Bone injury detection method and equipment based on ultrasonic guided wave and multi-branch convolutional neural network, and medium

Abstract

The invention discloses a bone injury detection method based on ultrasonic guided waves and a multi-branch convolutional neural network, and the method comprises the steps: firstly collecting UGW signals of bone tissues, and carrying out the standardization preprocessing of the UGW signals, so as to eliminate the amplitude difference between different samples; then, data enhancement strategies such as Gaussian noise disturbance and amplitude scaling are adopted, a finite sample set is effectively expanded, and the robustness and generalization ability of the model are improved; a multi-branch convolutional neural network (TB-CNN) structure is constructed, the network designs independent branches for time domain, frequency domain and time-frequency domain input, deep features are automatically extracted from the respective branches, and feature splicing and multi-level information integration are realized in a fusion module; finally, regression prediction of the bone injury depth is realized through a full connection layer, so that continuous value output is obtained, and the severity of the bone injury can be reflected more finely.

Description

Technical Field

[0001] This invention relates to the fields of biomedical engineering and artificial intelligence, specifically to a method, device, and medium for detecting bone injuries based on ultrasonic guided waves and multi-branch convolutional neural networks. Background Technology

[0002] Bone injuries and osteoporosis, among other skeletal diseases, have become significant threats to human health worldwide. With the increasing aging population, their incidence is rising annually, placing a heavy burden on public health systems. Currently, dual-energy X-ray absorptiometry (DXA) is widely used in clinical diagnosis as the gold standard for bone mineral density testing. However, DXA has significant limitations in reflecting changes in bone tissue microstructure and early bone injury. Furthermore, its large equipment size, high cost, and ionizing radiation restrict its application in real-time monitoring, preoperative assessment, and bedside testing.

[0003] In recent years, guided ultrasonic wave (UGW) technology has gradually become a research hotspot in the field of bone injury detection due to its advantages such as being non-destructive, portable, radiation-free, and highly sensitive to microstructural changes in bone tissue. When UGW propagates in bone structures, its propagation characteristics (such as amplitude, phase, and dispersion) are highly sensitive to micro-damage, geometric changes, and material properties within the bone, enabling effective responses in the early stages of bone injury.

[0004] However, existing UGW-based detection methods mostly rely on single features in the time or frequency domain, such as amplitude attenuation, propagation delay, or power spectrum distribution. These features are not robust in complex noise environments and are difficult to accurately characterize continuous changes in bone injury depth, limiting their widespread application in clinical practice. To overcome these limitations, some studies have attempted to combine UGW signal processing with machine learning (ML) methods.

[0005] Traditional machine learning methods, such as SVR, MLP, and linear regression, have shown promising results in early bone injury detection tasks. However, these methods generally rely on artificially constructed feature sets, and the model performance is highly dependent on the quality and experience of feature selection. Furthermore, they have limited ability to handle complex nonlinear signal variations, making it difficult to achieve refined and continuous prediction of the degree of bone injury.

[0006] To address this, this application proposes a multi-branch regression prediction model based on ultrasonic guided wave (UGW) and convolutional neural network (CNN). This model integrates time domain, frequency domain, and time-frequency domain data. Figure 3The model employs multi-scale automatic feature extraction from the enhanced UGW signal through independent branches, and integrates cross-modal information in the fusion module. Using the ratio of bone injury cross-sectional area as the regression target, the model achieves end-to-end quantitative prediction for samples with different injury degrees. To improve the model's generalization ability under small sample conditions, this application further introduces data augmentation strategies (such as Gaussian noise injection and amplitude scaling) to expand the diversity of training samples.

[0007] In summary, current research lacks an effective method that can simultaneously integrate multimodal UGW signal features, maintain high robustness under small sample conditions, and achieve continuous prediction of bone injury depth. Therefore, there is an urgent need to develop a new technical approach that fully utilizes the multi-scale information of UGW signals to significantly improve the accuracy, stability, and clinical applicability of bone injury detection. Summary of the Invention

[0008] This invention proposes a bone injury detection method based on ultrasonic guided waves and multi-branch convolutional neural networks, aiming to solve the problem of insufficient detection accuracy of existing methods under complex signal environments and small sample data conditions.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a bone injury detection method based on ultrasonic guided wave and multi-branch convolutional neural network. The method first uses ultrasonic guided wave (UGW) technology to acquire bone signals, obtaining both normal and damaged signals. The acquired signals are then standardized to effectively remove noise interference and retain key injury-related information. In the feature extraction stage, multimodal features are extracted from the signals, constructing multi-scale features in the time domain, frequency domain, and time-frequency domain. The time-domain features reflect transient changes in the bone signal, the frequency-domain features reveal the energy distribution pattern through Fast Fourier Transform (FFT), and the time-frequency domain features are generated by stitching together six time-frequency transforms (WVD, STFT, TRT, CWT, SLT, and SST) in a fixed order, comprehensively characterizing the dynamic evolution of the signal.

[0010] After feature extraction is completed, a multi-branch convolutional neural network model is constructed. The model specifically includes three independent branches: the first branch is a one-dimensional convolutional branch, used to process time-domain features; the second branch is a one-dimensional convolutional branch, used to process frequency-domain features; and the third branch is a two-dimensional convolutional branch, used to process the composite time-frequency image; receiving time-domain, frequency-domain, and time-frequency-domain features as inputs respectively. Each branch extracts deep features, and the multimodal features are spliced ​​and integrated in the fusion layer; The fused features are input into a fully connected layer for regression prediction, and the output is a continuous numerical result of the bone injury depth.

[0011] The convolutional neural network was trained using the Adam optimization algorithm, and Dropout and L2 regularization strategies were introduced to effectively prevent overfitting under small sample conditions. At the same time, mini-batch gradient descent (batch size=16) and K-fold cross-validation were used to train and validate the multi-branch convolutional neural network model to ensure the robustness and generalization ability of the prediction results.

[0012] The specific steps are as follows: Step 1: Use a UGW acquisition device to acquire raw sample signals of the bone, including normal signals and cutting damage signals, and convert the signals into digital format; Step 2: Perform Z-score normalization on the signals acquired in Step 1; Step 3: Divide the processed signal from Step 2 into a test set and a training set. Ensure that the test set does not participate in any augmentation operations and retains its original distribution for model performance evaluation. Perform Gaussian noise perturbation and amplitude scaling on the training set for data augmentation. Step four: After data augmentation in step three, the time-domain signal is converted into the frequency domain using Fast Fourier Transform (FFT). Various time-frequency analysis methods are then introduced to transform the time-domain signal into a two-dimensional time-frequency graph. Step 5: The signals transformed in Step 4 are fed into a multi-branch convolutional neural network model to automatically extract features, namely time-domain features, frequency-domain features, and time-frequency-domain features. Deep-level features are then extracted through convolution and pooling operations. Step 6: In the fusion module, the multi-branch features from Step 5 are spliced ​​and integrated, and then input into the fully connected layer for training. The Adam optimization algorithm is used to iteratively update the parameters, while Dropout and L2 regularization are introduced to prevent overfitting. Step 7: Train and validate the multi-branch convolutional neural network model from Step 6 using mini-batch gradient descent (batch size=16) and K-fold cross-validation to obtain stable prediction performance; Step 8: Use the multi-branch convolutional neural network model trained and optimized in Step 7 to predict the bone signal and output the regression result of the damage depth.

