A method for periodic signal detection that integrates rank statistics and residual convolutional neural networks

Abstract

This invention discloses a method for detecting periodic signals that integrates rank statistics and residual convolutional neural networks. The method includes: performing non-parametric processing on the original observation sequence to convert it into a rank statistics sequence to eliminate the influence of dimensions and enhance robustness; subsequently, calculating its covariance matrix based on this rank sequence to capture the structural dependency features within the data; finally, using this covariance matrix as a feature input to a pre-trained neural network detection model, which outputs the final classification or anomaly determination result. By using this invention, the key problem of performance degradation in traditional methods under unknown non-Gaussian impulse noise environments is effectively solved. This invention can be widely applied in the field of signal processing.

Description

Technical Field

[0001] This invention relates to the field of signal processing, and more particularly to a method for detecting periodic signals that integrates rank statistics and residual convolutional neural networks. Background Technology

[0002] Periodic signal detection is a core task in signal processing, aiming to accurately identify repetitive patterns in a signal amidst noise interference. Existing classical periodogram detection methods (with or without known frequencies) involve performing a Discrete Fourier Transform (DFT) on the observed signal, calculating the power spectrum, comparing peak values ​​at known frequencies, or scanning all frequencies to find the maximum value. A drawback is the need for known or estimated noise variance. For unknown frequencies, multiple verification corrections (such as Bonferroni correction) must be considered, but traditional corrections are often too conservative, reducing the detection probability. Therefore, existing methods have limitations in practical applications (such as non-Gaussian noise or unknown noise variance).

[0003] There is an urgent need for a new detection method that is insensitive to noise distribution, requires no variance estimation, and is applicable to the detection of both known and unknown frequencies. Summary of the Invention

[0004] In view of this, in order to address the technical problem that most existing periodic signal detection methods require prior knowledge of noise distribution or parameter estimation, leading to low detection accuracy in practical applications, this invention proposes a periodic signal detection method that integrates rank statistics and residual convolutional neural networks. This method includes the following steps: First, the original observation sequence is nonparametrically processed to convert it into a rank statistics sequence to eliminate the influence of dimensions and enhance robustness. Then, its covariance matrix is ​​calculated based on this rank sequence to capture the structural dependencies within the data. Finally, this covariance matrix is ​​used as a feature input into a pre-trained neural network detection model, which outputs the final classification or anomaly determination result.

[0005] In some embodiments, training of the neural network detection model is also included: Data generation and labeling: Based on the receiver system model, simulate and generate received signal data containing at least two typical categories (such as different modulation methods, presence or absence of interference, etc.), and assign corresponding labels to each type of data.

[0006] Feature extraction and construction: For the generated received signal data, observation sequences for training are extracted. Each observation sequence undergoes non-parametric processing, its rank statistic is calculated, and a sample rank covariance matrix characterizing the internal structure of the sequence is further constructed.

[0007] Dataset partitioning: All labeled sample rank covariance matrices are randomly divided into training and test sets according to a preset ratio (such as 7:3 or 8:2) to ensure that the two types of data are distributed in a balanced manner.

[0008] Model Training and Validation: The detection model is trained using the rank-covariance matrix of the samples in the training set as input features and their corresponding labels as supervision signals. During training, the test set is used for performance evaluation and hyperparameter tuning. Iterative optimization continues until the model loss converges and performance stabilizes, ultimately yielding the trained detection model.

