Non-contact material identification method based on residual network and millimeter wave radar

Abstract

The invention discloses a non-contact material identification method based on a residual network and a millimeter wave radar, which comprises the following steps of: playing an audio frequency by a loudspeaker to excite an object to generate micro vibration, extracting frequency, power and damping characteristics of a target material from an echo signal of the millimeter wave radar, and constructing a multi-characteristic fusion residual network model; according to the method, the extracted features are input into the network model to be processed, the weight of the network is continuously adjusted through a back propagation algorithm, a high-accuracy material recognition result is finally obtained, and the method has the advantages of low cost and simple operation process, does not need to strictly limit the distance between an object and a radar, and has higher flexibility.

Description

Technical Field

[0001] The present invention relates to material recognition technology, and specifically to a non-contact material recognition method based on a residual network and a millimeter-wave radar. Background Art

[0002] With the development of intelligent devices and smart cities, it has become increasingly important to perceive the environment around us. Identifying the materials of objects is an important part of environmental perception and has a wide range of applications in many fields. For example, the classification of waste materials in industrial automation, the detection of liquid and food safety, and the environmental modeling of smart homes. Compared with contact material recognition systems, the research on non-contact material recognition systems has received more and more attention due to their non-destructiveness and fewer usage restrictions.

[0003] The main existing methods for non-contact material recognition include near-infrared spectroscopy (NIR), optical sensing technology, and radio technology. NIR spectroscopy is used to detect the electromagnetic spectrum with wavelengths ranging from 780 nm to 2500 nm and has been applied to identify many organic materials. However, this special equipment usually costs tens of thousands of yuan, has a high cost, and is mainly used for laboratory analysis. Optical sensing technology identifies materials by analyzing the multi-spectral absorption and reflection characteristics of materials, but its accuracy is affected by the visibility of objects. Capturing material characteristics using radio signals can provide a cheap, non-invasive, and generally applicable material recognition method. WiMi uses the phase and amplitude changes generated when wireless signals penetrate liquids to identify different materials. mSense identifies materials by using millimeter-wave radios and measuring the reflection coefficients of wireless signals reflected from objects. Tagtag attaches RFID tags to targets and identifies materials by observing the high-resolution phase changes caused by differences in the impedance of tag antennas. However, these systems all have certain limitations, such as relying on tags, having fewer material-related features, and low material recognition accuracy. Summary of the Invention

[0004] To solve the problems existing in the prior art, the present invention proposes a non-contact material recognition method based on a residual network and a millimeter-wave radar, aiming to provide a non-contact, non-damaging, and high-precision material recognition method that can identify the materials of different objects without strictly restricting the distance between the radar and the target.

[0005] A non-contact material recognition method based on a residual network and a millimeter-wave radar includes the following steps:

[0006] Step 1: Use a speaker to play audio beside the object to excite micro-vibrations of the target object, and at the same time use a millimeter-wave radar to transmit a frequency-modulated continuous-wave signal to detect the target and obtain the radar echo;

[0007] Step 2: Further process the radar echo data, reorganize the reflected signals received by the millimeter-wave radar into a four-dimensional data block containing the characteristic change information in the time domain, the frequency characteristic distribution information of multiple samples, different receiving antennas, and different transmitting antennas, and correspondingly store the material label of the target;

[0008] Step 3: Perform frequency-domain analysis on the intermediate-frequency signal after mixing the transmitted signal and the echo signal of the frequency-modulated continuous-wave signal to obtain the distance information R of the target. Obtain the micro-motion change of the target according to the phase change of the target position, and extract the frequency, power, and damping characteristics of the target material, which can reflect the characteristics of the object material;

[0009] Step 4: Construct a multi-feature fusion residual network model with ResNet18 as the backbone network, input the extracted features into the network model for processing to obtain high-dimensional feature information, use the features of a large number of known material objects as training data, and continuously adjust the weights of the network through the backpropagation algorithm, so that the network can accurately identify the material of the object according to the input features, and finally output the recognition result of the object material.

