Transmission tower bolt looseness detection method based on airborne voiceprint detection technology
Through the on-board soundprint detection technology, the drone is used to collect and analyze the vibration soundprint signals of the transmission pole tower, and combine it with the SVM algorithm to identify the bolt loose state, the problems of low detection efficiency and high cost in the existing technology are solved, and the transmission is achieved The efficient and convenient detection of loose tower bolts is suitable for multi-region inspection in remote areas.
Patent Information
- Application Number
- CN202411040165.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-31
- Publication Date
- 2024-11-15
Smart Images

Figure BDA0004972177570000021 
Figure BDA0004972177570000051 
Figure BDA0004972177570000052
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of transmission tower fault diagnosis, and in particular to a method for detecting loose bolts of a transmission tower based on airborne voiceprint detection technology. Background Art
[0002] In recent years, ice dancing has occurred frequently and over a large area, which can easily cause the bolts of the tower to loosen and fall off, damage to hardware, insulators, jumpers, broken strands and wires, and damage to tower materials, leading to electrical failures. As far as towers are concerned, transmission towers widely use pressure-bearing bolt connections, and loose bolts are the most common form of dancing damage. Loose bolts are the main factor that ultimately causes the tower to deform or collapse, and the deformation of tower materials is a direct manifestation of the destruction of the overall stability of the tower structure. In addition, under the long-term action of wind loads, the bolts on the node plates of transmission towers will vibrate and loosen, causing damage to the bolts at the nodes. If not discovered in time, the damage to the nodes will continue to develop until the transmission tower collapses under the action of strong winds.
[0003] At present, the loosening detection of tower bolts is mainly carried out by regular inspections by operation and maintenance personnel. Some scholars have proposed some new detection technologies, including methods based on acoustic elastic signal detection, vibration signal detection, and piezoelectric impedance detection. The existing methods have the following problems:
[0004] (1) The operation and maintenance personnel regularly inspect the towers and manually check the bolts of each connection. Although this method is effective, it is labor-intensive, time-consuming, costly, and inefficient.
[0005] (2) Technologies such as torque wrench method, acoustic elastic signal detection, and piezoelectric impedance detection mainly rely on monitoring of bolt stress, torque and other state quantities to detect bolt loosening. Currently, they are all used to detect the looseness of a single bolt, and it is impossible to detect the looseness of multiple bolts at the same time. The detection range is limited. There are often thousands of bolts on a tower. If these detection methods are used, a large number of sensors need to be installed, which will result in high detection costs and great difficulty in implementation.
[0006] (3) Vibration signal detection is based on the acceleration data collected during the vibration of the tower steel structure to determine the loosening status of the tower bolts. Therefore, it is often necessary to climb the tower to install the acceleration sensor. However, many high-voltage transmission towers are located in remote mountains with inconvenient transportation. At the same time, there are safety hazards when climbing the tower during the vibration of the line, and the sensor is inconvenient to install. Therefore, it is difficult to carry out large-scale application at present. Summary of the invention
[0007] The purpose of the present invention is to solve the above-mentioned deficiencies in the prior art and thus provide a method for detecting loose bolts of a transmission tower based on airborne voiceprint detection technology.
[0008] A method for detecting loose bolts on a transmission tower based on airborne voiceprint detection technology comprises the following steps:
[0009] Step 1: Partition the transmission towers and build a partition identification model for the transmission towers;
[0010] Step 2: Voiceprint time signal collection
[0011] During the vibration of the transmission line, the drones equipped with the voiceprint time signal collection system are placed close to different areas of the transmission tower to be tested, and the voiceprint time signals of different areas of the transmission tower are collected;
[0012] Step 3: Preprocess the voiceprint time signals of different areas;
[0013] Step 4: Extract features of the voiceprint time signals in different areas after preprocessing to obtain feature vectors of the voiceprint time signals in different areas;
[0014] Step 5: Input the feature vectors of the voiceprint time signals in different areas into different partition recognition models to obtain the looseness status of bolts in different areas of the transmission tower to be detected.
[0015] The transmission tower is divided into four areas: tower head, cross arm, tower body and tower leg.
[0016] Construct a partition identification model for transmission towers, specifically:
[0017] (1) Select multiple transmission towers of the same type with different sizes as test objects, adjust the bolt tightness of different partitions of multiple test transmission towers respectively, and then carry out the transmission line galloping test under different wind speeds on the test transmission towers;
[0018] (2) In the real-life transmission line dancing test, drones equipped with a voiceprint time signal acquisition system were placed close to the same area of multiple test transmission towers to collect voiceprint time signals of the same area of the test transmission towers under different bolt loosening conditions. After preprocessing, feature extraction was performed on the voiceprint time signals of the same area of the test transmission towers under different bolt loosening conditions to obtain the corresponding feature vectors and obtain a training sample set.
