ABR signal extraction method based on wavelet analysis in combination with LM-RBF neural network

Abstract

The invention belongs to the technical field of bio-electricity signal processing, and particularly relates to an ABR signal extraction method based on wavelet analysis in combination with an LM-RBF neural network, and the method comprises the steps: S1, obtaining a noise-containing ABR signal through a hearing test detection system; s2, performing five-layer decomposition on the same ABR signal by using wavelet bases such as bior5.5, db4 and sym6, calculating a reconstruction error RMSE, and selecting a sym6 wavelet base most suitable for the method; s3, the selected sym6 is used for preprocessing the signals, original ABR signals are decomposed and reconstructed, useless physiological electric signals are filtered out, and the signal-to-noise ratio of the signals is increased; and S4, constructing an RBF (Radial Basis Function) neural network, and improving the RBFNN by using a Levenberg-Marquardt algorithm. According to the method, the signal-to-noise ratio of the ABR signal is improved by about 5dB, and the waveform definition is reduced from 0.45 mu V to 0.12 mu V; the recognition rate is increased from 78% to 95%; compared with a traditional superposition average method, the error is reduced by more than 30%.

Description

Technical Field

[0001] The present invention relates to the technical field of bioelectric signal processing, and in particular to an ABR signal extraction method based on wavelet analysis combined with LM-RBF neural network. Background Art

[0002] ABR is a type of electroencephalogram (EEG) evoked potential (ABR), representing transient changes in neural activity in the auditory pathway in response to certain sound stimuli. It consists of seven peaks, represented by Roman numerals: waves I, II, III, IV, V, VI, and VII. The information from these seven waves can be used to assess a subject's hearing ability. The ABR is essentially the electrical response of the auditory nerve and other related neurons to sound stimuli. Because ABR is objective and not influenced by subjective factors, and reflects brainstem activity, its detection is highly reliable. Because the amplitude of ABR is very small, ranging from approximately 0.1µV to 10µV, only 1% of spontaneous EEG, it has a low signal-to-noise ratio and is easily mixed with noise, drowning out other spontaneous EEG signals. Therefore, the extraction method of ABR is crucial.

[0003] While the widely used traditional ABR signal extraction method, the averaging and superposition method, is simple and easy to use, it takes a long time to detect ABRs, can easily cause neural fatigue in the subject, and can affect the accuracy of the test results. Therefore, various signal processing techniques are being used to extract ABRs to overcome these drawbacks, thereby improving the reliability and efficiency of ABR detection and reducing the difficulty of clinical application. To address the impact of factors such as the patient and the testing environment on experimental results in existing ABR detection, it is necessary to refine and optimize ABR signal extraction methods to effectively overcome signal noise. Wavelet analysis performs multi-resolution decomposition on the collected noisy signal, extracting useful information at the corresponding frequencies, filtering out irrelevant interference signals, and improving the negative signal-to-noise ratio of the ABR signal. RBF neural networks can improve ABR signal extraction efficiency, but their effectiveness is affected by the signal-to-noise ratio of the input signal: high SNR yields good results, while low SNR yields poor results. Summary of the Invention

[0004] (1) Technical problems solved

[0005] Traditional ABR signal extraction methods, such as the average superposition method, suffer from time-consuming detection and susceptibility to noise. This paper provides an ABR signal extraction method based on wavelet decomposition combined with an LM-RBF neural network. This method addresses the issues with traditional RBFNN extraction, which is significantly affected by the input signal-to-noise ratio, resulting in low extraction efficiency and poor accuracy. This paper proposes an algorithm fusion method that combines wavelet analysis with RBFNN optimization using the Levenberg-Marquardt algorithm, leveraging their complementary strengths to achieve rapid and accurate ABR signal extraction.