[0010] Further, in the above Step 1, use a millimeter-wave radar to transmit a frequency-modulated continuous wave and collect the echo data at different distances between the target and the radar:

[0011] The mathematical model of the radar transmitting FMCW signal can be expressed as:

[0012]

[0013] where A is the signal amplitude, f c is the starting frequency of the transmitted signal, B is the bandwidth of the transmitted signal, T is the sweep time of the transmitted signal, is the frequency modulation slope.

[0014] Set the distance between the target and the radar as R, and the echo signal formed by the signal reflected by the target and then received by the radar can be expressed as:

[0015]

[0016] The radar echo signal and the transmitted signal have the same signal waveform, and there is a delay between them in the time domain. t d is the echo time delay, c is the speed of light, and the calculation formula is:

[0017]

[0018] Further, in the above Step 2, reorganize the reflected signals received by the millimeter-wave radar into a four-dimensional data block containing the characteristic change information in the time domain, the frequency characteristic distribution information of multiple samples, different receiving antennas, and different transmitting antennas, and correspondingly store the material label of the target.

[0019] Furthermore, in the step 2, when parsing the radar echo data, the echo signal is reorganized into a four-dimensional data block containing feature change information in the time domain, frequency feature distribution information of multiple samples, different receiving antennas and different transmitting antennas according to the parameter setting of the FMCW signal, and the material label of the target is stored accordingly to construct a data set.

[0020] Furthermore, in step 3, the formula of the intermediate frequency signal after mixing is as follows:

[0021]

[0022] The intermediate frequency signal can be regarded as having a frequency of f IF and the phase is The sinusoidal signal is IF The frequency domain analysis can obtain the distance information R of the target, and the micro-motion change of the target can be obtained according to the phase change of the target position. If the micro-motion distance of the object is ΔR and the radar wavelength is λ, then the corresponding phase change It can be expressed by the following formula:

[0023]

[0024] Furthermore, in step 3, the frequency feature extraction method is:

[0025] 1. Analyze the radar echo signal and perform fast Fourier transform to locate the target position;

[0026] 2. Extract the phase information of different frames at the target position and unwrap the phase;

[0027] 3. Use short-time Fourier transform to analyze the phase change of the object:

[0028] STFT(t,ω)=∫x(τ)w(τ-t)e -jωτ dτ

[0029] Because the low-frequency components of 0-20HZ often represent more violent movements of objects and do not correspond to the vibration frequency range caused by the sound source, this part of invalid frequencies is filtered out; in addition, the target material has nothing to do with the time when the object resonates, so the time dimension is removed from the STFT and only the frequency information is retained;

[0030] Normalize each window of the resulting STFT to a standard normal distribution by subtracting the mean and dividing by the standard deviation.

[0031] 5. Sum the signal along the time window axis to obtain a frequency characteristic that does not change with time.

[0032] Furthermore, in step 3, the power feature extraction method is:

[0033] For an electromagnetic wave propagating in free space, the signal amplitude is inversely proportional to the propagation distance d, which can be expressed by the Friis transmission formula as:

[0034]

[0035] where A0 is the amplitude of the transmitted signal, A d is the received amplitude at the propagation distance d, g t , g r are the transmitting antenna gain and the receiving antenna gain respectively, and λ is the wavelength; since the target reflection signal is utilized and the radar transmitting and receiving antennas are at the same position, d is multiplied on the transmitting signal path and the echo signal path respectively to compensate for the amplitude loss, that is, after performing distance - dimension FFT on the original data, the amplitude is multiplied by the square of the distance corresponding to each distance bin as the power feature.

[0036] Further, in the third step, the damping feature extraction method is as follows:

[0037] The experiment uses a pseudo - pulse sequence as the audio signal to excite the object vibration. After each pulse, the object experiences a decay process from resonance to rest, that is, the damping phenomenon.