[0019] (3) The support vector machine algorithm (SVM) is used to train the training sample set to obtain the bolt loosening recognition model in the same area;
[0020] (4) Repeat steps (2) to (3) to obtain bolt loosening identification models for other areas of transmission towers in other areas.
[0021] The input of the bolt loosening identification model is the feature vector corresponding to the soundprint time signal under different bolt loosening states in the same area, and the output is the loosening state of the tower bolts in the area.
[0022] The tightness of the bolts is adjusted to three states: all bolts are fully tightened, a small number of bolts are loose, and serious bolts are loose.
[0023] Preprocessing includes:
[0024] The voiceprint time signal is processed by framing, and the entire audio of the voiceprint time signal is divided into short audios of 20ms to 30ms. Each short audio is called a frame of audio. The overlap ratio between two frames of audio is about 30%. After the framing is completed, the original continuous voiceprint time signal can be converted into a discrete voiceprint time series;
[0025] Perform windowing on the voiceprint time series.
[0026] The Hamming window is used for windowing operation, which is defined in the time domain as follows:
[0027]
[0028] Feature extraction includes:
[0029] Perform fast Fourier transform (FFT) on the pre-processed voiceprint time signal to obtain the frequency spectrum of each frame of audio;
[0030] Take the square of the audio spectrum of each frame to get the power spectrum X of each frame of audio m (k) 2 ;
[0031] The obtained power spectrum X m (k) 2 After filtering by the Mel filter bank and then summing the energy within each filter range separately, the output of the corresponding number of Mel filters will be obtained through one frame of audio. The logarithm of the output of the Mel filter is taken, and the obtained Log-Mel spectrum feature is the feature vector to be extracted.
[0032] A Mel filter bank typically contains about 20-40 triangular filter banks.
[0033] The present invention uses the feature recognition of soundprint time signals to realize bolt loosening detection. The signal comes from the soundprint time signal transmitted by the vibration of the tower itself under the dancing working condition. At the same time, the drone can freely collect the soundprint time signals of different areas of the tower, without installing sensors on the tower, and realize bolt loosening detection conveniently and efficiently.
[0034] The training sample data of the recognition model of the present invention comes from the real-type dancing test of the power transmission line. A number of tower-type poles of different sizes are selected as objects, and the tightness of the bolts in four areas, namely the tower head, cross arm, tower body and tower leg, are adjusted respectively. Taking the tower head as an example, the three states are adjusted to be all bolts fully tightened, a small number of bolts loosened and serious bolts loosened, and the real-type dancing test of the power transmission line under different wind speeds is carried out.
[0035] The present invention proposes Log-Mel feature vector as the feature quantity of voiceprint time signal, and uses SVM algorithm to establish recognition model, which can judge the three states of tower being completely tightened, partially loosened and severely loosened by region. The acquisition of training samples requires the implementation of dancing real-type test, which is difficult to obtain and the sample size is small. Therefore, the SVM algorithm suitable for small sample data is used for model training.
[0036] The present invention places a voiceprint time signal acquisition module in a drone, and uses the drone to approach a transmission tower to collect voiceprint time signals during the dancing of a transmission line. Voiceprints have the characteristics of non-contact, easy acquisition, and stable characteristics, so there is no need to install any sensors on the tower. It is simple and efficient to use, and is also suitable for tower detection in areas with inconvenient transportation. An artificial intelligence algorithm is used to complete the processes of voiceprint time signal denoising, feature extraction, and pattern recognition, to achieve overall detection of the loose state of bolts on transmission towers, reduce the scope of bolt loosening inspection, and speed up the efficiency of fault repair, which will significantly improve the operation and maintenance level of transmission towers and improve the overall economic benefits.
[0037] Compared with the prior art which uses a sound wave generator to transmit sound wave signals to the target tower and collects sound print signals at the screw connections of the target tower, and the transmitted sound wave signals need to adaptively determine the target transmission power of the sound wave generator based on real-time tower information such as temperature and wind speed, the present invention directly collects the sound print signals generated by the vibration of the tower itself during the dancing process, which is simple and efficient.