[0013] Furthermore, the signal acquired in step one is subjected to Z-score normalization, including the following steps: The formula for the first step of preprocessing is as follows: in The original guided wave signal, This is the signal after the first preprocessing step. M express length; The second step involves randomly dividing the standardized signal data into training and test sets in an 8:2 ratio. This ensures the test set is not subjected to any augmentation operations and retains its original distribution for model performance evaluation. Subsequently, two augmentation methods are applied to each sample in the training set: one is adding Gaussian noise, which... Gaussian noise is added to generate enhanced samples. Follows a mean of 0 and a standard deviation of Normal distribution: Enhanced signal Represented as: ,in, The intensity of noise is used to control its strength and can be adjusted according to actual needs. In this application, The value is set to 0.05 to preserve the main characteristics of the signal while enhancing data diversity. Secondly, amplitude scaling is used, specifically by multiplying each sample point of the signal by a random perturbation factor. It can be determined by Gaussian distribution. , Let represent the variance of the normal distribution. Let the original signal be... Thus, the perturbed signal is constructed. .

[0014] Furthermore, the augmented samples generated after data augmentation, combined with the previous test set samples, now result in a total sample dataset of... , The time series data is essentially a time-domain signal, so no additional conversion is needed. Next, we will convert the time-domain signal to the frequency domain, calculate the frequency domain representation of each sample using FFT, and generate a frequency axis for subsequent visualization. The calculation formula is as follows: In the formula This represents the time-domain signal at the nth sampling point, where N is the total length of the signal. This represents the complex frequency domain value corresponding to the k-th frequency point. To improve the efficiency of FFT calculation, each signal is first padded with zeros to the nearest zero. The points are then normalized. The resulting complex spectrum retains only its amplitude spectrum: Due to the original signal Since it is a real-valued signal, its spectrum has conjugate symmetry, therefore only the first part is retained. The points are treated as one-sided amplitude spectra and multiplied by 2 to maintain energy consistency. The final frequency axis is constructed using the following formula: ,in, Indicates the sampling frequency. , This represents the frequency value corresponding to the k-th frequency point. Frequency domain analysis is performed because damage causes changes in the frequency structure of the guided wave signal; spectral analysis can reveal these patterns of change, which is helpful for feature extraction and identification.

[0015] Furthermore, after converting the time-domain signal to the frequency domain, time-frequency analysis was conducted. This paper employs six typical time-frequency analysis methods: Wigner-Ville distribution (WVD), short-time Fourier transform (STFT), time redistribution transform (TRT) based on STFT, continuous wavelet transform (CWT), spectral local transform under multi-scale Morlet wavelet (SLT), and synchronous compression transform (SST). Each method generates a corresponding two-dimensional representation in a fixed order, ultimately forming a composite time-frequency image to more intuitively reveal the local variation characteristics of the signal in both time and frequency dimensions. 1. WVD for augmented time-domain signals Time-frequency analysis was performed to capture the fine structure of the signal in the time-frequency domain. Wavelength volume decomposition (WVD) is a typical nonlinear time-frequency analysis method that provides a high-resolution time-frequency representation, clearly showing the energy distribution of the signal. WVD exhibits a conjugate symmetric structure along the frequency axis. To highlight the energy distribution of the main frequency components, this paper only plots the positive frequency portion for analysis. Notably, two regions of concentrated energy can still be observed in the visualization results, reflecting the coexistence of multi-scale frequency components in the signal. This characteristic is significant for identifying the damage state of bone structures. The WVD of a signal can be represented as: Where n is the time index, representing a discrete time point; is the frequency index, representing a discrete frequency point; and is the time delay variable used to calculate the product of the signal at different time points. yes The complex conjugate of the signal. This formula reveals how WVD captures the energy distribution of a signal at different times and frequencies by calculating the product of signals at different time points and performing a Fourier transform.

[0016] 2. STFT is the most basic and commonly used time-frequency analysis method. This method analyzes the signal in the time domain... Apply a window function of length L. This method obtains short-time signal segments centered at a specific time point. By dividing the signal into a series of short time intervals, the local frequency characteristics of the signal at various time positions can be approximated. For each time point, the window function is slid to that position, and the discrete Fourier transform of the windowed signal is calculated to obtain the spectral information at that time. Its mathematical definition is as follows: ,in It is a signal with data augmentation. The STFT is a window function centered on the time position, where represents the frequency index and is the total length of the signal. The STFT reflects the frequency component changes of the signal at different time positions and can be used to analyze the characteristics of frequency changes over time in non-stationary signals.

[0017] 3. Time-Frequency Representation (TRT) is a novel method for time-frequency analysis that improves the resolution of time-frequency representations by redistributing the energy of the STFT (Spatial Time-Frequency Transform). The core idea of ​​TRT is to find local maxima along the time direction based on the STFT and redistribute energy to these local maxima points, thereby generating a sparser and more concentrated time-frequency representation. This involves identifying local maxima points in the STFT matrix; if a point... satisfy: and If the value is found to be local maximum, then it is determined to be a local maximum point. Finally, the TRT matrix is ​​constructed. Only retain the STFT coefficients at local maxima, and set the coefficients at other locations to zero: if If it is a local maximum point, then for In other cases, It is zero.

[0018] 4. Convolutional Wavelet (CWT) is a time-frequency tool for analyzing non-stationary signals, particularly suitable for scenarios with complex time-frequency characteristics, such as impulse signals. Its principle involves convolving the signal with a set of wavelet basis functions generated by scaling and shifting a mother wavelet function. This allows for simultaneous localization of signal features in both the time and frequency domains. The main advantage of CWT is its high time-frequency resolution. By selecting appropriate wavelet basis functions (such as complex Morlet wavelets), signal features can be better matched, enabling efficient signal analysis. CWT exhibits high frequency resolution in the low-frequency band and high time resolution in the high-frequency band. This characteristic allows it to more flexibly adapt to changes in signal frequency components over time, providing a clearer perspective for extracting the time-frequency features of the signal. In this study, we used complex Morlet wavelets as the mother wavelet function, which possesses good localization properties in the frequency domain and can effectively capture the frequency information of the signal. The calculation formula for CWT is as follows: The mother wavelet is a complex Morlet wavelet, and the calculation formula is as follows: .in, It is the input signal. It is a discrete-time mother wavelet function. It is its complex conjugate, where a is the scale parameter and b is the translation parameter. It is a scaling factor used to maintain energy normalization. It is the center angular frequency of the complex Morlet wavelet. By calculating the absolute values ​​of the wavelet coefficients, we obtained the time-frequency energy distribution of the signal.

[0019] 5. SLT is essentially a superposition of multi-scale wavelet modulus lengths, exhibiting good energy concentration and time-frequency localization characteristics, making it suitable for the analysis of non-stationary signals. Compared to the frequency ambiguity problem inherent in traditional CWT, SLT enhances frequency resolution by selecting appropriate wavelet scale ranges and wavelet functions (such as complex Morlet wavelets). Where k is the preset wavelet scale number, which is 32 here. For complex Morlet wavelets, the scale... and the corresponding center frequency There is an approximate relationship: Ultimately Visualize the signal on the frequency and time axes to form a two-dimensional time-frequency plot, which can intuitively reflect the distribution characteristics of signal energy in the time-frequency domain.