[0006] (2) Technical solution

[0007] In order to achieve the above-mentioned purpose, the present invention specifically adopts the following technical solutions:

[0008] An ABR signal extraction method based on wavelet analysis combined with LM-RBF neural network specifically includes the following steps:

[0009] S1. Acquire an ABR signal containing noise using an audiometry system;

[0010] S2. Use bior5.5, db4, sym6 and other wavelet bases to perform 5-layer decomposition on the same ABR signal, calculate the reconstruction error RMSE, and select the most suitable sym6 wavelet base of the present invention;

[0011] S3. Use the selected sym6 to preprocess the signal, decompose and reconstruct the original ABR signal, filter out useless physiological electrical signals, and improve the signal-to-noise ratio;

[0012] S4. Construct RBF neural network and improve it using Levenberg-Marquardt algorithm;

[0013] S5. The pre-processed ABR signal is used as the signal input of the optimized RBF neural network. After a series of signal processing, the final output signal signal-to-noise ratio is significantly improved and the signal extraction error is significantly reduced.

[0014] Furthermore, the hearing test detection system in S1 includes a Neuroscan SynAmps II amplifier, using a short click sound as the stimulation signal with a wave width of 0.1 ms and a stimulation intensity of 60-80 dB nHL; during signal acquisition, the stimulation is repeated 2000-5000 times, the sampling rate is ≥10 kHz, the recording time is 10 ms, and the latency of ABR peaks I-VII is covered.

[0015] Furthermore, the basis for selecting the sym6 wavelet basis in S2 is: by comparing different wavelet bases and calculating the reconstruction errors of bior5.5, db4, and sym6, the sym6 wavelet is selected due to its high symmetry and excellent frequency band separation ability.

[0016] Furthermore, the selected sym6 is used in S3 to preprocess the signal, decompose and reconstruct the original ABR signal, filter out useless physiological electrical signals, and improve the signal-to-noise ratio. The "two-decimation" is decomposed into the core operation of discrete wavelet transform, and the approximation coefficients and detail coefficients of each layer are downsampled, and the even-indexed samples are retained to achieve layer-by-layer separation of signal frequency bands; the number of decomposition layers is dynamically selected based on the energy entropy criterion, and the low-frequency approximation coefficients and the high-frequency coefficients after thresholding are retained to filter out the noise-dominated frequency bands.

[0017] Furthermore, the LM algorithm weight optimization is proposed in S4, and the objective function is to calculate the minimum mean square error (MSE); the Jacobian matrix of the error to weight in the weight update formula has a dimension of 512×22, and the damping factor is initially set to 0.01, which is adjusted according to the error change after each iteration; the termination condition is MSE < 0.001 or the number of iterations ≥ 100 times.

[0018] Furthermore, in S5, the signal reconstructed by wavelet is input into the optimized RBF neural network. In terms of network mapping, the denoised signal Y(k) is generated through hidden layer Gaussian function mapping and output layer linear weighting; the results are evaluated through waveform clarity and recognition rate as well as comparative experiments.

[0019] (3) Beneficial effects

[0020] Compared with the prior art, the present invention provides an ABR signal extraction method based on wavelet analysis combined with LM-RBF neural network, which has the following beneficial effects:

[0021] The method of the present invention improves the signal-to-noise ratio of the collected noisy ABR signal after wavelet analysis preprocessing, and enhances the anti-noise performance. This can first compensate for the disadvantage of the RBF neural network's unsatisfactory working efficiency when the signal-to-noise ratio is low; the RBF neural network algorithm optimized by the LM algorithm is used to significantly improve its convergence efficiency, so when the preprocessed ABR signal is used as the signal input of the optimized RBF neural network, a more accurate ABR signal can be obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 Flowchart of the ABR signal extraction algorithm of the present invention;

[0023] Figure 2 This is a schematic diagram of the decomposition of the "two extractions" of the present invention;

[0024] Figure 3 The schematic diagram of the RBF neural network extracting ABR signals in the present invention;

[0025] Figure 4 This is a flow chart of the ABR signal processing by the optimization algorithm of the present invention. DETAILED DESCRIPTION

[0026] The following will clearly describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.

