A signal recognition method based on multi-dimensional feature fusion and automatic learning framework

Abstract

This invention discloses a signal recognition method based on a multi-dimensional feature fusion and automatic learning framework, belonging to the field of wireless signal recognition technology. This method aims to address the shortcomings of existing signal recognition methods, such as single feature extraction, poor fusion effect, cumbersome model training, and unstable performance. The steps include: acquiring and preprocessing the baseband IQ of the target signal; extracting expert features based on prior knowledge and deep features based on 1DCNN through two parallel branches to obtain expert feature matrices and deep feature matrices; fusing the two feature matrices based on the SE attention mechanism to obtain a feature fusion matrix; training the dataset constructed from the feature fusion matrix under the AutoGluon automatic machine learning framework to obtain the optimal ensemble learning model; and using this model to complete target signal recognition inference. This invention fully integrates the advantages of traditional signal processing and deep learning, intelligently allocating feature weights, eliminating the need for extensive manual hyperparameter tuning, effectively improving signal recognition accuracy and efficiency, reducing computational complexity, saving significant parameter tuning time, and making it suitable for various signal recognition scenarios with high industrial application value.

Description

Technical Field

[0001] This invention relates to the field of wireless signal recognition technology, and in particular to a signal recognition method based on a multi-dimensional feature fusion and automatic learning framework. Background Technology

[0002] In the field of wireless signal recognition, traditional signal recognition methods often rely on single feature extraction approaches. These either employ expert feature extraction based solely on prior knowledge, resulting in insufficient feature mining and poor adaptability to complex scenarios; or they rely solely on deep learning models for feature extraction, leading to high training costs, insufficient utilization of prior experience, and feature redundancy that reduces recognition accuracy. Furthermore, existing feature fusion methods are relatively simple, failing to effectively balance the weights of different feature types. Moreover, model training requires extensive manual hyperparameter tuning, which is not only time-consuming and labor-intensive but also prone to performance instability due to human bias, making it difficult to meet the high accuracy, efficiency, and stability requirements of practical applications. Therefore, there is an urgent need for a signal recognition method that can fuse multi-dimensional features, achieve automatic learning and optimization, improve recognition performance, and reduce debugging costs, thereby addressing the shortcomings of existing technologies and improving the accuracy and practicality of signal recognition. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to overcome the shortcomings of existing signal recognition methods, such as single feature extraction, poor fusion effect, cumbersome model training and unstable performance. The present invention provides a signal recognition method based on multi-dimensional feature fusion and automatic learning framework, which can accurately and efficiently complete target signal recognition, reduce debugging costs, and improve model stability and recognition accuracy.

[0004] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0005] A signal recognition method based on a multi-dimensional feature fusion and automatic learning framework includes the following steps:

[0006] S1. Determine that the data type of the target signal to be identified is a baseband IQ signal, and collect the baseband IQ target signal data;

[0007] S2. The target signal is preprocessed, and then two parallel branches are used to extract signal features from the preprocessed target signal.

[0008] S3. One of the branches performs expert feature extraction on the target signal based on prior knowledge to obtain an expert feature matrix;

[0009] S4. Another branch extracts features from the target signal based on a deep feature learning method to obtain a deep feature matrix;

[0010] S5. Perform feature fusion on the expert feature matrix and the deep feature matrix to obtain a feature fusion matrix;

[0011] S6. The dataset constructed from the feature fusion matrix is ​​trained in the AutoGluon multi-model ensemble learning framework to obtain the optimal ensemble learning model.

[0012] S7. The obtained optimal ensemble learning model is used for the inference task of target signal recognition to complete the target signal recognition.