[0020] 6. SST is typically based on CWT. Its core idea is to redistribute the frequency axis after obtaining the time-frequency plot, thereby improving frequency resolution. CWT, when processing multi-component signals, leads to ambiguous frequency positioning due to frequency diffusion. SST, after calculating the wavelet transform, estimates the instantaneous frequency using phase information and recompresses the frequency energy back to its actual frequency position, resulting in a clearer and sparser time-frequency representation. First, the CWT of the signal is calculated, and then the corresponding instantaneous frequency estimate is calculated using the CWT coefficients. Next, the energy of the CWT is redistributed along the frequency direction based on the instantaneous frequency estimate, thus obtaining a synchronously compressed time-frequency representation. Synchronous compression is defined as... Its discrete synchronous compression redistribution formula is: and ,in, express phase angle, Where b is the scale, K is the translation parameter, and K is the number of scales (32). This indicates that energy will be redistributed to the instantaneous frequency. It is the k-th scale The wavelet coefficients at time point b.

[0021] Furthermore, multimodal features are extracted from the converted signal, including time-domain features, frequency-domain features, and composite time-frequency image features generated by six time-frequency transforms (WVD, STFT, TRT, CWT, SLT, SST); the converted data is then fed into a CNN model, which automatically extracts the features.

[0022] Furthermore, after extracting time-domain, frequency-domain, and time-frequency plot features using a CNN, the features from different modalities are deeply fused. The 128-dimensional vectors output by each of the three branches are concatenated along the channel direction to obtain a 384-dimensional fused feature. This vector is first subjected to Dropout (dropout rate 0.2) to reduce the risk of overfitting, then input into a fully connected layer with L2 regularization (128 neurons, ReLU activation), and then subjected to another Dropout (0.2) before entering a fully connected layer with 64 neurons (ReLU activation). Finally, a linear output layer is used to regress and predict the degree of bone damage.

[0023] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0024] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0025] Compared with the prior art, the present invention has the following advantages: This invention comprehensively utilizes time-domain, frequency-domain, and multi-scale time-frequency information, effectively overcoming the shortcomings of existing methods such as single feature and insufficient robustness. By employing a multi-branch convolutional neural network to achieve cross-modal deep fusion, this invention significantly outperforms existing models such as ResNet, SVR, MLP, and linear regression in prediction accuracy. Furthermore, this invention overcomes the limitations of discrete classification in traditional methods through continuous regression prediction, achieving a quantitative and precise assessment of bone injury severity.

[0026] The technical features of this invention also include: 1. Multimodal feature construction: Combining multi-dimensional signal processing methods such as time domain, frequency domain and time-frequency domain, and using six typical time-frequency analysis methods such as WVD, STFT, TRT, CWT, SLT and SST to generate composite time-frequency images in a fixed order, thereby comprehensively characterizing the bone injury features contained in the ultrasound guided wave signal.

[0027] 2. Construct a multi-branch convolutional neural network architecture: Design a multi-branch structure consisting of 1D-CNN branches (for time-domain and frequency-domain signals) and 2D-CNN branches (for time-frequency images). Each branch extracts deep features, and multi-level feature splicing and integration are performed in the fusion module to achieve joint modeling of cross-modal information.

[0028] 3. Robust training mechanism for small samples: Under limited data conditions, Gaussian noise perturbation and amplitude scaling data augmentation methods are used, combined with Dropout, L2 regularization and K-fold cross-validation techniques, to significantly alleviate the overfitting problem under small sample conditions.

[0029] 4. Continuous injury depth regression prediction: By outputting continuous values ​​through fully connected layers, the depth of bone injury can be quantitatively predicted, which breaks through the limitations of traditional methods of classifying injury as present or absent and can more accurately depict the severity of injury. Attached Figure Description

[0030] Figure 1 This is a flowchart of the algorithm of the present invention; Figure 2 This is a schematic diagram of the UGW signal acquisition device of the present invention; Figure 3 These are comparison images before and after the standardized preprocessing of this invention; Figure 4 This is a comparison chart of the data enhancement effects of the present invention; Figure 5 The present invention converts the time-domain signal into a spectrum using FFT; Figure 6 This is the time-frequency diagram of the time-domain signal after time-frequency domain conversion according to the present invention; Figure 7 This is a schematic diagram of the cutting process of the present invention; Figure 8 These are two 1D branch network structure diagrams in the CNN of this invention embodiment; Figure 9 These are two 2D branch network structure diagrams in the CNN of this invention embodiment; Figure 10 This is a diagram of the three-branch network structure according to an embodiment of the present invention; Figure 11 This is a graph showing the prediction performance of TB-CNN in an embodiment of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0032] This invention proposes an innovative method for bone injury detection, specifically a method based on ultrasonic guided waves and multi-branch convolutional neural networks. This method aims to avoid the radiation risks associated with traditional X-ray-based detection, reduce detection costs, improve the portability of the detection device, and enhance the detection and prediction capabilities for early bone injuries. Compared with existing technologies, this method removes environmental noise interference and expands the limited sample set by standardizing and preprocessing the UGW signal and performing data augmentation. In the feature construction process, this invention extracts bone injury-related features from three perspectives: time domain, frequency domain, and time-frequency domain. Time domain features reflect changes in signal amplitude and delay, frequency domain features reveal differences in energy distribution, and time-frequency domain features utilize methods such as WVD, STFT, TRT, CWT, SLT, and SST in a fixed order to form composite time-frequency images, comprehensively characterizing the microstructural features and injury patterns of bone.

[0033] This invention designs a TB-CNN model that constructs independent branches for different time-domain inputs to automatically extract multi-level features. The model includes three independent branches: the first branch is a one-dimensional convolutional branch for processing time-domain features; the second branch is a one-dimensional convolutional branch for processing frequency-domain features; and the third branch is a two-dimensional convolutional branch for processing the composite time-frequency image. Each branch extracts deep features, and the features are deeply integrated in the fusion layer. Finally, the regression prediction of bone injury depth is achieved through a fully connected layer, outputting a continuous numerical result of bone injury depth.

[0034] To prevent overfitting under small sample conditions, this invention introduces cross-validation and regularization techniques during model training, effectively improving the robustness and generalization performance of predictions. Experiments show that this method significantly outperforms traditional methods and existing deep learning models in predicting bone injuries under complex environments, enabling more accurate assessment of fracture severity and demonstrating promising application prospects in the early screening of osteoporosis and fractures.

[0035] Specifically, an embodiment of the present invention provides a bone injury detection method based on ultrasonic guided waves and multi-branch convolutional neural networks, comprising the following steps: (1) Raw signals of bone tissue were acquired using a UGW acquisition device, with excised bovine bone (approximately 22 cm in length) as the sample. In the experimental design, UGW signals from undamaged bone were first acquired as normal samples. Subsequently, a 1 mm cut was made on the bone surface using mechanical means, and damage depths ranging from 1 mm to 7 mm were created at the same location on each bone to simulate different degrees of bone tissue damage, thus establishing a complete damage dataset. A narrowband excitation signal with a center frequency of 65 kHz was used in the experiment, and a coupling layer was used to ensure good contact between the transducer and the sample surface to reduce signal loss. To improve sampling stability and signal-to-noise ratio, 30 sets of signals were repeatedly acquired at each damage depth, and the data were converted into digital format for storage, ultimately yielding a total of 240 sample signals.