[0038] 1. Extract the unwrapped phase of the received millimeter - wave radar signal, and then apply a Butterworth band - pass filter for filtering. Using the Welch power density spectrum estimation method, the frequency at the peak is selected as the pass - band cut - off frequency.

[0039] 2. Take the modulus of the Hilbert transform and calculate the envelope of the filtered phase signal. For each pulse emitted by the speaker, the object vibration has corresponding rises and falls. Find the peak corresponding to each rise in the envelope phase of the object: Assume there are approximately 250 phase samples between two pulses, so find the maximum value within every 250 points of the envelope phase. For each maximum value, take the next 125 samples, which approximately cover the entire time when the object stops vibrating before the start of the next pulse.

[0040] 3. Use these curves as the final damping features.

[0041] Further, in the fourth step, the constructed deep - learning network model uses ResNet18 as the backbone network, mainly including an input layer, a residual connection layer, a feature fusion layer, and an output layer. At the same time, the cross - entropy function is used as the loss function, and the Adam optimizer is used for parameter update. Select the model with the highest accuracy rate during the specified number of training epochs for testing:

[0042] The input layer is responsible for receiving the feature vectors obtained in step 3. These data are usually standardized to ensure that the data input to the network has the same scale. The input layer converts these data into a format suitable for processing by the convolution layer, that is, converting one-dimensional data into a multi-channel tensor form to facilitate subsequent high-dimensional feature extraction;

[0043] The residual connection layer is composed of 8 residual blocks, each of which is composed of two one-dimensional convolutional layers, and each convolutional layer is followed by a batch normalization layer and a ReLU activation layer. The short-circuit connection mechanism adds the input of the block directly to the output of the second convolutional layer, so that the gradient can bypass certain layers, thereby simplifying the training process of the deep network and improving the training efficiency and performance of the model. The present invention specially designs independent branches to process three features: frequency, power and damping. Each branch is composed of a series of residual blocks to deeply explore the unique properties of each feature. These branches not only generate results for the final classification, but also produce their own intermediate outputs, which play an important role in the subsequent loss function calculation and guide the optimization of model parameters;

[0044] The feature fusion layer first performs an adaptive average pooling operation on the features extracted from each independent branch to reduce the spatial dimension of the features and retain the most core information. Subsequently, the three pooled feature vectors are spliced ​​together along the feature dimension to form a comprehensive feature vector. This splicing method not only retains the unique information of each feature, but also promotes the interaction between different features, thereby generating a richer and more comprehensive feature representation. Finally, this comprehensive feature vector will be sent to the output layer to generate the final classification prediction;

[0045] In the output layer, for the result after feature fusion, the Dropout strategy is adopted to prevent overfitting and improve the generalization ability of the model by randomly discarding a part of neurons during the training process. In addition, the model outputs the corresponding prediction results for each feature, which are an important part of the loss function calculation;

[0046] The present invention uses the cross entropy function as the loss function, which not only calculates the loss value of the fusion feature prediction result, but also calculates the loss value for each feature separately. The final loss value is obtained by normalizing and summing the importance of the feature in proportions of 1, 0.9, 0.3, and 0.3. This design ensures that the model can be optimized in a targeted manner according to the actual contribution of the feature;

[0047] The present invention uses the Adam optimizer to perform parameter updates. The optimizer combines the advantages of momentum and adaptive learning rate, and can dynamically adjust the learning rate according to the change of gradient, accelerate the convergence speed of the model, and ensure the stability of the training process.

[0048] The present invention sets a total of 30 training cycles. After each round of training, the performance of the model on the validation set is evaluated, and the best-performing model is recorded. Finally, the model version with the highest accuracy rate on the validation set is selected for testing to ensure that the model has good generalization ability and practical application value.

[0049] Advantages:

[0050] 1. The material recognition method of the present invention has the characteristics of non-contact, non-damage, and low cost compared with the existing recognition methods. Since the millimeter-wave radar is very small in size, it is convenient to be transplanted into systems for detecting different materials. More importantly, even when the surface of the object to be measured is thinly covered by plastic film or paint, etc., this method can still maintain accurate material recognition ability.

