Iron tower voiceprint intelligent detection method and system

Abstract

The invention provides an iron tower voiceprint intelligent detection method and system, and belongs to the voiceprint detection technology. The method comprises the following steps: firstly, outputting a sweep frequency signal in a specific frequency range, collecting iron tower vibration response data, extracting candidate frequency through spectral analysis, and dynamically adjusting excitation frequency by using a gradient descent algorithm; hardware filtering is carried out on collected sound signals, frequency band signals related to excitation frequency are reserved, and space beam forming is carried out through a microphone array to enhance iron tower voiceprint signals. Separating iron tower vibration components from the mixed signals by utilizing independent component analysis, performing time-frequency analysis, extracting an energy ratio of a specified frequency band by adopting wavelet packet transformation, and performing deep learning processing in combination with a one-dimensional convolutional neural network to generate an energy ratio and zero-crossing rate feature vector; and finally, inputting the feature vectors into a support vector data description model, setting an initial threshold value and dynamically adjusting the initial threshold value, thereby realizing graded judgment of bolt looseness and improving detection efficiency and stability.

Description

Technical Field

[0001] The present invention relates to a method and system for intelligent detection of tower acoustic fingerprints, belonging to the technical field of acoustic fingerprint detection. Background Art

[0002] At present, the detection of tower bolt loosening is crucial for ensuring the safety of power and communication. Traditional detection techniques use fixed-frequency excitation, which is difficult to adapt to the complex vibration characteristics of towers, resulting in weak effective signals; manually setting thresholds lacks environmental adaptability and is prone to false positives and false negatives under noise interference; moreover, the power management mode is single, unable to effectively utilize environmental energy, and the battery life is insufficient. Although existing detection schemes have been improved, most lack a dynamic feedback mechanism and cannot adjust the excitation frequency and detection strategy based on real-time detection data; at the same time, there is a lack of multi-level collaborative design, and each detection link operates independently, making it difficult to exert the overall effectiveness, resulting in poor detection accuracy and stability. Therefore, there is an urgent need for an intelligent detection system for tower acoustic fingerprints that has dynamic feedback, multi-level collaboration, and can effectively overcome the defects of traditional detection techniques. Summary of the Invention

[0003] The object of the present invention is to provide a method and system for intelligent detection of tower acoustic fingerprints, which comprehensively improve the detection efficiency and stability, and provide a strong guarantee for the safe operation and maintenance of towers.

[0004] To achieve the above object, the present invention is realized through the following technical solutions: The method includes the following steps: Output a swept-frequency signal within a specific frequency range, collect tower vibration response data; extract candidate frequencies in the response signal through spectrum analysis, dynamically adjust the excitation frequency using the gradient descent algorithm, and calibrate the frequency range regularly based on temperature data; Perform hardware filtering on the collected sound signal to retain the frequency band signal related to the excitation frequency; perform spatial beamforming through a microphone array to enhance the tower acoustic fingerprint signal; use the independent component analysis algorithm to separate the tower vibration component from the mixed signal; Perform time-frequency analysis on the separated signal, extract the energy ratio of the specified frequency band after denoising using wavelet packet transform; at the same time, perform deep learning processing on the signal through a one-dimensional convolutional neural network to generate a feature vector composed of the energy ratio and the zero-crossing rate; Input the energy ratio and the zero-crossing rate in the time domain into the support vector data description model, set an initial threshold based on the statistical characteristics in the fault-free state, dynamically adjust the threshold in combination with environmental noise, and perform hierarchical determination of bolt loosening according to the scoring results.

[0005] Preferably, the spectrum analysis is Fourier transform; The objective function of the gradient descent algorithm is to maximize the vibration response amplitude, and the frequency update formula is: , Among them, represents the learning rate; The termination condition is that the amplitude change rate in three consecutive iterations is less than 5% or the maximum number of iterations is reached.

[0006] Preferably, the hardware filter uses a programmable filter with a center frequency of the current excitation frequency; The spatial beamforming is generated based on the generalized cross-correlation-phase transformation algorithm.

[0007] Preferably, the wavelet packet transform uses the Daubechies-4 wavelet basis for 4-layer decomposition to extract the energy ratio in the frequency band of 0.5 - 2 kHz; The input layer of the one-dimensional convolutional neural network contains 64 nodes, the convolutional kernel size is 3×1, and a 16-dimensional feature vector is output.

[0008] The calculation formula for the energy ratio is as follows: Among them, represents the energy ratio, represents the wavelet packet coefficient of the i-th sub-band; The zero-crossing rate is the frequency at which the time-domain signal crosses the zero point, and the calculation formula is: , Among them, is the indicator function, is the signal length.