[0036] (2) Before feature extraction, the collected UGW signals are preprocessed to improve data quality. The purpose is to eliminate environmental noise and individual differences, ensuring that the extracted features accurately reflect the true state of bone tissue, thereby improving the prediction performance of subsequent TB-CNN. First, the original signal is Z-score normalized. This method can eliminate the differences in amplitude and dimension between different samples, allowing feature comparison of each signal on a uniform scale. Its mathematical expression is as follows:

[0037] in The original guided wave signal, This is the signal after the first preprocessing step. M express length; (3) The standardized signal data is randomly divided into training and test sets in an 8:2 ratio, ensuring that the test set does not participate in any enhancement operations and retains its original distribution for model performance evaluation. Two enhancement methods are applied to each sample in the training set: one is to add Gaussian noise, which will affect the signal quality of the training set. Gaussian noise is added to generate enhanced samples. Follows a mean of 0 and a standard deviation of Normal distribution: Enhanced signal Represented as: ,in, Used to control noise intensity, it can be adjusted according to actual needs. The value is set to 0.05 to preserve the main characteristics of the signal while enhancing data diversity. Secondly, amplitude scaling is used, specifically by multiplying each sample point of the signal by a random perturbation factor. It can be determined by Gaussian distribution. , Let represent the variance of the normal distribution. Let the original signal be... Thus, the perturbed signal is constructed. .

[0038] (4) The augmented samples generated after data augmentation, plus the previous test set samples, now make up the total sample dataset. , The time series data is essentially a time-domain signal, so no additional conversion is needed. Next, we will convert the time-domain signal to the frequency domain, calculate the frequency domain representation of each sample using FFT, and generate a frequency axis for subsequent visualization. The calculation formula is as follows: In the formula This represents the time-domain signal at the nth sampling point, where N is the total length of the signal. This represents the complex frequency domain value corresponding to the k-th frequency point. To improve the efficiency of FFT calculation, each signal is first padded with zeros to the nearest zero. The points are then normalized. The resulting complex spectrum retains only its amplitude spectrum: Due to the original signal Since it is a real-valued signal, its spectrum has conjugate symmetry, therefore only the first part is retained. The points are treated as one-sided amplitude spectra and multiplied by 2 to maintain energy consistency. The final frequency axis is constructed using the following formula: ,in, Indicates the sampling frequency. , This represents the frequency value corresponding to the k-th frequency point. Frequency domain analysis is performed because damage causes changes in the frequency structure of the guided wave signal; spectral analysis can reveal these patterns of change, which is helpful for feature extraction and identification.

[0039] (5) After completing the conversion from the time domain signal to the frequency domain, time-frequency analysis is performed. This paper uses six typical time-frequency analysis methods, including WVD, STFT, TRT, CWT, SLT, and SST. Each method generates corresponding two-dimensional and three-dimensional time-frequency diagrams to more intuitively reveal the local variation characteristics of the signal in the two dimensions of time and frequency. A composite time-frequency image is generated by splicing the six time-frequency transformations WVD, STFT, TRT, CWT, SLT, and SST in a fixed order. WVD augmentation of time-domain signals Time-frequency analysis was performed to capture the fine structure of the signal in the time-frequency domain. Wavelength volume decomposition (WVD) is a typical nonlinear time-frequency analysis method that provides a high-resolution time-frequency representation, clearly showing the energy distribution of the signal. WVD exhibits a conjugate symmetric structure along the frequency axis. To highlight the energy distribution of the main frequency components, this paper only plots the positive frequency portion for analysis. Notably, two regions of concentrated energy can still be observed in the visualization results, reflecting the coexistence of multi-scale frequency components in the signal. This characteristic is significant for identifying the damage state of bone structures. The WVD of a signal can be represented as: Where n is the time index, representing a discrete time point; f is the frequency index, representing a discrete frequency point; and m is a time delay variable used to calculate the product of the signal at different time points. yes The complex conjugate of the signal. This formula reveals how WVD captures the energy distribution of a signal at different times and frequencies by calculating the product of signals at different time points and performing a Fourier transform.

[0040] STFT is the most basic and commonly used time-frequency analysis method. This method analyzes the signal in the time domain... Apply a window function of length L. This method obtains a short-time signal segment centered at time m, dividing the signal into a series of short time intervals to approximate the local frequency characteristics of the signal at various time positions. For each time m, the window function is slid to that position, and the discrete Fourier transform of the windowed signal is calculated to obtain the spectral information at that time. Its mathematical definition is as follows: ,in It is a signal with data augmentation. STFT is a window function centered at time position m, where f represents the frequency index and N is the total length of the signal. It reflects the frequency component changes of the signal at different time positions and can be used to analyze the frequency-time characteristics of non-stationary signals.

[0041] Time-Frequency Representation (TRT) is a novel method for time-frequency analysis that improves the resolution of time-frequency representations by redistributing the energy of the STFT (Spatial Time-Frequency Transform). The core idea of ​​TRT is to find local maxima along the time direction based on the STFT and redistribute energy to these local maxima points, thereby generating a sparser and more concentrated time-frequency representation. This involves identifying local maxima points in the STFT matrix; if a point... satisfy: and If the value is found to be local maximum, then it is determined to be a local maximum point. Finally, the TRT matrix is ​​constructed. Only retain the STFT coefficients at local maxima, and set the coefficients at other locations to zero: if If it is a local maximum point, then for In other cases, It is zero.

[0042] CWT (Convolutional Wavelet Wrapper) is a time-frequency tool for analyzing non-stationary signals, particularly suitable for scenarios with complex time-frequency characteristics, such as impulse signals. Its principle involves convolving the signal with a set of wavelet basis functions generated by scaling and shifting a mother wavelet function, thereby simultaneously locating signal features in both the time and frequency domains. The main advantage of CWT is its high time-frequency resolution. By selecting appropriate wavelet basis functions (such as complex Morlet wavelets), signal features can be better matched, enabling efficient signal analysis. CWT exhibits high frequency resolution in the low-frequency band and high time resolution in the high-frequency band. This characteristic allows it to more flexibly adapt to changes in signal frequency components over time, providing a clearer perspective for extracting the time-frequency features of the signal. In this invention, complex Morlet wavelets are used as the mother wavelet function, which has good localization characteristics in the frequency domain and can effectively capture the frequency information of the signal. The calculation formula for CWT is as follows: The mother wavelet is a complex Morlet wavelet, and the calculation formula is as follows: in, It is the input signal. It is a discrete-time mother wavelet function. It is its complex conjugate, where a is the scale parameter and b is the translation parameter. It is a scaling factor used to maintain energy normalization. It is the center angular frequency of the complex Morlet wavelet. By calculating the absolute values ​​of the wavelet coefficients, we obtained the time-frequency energy distribution of the signal.