[0013] The beneficial effects of this invention are as follows: By setting up two parallel branches, expert features based on prior knowledge and deep features based on deep learning are extracted respectively, fully integrating the advantages of traditional signal processing and the feature mining capabilities of deep learning, comprehensively exploring the multi-dimensional characteristics of baseband IQ signals; the use of the SE attention mechanism for feature fusion can intelligently allocate feature weights, effectively avoid feature redundancy, and improve feature representation capabilities; with the help of the AutoGluon automatic machine learning framework, no extensive manual tuning of hyperparameters is required, and model training and optimization are automatically completed, saving more than 90% of the tuning time compared to manual tuning, avoiding performance instability caused by human experience bias, and ensuring that the model always maintains its optimal state in real-world scenarios; the overall invention achieves high-precision and high-efficiency identification of baseband IQ signals, reduces computational complexity, improves the practicality and reliability of the method, and provides an efficient and stable solution for wireless signal identification, possessing high industrial application value. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the steps of a signal recognition method based on a multi-dimensional feature fusion and automatic learning framework according to an embodiment of the present invention.

[0015] Figure 2 This is a schematic diagram of the structure of a 1DCNN in a signal recognition method based on a multi-dimensional feature fusion and automatic learning framework according to an embodiment of the present invention;

[0016] Figure 3 This is a schematic diagram of the SE attention mechanism module in a signal recognition method based on a multi-dimensional feature fusion and automatic learning framework according to an embodiment of the present invention. Figure 4 This is a confusion matrix diagram of the wireless signal identification results collected by 8 mobile phones in Example 1. Detailed Implementation

[0017] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0018] Please refer to Figure 1This invention provides a signal recognition method based on a multi-dimensional feature fusion and automatic learning framework, and the specific implementation steps are as follows:

[0019] S1. Determine that the data type of the target signal to be identified is a baseband IQ signal, and collect the baseband IQ target signal data; wherein, the baseband IQ signal includes in-phase components (I components) and quadrature components (Q components), which is the most basic signal form in wireless signal transmission and is widely used in various wireless communication scenarios. The collected signal data must ensure integrity and accuracy, and avoid excessive noise interference that may affect the subsequent processing effect.

[0020] S2. Preprocess the target signal, specifically by: obtaining the maximum, minimum, and average values ​​of the target signal; subtracting the average value from the target signal data; and then dividing by the data fluctuation range (i.e., the maximum value minus the minimum value) to complete the signal normalization process. Normalization eliminates the influence of signal data of different magnitudes, unifies the data scale, avoids deviations in subsequent feature extraction and model training due to differences in signal amplitude, and improves the accuracy of feature extraction and the generalization ability of the model.

[0021] After preprocessing, two parallel branches are used to extract signal features from the preprocessed target signal. The two branches run synchronously to improve the feature extraction efficiency and ensure that the two types of features can comprehensively and accurately reflect the characteristics of the target signal.

[0022] S3. One branch performs expert feature extraction on the target signal based on prior knowledge to obtain an expert feature matrix. The specific method is as follows: extract the amplitude, phase, mean, standard deviation, maximum value, minimum value, mean absolute deviation, skewness, kurtosis, information entropy, energy, and root mean square of the target signal I and Q channels, a total of 12 expert features. Each signal corresponds to 12 features, and the expert features of the two signals are combined to form an expert feature matrix.

[0023] For I / Q signals (where I is the in-phase component and Q is the quadrature component), the specific calculation methods for each expert feature are as follows:

[0024] 1. Amplitude: Amplitude reflects the strength of the signal, and is calculated as follows:

[0025]

[0026] 2. Phase: Phase describes the angular position of a signal in the complex plane, and is of great significance for analyzing signal delay, modulation methods, etc. The calculation formula is:

[0027]

[0028] 3. Mean: The mean is used to reflect the average level of a signal. Let the sample sequence of the I / Q signal be... The mean is calculated as follows:

[0029]

[0030] 4. Standard Deviation: Standard deviation measures the dispersion of the real and imaginary parts of an I / Q signal. A larger standard deviation indicates more dispersed data. The formula for calculation is:

[0031]

[0032] 5. Maximum value: The largest value in the sample sequence, i.e. ;