[0009] Preferably, the specific method for dynamically adjusting the threshold is as follows: Among them, is the feature mean in the fault-free state, is the standard deviation, is the adjusted threshold; The hierarchical determination includes: when the score is greater than or equal to 1.5 and less than 2 it is mild looseness, when the score is greater than or equal to 2 and less than 3 it is moderate looseness, when the score is greater than or equal to 3 it is severe looseness.

[0010] Preferably, the SVDD model calculates the Mahalanobis distance between the denoised feature vector and the center of the hypersphere as the score : , Among them, is the normal sample mean vector, is the covariance matrix, represents the matrix transpose.

[0011] Preferably, the specific method of wavelet packet transform denoising is as follows: Collect environmental noise when there is no target signal, and calculate the root mean square value of the environmental noise , and count the number of wavelet packet coefficients of the subband , where is determined by the signal length and the decomposition level; Based on the noise characteristics of the wavelet packet decomposition subband, calculate the dynamic threshold , , The dynamic threshold without bandwidth correction The calculation formula is as follows: , If the subband bandwidths are significantly different, the threshold weight needs to be adjusted according to the frequency band width, and the dynamic threshold The calculation formula is as follows Among them, is the adjustment factor, is the bandwidth of the subband , is the total bandwidth.

[0012] A smart tower acoustic fingerprint detection system includes: an adaptive excitation frequency generation module, a hierarchical anti-interference acoustic fingerprint acquisition module, a multi-modal feature extraction module, a bolt loosening intelligent determination module, and a dynamic power management module; The dynamic power management module is respectively connected to the adaptive excitation frequency generation module, the hierarchical anti-interference acoustic fingerprint acquisition module, the multi-modal feature extraction module, and the bolt loosening intelligent determination module; The adaptive excitation frequency generation module is connected to the hierarchical anti-interference acoustic fingerprint acquisition module, and is used to output a dynamically adjusted excitation frequency signal and collect the tower vibration response; The hierarchical anti-interference acoustic fingerprint acquisition module is connected to the multi-modal feature extraction module, and is used to perform anti-interference processing on the collected acoustic fingerprint signal and separate the tower vibration component; The multi-modal feature extraction module is connected to the bolt loosening intelligent determination module, and is used to extract the time-frequency features and deep learning features of the signal; The bolt loosening intelligent determination module is based on a dynamic threshold model to classify and determine the bolt loosening state.

[0013] Preferably, the adaptive excitation frequency generation module includes: The AD9833 programmable waveform generator is used to output a swept-frequency signal of 10 Hz - 10 kHz; The ADXL357 triaxial vibration sensor is used to collect the vibration response data of the iron tower; The DS18B20 temperature sensor is used to obtain the ambient temperature data; The STM32H743 main control chip is used to dynamically adjust the excitation frequency through the gradient descent algorithm; The hierarchical anti-interference voiceprint acquisition module includes: The LTC1068 programmable filter is set to the band-pass mode, with a center frequency of the excitation frequency and a bandwidth of ±200 Hz; An array composed of 4 ADI-ADMP521-MEMS microphones realizes spatial beamforming based on the GCC-PHAT algorithm; The ADS1278 - 24-bit ADC is used for signal analog-to-digital conversion; The module separates the vibration components of the iron tower from the mixed signal through the FastICA algorithm.

[0014] Preferably, the dynamic power management module includes: The BQ25895 power management chip supports voltage switching in the excitation mode, acquisition mode, and standby mode. The module priority switching logic is: excitation mode > acquisition mode > standby mode; The LTC3588 energy recovery circuit converts the vibration energy into 5V DC power through a piezoelectric sensor; The CAP-XX super capacitor is used to store the recovered electric energy.

[0015] The advantages of the present invention are as follows: The present invention utilizes the gradient descent algorithm and the temperature calibration mechanism to be able to dynamically match the complex vibration characteristics of the iron tower. This characteristic enables the system to output the optimal excitation frequency, effectively enhancing the intensity of the iron tower vibration response signal and improving the sensitivity and accuracy of detection.

[0016] Adopt multi-level noise reduction technology, including steps such as hardware filtering, spatial beamforming, and signal separation. The combination of these technologies can significantly reduce the influence of environmental noise on the detection signal, improve the purity of the signal, and provide a high-quality data basis for subsequent feature extraction and determination.