[0051] 2. The material recognition method of the present invention realizes object material recognition based on a multi-feature fusion residual network combined with a millimeter-wave radar sensor. This method has strong flexibility and greater practical value, can more effectively reveal the physical characteristics of the target object, and has a very broad application prospect.

[0052] 3. The average accuracy rate of the present invention for material recognition of 4 common materials, namely metal, cardboard, wood, and glass, is 99.3%. It can be seen that the invention has a high material recognition accuracy, and the recognition accuracy is less interfered by the shape of the object, and is applicable to recognizing the materials of a variety of different-shaped objects. Description of the Drawings

[0053] Figure 1 is the overall structure diagram of the present invention;

[0054] Figure 2 is the flow chart of extracting frequency features, power features, and damping features of the present invention;

[0055] Figure 3 is the schematic diagram of the network structure of ResNet18 used in the present invention;

[0056] Figure 4 is the schematic diagram of the structure of the multi-feature fusion residual network model of the present invention;

[0057] Figure 5 is the feature map of different materials extracted by the present invention. Among them, (a) is the frequency feature result of metal, (b) is the power feature result of metal, (c) is the damping feature result of metal, (d) is the frequency feature result of wood, (e) is the power feature result of wood, and (f) is the damping feature result of wood;

[0058] Figure 6It is a photo of the experimental materials selected in the present invention. Among them, (a) is a metal material, (b) is a wood material, (c) is a cardboard material, and (d) is a glass material;

[0059] Figure 7 It is the confusion matrix of the material recognition results implemented in the present invention. Specific implementation manner

[0060] The purpose of the present invention is to solve the problems existing in the existing non-contact material recognition methods, such as high cost and low recognition accuracy, and provide a non-contact material recognition method based on a residual network and a millimeter-wave radar, realizing non-contact, non-destructive, and high-precision material recognition, and not requiring strict restrictions on the distance between the radar and the target.

[0061] Figure 1 It is the overall structure diagram of the method of the present invention. The speaker plays audio next to the target, exciting the object to generate micro-vibrations. At the same time, the millimeter-wave radar collects the echo data of the target, analyzes the target reflection signal, extracts three features of power, frequency, and damping from it, and finally builds a neural network model and fuses the three features to realize object material recognition.

[0062] Step 1: Use the speaker to play audio next to the object, exciting the target object to generate micro-vibrations. At the same time, use the millimeter-wave radar to transmit a frequency-modulated continuous wave signal to detect the target and obtain the radar echo:

[0063] In the experiment: The target is placed between the radar and the speaker. Since the sound signal attenuates very quickly, the speaker should be placed as close to the target as possible to ensure that the sound wave effectively hits the object and causes micro-vibrations of the object. The audio is played continuously for 10 seconds to fully excite the vibration of the object. At the same time, the radar transmits a frequency-modulated continuous wave to detect the target and collects the echo data at different distances between the target and the radar.

[0064] The FMCW echo signal at the position R from the radar is:

[0065]

[0066] Step 2: Further process the radar echo data, reorganize the reflected signal received by the millimeter-wave radar into a four-dimensional data block including the characteristic change information in the time domain, the frequency characteristic distribution information of multiple samples, and different receiving antennas and different transmitting antennas, and correspondingly store the material label of the target:

[0067] Analyze the radar echo data, reorganize the echo signal into a four-dimensional data block including the characteristic change information in the time domain, the frequency characteristic distribution information of multiple samples, and different receiving antennas and different transmitting antennas according to the parameter settings of the FMCW signal, and correspondingly store the material label of the target to construct a data set.

