using an abnormal broadcast recognition model based on neural network, through the construction of training data sets, data preprocessing and transfer learning, we automatically extract features from the original spectrum data, and realize all-weather monitoring of suspicious frequency points and automatic identification of abnormal broadcast transmission time distribution. Replace manual observation to improve analysis efficiency.
It realizes automatic identification of abnormal broadcast signals, reduces error detection and miss detection, improves the efficiency of transmission time distribution analysis of abnormal broadcast stations, and can train a highly accurate neural network classifier model under a small amount of spectrum data, reducing neural network Data processing pressure improves recognition accuracy.
Smart Images

Figure CN113269134B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to an abnormal broadcast identification technology, and in particular to an abnormal broadcast identification model and a construction method and a use method thereof. Background Art
[0002] With the continuous significant achievements in electromagnetic wave theory and application, electromagnetic waves have become the most important form of information and energy transmission for humans. The openness of radio waves makes them vulnerable to interference and illegal use, causing normal communication systems to be disturbed or even interrupted. In order to ensure the safety of residents and maintain normal production and living order, it is urgent to discover, locate and crack down on illegal frequency use.
[0003] At present, the traditional method is based on manual real-time observation of the spectrogram of the broadcast signal at the suspicious frequency point and the auxiliary judgment of the signal category by broadcast audio information, and then further statistics the distribution of the transmission time of abnormal radio stations based on the manual judgment results. Each expert can only focus on one suspicious frequency point in a single observation. In addition, compared with normal use, abnormal use is sporadic and only exists in a few times. A single observation is likely to consume a lot of time and energy but fail to capture abnormal broadcast signals, and thus it is impossible to start the problem of abnormal broadcast transmission time distribution. The traditional method requires the construction of a complete spectrum license database, the accumulation of sufficient spectrum monitoring data, and the understanding of the prior probability distribution of many frequency parameters. It is difficult to obtain, accumulate, analyze, and store this information. The input information characteristics that the traditional method needs to process are complex and diverse. Manual observation relies on the professional knowledge and experience of the observer. The observation results are highly subjective, consuming a lot of manpower and material resources and having low work efficiency. Summary of the invention
[0004] At least one purpose of the present application is to provide an abnormal broadcast identification model and a construction method and a use method thereof, which can automatically extract more expressive features from the original spectrum data to meet the end-to-end requirements in practical applications, directly determine the data category labels from the original data, and realize the analysis and identification of all-weather monitoring data of suspicious frequencies. The analysis and identification results can be reversed to obtain the abnormal broadcast transmission time distribution of the corresponding frequency points.
[0005] The abnormal broadcast identification model and its construction method and use method provided in this application can replace manual observation, reduce false detections and missed detections, and improve the efficiency of abnormal broadcast station transmission time distribution analysis and mining.
[0006] The present application provides a method for constructing an abnormal broadcast recognition model, including:
[0007] S1: build training dataset;
[0008] S2: training data preprocessing;
[0009] S3: Using the preprocessed training data set, transfer training is performed based on the neural network model that has been pre-trained on the image classification data set to obtain an abnormal broadcast recognition model.
[0010] Wherein, the training data in the training data set is a spectrum graph;
[0011] The training data preprocessing step includes performing the following processing on the spectrogram:
[0012] Traverse all the spectrograms and take the size of the smallest spectrogram as the standard size (L s ,W s );
[0013] For all except standard sizes (L s ,W s ) are cropped so that all the spectrum graphs have the standard size (L s ,W s ).
[0014] According to at least one embodiment of the present application, the step of clipping the spectrum graph includes:
[0015] Assume that the original size of the spectrum graph i is (L i ,W i ), cut off 0.5 (L i -L s ), cut off 0.5(W) on both the upper and lower sides of the spectrum graph i i -W s ).
[0016] According to at least one embodiment of the present application, the type of training data of the training data set constructed in step S1 is one of the following three types: abnormal broadcast signal; normal broadcast signal; no broadcast signal.
[0017] According to at least one embodiment of the present application, the numbers of the training data of the abnormal broadcast signal, the normal broadcast signal, and the non-broadcast signal type are equal.
[0018] According to at least one embodiment of the present application, the neural network model described in step S3 uses a ResNet18 model as a backbone network, and the ResNet18 model includes:
[0019] Layers 1 to 17 are all convolutional layers, used to extract input data features;
[0020] The 18th layer is a fully connected layer, which is used to combine the features extracted by the neural network model and output the predicted classification results.