[0033] 6. Minimum value: The smallest numerical value in the sample sequence, i.e. ;

[0034] 7. Mean Absolute Deviation (MAD): The average of the absolute deviations of each data point from the mean. If the mean of the signal is known... The calculation method is as follows:

[0035]

[0036] 8. Skewness: Used to describe the asymmetry of data distribution. If the mean of the sequence is known... and standard deviation The calculation method is as follows:

[0037]

[0038] 9. Kurt: Measured by the sharpness of the peaks in a data distribution compared to a normal distribution. A kurtosis greater than 3 indicates a more peaked and thicker-tailed distribution than a normal distribution; a kurtosis less than 3 indicates a flatter distribution. If the mean of the sequence is known... and standard deviation The calculation method is as follows:

[0039]

[0040] 10. Information Entropy: In information theory, entropy is a measure of uncertainty, used to gauge the complexity or information content of a signal. It is calculated as follows:

[0041]

[0042] in, Let B be the probability of the i-th signal state, and B be the total number of signal states.

[0043] 11. Energy: The calculation of signal energy is usually related to the square of the signal, which can reflect the energy distribution of the signal in both the I and Q channels. Let the sample sequence be... The energy calculation is as follows:

[0044]

[0045] 12. Root Mean Square (RMS): An effective measure of signal amplitude, related to signal energy, reflecting the average energy intensity of the signal in both the I and Q channels. If the energy E of the sequence is known, then the RMS is:

[0046]

[0047] By extracting the above 12 expert features and making full use of the prior experience of traditional signal processing, we can quickly capture the key statistical characteristics and time-domain and frequency-domain features of the signal, providing reliable basic feature support for subsequent feature fusion and model training.

[0048] S4. Another branch extracts features from the target signal based on a deep feature learning method to obtain a deep feature matrix. The deep feature learning method used is 1DCNN (One-Dimensional Convolutional Neural Network). The specific structure of this network is as follows: it contains a 64-channel convolutional layer, which is connected to a pooling layer, followed by a 128-channel convolutional layer and a pooling layer. The output of the above network is flattened and then concatenated into 65536-dimensional high-dimensional features. Then, it is output as 256-dimensional deep features through a fully connected layer. These 256-dimensional deep features constitute the deep feature matrix.

[0049] Furthermore, the 1DCNN structure network is as follows: Figure 2 As shown, the architecture mainly consists of convolutional layers, pooling layers, flattening layers, and fully connected layers. It is a convolutional neural network architecture for I / Q signal processing, employing a parallel branch design: the I and Q signals are first input into their respective branches, undergoing a feature extraction process of "input layer - 64-channel convolutional layer - pooling layer - 128-channel convolutional layer - pooling layer". The convolutional layers use learnable convolutional kernels to capture local signal features (such as frequency components and waveform patterns), and the pooling layers reduce dimensionality and enhance robustness through downsampling. Subsequently, the flattening layer flattens the multi-dimensional feature map into a one-dimensional vector, and the pooling layer merges the two feature maps. Finally, the fully connected layer learns the non-linear relationships between features, outputting a 256-dimensional deep feature matrix. This 1DCNN network can automatically mine hidden deep features in signals, compensating for the shortcomings of expert features in capturing complex signal features.

[0050] S5. Perform feature fusion on the expert feature matrix and the deep feature matrix to obtain a feature fusion matrix. Specifically, this is achieved by establishing a feature weight allocation model based on the SE (Squeeze-and-Excitation) attention mechanism to realize the deep fusion of expert features and deep features. The specific steps are as follows:

[0051] S51. For the input expert feature matrix and deep feature matrix (collectively referred to as input features), global average pooling is used to compress the spatial dimension and extract the overall channel response. The specific calculation formula is as follows:

[0052]

[0053] Where H and W are the height and width of the feature map, respectively. Let be the global statistic for the c-th channel, representing the global response of the spatial features of that channel. This step compresses the two-dimensional spatial information into the global statistic of a single channel, yielding the compressed feature vector. (c is the number of channels).

