A power equipment intelligent inspection system for multi-scene adaptation

Abstract

The application discloses a kind of for multi-scene adaptive power equipment intelligent inspection system, it is related to power equipment inspection field, wherein scene voiceprint modeling module is generated by laser radar scanning scene three-dimensional model and labels equipment identity data, and audio signal is collected to equipment, after extracting spectral feature, it is clustered, and it is constructed initial voiceprint model to remove anomaly;Adaptive learning module obtains real-time temperature with temperature sensor, calculates temperature influence coefficient and correction value, and obtains voiceprint baseline model by correcting initial model and is regularly updated;Inspection positioning module obtains the coordinates and attitude of inspection equipment, calculates its distance with equipment, azimuth angle, and generates predicted synthetic voiceprint data according to formula;Abnormality identification module compares real-time voiceprint data with prediction data, and identifies anomaly outside fluctuation range, and controls inspection equipment to move positioning abnormal equipment according to distance.The system solves the problem of multi-device sound source interference and temperature influence, and improves the reliability and abnormal positioning efficiency of inspection.

Description

Technical Field

[0001] This invention relates to the field of power equipment inspection, specifically a multi-scenario adaptive intelligent power equipment inspection system. Background Technology

[0002] As a core component of power transmission and transformation systems and new energy power plants, the operating status of power equipment directly determines the stability and security of energy supply. Traditional manual inspections are limited by low efficiency, poor environmental adaptability, and slow response, making it difficult to meet the real-time monitoring needs of digital power grids.

[0003] Voiceprint recognition technology, due to its advantages of being non-contact, easy to deploy, and capable of capturing early signs of faults, has become a core means of intelligent inspection. By analyzing the frequency, amplitude, and other characteristics of the operating sounds of equipment, it can achieve early warning of faults, which is of irreplaceable significance for ensuring the reliable operation of power systems. However, existing voiceprint inspection technology for power equipment faces multiple bottlenecks in multi-scenario applications: when multiple devices are running simultaneously in a scenario, mismatches in the direction angle of the sound source can easily occur, generating spurious sound sources and causing deviations in the location of abnormal equipment; during the movement of inspection equipment, instantaneous background noise and positional changes can easily trigger false alarms. These problems restrict the large-scale application of voiceprint inspection technology in multiple scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent inspection system for power equipment that is adaptive to multiple scenarios, in order to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A smart inspection system for power equipment with multi-scenario adaptation, the system includes a scene voiceprint modeling module, an adaptive learning module, an inspection positioning module and an anomaly recognition module;

[0007] The scene voiceprint modeling module is used to construct a spatial model of the power equipment in the scene, extract the voiceprint features of the power equipment, and construct an initial voiceprint model.

[0008] Furthermore, the specific steps for constructing the spatial model are as follows:

[0009] The scenario is the area where the power equipment is located when it is working. The scenario is scanned by LiDAR to obtain point cloud data, and a three-dimensional model of the scenario is generated based on the point cloud data. In the three-dimensional model, the type of power equipment and the three-dimensional coordinates of the power equipment are obtained by image recognition algorithm. The power equipment is numbered according to its type. Different power equipment in the scenario are associated and labeled according to the number and the three-dimensional coordinates of the power equipment to form equipment identity data.

[0010] The specific steps for constructing the initial voiceprint model are as follows:

[0011] The soundprint collection location for each power device in the preset scene is determined, and all soundprint collection locations are the same as the relative positions of their corresponding power devices. Sound level meters are placed in the scene to collect the audio signals and sound pressure levels of each power device during operation according to the preset sampling frequency and sampling duration, and the audio timestamps of the audio signals are obtained.

[0012] The audio signal is divided into several frames according to a preset frame length and frame shift; a fast Fourier transform is performed on each frame to convert each frame of the audio signal from the time domain to the frequency domain, generating a spectrum diagram; spectral features are extracted from the spectrum diagram, including the fundamental frequency, harmonic amplitude, frequency centroid, and spectral width.

[0013] The spectral features are concatenated sequentially to obtain the voiceprint feature vector. The voiceprint feature vector is then clustered according to the type of power equipment to obtain a cluster of voiceprint feature vectors for each type of power equipment. The mean and variance of each spectral feature in the cluster are calculated, and a feature fluctuation range is preset based on the mean and variance. Spectral features outside the feature fluctuation range are removed as abnormal features. The mean of each spectral feature after removing abnormal features is recalculated as the baseline data for each spectral feature. An initial voiceprint model is constructed based on the baseline data and the feature fluctuation range.

