Impact moment detection method based on MFCC feature extraction and CNN classification algorithm

Abstract

A collision moment detection method based on MFCC feature extraction and a CNN classification algorithm belongs to the technical field of signal processing, and comprises the following steps: an audio data acquisition and processing module acquires audio data of a target environment in real time; the audio data preprocessing module carries out preprocessing operation on the audio data to obtain standard audio data; an MFCC feature extraction module sequentially performs transformation, filtering and cepstrum analysis operation on the standard audio data to obtain 39-dimensional feature data; a CNN classification and recognition module performs classification and recognition operation on the 39-dimensional feature data to obtain an impact moment prediction value; and the impact moment detection output module compares the impact moment predicted value with a set impact threshold value, when the impact moment predicted value is larger than the set impact threshold value, it is indicated that the impact moment is detected, and otherwise, it is indicated that no impact event is detected. The method has the advantages that interference of noise on the impact moment recognition effect can be reduced, and the impact moment detection precision is improved.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, and in particular to a collision moment detection method based on MFCC feature extraction and CNN classification algorithm. Background Technology

[0002] Impact moment detection refers to the precise capture and location of the start point of transient events caused by impact, collision, or internal damage in a mechanical system, mechanical structure, or a specific external environment through various types of sensors and signal processing technologies. Its core objective is to achieve microsecond to nanosecond time resolution for short-duration, sudden impact signals against backgrounds of noise, harmonics, or multi-source aliasing. It is widely used in fields such as fault diagnosis, safety early warning, and engineering surveying.

[0003] In the fields of mechanical fault monitoring, safety protection systems, and engineering seismic surveying, accurate detection of the impact moment is crucial. In mechanical fault monitoring, the transient impact signals generated by early faults such as bearing spalling and gear cracks are weak and short-lived, and the time resolution of traditional vibration analysis methods is not high enough, leading to missed faults. In the field of safety protection systems, for high-speed impact events such as bird strikes and foreign object intrusions on rails, it is necessary to minimize the response time and achieve high-precision real-time triggering. In the field of engineering seismic surveying, the accurate extraction and capture of the first arrival waveform places extremely high demands on the real-time accuracy of the source energy release moment detection and its resistance to background noise interference.

[0004] Disadvantages of existing technologies: Existing collision moment detection algorithms typically employ methods such as threshold detection, STFT, wavelet transform, and HHT, which have limitations such as low time resolution, reliance on manual rules leading to poor generalization, weak anti-interference ability, low detection accuracy, and low computational efficiency resulting in poor real-time performance. Summary of the Invention

[0005] The present invention provides an impact moment detection method based on MFCC feature extraction and CNN classification algorithm, which can reduce the interference of noise on the impact moment recognition effect and improve the impact moment detection accuracy.

[0006] To achieve the above objectives, this invention provides a collision moment detection method based on MFCC feature extraction and CNN classification algorithm, which, as a key feature, includes the following steps:

[0007] Step 1: Construct an impact moment detection system based on 39-dimensional MFCC feature extraction and CNN classification algorithm. The impact moment detection system is equipped with an audio data acquisition and processing module, an audio data preprocessing module, an MFCC feature extraction module, a CNN classification and recognition module, and an impact moment detection output module connected in sequence.

[0008] Step 2: The audio data acquisition and processing module acquires audio data a from the target environment in real time and transmits it to the audio data preprocessing module;

[0009] Step 3: The audio data preprocessing module preprocesses the audio data a to obtain standard audio data b, and then passes it to the MFCC feature extraction module.

[0010] Step 4: The MFCC feature extraction module performs transformation, filtering, and cepstral analysis on the standard audio data b in sequence to obtain 39-dimensional feature data c, which is then passed to the CNN classification and recognition module.

[0011] Step 5: The CNN classification and recognition module performs classification and recognition operations on the 39-dimensional feature data c to obtain the predicted value d of the impact time, d∈[0,1], and passes it to the impact time detection output module;

[0012] Step 6: The impact moment detection output module compares the predicted impact moment value d with the set impact threshold. When the predicted impact moment value d is greater than the set impact threshold, it indicates that an impact moment has been detected; otherwise, no impact event has been detected.

