Method for adaptively classifying accidental abnormal signals in digital oscilloscope

Abstract

The invention discloses a method for adaptively classifying accidental abnormal signals in a digital oscilloscope, which comprises the following steps of: firstly, acquiring template signals and system noise, performing sample expansion on data, judging captured abnormal signal data through a reconstructed abnormal signal judgment model based on RNN and an automatic encoder, and judging whether the abnormal signal data is abnormal or not; judging whether the signal is of a new signal type or not based on the reconstruction error; and if the reconstruction error is smaller than a certain threshold value, considering that the abnormal signal type is known, then sending the abnormal signal data into a supervised signal classification model based on RNN, and finally obtaining the classification of the current abnormal signal.

Description

Technical Field

[0001] The present invention belongs to the technical field of oscilloscopes, and more specifically, relates to a method for classifying adaptive sporadic abnormal signals in a digital oscilloscope. Background Art

[0002] With the rapid development of electronic technologies such as communication and radar, the signal frequency has been continuously increasing, and abnormal phenomena under transient and non-stationary characteristics have become increasingly prominent. Against this background, on the premise that a high-speed acquisition system has the ability to detect complex sporadic abnormal signals, it is also necessary to effectively classify these sporadic abnormal signals, be able to quickly identify multiple sporadic abnormal types, and enhance the system's judgment and response capabilities for different sporadic abnormalities. This ability is crucial for real-time evaluating the impact of sporadic abnormalities on system performance, quickly locating system faults, and improving system reliability. However, in actual tests, the complexity, sporadity, and diversity of abnormal signals, as well as noise interference, etc., have brought many challenges to the classification of sporadic abnormal signals. The sporadity makes the appearance of abnormal signals random, and the sporadity of the change of their characteristics increases the difficulty of classification. In different scenarios, different sporadic abnormal signals may also exhibit similar characteristics, which may not only lead to inaccurate classification results, but also trigger repeated analysis, thus increasing the complexity of data processing. Therefore, this research aims to explore the classification method of complex sporadic signals, and by improving the accuracy and reliability of classifying complex sporadic abnormal signals, to achieve effective classification and analysis of sporadic abnormal signals, improve the processing efficiency of the acquisition system for sporadic abnormal signals, and enhance the accuracy of classification.

[0003] In recent years, driven by technologies such as compressive sensing and deep learning, significant progress has been made in the field of signal classification, not only improving the classification accuracy, but also enhancing the robustness of the algorithm in complex environments. Signal classification methods can be divided into three categories: unsupervised, supervised, and semi-supervised learning, each showing different advantages and limitations. Unsupervised classification does not rely on labeled data and mainly realizes automatic classification by analyzing the characteristics of data. Supervised learning relies on training data with clear labels and generates a model by learning the mapping relationship between data and labels. In the acquisition system, the sporadity of these abnormal signals results in insufficient sample quantity, and coupled with the noise interference in the acquisition environment, it makes it difficult for the classification model to accurately identify and distinguish sporadic abnormal types. In addition, new types of sporadic abnormal signals may appear in the acquisition system, which requires the classification method to have strong generalization ability and high-precision real-time classification performance to meet the requirements of uncertainty and complex environments. Therefore, in a high-speed acquisition system, the research focus of signal classification methods needs to pay attention to the adaptability and robustness of the algorithm in low-sample data sets, with noise interference, and when new types of complex and variable signals appear, in order to further improve the classification accuracy of sporadic abnormal signals to meet the classification requirements of abnormal sporadic signals in a complex and variable environment. Summary of the Invention

[0004] The object of the present invention is to overcome the deficiencies of the prior art and provide a method for adaptively classifying sporadic abnormal signals in a digital oscilloscope, an adaptive classification method for sporadic abnormal signals based on reconfigurable and self-classifying, to solve the problems of inaccurate classification and inability to adaptively classify caused by few complex sporadic signal samples and noise interference.