[0017] Integrate two technologies of wavelet packet transform and one-dimensional convolutional neural network, which can extract features from different angles for the signal. This multi-modal feature extraction method can more comprehensively describe the characteristics of the signal, provide richer and more accurate feature information for the intelligent determination of bolt loosening, and thus improve the accuracy and reliability of detection.

[0018] Based on the dynamic threshold model, it is able to dynamically adjust the determination threshold according to the real-time changes in environmental noise. This dynamic threshold determination method can significantly improve the system's adaptability to environmental noise, reduce the occurrence of false positives and false negatives, and enhance the robustness of detection.

[0019] The dynamic power management module realizes multi-mode power supply and energy recovery functions. It can automatically switch the voltage and current configurations according to the working state of the system, and convert the vibration energy into electrical energy through piezoelectric sensors for storage. This design effectively extends the battery life of the system and improves the practicality and economy of the system. Brief Description of the Drawings

[0020] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention.

[0021] Figure 1 It is a schematic flowchart of the method of the present invention.

[0022] Figure 2 It is a block diagram of an intelligent tower acoustic fingerprint detection system of the present invention; Detailed Embodiments The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0023] As Figure 1 shown, an intelligent tower acoustic fingerprint detection method includes the following steps: S1: Output a swept-frequency signal within a specific frequency range, collect tower vibration response data; extract candidate frequencies in the response signal through spectrum analysis, dynamically adjust the excitation frequency using the gradient descent algorithm, and calibrate the frequency range regularly based on temperature data.

[0024] As a refinement of the above embodiment, it specifically includes: Initial swept frequency: Output a 10Hz - 10kHz swept-frequency signal and collect the tower vibration response. Output a 10Hz - 10kHz swept-frequency signal through the AD9833 programmable waveform generator to excite the tower to generate a vibration response, and the ADXL357 three-axis vibration sensor collects these vibration data. This frequency range covers the common frequency bands where the tower may generate vibration responses, ensuring that the vibration characteristics of the tower can be fully excited.

[0025] Spectrum analysis: Extract the candidate frequencies corresponding to the energy peaks in the response signal through Fourier transform. Based on the main control chip STM32H743, the collected response signal is processed by fast Fourier transform to extract the candidate frequencies corresponding to the energy peaks. Fast Fourier transform is an efficient spectrum analysis algorithm that can convert the time-domain signal into a frequency-domain signal, thereby accurately finding the frequency points where the energy is concentrated, and these frequency points are the candidate frequencies for subsequent analysis.

[0026] The objective function of the gradient descent algorithm is to maximize the vibration response amplitude, that is, to minimize . The objective function here is the negative of the vibration response amplitude, that is, by minimizing to maximize the amplitude.

[0027] The frequency update formula is: , where represents the learning rate, which takes the value of 1 in this embodiment; The termination condition is that the amplitude change rate is less than 5% for 3 consecutive iterations or the maximum number of iterations is reached.

[0028] Dynamic adjustment: With the goal of maximizing the response amplitude, the excitation frequency is iteratively adjusted in steps of Δf = 10Hz. After each adjustment, the vibration response amplitude is calculated until the amplitude change rate for 3 consecutive iterations < 5% or the maximum number of iterations of 50 times is reached. This iterative adjustment method can gradually approach the frequency that causes the maximum vibration response of the iron tower, improving the sensitivity and accuracy of detection.

[0029] Calibration mechanism: Every 5 minutes, the system re-executes the frequency sweep operation and corrects the candidate frequency range according to the temperature data collected by the DS18B20 temperature sensor. The change in temperature will affect the physical properties of the iron tower material, and thus affect its vibration frequency. Therefore, temperature calibration can ensure the accuracy of the excitation frequency and make it adapt to the characteristics of the iron tower under different environmental temperatures.

[0030] S2: Perform hardware filtering on the collected sound signal to retain the frequency band signals related to the excitation frequency; perform spatial beamforming through the microphone array to enhance the iron tower soundprint signal; use the independent component analysis algorithm to separate the iron tower vibration component from the mixed signal.

[0031] As a refinement of the above embodiment, it specifically includes: Hardware filtering: Set the LTC1068 programmable filter to the band-pass mode, with the center frequency set to the excitation frequency obtained by the adaptive excitation frequency generation module and the bandwidth of ±200Hz. This can initially filter out the noise signals unrelated to the excitation frequency at the hardware level and only retain the signals within the frequency range related to the iron tower vibration, improving the purity of the signal.

