Photovoltaic power station inspection method and system based on unmanned aerial vehicle intelligent scheduling

Abstract

The invention discloses a photovoltaic power station inspection method and system based on unmanned aerial vehicle intelligent scheduling, and belongs to the technical field of photovoltaic power station inspection. Signals received by a radar on an unmanned aerial vehicle are preprocessed and digitally processed, a preset threshold value is set, and target existence or signal shadow is judged based on the preset threshold value; transmitting power is dynamically adjusted to enhance detection and judgment of a shielding area, multi-sensor data are fused to complement target information of a signal shadow area, the complemented target information is coded according to a radar point cloud format and is combined with radar original point cloud to be displayed, a display interface is developed, radar point cloud data are arranged on the bottom layer, and a visual recognition frame is overlaid on the upper layer to achieve visual display.

Description

Technical Field

[0001] The present invention relates to a photovoltaic power station inspection, and in particular to a photovoltaic power station inspection method and system based on unmanned aerial vehicle (UAV) intelligent scheduling, belonging to the technical field of photovoltaic power station inspection. Background Art

[0002] In the existing technology, when drones are used for inspection, millimeter-wave radar (30GHz-300GHz) has a short wavelength. Although its diffraction ability is weak, it has a certain degree of penetration into obstructions such as vegetation, walls, and module gaps. It can detect areas that traditional visible light / infrared cannot reach (such as rusted brackets covered with shrubs and hidden cracks on the back of modules). Therefore, it has a relatively large disadvantage in detecting vegetation obstructions in mountain power stations, shadow areas of distributed rooftop air-conditioning outdoor units, and edge defects of densely arranged modules. Therefore, a photovoltaic power station inspection method and system based on drone intelligent scheduling are designed to solve the above problems. Summary of the Invention

[0003] The main purpose of the present invention is to provide a photovoltaic power station inspection method and system based on unmanned aerial vehicle intelligent scheduling.

[0004] The purpose of the present invention can be achieved by adopting the following technical solutions: A photovoltaic power station inspection method based on drone intelligent scheduling includes the following steps: Step 1: Preprocess and digitize the signal received by the radar; Step 2: Set a preset threshold and determine the target presence or signal shadow based on the preset threshold; Step 3: Dynamically adjust the transmit power to enhance detection and judgment of blocked areas; Step 4: Fuse multi-sensor data to complete target information in the signal shadow area; Step 5: Encode the completed target information in the radar point cloud format and merge it with the original radar point cloud for display; Step 6: Develop a display interface with radar point cloud data at the bottom layer and a visual recognition frame superimposed on the top layer for intuitive display.

[0005] Preferably, in step 2, the dynamic threshold is calculated by a constant false alarm rate algorithm ; in, is the reference unit signal intensity; N is the number of reference units; CF is the adjustment factor; In step 2, the digitized signal strength S is compared with the dynamic threshold In contrast, if , mark the target existence and start the tracking algorithm, otherwise it is judged as a signal shadow area.

[0006] Preferably, when a signal shadow is detected When the transmission power is increased, a digitally controlled attenuator or a variable gain power amplifier is used to the preset high power level until the signal strength exceeds the threshold or reaches the hardware upper limit.

[0007] Preferably, in step 4, the signal shadow area is combined with the target detection result of the visual sensor to obtain the visual coordinates through spatial registration. Convert to physical coordinates in the radar coordinate system: , and generate a virtual point cloud to supplement the radar data.

[0008] Preferably, in step 4, the fusion of multi-sensor completion specifically includes the following steps: S11: Use GPS clock or timestamp interpolation to ensure time consistency between radar and visual data; S12: concatenate visual features and radar features into a multi-dimensional feature vector; S13: Optimize fused features via linear transformation or attention mechanism.

[0009] Preferably, the dynamic adjustment of the transmission power is based on the radar equation ; When shadowing causes path loss to increase When, through Calculate the required transmit power to maintain received signal power Constant.