[0013] The audio data acquisition and processing module uses a short-time frame signal cyclic acquisition method to make the acquisition of audio data at the impact moment more accurate and complete. The audio data preprocessing module performs pre-emphasis, framing, and windowing on the acquired sound data, balancing the high and low frequencies of the audio to make the sound signal more stable, while also trimming the audio signal and preventing leakage. The MFCC feature extraction module performs FFT transformation, filtering, cepstral analysis, and other operations on the preprocessed audio to provide feature parameters for the downstream classification model. The CNN classification and recognition module classifies and recognizes the 39-dimensional feature data output by the MFCC feature extraction module to determine whether the audio segment contains the impact moment signal.

[0014] The above design effectively reduces the interference of noise on the impact moment recognition effect, and improves the impact moment detection accuracy and calculation efficiency.

[0015] Preferably, the audio data acquisition and processing module uses an integrated audio decoding module to acquire audio data a, with a sampling frequency of 8KHz. It performs three 500ms loop acquisitions per startup and stores the audio file in .WAV format.

[0016] The audio data acquisition and processing module uses an integrated audio decoding module to acquire sound data at the moment of impact. The audio data acquisition and processing module uses a method of cyclic acquisition of short-time frame signals to make the audio data acquisition at the moment of impact more accurate and complete.

[0017] Preferably, in step 3, the audio data preprocessing module performs preprocessing operations on the audio data a, and the specific steps are as follows:

[0018] Step A1: The audio data preprocessing module pre-emphasizes the audio data a using a pre-emphasis filter to obtain pre-emphasized audio data a';

[0019] The expression for pre-emphasis calculation is:

[0020] ;

[0021] Where a' represents the pre-emphasized audio data; a represents the original audio data; For time; Indicates the pre-emphasis coefficient. ;

[0022] Step A2: The audio data preprocessing module divides the pre-emphasized data a' into segments according to a set time length. Each frame signal is then multiplied by a Hamming window function to obtain a Hamming frame signal. All Hamming frame signals together constitute the standard audio data b.

[0023] The Hamming window function expression is:

[0024] ;

[0025] in, This represents a weighted window function; Represents the window function coefficients. ; This represents the index of the sampling point of the window function. L represents the window length.

[0026] The audio data preprocessing module first pre-emphasizes the audio signal using a pre-emphasis filter, effectively balancing the high and low frequencies of the audio signal to avoid numerical problems in subsequent transformations and improve the signal-to-noise ratio. Then, the pre-emphasized audio signal is divided into short-time frame signals to make the signal stable and ensure the meaningfulness and effectiveness of subsequent feature extraction. Finally, by multiplying each frame of data after framing by a window function, data leakage caused by frame truncation is minimized as much as possible, ensuring signal quality.

[0027] Preferably, in step 4, the MFCC feature extraction module extracts the standard audio data. The transformation, filtering, and cepstral analysis operations are performed sequentially, with the specific steps as follows:

[0028] Step B1: The MFCC feature extraction module processes the standard audio data. Perform an FFT transform to obtain an audio data set, where each element represents the speech signal energy within one frame. The FFT transform expression is:

[0029] ;

[0030] in, This represents the frequency domain signal value at the k-th frequency point; Indicates k + Frequency domain signal values ​​at each frequency point; It is the FFT result of the first half of the sequence in the FFT butterfly decomposition; It is the FFT result of the second half of the sequence in the FFT butterfly decomposition; It is the k-th power of the fundamental twitch factor, expressed as: j is the imaginary unit; e is the base of the natural logarithm function; k is the frequency index; N is the FFT length, i.e., the total number of sampling points involved in the transformation;

[0031] Step B2: The MFCC feature extraction module constructs a Mel filter array of the same height as the audio data set, and selects 26 triangular filters to filter the audio data set to obtain the filtered sound signal. This makes the filtered sound signal closer to the human ear's auditory mechanism.