[0054] S52. Perform a nonlinear transformation on the compressed feature vector to learn the dependencies between channels and generate channel attention weights; specifically, this is implemented using a bottleneck structure consisting of two fully connected layers. First, the compressed feature vector is passed through a dimensionality-reducing fully connected layer (dimension: ...). (where r is the compression ratio), then it passes through the ReLU activation function, and then through an up-dimensional fully connected layer (dimension is...). Finally, the Sigmoid activation function is used to obtain the weight of each channel (a value between 0 and 1). The specific calculation formula is as follows:

[0055]

[0056] in, It is the compressed feature vector. It is a dimension reduction matrix (dimension is ). ), It is the ReLU activation function. It is an upgraded matrix (dimension is 1). ), It is the sigmoid function, which ultimately yields the feature vector for each channel. That is, channel attention weight.

[0057] S53. Multiply the learned channel weights s with the original input feature matrix channel by channel, that is, multiply the feature of each channel by the corresponding weight value, thereby realizing the recalibration of the features. The specific calculation formula is as follows:

[0058]

[0059] in, It is the weight of the c-th channel, which is a scalar; It is the c-th channel of the original feature map; This is the recalibrated feature map. This step enhances the features of important channels, suppresses the features of secondary channels, achieves an optimized combination of expert features and deep features, avoids feature redundancy, and improves the expressive power of the feature fusion matrix.

[0060] Furthermore, the SE attention mechanism module is as follows: Figure 3 As shown, it mainly consists of spatial feature compression, channel feature learning, and feature recalibration. It starts with an input feature cube (representing the multi-channel feature map before the fusion of the expert feature matrix and the deep feature matrix). The number of channels and spatial size are determined by the preceding feature extraction network. Through the above S51-S53 steps, the final output is a new feature cube that incorporates channel attention, i.e., the feature fusion matrix, which enables the subsequent model to pay more attention to key channel information and improve recognition accuracy.

[0061] S6. The dataset constructed from the feature fusion matrix is ​​trained in the AutoGluon multi-model ensemble learning framework to obtain the optimal ensemble learning model. The specific hyperparameters are set as follows: the optimizer is Adam, the number of training epochs is 3000, the batch size is 32, and the learning rate is 0.0001.

[0062] The specific implementation process is as follows: A dataset is constructed based on the feature fusion matrix, and the dataset is divided into a training set, a validation set, and a test set in a 6:2:2 ratio. Under the AutoGluon automatic machine learning framework, a multilayer perceptron classifier is used for end-to-end parameter optimization. The framework automatically completes the hyperparameter optimization, model selection, and ensemble learning of the multilayer perceptron classifier without manual intervention. The model is trained using the training and validation sets, and the model parameters are adjusted in real time using the validation set to avoid overfitting or underfitting. After training, the model performance is tested using the test set to ensure that the model achieves optimal recognition results, ultimately yielding the optimal ensemble learning model.

[0063] AutoGluon integrates a variety of advanced machine learning algorithms and automation tools, which can automatically complete the training and optimization of models according to the set training time, model performance and other constraints, avoiding the instability and failure to achieve optimal performance caused by manual parameter tuning. At the same time, it can flexibly set training parameters to meet the needs of practical applications from various dimensions.

[0064] S7. The obtained optimal ensemble learning model is used for the inference task of target signal recognition to complete the target signal recognition; specifically: new baseband IQ target signals are collected, preprocessed according to the method of step S2, expert feature matrix and deep feature matrix are extracted according to the methods of steps S3 and S4 respectively, feature fusion is performed according to the method of step S5 to obtain feature fusion matrix, the feature fusion matrix is ​​input into the optimal ensemble learning model, the model outputs the recognition result, and the target signal is quickly and accurately recognized.

