Internet of Things voice signal matching and recognition method

Abstract

The invention provides an Internet of Things voice signal matching and recognition method, which comprises the steps of analyzing voice features according to voice feature vectors in combination with a distance and angle position relationship between a sporter and a wall, establishing attenuation and interference models of voice signals under wall shielding, and obtaining an influence rule of wall shielding on voice recognition accuracy; if the voice recognition result is that the sporter sends a help-seeking instruction, an emergency rescue plan is triggered, alarm information is sent to a rescue center, an optimal rescue route is planned according to the position information of the sporter, and rescue workers are guided to quickly arrive at an accident site so as to reduce accidental injuries to the maximum extent; and continuously optimizing the wall shielding voice recognition model by adopting an incremental learning method, and dynamically updating the acoustic characteristic database and the voice attenuation interference model according to the recognition accuracy and the false alarm rate.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method for matching and recognizing voice signals of an Internet of Things. Background Art

[0002] In large-scale indoor sports scenes, the speech characteristics of athletes will change significantly when they exercise vigorously. Rapid breathing and physical exertion cause changes in the fundamental frequency, resonance peak position, energy distribution, etc. of the speech pattern, making feature extraction and pattern matching of speech recognition more difficult. At the same time, the wall material of the sports venue also has complex reflection and interference effects on the sound, long reverberation time, and multi-path effects, which seriously affect the speech clarity and recognition accuracy. The dynamic range of speech signals varies greatly and the non-stationarity is strong. The coupling effect of the above factors makes robust speech recognition in noisy indoor sports environments face huge technical challenges. However, both the dynamic changes of speech features and the complexity of environmental noise make this process full of challenges. In other words, at present, in large-scale indoor sports scenes, how to deal with the changes in athlete's speech characteristics and the interference of environmental noise while ensuring the accuracy of speech recognition is a technical problem that needs to be solved urgently. Summary of the invention

[0003] The present invention provides a method for matching and recognizing voice signals of an Internet of Things, which mainly includes:

[0004] Acquire sound reflection data of different wall materials, thicknesses, shapes, surface roughness, internal structures, hole distributions, attachments, and temperatures. Through acoustic modeling and simulation analysis, determine the effects of various wall properties on speech signal frequency, amplitude, duration, energy, spectral characteristics, resonance frequency, echo characteristics, and signal-to-noise ratio, and form a database of wall acoustic characteristics.

[0005] According to the wall acoustic characteristics database, the speech data of the athlete at different distances from the wall during exercise is obtained as speech samples, and a speech feature data set is constructed to extract the frequency, amplitude, duration, energy, spectral characteristics, resonance frequency, echo characteristics and signal-to-noise ratio parameters of the speech signal to obtain the speech feature vectors at different distances from the wall.

[0006] According to the speech feature vector, combined with the distance and angle position relationship between the athlete and the wall, the speech features are analyzed, and the attenuation and interference model of the speech signal under the occlusion of the wall is established, and the influence of the wall occlusion on the accuracy of speech recognition is obtained;

[0007] According to the wall material, thickness, shape, surface roughness, internal structure, hole distribution, attachments and temperature properties, the corresponding acoustic parameters are selected from the wall acoustic characteristics database, and combined with the voice signal attenuation and interference model to build a wall occlusion speech recognition model for accurately recognizing voice commands in complex wall environments;

[0008] Preprocess the speech signal in the wall-occluded speech recognition model to extract the frequency, amplitude, duration, energy, spectral characteristics, resonance frequency, echo characteristics and signal-to-noise ratio characteristic parameters of the speech signal, and optimize the feature subset to reduce the interference of the wall acoustic characteristics on speech recognition;

[0009] During the exercise, the voice data and position information of the athlete are obtained in real time and input into the wall-occluded speech recognition model. According to the distance and angle position relationship between the athlete and the wall, the acoustic parameters and speech feature weights of the wall-occluded speech recognition model are dynamically adjusted to adaptively optimize the speech recognition strategy and obtain the speech recognition result.

[0010] If the voice recognition result is that the athlete sends a distress command, the emergency rescue plan will be triggered, an alarm message will be sent to the rescue center, and the optimal rescue route will be planned based on the athlete's location information to guide rescuers to quickly reach the accident scene to minimize accidental injuries;

[0011] An incremental learning method is used to continuously optimize the wall-occluded speech recognition model, and the acoustic characteristic database and speech attenuation interference model are dynamically updated according to the recognition accuracy and false alarm rate.

[0012] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0013] The present invention discloses a method for matching and recognizing voice signals in the Internet of Things. The method establishes a database of wall acoustic characteristics, collects voice samples at different distances, constructs a voice feature data set, analyzes the impact of wall occlusion on voice, and establishes a voice signal attenuation interference model. Combining wall properties and acoustic parameters, a wall occlusion voice recognition model is constructed and trained. During the movement, voice data and position information are acquired in real time, model parameters are adjusted dynamically, and recognition strategies are adaptively optimized. When a distress command is recognized, a rescue plan is triggered. The present invention also uses an incremental learning algorithm to continuously optimize the model, improve the accuracy of voice recognition in complex environments, provide reliable technical support for rapid rescue in emergency situations, and effectively reduce accidental injuries. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 The present invention is a flow chart of a method for matching and recognizing voice signals of the Internet of Things.

[0015] Figure 2It is a schematic diagram of a method for matching and recognizing voice signals of the Internet of Things of the present invention.

[0016] Figure 3 This is another schematic diagram of an Internet of Things voice signal matching and recognition method of the present invention. DETAILED DESCRIPTION

[0017] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this specification.

[0018] like Figure 1-3 In this embodiment, a method for matching and identifying voice signals in the Internet of Things may specifically include:

[0019] S101. Obtain sound reflection data of walls with different materials, thicknesses, shapes, surface roughness, internal structures, hole distributions, attachments and temperatures. Through acoustic modeling and simulation analysis, determine the effects of various wall properties on speech signal frequency, amplitude, duration, energy, spectral characteristics, resonance frequency, echo characteristics and signal-to-noise ratio, and form a database of wall acoustic characteristics.

[0020] A laser scanner is used to perform three-dimensional scanning and modeling of the wall surface to obtain the wall surface roughness data; based on the wall surface roughness data, the wall temperature distribution map is collected, and the temperature gradient distribution inside the wall is calculated; the internal structure image of the wall is obtained from the X-ray tomography equipment, the hole distribution and attachment conditions are analyzed, and a three-dimensional model of the wall internal structure is generated; based on the three-dimensional wall surface model, the internal structure model and temperature distribution data, an acoustic simulation model is constructed using finite element analysis software; the sound ray tracing method is used to simulate the propagation path of sound waves within a specified frequency range in the wall, the sound wave energy attenuation and reflection characteristics are calculated, and the wall acoustic response function is obtained; a microphone array is used to collect reflected sound signals at different positions around the wall, and the spectral characteristics of the signal are extracted; the arrival time and direction of each frequency component are determined, and the independent component analysis method is used for blind source separation to eliminate the influence of multiple reflections and reverberation, and to construct a measured sound reflection characteristic data set.