[0005] To achieve the above object of the invention, a method for adaptively classifying sporadic abnormal signals in a digital oscilloscope of the present invention is characterized by comprising the following steps:

[0006] (1) Deploy a signal recognition model based on RNN and an autoencoder and a classification model based on RNN in the upper computer of the intelligent oscilloscope, and complete the offline training of the two models;

[0007] (2) Template signal acquisition;

[0008] (2.1) Input a normal signal into the intelligent oscilloscope, and obtain M-channel parallel sampling data through the acquisition system, which is represented in matrix form as:

[0009]

[0010] where x ij represents the i-th sampling value collected at the j-th sampling moment, i = 1, 2,..., M, j = 1, 2,..., N, N is the number of sampling moments, and the value of N is at least one complete signal period;

[0011] (2.2) When there is no signal input to the intelligent oscilloscope, obtain M-channel parallel sampling data through the acquisition system, which is represented in matrix form as:

[0012]

[0013] Normalize the matrix to obtain the normalized matrix where each element after normalization satisfies: n is the quantization bit number of the ADC in the acquisition system;

[0014] (2.3) Store the matrix X and the matrix as a template matrix in the upper computer of the intelligent oscilloscope;

[0015] (3) Randomly access the input signal and turn on the adaptive classification function for sporadic abnormal signals in the intelligent oscilloscope;

[0016] (3.1) Connect the input signal to the intelligent oscilloscope, and obtain M-channel parallel sampling data through the acquisition system, which is represented in matrix form as:

[0017]

[0018] (3.2) Convert the matrix into a serial vector

[0019] (3.3) Assume that the length of the input vector of the signal recognition model is L; compare the length MN of the serial vector with the input length L of the signal recognition model. If MN = L, then take the serial vector as the abnormal residual vector, and then jump to step (3.4); if MN > L, then fill 0 at the end of the serial vector to extend the length of the serial vector to λ times of L, then perform decimation by λ times to obtain an abnormal residual vector with length L, and jump to step (3.4); if MN < L, then fill 0 at the end of the serial vector to extend the length of the serial vector to L to obtain an abnormal residual vector with length L, and then enter step (3.4);

[0020] (3.4) Randomly generate a group of noise vectors subject to normal distribution, and the length of the noise vectors is L; superimpose the noise vectors on the abnormal residual vector obtained in step (3.3) to obtain the input vector of the signal recognition model;

[0021] (3.5) Normalize the input vector obtained by superimposition in step (3.4) and then input it into the signal recognition model to predict a serial reconstruction signal with length L;

[0022] (3.6) Represent the serial reconstruction signal in matrix form as:

[0023]

[0024] The host computer calculates the difference matrix ΔX according to the matrix X and the matrix , and its elements satisfy:

[0025]

[0026] Compare each element Δx in the difference matrix ΔX ij with the corresponding element in the template matrix . If each element at the corresponding position satisfies , then determine that the input signal is an occasional abnormal signal of a known type and enter step (3.8); otherwise, determine that the input signal is a new abnormal type signal and enter step (3.7);

[0027] (3.7) Add a new output class label to the signal recognition model and save the weight parameters corresponding to the signal recognition model under this class label;

[0028] (3.8) Normalize the serial reconstructed signal and input it into the RNN-based classification model to obtain the abnormal classification category of the current input signal, and send it to the host computer for display.

[0029] The invention object of the present invention is realized as follows:

[0030] A method for adaptive sporadic abnormal signal classification in a digital oscilloscope of the present invention first collects template signals and system noises, expands the data as samples, judges the captured abnormal signal data through a reconstructed abnormal signal determination model based on RNN and autoencoder, and judges whether the signal is a new signal type based on the reconstruction error; if the reconstruction error is less than a certain threshold, it is considered a known abnormal signal type, and then the abnormal signal data is sent to a supervised signal classification model based on RNN, and finally the classification of the current abnormal signal is obtained.

[0031] A method for adaptive sporadic abnormal signal classification in a digital oscilloscope of the present invention also has the following beneficial effects:

[0032] (1) The present invention adopts an adaptive classification method, and by combining the models of RNN and autoencoder, significantly improves the classification accuracy of the system for new types or unknown abnormal signals. Compared with traditional methods, the present invention can better identify and classify those new sporadic abnormal signals that have not been seen before, enhances the adaptability and recognition ability of the system, and avoids the problem of high misrecognition rate in traditional methods.

[0033] (2) Based on limited abnormal signal samples, the present invention expands the samples through digital enhancement processing technology of data. Solve the problem of poor model generalization ability caused by scarce samples. Solve the problem that sporadic abnormal signals are easily affected by noise.

[0034] processing technology to expand the samples. Solve the problem of poor model generalization ability caused by scarce samples. Solve the problem that sporadic abnormal signals are easily affected by noise.

[0035] (3) By adopting a difference threshold determination and anti-noise processing mechanism, the present invention effectively filters out noise interference and improves the abnormal signal detection ability of the system in a complex environment. By setting a difference threshold and combining system noise analysis, it can accurately distinguish normal signals from abnormal signals, ensuring that the system can still maintain high accuracy in the presence of noise.