[0043] SLT is essentially a superposition of multi-scale wavelet modulus lengths, exhibiting good energy concentration and time-frequency localization characteristics, making it suitable for the analysis of non-stationary signals. Compared to the frequency ambiguity problem inherent in traditional CWT, SLT enhances frequency resolution by selecting appropriate wavelet scale ranges and wavelet functions (such as complex Morlet wavelets). Where k is the preset wavelet scale number, which is 32 here. For complex Morlet wavelets, the scale... and the corresponding center frequency There is an approximate relationship: Ultimately Visualize the signal on the frequency and time axes to form a two-dimensional time-frequency plot, which can intuitively reflect the distribution characteristics of signal energy in the time-frequency domain.

[0044] SST (Synchronously Compressed Time-Frequency Transform) is typically based on CWT (Concurrent Wavelet Transform). Its core idea is to redistribute the frequency axis after obtaining the time-frequency plot, thereby improving frequency resolution. CWT, when processing multi-component signals, leads to ambiguous frequency localization due to frequency diffusion. SST, after calculating the wavelet transform, estimates the instantaneous frequency using phase information and recompresses the frequency energy back to its actual frequency position, resulting in a clearer and sparser time-frequency representation. First, the CWT of the signal is calculated, and then the corresponding instantaneous frequency estimate is calculated using the CWT coefficients. Next, the energy of the CWT is redistributed along the frequency direction based on the instantaneous frequency estimate, thus obtaining a synchronously compressed time-frequency representation. Synchronous compression is defined as... Its discrete synchronous compression redistribution formula is: and ,in, express phase angle, Here, 'b' is the scale, 'k' is the translation parameter, and 'K' is the number of scales (32). This indicates that energy will be redistributed to the instantaneous frequency. It is the k-th scale The wavelet coefficients at time point b.

[0045] (6) Following the content of (5) above, extract multimodal features from the converted signal, including time domain features, frequency domain features, and composite time-frequency images generated by six fixed-order time-frequency transforms (WVD, STFT, TRT, CWT, SLT, SST).

[0046] (7) After completing (6), the converted data is fed into the CNN model, and the CNN model automatically extracts features. CNN is a classic feedforward neural network. Its deep structure usually includes a data input layer, convolutional layer, activation layer, pooling layer, concatenation layer, fully connected layer and output layer. 1D-CNN is the same as traditional CNN in architecture, except that the dimension of convolution and pooling operations is reduced from two dimensions to one dimension, so it is more suitable for processing time series or one-dimensional signal data. After the CNN extracts the time domain, frequency domain and time-frequency plot features, the features of different modalities are deeply fused; the 128-dimensional vectors output by the three branches are concatenated in the channel direction to obtain a 384-dimensional fused feature. The vector is first dropped out (dropout rate 0.2) to reduce the risk of overfitting, and then input into a fully connected layer with L2 regularization (128 neurons, ReLU activation), and then dropped out (0.2) again before entering a fully connected layer with 64 neurons (ReLU activation). Finally, the degree of bone damage is predicted by a linear output layer regression.

[0047] In summary, to achieve high-precision prediction of bone injury severity, this invention designs a multimodal feature fusion and deep learning regression model TB-CNN based on UGW (User-Generated Wounds) data. The key technical feature of this invention lies in combining UGW signal acquisition, multimodal feature construction, and a multi-branch convolutional neural network to build a novel bone injury detection system. Specific improvements include: (1) Using UGW to replace traditional X-ray detection, achieving a radiation-free, low-cost and highly portable detection method, and effectively capturing subtle changes in the microstructure of bone tissue through a 65kHz high-frequency narrowband signal; (2) Proposing a multimodal feature fusion method, integrating time domain, frequency domain and six time-frequency domain transformations (WVD, STFT, TRT, CWT, SLT, SST) to form a composite feature input, significantly improving the ability to characterize bone injury patterns; (3) Designing a dedicated TB-CNN structure, constructing independent branches for feature extraction for different modal inputs, and realizing multi-level feature integration in the fusion module, outputting continuous prediction results of bone injury depth through a fully connected layer, maintaining extremely high prediction accuracy even under small sample conditions. Experimental results show that the method of the present invention achieves an RMSE of 0.0053 and an R² of 0.9998 in the bone injury depth prediction task, which is significantly better than traditional ResNet, SVR, linear regression and MLP methods. The TB-CNN model comprises three convolutional modules: a temporal input branch, a frequency domain input branch, and a time-frequency domain input branch. Each branch consists of multiple layers of convolutional and pooling units, used to progressively extract local and global features. The fusion module concatenates the deep features from the three branches and inputs them into a fully connected layer for prediction, ultimately outputting a continuous value of bone injury depth. During model training, the Adam optimization algorithm is used with an initial learning rate of 1e-3. Dropout and L2 regularization strategies are introduced into the network to effectively prevent overfitting. Mini-batch gradient descent (batch size=16) and K-fold cross-validation are employed during training to ensure the model's generalization ability and stability.

[0048] This invention significantly improves the robustness and small-sample adaptability of the model while maintaining prediction accuracy, making it particularly suitable for bone injury detection and prediction applications on small to medium-sized experimental datasets. Experimental results show that this method can more accurately assess the severity of bone injuries in complex environments, providing a reliable solution for early screening of osteoporosis and fractures.

[0049] In summary, this invention employs UGW (Underground Wild-Wide) technology to acquire signals. This technology has attracted widespread attention due to its radiation-free nature, cost-effectiveness, portability, and high sensitivity to early bone damage. Compared to traditional X-ray imaging techniques, UGW technology provides a safer and more economical means of assessing bone density and structural integrity, significantly reducing the risk of radiation exposure while improving accessibility and accuracy.

[0050] Specifically, after collecting the signal using UGW technology, the original signal is first standardized to have a mean of 0 and a variance of 1, thereby eliminating amplitude differences and noise interference between samples. Based on the standardized signal, further data augmentation can be performed, including adding Gaussian noise and amplitude scaling, to expand the finite sample set and improve the model's robustness to random disturbances and its adaptability to differences in signal strength. Considering that bone signals are affected by noise and environmental factors, a single feature is insufficient for stable and accurate damage prediction. This invention extracts features from three perspectives: time domain, frequency domain, and time-frequency domain, to comprehensively capture the multi-scale information and complexity of bone injury signals. Time domain features characterize amplitude attenuation and time delay characteristics, frequency domain features reflect energy distribution patterns, and time-frequency domain features generate composite images in a fixed order using methods such as WVD, STFT, TRT, CWT, SLT, and SST, revealing the energy accumulation and frequency shift characteristics caused by bone injury. This multimodal feature design can more comprehensively reflect the microstructural changes of bone tissue. After construction, this invention proposes a TB-CNN model. This model constructs independent branches for feature extraction based on different time-domain, frequency-domain, and time-frequency-domain inputs, and deeply integrates multimodal features in the fusion module. Finally, it outputs continuous predicted values ​​of bone injury depth through a fully connected layer. This structure can automatically learn multi-level features, effectively avoiding redundancy and instability in the manual feature selection process. During model training, this invention uses cross-validation to train and validate the TB-CNN model, dividing the dataset into several subsets to ensure that each sample participates in the training and validation process. Regularization and Dropout strategies are introduced during training, and the Adam optimization algorithm is used to update parameters, effectively preventing overfitting under small sample conditions. Experimental results show that the method of this invention exhibits excellent performance in bone injury prediction tasks, with an RMSE of 0.0053 and an R² of 0.9998, outperforming existing ResNet, SVR, MLP, and linear regression methods, thus verifying the robustness and generalization ability of the model.