[0021] Exemplarily, a laser scanner is used to perform three-dimensional scanning and modeling of the wall surface to obtain the roughness data of the wall surface. At the same time, infrared thermal imaging technology is used to collect the wall temperature distribution map, and the temperature gradient distribution inside the wall is calculated by combining with the pre-established material thermal property parameter database. The internal structure image of the wall is obtained from the X-ray tomography device, and the hole distribution and attachment conditions are analyzed to generate a three-dimensional model of the internal structure of the wall. According to the three-dimensional model of the wall surface, the internal structure model and the temperature distribution data, an acoustic simulation model is constructed by finite element analysis software, and the sound ray tracing method is used to simulate the propagation path of sound waves in the wall within the frequency range of 20Hz to 20kHz. The finite difference time domain method is used to calculate the attenuation coefficient and reflection coefficient of sound waves propagating in the wall at different frequencies. The attenuation coefficient and reflection coefficient at each frequency are substituted into the acoustic response function model to obtain the acoustic frequency response function of the wall.

[0022]

[0023] This formula is the propagation equation of sound waves in a uniform medium, where p represents the sound pressure, t represents the time, and c represents the speed of sound. is the Laplace operator, which represents the second derivative of space. The microphone array is used to collect the reflected sound signals at different positions around the wall. The spectral characteristics of the signal are extracted by Fourier transform. The arrival time and direction of each frequency component are determined by combining the delay estimation method and the direction estimation method. The independent component analysis is used for blind source separation to eliminate the influence of multiple reflections and reverberation. The measured sound reflection characteristic data set is constructed, including obtaining the spectral characteristics of the signal in different time periods by performing short-time Fourier transform on the sound signal. According to the phase difference of different frequency components in the spectral characteristics, the generalized cross-correlation method is used to estimate the arrival delay of each frequency component. According to the amplitude difference of different frequency components in the spectral characteristics, the multiple signal classification algorithm is used to estimate the arrival direction of each frequency component. It is judged whether the delay and direction of the frequency component meet the geometric relationship between direct sound and reflected sound. If so, the frequency component is considered to be direct sound, otherwise it is considered to be reflected sound. For the frequency components judged as reflected sound, the independent component analysis method is used for blind source separation to separate the reverberation signal into signals of different reflection sources. For the separated reflected sound signals, the corresponding reflection surface position and reflection characteristic parameters are determined according to their delay and direction information. The spectrum, delay, direction, reflection characteristics and other information of the direct sound signal and each reflected sound signal are annotated to construct a measured data set of acoustic reflection characteristics. The simulation model prediction results and the measured data are compared, and the multivariate regression analysis is used to preliminarily evaluate the influence of each wall attribute on the acoustic characteristics, and the initial parameter estimation value is established. The Markov chain Monte Carlo algorithm is used for Bayesian parameter estimation to optimize the acoustic model parameters and improve the prediction accuracy of the model for the acoustic characteristics of the wall. For different wall materials, thickness, shape, surface roughness, internal structure, hole distribution, attachments and temperature conditions, a wall acoustic characteristic database containing parameters such as speech signal frequency, amplitude, duration, energy, spectral characteristics, resonance frequency, echo characteristics and signal-to-noise ratio is generated. The laser scanner scans the 1m×1m wall sample with an accuracy of 0.1mm to obtain the surface three-dimensional point cloud data, and the surface roughness Ra value is calculated to be 5μm. The infrared thermal imager collected the temperature distribution map of the wall surface with a resolution of 0.1℃. Combined with the thermal conductivity of lime mortar in the pre-established material thermal property parameter database of 0.93W / (m·K), the temperature gradient inside the wall was calculated to be 0.5℃ / cm through the heat conduction equation. The X-ray tomography device obtained the internal structure image of the wall with a resolution of 0.5mm, and identified the bubble holes with a diameter of 2-5mm, a porosity of 3%, and 2cm×2cm metal attachments. Based on the acquired data, the finite element analysis software COMSOL Multiphysics was used to build an acoustic simulation model, and the ray tracing method was used to simulate the propagation of sound waves in the frequency range of 20Hz to 20kHz, with a sampling point every 100Hz, for a total of 201 frequency points.The finite difference time domain method is used, with a grid size of 1 / 10 of the wavelength and a time step of 1μs, to calculate the energy attenuation and reflection characteristics of the sound wave and obtain the acoustic response function of the wall. 16 omnidirectional microphones are arranged to form a 4×4 array with a spacing of 20cm. The reflected sound signals are collected at 1m, 2m, and 3m from the wall. The sampling rate is 48kHz and the sampling time is 5s. The collected signals are subjected to a 1024-point fast Fourier transform to extract the spectral features. The general cross correlation (GCC-PHAT) algorithm is used to estimate the arrival time of the sound wave with an accuracy of 0.1ms. The MUS IC algorithm is used to estimate the direction of the sound source with an azimuth resolution of 1°. The Fast I CA algorithm is used for independent component analysis to achieve blind source separation, eliminate the effects of multiple reflections and reverberation, and reconstruct the original sound source signal. A data set of measured sound reflection characteristics is constructed, including parameters such as reflection coefficient and phase delay at each frequency point. Through multiple linear regression analysis, it is preliminarily concluded that for every 1cm increase in wall thickness, the reflection coefficient of the intermediate frequency 1kHz increases by 0.02, and for every 1μm increase in surface roughness, the reflection coefficient of the high frequency 10kHz decreases by 0.01. The Metropolis-Hastings algorithm is used to implement Markov chain Monte Carlo sampling, and Bayesian parameter estimation is performed. After 10,000 iterations, the optimized acoustic model parameters are obtained. Finally, a database of wall acoustic characteristics is generated, including acoustic parameters such as the frequency response curve of speech signals in the 100Hz-8kHz frequency band, reverberation time RT60, early decay time EDT, and speech transmission index STI under different materials (brick, concrete, gypsum board, etc.), thickness (5-30cm), shape (plane, concave and convex surface), surface roughness 1-100μm, internal structure (homogeneous, porous), hole distribution 0-10%, attachments (metal, wood) and temperature 0-40℃.

[0024] S102. According to the wall acoustic characteristics database, the voice data of the athlete at different distances from the wall during exercise are obtained as voice samples, and a voice feature data set is constructed to extract the frequency, amplitude, duration, energy, spectral characteristics, resonance frequency, echo characteristics and signal-to-noise ratio parameters of the voice signal to obtain the voice feature vectors at different distances from the wall.

[0025] A plurality of omnidirectional microphone arrays are used to track the position of the athlete in real time; the change in the distance between the athlete and the wall is recorded according to the position of the athlete, and the athlete's voice data is collected synchronously to obtain an original voice sample set containing distance information; a preliminary quality assessment is performed on the original voice sample set, the signal-to-noise ratio is calculated and a spectrum analysis is performed, and valid voice samples are screened out according to a preset signal-to-noise ratio threshold; pre-processing is performed on the valid voice samples to remove environmental low-frequency noise, eliminate echo interference, extract valid voice segments, and obtain clear voice signals; feature parameters are extracted from the clear voice signal, the spectral features of the voice are calculated, the resonance peak frequency is estimated, the fundamental frequency period is extracted, the energy envelope is calculated, and the frequency, amplitude, duration, energy and other feature parameters of the voice are obtained; for the voice samples collected at different distances, the signal-to-noise ratio and echo characteristic parameters are calculated, and a voice feature vector containing distance information is constructed in combination with the voice feature parameters.