[0036] (4) The present invention supports the adaptive update and incremental training of the signal classification model. Whenever a new type of abnormal signal is recognized, the system can automatically add classification labels and continuously improve the classification model through incremental training. This feature enables the system to have dynamic adjustment capabilities, capable of coping with changing signal characteristics and avoiding the problem of the model "forgetting" previous signal types.

[0037] (5) With the iteration of training and the improvement of the model's recognition ability, the present invention can improve classification accuracy by dynamically adjusting the threshold of the difference. Especially in complex environments, as the classification model is continuously optimized, it can update the classification criteria in real time, improve classification accuracy, and avoid classification biases that may be caused by fixed thresholds.

[0038] (6) Through the adaptive classification function of the intelligent oscilloscope, the present invention realizes the automatic recognition and classification of signals without manual intervention. This method can automatically judge and classify various types of sporadic abnormal signals according to the characteristics of the signals, improving the automation level of the system and providing a more effective tool for real-time monitoring and fault diagnosis. Description of the Drawings

[0039] Figure 1 is the flowchart of the method for adaptive sporadic abnormal signal classification in a digital oscilloscope of the present invention;

[0040] Figure 2 is the structural diagram of the signal recognition model based on RNN and autoencoder;

[0041] Figure 3 is the structural diagram of the classification model based on RNN;

[0042] Figure 4 is the schematic diagram after the triangular wave is reconstructed;

[0043] Figure 5 is the flowchart of abnormal signal classification. Detailed Embodiments

[0044] The following describes the detailed embodiments of the present invention with reference to the drawings, so that those skilled in the art can better understand the present invention. It should be particularly noted that in the following description, when the detailed description of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.

[0045] Embodiment

[0046] In this embodiment, as Figure 1 shown, the method for adaptive sporadic abnormal signal classification in a digital oscilloscope of the present invention includes the following steps:

[0047] (1) Deploy a signal recognition model based on RNN and autoencoder and an RNN-based classification model in the upper computer of the intelligent oscilloscope, and complete the offline training of the two models;

[0048] In this embodiment, as Figure 2 shown, the signal recognition model based on RNN and autoencoder consists of an encoder and a decoder;

[0049] The encoder consists of a first-layer RNN, a dropout layer, a second-layer RNN, and a hidden layer; the first-layer RNN receives an input vector with a length of L and a dimension of 1 dimension, extracts 128 features at each time, and outputs a feature matrix with a dimension of (L, 128); the dropout probability of the dropout layer is set to 0.2, and 20% of the outputs of the RNN units are randomly discarded each time of update; the second-layer RNN receives the output from the first layer, with a dimension of (L, 128), and continues to extract features, and its output dimension is still (L, 128); the output of the encoder is reduced in dimension through the hidden layer to obtain a latent vector z with a length of 32;

[0050] The decoder consists of a first-layer RNN, a dropout layer, a second-layer RNN, and a fully connected layer; the input of the first-layer RNN in the decoder is the latent vector z, and it is decoded through the first-layer RNN containing 128 units to obtain an output dimension of (L, 128), and the dropout layer sets the dropout probability to 0.2, and 20% of the outputs of the RNN units are randomly discarded each time of update; the second-layer RNN also contains 128 RNN units, and the output is (L, 128); the output of the decoder is mapped back to the original signal space through a fully connected layer to obtain an output vector with a length of L and a dimension of 1 dimension.

[0051] As Figure 3 shown, the RNN-based classification model includes: the first-layer RNN receives an abnormal signal sequence with an input signal length of L and a dimension of 1 dimension, extracts 128 features at each time, and outputs a feature matrix with a dimension of (L, 128); the dropout layer sets the dropout probability to 0.2, and 20% of the outputs of the RNN units are randomly discarded each time of update; the second-layer RNN receives the output from the first layer, with a dimension of (L, 128), and continues to extract features, and its output dimension is still (L, 128); after the second-layer RNN layer, it enters a fully connected layer, and the fully connected layer contains 64 units, which is used to perform a non-linear transformation on the high-dimensional time series features and activate them with the ReLU activation function to obtain a feature vector with a dimension size of 64; the feature vector enters the output layer, and the Softmax activation function is used to classify the feature vector, and the category with the highest probability is selected as the abnormal type of the input signal.