[0032] Spatial beamforming: Based on the Generalized Cross-Correlation - Phase Transform algorithm, i.e., the GCC-PHAT algorithm, an acoustic wave beam pointing to the iron tower is generated using an array composed of 4 ADI-ADMP521-MEMS microphones. The GCC-PHAT algorithm determines the direction of the sound source by calculating the time delay of the signals received by different microphones, thereby forming a beam focused on the iron tower, enhancing the iron tower voiceprint signal, and suppressing noise signals in other directions.

[0033] Signal separation: The vibration component of the iron tower is separated from the mixed signals by means of the Fast Independent Component Analysis algorithm, i.e., the FastICA algorithm. Based on the assumption of signal independence, the FastICA algorithm can separate multiple mixed signal sources. In this system, the voiceprint signal generated by the vibration of the iron tower is effectively extracted, further suppressing noise interference.

[0034] S3: Perform time-frequency analysis on the separated signals, and use wavelet packet transform to extract the energy proportion of the specified frequency band; at the same time, perform deep learning processing on the signals through a one-dimensional convolutional neural network to generate a feature vector composed of the energy proportion and the zero-crossing rate.

[0035] As a refinement of the above embodiment, it specifically includes: Time-frequency analysis: Using Daubechies-4 wavelet packet transform, decompose 4 layers to generate 16 sub-bands, each sub-band width is 312.5Hz, extract the 5th - 8th sub-bands, that is, extract the energy proportion of the 0.5 - 2kHz frequency band.

[0036] The formula for calculating the energy proportion is as follows: Among them, represents the energy proportion, represents the wavelet packet coefficient of the i-th sub-band; the numerator is the sum of the squares of the wavelet packet coefficients of the 5th - 8th sub-bands, and the denominator is the sum of the squares of the wavelet packet coefficients of all 16 sub-bands. Divide the two and multiply by 100% to obtain the proportion of the energy of the 0.5 - 2kHz frequency band in the total signal energy.

[0037] The calculated energy proportion and the zero-crossing rate of the signal are combined to form a 2D feature vector. The zero-crossing rate refers to the number of times the signal crosses the zero level within a unit time, which reflects the degree of signal variation. Combining these two features can describe the characteristics of the signal from different angles, providing richer and more accurate feature information for the subsequent intelligent determination of bolt loosening.

[0038] The zero-crossing rate is the frequency at which the time-domain signal crosses zero, and the calculation formula is: , Among them, is an indicator function, is the signal length.

[0039] Deep learning: 1D-CNN, that is, one-dimensional convolutional neural network, with 64 nodes in the input layer, a convolutional kernel size of 3×1, and a 16-dimensional feature vector output. Performing deep learning processing on the signal can automatically extract the deep features in the signal and finally output a 16-dimensional feature vector. These feature vectors further enrich the description of the signal and, together with the features extracted by wavelet packet transform, help to more comprehensively and accurately judge the loosening condition of the tower bolts.

[0040] S4: Input the energy ratio and the zero-crossing rate in the time domain into the support vector data description model, set the initial threshold based on the statistical features in the fault-free state, dynamically adjust the threshold in combination with the environmental noise, and classify and judge the bolt loosening according to the scoring result.

[0041] As a refinement of the above embodiment, the denoised energy ratio and ZCR are combined into a 2D feature vector [E, ZCR] and used as the input of the support vector data description SVDD model. Through time-frequency joint analysis, it can capture both the frequency-domain energy anomaly, such as the structural resonance caused by bolt loosening, and detect the time-domain transient events, such as the instantaneous vibration caused by loosening, to improve the characterization ability of the features.

[0042] Specifically include: Feature fusion: Input the denoised energy ratio and the zero-crossing rate output by the multi-modal feature extraction module into the support vector data description model, that is, SVDD, which is a one-class classification algorithm.

[0043] Dynamic threshold: Adjust the judgment threshold according to the RMS value of the environmental noise. The specific method of dynamically adjusting the threshold is as follows: Among them, is the feature mean in the fault-free state, is the standard deviation, is the adjusted threshold; if the environmental noise fluctuates significantly, such as in strong wind weather, the system updates and in real time to ensure that the threshold adapts to the noise change.

[0044] The fault grading includes: a score greater than or equal to 1.5 less than 2 is mild loosening, a score greater than or equal to 2 less than 3 is moderate loosening, a score greater than or equal to 3 When it is severe looseness. Mild looseness indicates that the feature slightly deviates from the normal range, which may be initial looseness or environmental interference. Moderate looseness indicates that the feature significantly deviates and requires arranging manual inspections for confirmation. Severe looseness indicates that the feature is seriously abnormal, triggering an emergency alarm and locating the faulty bolt.