[0026] Exemplarily, based on the database of wall acoustic characteristics, multiple omnidirectional microphone arrays are arranged in a laboratory environment, and the position of the athlete is tracked in real time by triangulation, the distance change between the athlete and the wall is recorded, and the athlete's voice data is collected synchronously to form an original voice sample set containing distance information. A preliminary quality assessment is performed on the collected voice samples, the signal-to-noise ratio is calculated and the spectrum analysis is performed, the signal-to-noise ratio threshold is set to 10dB, and valid samples are screened out. The screened voice samples are preprocessed by removing environmental low-frequency noise through a high-pass filter with a cutoff frequency of 100Hz, eliminating echo interference using a nonlinear acoustic echo cancellation method, and extracting valid voice segments using the G.729B voice activity detection algorithm to obtain a clear voice signal. Feature parameters were extracted from the preprocessed speech signal, and the spectral features of the speech were calculated using short-time Fourier transform. The linear prediction analysis was performed through the Levinson-Durbin recursive algorithm to estimate the formant frequency. The FFT-based method was used to perform cepstrum analysis to extract the fundamental period, and the energy envelope of the speech signal was calculated. At the same time, the short-time zero-crossing rate and short-time energy were used to deal with the Doppler effect caused by movement, and the characteristic parameters of the speech, such as frequency, amplitude, duration, and energy, were obtained. For the speech samples collected at different distances, the signal-to-noise ratio and echo characteristic parameters were calculated. Combined with the previously extracted speech features, a speech feature vector containing distance information was constructed. Nonlinear dimensionality reduction was performed through kernel principal component analysis, and finally a low-dimensional representation reflecting the speech features at different distances from the wall was obtained. Polynomial regression was used to analyze the changing trend of speech features at different distances, and a mapping relationship between distance and speech features was established. 16 omnidirectional microphones were arranged in a 10m×10m laboratory to form a 4×4 array with a spacing of 2m. The position of the mover was tracked using triangulation with an accuracy of ±5cm. The participants were kept at different distances from the wall in the range of 0.5m to 5m, and a position point was recorded every 0.5m, for a total of 10 sampling positions. 5 seconds of speech was collected at each position, with a sampling rate of 44.1kHz and 16-bit quantization. The collected speech samples were preliminarily evaluated for quality, and the power spectral density was calculated using the Welch method, with 1024-point FFT, 50% overlap, and Hanning window. The signal-to-noise ratio was calculated, and a threshold of 10dB was set to filter out valid samples. A 4th-order Butterworth high-pass filter with a cutoff frequency of 100Hz was applied to the filtered samples to remove low-frequency noise. A frequency-shifted adaptive filtering algorithm was used for nonlinear acoustic echo cancellation, and the convergence step was set to 0.01. The G.729B algorithm was used for voice activity detection, with a frame length of 20ms and a frame shift of 10ms. A 512-point FFT was used for short-time Fourier transform, with a frame length of 25ms and a frame shift of 10ms, and the spectral characteristics of the speech were calculated. Linear prediction analysis was performed using a 10th-order Levinson-Durbin recursive algorithm to estimate the first three formant frequencies.The FFT-based method is used for cepstrum analysis to extract the fundamental frequency period, with a search range of 50-500Hz. The short-time energy and short-time zero-crossing rate are calculated with a frame length of 20ms to process the Doppler effect, and the thresholds are set to 0.015 and 0.1 respectively. For 10 different distance positions, 20-dimensional MFCC features are extracted, and a 40-dimensional feature vector is constructed by combining the previously obtained parameters such as frequency, amplitude, duration, and energy. Nonlinear dimensionality reduction is performed using the kernel principal component analysis of the radial basis function kernel, retaining the principal components that explain 95% of the variance, and obtaining a low-dimensional representation of about 15 dimensions. Finally, a third-order polynomial regression is used to analyze the changing trend of speech features at different distances, establish a mapping relationship between distance and speech features, and establish the determination coefficient R of the regression model. 2 Reached 0.85.

[0027] S103. Analyze the speech features according to the speech feature vector and the distance and angle position relationship between the athlete and the wall, establish an attenuation and interference model of the speech signal under wall occlusion, and obtain the influence of wall occlusion on the speech recognition accuracy.

[0028] The information of the position of the mover, the position of the wall and the voice propagation path is obtained, and the information is mapped to a pre-constructed three-dimensional coordinate system; based on the mapping information in the three-dimensional coordinate system, the propagation path length and the incident angle of the voice signal in space are calculated; the acoustic characteristic parameters of the wall material are measured using a material detector, and the acoustic characteristic parameters include the sound absorption coefficient and the scattering coefficient; based on the acoustic characteristic parameters, the propagation path length and the incident angle, an attenuation model of the voice signal in the process of spatial propagation is established; based on the multipath propagation theory, the time delay and phase difference between the direct sound and each reflected sound are analyzed, and an acoustic interference model under the condition of wall occlusion is established; using the attenuation model and the interference model, the original voice feature vector is transformed to obtain the voice feature vector after being blocked by the wall; the transformed voice feature vector is input into the pre-trained voice recognizer to obtain the recognition result; the recognition result is compared with the original voice content, and the word error rate is calculated; based on the word error rate, the influence of wall occlusion on the accuracy of voice recognition is determined.

[0029] For example, a three-dimensional coordinate system is constructed based on the voice feature vector and the distance and angular position relationship between the mover and the wall. The mover position, wall position and voice propagation path are mapped to the coordinate system. The Monte Carlo ray tracing algorithm is used to simulate the propagation path of the voice signal in space, and the path length and incident angle of the direct sound, reflected sound and diffracted sound are calculated. For complex wall structures, the triangular patch decomposition method is used to process irregular reflections. The acoustic characteristic parameters of the wall material, including the sound absorption coefficient and the scattering coefficient, are obtained using a material detector. The sound energy attenuation on each propagation path is calculated in combination with the incident angle and frequency. Considering air absorption, geometric divergence and surface reflection loss, the frequency dependence of air absorption is calculated using the ISO9613-1 standard, and an attenuation model of the voice signal in the process of spatial propagation is established. The actual sound pressure level is measured at different positions using a sound pressure level meter, and compared with the predicted results of the attenuation model to verify the accuracy of the model. Based on the multipath propagation theory, the time delay and phase difference between the direct sound and each reflected sound are analyzed, the frequency response function is calculated using the fast Fourier transform, and the influence of the comb filter effect and the Doppler effect on the speech signal is considered to establish an acoustic interference model under the condition of wall occlusion. Using the established attenuation model and interference model, the original speech feature vector is transformed to obtain the speech features after wall occlusion. The transformed features are input into the pre-trained DeepSpeech speech recognizer. By comparing the recognition results with the original speech content, the word error rate is calculated as a quantitative indicator of accuracy, and the recognition accuracy at different distances and angles is statistically analyzed. The differences in recognition accuracy of vowels, consonants and syllables are analyzed, and the confusion matrix is ​​used to evaluate the recognition performance. The influence of wall occlusion on the recognition accuracy of different types of speech content is obtained by multivariate regression analysis. In a 10m×10m×3m laboratory, a three-dimensional coordinate system is established, the position of the mover is set as the origin (0,0,0), and the wall is located at 5m on the x-axis. Using the Monte Carlo ray tracing algorithm, 10,000 rays were emitted, and each ray was calculated for a maximum of 5 reflections. For a complex wall structure of 0.5m×0.5m, the triangular patch decomposition method was used to decompose it into 200 small triangles. The sound absorption coefficient of the wall in the 125Hz-4kHz frequency band was measured using a material detector, and the average value was 0.3 and the scattering coefficient was 0.2. According to the ISO9613-1 standard, at 20℃ and 50% relative humidity, the attenuation coefficient of 500Hz sound waves in air was calculated to be 2.8×10^-3dB / m. After establishing the attenuation model, a sound pressure level meter was placed at 1m, 3m, and 5m away from the wall to measure the sound pressure level of a 1kHz pure tone, and compared with the model prediction value, the error was controlled within ±1dB. The frequency response function was calculated using a 1024-point FFT with a sampling rate of 44.1kHz, and the comb filter effect in the range of 0-22.05kHz was analyzed. Considering the Doppler effect when a person moves at a speed of 1 m / s, the maximum frequency shift is about 3 Hz.After the original speech feature vector is transformed by attenuation and interference model, it is input into the pre-trained DeepSpeech speech recognizer. The recognizer is trained based on 5000 hours of Chinese speech data, with 5 hidden layers and 1024 neurons in each layer. The recognition results are compared with the original speech content, and the word error rate (WER) is calculated. At the distances of 1m, 3m, and 5m, the average WER of 10%, 15%, and 25% is obtained respectively. Through confusion matrix analysis, it is found that the accuracy of vowel recognition is 10% higher than that of consonants, and the error rate of monosyllabic words is 5% higher than that of polysyllabic words. Using multiple regression analysis, it is found that the linear relationship, R, is that for every 1m increase in distance, the WER increases by 3%. 2 The value is 0.92.