[0032] The formula for the equal-height Mel filter is:

[0033] ;

[0034] in, This represents the filtering weight of the m-th Mel filter at the k-th frequency point; This represents the center frequency index of the m-th Mel filter; k represents the frequency index in the frequency domain; and m represents the index of the Mel filter, i.e., the m-th Mel filter.

[0035] Step B3: The MFCC feature extraction module processes the filtered audio signal. Taking the logarithm and then performing the discrete cosine transform, the expression is:

[0036] ;

[0037] in, This represents the MFCC feature value in the nth dimension of the i-th frame; n represents the dimension index of the MFCC; i represents the index of the time frame; m represents the index of the Mel filter; M represents the total number of Mel filters. This represents the Mel power spectrum value of the i-th frame after passing through the m-th Mel filter;

[0038] Step B4: The MFCC feature extraction module extracts the spectral envelope from the discrete cosine transform (DCT) sound signal as the sound signal's feature parameters. The first 13 dimensions of these feature parameters are then used as the first-order features of the MFCC. First-order and second-order difference operations are then performed on these first-order features to obtain 13-dimensional first-order and second-order difference features, respectively. These first-order features, first-order difference features, and second-order difference features together constitute 39-dimensional feature data. ;

[0039] The expression for the first-order difference operation is:

[0040] ;

[0041] The expression for the second-order difference operation is:

[0042] ;

[0043] in, Represents the original first-order feature. Represents the first-order difference feature. Second-order difference features, where i represents the time frame index and j represents the dimension index of MFCC.

[0044] Preferably, the CNN classification and recognition module is provided with an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a first fully connected layer, a second fully connected layer, and an output layer connected in sequence. The output terminals of the first convolutional layer, the second convolutional layer, and the first fully connected layer are all provided with ReLU activation functions, and the output terminal of the output layer is provided with a Sigmoid activation function.

[0045] Preferably, the CNN classification and recognition module processes the 39-dimensional feature data. The specific steps for performing classification and recognition are as follows:

[0046] Step C1: The input layer of the CNN classification and recognition module acquires data with the following dimensions. 39-dimensional feature data and reconstruct it into The one-dimensional vector is then passed to the first convolutional layer; This represents the total number of frame signals, with each frame containing 39 MFCC features.

[0047] Step C2: The first convolutional layer uses 12 convolutional kernels of size 1×6 to perform one-dimensional convolution on the reconstructed one-dimensional vector along the time dimension. The convolution stride is 1 and the padding method is "same". Then, the output of the convolutional layer is non-linearly transformed by the ReLU activation function. Then, the first pooling layer performs max pooling operation with a pooling window size of 1×2 and a stride of 2 to obtain the first pooled data.

[0048] Step C3: The second convolutional layer uses 12 convolutional kernels of size 1×6 to perform a one-dimensional convolution operation on the first pooling data along the time dimension. The convolution stride is 1 and the padding method is "same". Then, the output of the convolutional layer is non-linearly transformed by the ReLU activation function. Then, the second pooling layer performs a max pooling operation with a pooling window size of 1×2 and a stride of 2 to obtain the second pooling data.

[0049] Step C4: The first fully connected layer flattens the second pooled data and performs a full connection operation, then activates it through the ReLU activation function, and then performs a full connection operation through the second fully connected layer to obtain fully connected data;

[0050] Step C5: The output layer uses a single neuron and uses the Sigmoid activation function to map the fully connected data to the range of 0 to 1 to obtain the predicted value d for the impact time.

[0051] Preferably, when an impact is detected at any time, the impact moment detection output module performs time compensation by calculating the program running time, locates the time point to which the impact moment belongs on the system time axis through compensation calculation, and outputs an electrical signal to the signal receiving system to complete the impact moment detection output function.

[0052] The beneficial effects of this invention are: This invention provides an impact moment detection method based on MFCC feature extraction and CNN classification algorithm, which can effectively reduce the interference of noise on the impact moment recognition effect, improve data quality, and greatly improve the impact moment detection accuracy and computational efficiency. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the data preprocessing and feature extraction process in the embodiment;

[0054] Figure 2 This is a flowchart of the MFCC feature extraction procedure in the embodiment;

[0055] Figure 3 This is a graph of MFCC eigenvalue data from an embodiment.