[0045] The SVDD model calculates the Mahalanobis distance between the denoised feature vector and the center of the hypersphere as the score : , Among them, is the normal sample mean vector, is the covariance matrix, represents the matrix transpose.

[0046] The wavelet packet transform denoising is based on a dynamic threshold model and is implemented through the following steps: Step 1: Noise measurement: Collect environmental noise when there is no target signal, calculate the root mean square value of the environmental noise , and count the number of wavelet packet coefficients in the subband where is determined by the signal length and the decomposition level.

[0047] Step 2: Dynamic threshold calculation: Based on the noise characteristics of the wavelet packet decomposition subbands, the dynamic threshold is generated according to the following formula: .

[0048] Among them, is the adjustment factor, and its value range is , and the default setting is 2.0.

[0049] Step 3: Non-uniform bandwidth correction If the subband bandwidths vary significantly, add a bandwidth weight to correct the threshold .

[0050] Among them, is the bandwidth of the subband , is the total bandwidth.

[0051] Step 4: Real-time adjustment: According to the real-time change of the environmental noise , dynamically update the threshold: .

[0052] Step 5: Threshold application: For the subband wavelet packet coefficients output by the multi-modal feature extraction module , set the coefficients with absolute values less than to zero or shrink them to achieve noise suppression and feature enhancement.

[0053] ​During the wavelet packet decomposition stage, sub-band thresholds are dynamically generated to suppress environmental noise and improve the signal-to-noise ratio of feature extraction.

[0054] First, noise measurement is carried out. The data source is the preprocessed signal output by the hierarchical anti-interference voiceprint acquisition module when there is no tower vibration.

[0055] Parameter calculation: RMS value of environmental noise , which reflects the overall energy of the noise, and the unit is voltage (V) or sound pressure level (dB).

[0056] The number of sub-band coefficients is determined by the signal length L and the decomposition level J. The calculation formula is: .

[0057] Dynamic threshold calculation, basic formula: .

[0058] The physical meaning is the statistical derivation based on the extreme value distribution of Gaussian noise, ensuring that 99% of the noise coefficients are suppressed.

[0059] : Adjustment factor, default value 2.0, which can be adjusted between 1.0 (loose) and 3.0 (strict) through the system interface.

[0060] Non-uniform bandwidth correction: When the sub-band bandwidths vary significantly, such as in non-uniform decomposition, the threshold weights need to be corrected. The correction formula is .

[0061] Dynamic update mechanism: Real-time collect environmental noise signals, calculate , and dynamically update the threshold: .

[0062] Update frequency: Calculate once every 1 second to ensure a quick response to noise fluctuations, such as sudden electromagnetic interference.

[0063] Threshold application: When performing hard threshold processing, if the absolute value of the sub-band coefficient is less than , then set it to zero: .

[0064] Applicable scenario: In a strong noise environment, give priority to ensuring feature stability.

[0065] Soft threshold processing: If the absolute value of the sub-band coefficient is less than, perform shrinkage: .

[0066] Applicable scenario: In a weak noise environment, retain some low-amplitude feature details.

[0067] Example 2 An intelligent detection system for the sound pattern of iron towers, comprising: an adaptive excitation frequency generation module, a hierarchical anti-interference sound pattern acquisition module, a multi-modal feature extraction module, a bolt loosening intelligent determination module, and a dynamic power management module; the adaptive excitation frequency generation module is connected to the hierarchical anti-interference sound pattern acquisition module, the hierarchical anti-interference sound pattern acquisition module is connected to the multi-modal feature extraction module, and the multi-modal feature extraction module is connected to the bolt loosening intelligent determination module.

[0068] The adaptive excitation frequency generation module outputs a swept-frequency signal within a specific frequency range, collects the vibration response data of the iron tower, obtains candidate frequencies through spectrum analysis, dynamically adjusts the excitation frequency using the gradient descent algorithm, and calibrates the frequency range regularly according to the temperature data.

[0069] The hierarchical anti-interference sound pattern acquisition module first performs hardware filtering on the collected sound signal in a band-pass mode, then realizes spatial beamforming through a microphone array based on the GCC-PHAT algorithm, and finally separates the vibration components of the iron tower from the mixed signal and suppresses noise with the help of the FastICA algorithm.

[0070] The multi-modal feature extraction module, during the extraction process, on the iron tower, on the one hand, uses Daubechies4 wavelet packet transform to decompose the signal into 4 layers, and extracts the energy proportion of a specific frequency band; on the other hand, performs deep learning processing on the signal through a one-dimensional convolutional neural network and outputs a feature vector.