[0009] Based on the above scheme, this invention provides a periodic signal detection method that integrates rank statistics and residual convolutional neural networks. By using rank transformation to suppress the contamination of the covariance matrix by extreme values ​​of impulse noise, and without requiring prior noise variance, it achieves model-independent robust feature extraction. This method is applicable to fields such as communication, radar, mechanical fault diagnosis, and biomedical signal processing. Especially in complex environments with strong impulse interference or unknown noise statistical characteristics, it significantly improves detection accuracy compared to traditional periodogram methods, generalized likelihood ratio tests, and sample covariance convolutional neural networks, demonstrating broad prospects for industrial application. Attached Figure Description

[0010] Figure 1 This is a flowchart of the steps of a periodic signal detection method that integrates rank statistics and residual convolutional neural networks according to the present invention. Figure 2 This is a schematic diagram of the data flow during the detection process of the method of the present invention; Figure 3 This is a schematic diagram of the residual convolutional neural network structure designed in this invention; Figure 4 This is a statistical graph of the confusion matrix of five types of samples under Bernoulli Gaussian noise using the method of this invention; Figure 5 This is a statistical graph of the confusion matrix of five classes of samples under Bernoulli Gaussian noise, using a traditional method based on convolutional neural networks using the sample covariance matrix. Detailed Implementation

[0011] In addition to the problems mentioned in the background section, existing detection methods also include detection methods based on model parameter estimation. Specifically, these methods assume a signal model and calculate detection statistics, such as the generalized likelihood ratio test (GLRT), through maximum likelihood estimation. However, their drawbacks include reliance on the parameter model of the noise distribution, computational complexity, and sensitivity to model mismatch.

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] It should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0014] It should be understood that the terms "system," "apparatus," "unit," and / or "module" used in this application are a method of distinguishing different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they may be replaced by other expressions.

[0015] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "a," and / or "the" are not specifically singular and may include the plural. Generally, the terms "comprising" and "including" only indicate the inclusion of expressly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements. An element defined by the phrase "comprising an..." does not exclude the presence of other identical elements in the process, method, product, or apparatus that includes the element.

[0016] In the description of the embodiments of this application, "a plurality of" refers to two or more. The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0017] Furthermore, flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, the steps can be processed in reverse order or simultaneously. Additionally, other operations can be added to these processes, or one or more steps can be removed from them.

[0018] Reference Figure 1 The diagram below illustrates an optional example of the periodic signal detection method that integrates rank statistics and residual convolutional neural networks proposed in this invention. This method can be applied to computer devices, and the detection method proposed in this embodiment may include, but is not limited to, the following steps: Step S1: Convert the observation sequence into a rank statistic sequence; Step S2: Calculate the covariance matrix of the rank statistic sequence; Step S3: Construct and train a detection model based on a convolutional neural network; Step S4: Input the covariance matrix into the trained detection network to obtain the detection results.

[0019] Data flow direction of the overall process of this invention (refer to) Figure 2 It simultaneously realizes signal presence detection and unknown frequency identification, meeting the dual requirements of sensitivity to weak periodic components and classification accuracy in practical engineering.

[0020] In some feasible embodiments, step S1 specifically includes: Observations Convert to rank ,Right now yes In the sample The ascending order ranking (minimum value is 1, maximum value is 1) When the observations are the same, the average rank strategy is used.

[0021] In some feasible embodiments, step S2 specifically includes: set up for The corresponding rank statistic vector, according to the definition of the sample covariance matrix, the formula for calculating the sample rank covariance matrix is ​​as follows: in Indicates transpose. This indicates the calculation of the mean. The covariance matrix in this scheme has not undergone additional normalization.

[0022] In some feasible embodiments, step S3 specifically includes: Receiver system model: in This refers to the sequence of signals received by the receiver. To simulate the impulse interference noise in the environment, this study uses Bernoulli Gaussian noise with the following probability density function: in This represents the first Gaussian noise (variance is...). The probability of ) occurring, This represents the second Gaussian noise (variance is...). The probability of ) occurring, and . For a periodic signal of unknown frequency to be detected, it is defined as follows: in ,and and Coprime.

[0023] Based on the receiver system model, five types of tags ("0", "1", "2", "3", and "4") are generated for the received signals. The data. The label "0" represents the periodic signal to be detected. The received signal data when it is not present; label "1" represents the periodic signal to be detected. It exists and has a frequency of Label "2" represents the periodic signal to be detected. It exists and has a frequency of The label "3" represents the periodic signal to be detected. It exists and has a frequency of Label "4" represents the periodic signal to be detected. It exists and has a frequency of .