[0038] Compared with the prior art, which emits sound waves, collects soundprint signals at the screw connection, and then analyzes and judges the signals, that is, each time the detection object is a bolt, when detecting other bolts, it is necessary to re-determine the power of the sound wave generator, etc., and transmit signals, collect, and analyze again. However, a line often has hundreds or even thousands of pole towers, and each pole tower has thousands of bolts. Therefore, this invention is difficult to apply to actual line inspection and early detection of tower hidden dangers. In comparison, the present invention uses an airborne soundprint collection device to divide the pole tower into 4 areas, detect soundprint signals in different areas, and the detection object is one area. Through signal analysis, it is determined whether there is a loose bolt phenomenon in the area. The detection is efficient, and it can realize the rapid diagnosis of the bolt status of multiple pole towers with the help of drones.
[0039] Compared with the prior art, in which online monitoring equipment can be installed on the tower to be tested in the form of an explosion-proof box, and portable monitoring equipment can be transported to the bottom of the tower in the form of an engineering equipment box, and the target tower is also required to be installed with a patch-type bone conduction voiceprint online sensor, a temperature sensor, a wind speed sensor, a sound wave generator, etc., the present invention does not require the installation of sensors on the tower, and is simple and convenient to use. At the same time, the inspection personnel can remotely control the drone to approach the tower without the need to install it on the tower or operate it under the tower.
[0040] The present invention adopts the SVM algorithm to establish the correspondence between the characteristic vector of the soundprint time signal under different bolt looseness states and the looseness state of the tower bolts, but the target file only mentions the acoustic model, and does not mention the specific algorithm of the acoustic model or how to establish it.
[0041] The present invention adopts airborne voiceprint detection technology to detect the status of tower bolts in different areas, avoiding the installation of a large number of monitoring sensors on the tower. At the same time, the drone can easily realize line inspection and signal collection at different positions (if the sensor is installed, it can only detect the position of the sensor). It can conveniently and efficiently realize bolt loosening detection, reduce the scope of bolt loosening inspection, speed up the efficiency of fault repair, and significantly improve the operation and maintenance level of transmission towers and improve the overall economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 is a flow chart of the inventive method;
[0043] Figure 2 Flowchart for extracting Log-Mel spectrum features of voiceprint time signals. DETAILED DESCRIPTION
[0044] See also Figure 1 , the method of the present invention comprises the following steps:
[0045] Step 1: Divide the transmission tower into four areas: tower head, cross arm, tower body and tower leg, and build a transmission tower partition recognition model
[0046] (1) Taking a certain type of tower as an example (e.g. cathead tower), several towers of this type with different sizes were selected as the objects, and the bolt tightness of the four areas of the tower head, cross arm, tower body and tower leg was adjusted respectively. Taking the tower head as an example, the three states of all bolts being fully tightened, a small number of bolts being loosened, and severe bolts being loosened were adjusted respectively, and a real-life transmission line galloping test under different wind speeds was carried out.
[0047] (2) In the real-type test of transmission line galloping, the soundprint time signals under different bolt looseness states are collected. After preprocessing, multiple groups of Log-Mel feature vectors of the tower head area are obtained. The Log-Mel feature vector obtained for the fully tightened bolt state is numbered 1; the Log-Mel feature vector obtained for the partially loose bolt state is numbered 2; and the Log-Mel feature vector obtained for the severe bolt loose state is numbered 3.
[0048] (3) The training sample data is trained using an artificial intelligence algorithm. The model input is the Log-Mel feature vector of the voiceprint time signal, and the output is the loosening state of the tower bolts, which is divided into three types: fully tightened, partially loose, and severely loose. The artificial intelligence algorithm for pattern recognition uses a conventional SVM (support vector machine) algorithm because the acquisition of training samples requires a real-type dancing test, which is difficult to obtain and the sample size is small. Therefore, an SVM algorithm suitable for small sample data is used for model training. After the training is completed, the bolt loosening recognition model Model 1 at the tower head is obtained.
[0049] (4) Repeat steps (2) to (3) for the training samples of the other three areas, namely the cross arm, tower body and tower leg, and use the SVM algorithm to perform pattern recognition training to obtain recognition models Model 2, Model 3 and Model 4.
[0050] Step 2: Voiceprint time signal collection
[0051] The non-contact microphone and the voiceprint time signal collection device are integrated together to form a transmission tower voiceprint time signal collection system, which is connected to the drone gimbal, and the drone gimbal realizes the power supply, recording and signal transmission of the collection system. During the dancing of the transmission line, the transmission tower vibrates greatly. At this time, the drone is placed close to the tower head, cross arm, tower body and tower leg to collect the voiceprint time signals I1, I2, I3 and I4 during the vibration of the tower.