[0052] (2) Template signal acquisition;

[0053] (2.1) In this embodiment, a sine signal with a frequency of 100 MHz is given as a template signal, and then it is input into the intelligent oscilloscope. M parallel sampling data are obtained through the acquisition system and represented in matrix form as:

[0054]

[0055] Where x ij represents the i-th sampling value collected at the j-th sampling moment, i = 1, 2,..., M, j = 1, 2,..., N, N is the number of sampling moments, and the value of N is at least one complete period of the signal;

[0056] In this embodiment, the sampling rate is 20 Gsps, the number of M parallel channels is 80 at a 250 M system clock, and there are 200 points in one period, so N is at least 3.

[0057] (2.2) When there is no signal input to the intelligent oscilloscope, M parallel sampling data are obtained through the acquisition system and represented in matrix form as:

[0058]

[0059] The matrix is normalized to obtain the normalized matrix Where each element after normalization satisfies: n is the quantization bit number of the ADC in the acquisition system;

[0060] In this embodiment, n is 12 bits.

[0061] (2.3) The matrices X and are stored as template matrices in the upper computer of the intelligent oscilloscope;

[0062] (3) Randomly access the input signal, and turn on the adaptive classification function for occasional abnormal signals in the intelligent oscilloscope. The classification process is as Figure 5 shown;

[0063] (3.1) In this embodiment, the input signal accessed is a triangular wave of 100 MHz to the intelligent oscilloscope. M parallel sampling data are obtained through the acquisition system and represented in matrix form as:

[0064]

[0065] (3.2) The matrix is converted into a serial vector

[0066] (3.3) Assume that the length of the input vector of the signal recognition model is L; compare the serial vector The length MN of and the input length L of the signal recognition model. If MN = L, then the serial vector is used as the abnormal residual vector, and then it jumps to step (3.4); if MN > L, then 0 is padded at the end of the serial vector to expand the length of the serial vector to λ times of L, then perform decimation by λ times to obtain an abnormal residual vector with a length of L, and jump to step (3.4); if MN < L, then 0 is padded at the end of the serial vector to expand the length of the serial vector to L to obtain an abnormal residual vector with a length of L, and then enter step (3.4);

[0067] In this embodiment, L = 256 and MN = 240, then it is padded with zeros to 256 first to obtain an abnormal residual vector with a length of 256.

[0068] (3.4) Randomly generate a group of noise vectors that follow a normal distribution, and the length of the noise vectors is L; superimpose the noise vectors on the abnormal residual vector obtained in step (3.3) to obtain the input vector of the signal recognition model;

[0069] In this embodiment, Gaussian white noise with a mean of 0 and a standard deviation of 0.03 is superimposed.

[0070] (3.5) Normalize the input vector obtained by superimposition in step (3.4) and then input it into the signal recognition model to predict a serial reconstruction signal with a length of L;

[0071] In this embodiment, the reconstructed signal obtained as output is as Figure 4 shown, and the error is 0.0215.

[0072] (3.6) Represent the serial reconstruction signal in matrix form as:

[0073]

[0074] The host computer calculates the difference matrix ΔX according to the matrix X and the matrix , and its elements satisfy:

[0075]

[0076] Compare the elements Δx ij in the difference matrix ΔX with the corresponding elements in the template matrix . If the elements at the corresponding positions all satisfy If it is determined that the input signal is an occasional abnormal signal of a known type, proceed to step (3.8); otherwise, determine that the input signal is a new abnormal type signal and proceed to step (3.7).

[0077] In this embodiment, it is determined as a new abnormal type signal.

[0078] (3.7) Add a new output category label to the signal recognition model and save the weight parameters corresponding to the signal recognition model under this category label.

[0079] (3.8) After normalizing the serial reconstruction signal, input it into the RNN-based classification model to obtain the abnormal classification category of the current input signal and send it to the host computer for display.

[0080] Although the above describes the illustrative specific embodiments of the present invention for the convenience of those skilled in the art to understand the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are within the scope of protection.

