Rapid non-uniform sampling signal period extraction method and device
The non-uniform sampling signal period is extracted through the neural network model, which solves the problem of high computational complexity in the existing technology, and realizes fast and real-time signal period extraction, which is suitable for non-uniform sampling scenarios.
Patent Information
- Application Number
- CN202410997188.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-24
- Publication Date
- 2024-11-15
AI Technical Summary
Under non-uniform sampling conditions, existing signal period extraction methods such as fast Fourier transform and discrete Fourier transform have high computational complexity and delay problems, making it difficult to meet real-time requirements and other fields such as radar detection and health detection. need.
The neural network model is adopted to generate signals and tags through simulation, train neural networks, preprocess signals, and use neural networks to extract signal cycles to reduce the computational complexity.
Reducing the computational complexity from O(N2) to O(N), significantly reducing the computational time-consuming and improving real-time and efficiency.
Smart Images

Figure CN118964931A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of signal processing, and in particular relates to a method and a device for quickly extracting a period of a non-uniform sampling signal. Background Art
[0002] The period of the signal contains rich target information. By studying the period of the signal, a lot of valuable intelligence can be mined. For example, the period extraction of the heart sound signal can reflect the health and emotional state of the human body. The analysis of the sound period of the train track can reflect the wear of the wheels. In the field of radar detection, the radar cross-sectional area (RCS) period of the space target can reflect the load and motion state of the space target, and the RCS period of the ballistic target can reflect the attitude control of the target. Therefore, how to effectively extract the periodic information of the signal is of great significance in the fields of health, manufacturing, national defense, etc.
[0003] In actual scenarios, due to the limitation of sensor resources or the shielding of the detection target, the sampled signals usually obtained are non-uniform, that is, the time intervals between each sampling point are not completely equal. In a non-uniform sampling environment, the commonly used period extraction method is no longer applicable. For example, the common spectrum analysis methods under uniform sampling conditions include the fast Fourier transform method (FFT) and the autocorrelation function method. However, in non-uniform sampling, the phases that need to be compensated for each sampling point are not equal, and the time intervals between adjacent sampling points are different. The above two methods are not applicable. Usually, for non-uniform sampling scenarios, the methods used include the discrete Fourier transform method (DFT) and the sparse reconstruction method. The discrete Fourier transform method needs to search the signal in the frequency domain to obtain the spectrum of the signal, and calculate the corresponding signal period through processes such as peak finding. However, since the discrete Fourier transform needs to search the frequency domain, it has problems such as high computational complexity. For a sequence with a signal length of N, its computational complexity is O(N2), and for scenarios such as radar detection with high real-time requirements, there will be obvious delays. Therefore, finding a faster method suitable for non-uniform sampling is of great significance for radar detection, health detection, industrial control and other fields. Summary of the invention
[0004] In order to solve the above problems, the present invention proposes a fast method for extracting the period of a non-uniform sampling signal, comprising the following steps:
[0005] (10) Generate simulated periodic signals and labels;
[0006] (20) Constructing a neural network model;
[0007] (30) Training a neural network model;
[0008] (40) preprocessing the obtained signal;
[0009] (50) Signal period extraction.
[0010] Specifically, the step (10) comprises the following steps:
[0011] (11) Determine signal parameters: According to the pattern of the real signal to be measured, determine the waveform W, period T, amplitude range A, original sampling rate Fs, and signal length L of the simulated signal; the period T and the original sampling rate Fs are selected according to the situation of the real signal; the signal length L contains more than one signal period, that is, L>T×Fs;
[0012] (12) Generate original signal and label: Generate original signal and label according to the signal parameters determined in (11); the original signal includes signal amplitude value and sampling time, and the label is the period T when the signal is generated, wherein each original signal is in the form of an L×2 array, the first column is the signal amplitude value, the second column is the corresponding sampling time, and the label is the value of the signal period T;
[0013] (13) First difference of time series: perform a difference on the sampling time in the second column of the L×2 array in step (12), then add 0 before the difference result sequence to ensure that the length is still L, and fill the new sequence into the second column of the original L×2 array, covering the original sampling time;
[0014] (14) Adding noise: adding noise to the signal amplitude value of the original signal obtained in (13);
[0015] (15) Perform non-uniform sampling on the signal: For the signal obtained in step (14), some sampling points are randomly discarded to generate a non-uniform sampling sequence.