[0056] Figure 4 This is a network structure diagram of the CNN classification and recognition module in the embodiment;

[0057] Figure 5 This is a block diagram of the impact moment detection system in the embodiment. Detailed Implementation

[0058] The present invention will be further described in detail below with reference to the accompanying drawings and specific examples. The following embodiments or drawings are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0059] A collision moment detection method based on MFCC feature extraction and CNN classification algorithm includes the following steps:

[0060] Step 1: Construct an impact moment detection system based on 39-dimensional MFCC feature extraction and CNN classification algorithm, such as... Figure 5 As shown, the impact moment detection system is equipped with an audio data acquisition and processing module, an audio data preprocessing module, an MFCC feature extraction module, a CNN classification and recognition module, and an impact moment detection output module connected in sequence.

[0061] Step 2: The audio data acquisition and processing module acquires audio data a from the target environment in real time and transmits it to the audio data preprocessing module;

[0062] Step 3: The audio data preprocessing module preprocesses the audio data a to obtain standard audio data b, and then passes it to the MFCC feature extraction module.

[0063] Step 4: The MFCC feature extraction module performs transformation, filtering, and cepstral analysis on the standard audio data b in sequence to obtain 39-dimensional feature data c, which is then passed to the CNN classification and recognition module.

[0064] Step 5: The CNN classification and recognition module performs classification and recognition operations on the 39-dimensional feature data c to obtain the predicted value d of the impact time, d∈[0,1], and passes it to the impact time detection output module;

[0065] Step 6: The impact moment detection output module compares the predicted impact moment value d with the set impact threshold. When the predicted impact moment value d is greater than the set impact threshold, it indicates that an impact moment has been detected; otherwise, no impact event has been detected.

[0066] The audio data acquisition and processing module uses an integrated audio decoding module to acquire audio data a, with a sampling frequency of 8KHz. It performs three 500ms loop acquisitions per startup and stores the audio file in .WAV format.

[0067] like Figure 1 , Figure 2As shown, the audio data preprocessing module performs preprocessing operations on the audio data a. The specific steps are as follows:

[0068] Step A1: The audio data preprocessing module pre-emphasizes the audio data a using a pre-emphasis filter to obtain pre-emphasized audio data a';

[0069] The expression for pre-emphasis calculation is:

[0070] ;

[0071] Where a' represents the pre-emphasized audio data; a represents the original audio data; For time; Indicates the pre-emphasis coefficient;

[0072] Step A2: The audio data preprocessing module divides the pre-emphasized data a' into segments according to a set time length. Each frame signal is multiplied by a Hamming window function to obtain a Hamming frame signal. All Hamming frame signals together constitute the standard audio data b. In this embodiment, the frame length of each frame signal is set to 25ms, resulting in 33 frame signals.

[0073] The Hamming window function expression is:

[0074] ;

[0075] in, This represents a weighted window function; Represents the window function coefficients. ; This represents the index of the sampling point of the window function. L represents the window length.

[0076] The audio data preprocessing module first pre-emphasizes the audio signal using a pre-emphasis filter, effectively balancing the high and low frequencies of the audio signal to avoid numerical problems in subsequent transformations and improve the signal-to-noise ratio. Then, the pre-emphasized audio signal is divided into short-time frame signals to make the signal stable and ensure the meaningfulness and effectiveness of subsequent feature extraction. Finally, by multiplying each frame of data after framing by a window function, data leakage caused by frame truncation is minimized as much as possible, ensuring signal quality.