[0071] When the bolt loosening intelligent determination module makes a determination, it inputs the energy proportion and the zero-crossing rate in the time domain obtained from time-frequency analysis into the support vector data description model, first calculates the mean and standard deviation in the fault-free state to determine the initial threshold, then dynamically scales the threshold according to the RMS value of the environmental noise, and finally grades the bolt loosening situation according to the score.

[0072] The dynamic power management part enables the system to switch between working modes with different voltage and current configurations, and at the same time uses the piezoelectric principle to convert vibration energy into electrical energy to charge the energy storage device and monitors the battery level in real time.

[0073] As a refinement of the above embodiment, the adaptive excitation frequency generation module includes: an AD9833 programmable waveform generator, an ADXL357 three-axis vibration sensor, a DS18B20 temperature sensor, and an STM32H743 main control chip, which is used to dynamically adjust the excitation frequency through the gradient descent algorithm.

[0074] The AD9833 programmable waveform generator is used to output a swept-frequency signal of 10Hz - 10kHz; it excites the iron tower to generate vibration responses, and the ADXL357 three-axis vibration sensor collects these vibration data. This frequency range covers the common frequency bands where the iron tower may generate vibration responses, ensuring that the vibration characteristics of the iron tower can be fully excited.

[0075] ADXL357 three-axis vibration sensor, used to collect the vibration response data of the iron tower.

[0076] DS18B20 temperature sensor, used to obtain the ambient temperature data.

[0077] STM32H743 main control chip, used to dynamically adjust the excitation frequency through the gradient descent algorithm; process the collected response signals through the fast Fourier transform, and extract the candidate frequencies corresponding to the energy peaks therein. The fast Fourier transform is an efficient spectrum analysis algorithm that can convert the time-domain signal into the frequency-domain signal, so as to accurately find the frequency points where the energy is concentrated, and these frequency points are the candidate frequencies for subsequent analysis.

[0078] As a refinement of the above embodiment, the hierarchical anti-interference voiceprint acquisition module includes: an array composed of 4 ADI-ADMP521-MEMS microphones, an LTC1068 programmable filter, and an ADS1278-24-bit ADC, used to suppress noise through the following steps.

[0079] LTC1068 programmable filter, set the LTC1068 programmable filter to the band-pass mode, the center frequency is set to the excitation frequency obtained by the adaptive excitation frequency generation module, and the bandwidth is ±200Hz. This can initially filter out the noise signals irrelevant to the excitation frequency at the hardware level, and only retain the signals within the frequency range related to the vibration of the iron tower, improving the purity of the signals.

[0080] The array composed of 4 ADI-ADMP521-MEMS microphones realizes spatial beamforming based on the GCC-PHAT algorithm; uses the array composed of 4 ADI-ADMP521-MEMS microphones to generate an acoustic wave beam pointing to the iron tower. The GCC-PHAT algorithm determines the direction of the sound source by calculating the time delay of the signals received by different microphones, so as to form a beam focused on the iron tower, enhance the voiceprint signal of the iron tower, and suppress the noise signals in other directions.

[0081] ADS1278-24-bit ADC, used for signal analog-to-digital conversion; The hierarchical anti-interference voiceprint acquisition module separates the vibration components of the iron tower from the mixed signals through the FastICA algorithm; the FastICA algorithm is based on the independence assumption of the signals and can separate multiple signal sources mixed together. In this system, it effectively extracts the voiceprint signals generated by the vibration of the iron tower and further suppresses the noise interference.

[0082] As a refinement of the above embodiment, the multi-modal feature extraction module is integrated in the Xilinx-Artix-7-FPGA chip, and uses its powerful processing ability and reconfigurability to provide efficient hardware support for multi-modal feature extraction.

[0083] As a refinement of the above embodiment, the intelligent bolt loosening determination module is based on a dynamic threshold model and, through time-frequency joint analysis, captures both frequency-domain energy anomalies, such as structural resonance caused by bolt loosening, and detects time-domain transient events, such as instantaneous vibrations caused by loosening, to enhance the characterization ability of features.

[0084] As a refinement of the above embodiment, the dynamic power management module includes a BQ25895 power management chip, a piezoelectric sensor, an LTC3588 energy recovery circuit, a PZT-5H piezoelectric ceramic sheet, and a CAP-XX supercapacitor, and supports: Voltage switching: The excitation mode outputs 12V / 3A, the acquisition mode outputs 5V / 1A, and the standby mode outputs 3.3V / 0.5A.

[0085] Energy recovery: The vibration energy is converted into 5V DC power and stored through the piezoelectric sensor.