[0016] Specifically, the step (20) of constructing a neural network model includes:
[0017] (21) Determine the model input length: Determine the length Lm of the signal sequence when it is input into the model. This length should be greater than the product of the maximum value of the period T selected in (11) and the maximum value of the original sampling rate Fs, that is, Lm>max(T)×max(Fs), and less than the signal length, that is, Lm <L;
[0018] (22) Selecting a model type: selecting a neural network model type, wherein the neural network model types include a recurrent neural network (RNN) and a convolutional neural network (CNN);
[0019] (23) Setting model parameters: Determine the corresponding parameters of the model selected in (22), including the number of network layers and nodes;
[0020] Specifically, the step (30) of training the neural network model includes:
[0021] (31) Sliding window interception training samples: According to the model input length Lm determined in (21), the signal generated by (10) is subjected to sliding window interception, the window length is Lm, and the sliding window step size is selected as 1;
[0022] (32) Normalizing the signal amplitude value: For each signal segment of length Lm obtained by the sliding window in step (31), the amplitude value is normalized; the normalization interval is taken as the amplitude range A determined in step (11);
[0023] (33) Dividing the training set and the test set: Determine the division ratio, and divide the training samples generated in (32) into the training set and the test set according to the ratio;
[0024] (34) Setting training parameters: The training parameters include the number of training times, loss function, and early termination conditions; the loss function of the neural network model includes the mean absolute difference function MAE and the mean square error function MSE;
[0025] (35) training the model: according to the training parameters set in (34), the model generated in step (20) is trained using the training set and the test set generated in (33) to obtain a trained model;
[0026] Specifically, the step (40) of preprocessing the obtained signal includes:
[0027] (41) Obtaining the real signal to be tested: Record the signal collected by the sensor to obtain a sequence of length Ls, including the signal amplitude value and sampling time, where Ls ≥ Lm.
[0028] (42) Sliding window signal extraction: Sliding window is performed on the sequence obtained in step (41), the window length is set to be the length Lm determined in step (21), the sliding window step length is set to be 1, and Ls-Lm+1 signal segments of length Lm are obtained by sliding window;
[0029] (43) Normalizing the signal amplitude value: normalizing the amplitude value of the signal segment with a length of Lm obtained by the sliding window in step (42); the normalization interval is taken as the amplitude range A determined in step (11);
[0030] (44) Time series first difference: For the signal segment of length Lm obtained by sliding window in step (42), first difference is performed in sampling time. The length of the sequence obtained after first difference of Lm sampling times is Lm-1. A zero is added before the sequence to ensure that the length is still Lm.
[0031] (45) Sequence merging: Combine the amplitude value sequence of the signal segment with a length of Lm obtained in step (43) and step (44) with the time difference sequence with a length of Lm to obtain an array with a size of Lm×2.
[0032] Specifically, the (50) signal period extraction step includes:
[0033] (51) Inputting the sequence into the model: Input the result generated in step (40) into the trained model obtained in step (30), inputting each signal segment once, for a total of Ls-Lm+1 inputs;
[0034] (52) Extracting the periodic value: Check the output value of the model. There are Ls-Lm+1 outputs in total. Average all the output values and the value obtained is the periodic value of the signal.
[0035] Specifically, step (14) comprises the following steps:
[0036] (141) Determining noise parameters: determining the probability distribution and parameters of the noise; the probability distribution includes Gaussian distribution and chi-square distribution, and the parameters include mean and standard deviation;
[0037] (142) Adding noise: Based on the signal length L in step (12) and the noise parameters determined in (141), a noise sequence of length L is generated and added to the signal amplitude value obtained in (13).
[0038] Specifically, step (15) comprises the following steps:
[0039] (151) Determine the discard ratio: determine the ratio of the total length of the sampling points to be discarded as R;
[0040] (152) Generate a discard index: According to the discard ratio determined in (151), determine the discard number Ld=round(L×R) and randomly generate Ld non-repeating positive integer sequences Idx in the interval [1, L];
[0041] (153) Deleting sampling points: According to the positive integer sequence Idx obtained in (152), delete the row corresponding to Idx from the signal of size L×2 obtained in step (14) to obtain an array of size (L-Ld)×2;
[0042] Specifically, the loss function in step (34) is the mean absolute difference function MAE, which is expressed as follows:
[0043]
[0044] where y i is the model prediction value, is the label value, N is the number of samples, and the early termination condition is set as MAE does not decrease for 10 consecutive times.