[0051] The following are examples: Example

[0052] like Figure 1 As shown, this invention proposes a method for ultrasound detection of bone injuries based on multimodal features and TB-CNN. The overall process includes key steps such as signal acquisition, signal preprocessing, data augmentation, feature extraction, feature fusion, model training, and prediction. Through this process, UGW signals can be systematically converted into multimodal features usable by deep learning models, thereby achieving high-precision prediction of the degree of bone injury. Firstly, as... Figure 2As shown, the UGW signal acquisition device used in this invention includes a signal excitation module, a sensor, a transducer, and a data acquisition unit. This device is used to detect isolated bovine bones and acquire bone tissue signals under normal and damaged conditions. Unlike traditional X-ray detection, UGW technology has the advantages of being radiation-free, portable, low-cost, and highly sensitive to bone microstructure, thus enabling the acquisition of high-quality bone injury information without incurring additional risks. After signal acquisition, preprocessing is required. The acquired raw signals often contain noise and amplitude differences; if directly used for feature extraction, this can lead to unstable feature representation or even misjudgment. This invention uses the Z-score normalization method to normalize the raw signals, making their mean 0 and variance 1, thereby eliminating dimensional differences between different samples. A comparison of the signal effects before and after preprocessing is provided. Figure 3 As shown, this facilitates subsequent feature extraction. To further expand the limited dataset and improve the model's robustness under small sample conditions, this invention introduces data augmentation strategies after signal preprocessing, mainly including Gaussian noise perturbation and amplitude scaling. The aim is to simulate signal changes that may occur under different experimental environments, enabling the model to learn more representative and widespread feature patterns. Specifically, Gaussian noise perturbation involves superimposing random noise with a mean of 0 and adjustable variance onto the original signal, thereby simulating the effects of electronic interference, poor contact, and environmental noise that the signal may encounter during acquisition. The augmented signal generated in this way is highly consistent with the original signal in waveform, but exhibits random perturbations in local details, enabling the model to improve its adaptability to environmental noise and its anti-interference ability.

[0053] Amplitude scaling, on the other hand, amplifies or reduces the overall amplitude of the original signal to simulate differences in signal amplitude under varying excitation intensities, sensor sensitivities, or coupling conditions. The signal after amplitude scaling maintains a consistent overall shape, but its peak value and energy level change, thus enabling the trained model to generalize better to changes in signal amplitude. For example... Figure 4 As shown, the signal comparison results before and after data augmentation clearly demonstrate the effectiveness of the proposed method. (a) shows the signal after amplitude scaling; the waveform trend remains unchanged, but the overall amplitude level changes, improving the model's adaptability to different signal intensities. (b) shows the signal after Gaussian noise perturbation; it can be observed that while maintaining the main structure, local perturbations are superimposed, enhancing the model's robustness. Through the above augmentation method, the limited dataset is effectively expanded, allowing the model to encounter more types of signal samples during training, thereby improving its adaptability and stability to actual detection environments.

[0054] After data augmentation, this invention further performs feature transformation and extraction on the time-domain signal to comprehensively capture the multi-scale information and complex changes of bone injury signals. Specifically, this invention constructs multimodal feature inputs from three perspectives: time domain, frequency domain, and time-frequency domain, to provide rich feature representations for subsequent TB-CNN. In frequency domain analysis, FFT is first used to transform the augmented time-domain signal into an energy spectrum distribution, such as... Figure 5 As shown in the figure, this spectrum can reveal the differences in energy migration and frequency response caused by bone injury. Normal bone tissue and damaged bone tissue exhibit significant differences in spectral peaks and distribution patterns, thus providing an effective basis for identifying the injury state. In time-frequency domain analysis, this invention employs six classical time-frequency methods, including WVD, STFT, TRT, CWT, SLT, and SST, to characterize the local variation characteristics of the signal in both time and frequency dimensions. The time-frequency images obtained through these methods are shown in the figure. Figure 6 As shown, each method has its own advantages in terms of energy concentration, temporal resolution, and frequency resolution: WVD can highlight the overall energy distribution, but it suffers from cross-term interference; STFT balances time and frequency resolution and is suitable for the analysis of stable signals; TRT can enhance the clarity of transient features; CWT is suitable for multi-scale non-stationary signal analysis; SLT offers flexibility in both time and frequency localization; and SST achieves higher energy concentration through frequency redistribution, highlighting the features of the main components. Ultimately, this invention stitches six types of time-frequency images together in a fixed order to form a composite input, thereby completely preserving the dynamic changes of bone injury signals at the time, frequency, and time-frequency domains, providing a solid data foundation for the model's deep learning.

[0055] To simulate signal changes in bone tissue under different injury states, this invention uses excised bovine bone samples and creates injuries of varying depths through cutting. A detailed illustration of the bone sample cutting is shown below. Figure 7 As shown, the cutting depth ranged from 1 mm to 7 mm, and multiple sets of signals were collected at each depth to establish a complete damage dataset. This experimental design ensured the diversity of sample inputs required for model training, providing a solid data foundation for subsequent prediction tasks.

[0056] In terms of model design, this invention proposes a TB-CNN specifically for multimodal feature fusion and bone injury prediction. This network structure fully considers the differences in representational capabilities of time-domain, frequency-domain, and time-frequency-domain signals, constructing three independent input branch modules to achieve targeted extraction and deep learning of features from different modalities. In the time-domain and frequency-domain branches, this invention employs a 1D-CNN structure, such as... Figure 8As shown, this structure can perform convolution and pooling operations on one-dimensional signal sequences, capturing local features of the signal in the time series or frequency distribution layer by layer, while preserving the overall trend. Correspondingly, the time-frequency domain branch adopts a two-dimensional convolutional neural network structure, such as... Figure 9 As shown, this branch is used to process time-frequency maps input in image form. By sliding the two-dimensional convolutional kernel, this branch can simultaneously extract the spatial correlation and energy distribution features of the signal in both time and frequency dimensions, thereby effectively capturing the differences in time-frequency patterns under different injury states. In the three branches, the convolutional layer is used to extract local pattern features step by step, and the pooling layer is used for dimensionality reduction and feature compression, thereby reducing the amount of computation while enhancing the model's robustness to translation and noise. After multiple convolutional and pooling units, each branch can form a deep representation of the input modality. Subsequently, in the fusion layer, the feature vectors output from the three branches are concatenated and integrated to obtain a unified multimodal feature representation, so as to fully combine information from the time domain, frequency domain, and time-frequency domain. This fused feature can simultaneously reflect the instantaneous changes, spectral distribution, and time-frequency local patterns of the signal, thus providing a more comprehensive and robust input for regression prediction. Finally, the fused features are fed into a fully connected layer, and continuous value prediction of bone injury depth is achieved through nonlinear mapping. Figure 10 As shown, the overall structure of TB-CNN includes an input layer, three parallel convolutional branches, a feature fusion module, and an output layer, fully demonstrating the process of feature input, branch learning, feature integration, and regression prediction. This design not only effectively avoids the limitations of single-modality features in terms of expressive power, but also improves prediction accuracy and model generalization ability through multi-branch parallel learning and fusion, making it particularly suitable for bone injury detection tasks under small sample conditions.