[0024] Consistent with the process in steps S1 and S2, generate the rank statistic and the sample rank covariance; Split the training and test sets: Divide the dataset of sample rank covariance matrices of the five-class labels into training set:test set = 8:2.

[0025] The training set is fed into the designed residual convolutional neural network for model training. The loss function used is the cross-entropy loss function, defined as: in It refers to the batch size. It is the number of categories. It is the one-hot encoding of the real label (if the sample Category The value is 1 if it is 1, otherwise it is 0. It is the model on the sample Predicted as The probability is calculated using the Softmax function: in It is the original score output by the last layer (FC3) of the model.

[0026] The residual convolutional neural network structure is as follows: To achieve the above two objectives, designing a suitable neural network structure is necessary. Since the sample rank covariance matrix is... Given the two-dimensional matrix data, we introduce a convolutional neural network (CNN) to extract features from it. Furthermore, to further improve detection performance and make model training more stable, this embodiment uses two residual blocks to optimize the original CNN. The specific residual convolutional network structure diagram is shown below. Figure 3 As shown in Table 1, the number of parameters for each layer is as follows: express The size is The convolution kernel.

[0027] Table 1. Number of parameters per layer in residual convolutional neural networks Data flow in residual convolutional neural networks: Let n=40, and the input layer input a 40×40 two-dimensional matrix.

[0028] Initial convolutional layers: (1) The first convolutional layer (Conv1) takes a 40×40 two-dimensional matrix as input and outputs 32 40×40 two-dimensional matrices. It includes a batch normalization (BN) layer and a ReLU activation function layer (ReLU). (2) The first pooling layer (Pool1) takes 32 40×40 two-dimensional matrices as input and outputs 32 20×20 two-dimensional matrices.

[0029] Residual Block 1: (1) The second convolutional layer (Conv2) takes 32 20×20 two-dimensional matrices as input and outputs 32 20×20 two-dimensional matrices. It includes one batch normalization layer (BN) and one ReLU activation function layer (ReLU). (2) The third convolutional layer (Conv3) takes 32 20×20 two-dimensional matrices as input and outputs 32 20×20 two-dimensional matrices. It includes one batch normalization layer (BN). (3) The 32 20×20 two-dimensional matrices output by the third convolutional layer (Conv3) are added to the 32 20×20 two-dimensional matrices output by the first pooling layer (Pool1) to output 32 20×20 two-dimensional matrices. (4) The first independent ReLU activation function layer (ReLU1) takes 32 20×20 two-dimensional matrices as input and outputs 32 20×20 two-dimensional matrices. (5) The second pooling layer (Pool2) takes 32 20×20 two-dimensional matrices as input and outputs 32 10×10 two-dimensional matrices. At this time, the 32 10×10 two-dimensional matrices output by the second pooling layer (Pool2) are the output of residual block 1.

[0030] Residual Block 2: (1) The fourth convolutional layer (Conv4) takes 32 10×10 two-dimensional matrices as input and outputs 64 10×10 two-dimensional matrices. It includes one batch normalization layer (BN) and one ReLU activation function layer (ReLU). (2) The fifth convolutional layer (Conv5) takes 32 10×10 two-dimensional matrices as input and outputs 64 10×10 two-dimensional matrices. This layer has a skip connection structure and is used for residual block dimension matching. (3) The sixth convolutional layer (Conv6) takes 64 10×10 two-dimensional matrices as input and outputs 64 10×10 two-dimensional matrices. It includes one batch normalization layer (BN). (4) The 64 10×10 two-dimensional matrices output by the fifth convolutional layer (Conv5) are added to the 64 10×10 two-dimensional matrices output by the sixth convolutional layer (Conv6) to output 64 10×10 two-dimensional matrices. (5) The second independent ReLU activation function layer (ReLU2) takes 64 10×10 two-dimensional matrices as input and outputs 64 10×10 two-dimensional matrices. (6) The third pooling layer (Pool3) takes 64 10×10 two-dimensional matrices as input and outputs 64 5×5 two-dimensional matrices. At this time, the 64 5×5 two-dimensional matrices output by the third pooling layer (Pool3) are the output of residual block 2.