[0068] Step 3: Perform frequency-domain analysis on the intermediate-frequency signal obtained by mixing the transmitted signal and the echo signal of the frequency-modulated continuous-wave signal to obtain the distance information R of the target. Obtain the micro-motion change of the target based on the phase change of the target position, and extract the frequency, power, and damping characteristics of the target material. These characteristics can reflect the properties of the object material:

[0069] As Figure 2 shown, the steps for frequency characteristic extraction are as follows: First, parse the radar echo data, determine the region where the target is located, perform phase extraction and unwrap the phase at the target, perform short-time Fourier transform on the unwrapped phase, and analyze the phase change.

[0070] STFT(t,ω) = ∫x(τ)w(τ - t)e -jωτ dτ

[0071] Since the target material has nothing to do with the time when the object resonates, the time dimension is removed from the STFT, and only the frequency information is retained. To this end, each window of the obtained STFT is normalized to a standard normal distribution by subtracting the mean and dividing by the standard deviation, and then the signal is summed along the time-window axis to obtain a frequency characteristic that does not change with time.

[0072] As Figure 2 shown, the steps for power characteristic extraction are as follows: First, parse the radar echo data, perform fast Fourier transform on the original data in the range dimension to obtain range bin information.

[0073] According to the Friis transmission formula, multiply the amplitude in each range bin by the square of the corresponding range to complete range compensation, and calculate the amplitude of the reflected signal after range compensation as the power characteristic.

[0074]

[0075] As Figure 2 shown, the steps for damping characteristic extraction are as follows: First, parse the radar echo data, determine the region where the target is located, and extract the unwrapped phase of the echo signal. To better capture the damping characteristics of the target, apply a Butterworth band-pass filter for filtering. Using the Welch power density spectrum estimation method, select the frequency at the peak as the passband cut-off frequency. Then take the modulus of the Hilbert transform and calculate the envelope of the filtered phase signal.

[0076] Step 4: Construct a multi-feature fusion residual network model with ResNet18 as the backbone network, input the extracted features into the network model for processing to obtain high-dimensional feature information, use the features of a large number of objects with known materials as training data, and continuously adjust the weights of the network through the backpropagation algorithm so that the network can accurately identify the material of the object according to the input features, and finally output the recognition result of the object material:

[0077] As Figure 3 shown, the constructed deep learning network model uses ResNet18 as the backbone network, mainly including an input layer, a residual connection layer, a feature fusion layer, and an output layer:

[0078] Among them, the input layer converts the feature vectors of the three features into a multi-channel tensor form with the same number of channels for subsequent feature extraction; the residual connection layer consists of 8 residual blocks, each residual block contains two one-dimensional convolutional layers, and each convolutional layer is followed by a batch normalization layer and a ReLU activation layer. The short-circuit connection mechanism directly adds the input of the block to the output of the second convolutional layer, and when the number of input channels is inconsistent with the number of output channels, a 1×1 convolutional layer is used to match the dimensions to ensure the feasibility of the residual connection. The network model of the present invention designs independent branches to process the three features of frequency, power, and damping, as Figure 4 shown, each branch is composed of a series of residual blocks inside; the feature fusion layer first performs an adaptive average pooling operation on the features extracted from the three independent branches to reduce the spatial dimension of the features. Subsequently, these three pooled feature vectors are concatenated together along the feature dimension to obtain a one-dimensional comprehensive material category feature; for the result after feature fusion in the output layer, the Dropout strategy is adopted, and the Dropout probability is set to 0.25. In addition, the model outputs the corresponding prediction results for each feature respectively for the calculation of the loss function.

[0079] In this embodiment, the cross-entropy function is used as the loss function, which not only calculates the loss value of the prediction result of the fusion feature, but also calculates the loss value for each feature separately. The final loss value is obtained by normalizing and summing according to the importance of the features in the ratio of 1, 0.9, 0.3, 0.3.

[0080] In this embodiment, the Adam optimizer is used to perform parameter update, and the learning rate is set to 0.001.

[0081] In this embodiment, a total of 30 training cycles are set, and the batch size is 32 during the training process. After each round of training ends, the performance of the model on the validation set will be evaluated. Finally, the model with the highest accuracy rate on the validation set is selected for testing.