[0021] According to at least one embodiment of the present application, step S3 includes:
[0022] Load the ResNet18 model that has been pre-trained on the image classification dataset;
[0023] Adding 18th to 21st layers between the 17th layer and the fully connected layer of the ResNet18 model to construct a deep migration network model, wherein the deep migration network model includes: 1st to 3rd layers for learning general features; 4th to 17th layers for learning task-specific features; 18th to 21st layers are 4 adaptive layers to make the target domain data distribution closer to the source domain;
[0024] The deep migration network model is used to perform migration training.
[0025] According to at least one embodiment of the present application, the optimization objective function is
[0026] Where l represents the total loss, the first term is the CrossEntropyLoss loss, and the second term represents the adaptive loss of the network, D s is the source domain, D t is the target domain, λ is the weight parameter, l 1 , l 2 Which equals 18 and 21 respectively.
[0027] According to at least one embodiment of the present application, the image classification dataset is an ImageNet dataset.
[0028] The present application also provides an abnormal broadcast identification model, which is constructed by the above method.
[0029] The present application also provides a method for using an abnormal broadcast identification model, which utilizes the abnormal broadcast identification model to classify broadcast signals to identify abnormal broadcast signals.
[0030] The abnormal broadcast recognition model provided by the present application can replace manual observation to realize automatic recognition of abnormal broadcast signals. In the technical solution provided by the present application, the abnormal broadcast recognition problem is converted into an image classification problem, thereby making the complex multi-parameter problem intuitive and simple.
[0031] In the technical solution provided by the present application, only a small amount of spectrum data is needed to train a neural network classifier model with sufficiently high accuracy. In accordance with the abnormal broadcast recognition model provided by the present application, the present application also provides a data preprocessing method for making the spectrogram size consistent, which can effectively reduce the data processing pressure of the neural network and improve the recognition accuracy of the abnormal broadcast recognition model trained by it. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] The following will further explain the above characteristics, technical features, advantages and implementation methods of the present application in a clear and understandable manner through the description of the preferred embodiments and in combination with the accompanying drawings. The following drawings are only intended to illustrate and explain the present application, and do not limit the scope of the present application. Among them:
[0033] Figure 1 A flowchart of a method for constructing an abnormal broadcast recognition model according to an embodiment of the present application is shown;
[0034] Figure 2 shows a distribution diagram of abnormal broadcast signal transmission time obtained according to an embodiment of the present application;
[0035] Figure 3 A distribution diagram of abnormal broadcast signal transmission time obtained according to another embodiment of the present application is shown. DETAILED DESCRIPTION
[0036] In order to have a clearer understanding of the technical features, purposes and effects of the present application, the specific implementation methods of the present application are now described with reference to the accompanying drawings.
[0037] The abnormal broadcast signal usually has a slight frequency deviation (the frequency deviation usually does not exceed 0.1KHz) compared to the frequency point (usually in MHz) of the authorized broadcast signal. In this application, the broadcast signal with a slight frequency deviation (the frequency deviation usually does not exceed 0.1KHz) at the known authorized broadcast signal frequency point (usually in MHz) is defined as an abnormal broadcast signal. By using this slight frequency deviation between the abnormal broadcast signal and the authorized broadcast signal, the abnormal broadcast signal can be identified from the authorized broadcast signal.
[0038] In the prior art, abnormal broadcast signals are usually identified by manually observing frequency deviations. However, since such subtle frequency deviations are usually difficult to detect, manual observation results are highly subjective and rely on the professional knowledge and experience of the observers, which can easily lead to false detections and missed detections.
[0039] In order to replace manual observation, the present application provides an abnormal broadcast recognition model based on a neural network and a construction method and a use method thereof to realize automatic recognition of abnormal broadcast signals.
[0040] In the technical solution of the present application, an abnormal broadcast recognition model is constructed based on the spectrum of the received broadcast using an artificial neural network, and the abnormal broadcast recognition model is trained using a training data set. Using the trained abnormal broadcast recognition model, the spectrum of the received broadcast signal can be classified to identify whether the broadcast signal is abnormal / normal / no broadcast signal.
[0041] It can be seen that in the technical solution provided by this application, the abnormal broadcast identification problem is converted into an image classification problem, thereby making the complex multi-parameter problem intuitive and simple.
