A state monitoring method based on signal intelligent noise reduction algorithm

Abstract

A state monitoring method based on a signal intelligent denoising algorithm comprises: dividing the collected signal samples under different states into a training set and a test set; decomposing the training set signal into multiple layers of wavelet coefficients by using FDWT; extracting the maximum value and average value of each layer of wavelet coefficients as signal features and evaluating their effectiveness for state recognition; recording the decomposition layer numbers corresponding to the signal features with high effectiveness; establishing a signal intelligent denoising model and training the model by using the training set signal, wherein the model decomposes the signal by using FDWT for the decomposition layer corresponding to the signal features with high effectiveness and decomposes the signal by using the intelligent denoising algorithm for the remaining decomposition layers; decomposing the training set and test set signals into multiple layers of wavelet coefficients after denoising by using the trained intelligent denoising model and extracting the maximum value and average value of each layer of wavelet coefficients as signal features; training a classification model according to the features of the training set signal and testing the state recognition effect by using the features of the test set signal.

Description

Technical Field

[0001] The present invention relates to the technical field of signal processing and machine learning, and more specifically, to a state monitoring method based on a signal intelligent noise reduction algorithm. Background Art

[0002] Condition monitoring is widely used in various fields such as civil engineering, mechanical engineering, electrical engineering, and vehicle engineering. Practitioners in these industries often need to monitor the condition of infrastructure or industrial equipment based on signals collected by sensors in order to identify abnormal conditions in a timely manner and avoid accidents.

[0003] Commonly used condition monitoring methods first transform the signals collected by the sensor into the time-frequency domain or wavelet domain using methods such as short-time Fourier transform (STFT), Mel spectrogram, wavelet packet decomposition (WPT) or FDWT, then perform feature extraction, train a classification model based on the extracted signal features, and use the trained model for condition recognition. These condition monitoring methods use traditional algorithms to transform the signals, and the extracted signal features are sensitive to unexpected noise or changing working conditions, which reduces the accuracy of condition recognition.

[0004] In order to solve the above problems, some scholars at home and abroad use convolutional neural networks to learn convolution kernels similar to Fourier transform or wavelet transform, so as to transform the signal into the time-frequency domain or wavelet domain under supervised machine learning and give the signal noise reduction effect. However, these methods not only require manual labeling due to the use of supervised machine learning, but also have high requirements for computer configuration. Later, signal transformation algorithms based on unsupervised machine learning were gradually developed, eliminating human intervention in such learning tasks and alleviating the high demand for computer configuration. However, when the number of signals or the length of the signal increases, the model training time of unsupervised machine learning will become unacceptable. In addition, if the number of training set samples is small, the risk of overfitting in unsupervised learning will also increase greatly.

[0005] Combining traditional signal transformation algorithms with unsupervised machine learning algorithms can perform targeted noise reduction and decomposition of signal components, reduce computing costs and the risk of model overfitting, and provide new ideas for state recognition tasks under noise interference. Summary of the invention

[0006] The purpose of the present invention is to provide a state monitoring method based on a signal intelligent noise reduction algorithm to reduce the calculation cost and improve the accuracy of state recognition.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] A state monitoring method based on a signal intelligent noise reduction algorithm comprises the following steps:

[0009] S1, dividing the collected signal samples in different states into training sets and test sets;

[0010] S2, using fast discrete wavelet transform (FDWT) to decompose the training set signal into multiple layers of wavelet coefficients;

[0011] S3, extract the maximum and average values ​​of the wavelet coefficients of each layer as signal features and evaluate their effectiveness for state recognition;

[0012] S4, record the decomposition layer number corresponding to the signal feature with high validity;

[0013] S5. Establish a signal intelligent denoising model and use the training set signal to train the model. The model uses FDWT to decompose the signal for the decomposition layer corresponding to the signal features with high effectiveness, and uses the intelligent denoising algorithm to decompose the signal for the other decomposition layers;

[0014] S6. Decomposing the training set and test set signals into multiple layers of denoised wavelet coefficients using the trained intelligent denoising model and extracting the maximum and average values ​​of each layer of wavelet coefficients as signal features;

[0015] S7. Train the classification model according to the characteristics of the training set signal, and use the characteristics of the test set signal to test the state recognition effect.

[0016] Preferably, in step S1, the ratio of the number of samples in the training set to that in the test set is between 2:1 and 5:1.