[0030] S104. According to the wall material, thickness, shape, surface roughness, internal structure, hole distribution, attachments and temperature properties, corresponding acoustic parameters are selected from the wall acoustic characteristic database, and combined with the speech signal attenuation and interference model to construct a wall occlusion speech recognition model for accurately recognizing voice commands in complex wall environments.

[0031] Acoustic parameters are obtained from the wall acoustic characteristic database, including sound absorption coefficient, scattering coefficient, transmission loss and acoustic impedance. The missing values ​​in the acoustic parameters are supplemented by the Kriging interpolation method to obtain a complete set of wall acoustic characteristic parameters. According to the wall acoustic characteristic parameter set, the propagation path of the voice signal in a complex wall environment is simulated; if multiple reflections and diffraction effects are detected, the sound wave diffraction of the irregular surface is processed. The frequency response function of the voice signal at the receiving point is calculated through the propagation path model; it is determined whether there is a comb filter effect and a Doppler effect. If so, a harmonic distortion model is introduced to process the nonlinear acoustic effect; a time-frequency characteristic transformation model of the voice signal is constructed, and the original voice features are transformed to obtain the voice features after being blocked by the wall. The signal-to-noise ratio of the transformed voice features is calculated to determine the threshold of feature correlation analysis; the effective features are screened according to the threshold to generate voice data under the blocked environment. The screened voice features are input into the pre-trained end-to-end speech recognition model based on Transformer; the difference in recognition accuracy of different types of voice commands is determined to obtain the voice recognition accuracy of the model in a complex wall environment.

[0032] Exemplarily, according to the wall material, thickness, shape, surface roughness, internal structure, hole distribution, attachments and temperature properties, the corresponding acoustic parameters, including sound absorption coefficient, scattering coefficient, transmission loss and acoustic impedance, are extracted from the wall acoustic characteristic database, and the missing parameters are supplemented by the Kriging interpolation method to construct a complete wall acoustic characteristic parameter set. A temperature sensor is used to monitor the wall temperature change in real time and dynamically update the acoustic parameters. The frequency-dependent ray tracing method is used to simulate the propagation path of the voice signal in a complex wall environment. Combined with the wall acoustic characteristic parameter set, the energy attenuation and delay on each propagation path are calculated, considering multiple reflections and diffraction effects, and the Kirchhoff approximation method is used to process the sound wave diffraction on irregular surfaces to construct a spatial propagation model of the voice signal. Based on the spatial propagation model, combined with the indoor sound field theory, the frequency response function of the voice signal at the receiving point is calculated, the comb filter effect and the Doppler effect are considered, and the harmonic distortion model is introduced to deal with nonlinear acoustic effects. The time-frequency characteristic transformation model of the voice signal is constructed, and the original voice features are transformed to obtain the voice features after being blocked by the wall. The signal-to-noise ratio and feature correlation analysis of the transformed speech features are performed, and the threshold is set to filter out effective features. Data enhancement techniques are used, such as adding simulated reflection and attenuation effects, to generate speech data in an occluded environment. The filtered speech features are input into the end-to-end speech recognition model based on Transformer. The progressive transfer learning method is used to fine-tune the network parameters based on the pre-trained model using speech data in an occluded environment. The differences in recognition accuracy of different types of voice commands (short commands, long sentences) are analyzed, and the model performance is evaluated using perplexity to improve the model's speech recognition accuracy in complex wall environments. In a 10m×10m×3m laboratory, the acoustic material database is used to extract wall acoustic parameters, such as the sound absorption coefficient of 0.1-0.9, the scattering coefficient of 0.05-0.5, the transmission loss of 20-60dB, and the acoustic impedance of 2000-8000kg / m in the 125Hz-4kHz frequency band. 2 s. Kriging interpolation method is used, with 0.1m 3The grid density is used to complete the missing parameters. 10 temperature sensors are arranged, sampling every 30 seconds with an accuracy of ±0.1℃, and the acoustic parameters are updated dynamically. A frequency-dependent ray tracing algorithm is used to emit 1000 rays in the range of 100Hz-8kHz, with a frequency point of every 100Hz, and a maximum of 5 reflections are calculated. The Kirchhoff approximation method is used to process the sound wave diffraction of the 0.5m×0.5m irregular wall. A 16-order FIR filter is constructed to simulate the comb filtering effect, considering the Doppler frequency shift of ±10Hz. A 3rd-order harmonic distortion model is introduced, and the distortion coefficient is set to 0.05. The signal-to-noise ratio is calculated for the transformed speech features, and a threshold of 6dB is set to retain features with a signal-to-noise ratio greater than the threshold. Feature correlation analysis is used to eliminate redundant features with a correlation coefficient greater than 0.95. 10,000 speech data in an occluded environment are generated by adding random reflections of 0-10dB and random attenuation of 0-20dB. An end-to-end speech recognition model with a 12-layer Transformer structure, a hidden layer dimension of 512, and 8 attention heads. Progressive transfer learning is used, and the learning rate is gradually increased from 1e-5 to 1e-4. The difference in recognition accuracy between short commands of 1-3 seconds and long sentences of 5-10 seconds is analyzed, and the perplexity is used to evaluate the model performance. The perplexity of short commands and long sentences is controlled within 50 and 100 respectively. The final model has a speech recognition accuracy of 95% in a complex wall environment.

[0033] S105. Preprocess the speech signal in the wall-occluded speech recognition model to extract the frequency, amplitude, duration, energy, spectral characteristics, resonance frequency, echo characteristics and signal-to-noise ratio characteristic parameters of the speech signal, and optimize the feature subset to reduce the interference of the wall acoustic characteristics on speech recognition.

[0034] After preprocessing the input speech signal, voice activity detection is performed to obtain valid speech segments. Feature parameters are extracted from the valid speech segments, and the feature parameters include: calculating the spectrum features of the valid speech segments; performing linear prediction analysis through the Levinson-Durbin recursive algorithm to obtain the formant frequency of the valid speech segments; extracting the pitch period of the valid speech segments using the real cepstrum method; and calculating the energy envelope of the valid speech segments. Modeling the feature parameters includes: using a Gaussian mixture model to model the sound features of different speakers; calculating the signal-to-noise ratio and echo characteristic parameters; combining the spectrum features, formant frequency, pitch period and energy envelope to construct a high-dimensional feature vector. The high-dimensional feature vector is optimized, and the optimization includes: normalizing the high-dimensional feature vector using the Z-score standardization method; performing dimensionality reduction and correlation analysis on the normalized feature vector using principal component analysis; and compensating for the influence of the acoustic characteristics of the wall. Feature selection is performed on the feature vector after optimization to obtain an optimized feature vector.