[0082] To verify the material recognition ability of the contactless material recognition method based on the residual network and millimeter-wave radar proposed by the present invention, 4 types of objects with different materials as Figure 6 shown are selected, where: (a) is a metal material, (b) is a wood material, (c) is a cardboard material, and (d) is a glass material. Table 1 shows the key parameter settings of the radar in the experiment. Figure 7 The confusion matrix of the network model is shown. The experimental results show that the method of the present invention has a material recognition accuracy rate as high as 99.3%.

[0083] Table 1 Radar Parameter Settings

[0084]

[0085]

[0086] In summary, the above is only the preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A non-contact material recognition method based on residual network and millimeter wave radar, characterized in that: The following steps are involved: Step 1: Use a speaker to play audio next to the object to stimulate the target object to produce micro-vibration, and use the millimeter wave radar to transmit a frequency modulated continuous wave signal to detect the target and obtain the radar echo; Step 2: further process the radar echo data, reorganize the reflected signal received by the millimeter-wave radar into a four-dimensional data block containing feature change information in the time domain, frequency feature distribution information of multiple samples, different receiving antennas and different transmitting antennas, and store the material label of the target accordingly; Step 3: Perform frequency domain analysis on the intermediate frequency signal after mixing the transmission signal of the frequency modulated continuous wave signal and the echo signal to obtain the distance information R of the target, obtain the micro-motion change of the target according to the phase change of the target position, and extract the frequency, power and damping characteristics of the target material. These characteristics can reflect the characteristics of the object material. Step 4: Build a multi-feature fusion residual network model with ResNet18 as the backbone network, input the extracted features into the network model for processing, obtain high-dimensional feature information, use a large number of features of objects with known materials as training data, and continuously adjust the network weights through the back propagation algorithm, so that the network can accurately identify the material of the object based on the input features, and finally output the recognition result of the object material.

2. The method according to claim 1, characterized in that In step 1, a millimeter wave radar is used to transmit a frequency modulated continuous wave and collect echo data at different distances between the target and the radar: The mathematical model of the radar transmitting FMCW signal is expressed as: Where A is the signal amplitude, f c is the starting frequency of the transmitting signal, B is the bandwidth of the transmitting signal, T is the frequency sweep time of the transmitting signal, It is the frequency modulation slope; Assuming the distance between the target and the radar is R, the echo signal formed by the signal reflected by the target and then received by the radar can be expressed as: The radar echo signal and the transmitted signal have the same signal waveform, but there is a delay between them in the time domain. d is the echo delay, and the calculation formula is: Where R is the distance between the target and the radar, and the speed of electromagnetic waves propagating in space is the speed of light c.

3. The method according to claim 1, characterized in that In the step 2, the reflected signal received by the millimeter wave radar is reorganized into a four-dimensional data block containing feature change information in the time domain, frequency feature distribution information of multiple samples, different receiving antennas and different transmitting antennas, and the material label of the target is stored accordingly.

4. The method according to claim 1, characterized in that In the step 2, when parsing the radar echo data, the echo signal is reorganized into a four-dimensional data block containing feature change information in the time domain, frequency feature distribution information of multiple samples, different receiving antennas and different transmitting antennas according to the parameter setting of the FMCW signal, and the material label of the target is stored accordingly to construct a data set.

5. The method according to claim 1, characterized in that In step 3, the formula of the intermediate frequency signal after mixing is as follows: The intermediate frequency signal can be regarded as having a frequency of f IF and the phase is The sinusoidal signal is IF The frequency domain analysis can obtain the distance information R of the target, and the micro-motion change of the target can be obtained according to the phase change of the target position. If the micro-motion distance of the object is ΔR and the radar wavelength is λ, then the corresponding phase change It can be expressed by the following formula:

6. The method according to claim 5, characterized in that In the step 3, the frequency feature extraction method is:

1. Analyze the radar echo signal and perform fast Fourier transform to locate the target position; 2. Extract the phase information of different frames at the target position and unwrap the phase; 3. Use short-time Fourier transform to analyze the phase change of the object: STFT(t,ω)=∫x(τ)w(τ-t)e -jωτ dt Among them, x(t) is the original signal, w(t) is the window function, which is used to limit the time range of analysis, t is the time variable, which indicates the center position of the window function, and ω is the angular frequency, which indicates the frequency component of the signal. Because the low-frequency component of 0-20HZ often represents more violent movement of the object, it does not correspond to the vibration frequency range caused by the sound source, so this part of invalid frequency is filtered out; in addition, the target material has nothing to do with the time when the object resonates, so the time dimension is removed from the STFT, and only the frequency information is retained; Normalize each window of the resulting STFT to a standard normal distribution by subtracting the mean and dividing by the standard deviation.

5. Sum the signal along the time window axis to obtain a frequency characteristic that does not change with time.

7. The method according to claim 5, characterized in that In step 3, the power feature extraction method is: For electromagnetic waves propagating in free space, the signal amplitude is inversely proportional to the propagation distance d, which can be expressed by the Friis transmission formula: Where A0 is the amplitude of the transmitted signal, A d is the received amplitude at the propagation distance d, g t , g r are the transmitting antenna gain and the receiving antenna gain, respectively, and λ is the wavelength. Since the target reflected signal is used and the radar transmitting and receiving antennas are in the same position, the transmitting signal path and the echo signal path are multiplied by d to compensate for the amplitude loss. That is, after performing range-dimensional FFT on the original data, the amplitude is multiplied by the square of the corresponding distance of each range bin as the power feature.

8. The method according to claim 5, characterized in that In step 3, the damping feature extraction method is:

1. Extract the unwrapped phase of the received millimeter-wave radar signal, and then apply a Butterworth bandpass filter to filter it; use the Welch power density spectrum estimation method to select the frequency at the peak as the passband cutoff frequency; 2. Take the modulus of the Hilbert transform and calculate the envelope of the filtered phase signal; for each pulse emitted by the speaker, the object vibration has a corresponding rise and fall; find the peak corresponding to each rise in the object's envelope phase; 3. Use these curves as the final damping characteristics.

9. The method according to claim 1, characterized in that In the fourth step, the constructed deep learning network model uses ResNet18 as the backbone network, mainly including an input layer, a residual connection layer, a feature fusion layer and an output layer. At the same time, the cross entropy function is used as the loss function, and the Adam optimizer is used for parameter update.

10. The method according to claim 9, characterized in that In step 4, the residual connection layer consists of 8 residual blocks, each residual block consists of two one-dimensional convolutional layers, and each convolutional layer is followed by a batch normalization layer and a ReLU activation layer.

11. The method according to claim 9, characterized in that In the fourth step, the feature fusion layer first performs an adaptive average pooling operation on the features extracted from each independent branch to reduce the spatial dimension of the features and retain the most core information; then, the three pooled feature vectors are spliced ​​together along the feature dimension to form a comprehensive feature vector; finally, this comprehensive feature vector will be sent to the output layer to generate the final classification prediction.

12. The method according to claim 9, characterized in that In the step 4, in the output layer, for the result after feature fusion, the Dropout strategy is adopted to prevent overfitting by randomly discarding a part of neurons during the training process, thereby improving the generalization ability of the model.

13. The method according to claim 9, characterized in that In the fourth step, the cross entropy function is used as the loss function, which not only calculates the loss value of the fusion feature prediction result, but also calculates the loss value for each feature separately; the final loss value is obtained by normalizing and summing the features in proportion to 1, 0.9, 0.3, and 0.3 according to their importance.

14. The method according to claim 9, characterized in that In the fourth step, a total of 30 training cycles are set. After each round of training, the performance of the model on the validation set is evaluated and the best performing model is recorded; finally, the model version with the highest accuracy on the validation set is selected for testing.

Images