[0077] The MFCC feature extraction module performs transformation, filtering, and cepstral analysis operations on the standard audio data b in sequence. The specific steps are as follows:

[0078] Step B1: The MFCC feature extraction module performs an FFT transform on the standard audio data b to obtain an audio data set, where each element represents the speech signal energy within a frame; the FFT transform expression is:

[0079] ;

[0080] in, This represents the frequency domain signal value at the k-th frequency point; Indicates k + Frequency domain signal values ​​at each frequency point; It is the FFT result of the first half of the sequence in the FFT butterfly decomposition; It is the FFT result of the second half of the sequence in the FFT butterfly decomposition; It is the k-th power of the fundamental twitch factor, expressed as: j is the imaginary unit; e is the base of the natural logarithm function; k is the frequency index; N is the FFT length, i.e., the total number of sampling points involved in the transformation;

[0081] Step B2: The MFCC feature extraction module constructs a Mel filter array of the same height as the audio data set, and selects 26 triangular filters to filter the audio data set to obtain the filtered sound signal. This makes the filtered sound signal closer to the human ear's auditory mechanism.

[0082] The formula for the equal-height Mel filter is:

[0083] ;

[0084] in, This represents the filtering weight of the m-th Mel filter at the k-th frequency point; This represents the center frequency index of the m-th Mel filter; k represents the frequency index in the frequency domain; and m represents the index of the Mel filter, i.e., the m-th Mel filter.

[0085] Step B3: The MFCC feature extraction module processes the filtered audio signal. Taking the logarithm and then performing the discrete cosine transform, the expression is:

[0086] ;

[0087] in, This represents the MFCC feature value in the nth dimension of the i-th frame; n represents the dimension index of the MFCC; i represents the index of the time frame; m represents the index of the Mel filter; M represents the total number of Mel filters. This represents the Mel power spectrum value of the i-th frame after passing through the m-th Mel filter;

[0088] Step B4: The MFCC feature extraction module extracts the spectral envelope from the discrete cosine transform (DCT) sound signal as the sound signal's feature parameters. The first 13 dimensions of these feature parameters are then used as the first-order features of the MFCC. First-order and second-order difference operations are then performed on these first-order features to obtain 13-dimensional first-order and second-order difference features, respectively. These first-order features, first-order difference features, and second-order difference features together constitute 39-dimensional feature data. ;

[0089] The expression for the first-order difference operation is:

[0090] ;

[0091] The expression for the second-order difference operation is:

[0092] ;

[0093] in, Represents the original first-order feature. Represents the first-order difference feature. Let i represent the second-order difference feature, i represent the time frame index, and j represent the dimension index of MFCC.

[0094] like Figure 4 As shown, the CNN classification and recognition module is configured with an input layer (Input), a first convolutional layer (Cov1), a first pooling layer (Pool1), a second convolutional layer (Cov2), a second pooling layer (Pool2), a first fully connected layer (Fc1), a second fully connected layer (Fc2), and an output layer (Output) connected in sequence. The outputs of the first convolutional layer (Cov1), the second convolutional layer (Cov2), and the first fully connected layer (Fc1) are all equipped with ReLU activation functions, and the output of the output layer (Output) is equipped with a Sigmoid activation function.

[0095] The CNN classification and recognition module uses the Adam optimizer for network training, employs binary cross-entropy as the loss function, and uses labeled collision / non-collision audio datasets for supervised learning, ultimately achieving accurate identification of the collision moment.

[0096] The CNN classification and recognition module processes the 39-dimensional feature data. The specific steps for performing classification and recognition are as follows:

[0097] Step C1: The input layer of the CNN classification and recognition module acquires feature data with a data dimension of 33 frames × 39 dimensions, such as... Figure 3 As shown, it is reconstructed into a 1×1287 one-dimensional vector and then passed to the first convolutional layer;

[0098] Step C2: The first convolutional layer uses 12 convolutional kernels of size 1×6 to perform one-dimensional convolution on the reconstructed one-dimensional vector along the time dimension. The convolution stride is 1 and the padding method is "same". Then, the output of the convolutional layer is non-linearly transformed by the ReLU activation function. Then, the first pooling layer performs max pooling operation with a pooling window size of 1×2 and a stride of 2 to obtain the first pooled data.