[0042] However, ordinary neural network classifier models need to be trained with a large number of spectrograms to achieve sufficient classification accuracy. Since abnormal broadcasts are sporadic and only exist for a few times, the existing accumulation of spectrograms is not enough to train a neural network classifier model that can capture fine features such as frequency deviation, and the superior performance of neural networks is not brought into play.
[0043] In order to solve the problem of insufficient accumulation of spectrum data and inability to train a neural network classifier model with sufficiently high accuracy, the present application provides an abnormal broadcast recognition model based on transfer learning and a construction method and a use method thereof.
[0044] like Figure 1 As shown, the method for constructing the abnormal broadcast recognition model provided by this application includes the following steps:
[0045] S1: Building a training dataset
[0046] Since deep learning is a data-driven representation learning method, the performance of the model depends on the quality of the training data, so a high-quality training dataset is crucial.
[0047] According to one embodiment of the present application, a method for constructing a high-quality training data set is provided, the method comprising:
[0048] 1. Data Acquisition
[0049] The sampling duration is nearly 72 hours (08:50:32 on October 30, 2020-08:50:12 on November 1, 2020), and the sampling interval is 10 seconds to obtain the original data sample. The specific form of the data sample is a signal spectrum diagram, with the horizontal axis being the frequency and the vertical axis being the amplitude.
[0050] 2. Data cleaning
[0051] In order to ensure the quality of the training data set, data samples with interfering signals are first removed.
[0052] 3. Data classification
[0053] According to the number of categories of the input data samples, the problem can be understood as a three-classification problem, which classifies the signal data samples into three categories: abnormal / normal / no broadcast signal. As mentioned above, the amount of data accumulated in the existing spectrum graph is not enough to train a neural network classifier model that can capture fine features such as frequency deviation, and the superior performance of the neural network is not brought into play.
[0054] Transfer learning has the ability to learn with small samples. It only needs a small number of data samples of abnormal / normal / no broadcast signals at the frequency point to train a neural network classifier model that can capture fine features such as frequency deviation to be used as an abnormal broadcast recognition model. This trained abnormal broadcast recognition model can be applied to the problem of broadcast signal recognition and classification to classify unknown types of broadcast signals into three types: abnormal broadcast signals, normal broadcast signals, and no broadcast signals.
[0055] In this embodiment, according to the style of the typical abnormal / normal / no broadcast signal spectrum diagram of the known frequency point A, the original data samples (i.e., the spectrum diagram of the signal) are artificially divided into three categories: abnormal broadcast signal, normal broadcast signal, and no broadcast signal, wherein the number of data samples contained in each type of signal is 3222, 2783, and 11218 respectively. It can be seen that the distribution of data sample categories is very uneven. Among these three types of data samples, the signals belonging to the type of no broadcast signal occupy the vast majority (more than twice the sum of abnormal broadcast signals and normal broadcast signals).
[0056] The inventors have found that when model training is performed based on transfer learning, if the sample type distribution is uneven, it will have an adverse effect on the training efficiency and training results. In order to avoid the impact of uneven distribution of training data types on model training and improve the model training efficiency, preferably, data samples of similar numbers are randomly selected from the three types of abnormal / normal / no broadcast signal. In this embodiment, 1000 samples are randomly selected from each of the three types of abnormal / normal / no broadcast signal data to form a training data set.
[0057] S2: Training data preprocessing
[0058] As mentioned above, the frequency deviation between abnormal and normal broadcast signals is subtle. For the problem of extracting and identifying such fine features, it is necessary to improve the sensitivity of the classification model to subtle differences from two aspects: the neural network structure and the activation function selection. Therefore, the backbone neural network selects the convolutional neural network (CNN), and the activation function takes the logarithm based on SoftMax, and selects LogSoftMax.
[0059] Since the convolutional neural network requires the input image to be of consistent size, the image size is firstly uniformed. In view of the construction of the training data set required by the abnormal broadcast recognition model based on transfer learning provided by the present application, the present application proposes a data preprocessing method to make the spectrogram size consistent. The spectrogram processed by this method can meet the requirements of the convolutional neural network for the consistency of the size of the input image. In addition, the training data set composed of the spectrogram processed by this method can improve the recognition accuracy of the abnormal broadcast recognition model trained by it.