[0014] The adaptive learning module is used to correct the initial voiceprint model according to the temperature in the scene, obtain the voiceprint baseline model, and update the voiceprint baseline model.

[0015] Furthermore, the specific steps for correcting the initial voiceprint model are as follows:

[0016] Several temperature sensors are deployed in the scene to acquire the real-time average temperature of the scene at a preset frequency, and the temperature timestamp of the real-time average temperature is also acquired. A preset standard temperature is used to calculate the difference ΔT between the real-time average temperature and the standard temperature. The temperature timestamp of the difference ΔT is correlated with the audio timestamp to obtain the spectral characteristics corresponding to the difference. The rate of change of the spectral characteristics with temperature is calculated based on the difference, which is used as the temperature influence coefficient α. A correction value E is calculated based on the difference and the temperature influence coefficient. i =ΔT×α i E i The correction value α represents the i-th spectral feature. i The temperature influence coefficient representing the i-th spectral characteristic;

[0017] The baseline data in the initial voiceprint model is added to the corresponding correction value to obtain the correction data, and the corresponding feature fluctuation range is adjusted according to the ratio of the change between the correction data and the baseline data to obtain the correction fluctuation range; a voiceprint baseline model is constructed based on the correction data and the correction fluctuation range; the voiceprint baseline model is associated with the device identity data;

[0018] The specific steps for updating the voiceprint baseline model are as follows:

[0019] A preset update cycle is set. At the end of each update cycle, the mean value of each spectral feature after removing abnormal features in that update cycle is obtained according to the audio timestamp, and used as new baseline data. The voiceprint baseline model is then updated based on the new baseline data.

[0020] The inspection and positioning module is used to control the inspection equipment to move along a preset route, obtain the inspection coordinates of the inspection equipment in the spatial model, and obtain the predicted synthetic voiceprint data of the current inspection coordinates based on the inspection coordinates and the voiceprint baseline model of the power equipment.

[0021] Furthermore, the specific steps for obtaining the inspection coordinates of the inspection equipment are as follows:

[0022] When the inspection equipment enters the scene, the coordinates of the inspection equipment in the three-dimensional model are acquired in real time. The acceleration and angular velocity of the inspection equipment are collected by the IMU mounted on the inspection equipment, and the attitude data of the inspection equipment is calculated. The attitude data includes yaw angle, pitch angle and roll angle. The inspection coordinates include the coordinates and attitude data of the inspection equipment in the three-dimensional model.

[0023] The specific steps for obtaining the predicted synthetic voiceprint of the current inspection coordinates are as follows:

[0024] Calculate the straight-line distance between the inspection coordinates and the three-dimensional coordinates of the power equipment; establish a horizontal coordinate system with the inspection coordinates as the origin, and calculate the azimuth angle between the inspection coordinates and the three-dimensional coordinates of the power equipment in the horizontal coordinate system; establish a location information dataset, which includes the equipment identification data of the power equipment and the straight-line distance and azimuth angle corresponding to the equipment identification data.

[0025] Calculate the theoretical voiceprint data at the inspection coordinates for each type of correction data in the voiceprint baseline model based on the straight-line distance:

[0026] ;

[0027] Where A represents theoretical voiceprint data, A0 represents the correction data corresponding to the theoretical voiceprint data, lg represents logarithmic operation, r0 represents the distance between the voiceprint acquisition location and the power equipment, and r represents the straight-line distance in the location information dataset of the power equipment.

[0028] A three-dimensional coordinate system centered on the inspection equipment is established. The sound propagation direction of each power device is obtained according to the azimuth angle. The theoretical voiceprint data of each power device is vector-superimposed according to the type based on the sound propagation direction to obtain the predicted synthetic voiceprint data. The average value and variance of the modulus of the predicted synthetic voiceprint data of the inspection coordinate in several update cycles are calculated. The fluctuation range of the synthetic voiceprint data is preset according to the average value and variance of the modulus.

[0029] The anomaly identification module is used to collect real-time synthesized voiceprint data and compare it with the predicted synthesized voiceprint data. Based on the comparison results, anomalies are identified and located.

[0030] Furthermore, the specific steps for collecting real-time synthesized voiceprint data are as follows:

[0031] A sound level meter is installed on the inspection equipment, and the soundprint feature vector at the inspection coordinate is obtained in the same way as in the scene soundprint modeling module, which is used as real-time synthesized soundprint data.