[0052] Step 3: Voiceprint time signal preprocessing. The preprocessing process of this step is the same as the preprocessing process in step (2).
[0053] Frame processing is performed on individual voiceprint time signals, and the entire audio is divided into short audio segments of 20ms to 30ms. Each short audio segment is called an audio frame. When framing, the time-varying nature of the audio signal is still taken into account, and there must be overlapping parts between adjacent audio frames, so that the extracted features can be more accurate. Preferably, the overlapping ratio between two frames is about 30%. After the framing is completed, the original continuous voiceprint time signal can be converted into a discrete time series vector.
[0054] The voiceprint time series is windowed because: after the audio signal is framed, the signal at the boundary of each frame will drop sharply to zero, causing spectrum leakage, and the purpose of windowing is to allow the signals at both ends of the audio to smoothly transition to zero and reduce the truncation effect. The present invention uses a Hamming window for windowing, which is defined in the time domain as follows:
[0055]
[0056] Step 4: Feature extraction of voiceprint time signal
[0057] Log-Mel spectrum features are used to represent the characteristics of the perceived voiceprint in a given sample. The mapping relationship between Mel frequency and Hz frequency is as follows:
[0058]
[0059] The extraction process of Log-Mel spectrum features is as follows Figure 2 shown.
[0060] The specific process is as follows: Perform a 2048-point Fast Fourier Transform (FFT) on the preprocessed sound signal. The reason for choosing 2048 points is that the FFT transformation requires the input data length to be an exponential multiple of 2, and when framing, one frame is 1200 sampling points, so zeros are padded to 2048 points and then FFT transformation is performed to obtain the spectrum of each frame. Then take the square of the audio signal spectrum to obtain the power spectrum Xm(k)2 of each point of the speech signal.
[0061] After obtaining the power spectrum Xm(k)2, it is filtered through a Mel filter bank, which usually contains about 20-40 (usually 26) triangular filter banks. The energy within each filter range is then summed separately, and the output of the corresponding number of Mel filters is obtained through a frame of audio signal. Finally, the logarithm of these outputs is taken to obtain the Log-Mel spectrum feature.
[0062] Taking the tower top voiceprint time signal I1 collected in step 2 as an example, assuming that I1 is divided into M frames, Log-Mel spectrum feature extraction is performed on each frame signal, and M groups of Log-Mel feature vectors are obtained, which are used as the feature quantity of the voiceprint time signal for subsequent identification.
[0063] Step 5: Practical application of voiceprint recognition
[0064] The characteristic vectors of the voiceprint time signals of different areas obtained in step 4 are input into the looseness recognition models Model1 to Model4 trained in advance, and the looseness status of the bolts in different areas of the transmission tower can be output.
[0065] Compared with the traditional transmission tower bolt loosening detection technology, the present invention does not need to install any sensors on the tower, is simple and efficient to use, and is also suitable for tower detection in areas with inconvenient transportation. It uses artificial intelligence algorithms to complete the processes of voiceprint time signal denoising, feature extraction and pattern recognition, to achieve overall detection of the loose state of transmission tower bolts, reduce the scope of bolt loosening investigation, speed up fault repair efficiency, and significantly improve the operation and maintenance level of transmission towers and improve overall economic benefits.
[0066] The present invention is airborne, and a voiceprint time signal acquisition system composed of a non-contact microphone and a voiceprint signal acquisition unit is integrated on a drone, and the power supply is provided by a pan-tilt platform; since the voiceprint acquisition device is on the drone, the drone can freely move close to different areas of the pole tower (compared to the method of directly installing a voiceprint or vibration signal sensor on the pole tower), and it is simple and efficient to realize the partition judgment of whether the bolts of the pole tower are loose; the generation of the voiceprint signal does not rely on external objects, and the prior art needs to use a sound wave generator, and use a sledgehammer to hit the bottom of the tower to generate the voiceprint signal. The present invention directly uses the voiceprint signal of the pole tower's own vibration during the dancing process. The line dancing is a self-excited vibration with a fixed frequency. The loosening of the pole tower bolts will change the frequency of the vibration sound wave and other characteristics. The training sample is obtained based on the real dancing test, which is difficult to obtain and the sample size is small. Therefore, the SVM algorithm suitable for small sample data is used for model training.