[0060] The spectrogram is generated by collecting and generating the focused frequency points, and the focused frequency points are all located in the middle of the spectrogram, so the important characteristic elements in the spectrogram are all located in the central area of the picture. The data preprocessing method proposed in this application to make the spectrogram size consistent includes:
[0061] 1. Traverse all the spectrum graphs, count the size of each spectrum graph, and take the smallest graph size as the standard size, recorded as the standard size (L s ,W s );
[0062] 2. All other spectrum graphs except those with standard size are cropped to the standard size. The specific cropping method is: s ,W s ) is the origin (0,0) at the lower left corner of the spectrum graph, and the center coordinates of the spectrum graph are marked as C(0.5L s ,0.5W s Then, all the spectrograms except those with standard size are centered at point C (0.5L s ,0.5W s ) as the center for uniform cropping, assuming that the original size of the spectrum graph i is (L i ,W i ), the left and right sides of the spectrum i are cut off by 0.5 (L i -L s ), cut off 0.5(W) on both the upper and lower sides of the spectrum graph i i -W s ). In this way, the size uniformity of the spectrum graph is completed.
[0063] Using the smallest graphic size as the standard size can also achieve the following technical effects: 1) reducing the pressure of neural network data processing, shortening the result calculation time, and meeting real-time requirements; 2) this cropping method will not lose useful information.
[0064] Since the fine features that need to be extracted are all located in the central area of the spectrum graph, uniformly cropping the four sides of the spectrum graph will not lose useful information. In addition, the training data set composed of the spectrum graphs cropped by this method can improve the recognition accuracy of the abnormal broadcast recognition model trained by it.
[0065] S3: Fine-tuning Adaptive Deep Network Transfer Training
[0066] In order to improve the training efficiency of the abnormal broadcast recognition model and enhance the model promotion capability, transfer learning is introduced in the technical solution provided in this application.
[0067] Ordinary neural network classifier models need to be trained from scratch, and all weight parameters are initialized. As mentioned above, due to the limited number of training samples, the existing accumulation of spectrum graphs is not enough to train a neural network classifier model that can capture fine features such as frequency deviation.
[0068] Compared with the training method of starting from scratch and initializing all weight parameters, the transfer learning method using the pre-trained model requires smaller training data samples, has faster model convergence, shorter training time, and higher accuracy.
[0069] ResNet alleviates the gradient vanishing or exploding problem faced in deep neural network training by introducing residual learning, which can further deepen the network and extract more complex feature patterns, which are more sensitive to fine features. Therefore, in the transfer learning solution provided in this application, the ResNet18 model is used as the backbone network for transfer training. ResNet18 consists of 17 convolutional layers and 1 fully connected layer. The fully connected layer is used to combine the features extracted by the ResNet18 network.
[0070] The data used for migration training comes from the training data set constructed by the above step S1 and has been preprocessed by the above step S2.
[0071] Before transfer training, data augmentation operations are performed on the preprocessed training data by cropping and horizontally rotating the images to improve the richness of the training data. During transfer training, the ResNet18 model that has been pre-trained on the ImageNet dataset is loaded. The classifier parameters trained on the ImageNet dataset are used as the initialization parameters of the corresponding network layer, and the remaining parameters are randomly initialized.
[0072] Since the source domain optical image and the target domain spectrum are quite different, in order to make the data distribution of the source domain and the target domain closer, the technical solution provided by this application introduces an adaptive layer in the ResNet18 model to adapt the data distribution, and adds an adaptive metric to the 18th layer of the ResNet18 model. Since the feature extraction network is deep, in order to adapt to its depth, 3 additional adaptive layers are added before the fully connected layer of the ResNet18 model, that is, a total of 4 adaptive layers are added to make the target domain data distribution closer to the source domain. Among them, each adaptive layer adopts the MMD criterion.
[0073] A key point of neural network is the definition of network loss. In the technical solution provided by this application, a new neural network optimization objective function is defined:
[0074] Where l represents the total loss, the first term is the CrossEntropyLoss loss, which is the conventional classification loss, and the second term represents the adaptive loss of the network, D s is the source domain, D t is the target domain, λ is the weight parameter that weighs the first and second terms, l 1 , l 2 They are equal to 18 and 21 respectively, indicating that the network adaptation is from layer 18 to layer 21, and no adaptation is performed before. The optimization goal is to make l as small as possible.
[0075] Through the selection of hyperparameters such as learning rate and batch size, the classification model is further optimized (epoch=64, batch_size=4). In the transfer training, the network weight parameters of the 1st to 3rd layers of the classification model network remain unchanged (the 1st to 3rd layers of the network learn common features), the network weights of the 4th to 17th layers will be updated and fine-tuned as the error back propagates (as the number of network layers increases, the latter networks are more inclined to learn task-specific features), the 18th to 21st layers are 4 adaptive layers (the purpose is to make the target domain data distribution closer to the source domain), and finally a fully connected layer is connected to output the predicted classification results.