[0032] The specific steps for comparing and identifying anomalies based on the comparison results are as follows: obtain the fluctuation range of the synthetic voiceprint data of the inspection coordinate, calculate the difference between the modulus of the predicted synthetic voiceprint data and the real-time synthetic voiceprint data, and compare whether the difference is within the fluctuation range of the synthetic voiceprint data; if there is a difference outside the fluctuation range of the synthetic voiceprint data, it is identified as an anomaly, otherwise it is identified as normal; when it is identified as normal, the inspection equipment continues to move.

[0033] The specific steps for locating the anomaly are as follows:

[0034] Set a movement point for each power device in the scene; use the spectral features corresponding to the differences identified as abnormal as abnormal features; sort the power devices in ascending order of the straight-line distance between the current inspection coordinates and each power device, and control the inspection device to move to the movement point of the power device in ascending order of the straight-line distance;

[0035] The difference between the modulus of the predicted synthetic voiceprint data and the real-time synthetic voiceprint data corresponding to the abnormal features at the moving point is calculated as the positioning difference. A moving cycle is set. When the number of moving points traversed by the inspection equipment reaches the moving cycle, the moving point with the largest positioning difference within the moving cycle is compared. The power equipment corresponding to this moving point is identified as an abnormal device. The device identity data of the abnormal device is obtained, and the three-dimensional coordinates and number of the power equipment in the device identity data are extracted as the abnormal coordinates and abnormal number.

[0036] The output of the scene voiceprint modeling module is connected to the input of the adaptive learning module; the output of the adaptive learning module is connected to the input of the inspection and positioning module; and the output of the inspection and positioning module is connected to the input of the anomaly recognition module.

[0037] Compared with the prior art, the beneficial effects of the present invention are:

[0038] 1. This invention effectively solves the problem of sound source interference when multiple devices are running simultaneously by constructing a three-dimensional model and superimposing theoretical voiceprint data vectors. The system can accurately separate mixed voiceprints and match the direction of the sound source, significantly reducing the recognition rate of spurious sound sources, achieving accurate positioning of abnormal devices, and improving the reliability of inspection in complex scenarios.

[0039] 2. This invention achieves adaptive adjustment of the voiceprint baseline model with ambient temperature by introducing a temperature sensor and a dynamic correction mechanism for the voiceprint model. By calculating the temperature influence coefficient of the spectral characteristics and generating the corrected voiceprint baseline model, the interference caused by temperature changes on voiceprint characteristics is overcome, improving the accuracy of anomaly identification and the system's environmental adaptability.

[0040] 3. This invention achieves rapid tracking and identification of abnormal equipment by designing a method for comparing moving point scanning and positioning differences. The system controls the inspection equipment to move according to distance priority, and combines multi-cycle data comparison to lock the coordinates and number of the equipment with the greatest abnormal characteristics, which greatly improves the efficiency of fault location and reduces the risk of missed detection. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating an intelligent power equipment inspection system for multi-scenario adaptation according to the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] Example: Figure 1 As shown, the present invention provides a technical solution: an intelligent inspection system for power equipment with multi-scenario adaptability. The system includes a scene voiceprint modeling module, an adaptive learning module, an inspection positioning module, and an anomaly recognition module.

[0044] The scene voiceprint modeling module is used to construct a spatial model of the power equipment in the scene, extract the voiceprint features of the power equipment, and construct an initial voiceprint model.

[0045] The specific steps for constructing the spatial model are as follows:

[0046] The scenario is the area where the power equipment is located when it is working. The scenario is scanned by LiDAR to obtain point cloud data, and a three-dimensional model of the scenario is generated based on the point cloud data. In the three-dimensional model, the type of power equipment and the three-dimensional coordinates of the power equipment are obtained by image recognition algorithm. The power equipment is numbered according to its type. Different power equipment in the scenario are associated and labeled according to the number and the three-dimensional coordinates of the power equipment to form equipment identity data.

[0047] The specific steps for constructing the initial voiceprint model are as follows:

[0048] The soundprint collection location for each power device in the preset scene is determined, and all soundprint collection locations are the same as the relative positions of their corresponding power devices. Sound level meters are placed in the scene to collect the audio signals and sound pressure levels of each power device during operation according to the preset sampling frequency and sampling duration, and the audio timestamps of the audio signals are obtained.

[0049] The audio signal is divided into several frames according to a preset frame length and frame shift; a fast Fourier transform is performed on each frame to convert each frame of the audio signal from the time domain to the frequency domain, generating a spectrum diagram; spectral features are extracted from the spectrum diagram, including the fundamental frequency, harmonic amplitude, frequency centroid, and spectral width.