[0045] The present invention also provides a fast non-uniform sampling signal period extraction device, including a simulated periodic signal and label generation module, a neural network model construction module, a neural network model training module, a signal preprocessing module and a signal period extraction module; the simulated periodic signal and label generation module, the neural network model construction module, the neural network model training module, the signal preprocessing module and the signal period extraction module perform fast non-uniform sampling signal period extraction according to the steps of any of the above methods.
[0046] Compared with the prior art, the present invention has the following advantages:
[0047] The present invention uses a neural network for the first time to extract the period of non-uniformly sampled signals, overcoming the problem of failure of methods such as fast Fourier transform under non-uniform sampling. Moreover, compared with the traditional discrete Fourier transform method, the method provided by the present invention reduces the amount of computation from O(N2) to O(N), and the computational time is significantly reduced. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 It is an implementation flow chart of an embodiment of the present invention.
[0049] Figure 2 This is one of the periodic signal graphs generated by the simulation.
[0050] Figure 3 This is the second periodic signal graph generated by simulation.
[0051] Figure 4 It is a schematic diagram of the convolutional neural network structure.
[0052] Figure 5 It is a graph of the loss changes on the training set and the test set during network training.
[0053] Figure 6 It is a time series diagram of the measured radar RCS of a certain target.
[0054] Figure 7 for Figure 6 The spectrum diagram of the measured radar RCS time series after DFT analysis.
[0055] Figure 8 It is a time consumption comparison diagram between the method of the present invention and the traditional DFT method. DETAILED DESCRIPTION
[0056] In view of the shortcomings of the prior art, the purpose of the present invention is to propose a method for quickly extracting the period of a non-uniformly sampled signal, which can effectively extract the period of a non-uniformly sampled signal and reduce the computational complexity and improve the computational speed compared to traditional algorithms. This method designs a neural network model and uses simulation to generate training data to train the model. The signal period is extracted using the neural network model and verified by measured data. Compared with the traditional discrete Fourier transform method, this method has the advantages of low computational complexity and good real-time performance.
[0057] In order to facilitate understanding of the present application, the present application will be described more fully below with reference to the relevant drawings. Embodiments of the present application are provided in the drawings. However, the present application can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive.
[0058] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein in the specification of this application are only for the purpose of describing specific embodiments and are not intended to limit this application.
[0059] Reference Figure 1 The present invention discloses a fast method for extracting the period of a non-uniform sampling signal, and the specific implementation steps are as follows:
[0060] (10) Generate simulated periodic signals and labels.
[0061] The steps to generate simulated periodic signals and labels include:
[0062] (11) Determine the signal parameters: According to the style of the real signal to be tested, determine the waveform W, period T, amplitude range A, original sampling rate Fs, and signal length L of the simulated signal. Commonly used waveforms W include sine waves, sawtooth waves, and square waves. The period T and the original sampling rate Fs are selected according to the situation of the real signal. The amplitude range A can usually be -1 to 1 or 0 to 1; the signal length L must contain more than one signal period, that is, L>T×Fs. In this example, the signal waveform W used includes sine waves and sawtooth waves; the period T ranges from 0.5 seconds to 10 seconds, with an interval of 0.5 seconds; the amplitude range A is -1 to 1; the original sampling rate Fs is 10Hz, 20Hz, and 30Hz; the signal length L is fixed to 1000.
[0063] (12) Generate original signal and label: Generate original signal and label according to the signal parameters determined in (11). The original signal includes signal amplitude value and sampling time, and the label is the period T when the signal is generated. In this example, each original signal is in the form of an L×2 array, the first column is the signal amplitude value, the second column is the corresponding sampling time, and the label is the value of the signal period T.
[0064] (13) First difference of time series: perform a difference on the sampling time in the second column of the L×2 array in step (12), and then add a 0 before the difference result sequence to ensure that the length is still L. Fill the new sequence into the second column of the original L×2 array, covering the original sampling time.
[0065] (14) Adding noise: For the original signal obtained in (13), add noise to the signal amplitude value. Its sub-steps include determining noise parameters and adding noise. The two sub-steps are as follows:
[0066] (141) Determine noise parameters: Determine the probability distribution and parameters of the noise. Common distributions include Gaussian distribution and chi-square distribution. Parameters usually include mean and standard deviation. In this example, Gaussian noise is used with a mean of 0 and a standard deviation of 0.1 to 0.4, with an interval of 0.1.
[0067] (142) Adding noise: Based on the signal length L in step (11) and the noise parameters determined in (141), a noise sequence of length L is generated and added to the signal amplitude value obtained in (13).