[0035] Exemplarily, the input speech signal is preprocessed, and the low-frequency noise of the environment is removed by setting a high-pass filter with a cutoff frequency of 100Hz, and the echo interference is eliminated by using a minimum mean square error adaptive filter. The G.729B algorithm is used for voice activity detection, and effective voice segments are extracted to obtain a clear voice signal. Feature parameters are extracted from the preprocessed voice signal, and the spectral features of the voice are calculated using the short-time Fourier transform. The linear prediction analysis is performed by the Levinson-Durbin recursive algorithm to estimate the resonance peak frequency, and the pitch period is extracted by the real cepstrum method. The energy envelope of the speech signal is calculated to obtain the frequency, amplitude, duration, energy and other feature parameters of the voice. The sound features of different speakers are modeled using a Gaussian mixture model to improve the robustness of feature extraction. For the extracted feature parameters, the signal-to-noise ratio and echo characteristic parameters are calculated, and the high-dimensional feature vector is constructed by combining the previously obtained frequency, amplitude, duration, energy and other features. The feature parameters of different scales are normalized using the Z-score standardization method to ensure the comparability of the data of each dimension in the feature vector. The principal component analysis was used to perform preliminary dimension reduction and correlation analysis on the feature vector to remove redundant information. The inverse filtering method of the acoustic transfer function was used to compensate for the influence of the acoustic characteristics of the wall and reduce the interference of the wall occlusion on the speech features. The Lasso regression based on L1 regularization was used for feature selection. The performance of different feature subsets was evaluated by cross-validation. The cross-validation error increase was set to no more than 1% as the stopping criterion, and the optimal feature subset was selected to reduce the influence of redundant features on the recognition performance and obtain the optimized feature vector. In the laboratory environment, the speech signal with a sampling rate of 16kHz was preprocessed, and a 4th-order Butterworth high-pass filter was used with a cutoff frequency of 100Hz to remove environmental low-frequency noise. A 32nd-order minimum mean square error adaptive filter with a step size of 0.01 was used to eliminate echo interference. The improved G.729B algorithm was used with a frame length of 20ms and a frame shift of 10ms to perform voice activity detection and extract valid voice segments. The preprocessed signal was subjected to a 512-point short-time Fourier transform with a frame length of 25ms and a frame shift of 10ms to calculate the spectral characteristics of the speech. A 10th-order Levinson-Durbin recursive algorithm was used for linear prediction analysis to estimate the frequencies of the first three resonance peaks. The real cepstrum method was used with a window length of 30ms to extract the fundamental frequency period, and the search range was 50-500Hz. The short-time energy of a 25ms frame length was calculated to obtain the energy envelope. A 32-component Gaussian mixture model was used to model the voice features of 5 different speakers to improve the robustness of feature extraction. The signal-to-noise ratio was calculated, and a threshold of 6dB was set to retain features with a signal-to-noise ratio greater than the threshold. A high-dimensional vector containing 40-dimensional features was constructed and normalized using the Z-score normalization method. Principal component analysis was used for dimensionality reduction, retaining the principal components that explained 95% of the variance and reducing it to 15 dimensions.The acoustic transfer function inverse filtering technology based on the full-pole model is used with an order of 20 to compensate for the influence of the acoustic characteristics of the wall. Lasso regression is used for feature selection, and the regularization parameter λ starts from 0.1 and increases by 0.1 each iteration until the cross-validation error increases by more than 1%. Finally, the 10-dimensional optimal feature subset is selected to form the optimized feature vector.

[0036] S106. Acquire the athlete's voice data and position information in real time during the exercise process, and input them into the wall-occluded speech recognition model. Dynamically adjust the acoustic parameters and speech feature weights of the wall-occluded speech recognition model according to the distance and angular position relationship between the athlete and the wall, adaptively optimize the speech recognition strategy, and obtain the speech recognition result.

[0037] The three-dimensional position coordinates of the athlete are obtained. The three-dimensional position coordinates are calculated in real time by a multi-sensor fusion positioning technology using UWB, inertial sensors and cameras; the athlete's voice signal is collected according to the three-dimensional position coordinates. The voice signal is collected by a microphone array and subjected to signal preprocessing and feature extraction to obtain a voice feature vector; the distance and angle relationship between the athlete and the wall are calculated. The distance and angle relationship is determined by the three-dimensional position coordinates; the sound wave propagation path is simulated. The sound wave propagation path is calculated based on the distance and angle relationship; the voice feature vector is dynamically compensated. The dynamic compensation is performed based on an acoustic propagation model updated by the sound wave propagation path. The compensated voice feature vector is obtained by adjusting the frequency response, energy distribution and delay characteristics; the compensated voice feature vector is input into an adaptive speech recognizer. The adaptive speech recognizer uses a recursive least squares algorithm for online learning. The speech recognition result is obtained by updating the recognizer parameters.

[0038] Exemplarily, the multi-sensor fusion positioning technology of UWB, inertial sensor and camera is used to obtain the three-dimensional position coordinates of the athlete in real time, and the extended Kalman filter is used for data fusion. At the same time, the microphone array is used to collect the athlete's voice signal, and the segmentation and splicing of the continuous voice signal are processed by the sliding window method, and the signal preprocessing and feature extraction are performed to obtain the voice feature vector and position information. According to the real-time position coordinates of the athlete, the distance and angle relationship between the athlete and the wall are calculated, and the propagation path of sound waves of different frequencies is simulated by the frequency band ray tracing algorithm, and the multipath effect caused by the complex wall structure is processed. Combined with the wall acoustic characteristic database, the attenuation coefficient and reflection coefficient in the acoustic propagation model are updated. A sound pressure level meter is used to perform real-time measurements at key locations to verify the accuracy of the updated acoustic propagation model. Based on the updated acoustic propagation model, the voice feature vector is dynamically compensated, the frequency response, energy distribution and delay characteristics are adjusted, and the weight coefficients of each dimension of the voice feature are dynamically adjusted by using an exponential decay function based on distance and angle. An adaptive filter is used to compensate for the Doppler effect caused by different movement speeds and direction changes to improve the adaptability of the model to the motion state. The compensated speech feature vector is input into the adaptive speech recognizer, and the recursive least squares algorithm is used for online learning to update the recognizer parameters in real time. The dynamic time warping algorithm is used to balance the real-time performance and recognition accuracy. The decision threshold is dynamically adjusted according to the confidence of the recognition result, and the speech recognition result is finally output. In a 10m×10m×3m laboratory, 4 UWB base stations, 6 high-definition cameras and 1 inertial measurement unit are deployed. The extended Kalman filter algorithm is used to fuse multi-sensor data, and the positioning accuracy reaches ±5cm. 16 omnidirectional microphones are used to form a 4×4 array with a sampling rate of 48kHz. A 200ms sliding window and an overlap rate of 50% are used to process continuous speech. The ray tracing algorithm is used to simulate the propagation of sound waves in the range of 100Hz-8kHz. Each 100Hz is a frequency band, 1000 rays are emitted, and up to 5 reflections are calculated. The 15th-order full-pole model is used to construct the acoustic transfer function, and the model parameters are updated every 0.5s. A sound pressure level meter was placed at 1m, 2m, and 3m away from the wall to verify the accuracy of the model in real time, and the error was controlled within ±1dB. The exponential decay function w=exp(-0.1d-0.05θ) based on the distance d and the angle θ was used to dynamically adjust the feature weight. Exp represents the natural exponential function, which is an exponential function with the mathematical constant e (approximately equal to 2.71828) as the base. A 10th-order adaptive F IR filter with a step size of 0.01 was used to compensate for the Doppler frequency shift of ±10Hz. The adaptive speech recognizer adopts a 12-layer Transformer structure with a hidden layer dimension of 512 and 8 attention heads. The learning rate of the recursive least squares algorithm was set to 0.1 and the forgetting factor was 0.98. The dynamic time warping algorithm uses the Sakoe-Chiba band with a bandwidth of 10% of the sequence length.The initial value of the decision threshold is set to 0.8, and it is dynamically adjusted according to the average confidence of the previous 10 recognition results, with an adjustment step of 0.02. Finally, within the range of motion speed 0-2m / s and steering angle ±90°, the speech recognition accuracy remains above 95%.