[0065] The signal recognition method based on multi-dimensional feature fusion and automatic learning framework described above is applicable to various scenarios requiring the recognition of baseband IQ signals. The following specific embodiments further illustrate this method:

[0066] Example 1

[0067] Please refer to Figure 1-3 The specific implementation steps of the signal recognition method based on multi-dimensional feature fusion and automatic learning framework in this embodiment are as follows:

[0068] S1. Eight mobile phones were selected for Bluetooth and WiFi signal collection. The selected mobile phone models were Apple iPhone 5 / 5s / 6 / 6s, LG G4, Samsung S5 / Note5, and Sony Xperia M5.

[0069] S2. Collect baseband IQ signals as the target signals to be identified. The collection scenario is an indoor wireless communication scenario. A total of 8 sets of Bluetooth signals and 2 sets of WiFi signals are collected. Each set of signals has 5250 signal samples, for a total of 52500 signal samples. Each signal sample consists of 1024 I points and 1024 Q points.

[0070] S3. Preprocess the acquired target signals: Calculate the maximum, minimum, and average values ​​of each group of signals using the formula... Complete the normalization process, among which The original signal data, The average value of the signal. The maximum value of the signal. The minimum value of the signal is used; the normalized data range is unified to [0,1] to eliminate the influence of differences in the amplitude of different signals.

[0071] S4. Extracting expert features based on prior knowledge: For the normalized I / Q signals, extract 12 expert features respectively, including amplitude, phase, mean, standard deviation, maximum value, minimum value, mean absolute deviation, skewness, kurtosis, information entropy, energy, and root mean square. The calculation method of each feature is strictly performed according to the formula described in the specific implementation of this invention, resulting in a 24-dimensional (12-dimensional I channel + 12-dimensional Q channel) expert feature matrix for each group.

[0072] S5. Extracting deep features based on 1DCNN: Construct a 1DCNN network with the following structure: Input layer → 64-channel convolutional layer (kernel size 3, stride 1) → pooling layer (max pooling, kernel size 2, stride 2) → 128-channel convolutional layer (kernel size 3, stride 1) → pooling layer (max pooling, kernel size 2, stride 2) → Flatten layer → fully connected layer (output dimension 256); Input the normalized I / Q signals into the 1DCNN network respectively. After network processing, 256-dimensional deep features are output, resulting in a 256-dimensional deep feature matrix for each group.

[0073] S5. Feature fusion based on SE attention mechanism: The expert feature matrix and deep feature matrix are used as input features. The compression ratio of the SE attention mechanism is set to r=16. The spatial dimension is compressed by global average pooling to obtain the compressed feature vector. Channel attention weights are generated by passing through a dimension reduction fully connected layer, ReLU activation function, dimension increase fully connected layer and Sigmoid activation function. The attention weights are multiplied with the input features channel by channel to complete the feature recalibration and obtain a 280-dimensional (24-dimensional + 256-dimensional) feature fusion matrix for each group.

[0074] S6. Train the optimal ensemble learning model based on the AutoGluon framework: Divide the dataset constructed from the feature fusion matrix into training, validation, and test sets in a 6:2:2 ratio; Set hyperparameters: optimizer is Adam, epoch=3000, batch size=32, lr=0.0001; Start model training under the AutoGluon framework. The framework automatically completes hyperparameter optimization, model selection, and ensemble learning. During training, monitor model performance through the validation set. When the model's accuracy on the validation set does not improve for 50 consecutive epochs, stop training to obtain the optimal ensemble learning model.

[0075] S7. Model Inference and Performance Testing: The feature fusion matrix of the test set is input into the optimal ensemble learning model, and the model outputs the recognition results. Testing shows that the model's recognition accuracy reaches 99.8%. The confusion matrix for the recognition results of 8 signal devices and 10 signals is shown in the attached diagram. Figure 4As shown, compared with traditional signal recognition methods that extract only one feature, the accuracy is improved by more than 15%, and the model training time is shortened by 90% compared with manual parameter tuning, demonstrating high recognition accuracy and efficiency.