[0057] During model training, this invention employs K-fold cross-validation, dividing the finite dataset into several non-overlapping subsets. Each time, one subset is selected as the validation set, and the remaining subsets are used as the training set, repeated K times to ensure all samples can be used for both training and validation. This method effectively avoids bias caused by imbalanced data partitioning and improves the model's generalization ability and stability. In terms of specific training strategies, this invention uses the Adam optimization algorithm to iteratively update network parameters. This algorithm achieves fast convergence at a low learning rate and is adaptable to the optimization process of non-stationary objective functions. To further prevent overfitting, this invention introduces Dropout technology in the fully connected layers, randomly discarding some neuron connections to reduce the model's over-reliance on local features. Simultaneously, an L2 regularization term is added to the weight updates to constrain parameter magnitudes and improve the model's generalization performance. During training, a mini-batch gradient descent strategy is used, with a batch size of 16 and 50 iterations. In each iteration, the network is progressively optimized through three stages: forward propagation, backpropagation of error, and parameter update. After multiple rounds of iterative training, the model's loss function gradually converged, and the prediction performance tended to stabilize, indicating that the training method of this invention can achieve robust optimization results under small sample conditions. Experimental results are as follows: Figure 11 As shown, the TB-CNN model proposed in this invention achieves significant advantages in bone injury prediction tasks. In regression prediction at different injury depths, the model's RMSE is only 0.0053, and the R² is as high as 0.9998, outperforming existing methods such as ResNet, SVR, MLP, and linear regression. Compared with these methods, the model of this invention can more accurately capture subtle differences in bone tissue signals caused by injury and maintain high stability and robustness under small sample conditions. This result fully verifies the superior performance of TB-CNN after fusing multimodal features, proving that it can still achieve high-precision predictions in complex environments and with limited sample sizes. In summary, this invention not only theoretically proposes an innovative bone injury detection method based on UGW and multi-branch convolutional neural networks, but also demonstrates excellent detection accuracy and generalization ability in experiments, showing great potential for widespread application in the early diagnosis of osteoporosis and fractures.

[0058] The ultrasound prediction method for bone injury proposed in this invention is significantly superior to existing methods such as ResNet, Support Vector Regression (SVR), Multilayer Perceptron (MLP), and Linear Regression in both root mean square error (RMSE) and coefficient of determination (R²).

[0059] In the comparative experiments, the proposed TB-CNN model was compared with ResNet, SVR, MLP, and linear regression models. The comparison metrics included RMSE and R², and the results are shown in the table.

[0060]

[0061] Analysis of the experimental results shows that the TB-CNN model of this invention achieves the best performance in both RMSE and R² metrics. Its RMSE is significantly lower than other models, and its R² is close to 1, indicating that the prediction results are highly consistent with the true values. In comparison, although ResNet performs better as a deep learning model, it is still inferior to the method of this invention. Furthermore, the prediction performance of traditional SVR, MLP, and linear regression declines significantly under small sample conditions, verifying the robustness and superiority of this invention in complex signal and small sample environments.

[0062] This method not only maintains extremely high prediction accuracy in small sample scenarios (RMSE reaches 0.0053, R² reaches 0.9998), but also demonstrates good robustness and application value in complex signal environments, providing an efficient and reliable solution for the early diagnosis of osteoporosis and fractures.

[0063] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0064] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0065] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the ultrasonic guided wave signal compensation methods for improving damage detection accuracy described in the above embodiments.

[0066] It is understood that the system provided in the embodiments of the present invention corresponds to the method provided in the embodiments of the present invention, and the explanation, examples and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.

[0067] This application also provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, communication interface, and memory communicate with each other via the communication bus. Memory, used to store computer programs; When the processor executes the program stored in the memory, it implements the ultrasonic guided wave signal compensation method described above for improving the accuracy of damage detection.

[0068] The communication bus mentioned in the aforementioned electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc.

[0069] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0070] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0071] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0072] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0073] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting bone injuries based on ultrasonic guided waves and multi-branch convolutional neural networks, characterized in that, This method first uses ultrasonic guided wave technology to acquire bone signals, obtain normal signals and damaged signals respectively, and then performs standardization processing on the acquired signals to remove noise interference and retain key damage-related information. In the feature extraction stage, multimodal features are extracted from the signal, including: multi-scale feature construction from three dimensions: time domain, frequency domain and time-frequency domain. Among them, time domain features are used to reflect the transient changes of bone signal, frequency domain features reveal the energy distribution law through fast Fourier transform, and time-frequency domain features generate composite time-frequency image based on multiple time-frequency domain transformations. After feature extraction is completed, a multi-branch convolutional neural network model is constructed. The model includes at least three independent branches: the first branch is a one-dimensional convolutional branch, which is used to process time-domain features; the second branch is a one-dimensional convolutional branch, which is used to process frequency-domain features; and the third branch is a two-dimensional convolutional branch, which is used to process the composite time-frequency image. Each branch extracts deep features, and the multimodal features are spliced ​​and integrated in the fusion layer; The fused features are input into a fully connected layer for regression prediction, and the output is a continuous numerical result of the bone injury depth.

2. The bone injury detection method based on ultrasonic guided waves and multi-branch convolutional neural networks according to claim 1, characterized in that, Specifically, it includes the following steps: Step 1: Use a UGW acquisition device to acquire raw sample signals of the bone, including normal signals and cutting damage signals, and convert the signals into digital format; Step 2: Perform Z-score normalization on the signals acquired in Step 1; Step 3: Divide the processed signal from Step 2 into a test set and a training set. Ensure that the test set does not participate in any augmentation operations and retains its original distribution for model performance evaluation. Perform Gaussian noise perturbation and amplitude scaling on the training set for data augmentation. Step four: After data augmentation in step three, the time-domain signal is converted into the frequency domain through a fast Fourier transform. Various time-frequency analysis methods are introduced to transform the time-domain signal into a two-dimensional time-frequency graph. Step 5: The signals transformed in Step 4 are fed into a multi-branch convolutional neural network model to automatically extract features, namely time-domain features, frequency-domain features, and time-frequency-domain features. Deep-level features are then extracted through convolution and pooling operations. Step 6: In the fusion module, the features from the three branches in Step 5 are concatenated and integrated, and then input into the fully connected layer for training. The Adam optimization algorithm is used to iteratively update the parameters, while Dropout and L2 regularization are introduced to prevent overfitting. Step 7: Train and validate the multi-branch convolutional neural network model from Step 6 using mini-batch gradient descent and K-fold cross-validation to obtain stable prediction performance; Step 8: Use the multi-branch convolutional neural network model trained and optimized in Step 7 to predict the bone signal and output the regression result of the damage depth.

3. The bone injury detection method based on ultrasonic guided waves and multi-branch convolutional neural networks according to claim 2, characterized in that, In step two, the signal acquired in step one is subjected to Z-score normalization preprocessing. The formula for normalization preprocessing is as follows: Where M represents the signal before standardization. The length of the data, after Z-score processing, yields data with a mean of 0 and a standard deviation of 1. .