Claims

1. A method for adaptively classifying occasional abnormal signals in a digital oscilloscope, characterized in that: The following steps are involved: (1) Deploy a signal recognition model based on RNN and autoencoder and a classification model based on RNN in the host computer of the intelligent oscilloscope, and complete offline training of the two models; (2) Template signal acquisition; (2.1) Input the normal signal into the intelligent oscilloscope, and obtain M parallel sampling data through the acquisition system, which is expressed in matrix form as follows: Among them, x ij represents the i-th sampling value collected at the j-th sampling time, i = 1, 2, ..., M, j = 1, 2, ..., N, N is the number of sampling times, and the value of N contains at least one complete signal period; (2.2) When there is no signal input to the intelligent oscilloscope, M parallel sampling data are obtained through sampling by the acquisition system, which is expressed in matrix form as follows: Pair Matrix Perform normalization to obtain the normalized matrix Among them, each element after normalization satisfies: n is the number of quantization bits of the ADC in the acquisition system; (2.3), matrix X and matrix It is stored as a template matrix in the host computer of the intelligent oscilloscope; (3) Randomly access the input signal and enable the adaptive classification function of occasional abnormal signals in the intelligent oscilloscope; (3.1) The input signal is connected to the intelligent oscilloscope, and M parallel sampling data are obtained through sampling by the acquisition system, which is expressed in matrix form as follows: (3.2), the matrix Convert to a serial vector (3.3), assuming that the length of the input vector of the signal recognition model is L; compare the serial vector The length MN of the signal recognition model is the same as the input length L. If MN = L, then the serial vector As the abnormal residual vector, jump to step (3.4); if MN>L, then in the serial vector The end of the serial vector is filled with 0 The length of L is extended to λ times, Then perform λ times extraction to obtain an abnormal residual vector of length L, and jump to step (3.4); if MN < L, then in the serial vector The end of the serial vector is filled with 0 The length of is expanded to L, and an abnormal residual vector of length L is obtained, and then step (3.4) is entered; (3.4) randomly generate a set of noise vectors that obey the normal distribution, and the length of the noise vector is L; superimpose the noise vector with the abnormal residual vector obtained in step (3.3) to obtain the input vector of the signal recognition model; (3.5) The input vector obtained by superposition in step (3.4) is normalized and then input into the signal recognition model, thereby predicting a serial reconstructed signal with a length of L; (3.6), the serial reconstructed signal is expressed in matrix form as: The host computer calculates the matrix X and the matrix Calculate the difference matrix ΔX, whose elements satisfy: Compare the elements Δx in the difference matrix ΔX ij With the template matrix The corresponding element The size of, if the elements at the corresponding positions all satisfy Then the input signal is determined to be an occasional abnormal signal of a known type, and the process goes to step (3.8); otherwise, the input signal is determined to be a signal of a new abnormal type, and the process goes to step (3.7); (3.7) Add an output category label to the signal recognition model and save the weight parameters corresponding to the signal recognition model under the category label; (3.8) The serial reconstructed signal is normalized and then input into the RNN-based classification model to obtain the abnormal classification category of the current input signal and send it to the host computer for display.

2. The method for adaptively classifying occasional abnormal signals in a digital oscilloscope according to claim 1, characterized in that: The signal recognition model based on RNN and autoencoder consists of an encoder and a decoder; The encoder is composed of a first-layer RNN, a drop layer, a second-layer RNN and a hidden layer; the first-layer RNN receives an input vector of length L and dimension 1, extracts 128 features at each time, and outputs a feature matrix of dimension (L, 128); the drop probability of the drop layer is set to 0.2, and 20% of the outputs of the RNN units are randomly dropped at each update; the second-layer RNN receives the output from the first layer, the dimension is (L, 128), and continues to extract features, and its output dimension is still (L, 128); the output of the encoder is reduced in dimension through the hidden layer to obtain an implicit vector z of length 32; The decoder consists of a first-layer RNN, a drop layer, a second-layer RNN and a fully connected layer; the first-layer RNN input in the decoder is an implicit vector z, which is decoded by the first-layer RNN containing 128 units to obtain an output dimension of (L, 128); the drop layer sets the drop probability to 0.2, and randomly drops 20% of the outputs of the RNN units during each update; the second-layer RNN also contains 128 RNN units, and the output is (L, 128); the output of the decoder is mapped back to the original signal space through a fully connected layer to obtain an output vector with a length of L and a dimension of 1.

3. The method for adaptively classifying occasional abnormal signals in a digital oscilloscope according to claim 1, characterized in that: The RNN-based classification model includes: the first layer of RNN accepts an abnormal signal sequence with an input signal length of L and a dimension of 1, extracts 128 features at each time, and outputs a feature matrix with a dimension of (L, 128); the discard layer sets the discard probability to 0.2, and randomly discards 20% of the outputs of the RNN units at each time update; the second layer of RNN receives the output from the first layer, whose dimension is (L, 128), and continues to extract features, and its output dimension is still (L, 128); after the second layer of RNN layer, a fully connected layer is entered, the fully connected layer includes 64 units, which are used to perform nonlinear transformation on high-dimensional time series features, and are activated by ReLU activation function to obtain a feature vector with a dimension size of 64; the feature vector enters the output layer, and the feature vector is classified using the Softmax activation function, and the category with the highest probability is selected as the abnormal type of the input signal.

Images