[0027] Example

[0028] As shown in the accompanying drawings, an embodiment of the present invention proposes an ABR signal extraction method based on wavelet analysis combined with LM-RBF neural network, which includes the following steps:

[0029] S1: ABR signal acquisition system was used to extract ABR signals. A Neuroscan SynAmps II EEG amplifier was used, equipped with Ag / AgCl disk electrodes. The international 10-20 system was strictly followed, with the recording electrode placed at the center of the forehead (Fz), reference electrodes placed at the left and right earlobes (A1 and A2), and a ground electrode placed at the center of the forehead (FPz). The system primarily consisted of an electrode cap, power supply box, head box, control box, acquisition and analysis tools, and stimulation tools. The sampling rate was set to 10 kHz, and the recording duration was 10 ms, covering the latencies of ABR peaks I-VII (1.5 ms to 8 ms).

[0030] Stimulation parameters: The stimulus type used was a short sound (click), with a wave width of 0.1 ms and an intensity of 70 dB nHL (normalized hearing level). The number of repetitions was 3000, and the stimulation interval was randomized (20-30 ms) to avoid subject adaptation interference. The sampling rate was set to 10 kHz, and the signal recording time was 10 ms (covering the latency of ABR peaks I-VII). Noise model: The original signal contains electrooculogram (0.5-5 Hz), electromyography (20-500 Hz) and environmental noise (50 Hz power frequency interference), and the initial signal-to-noise ratio (SNR) value is -10 dB. Original signal storage: The original noisy signal is recorded as , is an ABR signal, The data is stored in .mat format, with 512 sampling points per channel (corresponding to 10ms duration).

[0031] S2: When selecting the wavelet basis, the sym6 wavelet was selected. By comparing the ABR signal reconstruction errors of bior5.5, db4, sym6 and other wavelets, the sym6 wavelet was selected as the decomposition basis function due to its high symmetry and excellent frequency band separation ability.

[0032] In the comparative experiment, the same ABR signal was decomposed into 5 layers using wavelet bases such as bior5.5, db4, and sym6, and the reconstruction error RMSE was calculated. The calculation formula is:

[0033] RMSE= ,

[0034] in is the i-th true value of the original ABR signal, is the i-th predicted value output after wavelet analysis, and n is the number of samples.

[0035] Below are the RMSE values ​​of the three wavelet bases:

[0036] bior5.5: RMSE = 0.32 μV

[0037] db4: RMSE = 0.28 μV

[0038] sym6: RMSE = 0.15 μV (selected)

[0039] The experimental data obtained through calculation show that the sym6 wavelet has high symmetry, can effectively reduce phase distortion, has excellent frequency band separation ability, and has a significant effect on suppressing high-frequency noise. Therefore, sym6 is selected as the wavelet basis for processing ABR signals in the present invention.

[0040] S3: Select the sym6 wavelet basis to preprocess the signal, decompose and reconstruct the original ABR signal, so as to obtain a pure ABR signal with high signal-to-noise ratio and retained effective features.

[0041] About signal decomposition: In wavelet analysis, "layer" can be used to represent the number of times the signal is decomposed, and the signal can be divided into multiple levels according to the frequency band. In this design, the signal sampling frequency is set to 100KHz, and the center frequency is used as the dividing point for "two extractions" to divide the signal into high-frequency and low-frequency parts. Because the low-frequency ABR signal is to be extracted, it is only necessary to use "two extractions" to decompose the low-frequency signal part during the wavelet analysis process. After each decomposition, the existing frequency range band will become half of the previous frequency range, and it will be repeated until the desired signal frequency is decomposed. The process of "two extractions" to decompose the signal is as follows. Figure 2 .

[0042] The original noisy signal is denoted as , is the useful ABR signal, that is, the signal to be extracted, is the noise, i=1, 2, ..., n. When performing wavelet decomposition, we can use and At different scales, retain The details or approximation signals of the scale are set to zero, and then the wavelet reconstruction is performed to remove the noise. purpose.