[0031] Global average pooling layer: The fourth pooling layer (Pool4) takes 64 5×5 two-dimensional matrices as input and outputs 64 1×1 scalars.

[0032] Fully connected layers: (1) The first fully connected layer (FC1) takes a 1×64 one-dimensional vector as input and outputs a 1×64 one-dimensional vector, which includes a ReLU activation function layer (ReLU). (2) The second fully connected layer (FC2) takes a 1×128 one-dimensional vector as input and outputs a 1×64 one-dimensional vector, which includes a ReLU activation function layer (ReLU).

[0033] Output layer: The third fully connected layer (FC3) takes a 1×64 one-dimensional vector as input and outputs a 1×5 one-dimensional vector. The final output is calculated by the softmax function and then output.

[0034] The convolutional neural network architecture with dual residual blocks designed in this embodiment, combined with batch normalization and Dropout regularization, effectively captures the deep statistical features of periodic signals and accelerates model convergence.

[0035] Based on the overall process of the above method, this invention also provides relevant data examples: Unknown frequency detection – Rotating machinery fault detection Scenario: Vibration signal of rotating machinery, sampling rate Hz, analysis length Pre-set Divide into 50 equal parts, each part has a sampling length of It is necessary to detect unknown fault characteristic frequencies.

[0036] step: Acquire vibration signal sequence , ,..., .

[0037] Calculate the rank statistic sequence for each sample segment. , ,..., .

[0038] Calculate the sample rank covariance matrix: The dataset was integrated. For each of the five label classes "0", "1", "2", "3", and "4", 10,000 samples were collected and constructed, resulting in a dataset of 50,000 samples.

[0039] Offline training, sending the dataset in Figure 3 The residual convolutional neural network designed in the middle is used to train the model until the loss function converges and the training is completed.

[0040] Online detection, length of each analysis The signal data is preset to... Divide into 50 equal parts, each part has a sampling length of Then, by converting the original sample signals into corresponding rank statistics and calculating the corresponding sample rank covariance matrix, the sample rank covariance matrix is ​​fed into the trained residual convolutional neural network to identify the presence and frequency of periodic signals.

[0041] Results: It can still accurately detect the existence of periodic signals and reliably identify periodic components of unknown frequencies even under non-Gaussian impulse noise.

[0042] Experimental conditions: Collect "0" (periodic signal does not exist) and "1" (periodic signal exists and the signal frequency is high). ), "2" (a periodic signal exists and the signal frequency) ), "3" (a periodic signal exists and the signal frequency) ), "4" (a periodic signal exists and the signal frequency) The dataset consists of five labeled classes, with 10,000 samples collected for each class, and these samples are combined to construct a dataset of 50,000. The noise is Bernoulli Gaussian noise, with the parameters set as follows: , , The training parameters were set as follows: batch size 64, training epochs 50, learning rate 0.001, and test set split ratio 0.2. The optimizer used the Adam algorithm with a first-order moment decay coefficient. Second-order moment attenuation coefficient The convolutional layers are initialized using the Kaiming normal distribution, and the batch normalized layer weights are initialized to 1 and the biases are initialized to 0. Other parameters are shown in Table 1.