[0039] S107. If the voice recognition result is that the athlete sends a distress command, the emergency rescue plan is triggered, an alarm message is sent to the rescue center, and the optimal rescue route is planned according to the athlete's location information to guide rescue personnel to quickly reach the accident scene to minimize accidental injuries.

[0040] The BERT model is used to perform semantic analysis on the speech recognition results, extract keywords and match them with the preset rescue command vocabulary to obtain the semantic matching score. According to the semantic matching score and emotional urgency score, it is determined whether the rescue command is triggered. If the rescue command is triggered, a three-dimensional scene model is constructed and updated in real time. According to the three-dimensional scene model, a detailed scene map containing information such as obstacles, passages, and entrances and exits is generated in combination with the geographic information system. The rescue command, the position of the athlete, and the scene map information are packaged and sent to the rescue center. After receiving the rescue command sent by the rescue center, the A* algorithm is used for multi-objective path planning. According to the multi-objective path planning results, multiple candidate rescue routes are calculated, and the optimal rescue route is selected from the multiple candidate rescue routes. The rescue effect data of the optimal rescue route is obtained, and a rescue effect evaluation system is established. According to the rescue effect evaluation system, the response time, arrival time, and rescue success rate indicators are recorded to obtain the key factor analysis results of the rescue process.

[0041] Exemplarily, the BERT model is used to perform semantic analysis on the speech recognition results, extract keywords and match them with the preset distress command word library, and use speech emotion recognition technology to analyze the urgency of the speech. If the comprehensive score of semantic matching and emotional urgency exceeds the preset threshold, it is determined to be a distress command and trigger the emergency rescue plan. According to the real-time location information of the athlete and the surrounding environment data, the laser radar SLAM technology is used to build and update the three-dimensional scene model in real time, and a detailed scene map containing information such as obstacles, channels and entrances and exits is generated in combination with the geographic information system. Through the MQTT protocol and 5G network, the distress command, athlete location, scene map and other information are packaged and sent to the rescue center using an encrypted transmission protocol, and the emergency equipment on the scene, such as sirens and emergency lighting, is activated at the same time. After receiving the information, the rescue center immediately parses and verifies the information, confirms the accident level and required resources, and dispatches the nearest rescue personnel and equipment. The improved A* algorithm is used for multi-objective path planning, taking into account the location of the rescue personnel, real-time traffic conditions and scene obstacles, and multiple candidate rescue routes are calculated. A comprehensive evaluation is performed based on factors such as time and safety to select the optimal rescue route. Through navigation, rescuers are guided in real time, while road conditions are continuously updated and routes are adjusted dynamically. Rescue effect evaluation is carried out, and indicators such as response time, arrival time, and rescue success rate are recorded. Data mining technology is used to analyze key factors in the rescue process, and rescue strategies and resource allocation are continuously optimized. In practical applications, the pre-trained BERT-base-ch i nese model is used to perform semantic analysis on speech recognition results, extract top-5 keywords, and calculate cosine similarity with a vocabulary containing 200 words related to help, with a threshold set to 0.8. At the same time, a speech emotion recognition model based on CNN-LSTM is used to divide speech into three categories: "calm", "anxious", and "fearful", and the urgency is assigned to 0, 0.5, and 1 respectively. Rescue is triggered when the comprehensive score exceeds 0.9. Velodyne VLP-16 laser radar is used for SLAM mapping, with a scanning frequency of 10Hz, a measurement range of 100m, an accuracy of ±3cm, and real-time updates of 10m×10m area three-dimensional scenes. The MQTT protocol is used, and the QoS level is set to 2. The data packet is transmitted through the 5G network (uplink rate 100Mbps), which is about 5MB and contains 1080p scene images, location coordinates, and voice files to the rescue center. The rescue center server has a 64-core CPU and 256GB memory to complete information parsing within 100ms, and the accident level 1-5 is determined by the decision tree algorithm. The improved A* algorithm is used for path planning, considering 5 goals, including time, distance, safety, road conditions, and resource consumption. A 100×100 grid map is set, and the road condition information is updated every 500ms. The navigation system uses Ka lman filtering to fuse GPS and IMU data, and the positioning accuracy reaches ±1m.The rescue effect evaluation system records 10 indicators, including the average response time target <3 minutes, the arrival time target <10 minutes, and the rescue success rate target >95%. The random forest algorithm is used to analyze historical data with a sample size of >10,000 rescue records to identify key factors affecting rescue effects and update the rescue strategy once a month.

[0042] S108. Use an incremental learning method to continuously optimize the wall-occluded speech recognition model, and dynamically update the acoustic characteristic database and speech attenuation interference model based on the recognition accuracy and false alarm rate.

[0043] The sliding window method is used to collect speech recognition results, and the F1 score is calculated based on the speech recognition results. Based on the F1 score, Gaussian process regression is used for Bayesian optimization, and the weights of each parameter in the acoustic characteristic database are dynamically adjusted. Based on the acoustic characteristic database, the attenuation model and interference model of the speech signal are reconstructed, and the recursive least squares method is used to adjust the model parameters in real time. Incremental learning is performed on the wall occlusion speech recognition model, and the newly collected speech data is combined with historical data to dynamically adjust the model structure and parameters. A model performance evaluation system is established, including accuracy, recall rate, and F1 score indicators, and a dynamic threshold is set. When the performance indicator is lower than the threshold, the model retraining is triggered.