[0068] (15) Perform non-uniform sampling on the signal: For the signal obtained in step (14), randomly discard some sampling points to generate a non-uniform sampling sequence. The steps include determining the discard ratio, generating a discard index, and deleting the sampling points. The generated partial signal is as follows Figure 2 and Figure 3 The sub-steps of this step are as follows:
[0069] (151) Determine the discard ratio: Determine the ratio R of the total length of the sampling points to be discarded. In this example, it is fixed at 30%.
[0070] (152) Generate a discard index: According to the discard ratio determined in (151), determine the discard number Ld = round (L × R). Randomly generate Ld non-repeating positive integer sequences Idx in the interval [1, L].
[0071] (153) Deleting sampling points: According to the positive integer sequence Idx obtained in (152), delete the row corresponding to Idx from the signal of size L×2 obtained in step (14) to obtain an array of size (L-Ld)×2.
[0072] (20)Construct a neural network model: Determine the input length, model type, and model parameters of the network model. Figure 4 It is a schematic diagram of the constructed network model. The sub-steps include the following steps:
[0073] (21)Determine the model input length: Determine the length Lm when the signal sequence is input into the model. This length should be greater than the product of the maximum period T selected in (11) and the maximum original sampling rate Fs, that is, Lm > max(T) × max(Fs), and at the same time less than the signal length, that is, Lm < L. In this example, Lm is taken as 400.
[0074] (22)Select the model type: Select the model type of the neural network. Commonly used model types include the recurrent neural network RNN and the convolutional neural network CNN, etc. In this example, the CNN model is adopted.
[0075] (23)Set the model parameters: Determine the corresponding parameters of the model selected in (22), including the number of network layers, the number of nodes, etc. In this example, it is determined to be a convolutional layer and a fully connected layer. The convolutional layer contains 64 one-dimensional convolutional kernels, and the size of a single convolutional kernel is 9×2. The fully connected layer contains 128 neuron nodes.
[0076] (30)Train the neural network model: Use the data generated in (10) to train the model constructed in (20). The change of loss during the process is as Figure 5 . This step includes the following steps:
[0077] (31)Slide the window to intercept training samples: According to the model input length Lm determined in (21), slide the window to intercept the signal generated in (10). The window length is Lm, and the sliding window step size is selected as 1.
[0078] (32)Normalize the signal amplitude value: For each signal segment with a length of Lm intercepted by sliding the window in step (31), perform amplitude value normalization. The normalization interval is taken as the amplitude range A determined in step (11). In this example, it is -1 to 1.
[0079] (33)Divide the training set and the test set: Determine the division ratio, and divide the training samples generated in (32) into the training set and the test set according to the ratio. In this example, 70% is selected as the training set and 30% as the test set.
[0080] (34)Set the training parameters: Set training parameters such as the number of training times, the loss function, and the early termination condition. Since the period prediction in the task of the present invention belongs to a regression problem, the loss function is usually selected as the mean absolute error function MAE, the mean squared error function MSE, etc. In this example, the number of training times is selected as 300 times, and the loss function is selected as the mean absolute error function MAE. The expression is as shown in the following formula, where y i is the model prediction value. is the label value, N is the number of samples. The early termination condition is set to MAE does not decrease for 10 consecutive times.
[0081]
[0082] (35) Training model: According to the training parameters set in (34), the model generated in step (20) is trained using the training set and test set generated in (33). A trained model is obtained.
[0083] (40) Preprocess the obtained signal. Use the trained model obtained in (30) to extract the period of the real signal to be measured. The following steps are included:
[0084] (41) Obtaining the real signal to be tested: Record the signal collected by the sensor to obtain a sequence of length Ls, including the signal amplitude value and sampling time, where Ls ≥ Lm.
[0085] (42) Sliding window signal extraction: Sliding window is performed on the sequence obtained in (41), and the window length is taken as the length Lm determined in step (21). The sliding window step is taken as 1. By sliding the window, Ls-Lm+1 signal segments with a length of Lm can be obtained.
[0086] (43) Signal amplitude normalization: The signal segment of length Lm obtained by sliding window cutting in step (42) is normalized in amplitude. The normalization interval is taken as the amplitude range A determined in step (11). In this example, it is -1 to 1.
[0087] (44) Time series first difference: For the signal segment of length Lm obtained by sliding window in step (42), first difference is performed in sampling time. The length of the sequence obtained after first difference of Lm sampling times is Lm-1. A zero is added before the sequence to ensure that the length is still Lm.