[0043] Experimental Results: To better demonstrate the advantages of using the sample rank covariance matrix as data preprocessing, we compared the accuracy of this method with the traditional method that uses the sample covariance matrix as the dataset in identifying periodic signals in Bernoulli Gaussian noise. Tables 2 and 3 show that the method of this invention achieves precision, recall, and F1 score all above 0.9 for the five classes of samples, significantly higher than the traditional method based on convolutional neural networks using the sample covariance matrix. Figure 4 and Figure 5 The confusion matrices of five classes of samples are presented using two methods, and the results further confirm that the method of this invention is significantly superior to traditional methods. The solution of this invention can achieve accurate detection and frequency identification of periodic signals with unknown frequencies under Bernoulli Gaussian noise.

[0044] Table 2. Accuracy index of the method of the present invention for identifying five types of samples under Bernoulli Gaussian noise. Table 3. Accuracy metrics of traditional convolutional neural network-based methods using sample covariance matrices for identifying five classes of samples under Bernoulli Gaussian noise. A periodic signal detection system integrating rank statistics and residual convolutional neural networks, comprising: The conversion module is used to perform step S1; The covariance matrix calculation module is used to execute step S2; The training module is used to execute step S3; The online detection module is used to perform step S4.

[0045] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0046] A periodic signal detection device that integrates rank statistics and residual convolutional neural networks: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements a periodic signal detection method that integrates rank statistics and residual convolutional neural networks as described above.

[0047] The content of the above method embodiments is applicable to the device embodiments. The specific functions implemented by the device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0048] A storage medium storing processor-executable instructions, which, when executed by a processor, are used to implement a periodic signal detection method that integrates rank statistics and residual convolutional neural networks as described above.

[0049] The content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0050] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for detecting periodic signals that integrates rank statistics and residual convolutional neural networks, characterized in that, Includes the following steps: Convert the observation sequence into a rank statistic sequence; Calculate the covariance matrix of the rank statistic sequence; The covariance matrix is ​​input into a pre-trained detection network to obtain the detection results.

2. The method for detecting periodic signals by fusing rank statistics and residual convolutional neural networks according to claim 1, characterized in that, The receiver's system model is as follows: in, Indicates the observed signal, For an unknown frequency periodic signal, This represents a sequence of interfering noise.

3. The method for detecting periodic signals by fusing rank statistics and residual convolutional neural networks according to claim 2, characterized in that, The step of converting the observed sequence into a rank statistic sequence specifically includes: Obtain the observation sequence; The observed values ​​in the observation sequence are converted into ranks. When the observed values ​​are the same, the average rank strategy is used to obtain the rank statistical sequence.

4. The method for detecting periodic signals by fusing rank statistics and residual convolutional neural networks according to claim 2, characterized in that, The formula for calculating the covariance matrix is ​​as follows: in, This indicates calculating the mean. Indicates transpose. This represents a sequence of rank statistics.

5. The method for detecting periodic signals by fusing rank statistics and residual convolutional neural networks according to claim 2, characterized in that, The detection network includes an input layer, an initial convolutional layer, a first residual block, a second residual block, a global average pooling layer, a fully connected layer, and an output layer.

6. The method for detecting periodic signals by fusing rank statistics and residual convolutional neural networks according to claim 1, characterized in that, The training process of the detection model includes: Based on the receiver system model, generate data for the received signals of at least two types of tags; Based on the data, the training observation sequence is obtained, the rank statistic is calculated, and the sample rank covariance matrix is ​​constructed. The sample rank covariance matrix corresponding to the label is divided to obtain the training set and the test set; The detection model is trained based on the training set and the test set until the model converges, resulting in a fully trained detection model.

7. The method for detecting periodic signals by fusing rank statistics and residual convolutional neural networks according to claim 4, characterized in that, The loss function during the training process of the detection network is: in, Indicates batch size. Indicates the number of categories. Indicates the true label, The model represents the samples Predicted as The probability of.

8. A periodic signal detection system integrating rank statistics and residual convolutional neural networks, characterized in that, include The conversion module is used to convert the observed sequence into a rank statistic sequence; The covariance matrix calculation module is used to calculate the covariance matrix of the rank statistic sequence; An online detection module is used to input the covariance matrix into a pre-trained detection network to obtain detection results.

Images