[0010] Preferably, the determination of the signal shadow in step 2 is performed by the following steps: S21: After the radar antenna receives the electromagnetic wave signal reflected by the target, it uses a low-noise amplifier to amplify the weak signal to increase the signal strength; S22: Filter out-of-band noise and interference through a bandpass filter to improve the signal-to-noise ratio and avoid misjudgment caused by noise; S23: using an analog-to-digital converter to convert the pre-processed analog signal into a digital signal for subsequent digital signal processing; S24: The ADC samples the signal at a certain sampling frequency and quantizes the signal amplitude into digital code; S25: Preset a fixed value based on the noise level of typical radar application scenarios; S26: Determine a fixed threshold by measuring noise power ; S27: Calculate the average noise power of N reference units around the detection unit ; in, is the reference unit signal strength, multiplied by the adjustment factor CF to obtain the dynamic threshold , used in uniform clutter environments.

[0011] Preferably, step 4 also includes performing target detection on N consecutive frames of visual images and recording the existence status of the target in each frame. ; in, Indicates that the target is detected in the i-th frame, on the contrary; Calculate the total number of votes for the target in N frames ; like ; in, If is the voting threshold, the decision is that the target exists, so the visual decision is used as a supplement.

[0012] A photovoltaic power station inspection system based on intelligent dispatching of drones includes a signal processing module containing a low-noise amplifier, a bandpass filter, and an analog-to-digital converter for signal preprocessing and digitization; Threshold calculation module, configured to calculate static thresholds or dynamic threshold , and realize real-time comparison between signal strength and threshold; Transmit power adjustment module, integrated with digitally controlled attenuator or variable gain power amplifier, dynamically adjusts the transmit power according to the signal shadow detection results ; Multi-sensor fusion module, including visual sensors, coordinate calibration unit and feature fusion algorithm unit, for spatiotemporal registration and feature-level fusion of visual and radar data; The display and completion module combines the completed virtual point cloud with the original radar point cloud and overlays the visual recognition box through a dedicated interface.

[0013] Preferably, the feature fusion algorithm unit supports principal component analysis dimensionality reduction, compresses the fused feature dimension to M dimensions, and inputs it into a generative adversarial network or a variational autoencoder to generate a radar target representation of the signal shadow area; The threshold calculation module supports mean-value CFAR and ordered statistics CFAR, the latter of which estimates noise power by sorting reference cell data and taking the median.

[0014] Beneficial technical effects of the present invention: The present invention provides a photovoltaic power station inspection method and system based on intelligent scheduling of drones. The millimeter-wave radar has a certain degree of penetration into obstructions such as vegetation, walls, and component gaps. Combined with dynamic transmission power adjustment technology, it can automatically enhance signal strength according to the degree of obstruction, effectively covering traditional blind spots such as the shrub-covered brackets of mountain power stations and the shadow areas of distributed rooftop air-conditioning outdoor units. For signal shadow areas that cannot be directly detected by radar, the target contour features are extracted by visual sensors, and a virtual point cloud is generated through spatial alignment and merged with the original radar data to achieve "visualization" of defects.

[0015] Through pre-processing methods such as low-noise amplification and bandpass filtering, electromagnetic interference and background noise are effectively suppressed, and signal quality is significantly improved. Even in strong interference scenarios, a high signal-to-noise ratio can be maintained, ensuring reliable detection of weak target signals.

[0016] When a signal shadow caused by an obstruction is detected, the system automatically increases the transmission power to enhance signal penetration capability, increases the signal sampling frequency in high-power mode, and further improves the detection capability of weakly reflective targets behind the obstruction through multi-frame data accumulation and analysis. Even if the target reflection signal is weak, reliable identification can be achieved through time series data enhancement.

[0017] High-precision clock synchronization and coordinate calibration technology ensure that radar and visual data are accurately aligned in time and space. A deep learning model is used to generate a virtual point cloud of the occluded area, which is merged and displayed with the original radar point cloud in real time, and a visual recognition box is superimposed. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 The present invention is a flowchart of a preferred embodiment of a photovoltaic power station inspection method and system based on intelligent scheduling of drones. DETAILED DESCRIPTION

[0019] In order to make the technical solution of the present invention more clear and specific to those skilled in the art, the present invention is further described in detail below with reference to embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0020] Cameras and millimeter-wave radars are installed on drones. Millimeter-wave wavelengths are often short (30GHz-300GHz corresponds to a wavelength of 1mm-10mm) and have weak diffraction capabilities, making it difficult to penetrate or bypass obstacles (such as building corners, walls, and dense vegetation) in drone inspection systems.