[0050] The spectral features are concatenated sequentially to obtain the voiceprint feature vector. The voiceprint feature vector is then clustered according to the type of power equipment to obtain a cluster of voiceprint feature vectors for each type of power equipment. The mean and variance of each spectral feature in the cluster are calculated, and a feature fluctuation range is preset based on the mean and variance. Spectral features outside the feature fluctuation range are removed as abnormal features. The mean of each spectral feature after removing abnormal features is recalculated as the baseline data for each spectral feature. An initial voiceprint model is constructed based on the baseline data and the feature fluctuation range.

[0051] The adaptive learning module is used to correct the initial voiceprint model according to the temperature in the scene, obtain the voiceprint baseline model, and update the voiceprint baseline model.

[0052] The specific steps for correcting the initial voiceprint model are as follows:

[0053] Several temperature sensors are deployed in the scene to acquire the real-time average temperature of the scene at a preset frequency, and the temperature timestamp of the real-time average temperature is also acquired. A preset standard temperature is used to calculate the difference ΔT between the real-time average temperature and the standard temperature. The temperature timestamp of the difference ΔT is correlated with the audio timestamp to obtain the spectral characteristics corresponding to the difference. The rate of change of the spectral characteristics with temperature is calculated based on the difference, which is used as the temperature influence coefficient α. A correction value E is calculated based on the difference and the temperature influence coefficient. i =ΔT×α i E i The correction value α represents the i-th spectral feature. i The temperature influence coefficient representing the i-th spectral characteristic;

[0054] The baseline data in the initial voiceprint model is added to the corresponding correction value to obtain the correction data, and the corresponding feature fluctuation range is adjusted according to the ratio of the change between the correction data and the baseline data to obtain the correction fluctuation range; a voiceprint baseline model is constructed based on the correction data and the correction fluctuation range; the voiceprint baseline model is associated with the device identity data;

[0055] The specific steps for updating the voiceprint baseline model are as follows:

[0056] A preset update cycle is set. At the end of each update cycle, the mean value of each spectral feature after removing abnormal features in that update cycle is obtained according to the audio timestamp, and used as new baseline data. The voiceprint baseline model is then updated based on the new baseline data.

[0057] The inspection and positioning module is used to control the inspection equipment to move along a preset route, obtain the inspection coordinates of the inspection equipment in the spatial model, and obtain the predicted synthetic voiceprint data of the current inspection coordinates based on the inspection coordinates and the voiceprint baseline model of the power equipment.

[0058] The specific steps for obtaining the inspection coordinates of the inspection equipment are as follows:

[0059] When the inspection equipment enters the scene, the coordinates of the inspection equipment in the three-dimensional model are acquired in real time. The acceleration and angular velocity of the inspection equipment are collected by the IMU mounted on the inspection equipment, and the attitude data of the inspection equipment is calculated. The attitude data includes yaw angle, pitch angle and roll angle. The inspection coordinates include the coordinates and attitude data of the inspection equipment in the three-dimensional model.

[0060] The specific steps for obtaining the predicted synthetic voiceprint of the current inspection coordinates are as follows:

[0061] Calculate the straight-line distance between the inspection coordinates and the three-dimensional coordinates of the power equipment; establish a horizontal coordinate system with the inspection coordinates as the origin, and calculate the azimuth angle between the inspection coordinates and the three-dimensional coordinates of the power equipment in the horizontal coordinate system; establish a location information dataset, which includes the equipment identification data of the power equipment and the straight-line distance and azimuth angle corresponding to the equipment identification data.

[0062] Calculate the theoretical voiceprint data at the inspection coordinates for each type of correction data in the voiceprint baseline model based on the straight-line distance:

[0063] ;

[0064] Where A represents theoretical voiceprint data, A0 represents the correction data corresponding to the theoretical voiceprint data, lg represents logarithmic operation, r0 represents the distance between the voiceprint acquisition location and the power equipment, and r represents the straight-line distance in the location information dataset of the power equipment.

[0065] A three-dimensional coordinate system centered on the inspection equipment is established. The sound propagation direction of each power device is obtained according to the azimuth angle. The theoretical voiceprint data of each power device is vector-superimposed according to the type based on the sound propagation direction to obtain the predicted synthetic voiceprint data. The average value and variance of the modulus of the predicted synthetic voiceprint data of the inspection coordinate in several update cycles are calculated. The fluctuation range of the synthetic voiceprint data is preset according to the average value and variance of the modulus.

[0066] The anomaly identification module is used to collect real-time synthesized voiceprint data and compare it with the predicted synthesized voiceprint data. Based on the comparison results, anomalies are identified and located.