[0088] (45) Sequence merging: Combine the amplitude value sequence of the signal segment with a length of Lm obtained in step (43) and step (44) with the time difference sequence with a length of Lm to obtain an array with a size of Lm×2.
[0089] (50) Signal period extraction: input the signal preprocessing result obtained in step (40) into the model obtained in step (30), and output the signal period value. The steps include:
[0090] (51) Input the sequence into the model: Input the result generated in step (40) into the trained model obtained in step (30), input each signal segment once, for a total of Ls-Lm+1 inputs.
[0091] (52) Extracting the period value: Check the output value of the model. There are Ls-Lm+1 outputs in total. Average all the output values and the value obtained is the period value of the signal. Figure 6 This is the time series diagram of the measured radar RCS of a certain target. Figure 7 This is the spectrum diagram of the RCS sequence analyzed by the DFT method. It can be seen that the period is 2.32 seconds (1 / 0.43 Hz). The predicted period output by the method in this embodiment is 2.45 seconds, which is in good agreement.
[0092] Table 1 Comparison of the running time of the method of this embodiment and the traditional DFT method
[0093]
[0094] Figure 8 Table 1 is a time-consuming comparison between the method of this embodiment and the traditional DFT method. Figure 8 As can be seen from Table 1, the computation time of the method in this embodiment is greatly reduced.
[0095] The present invention also provides a fast non-uniform sampling signal period extraction device, including a simulated periodic signal and label generation module, a neural network model construction module, a neural network model training module, a signal preprocessing module and a signal period extraction module; the simulated periodic signal and label generation module, the neural network model construction module, the neural network model training module, the signal preprocessing module and the signal period extraction module perform fast non-uniform sampling signal period extraction according to the steps of any of the above methods.
[0096] The present invention uses a neural network for the first time to extract the period of non-uniformly sampled signals, overcoming the problem of failure of methods such as fast Fourier transform under non-uniform sampling. Moreover, compared with the traditional discrete Fourier transform method, the method provided by the present invention reduces the amount of computation from O(N2) to O(N), and the computational time is significantly reduced.
[0097] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. A fast method for extracting the period of a non-uniformly sampled signal, characterized in that: The following steps are involved: (10) Generate simulated periodic signals and labels; (20) Constructing a neural network model; (30) Training a neural network model; (40) preprocessing the obtained signal; (50) Signal period extraction.
2. The fast non-uniform sampling signal period extraction method according to claim 1, characterized in that: The step (10) specifically comprises the following steps: (11) Determine signal parameters: According to the pattern of the real signal to be measured, determine the waveform W, period T, amplitude range A, original sampling rate Fs, and signal length L of the simulated signal; the period T and the original sampling rate Fs are selected according to the situation of the real signal; the signal length L contains more than one signal period, that is, L>T×Fs; (12) Generate original signal and label: Generate original signal and label according to the signal parameters determined in (11); the original signal includes signal amplitude value and sampling time, and the label is the period T when the signal is generated, wherein each original signal is in the form of an L×2 array, the first column is the signal amplitude value, the second column is the corresponding sampling time, and the label is the value of the signal period T; (13) First difference of time series: perform a difference on the sampling time in the second column of the L×2 array in step (12), then add 0 before the difference result sequence to ensure that the length is still L, and fill the new sequence into the second column of the original L×2 array, covering the original sampling time; (14) Adding noise: adding noise to the signal amplitude value of the original signal obtained in (13); (15) Perform non-uniform sampling on the signal: For the signal obtained in step (14), some sampling points are randomly discarded to generate a non-uniform sampling sequence.
3. The fast non-uniform sampling signal period extraction method according to claim 1, characterized in that: The step (20) of constructing a neural network model comprises: (21) Determine the model input length: Determine the length Lm of the signal sequence when it is input into the model. This length should be greater than the product of the maximum value of the period T selected in (11) and the maximum value of the original sampling rate Fs, that is, Lm>max(T)×max(Fs), and less than the signal length, that is, Lm <L; (22) Selecting a model type: selecting a neural network model type, wherein the neural network model types include a recurrent neural network (RNN) and a convolutional neural network (CNN); (23) Setting model parameters: Determine the corresponding parameters of the model selected in (22), including the number of network layers and nodes.