[0021] When the target is behind an obstruction, a "signal shadow" will be formed in the obstruction area, causing the radar to be unable to receive the echo reflected by the target, resulting in target loss. During perimeter drone inspections, pedestrians hiding behind walls or trees may not be detected by the radar.

[0022] Therefore, the problem of "signal shadow" is constituted. The design of the present invention is that the radar makes a judgment by comparing the received signal strength with a preset threshold.

[0023] When the echo signal strength is significantly lower than normal and persists for a period of time, it may indicate signal attenuation or loss caused by obstruction. The specific solution is as follows: After the radar antenna receives the electromagnetic wave signal reflected by the target, it uses a low-noise amplifier to amplify the weak signal to increase the signal strength.

[0024] Bandpass filters are used to filter out noise and interference outside the frequency band (such as electromagnetic noise and clutter), thereby improving the signal-to-noise ratio (SNR) and avoiding misjudgment caused by noise.

[0025] An analog-to-digital converter is used to convert the preprocessed analog signal into a digital signal to facilitate subsequent digital signal processing.

[0026] The ADC samples the signal at a certain sampling frequency (such as millions of times per second) and quantizes the signal amplitude into digital code.

[0027] Fixed values are pre-set based on the noise level of typical radar application scenarios.

[0028] In a laboratory environment, a fixed threshold is determined by measuring the noise power .

[0029] Calculate the average noise power of N reference units around the detection unit ; is the reference unit signal strength, multiplied by the adjustment factor CF to obtain the dynamic threshold , used in uniform clutter environments.

[0030] The reference unit data are sorted and the median is selected as the noise estimate, which is suitable for non-uniform clutter environments and avoids threshold deviation caused by strong clutter interference.

[0031] Only the comparator circuit is used to compare the digitized signal amplitude with the threshold in real time. If the signal strength S exceeds the dynamic threshold , the comparator outputs a high level or triggers an interrupt, directly indicating the target detection result, which is suitable for scenarios with high real-time requirements (such as missile guidance radar).

[0032] In a digital signal processing unit (such as FPGA, DSP), point-by-point comparison is performed through an algorithm.

[0033] Write Verilog code in FPGA, or use C language to implement loop judgment in DSP. , then the target is marked as existing, which facilitates flexible adjustment of the processing logic.

[0034] like , start the target tracking algorithm (such as Kalman filter) and continuously monitor the target motion state; like , continue to receive new echo signals and repeat the above process.

[0035] Mean class CFAR threshold calculation: ; Where N is the number of reference units, is the signal strength of the i-th reference unit, and CF is the adjustment factor (set according to scene experience).

[0036] Signal judgment logic: target exists ; Where S is the signal strength of the current detection unit, is a preset threshold (static or dynamic CFAR threshold); Through the above technical solutions and formulas, the radar can accurately compare the signal strength and threshold in different environments to determine whether the target exists.

[0037] Furthermore, the present invention integrates a digitally controlled attenuator or a variable gain power amplifier in the radar transmitter module, and adjusts the transmission power in real time through software instructions. .

[0038] Low power mode (normal scenario): , reduce power consumption; High power mode (occlusion scenario): , enhancing signal penetration capability.

[0039] When the radar determines that there is a signal shadow in a certain area through signal strength comparison When , the following process is triggered: Send the command to the transmitter module. Boost to the preset high power gear; Continuously monitor the signal strength in the area. Then keep the current power, otherwise further increase Until a threshold is reached or the hardware limit is reached.

[0040] Combined with the "multi-frame signal accumulation detection" logic, the number of signal sampling frames is increased in high-power mode (for example, from 10 frames / second to 20 frames / second), and the signal-to-noise ratio is improved through time dimension accumulation.