[0043] Regarding wavelet reconstruction: In wavelet analysis, the original noisy signal is decomposed using wavelets to obtain low-frequency signal components that eliminate interfering signals. However, for the filtered signal, we usually need to restore it based on the wavelet decomposition factors. This restoration method is called wavelet reconstruction. The low-frequency and high-frequency signals after "double decimation" are combined into the previous level signal, and this combination is repeated multiple times until a signal with the same length as the original signal is obtained. As a result, the reconstructed signal is a pure signal with the noise components removed while retaining the effective features. The experimental results show that the signal-to-noise ratio is improved to -5dB, an improvement of 5dB.

[0044] S4: Build RBF neural network and use LM algorithm to improve RBFNN:

[0045] About constructing RBF neural network: Using Gaussian basis function as the basis function of RBF neural network, assuming that there are M neuron nodes in the hidden layer, the output formula of the j-th neuron node is:

[0046] ,in is the Euclidean norm, which is the Euclidean distance between the input sample and the central node. The size of this distance is inversely proportional to the output response of the hidden layer; is the basis function center, is the basis function width, these two parameters are needed in the improvement optimization process; is the decay rate of the Gaussian kernel function of the jth node. The smaller the value, the narrower the width of the function and the faster the decay rate of the function.

[0047] There is one neuron node in the output layer, and the final output expression of the k-th neuron node is:

[0048] , where: X is the input vector; k is the kth node of the output layer; is the weight connecting the jth hidden layer neuron and the kth output neuron; is the threshold of the output node.

[0049] When extracting ABR signals, the noisy ABR signal is used as the ABR signal to be extracted. ( , is the ABR signal to be extracted, is the noise component) is input into the network from the input layer and the output is obtained after calculation by the hidden layer By adjusting the weight vector to make the network close to the ideal ABR signal, the approximation error between the actual output and the noisy signal is calculated. ,when When it is minimum, it is considered that an ABR signal is obtained.

[0050] The following describes improvements to the LM algorithm for RBF neural networks: The core of the LM algorithm is to iteratively update parameters, using approximations of the Jacobian and Hessian matrices to adjust the step size. In each iteration, the error in the current parameters is calculated, and then the parameters are adjusted according to the LM algorithm's update rule. The construction of the Jacobian matrix and how to incorporate the damping factor to balance the characteristics of gradient descent and the Gauss-Newton method require detailed explanation.

[0051] The first step is to initialize the network parameters. In the basic RBF neural network formula, the center of the radial basis function and width As one of the components of the RBF neural network, their selection will also affect the extraction effect of the RBF neural network. The following is the specific formula:

[0052] =1+(j-1)× ,j=1,2,...,22

[0053]

[0054] Then the network output and error are calculated, that is, the mean square error (MSE) is minimized:

[0055] ,in is the input signal, is the reference signal after wavelet preprocessing.

[0056] Then calculate the weight update formula:

[0057] ,in is the error vector, through Calculation; J is the Jacobian matrix of error versus weight (dimension 512×22); is the damping factor, with an initial value of 0.01, which is adjusted according to the error after each iteration:

[0058]

[0059] The optimization algorithm terminates when MSE < 0.001 or when the number of iterations is ≥ 100 (typically convergence requires 40-60 iterations). When the algorithm terminates, the output is considered the desired ABR signal.

[0060] The LM algorithm is a hybrid optimization method that combines the gradient descent method: strong robustness, but slow convergence speed; Gauss-Newton method: fast convergence speed, but sensitive to initial values. The core of the algorithm is to use the damping factor Dynamically adjust the weights of the two methods. When it is large, it degenerates into gradient descent to ensure stability; when When it is small, it is close to the Gauss-Newton method, which accelerates convergence and significantly improves the optimization efficiency of the RBF neural network.