[0067] The specific steps for collecting real-time synthesized voiceprint data are as follows:

[0068] A sound level meter is installed on the inspection equipment, and the soundprint feature vector at the inspection coordinate is obtained in the same way as in the scene soundprint modeling module, which is used as real-time synthesized soundprint data.

[0069] The specific steps for comparing and identifying anomalies based on the comparison results are as follows: obtain the fluctuation range of the synthetic voiceprint data of the inspection coordinate, calculate the difference between the modulus of the predicted synthetic voiceprint data and the real-time synthetic voiceprint data, and compare whether the difference is within the fluctuation range of the synthetic voiceprint data; if there is a difference outside the fluctuation range of the synthetic voiceprint data, it is identified as an anomaly, otherwise it is identified as normal; when it is identified as normal, the inspection equipment continues to move.

[0070] The specific steps for locating the anomaly are as follows:

[0071] Set a movement point for each power device in the scene; use the spectral features corresponding to the differences identified as abnormal as abnormal features; sort the power devices in ascending order of the straight-line distance between the current inspection coordinates and each power device, and control the inspection device to move to the movement point of the power device in ascending order of the straight-line distance;

[0072] The difference between the modulus of the predicted synthetic voiceprint data and the real-time synthetic voiceprint data corresponding to the abnormal features at the moving point is calculated as the positioning difference. A moving cycle is set. When the number of moving points traversed by the inspection equipment reaches the moving cycle, the moving point with the largest positioning difference within the moving cycle is compared. The power equipment corresponding to this moving point is identified as an abnormal device. The device identity data of the abnormal device is obtained, and the three-dimensional coordinates and number of the power equipment in the device identity data are extracted as the abnormal coordinates and abnormal number.

[0073] The output of the scene voiceprint modeling module is connected to the input of the adaptive learning module; the output of the adaptive learning module is connected to the input of the inspection and positioning module; and the output of the inspection and positioning module is connected to the input of the anomaly recognition module.

[0074] For example:

[0075] Spatial model construction:

[0076] A LiDAR with a scanning frequency of 10Hz and a point cloud data density of 500 points / ㎡ was used to scan the equipment area of ​​the substation to generate a 3D model;

[0077] Three types of electrical equipment were identified using the YOLOv8 image recognition algorithm:

[0078] Main transformer, numbered T1: 3D coordinates (20.5m, 15.3m, 5.2m), T2: (25.8m, 15.5m, 5.3m);

[0079] High-voltage circuit breakers, numbered B1: (19.5m, 16.2m, 5.3m), B2: (21.2m, 18.0m, 5.2m), B3: (24.0m, 18.3m, 5.4m), B4: (26.5m, 16.5m, 5.3m);

[0080] Disconnecting switches, numbered S1-S6, have the following coordinate range:

[0081] (18.0m-27.0m,14.0m-19.0m,4.8m-5.0m);

[0082] The equipment identification dataset is formed by associating equipment type number with three-dimensional coordinates.

[0083] Initial voiceprint model construction:

[0084] The preset location for collecting voiceprints from all devices is 2m away from the device. Sound level meters are set up at the collection locations, the sampling frequency is set to 44.1kHz, the sampling duration is 60s, and the audio signals and sound pressure levels of each device are collected. The audio timestamps are accurate to milliseconds.

[0085] The audio signal is divided into frames with a frame length of 20ms and a frame shift of 10ms. A fast Fourier transform is performed on each frame to generate a spectrum. The spectrum features are extracted. Taking T1 as an example: the average fundamental frequency is 50Hz, the average harmonic amplitude is 80dB, the average frequency centroid is 1000Hz, and the average spectrum width is 500Hz. The variances of each feature are 0.5Hz, 2dB, 50Hz, and 20Hz, respectively.

[0086] The voiceprint feature vectors are clustered according to the device type. For example, T1 and T2 are in one class, B1-B4 are in another class, and S1-S6 are in another class. The preset feature fluctuation range is the mean ± 2 times the variance. Abnormal features that exceed the range are removed, and the mean is recalculated as the baseline data. The initial voiceprint model is constructed by combining the feature fluctuation range.