[0041] According to the radar equation, the received signal power and transmit power The relationship is: ; is the transmit / receive antenna gain; is the target radar cross section; R is the target distance; L is the path loss caused by obstruction.

[0042] Assume that the original transmission power is , the path loss increases due to the shadowing (such as vegetation blocking ), in order to make the received signal power remains unchanged, the transmit power needs to be increased to ; According to the dynamic threshold formula , when the transmit power increases, the average signal strength of the reference unit increases synchronously, so: If the detection sensitivity is kept constant (false alarm rate is constant), CF needs to be kept constant. Follow Linear growth; If you want to increase the detection probability, you can appropriately reduce CF in high power mode (for example, from the default value of 3 to 2.5) to make the threshold The increase is smaller than the signal strength increase, thus improving probability.

[0043] Table 1 Key parameters and implementation effects

[0044] Combined with "multi-sensor fusion" technology, if the radar still cannot detect the target in high-power mode (such as extreme obstruction), the visual sensor completion is triggered (such as the camera recognizes the target behind the wall and fuses it with the radar data to display a virtual point cloud).

[0045] The specific combination of "multi-sensor fusion" technology includes the following solutions: Use hardware clocks (such as GPS synchronized clocks) or software algorithms (such as timestamp-based interpolation) to ensure that the data collection times of millimeter-wave radar and visual sensors (cameras, thermal imagers, etc.) are consistent, avoiding target position deviations due to time differences.

[0046] Through calibration plates or known scene features, the coordinate transformation relationship between the visual sensor and the millimeter-wave radar (such as rotation matrix and translation vector) is established, and the visual target coordinates are mapped to the radar coordinate system to achieve spatial unification.

[0047] A deep learning model is used to detect objects in camera images, extracting features such as the object's outline, color, and texture. For example, the visual outline of a pedestrian behind a wall can be detected.

[0048] Preprocess the millimeter-wave radar data (such as denoising and filtering) to extract information such as distance and speed.

[0049] For the signal shadow area, mark the spatial location of the signal missing area.

[0050] Combine visual features (such as target outline vectors) and radar features (such as distance information) into a multidimensional feature vector. Combine the pedestrian's visual outline coordinates with the surrounding environment distance information measured by the radar to form complementary features. The specific technical solution is as follows: Use a GPS clock or dedicated synchronization module (such as IEEE1588) to ensure that the sampling time error between the radar and camera is less than 1ms, avoiding feature mismatch caused by time misalignment.

[0051] For asynchronously collected data, alignment is performed based on timestamps using linear interpolation or nearest neighbor interpolation. The formula is: ; in, , are the timestamps of adjacent frames, is the interpolated eigenvalue.

[0052] Use Zhang's calibration method or a 3D calibration plate to solve the extrinsic parameters of the camera and radar (rotation matrix R, translation vector t).

[0053] The pixel coordinates of the visual target Convert the camera's intrinsic parameter K into the three-dimensional coordinates in the radar coordinate system; ; in, is the inverse camera intrinsic parameter matrix, which is used to convert pixel coordinates into normalized camera coordinates.

[0054] Use deep learning models (such as Mask R-CNN) to segment the target and extract the coordinates of contour key points (Unit: pixel) and converted into physical coordinates in the radar coordinate system through spatial registration .

[0055] Extract the unique hot encoding of the target category (such as "pedestrian" and "vehicle") through CNN (such as ResNet) (C is the number of categories).

[0056] Input vector: (N is the number of contour key points); Extract the target's distance d and azimuth from the radar point cloud or RD map , pitch angle .

[0057] For obstacles within the radar scanning range, extract their distance distribution characteristics (such as the distance to the nearest obstacle , average distance ).

[0058] Output vector: ; The visual features and radar features are concatenated by dimension to form a complementary feature vector: ; Eliminate scale differences between features through linear transformation or nonlinear mapping. The formula is: ; , : learnable weight matrix; is the activation function (such as ReLU); Suitable for scenarios where neural networks are subsequently connected (such as fusion features input to Transformer for target detection); Introducing attention weights to adaptively adjust feature importance: ; ; is the attention weight vector, learned through training; It is an element-wise product and is used to enhance key modal features (such as vision dominance in occluded scenes and radar dominance in non-occluded scenes).