[0099] Step C3: The second convolutional layer uses 12 convolutional kernels of size 1×6 to perform a one-dimensional convolution operation on the first pooling data along the time dimension. The convolution stride is 1 and the padding method is "same". Then, the output of the convolutional layer is non-linearly transformed by the ReLU activation function. Then, the second pooling layer performs a max pooling operation with a pooling window size of 1×2 and a stride of 2 to obtain the second pooling data.

[0100] Step C4: The first fully connected layer flattens the second pooled data and performs a full connection operation, then activates it through the ReLU activation function, and then performs a full connection operation through the second fully connected layer to obtain fully connected data;

[0101] Step C5: The output layer uses a single neuron and uses the Sigmoid activation function to map the fully connected data to the range of 0 to 1 to obtain the predicted value d for the impact time.

[0102] When an impact is detected at any time, the impact moment detection output module performs time compensation by calculating the program running time. It locates the time point to which the impact moment belongs on the system time axis through compensation calculation and outputs an electrical signal to the signal receiving system to complete the impact moment detection output function.

[0103] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A collision moment detection method based on MFCC feature extraction and CNN classification algorithm, characterized in that, Includes the following steps: Step 1: Construct an impact moment detection system based on 39-dimensional MFCC feature extraction and CNN classification algorithm. The impact moment detection system is equipped with an audio data acquisition and processing module, an audio data preprocessing module, an MFCC feature extraction module, a CNN classification and recognition module, and an impact moment detection output module connected in sequence. Step 2: The audio data acquisition and processing module acquires audio data a from the target environment in real time and transmits it to the audio data preprocessing module; Step 3: The audio data preprocessing module preprocesses the audio data a to obtain standard audio data b, and then passes it to the MFCC feature extraction module. Step 4: The MFCC feature extraction module performs transformation, filtering, and cepstral analysis on the standard audio data b in sequence to obtain 39-dimensional feature data c, which is then passed to the CNN classification and recognition module. Step 5: The CNN classification and recognition module performs classification and recognition operations on the 39-dimensional feature data c to obtain the predicted value d of the impact time, d∈[0,1], and passes it to the impact time detection output module; Step 6: The impact moment detection output module compares the predicted impact moment value d with the set impact threshold. When the predicted impact moment value d is greater than the set impact threshold, it indicates that an impact moment has been detected; otherwise, no impact event has been detected.

2. The collision time detection method based on MFCC feature extraction and CNN classification algorithm according to claim 1, characterized in that: The audio data acquisition and processing module uses an integrated audio decoding module to acquire audio data a, with a sampling frequency of 8KHz. It performs three 500ms loop acquisitions per startup and stores the audio file in .WAV format.

3. The collision moment detection method based on MFCC feature extraction and CNN classification algorithm according to claim 1, characterized in that: In step 3, the audio data preprocessing module performs preprocessing operations on the audio data a. The specific steps are as follows: Step A1: The audio data preprocessing module pre-emphasizes the audio data a using a pre-emphasis filter to obtain pre-emphasized audio data a'; The expression for pre-emphasis calculation is: ； Where a' represents the pre-emphasized audio data; a represents the original audio data; For time; Indicates the pre-emphasis coefficient; Step A2: The audio data preprocessing module divides the pre-emphasized data a' into segments according to a set time length. Each frame signal is then multiplied by a Hamming window function to obtain a Hamming frame signal. All Hamming frame signals together constitute the standard audio data b. The Hamming window function expression is: ； in, This represents a weighted window function; Represents the window function coefficients. ; This represents the index of the sampling point of the window function. L represents the window length.