[0017] Preferably, the number of decomposition levels of the multi-layer wavelet coefficients in step S2 is determined as a value L rounded down from the result of log2N calculation according to the number of signal sampling points N.

[0018] Preferably, the Fisher Score indicator is used to evaluate the effectiveness of the signal feature for state recognition in step S3, and the indicator is expressed as:

[0019]

[0020] Among them, Q i is the Fisher Score value of the i-th signal feature, n j is the number of samples in the jth state category, μ ij and σ ij is the mean and standard deviation of the i-th signal feature in the j-th state category, μ i is the mean of the i-th signal feature under all state categories, and J is the total number of state categories.

[0021] Preferably, the signal feature with high validity in step S4 refers to a feature whose corresponding Fisher Score value exceeds the average value of the Fisher Score values ​​of all features.

[0022] Preferably, the establishment and training of the intelligent noise reduction model in step S5 includes the following sub-steps:

[0023] S5.1. Build an L-layer signal decomposition framework that is consistent with the decomposition process of the FDWT algorithm, that is, a signal decomposition framework that can continuously perform L high-pass and low-pass filtering on the input signal to obtain wavelet coefficients; except for the first high-pass and low-pass filtering on the original signal, the remaining (L-1) high-pass and low-pass filtering are all for the signal after the previous low-pass filtering;

[0024] S5.2. In the decomposition layer corresponding to the signal features with high validity, a filter coefficient vector consistent with the wavelet basis function in the FDWT algorithm is used as a convolution operation factor for signal decomposition;

[0025] S5.3. A neural network model is built in each layer of the remaining decomposition layers to learn the filter coefficient vector and perform signal noise reduction. The activation function of the neural network model adopts a hard threshold activation function with noise reduction function, which is expressed as:

[0026]

[0027] Among them, b1 and b2 represent the learnable activation function bias.

[0028] S5.4. Use the training set signal to train the intelligent noise reduction model corresponding to the entire signal decomposition framework. The objective function of the model training is expressed as:

[0029]

[0030] Where s represents the original signal, It represents the reconstructed signal obtained by the step-by-step inverse operation after the original signal is decomposed by this decomposition framework, C L Represents the set of wavelet coefficients obtained after the original signal is decomposed by L layers.

[0031] Preferably, the number of decomposition layers corresponding to the multi-layer wavelet coefficients in step S6 is consistent with the number of decomposition layers in step S2.

[0032] Preferably, the classification model is trained using a support vector machine in step S7.

[0033] The beneficial effects of the present invention are as follows:

[0034] The present invention can carry out targeted noise reduction on signal components and improve the accuracy of state recognition; the intelligent noise reduction model has a simple architecture, and the entire model training process does not require manual intervention and is adaptive;

[0035] The technical solution of the present invention combines the traditional FDWT algorithm and the machine learning algorithm, has low overfitting risk, low computational cost, and broad application scenarios.

[0036] The technical solution of the present invention is adaptive, has broad application scenarios, and can improve the accuracy of state recognition under noise interference. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A framework flow chart showing a state monitoring method based on a signal intelligent noise reduction algorithm provided by the present invention

[0038] Figure 2 A framework diagram showing a signal intelligent noise reduction model provided by the present invention DETAILED DESCRIPTION

[0039] In order to make the technical solution of the present invention easier to understand, the specific implementation of the present invention will be described more clearly below in conjunction with embodiments.

[0040] Example

[0041] like Figure 1 The urban rail transit slab track state monitoring method based on the sound signal intelligent noise reduction algorithm shown includes the following steps:

[0042] S1, dividing the collected sound signal samples of the slab track in different states into a training set and a test set;

[0043] S2, using FDWT to decompose the training set signal into multiple layers of wavelet coefficients;

[0044] S3, extract the maximum and average values ​​of the wavelet coefficients of each layer as signal features and evaluate their effectiveness for state recognition;

[0045] S4, record the decomposition layer number corresponding to the signal feature with high validity;

[0046] S5. Establish a signal intelligent denoising model and use the training set signal to train the model. The model uses FDWT to decompose the signal for the decomposition layer corresponding to the signal features with high effectiveness, and uses the intelligent denoising algorithm to decompose the signal for the other decomposition layers;

[0047] S6. Decomposing the training set and test set signals into multiple layers of denoised wavelet coefficients using the trained intelligent denoising model and extracting the maximum and average values ​​of each layer of wavelet coefficients as signal features;

[0048] S7. Train the classification model according to the features of the training set signal, and use the features of the test set signal to test the state recognition effect of the slab track.