[0086] Functions of the BQ25895 power management chip: Multi-mode output: Supports switching of voltage / current configurations for the excitation mode (12V / 3A), acquisition mode (5V / 1A), and standby mode (3.3V / 0.5A).

[0087] Dynamic load adjustment: Automatically adjusts the output according to the power consumption requirements of the module.

[0088] The voltage switching mechanism for the excitation mode: The trigger condition is when the adaptive excitation frequency generation module starts to generate a frequency sweep signal. The output configuration is 12V / 3A, which powers the AD9833 waveform generator and the vibration sensor.

[0089] Acquisition mode: The trigger condition is when the hierarchical anti-interference voiceprint acquisition module is activated. The output configuration is 5V / 1A, which powers the MEMS microphone array and the ADS1278 24-bit ADC.

[0090] Standby mode: The trigger condition is that the system has no operation timeout, with a default of 5 minutes. Output configuration: 3.3V / 0.5A, which only maintains the basic operation of the main control chip STM32H743.

[0091] The switching logic is priority control: excitation mode > acquisition mode > standby mode.

[0092] Seamless switching: Through the dynamic voltage scaling function of the BQ25895, the mode switching is completed within <1ms to avoid power interruption.

[0093] Functions of the LTC3588 energy recovery circuit: Piezoelectric energy conversion: Converts the vibrational mechanical energy captured by the piezoelectric sensor into 5V DC power.

[0094] Energy Storage Management: Optimize energy storage efficiency through Pulse Frequency Modulation (PFM).

[0095] CAP-XX Supercapacitor Function: High power density energy storage: Quickly store the electrical energy converted by LTC3588 to support instantaneous high power demands.

[0096] Long cycle life: The number of charge and discharge cycles > 1 million times, adapting to frequent energy fluctuations.

[0097] The energy recovery process is that first, vibration energy is captured by installing PZT-5H piezoelectric ceramic chips at the key vibration nodes of the iron tower, with a sensitivity of 12 mV / N.

[0098] The energy conversion process is that the mechanical energy of vibration is converted into alternating current by the piezoelectric sensor, then converted into 5V direct current by LTC3588 after rectification.

[0099] Energy storage path: The output of LTC3588 is connected to the CAP-XX supercapacitor for real-time charging.

[0100] Power supply strategy: Give priority to using the supercapacitor: During high power consumption periods in the excitation mode or acquisition mode, the supercapacitor directly supplies power.

[0101] External power supply supplement: When the voltage of the supercapacitor < 3.9V, switch to an external solar cell or lithium battery.

[0102] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, for those skilled in the art, they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An intelligent detection method for the voiceprint of a steel tower, characterized in that It includes the following steps: Output a swept-frequency signal in a specific frequency range, and collect the vibration response data of the iron tower; extract the candidate frequencies in the response signal through spectrum analysis, dynamically adjust the excitation frequency using the gradient descent algorithm, and calibrate the frequency range regularly based on the temperature data; Perform hardware filtering on the collected sound signal to retain the frequency band signal related to the excitation frequency; Perform spatial beamforming through a microphone array to enhance the soundprint signal of the iron tower; use the independent component analysis algorithm to separate the vibration components of the iron tower from the mixed signal; Perform time-frequency analysis on the separated signal, extract the energy ratio of the specified frequency band after denoising using wavelet packet transform; at the same time, perform deep learning processing on the signal through a one-dimensional convolutional neural network to generate a feature vector composed of the energy ratio and the zero-crossing rate; Input the energy ratio and the zero-crossing rate in the time domain into the support vector data description model, set the initial threshold based on the statistical characteristics in the fault-free state, dynamically adjust the threshold in combination with the environmental noise, and classify and determine the bolt looseness according to the scoring result.

2. The intelligent detection method of tower acoustic fingerprint according to claim 1, characterized in that The spectrum analysis is Fourier transform; The objective function of the gradient descent algorithm is to maximize the vibration response amplitude, and the frequency update formula is: , Among them, represents the learning rate; The termination condition is that the amplitude change rate is less than 5% for three consecutive iterations or the maximum number of iterations is reached.

3. The intelligent detection method for tower acoustic fingerprint according to claim 1, wherein The hardware filtering uses a programmable filter with the center frequency being the current excitation frequency; The spatial beamforming is generated based on the generalized cross-correlation-phase transformation algorithm.