4. The fast non-uniform sampling signal period extraction method according to claim 1, characterized in that: The (30) step of training the neural network model comprises: (31) Sliding window interception training samples: According to the model input length Lm determined in (21), the signal generated by (10) is subjected to sliding window interception, the window length is Lm, and the sliding window step size is selected as 1; (32) Normalizing the signal amplitude value: For each signal segment of length Lm obtained by the sliding window in step (31), the amplitude value is normalized; the normalization interval is taken as the amplitude range A determined in step (11); (33) Dividing the training set and the test set: Determine the division ratio, and divide the training samples generated in (32) into the training set and the test set according to the ratio; (34) Setting training parameters: The training parameters include the number of training times, loss function, and early termination conditions; the loss function of the neural network model includes the mean absolute difference function MAE and the mean square error function MSE; (35) Training model: According to the training parameters set in (34), the model generated in step (20) is trained using the training set and test set generated in (33) to obtain a trained model.
5. The fast non-uniform sampling signal period extraction method according to claim 1, characterized in that: The step (40) of preprocessing the obtained signal comprises: (41) Obtaining the real signal to be measured: Recording the signal collected by the sensor to obtain a sequence of length Ls, including the signal amplitude value and sampling time, where Ls ≥ Lm; (42) Sliding window signal extraction: Sliding window is performed on the sequence obtained in step (41), the window length is set to be the length Lm determined in step (21), the sliding window step length is set to be 1, and Ls-Lm+1 signal segments of length Lm are obtained by sliding window; (43) Normalizing the signal amplitude value: normalizing the amplitude value of the signal segment with a length of Lm obtained by the sliding window in step (42); the normalization interval is taken as the amplitude range A determined in step (11); (44) Time series first difference: For the signal segment of length Lm obtained by sliding window in step (42), first difference is performed in sampling time. The length of the sequence obtained after first difference of Lm sampling times is Lm-1. A zero is added before the sequence to ensure that the length is still Lm. (45) Sequence merging: Combine the amplitude value sequence of the signal segment with a length of Lm obtained in step (43) and step (44) with the time difference sequence with a length of Lm to obtain an array with a size of Lm×2.
6. A fast method for extracting period of non-uniform sampling signal as claimed in claim 1, characterized in that: The (50) signal period extraction step comprises: (51) Inputting the sequence into the model: Input the result generated in step (40) into the trained model obtained in step (30), inputting each signal segment once, for a total of Ls-Lm+1 inputs; (52) Extracting the periodic value: Check the output value of the model. There are Ls-Lm+1 outputs in total. Average all the output values and the value obtained is the periodic value of the signal.
7. The fast non-uniform sampling signal period extraction method as claimed in claim 2, characterized in that: Step (14) specifically comprises the following steps: (141) Determining noise parameters: determining the probability distribution and parameters of the noise; the probability distribution includes Gaussian distribution and chi-square distribution, and the parameters include mean and standard deviation; (142) Adding noise: Based on the signal length L in step (12) and the noise parameters determined in (141), a noise sequence of length L is generated and added to the signal amplitude value obtained in (13).
8. A fast method for extracting period of non-uniform sampling signal as claimed in claim 2, characterized in that: Step (15) specifically comprises the following steps: (151) Determine the discard ratio: determine the ratio of the total length of the sampling points to be discarded as R; (152) Generate a discard index: According to the discard ratio determined in (151), determine the discard number Ld=round(L×R) and randomly generate Ld non-repeating positive integer sequences Idx in the interval [1, L]; (153) Deleting sampling points: According to the positive integer sequence Idx obtained in (152), delete the row corresponding to Idx from the signal of size L×2 obtained in step (14) to obtain an array of size (L-Ld)×2.
9. The fast non-uniform sampling signal period extraction method according to claim 4, characterized in that: The loss function in step (34) is the mean absolute difference function MAE, which is expressed as follows: where y i is the model prediction value, is the label value, N is the number of samples, and the early termination condition is set as MAE does not decrease for 10 consecutive times.
10. A fast non-uniform sampling signal period extraction device, characterized in that: It includes a simulated periodic signal and label generation module, a neural network model construction module, a neural network model training module, a signal preprocessing module and a signal period extraction module; the simulated periodic signal and label generation module, the neural network model construction module, the neural network model training module, the signal preprocessing module and the signal period extraction module perform fast extraction of non-uniform sampling signal periods according to the steps of the method described in any one of claims 1 to 9.
Citation Information
Cited By
Voiceprint detection method for internal abnormity of wind driven generator blade
CN120220697A