[0049] Specifically, the sound signal samples collected in step S1 are collected by using acoustic sensors set beside the slab track in three different states (i.e., normal state, moderately damaged state, and severely damaged state) to collect the sound signals generated when the train passes by, and the number of sampling points for each signal segment is 512415. 144 sound signal samples (a total of 432 samples) are collected in each slab track state, and they are divided into a training set and a test set at a ratio of 2:1. Finally, the training set and the test set contain 288 and 144 signal samples, respectively;

[0050] Specifically, in step S2, since the number of sampling points of each signal segment is 512415, the number of signal decomposition layers is determined to be 18 according to the rounded value 18 of the log2512415 calculation result, and the sound signal of the training set is decomposed into 18 layers of wavelet coefficients using FDWT;

[0051] Specifically, the effectiveness of the signal feature for state recognition in step S3 is evaluated using the Fisher Score indicator, which is expressed as:

[0052]

[0053] Among them, Q i is the Fisher Score value of the i-th signal feature, n j is the number of samples in the jth state category, μ ij and σ ij is the mean and standard deviation of the i-th signal feature in the j-th state category, μ i is the mean of the i-th signal feature under all state categories, and J is the total number of state categories.

[0054] The Fisher Score values ​​of the two signal features corresponding to each decomposition layer are calculated and shown in Tables 1 and 2.

[0055] Table 1 Fisher Score values ​​of signal features (1-9 decomposition levels)

[0056]

[0057] Table 2 Fisher Score values ​​of signal features (10-18 decomposition level)

[0058]

[0059] Specifically, the decomposition layer numbers corresponding to the signal features with high validity in step S4 are selected according to the results of Table 1 and Table 2. The average Fisher Score value of all features in Table 1 and Table 2 is 0.136, and the features with Fisher Score values ​​greater than 0.136 include the maximum value features of the 3rd, 4th and 5th layers, and the maximum value features and average value features of the 7th and 8th layers. Therefore, the decomposition layer numbers corresponding to the signal features with high validity are determined to be 3, 4, 5, 7, 8;

[0060] Specifically, the framework of the intelligent noise reduction model in step S5 is as follows: Figure 2 As shown, the establishment and training of the model includes the following sub-steps:

[0061] S5.1. Build an 18-layer signal decomposition framework that is consistent with the decomposition process of the FDWT algorithm, that is, a signal decomposition framework that can continuously perform 18 high-pass and low-pass filtering on the input signal to obtain wavelet coefficients; except for the first high-pass and low-pass filtering on the original signal, the remaining 17 high-pass and low-pass filtering are all for the signal after the previous low-pass filtering;

[0062] S5.2. In the decomposition layers corresponding to the signal features with high validity, i.e., layers 3, 4, 5, 7, and 8, the filter coefficient vector [-0.0106, 0.0329, 0.0308, -0.1870, -0.0280, 0.6309, 0.7148, 0.2304] consistent with the function of the 'db4' wavelet basis function in the FDWT algorithm is used as the convolution operation factor for signal decomposition;

[0063] S5.3. A neural network model is built in each layer of the remaining decomposition layers to learn the filter coefficient vector and perform signal noise reduction. The activation function of the neural network model adopts a hard threshold activation function with noise reduction function, which is expressed as:

[0064]

[0065] Among them, b1 and b2 represent the learnable activation function bias.

[0066] S5.4. Use the training set signal to train the intelligent noise reduction model corresponding to the entire signal decomposition framework. The objective function of the model training is expressed as:

[0067]

[0068] Among them, s represents the original signal, It represents the reconstructed signal obtained by the step-by-step inverse operation after the original signal is decomposed by this decomposition framework, C L Represents the set of wavelet coefficients obtained after the original signal is decomposed by L layers.

[0069] Specifically, the number of decomposition layers corresponding to the multi-layer wavelet coefficients in step S6 is consistent with the number of decomposition layers in step S2, that is, 18 layers.

[0070] Specifically, the classification model in step S7 is trained using a support vector machine.