[0087] Initial voiceprint model correction: Five temperature sensors are placed at the four corners and center of the scene, with a sampling frequency of 1 time / minute and a preset standard temperature of 25℃. At a certain moment, the real-time average temperature is collected at 28℃, and ΔT=3℃ is calculated. The temperature timestamp is associated with the audio timestamp of T1. The temperature influence coefficient is calculated as follows: the fundamental frequency of T1 is α1=0.1Hz / ℃, the harmonic amplitude is α2=0.5dB / ℃, the frequency centroid is α3=10Hz / ℃, and the spectrum width is α4=5Hz / ℃. The correction values ​​are calculated as follows: E1=3×0.1=0.3Hz, E2=3×0.5=1.5dB, E3=3×10=30Hz, and E4=3×5=15Hz.

[0088] The baseline data of the initial voiceprint model is added to the correction value. The feature fluctuation range is adjusted according to the change ratio between the correction data and the baseline data. The voiceprint baseline model is constructed and associated with the device identity data. For example, the voiceprint baseline model of T1 is associated with the number T1 and the coordinates (20.5m, 15.3m, 5.2m).

[0089] Voiceprint baseline model update: The preset update cycle is 24 hours. After each cycle, the mean value of the spectral features after removing abnormal features is extracted based on the audio timestamp. For example, the new fundamental frequency mean value of T1 is 50.2Hz. The voiceprint baseline model is updated with the new mean value.

[0090] Inspection coordinate acquisition: The inspection equipment uses a drone, which acquires its coordinates in the 3D model in real time after entering the scene, for example, (18.2m, 14.5m, 6.0m) at a certain moment; the drone's onboard IMU collects acceleration and angular velocity, for example, acceleration: 0.2m / s², 0.1m / s², -9.8m / s²; angular velocity: 0.05rad / s, 0.03rad / s, 0.01rad / s;

[0091] Calculate attitude data: yaw angle 15°, pitch angle -5°, roll angle 2°;

[0092] The inspection coordinates are set at (18.2m, 14.5m, 6.0m), 15°, -5°, and 2°.

[0093] Predictive synthetic voiceprint data acquisition: Calculate the straight-line distance between the inspection coordinates and T1: ;

[0094] A horizontal coordinate system is established with the inspection coordinates as the origin. The azimuth angle is calculated to be 30°. A location information dataset is constructed, including T1 number, 2.5m distance, and 30° azimuth angle. The theoretical voiceprint data is calculated according to the formula A=A0×10^lg(r0 / r), where r0=2m is the distance between the voiceprint acquisition location and the device, r=2.5m is the distance between the inspection coordinates and the device, T1 fundamental frequency A0=50.5Hz, lg(2 / 2.5)=lg0.8≈-0.0969, so A=50.5×10^(-0.0969)≈50.5×0.8≈40.4Hz;

[0095] The harmonic amplitude of T1 is A0 = 81.5dB, and similarly, A = 81.5 × 0.8 ≈ 65.2dB. A three-dimensional coordinate system is established with the UAV as the center, and the sound propagation direction of each device is determined according to the azimuth angle. The theoretical voiceprint data of T1, T2, and B1 are vector-superimposed according to the type to obtain the predicted synthetic voiceprint data, for example, the fundamental frequency is 48.6Hz and the harmonic amplitude is 63.5dB.

[0096] With a 24-hour update cycle, the mean fundamental frequency of the predicted synthetic voiceprint modulus for the inspection coordinates is calculated to be 48.6Hz and the variance is 0.2Hz within 10 update cycles; the preset synthetic voiceprint fluctuation range is ±0.4Hz.

[0097] Real-time synthesized voiceprint acquisition and anomaly identification: A sound level meter is deployed on the drone, and real-time synthesized voiceprint data is acquired using the same method as the scene voiceprint modeling module; the difference in modulus between the calculated and predicted data, such as the fundamental frequency difference |49.7-48.6|=1.1Hz, exceeds the ±0.4Hz fluctuation range and is identified as an anomaly, with the fundamental frequency as the anomaly characteristic.

[0098] Anomaly localization: A movement point is set for each device, 1m away from the device. For example, the coordinates of movement point B1 are (19.2m, 16.2m, 5.3m). Devices are sorted by straight-line distance from their inspection coordinates, with B1 being the closest at 1.8m, followed by T1 at 2.5m. The drone is controlled to move towards the movement points in ascending order of distance. A movement cycle of 3 movement points is set, and the positioning difference at each movement point is calculated.

[0099] B1 moving point: predicted fundamental frequency 47.2Hz, real-time fundamental frequency 49.7Hz, difference 2.5Hz;

[0100] T1 moving point: predicted base frequency 48.0Hz, real-time base frequency 49.7Hz, difference 1.7Hz;

[0101] B2 moving point: predicted base frequency 49.0Hz, real-time base frequency 49.7Hz, difference 0.7Hz;

[0102] The comparison showed that the B1 movement point had the largest positioning difference, thus identifying B1 as an abnormal device. The device identification data of B1 was extracted: device number B1, three-dimensional coordinates (19.5m, 16.2m, 5.3m). The abnormal location was completed.