[0076] As can be seen from this embodiment, the signal recognition method of the present invention can effectively integrate expert features and deep features, and achieve automatic learning and optimization with the help of the AutoGluon framework. This not only improves the accuracy and efficiency of signal recognition, but also reduces the model debugging cost, solves the shortcomings of the prior art, and has high practicality and innovation, meeting the requirements for patent authorization.

Claims

1. A signal recognition method based on a multi-dimensional feature fusion and automatic learning framework, characterized in that, Including the following steps: The target signal data type to be identified is a baseband IQ signal; The target signal is preprocessed, and then signal features are extracted using two parallel branches. One branch performs expert feature extraction on the target signal based on prior knowledge to obtain an expert feature matrix; Another branch performs feature extraction on the target signal based on a deep feature learning method to obtain a deep feature matrix; The expert feature matrix and the deep feature matrix are fused to obtain a feature fusion matrix; The dataset constructed from the feature matrix is ​​trained on the AutoGluon multi-model ensemble learning framework to obtain the optimal ensemble learning model. The obtained optimal ensemble learning model is used for the inference task of target signal recognition.

2. The signal recognition method based on a multi-dimensional feature fusion and automatic learning framework according to claim 1, characterized in that: The specific implementation of the preprocessing of the target signal is as follows: obtain the maximum and minimum values ​​and the average value of the target signal, subtract the average value from the data and divide by the data fluctuation range (maximum value minus minimum value), and then perform normalization.

3. The signal recognition method based on a multi-dimensional feature fusion and automatic learning framework according to claim 1, characterized in that: The method for extracting expert features from the target signal based on prior knowledge is as follows: extracting 12 expert features from the I / Q channels of the target signal, including amplitude, phase, mean, standard deviation, maximum value, minimum value, mean absolute deviation, skewness, kurtosis, information entropy, energy, and root mean square.

4. The signal recognition method based on a multi-dimensional feature fusion and automatic learning framework according to claim 1, characterized in that: The target signal is subjected to feature extraction using a deep feature learning method to obtain a deep feature matrix. The deep feature learning method used is a 1DCNN network, which contains a 64-channel convolutional layer connected to a pooling layer, followed by a 128-channel convolutional layer and a pooling layer. The output is flattened and then concatenated into a 65536-dimensional high-dimensional feature, which is then output as a 256-dimensional deep feature through a fully connected layer.

5. The signal recognition method based on a multi-dimensional feature fusion and automatic learning framework according to claim 1, characterized in that: The expert feature matrix and the deep feature matrix are fused to obtain a feature fusion matrix. The specific implementation method is to establish a feature weight allocation model based on the SE (Squeeze-and-Excitation) attention mechanism to achieve deep fusion of expert features and deep features. For input features Compress spatial dimensions using global average pooling: Where H and W are the height and width of the feature map, respectively. is the global statistic for the c-th channel, representing the global response of the spatial characteristics of that channel. in, It is the compressed feature vector. It is a dimension reduction matrix (dimension is ). ), It is the ReLU activation function. It is an upgraded matrix (dimension is 1). ), It is the sigmoid function, which ultimately yields the feature vector for each channel. . The learned channel weights s are multiplied with the original feature matrix channel by channel, that is, the feature of each channel is multiplied by the corresponding weight value, so as to achieve feature recalibration. in, It is the weight of the c-th channel, which is a scalar; It is the c-th channel of the original feature map; It is the recalibrated feature map.

6. The signal recognition method based on a multi-dimensional feature fusion and automatic learning framework according to claim 1, characterized in that: The dataset constructed from the feature matrix is ​​trained on the AutoGluon multi-model ensemble learning framework to obtain the optimal ensemble learning model. The specific hyperparameters are set as follows: the optimizer is Adam, the epoch is 3000, the batch size is 32, and the learning rate is 0.0001.

Images