[0044] Exemplarily, the sliding window method is used to collect speech recognition results in the recent period of time, and the F1 score is calculated to evaluate the recognition performance. At the same time, the sensor network is used to collect environmental parameters in real time, including wall material, thickness, shape, surface roughness, internal structure, hole distribution, attachments and temperature. According to the calculated F1 score, Gaussian process regression is used for Bayesian optimization, the weights of each parameter in the acoustic characteristic database are dynamically adjusted, and the parameters with greater influence on the recognition effect are optimized. The acoustic characteristic database is updated, and the confidence interval of the parameter update is calculated. The k-fold cross-validation method is used to verify the validity of the updated database to ensure the reasonable integration of new data and historical data. Based on the updated acoustic characteristic database, the attenuation model and interference model of the speech signal are reconstructed, the recursive least squares method is used to adjust the model parameters in real time, and the minimum description length criterion is used to select the optimal model structure, reduce the computational complexity, and improve the adaptability of the model to environmental changes. The Learn++ algorithm is used to perform incremental learning on the wall occlusion speech recognition model, combine the newly collected speech data with the historical data, dynamically adjust the model structure and parameters, add the L2 regularization term to prevent overfitting, and continuously optimize the model performance. A model performance evaluation system was established, including indicators such as accuracy, recall, and F1 score. Dynamic thresholds were set. When the performance indicators were lower than the thresholds, model retraining was triggered to ensure that the recognition accuracy continued to improve and the false alarm rate was reduced. In actual applications, a 30-minute sliding window was used, updated every 5 minutes, to collect speech recognition results and calculate the F1 score to evaluate performance. 20 environmental sensors were deployed, including 5 material analyzers, 5 thickness gauges, 5 roughness sensors, and 5 temperature sensors, with a sampling frequency of 1Hz. Bayesian optimization was performed using Gaussian process regression, with 100 iterations and a learning rate of 0.01, to optimize 50 key parameters in the acoustic characteristics database. A 5-fold cross validation was used, and the validation threshold was set to 0.85. The reconstructed attenuation model and interference model used a 20th-order all-pole structure, the forgetting factor of the recursive least squares method was set to 0.98, and the penalty coefficient of the minimum description length criterion was 0.1. The Learn++ algorithm sets the number of base classifiers to 10, uses a decision tree as the base classifier, and has a maximum depth of 5. An L2 regularization term of 0.001 was added to prevent overfitting. In the model performance evaluation system, the weights of accuracy, recall, and F1 scores were 0.3, 0.3, and 0.4, respectively. The initial value of the dynamic threshold was set to 0.9, and it was adjusted in steps of 0.01 after each evaluation. When the comprehensive performance index was lower than the threshold for three consecutive times, the model retraining was triggered. The speech recognition accuracy of the wall-occluded speech recognition model in a complex wall environment increased from the initial 85% to 95%, the false alarm rate decreased from 10% to 3%, and the F1 score increased from 0.87 to 0.96.

[0045] It will be apparent to those skilled in the art that the present application is not limited to the details of the exemplary embodiments described above, and that the present application can be implemented in other specific forms without departing from the spirit or essential features of the present application. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the present application is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims be included in the present application. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A method for matching and identifying voice signals in the Internet of Things, characterized in that: The method comprises: Acquire sound reflection data of different wall materials, thicknesses, shapes, surface roughness, internal structures, hole distributions, attachments, and temperatures. Through acoustic modeling and simulation analysis, determine the effects of various wall properties on speech signal frequency, amplitude, duration, energy, spectral characteristics, resonance frequency, echo characteristics, and signal-to-noise ratio, and form a database of wall acoustic characteristics. According to the wall acoustic characteristics database, the speech data of the athlete at different distances from the wall during exercise is obtained as speech samples, and a speech feature data set is constructed to extract the frequency, amplitude, duration, energy, spectral characteristics, resonance frequency, echo characteristics and signal-to-noise ratio parameters of the speech signal to obtain the speech feature vectors at different distances from the wall. According to the speech feature vector, combined with the distance and angle position relationship between the athlete and the wall, the speech features are analyzed, and the attenuation and interference model of the speech signal under the occlusion of the wall is established, and the influence of the wall occlusion on the accuracy of speech recognition is obtained; According to the wall material, thickness, shape, surface roughness, internal structure, hole distribution, attachments and temperature properties, the corresponding acoustic parameters are selected from the wall acoustic characteristics database, and combined with the voice signal attenuation and interference model to build a wall occlusion speech recognition model for accurately recognizing voice commands in complex wall environments; Preprocess the speech signal in the wall-occluded speech recognition model to extract the frequency, amplitude, duration, energy, spectral characteristics, resonance frequency, echo characteristics and signal-to-noise ratio characteristic parameters of the speech signal, and optimize the feature subset to reduce the interference of the wall acoustic characteristics on speech recognition; During the exercise, the voice data and position information of the athlete are obtained in real time and input into the wall-occluded speech recognition model. According to the distance and angle position relationship between the athlete and the wall, the acoustic parameters and speech feature weights of the wall-occluded speech recognition model are dynamically adjusted to adaptively optimize the speech recognition strategy and obtain the speech recognition result. If the voice recognition result is that the athlete sends a distress command, the emergency rescue plan will be triggered, an alarm message will be sent to the rescue center, and the optimal rescue route will be planned based on the athlete's location information to guide rescuers to quickly reach the accident scene to minimize accidental injuries; An incremental learning method is used to continuously optimize the wall-occluded speech recognition model, and the acoustic characteristic database and speech attenuation interference model are dynamically updated according to the recognition accuracy and false alarm rate.

2. The method according to claim 1, characterized in that The method obtains sound reflection data of different wall materials, thicknesses, shapes, surface roughness, internal structures, hole distributions, attachments and temperatures, and determines the influence of various wall attributes on the frequency, amplitude, duration, energy, spectral characteristics, resonance frequency, echo characteristics and signal-to-noise ratio of speech signals through acoustic modeling and simulation analysis, so as to form a database of wall acoustic characteristics, including: Use a laser scanner to perform three-dimensional scanning and modeling of the wall surface to obtain the wall surface roughness data; According to the wall surface roughness data, the wall temperature distribution map is collected and the temperature gradient distribution inside the wall is calculated; Obtain the internal structure image of the wall from the X-ray tomography device, analyze the hole distribution and attachment conditions, and generate a three-dimensional model of the internal structure of the wall; According to the three-dimensional model of the wall surface, the internal structure model and the temperature distribution data, an acoustic simulation model is constructed using finite element analysis software; The sound ray tracing method is used to simulate the propagation path of sound waves within a specified frequency range in the wall, calculate the sound wave energy attenuation and reflection characteristics, and obtain the wall acoustic response function; Use microphone arrays to collect reflected sound signals at different locations around the wall and extract the spectral characteristics of the signals; The arrival time and direction of each frequency component are determined, and the independent component analysis method is used for blind source separation to eliminate the effects of multiple reflections and reverberation, and to construct a data set of measured acoustic reflection characteristics.

3. The method according to claim 1, characterized in that According to the wall acoustic characteristic database, the speech data of the athlete at different distances from the wall during exercise is obtained as a speech sample, a speech feature data set is constructed, and the frequency, amplitude, duration, energy, spectrum characteristics, resonance frequency, echo characteristics and signal-to-noise ratio parameters of the speech signal are extracted to obtain speech feature vectors at different distances from the wall, including: Use multiple omnidirectional microphone arrays to track the athlete's position in real time; Record the distance change between the athlete and the wall according to the position of the athlete, and synchronously collect the athlete's voice data to obtain an original voice sample set containing distance information; Perform preliminary quality assessment on the original speech sample set, calculate the signal-to-noise ratio and perform spectrum analysis, and filter out valid speech samples based on the preset signal-to-noise ratio threshold; Preprocess the effective speech samples, remove the low-frequency noise in the environment, eliminate the echo interference, extract the effective speech segments, and obtain the clear speech signals; Extract characteristic parameters from clear speech signals, calculate the spectral characteristics of speech, estimate the formant frequency, extract the pitch period, calculate the energy envelope, and obtain characteristic parameters such as frequency, amplitude, duration and energy of speech; For speech samples collected at different distances, the signal-to-noise ratio and echo characteristic parameters are calculated, and a speech feature vector containing distance information is constructed by combining the speech feature parameters.