4. The collision time detection method based on MFCC feature extraction and CNN classification algorithm according to claim 1, characterized in that: In step 4, the MFCC feature extraction module performs transformation, filtering, and cepstral analysis on the standard audio data b in sequence. The specific steps are as follows: Step B1: The MFCC feature extraction module performs an FFT transform on the standard audio data b to obtain an audio data set, where each element represents the speech signal energy within a frame; the FFT transform expression is: ； in, This represents the frequency domain signal value at the k-th frequency point; Indicates k + Frequency domain signal values ​​at each frequency point; It is the FFT result of the first half of the sequence in the FFT butterfly decomposition; It is the FFT result of the second half of the sequence in the FFT butterfly decomposition; It is the k-th power of the fundamental twitch factor, expressed as: j is the imaginary unit; e is the base of the natural logarithm function; k is the frequency index; N is the FFT length, i.e., the total number of sampling points involved in the transformation; Step B2: The MFCC feature extraction module constructs a Mel filter array of the same height as the audio data set, and selects 26 triangular filters to filter the audio data set to obtain the filtered sound signal. ; The formula for the equal-height Mel filter is: ； in, This represents the filtering weight of the m-th Mel filter at the k-th frequency point; This represents the center frequency index of the m-th Mel filter; k represents the frequency index in the frequency domain; and m represents the index of the Mel filter, i.e., the m-th Mel filter. Step B3: The MFCC feature extraction module processes the filtered audio signal. Taking the logarithm and then performing the discrete cosine transform, the expression is: ； in, This represents the MFCC feature value in the nth dimension of the i-th frame; n represents the dimension index of the MFCC; i represents the index of the time frame; m represents the index of the Mel filter; M represents the total number of Mel filters. This represents the Mel power spectrum value of the i-th frame after passing through the m-th Mel filter; Step B4: The MFCC feature extraction module extracts the spectral envelope from the discrete cosine transform (DCT) sound signal as the sound signal's feature parameters. The first 13 dimensions of these feature parameters are then used as the first-order features of the MFCC. First-order and second-order difference operations are then performed on these first-order features to obtain 13-dimensional first-order and second-order difference features, respectively. These first-order features, first-order difference features, and second-order difference features together constitute 39-dimensional feature data. ; The expression for the first-order difference operation is: ; The expression for the second-order difference operation is: ; in, Represents the original first-order feature. Represents the first-order difference feature. Second-order difference features, where i represents the time frame index and j represents the dimension index of MFCC.

5. The collision moment detection method based on MFCC feature extraction and CNN classification algorithm according to claim 1, characterized in that: The CNN classification and recognition module is configured with an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a first fully connected layer, a second fully connected layer, and an output layer connected in sequence. The output terminals of the first convolutional layer, the second convolutional layer, and the first fully connected layer are all equipped with ReLU activation functions, and the output terminal of the output layer is equipped with a Sigmoid activation function.

6. The collision moment detection method based on MFCC feature extraction and CNN classification algorithm according to claim 5, characterized in that: The CNN classification and recognition module processes the 39-dimensional feature data. The specific steps for performing classification and recognition are as follows: Step C1: The input layer of the CNN classification and recognition module acquires data with the following dimensions. 39-dimensional feature data and reconstruct it into 1×( The one-dimensional vector is then passed to the first convolutional layer; Step C2: The first convolutional layer uses 12 convolutional kernels of size 1×6 to perform one-dimensional convolution operation on the reconstructed one-dimensional vector along the time dimension, and then performs non-linear transformation on the output of the convolutional layer through the ReLU activation function; then it performs max pooling operation through the first pooling layer with a pooling window size of 1×2 and a stride of 2 to obtain the first pooling data. Step C3: The second convolutional layer uses 12 convolutional kernels of size 1×6 to perform a one-dimensional convolution operation on the first pooling data along the time dimension, and then performs a non-linear transformation on the output of the convolutional layer through the ReLU activation function; then it performs a max pooling operation through the second pooling layer to obtain the second pooling data; Step C4: The first fully connected layer flattens the second pooled data and performs a full connection operation, then activates it through the ReLU activation function, and then performs a full connection operation through the second fully connected layer to obtain fully connected data; Step C5: The output layer uses a single neuron and uses the Sigmoid activation function to map the fully connected data to the range of 0 to 1 to obtain the predicted value d for the impact time.

7. The collision moment detection method based on MFCC feature extraction and CNN classification algorithm according to claim 1, characterized in that: When an impact is detected at any time, the impact moment detection output module performs time compensation by calculating the program running time. It locates the time point to which the impact moment belongs on the system time axis through compensation calculation and outputs an electrical signal to the signal receiving system to complete the impact moment detection output function.

Images