[0103] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A smart inspection system for power equipment with multi-scenario adaptability, characterized in that: The system includes a scene voiceprint modeling module, an adaptive learning module, an inspection and positioning module, and an anomaly recognition module; The scene voiceprint modeling module is used to construct a spatial model of the power equipment in the scene, extract the voiceprint features of the power equipment, and construct an initial voiceprint model. The adaptive learning module is used to correct the initial voiceprint model according to the temperature in the scene, obtain the voiceprint baseline model, and update the voiceprint baseline model. The inspection and positioning module is used to control the inspection equipment to move along a preset route, obtain the inspection coordinates of the inspection equipment in the spatial model, and obtain the predicted synthetic voiceprint data of the current inspection coordinates based on the inspection coordinates and the voiceprint baseline model of the power equipment. The anomaly identification module is used to collect real-time synthesized voiceprint data and compare it with the predicted synthesized voiceprint data. Based on the comparison results, anomalies are identified and located. The output of the scene voiceprint modeling module is connected to the input of the adaptive learning module; the output of the adaptive learning module is connected to the input of the inspection and positioning module; and the output of the inspection and positioning module is connected to the input of the anomaly recognition module.

2. The intelligent inspection system for power equipment with multi-scenario adaptation according to claim 1, characterized in that: The specific steps for constructing the spatial model in the scene voiceprint modeling module are as follows: The scenario is the area where the power equipment is located when it is working. The scenario is scanned by LiDAR to obtain point cloud data, and a three-dimensional model of the scenario is generated based on the point cloud data. In the three-dimensional model, the type of power equipment and the three-dimensional coordinates of the power equipment are obtained by image recognition algorithm. The power equipment is numbered according to its type. Different power equipment in the scenario are associated and labeled according to the number and the three-dimensional coordinates of the power equipment to form equipment identity data.

3. The intelligent inspection system for power equipment with multi-scenario adaptation according to claim 2, characterized in that: The specific steps for constructing the initial voiceprint model in the scene voiceprint modeling module are as follows: The soundprint collection location for each power device in the preset scene is determined, and all soundprint collection locations are the same as the relative positions of their corresponding power devices. Sound level meters are placed in the scene to collect the audio signals and sound pressure levels of each power device during operation according to the preset sampling frequency and sampling duration, and the audio timestamps of the audio signals are obtained. The audio signal is divided into several frames according to a preset frame length and frame shift; a fast Fourier transform is performed on each frame to convert each frame of the audio signal from the time domain to the frequency domain, generating a spectrum diagram; spectral features are extracted from the spectrum diagram, including the fundamental frequency, harmonic amplitude, frequency centroid, and spectral width. The spectral features are concatenated sequentially to obtain the voiceprint feature vector. The voiceprint feature vector is then clustered according to the type of power equipment to obtain a cluster of voiceprint feature vectors for each type of power equipment. The mean and variance of each spectral feature in the cluster are calculated, and a feature fluctuation range is preset based on the mean and variance. Spectral features outside the feature fluctuation range are removed as abnormal features. The mean of each spectral feature after removing abnormal features is recalculated as the baseline data for each spectral feature. An initial voiceprint model is constructed based on the baseline data and the feature fluctuation range.

4. The intelligent inspection system for power equipment with multi-scenario adaptation according to claim 3, characterized in that: The specific steps for correcting the initial voiceprint model in the adaptive learning module are as follows: Several temperature sensors are deployed in the scene to obtain the real-time average temperature of the scene at a preset frequency, and the temperature timestamp of the real-time average temperature is obtained. A preset standard temperature is used to calculate the difference ΔT between the real-time average temperature and the standard temperature. The temperature timestamp of the difference ΔT is correlated with the audio timestamp to obtain the spectral characteristics corresponding to the difference. The rate of change of the spectral characteristics with temperature is calculated based on the difference, which is used as the temperature influence coefficient α. A correction value E is calculated based on the difference and the temperature influence coefficient. i =ΔT×α i E i The correction value α represents the i-th spectral feature. i The temperature influence coefficient representing the i-th spectral characteristic; The baseline data in the initial voiceprint model is added to the corresponding correction value to obtain the correction data, and the corresponding feature fluctuation range is adjusted according to the ratio of the change between the correction data and the baseline data to obtain the correction fluctuation range; a voiceprint baseline model is constructed based on the correction data and the correction fluctuation range; the voiceprint baseline model is associated with the device identity data; The specific steps for updating the voiceprint baseline model are as follows: A preset update cycle is set. At the end of each update cycle, the mean value of each spectral feature after removing abnormal features in that update cycle is obtained according to the audio timestamp, and used as new baseline data. The voiceprint baseline model is then updated based on the new baseline data.