4. The method according to claim 1, characterized in that: The speech feature vector is combined with the distance and angle position relationship between the athlete and the wall to analyze the speech feature, establish the attenuation and interference model of the speech signal under the wall occlusion, and obtain the influence of the wall occlusion on the speech recognition accuracy, including: Obtain information about the position of the athlete, the position of the wall, and the voice propagation path, and map the information into a pre-constructed three-dimensional coordinate system; According to the mapping information in the three-dimensional coordinate system, the propagation path length and incident angle of the voice signal in space are calculated; Use a material detector to measure the acoustic characteristic parameters of the wall material, including the sound absorption coefficient and the scattering coefficient; According to the acoustic characteristic parameters, propagation path length and incident angle, the attenuation model of speech signal in the process of spatial propagation is established; Based on the multipath propagation theory, the time delay and phase difference between the direct sound and each reflected sound are analyzed, and the acoustic interference model under the condition of wall shielding is established; Using the attenuation model and the interference model, the original speech feature vector is transformed to obtain the speech feature vector after being blocked by the wall; Input the transformed speech feature vector into a pre-trained speech recognizer to obtain a recognition result; Compare the recognition results with the original speech content and calculate the word error rate; Based on the word error rate, determine the influence of wall occlusion on speech recognition accuracy.

5. The method according to claim 1, characterized in that The method selects corresponding acoustic parameters from the wall acoustic characteristic database according to the wall material, thickness, shape, surface roughness, internal structure, hole distribution, attachments and temperature properties, and combines them with the voice signal attenuation and interference model to construct a wall occlusion voice recognition model for accurately recognizing voice commands in a complex wall environment, including: Acquiring acoustic parameters from a wall acoustic property database, the acoustic parameters including sound absorption coefficient, scattering coefficient, transmission loss and acoustic impedance; The missing values ​​in the acoustic parameters are supplemented by the Kriging interpolation method to obtain a complete set of wall acoustic characteristic parameters. According to the wall acoustic characteristic parameter set, the propagation path of the speech signal in a complex wall environment is simulated; If multiple reflections and diffraction effects are detected, the diffraction of sound waves on irregular surfaces is handled; Calculate the frequency response function of the speech signal at the receiving point through the propagation path model; Determine whether comb filter effect and Doppler effect exist. If so, introduce harmonic distortion model to deal with nonlinear acoustic effects. Construct a time-frequency characteristic transformation model of speech signals, transform the original speech features, and obtain the speech features after being blocked by the wall; Calculate the signal-to-noise ratio of the transformed speech features and determine the threshold of feature correlation analysis; Filter effective features according to the threshold and generate speech data in an occluded environment; Input the filtered speech features into the pre-trained Transformer-based end-to-end speech recognition model; Determine the differences in recognition accuracy of different types of voice commands and obtain the model's voice recognition accuracy in a complex wall environment.

6. The method according to claim 1, characterized in that The method preprocesses the speech signal in the wall-occluded speech recognition model, extracts the frequency, amplitude, duration, energy, spectral characteristics, resonance frequency, echo characteristics and signal-to-noise ratio characteristic parameters of the speech signal, and optimizes the feature subset to reduce the interference of the wall acoustic characteristics on speech recognition, including: After preprocessing the input voice signal, voice activity detection is performed to obtain valid voice segments; Extracting feature parameters from the effective speech segment, the feature parameters comprising: calculating the spectrum features of the effective speech segment; The linear prediction analysis is performed through the Levinson-Durbin recursive algorithm to obtain the formant frequency of the effective speech segment; The real cepstrum method is used to extract the pitch period of the effective speech segment; Calculate the energy envelope of the valid speech segment; Modeling the feature parameters, including: using a Gaussian mixture model to model the voice features of different speakers; Calculate signal-to-noise ratio and echo characteristic parameters; Combining spectral features, formant frequency, pitch period and energy envelope, a high-dimensional feature vector is constructed; The high-dimensional feature vector is optimized, and the optimization process includes: normalizing the high-dimensional feature vector by using a Z-score normalization method; Use principal component analysis to perform dimension reduction and correlation analysis on the normalized feature vector; Compensation for the effects of wall acoustic properties; Feature selection is performed on the optimized feature vector to obtain the optimized feature vector.

7. The method according to claim 1, characterized in that The method acquires the voice data and position information of the athlete in real time during the exercise process, and inputs them into the wall-blocked voice recognition model. According to the distance and angle position relationship between the athlete and the wall, the acoustic parameters and voice feature weights of the wall-blocked voice recognition model are dynamically adjusted, and the voice recognition strategy is adaptively optimized to obtain the voice recognition result, including: Obtain the three-dimensional position coordinates of the athlete, which are calculated in real time by multi-sensor fusion positioning technology using UWB, inertial sensors and cameras; The speech signal of the athlete is collected according to the three-dimensional position coordinates, and the speech signal is collected by a microphone array and subjected to signal preprocessing and feature extraction to obtain a speech feature vector; Calculate the distance and angle relationship between the athlete and the wall, the distance and angle relationship are determined by the three-dimensional position coordinates; Simulate the sound wave propagation path, which is calculated based on the relationship between distance and angle; Dynamically compensate the speech feature vector. The dynamic compensation is performed based on the acoustic propagation model updated by the sound wave propagation path. The compensated speech feature vector is obtained by adjusting the frequency response, energy distribution and delay characteristics. The compensated speech feature vector is input into an adaptive speech recognizer, which adopts a recursive least squares algorithm for online learning and obtains a speech recognition result by updating recognizer parameters.

8. The method according to claim 1, characterized in that If the voice recognition result is that the athlete sends a distress command, the emergency rescue plan is triggered, an alarm message is sent to the rescue center, and the optimal rescue route is planned according to the athlete's location information to guide rescue personnel to quickly reach the accident scene to minimize accidental injuries, including: The BERT model is used to perform semantic analysis on the speech recognition results, extract keywords and match them with the preset rescue command vocabulary to obtain a semantic matching score; According to the semantic matching score and emotional urgency score, determine whether to trigger the help command; If a distress command is triggered, a three-dimensional scene model is constructed and updated in real time; Based on the 3D scene model and in combination with the geographic information system, a detailed scene map containing information such as obstacles, passages, and entrances and exits is generated; Package the distress instructions, the athlete's location and the scene map information and send them to the rescue center; Receive rescue instructions from the rescue center and use the A* algorithm for multi-objective path planning; According to the multi-objective path planning results, multiple candidate rescue routes are calculated, and the optimal rescue route is selected from the multiple candidate rescue routes; Obtain rescue effect data of the optimal rescue route and establish a rescue effect evaluation system; According to the rescue effect evaluation system, the response time, arrival time and rescue success rate indicators are recorded to obtain the key factor analysis results of the rescue process.

9. The method according to claim 1, characterized in that: The incremental learning method is used to continuously optimize the wall-occluded speech recognition model, and dynamically update the acoustic characteristic database and the speech attenuation interference model according to the recognition accuracy and false alarm rate, including: The sliding window method is used to collect speech recognition results, and the F1 score is calculated based on the speech recognition results; According to the F1 score, Gaussian process regression is used to perform Bayesian optimization to dynamically adjust the weights of each parameter in the acoustic characteristics database; Based on the acoustic characteristics database, the attenuation model and interference model of the speech signal are reconstructed, and the model parameters are adjusted in real time using the recursive least squares method; Incremental learning of the wall-occluded speech recognition model is performed to combine newly collected speech data with historical data and dynamically adjust the model structure and parameters; Establish a model performance evaluation system, including accuracy, recall and F1 score indicators, set dynamic thresholds, and trigger model retraining when the performance indicator is lower than the threshold.

Images