4. The bone injury detection method based on ultrasonic guided waves and multi-branch convolutional neural networks according to claim 2, characterized in that, In step five, after the signal conversion in step four is completed, the converted signal is sent into the convolutional neural network model, and the convolutional neural network model automatically extracts features. The convolutional neural network includes a data input layer, a convolutional layer, an activation layer, a pooling layer, a concatenation layer, a fully connected layer, and an output layer; after extracting time-domain, frequency-domain, and time-frequency plot features by the convolutional neural network, the features of different modalities are deeply fused.

5. The bone injury detection method based on ultrasonic guided waves and multi-branch convolutional neural networks according to claim 3, characterized in that, The Gaussian noise perturbation and amplitude scaling enhancement method includes: The Gaussian noise perturbation enhancement method is used to simulate random perturbations that may occur in actual signal acquisition, and will [improve / enhance] the [signal acquisition process]. Add Gaussian noise to generate enhanced samples; Gaussian noise Follows a mean of 0 and a standard deviation of Normal distribution: Enhanced signal Represented as: in, Used to control the intensity of noise, which can be adjusted according to actual needs; The amplitude scaling enhancement method multiplies each sampling point of the signal by a random perturbation factor. Determined by Gaussian distribution , Let the variance of the normal distribution be denoted by , and let the original signal be . Thus, the perturbed signal is constructed. .

6. The bone injury detection method based on ultrasonic guided waves and multi-branch convolutional neural networks according to claim 5, characterized in that, The augmented samples generated after step three, plus the previous test set samples, make up the total sample dataset. , Time series data is currently a time-domain signal. The time-domain signal is converted to the frequency domain, and the frequency domain representation of each sample is calculated by FFT. A frequency axis is generated for subsequent visualization. After the conversion from time-domain signal to frequency domain is completed, time-frequency analysis is carried out.

7. The bone injury detection method based on ultrasonic guided waves and multi-branch convolutional neural networks according to claim 6, characterized in that, The formula for converting the time-domain signal to the frequency domain is as follows: In the formula The complex frequency domain value corresponding to the k-th frequency point. Let N represent the time-domain signal at the nth sampling point, where N is the total length of the signal. To improve the efficiency of FFT calculation, each signal is first padded with zeros to the nearest zero. The points are then normalized; the resulting complex spectrum retains only its amplitude spectrum. Due to the original signal Since it is a real-valued signal, its spectrum has conjugate symmetry, therefore only the first part is retained. The point is used as a one-sided amplitude spectrum and multiplied by 2 to maintain energy consistency; the final frequency axis is constructed using the following formula: in, Indicates the sampling frequency. , This represents the frequency value corresponding to the k-th frequency point.

8. A bone injury detection method based on ultrasonic guided waves and multi-branch convolutional neural networks as described in any one of claims 1 to 7, characterized in that, The composite time-frequency image is composed of six time-frequency transformations—WVD, STFT, TRT, CWT, SLT, and SST—in a fixed order. The time-frequency transformation includes: Wigner-Ville distribution (WVD); Short-time Fourier transform (STFT); Time redistribution transform based on STFT (TRT); Continuous wavelet transform (CWT); Spectral local transform under multi-scale Morlet wavelet (SLT); Synchronous compression transform (SST). Each method generates a corresponding two-dimensional time-frequency plot, forming a composite time-frequency image; WVD augmentation of time-domain signals Time-frequency analysis is performed to capture the fine structure of the signal in the time-frequency domain, based on the fact that WVD exhibits a conjugate symmetric structure on the frequency axis. The WVD of a signal can be represented as: Where n is the time index, representing a discrete time point; f is the frequency index, representing a discrete frequency point; and m is a time delay variable used to calculate the product of the signal at different time points. yes The complex conjugate; STFT processes signals in the time domain. Apply a window function of length L. This process yields a short signal segment centered at time m. By dividing the signal into a series of short time intervals, the local frequency characteristics of the signal at each time position can be approximated. For each time m, the window function is slid to that position, and the discrete Fourier transform of the windowed signal is calculated to obtain the spectral information at that time. Its mathematical definition is as follows: ,in It is a signal with data augmentation. Let f be a window function centered at time position m, where f represents the frequency index and N is the total length of the signal; TRT improves the resolution of time-frequency representation by reallocating the energy of the STFT. Based on the STFT, it searches for local maxima along the time direction and redistributes energy to these local maxima points, thereby generating a sparser and more concentrated time-frequency representation. It identifies local maxima points in the STFT matrix; if a point... satisfy: and If it is a local maximum point, then it is determined to be a local maximum point; construct the TRT matrix. Only retain the STFT coefficients at local maxima, and set the coefficients at other locations to zero: if If it is a local maximum point, then for In other cases, It is zero; CWT convolves the signal with a set of wavelet basis functions generated by scaling and translating the mother wavelet function, thereby simultaneously locating signal features in the time and frequency domains. By selecting appropriate wavelet basis functions to match signal features, efficient signal analysis is achieved. The complex Morlet wavelet is used as the mother wavelet function, which has good localization properties in the frequency domain and can effectively capture the frequency information of the signal. The calculation formula for CWT is as follows: The mother wavelet is a complex Morlet wavelet, and the calculation formula is as follows: , It is the input signal. It is a discrete-time mother wavelet function. It is its complex conjugate, where a is the scale parameter and b is the translation parameter. It is a scaling factor used to maintain energy normalization. It is the center angular frequency of the complex Morlet wavelet. By calculating the absolute values ​​of the wavelet coefficients, the time-frequency energy distribution of the signal was obtained; SLT enhances frequency resolution by selecting appropriate wavelet scale ranges and wavelet functions: Where k is the preset wavelet scale number, and for complex Morlet wavelets, the scale is... and the corresponding center frequency There is an approximate relationship: Ultimately Visualize the signal on the frequency and time axes to form a two-dimensional time-frequency plot, which can intuitively reflect the distribution characteristics of signal energy in the time-frequency domain; SST is based on CWT. After obtaining the time-frequency diagram, the frequency axis is redistributed to improve the frequency resolution. When CWT processes multi-component signals, frequency diffusion can lead to fuzzy frequency positioning. However, after calculating the wavelet transform, SST estimates the instantaneous frequency through phase information and recompresses the frequency energy to its actual frequency position, thus obtaining a clearer and sparser time-frequency representation. First, the CWT of the signal is calculated. Then, the corresponding instantaneous frequency estimate is calculated using the CWT coefficients. Next, the energy of the CWT is redistributed in the frequency direction based on the instantaneous frequency estimate, thus obtaining the time-frequency representation of synchronous compression. Synchronous compression is defined as... Its discrete synchronous compression redistribution formula is: and ,in, express phase angle, Here, 'b' is the scale, 'k' is the translation parameter, and 'K' is the number of scales (32). This indicates that energy will be redistributed to the instantaneous frequency. It is the k-th scale The wavelet coefficients at time point b.

9. A computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the steps of the method as claimed in any one of claims 1 to 8.

10. A computer device comprising a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method as claimed in any one of claims 1 to 8.

Images