[0076] After testing, the three-category accuracy of the fine-tuned adaptive transfer learning classification model provided in this application can reach 95.8%.
[0077] By using the abnormal broadcast recognition model provided by the present application, it is possible to automatically classify unknown types of broadcast signals into one of the following three types: abnormal broadcast signals; normal broadcast signals; no broadcast signals, thereby automatically identifying abnormal broadcast signals. Based on the recognition results, a distribution diagram of the emission time of abnormal broadcast signals can be drawn to visualize the results. Figure 2 and Figure 3The figure shows a time distribution diagram of abnormal broadcast signal transmission. The original data samples are named by the sampling time, with a time resolution of 10 seconds. The recognition and classification results are written into abnormal.txt, normal.txt, and no signal.txt files, that is, the file name of the sample belonging to this category is written into each txt file, and then the plot is drawn in units of 'days' according to the contents of the txt files (see Figure 2 ) and plotted in 'hours' (see Figure 3 ). The abnormal broadcast signal and the normal broadcast signal are represented by line segments with different grayscales, respectively.
[0078] In summary, the present application provides an abnormal broadcast identification model and a construction method and a use method thereof, which can replace manual observation to realize the automatic identification of abnormal broadcast signals. In the technical solution provided by the present application, the abnormal broadcast identification problem is converted into an image classification problem, thereby making the complex multi-parameter problem intuitive and simple. In the technical solution provided by the present application, only a small amount of spectral data is needed to train a neural network classifier model with sufficiently high accuracy. In accordance with the abnormal broadcast identification model provided by the present application, the present application also provides a data preprocessing method for unifying the size of the spectrum graph, which can effectively reduce the pressure of neural network data processing and can improve the recognition accuracy of the abnormal broadcast identification model trained by it.
[0079] It should be understood that although this specification is described according to various embodiments, not every embodiment contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.
[0080] The above description is only an illustrative embodiment of the present application and is not intended to limit the scope of the present application. Any equivalent changes, modifications and combinations made by any technician in the field without departing from the concept and principle of the present application shall fall within the scope of protection of the present application.
Claims
1. A method for constructing an abnormal broadcast recognition model, comprising: S1: Construct a training data set, where the types of training data in the training data set include: abnormal broadcast signal; normal broadcast signal; no broadcast signal; S2: training data preprocessing; S3: Using the preprocessed training data set, transfer training is performed based on the neural network model that has been pre-trained on the image classification data set to obtain an abnormal broadcast recognition model. Wherein, the training data in the training data set is a spectrum graph; The training data preprocessing step includes performing the following processing on the spectrogram: Traverse all the spectrograms and take the size of the smallest spectrogram as the standard size (L s , W s ); For all except standard sizes (L s , W s ) are cropped so that all the spectrum graphs have the standard size (L s , W s ), The neural network model described in step S3 uses the ResNet18 model as the backbone network, and the ResNet18 model includes: Layers 1 to 17 are all convolutional layers, used to extract input data features; The 18th layer is a fully connected layer, which is used to combine the features extracted by the neural network model and output the predicted classification results. Wherein, the step S3 comprises: Load the ResNet18 model that has been pre-trained on the image classification dataset; Adding 18th to 21st layers between the 17th layer and the fully connected layer of the ResNet18 model to construct a deep migration network model, wherein the deep migration network model includes: 1st to 3rd layers for learning general features; 4th to 17th layers for learning task-specific features; 18th to 21st layers are 4 adaptive layers to make the target domain data distribution closer to the source domain; The deep migration network model is used to perform migration training.
2. The method according to claim 1, wherein: The steps to crop the spectrum graph include: Assume that the original size of the spectrum graph i is (L i , W i ), cut off 0.5 (L i -L s ), cut off 0.5(W) on both the upper and lower sides of the spectrum graph i i -W s ).
3. The method according to claim 1, wherein: The numbers of the training data of the abnormal broadcast signal, the normal broadcast signal, and the non-broadcast signal type are equal.
4. The method according to claim 1, wherein: The image classification dataset is the ImageNet dataset.
5. A method for using an abnormal broadcast identification model, using the abnormal broadcast identification model constructed according to any one of claims 1 to 4 to classify broadcast signals to identify abnormal broadcast signals.