[0059] Normalize features of different dimensions (such as pixel coordinates and distances): ; , is the mean and standard deviation of the feature, obtained through training data statistics.

[0060] If the feature dimension after fusion is too high (e.g. > 100 dimensions), it is compressed to M dimensions through PCA: ; U is the projection matrix composed of the first M principal components of the training data feature covariance matrix.

[0061] Example 1: Pedestrian outline coordinates behind a wall (estimated through camera perspective transformation); The signal strength in the wall area is below the threshold , marked as signal shadow.

[0062] The visual profile coordinates are converted into the physical position (X, Y) in the radar coordinate system through spatial registration; The radar provides the environmental distance information of the location (such as the distance to the wall ); The fusion vector is [ ,category code “pedestrian”]; The input classifier determines whether there is an occluded target at the location (such as outputting a probability value through SVM or neural network).

[0063] After completing feature-level fusion using the aforementioned solution, a generative adversarial network (GAN) or variational autoencoder (VAE) is trained. This input combines visual features with radar data from non-shadowed areas to generate radar target representations for signal shadowed areas. For example, the model learns the correlation between the features of a pedestrian behind a wall in both the visual and radar-free areas, generating corresponding radar point cloud features.

[0064] The visual sensor independently identifies the target (such as determining that there is a pedestrian behind a wall), and the radar determines that the signal is missing in this area. Through a rule engine or voting mechanism, the visual decision is supplemented and the corresponding target information (such as a virtual point cloud) is directly added to the radar data.

[0065] Perform target detection on N consecutive frames of visual images and record the existence status of the target in each frame ( Indicates that the target is detected in the i-th frame, on the contrary).

[0066] Calculate the total number of votes for the target in N frames .

[0067] like ( is the voting threshold), the decision is that the target exists, so the visual decision is used as a supplement.

[0068] The completed target information (such as a virtual point cloud) is encoded in the radar point cloud format and displayed together with the original radar point cloud. For example, in the radar monitoring interface, pedestrians behind a wall are displayed in a point cloud with a specific color, which is displayed in unison with the point cloud directly detected by the radar.

[0069] A dedicated display interface has been developed, with radar point cloud data as the bottom layer and visual recognition boxes (such as BoundingBox) superimposed on top. Targets are annotated with visual recognition boxes in signal shadow areas, while virtual point clouds are dynamically generated in areas where radar point clouds are missing, enabling intuitive fusion display.

[0070] This is a further embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes based on the technical solutions and concepts of the present invention within the scope disclosed by the present invention, which fall within the scope of protection of the present invention.

Claims

1. A photovoltaic power station inspection method based on intelligent dispatching of drones, characterized by: The steps include: Step 1: Preprocess and digitize the signal received by the radar; Step 2: Set a preset threshold and determine the target presence or signal shadow based on the preset threshold; Step 3: Dynamically adjust the transmit power to enhance detection and judgment of blocked areas; Step 4: Fuse multi-sensor data to complete target information in the signal shadow area; Step 5: Encode the completed target information in the radar point cloud format and merge it with the original radar point cloud for display; Step 6: Develop a display interface with radar point cloud data at the bottom layer and a visual recognition frame superimposed on the top layer for intuitive display.

2. The photovoltaic power station inspection method based on intelligent dispatching of drones according to claim 1 is characterized by: In step 2, the dynamic threshold is calculated by the constant false alarm rate algorithm ; in, is the reference unit signal intensity; N is the number of reference units; CF is the adjustment factor; In step 2, the digitized signal strength S is compared with the dynamic threshold In contrast, if , mark the target existence and start the tracking algorithm, otherwise it is judged as a signal shadow area.

3. The photovoltaic power station inspection method based on intelligent dispatching of drones according to claim 2 is characterized by: When a signal shadow is detected When the transmission power is increased, a digitally controlled attenuator or a variable gain power amplifier is used to the preset high power level until the signal strength exceeds the threshold or reaches the hardware upper limit.