[0071] The state recognition accuracy of the urban rail transit slab track in this embodiment is shown in Table 3 by adopting the technical solution of the present invention and the existing common technical solution (i.e., transforming the signal into the time-frequency domain or wavelet domain through STFT, Melspectrogram, WPT or FDWT, and then performing feature extraction and state recognition). It can be seen that the state recognition accuracy of the state monitoring method based on the signal intelligent noise reduction algorithm provided by the present invention is significantly improved compared with the existing common method.

[0072] Table 3 Comparison of the state recognition accuracy of the state monitoring method provided by the present invention and the existing common state monitoring method

[0073]

[0074]

[0075] The above embodiments of the present invention are only preferred embodiments of the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, several changes or modifications may be made without departing from the spirit and scope of the present invention. It is impossible to list all the embodiments here, and any obvious changes or modifications made on the basis of the technical solution of the present invention are still within the scope of protection of the present invention.

Claims

1. A state monitoring method based on a signal intelligent noise reduction algorithm comprises the following steps: S1, dividing the collected signal samples in different states into training sets and test sets; S2, using fast discrete wavelet transform FDWT to decompose the training set signal into multiple layers of wavelet coefficients; S3, extract the maximum and average values ​​of the wavelet coefficients of each layer as signal features and evaluate their effectiveness for state recognition; S4, record the decomposition layer number corresponding to the signal feature with high validity; S5. Establish a signal intelligent denoising model and use the training set signal to train the model. The model uses FDWT to decompose the signal for the decomposition layer corresponding to the signal features with high effectiveness, and uses the intelligent denoising algorithm to decompose the signal for the other decomposition layers; S6. Decomposing the training set and test set signals into multiple layers of denoised wavelet coefficients using the trained intelligent denoising model and extracting the maximum and average values ​​of each layer of wavelet coefficients as signal features; S7, training the classification model according to the characteristics of the training set signal, and testing the state recognition effect using the characteristics of the test set signal; In step S3, the effectiveness of the signal feature for state recognition is evaluated using the Fisher Score indicator, which is expressed as: (1) in, It is i Fisher Score value of signal features, It is j The number of samples in each state category, and It is i The signal characteristics are j The mean and standard deviation of the status categories, It is i The mean of the signal features under all state categories, J is the total number of status categories; The establishment and training of the intelligent noise reduction model in step S5 includes the following sub-steps: S5.

1. Build a process consistent with the FDWT algorithm decomposition process L Layer signal decomposition framework, that is, it can continuously perform L The signal decomposition framework of wavelet coefficients is obtained by high-pass and low-pass filtering. In addition to the first high-pass and low-pass filtering for the original signal, the remaining ( L -1) Both high-pass and low-pass filtering are for the signal after the previous low-pass filtering; S5.

2. In the decomposition layer corresponding to the signal features with high validity, a filter coefficient vector consistent with the wavelet basis function in the FDWT algorithm is used as a convolution operation factor for signal decomposition; S5.

3. A neural network model is built in each layer of the remaining decomposition layers to learn the filter coefficient vector and perform signal noise reduction. The activation function of the neural network model adopts a hard threshold activation function with noise reduction function, which is expressed as: (2) in, and represents the learnable activation function bias; S5.

4. Use the training set signal to train the intelligent noise reduction model corresponding to the entire signal decomposition framework. The objective function of the model training is expressed as: (3) in, represents the original signal, It means the reconstructed signal obtained by the step-by-step inverse operation after the original signal is decomposed by this decomposition framework. Represents the original signal L The wavelet coefficient set obtained after layer decomposition.

2. A state monitoring method based on a signal intelligent noise reduction algorithm as claimed in claim 1, characterized in that: The ratio of the number of samples in the training set to the number of samples in the test set in step S1 is between 2:1 and 5:

1.

3. A state monitoring method based on a signal intelligent noise reduction algorithm as claimed in claim 1, characterized in that: The number of decomposition layers of the multi-layer wavelet coefficients in step S2 is based on the number of signal sampling points. N Determined as The floor value of the calculation result L .

4. A state monitoring method based on a signal intelligent noise reduction algorithm as claimed in claim 1, characterized in that: The signal feature with high validity in step S4 refers to a feature whose corresponding Fisher Score value exceeds the average value of the Fisher Score values ​​of all features.

5. The state monitoring method based on the signal intelligent noise reduction algorithm according to claim 1, characterized in that: The number of decomposition layers corresponding to the multi-layer wavelet coefficients in step S6 is consistent with the number of decomposition layers in step S2.

Images