Claims
1. A method for detecting loose bolts on a transmission tower based on airborne voiceprint detection technology, characterized in that: The following steps are involved: Step 1: Partition the transmission towers and build a partition identification model for the transmission towers; Step 2: Voiceprint time signal collection During the vibration of the transmission line, the drones equipped with the voiceprint time signal collection system are placed close to different areas of the transmission tower to be tested, and the voiceprint time signals of different areas of the transmission tower are collected; Step 3: Preprocess the voiceprint time signals of different areas; Step 4: Extract features of the voiceprint time signals in different areas after preprocessing to obtain feature vectors of the voiceprint time signals in different areas; Step 5: Input the feature vectors of the voiceprint time signals in different areas into different partition recognition models to obtain the looseness status of bolts in different areas of the transmission tower to be detected.
2. According to claim 1, a method for detecting loose bolts of transmission towers based on airborne voiceprint detection technology is characterized in that: The transmission tower is divided into four areas: tower head, cross arm, tower body and tower leg.
3. According to claim 1, a method for detecting loose bolts of transmission towers based on airborne voiceprint detection technology is characterized in that: Construct a partition identification model for transmission towers, specifically: (1) Select multiple transmission towers of the same type with different sizes as test objects, adjust the bolt tightness of different partitions of multiple test transmission towers respectively, and then carry out the transmission line galloping test under different wind speeds on the test transmission towers; (2) In the real-life transmission line dancing test, drones equipped with a voiceprint time signal acquisition system were placed close to the same area of multiple test transmission towers to collect voiceprint time signals of the same area of the test transmission towers under different bolt loosening conditions. After preprocessing, feature extraction was performed on the voiceprint time signals of the same area of the test transmission towers under different bolt loosening conditions to obtain the corresponding feature vectors and obtain a training sample set. (3) The support vector machine algorithm (SVM) is used to train the training sample set to obtain the bolt loosening recognition model in the same area; (4) Repeat steps (2) to (3) to obtain bolt loosening identification models for other areas of transmission towers in other areas.
4. The method for detecting loose bolts of transmission towers based on airborne voiceprint detection technology according to claim 3 is characterized in that: The input of the bolt loosening identification model is the feature vector corresponding to the soundprint time signal under different bolt loosening states in the same area, and the output is the loosening state of the tower bolts in the area.
5. According to claim 3 or a method for detecting loose bolts on transmission towers based on airborne voiceprint detection technology, it is characterized in that: The tightness of the bolts is adjusted to three states: all bolts are fully tightened, a small number of bolts are loose, and serious bolts are loose.
6. A method for detecting loose bolts on transmission towers based on airborne voiceprint detection technology according to any one of claims 1 to 5, characterized in that: Preprocessing includes: The voiceprint time signal is processed by framing, and the entire audio of the voiceprint time signal is divided into short audios of 20ms to 30ms. Each short audio is called a frame of audio. The overlap ratio between two frames of audio is about 30%. After the framing is completed, the original continuous voiceprint time signal can be converted into a discrete voiceprint time series; Perform windowing on the voiceprint time series.
7. According to claim 6 or a method for detecting loose bolts on transmission towers based on airborne voiceprint detection technology, it is characterized in that: The Hamming window is used for windowing operation, and its definition in the time domain is as follows:
8. A method for detecting loose bolts on transmission towers based on airborne voiceprint detection technology according to claim 1-5 or 7, characterized in that: Feature extraction includes: Perform fast Fourier transform (FFT) on the pre-processed voiceprint time signal to obtain the frequency spectrum of each frame of audio; Take the square of the audio spectrum of each frame to get the power spectrum X of each frame of audio m (k) 2 ; The obtained power spectrum X m (k) 2 After filtering by the Mel filter bank and then summing the energy within each filter range separately, the output of the corresponding number of Mel filters will be obtained through one frame of audio. The logarithm of the output of the Mel filter is taken, and the obtained Log-Mel spectrum feature is the feature vector to be extracted.
9. The method for detecting loose bolts of transmission towers based on airborne voiceprint detection technology according to claim 8, characterized in that: A Mel filter bank typically contains about 20-40 triangular filter banks.
Citation Information
Cited By
Transmission tower bolt looseness detection and positioning method and system based on fractional Fourier transform
CN119245912A
A method and system for detecting and locating loose bolts of transmission towers based on fractional Fourier transform
CN119245912B
Target detection method for patrol inspection of broken strand conductor of power transmission line by unmanned aerial vehicle
CN120544084A