4. The photovoltaic power station inspection method based on intelligent dispatching of drones according to claim 1 is characterized by: In the signal shadow area of step 4, the visual coordinates are aligned with the target detection results of the visual sensor through spatial registration. Convert to physical coordinates in the radar coordinate system: , and generate a virtual point cloud to supplement the radar data; Where R is the rotation matrix; t is the translation vector; K is the camera internal parameter.

5. The photovoltaic power station inspection method based on drone intelligent scheduling according to claim 1 is characterized by: In step 4, the fusion of multi-sensor completion specifically includes the following steps: S11: Use GPS clock or timestamp interpolation to ensure time consistency between radar and visual data; S12: concatenate visual features and radar features into a multi-dimensional feature vector; S13: Optimize fused features via linear transformation or attention mechanism.

6. The photovoltaic power station inspection method based on drone intelligent scheduling according to claim 1 is characterized by: The dynamic adjustment of the transmission power is based on the radar equation ; When shadowing causes path loss to increase When, through Calculate the required transmit power to maintain received signal power constant; is the transmit / receive antenna gain; is the target radar cross section; R is the target distance; L is the path loss caused by obstruction; is the received signal power; is the transmit power.

7. The photovoltaic power station inspection method based on UAV intelligent scheduling according to claim 1 is characterized by: The signal shadow in step 2 is determined by the following steps: S21: After the radar antenna receives the electromagnetic wave signal reflected by the target, it uses a low-noise amplifier to amplify the weak signal to increase the signal strength; S22: Filter out-of-band noise and interference through a bandpass filter to improve the signal-to-noise ratio and avoid misjudgment caused by noise; S23: using an analog-to-digital converter to convert the pre-processed analog signal into a digital signal for subsequent digital signal processing; S24: The ADC samples the signal at a certain sampling frequency and quantizes the signal amplitude into digital code; S25: Preset a fixed value based on the noise level of typical radar application scenarios; S26: Determine a fixed threshold by measuring noise power ; S27: Calculate the average noise power of N reference units around the detection unit ; in, is the reference unit signal strength, multiplied by the adjustment factor CF to obtain the dynamic threshold , used in uniform clutter environments.

8. The photovoltaic power station inspection method based on intelligent dispatching of drones according to claim 1 is characterized by: Step 4 also includes target detection for N consecutive frames of visual images, recording the existence status of the target in each frame. ; in, Indicates that the target is detected in the i-th frame, on the contrary; Calculate the total number of votes for the target in N frames ; like ; in, If is the voting threshold, the decision is that the target exists, so the visual decision is used as a supplement.

9. A photovoltaic power station inspection system based on drone intelligent scheduling, based on the photovoltaic power station inspection method based on drone intelligent scheduling according to any one of claims 1 to 8, characterized in that: include: Signal processing module, including low-noise amplifier, bandpass filter and analog-to-digital converter for signal preprocessing and digitization; Threshold calculation module, configured to calculate static thresholds or dynamic threshold , and realize real-time comparison between signal strength and threshold; Transmit power adjustment module, integrated with digitally controlled attenuator or variable gain power amplifier, dynamically adjusts the transmit power according to the signal shadow detection results ; Multi-sensor fusion module, including visual sensors, coordinate calibration unit and feature fusion algorithm unit, for spatiotemporal registration and feature-level fusion of visual and radar data; The display and completion module combines the completed virtual point cloud with the original radar point cloud and overlays the visual recognition box through a dedicated interface.

10. The photovoltaic power station inspection system based on intelligent dispatching of drones according to claim 9 is characterized by: The feature fusion algorithm unit supports principal component analysis dimensionality reduction, compresses the fused feature dimension to M dimensions, and inputs it into a generative adversarial network or a variational autoencoder to generate a radar target representation of the signal shadow area; The threshold calculation module supports mean-value CFAR and ordered statistics CFAR, the latter of which estimates noise power by sorting reference cell data and taking the median.

Images