4. The intelligent detection method for the soundprint of the iron tower according to claim 1, wherein, The wavelet packet transform uses the Daubechies-4 wavelet basis for 4-layer decomposition to extract the energy ratio of the 0.5 - 2 kHz frequency band; The input layer of the one-dimensional convolutional neural network contains 64 nodes, the convolution kernel size is 3×1, and a 16-dimensional feature vector is output; The formula for calculating the energy ratio is as follows: ； Among them, represents the energy ratio, represents the wavelet packet coefficient of the \(i\)-th sub-band; The zero-crossing rate is the frequency at which the time-domain signal crosses zero, and the calculation formula is as follows: ， Among them, is an indicator function, is the signal length.

5. The intelligent detection method for the soundprint of a steel tower according to claim 1, wherein The specific method for dynamically adjusting the threshold is as follows: wherein, is the characteristic mean value in the fault-free state, is the standard deviation, is the adjusted threshold value; The grading determination includes: the score is greater than or equal to 1.5 less than 2 When it is mild loosening, the score is greater than or equal to 2 less than 3 When it is moderate loosening, the score is greater than or equal to 3 When it is severe loosening.

6. The intelligent detection method for the soundprint of a iron tower according to claim 5, characterized in that, The support vector data description model calculates the Mahalanobis distance between the denoised feature vector and the center of the hypersphere as the score : ， wherein, is the normal sample mean vector, is the covariance matrix, represents matrix transpose.

7. The intelligent detection method for the soundprint of the iron tower according to claim 6, wherein The specific method for wavelet packet transform denoising is as follows: Collect environmental noise when there is no target signal, and calculate the root mean square value of the environmental noise , and count the number of wavelet packet coefficients in the subband , where is determined by the signal length and the decomposition level; ​ Calculate the dynamic threshold based on the noise characteristics of the subbands obtained by wavelet packet decomposition , Dynamic Threshold without Bandwidth Correction The calculation formula is as follows: ， The sub-band bandwidths vary significantly, and the threshold weight needs to be adjusted according to the frequency band width, with a dynamic threshold The calculation formula is as follows Among them, is a regulatory factor, is the sub-band bandwidth, is the total bandwidth.

8. An intelligent detection system for the sound pattern of iron towers, characterized in that, Execute the intelligent soundprint detection method for iron towers as described in any one of claims 1 - 7, including: an adaptive excitation frequency generation module, a hierarchical anti-interference soundprint acquisition module, a multi-modal feature extraction module, a bolt looseness intelligent determination module, and a dynamic power management module; The dynamic power management module is respectively connected to the adaptive excitation frequency generation module, the hierarchical anti-interference soundprint acquisition module, the multi-modal feature extraction module, and the bolt looseness intelligent determination module; The adaptive excitation frequency generation module is connected to the hierarchical anti-interference soundprint acquisition module, and is used to output a dynamically adjusted excitation frequency signal and collect the vibration response of the iron tower; The hierarchical anti-interference soundprint acquisition module is connected to the multi-modal feature extraction module, and is used to perform anti-interference processing on the collected soundprint signal and separate the vibration components of the iron tower; The multi-modal feature extraction module is connected to the bolt looseness intelligent determination module, and is used to extract the time-frequency features and deep learning features of the signal; The bolt looseness intelligent determination module is based on a dynamic threshold model to classify and determine the bolt looseness state.

9. The intelligent detection system for tower acoustic fingerprint according to claim 8, wherein The adaptive excitation frequency generation module includes: An AD9833 programmable waveform generator, which is used to output a swept-frequency signal of 10 Hz - 10 kHz; An ADXL357 three-axis vibration sensor, which is used to collect the vibration response data of the iron tower; DS18B20 temperature sensor, used to obtain ambient temperature data; STM32H743 main control chip, used to dynamically adjust the excitation frequency through the gradient descent algorithm; The hierarchical anti-interference voiceprint acquisition module includes: LTC1068 programmable filter, set to band-pass mode, with a center frequency of the excitation frequency and a bandwidth of ±200 Hz; An array composed of 4 ADI-ADMP521-MEMS microphones, implementing spatial beamforming based on the GCC-PHAT algorithm; ADS1278 - 24-bit ADC, used for signal analog-to-digital conversion; The module separates the tower vibration component from the mixed signal through the FastICA algorithm.

10. The intelligent detection system for tower acoustic fingerprint according to claim 8, characterized in that, The dynamic power management module includes: BQ25895 power management chip, supporting voltage switching in the excitation mode, acquisition mode, and standby mode, and the module priority switching logic is: excitation mode > acquisition mode > standby mode; LTC3588 energy recovery circuit, converting vibration energy into 5V DC power through a piezoelectric sensor; CAP-XX supercapacitor, used to store the recovered electrical energy.

Images