[0061] S5: Input Preprocessed Signal: The wavelet-reconstructed signal is fed into the optimized RBF neural network. In terms of network mapping, the denoised signal Y(k) is generated through hidden layer Gaussian function mapping and output layer linear weighting. In terms of result evaluation, waveform clarity, as measured by the standard deviation of the amplitudes of peaks I, II, III, IV, V, VI, and VII, was reduced from 0.45 μV with the traditional method to 0.12 μV. The recognition rate, based on the peak detection algorithm (threshold + slope method), increased from 78% with the traditional method to 95%. In a comparative experiment, the traditional superposition averaging method required 5,000 superpositions, took 5 minutes, and resulted in a final error of 0.35 μV. However, this method only required 3,000 stimulations, took 3 minutes, and achieved an error of 0.02 μV (a reduction of over 30%).

[0062] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for extracting ABR signals based on wavelet analysis combined with LM-RBF neural network, comprising the following steps: S1. Acquire an ABR signal containing noise using an audiometry system; S2. Use bior5.5, db4, sym6 and other wavelet bases to perform 5-layer decomposition on the same ABR signal, calculate the reconstruction error RMSE, and select the most suitable sym6 wavelet base of the present invention; S3. Use the selected sym6 to preprocess the signal, decompose and reconstruct the original ABR signal, filter out useless physiological electrical signals, and improve the signal-to-noise ratio; S4. Construct RBF neural network and improve it using Levenberg-Marquardt algorithm; S5. The pre-processed ABR signal is used as the signal input of the optimized RBF neural network. After a series of signal processing, the final output signal signal-to-noise ratio is significantly improved and the signal extraction error is significantly reduced.

2. The ABR signal extraction method based on wavelet analysis combined with LM-RBF neural network according to claim 1, characterized in that: The hearing test detection system in S1 includes a Neuroscan SynAmps II amplifier, using short click sounds as stimulation signals with a wave width of 0.1 ms and a stimulation intensity of 60-80 dB nHL. During signal acquisition, the stimulation is repeated 2000-5000 times, with a sampling rate of ≥10 kHz and a recording duration of 10 ms, covering the latency of ABR peaks I-VII.

3. The ABR signal extraction method based on wavelet analysis combined with LM-RBF neural network according to claim 1, characterized in that: The basis for selecting the sym6 wavelet basis in S2 is: by comparing different wavelet bases and calculating the reconstruction errors of bior5.5, db4, and sym6, the sym6 wavelet is selected due to its high symmetry and excellent frequency band separation ability.

4. The ABR signal extraction method based on wavelet analysis combined with LM-RBF neural network according to claim 1, characterized in that: In S3, the selected sym6 is used to preprocess the signal, decompose and reconstruct the original ABR signal, filter out useless physiological electrical signals, and improve the signal-to-noise ratio. "Two-decimation" is decomposed into the core operation of discrete wavelet transform, downsampling the approximation coefficients and detail coefficients of each layer, retaining even-indexed samples, and realizing layer-by-layer separation of signal frequency bands; the number of decomposition layers is dynamically selected based on the energy entropy criterion, retaining low-frequency approximation coefficients and high-frequency coefficients after thresholding, and filtering out the noise-dominated frequency band.

5. The ABR signal extraction method based on wavelet analysis combined with LM-RBF neural network according to claim 1, characterized in that: The LM algorithm weight optimization is proposed in S4, and the objective function is to minimize the mean square error (MSE). The Jacobian matrix of the error to weight in the weight update formula has a dimension of 512×22, and the damping factor is initially set to 0.01 and adjusted according to the error change after each iteration. The termination condition is MSE < 0.001 or the number of iterations ≥ 100.

6. The ABR signal extraction method based on wavelet analysis combined with LM-RBF neural network according to claim 1, characterized in that: In S5, the signal reconstructed by wavelet is input into the optimized RBF neural network. In terms of network mapping, the denoised signal Y(k) is generated through hidden layer Gaussian function mapping and output layer linear weighting. The results are evaluated through waveform clarity and recognition rate as well as comparative experiments.

Images