5. The intelligent inspection system for power equipment with multi-scenario adaptation according to claim 4, characterized in that: The specific steps for obtaining the inspection coordinates of the inspection equipment in the inspection positioning module are as follows: When the inspection equipment enters the scene, the coordinates of the inspection equipment in the three-dimensional model are acquired in real time. The acceleration and angular velocity of the inspection equipment are collected by the IMU mounted on the inspection equipment, and the attitude data of the inspection equipment is calculated. The attitude data includes yaw angle, pitch angle and roll angle. The inspection coordinates include the coordinates and attitude data of the inspection equipment in the three-dimensional model.

6. The intelligent inspection system for power equipment with multi-scenario adaptation according to claim 5, characterized in that: The specific steps for obtaining the predicted synthetic voiceprint of the current inspection coordinates in the inspection and positioning module are as follows: Calculate the straight-line distance between the inspection coordinates and the three-dimensional coordinates of the power equipment; establish a horizontal coordinate system with the inspection coordinates as the origin, and calculate the azimuth angle between the inspection coordinates and the three-dimensional coordinates of the power equipment in the horizontal coordinate system. Establish a location information dataset, which includes equipment identification data of power equipment and the straight-line distance and azimuth angle corresponding to the equipment identification data; Calculate the theoretical voiceprint data at the inspection coordinates for each type of correction data in the voiceprint baseline model based on the straight-line distance: ； Where A represents theoretical voiceprint data, A0 represents the correction data corresponding to the theoretical voiceprint data, lg represents logarithmic operation, r0 represents the distance between the voiceprint acquisition location and the power equipment, and r represents the straight-line distance in the location information dataset of the power equipment. A three-dimensional coordinate system centered on the inspection equipment is established. The sound propagation direction of each power device is obtained according to the azimuth angle. The theoretical voiceprint data of each power device is vector-superimposed according to the type based on the sound propagation direction to obtain the predicted synthetic voiceprint data. The average value and variance of the modulus of the predicted synthetic voiceprint data of the inspection coordinate in several update cycles are calculated. The fluctuation range of the synthetic voiceprint data is preset according to the average value and variance of the modulus.

7. The intelligent inspection system for power equipment with multi-scenario adaptation according to claim 6, characterized in that: The specific steps for collecting real-time synthesized voiceprint data in the anomaly detection module are as follows: A sound level meter is installed on the inspection equipment, and the soundprint feature vector at the inspection coordinate is obtained in the same way as in the scene soundprint modeling module, which is used as real-time synthesized soundprint data. The specific steps for comparing and identifying anomalies based on the comparison results are as follows: obtain the fluctuation range of the synthetic voiceprint data of the inspection coordinate, calculate the difference between the modulus of the predicted synthetic voiceprint data and the real-time synthetic voiceprint data, and compare whether the difference is within the fluctuation range of the synthetic voiceprint data; if there is a difference outside the fluctuation range of the synthetic voiceprint data, it is identified as an anomaly, otherwise it is identified as normal; when it is identified as normal, the inspection equipment continues to move.

8. The intelligent inspection system for power equipment with multi-scenario adaptation according to claim 7, characterized in that: The specific steps for locating anomalies in the anomaly identification module are as follows: Set a movement point for each power device in the scene; use the spectral features corresponding to the differences identified as anomalies as anomaly features; The power equipment is sorted in ascending order of the straight-line distance between the current inspection coordinates and each power equipment, and the inspection equipment is controlled to move sequentially to the movement point of the power equipment in ascending order of the straight-line distance. The difference between the modulus of the predicted synthetic voiceprint data and the real-time synthetic voiceprint data corresponding to the abnormal features at the moving point is calculated as the positioning difference. A moving cycle is set. When the number of moving points traversed by the inspection equipment reaches the moving cycle, the moving point with the largest positioning difference within the moving cycle is compared. The power equipment corresponding to this moving point is identified as an abnormal device. The device identity data of the abnormal device is obtained, and the three-dimensional coordinates and number of the power equipment in the device identity data are extracted as the abnormal coordinates and abnormal number.

Images