A method, system for visual enhancement under transformer oil
By acquiring the optical state feature vector of transformer oil and dynamically adjusting the light source and camera parameters, the problem of image quality degradation in transformer oil was solved, and high-quality imaging and fault identification were achieved under different turbidity environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA UNIV OF MINING & TECH
- Filing Date
- 2026-06-17
- Publication Date
- 2026-07-14
AI Technical Summary
Existing robot vision systems suffer from image quality degradation in transformer oil, including decreased contrast, blurred details, color distortion, and shortened observation distance. They also cannot dynamically adjust according to real-time changes in the optical state of the oil, resulting in limited fault identification capabilities.
By acquiring the extinction coefficient, volume scattering function, and turbidity scalar parameters of transformer oil, an environmental optical state feature vector is generated. The light source power, spectral ratio, camera polarization angle, and time gate width are dynamically adjusted. Combined with integrated acquisition of multi-dimensional parameters and temperature drift compensation, differentiated light field control is achieved to adapt to environments with different turbidity levels.
Achieving high-quality imaging under different oil degradation levels highlights target texture and gap features, improves target three-dimensional recognition and fault detection rate, and ensures the stability of imaging quality during the inspection robot's movement.
Smart Images
Figure CN122385489A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of detection technology, specifically relating to a method and system for visual enhancement under transformer oil. Background Technology
[0002] As a core component of the power system, the transformer's internal insulation and the operating condition of its mechanical parts directly affect the stability and security of the power grid. Transformer oil not only provides insulation and cooling but also carries information about its internal condition. The physicochemical properties and cleanliness of the oil directly reflect the transformer's operating conditions.
[0003] In recent years, with the development of robotics technology, infiltrative inspection robots equipped with vision sensors have been gradually applied to online inspection inside transformers. However, during long-term operation, transformer oil becomes increasingly turbid due to factors such as cellulose particle shedding, metal wear particles, moisture intrusion, and the oil's own oxidation and aging. This turbid oil produces strong absorption and Mie scattering effects on light signals, causing severe problems such as decreased contrast, blurred details, color distortion, and a sharp reduction in observation distance in images captured by traditional vision systems.
[0004] Most existing robot vision systems employ lighting schemes with fixed parameters (such as constant power, fixed spectrum, and uniform illumination), which cannot be dynamically adjusted according to real-time changes in the optical state of the oil. In low-turbidity oil, scattering noise is not effectively suppressed, resulting in insufficient differentiation between target edges and background. In medium-to-high turbidity or aged oil, short-wavelength light is severely absorbed, leading to dim images, insufficient penetration, and difficulty in observing deep components. Under extreme turbidity conditions, it may even be impossible to form an effective target image, severely limiting the inspection robot's ability to identify early faults (such as winding deformation, local overheating, particle deposition, and loose interfaces).
[0005] Patent CN202311192748.1 discloses an image enhancement method for transformer oil based on a gridded gray-world white balance algorithm. It is mainly used to solve the problem that the white balance effect of the traditional gray-world white balance algorithm is not good when the image contains a large area of single color block or the scene color is not rich. It cannot address the physical root causes of the degradation of the image quality under oil (scattering, absorption, uneven illumination) and the algorithm parameters are fixed, so it cannot be adaptively adjusted according to the dynamic environmental factors such as oil turbidity and target distance.
[0006] Patent CN202311193005.6 discloses a method for enhancing transformer oil images based on adaptive channel compensation and standard deviation weighted MSRCR fusion. It mainly targets the acquired oil images with overall degraded quality. It does not require the imaging system (light source, camera) itself and cannot fundamentally improve the physical quality degradation of the image caused by oil scattering, absorption and uneven illumination. Moreover, the algorithm parameters and fusion weights are mostly preset and cannot be adaptively adjusted online according to dynamic environmental factors such as oil turbidity, target distance and changes in on-site lighting. Summary of the Invention
[0007] The purpose of this invention is to provide a method and system for enhancing visual perception under transformer oil.
[0008] To achieve the above objectives, the technical solution of the present invention is as follows: A method for enhancing visual perception under transformer oil includes the following steps: S1: Obtain the extinction coefficient, volume scattering function, and turbidity scalar parameters of the transformer oil in the transformer to be tested, and generate an environmental optical state feature vector after temperature drift compensation.
[0009] S2: Based on the environmental optical state feature vector, target distance, and robot moving speed, the light source power, spectral ratio, polarization angle, and time gate width of the light source on the robot are matched.
[0010] S3: When the turbidity scalar parameter of the transformer oil is less than the first turbidity threshold, adjust the angle between the light source and the camera polarizer to the polarization angle matched in step S2, acquire orthogonal polarization images and parallel polarization images, and suppress the depolarization scattering component based on the difference coefficient proportional to the extinction coefficient.
[0011] When the turbidity scalar parameter of the transformer oil is greater than or equal to the first turbidity threshold and less than the second turbidity threshold, the spectral ratio of the light source mounted on the robot is adjusted so that the spectral centroid wavelength of the light emitted by the light source shifts towards the longer wavelength direction.
[0012] When the turbidity scalar parameter of the transformer oil is greater than or equal to the second turbidity threshold, a nanosecond-level pulsed laser is activated as the light source. Based on the propagation speed of light in the transformer oil, the round-trip time of the light is calculated according to the target distance in step S2. Based on the time gating width, the camera shutter time window is controlled.
[0013] S4: Based on the geometry of the target, determine the brightness weight of the light source on the robot in different regions. Calculate the spatial noise reduction coefficient based on the average pixel intensity of different regions of the target and the reflected light intensity of the target. After adjusting the brightness weight to make the spatial noise reduction coefficient greater than a preset threshold, acquire and output the target image.
[0014] Furthermore, after S4, step S5 is also included: extracting three quality indicators of the target image in real time: signal-to-noise ratio, information entropy, and Tenengrad gradient value, establishing a negative feedback adjustment loop, and when any quality indicator fails to reach the preset threshold, sequentially triggering the light source power adjustment, multi-frame image fusion, and robot slow-down exposure compensation strategy, updating control parameters, re-acquiring images, and outputting the target image.
[0015] Furthermore, in S3, when the turbidity scalar parameter of the transformer oil is less than the first turbidity threshold, the adjustment method is achieved through the following formula: in, It is an orthogonal polarization image. This is a parallel polarization image. The difference coefficients are determined using the following formula: in, The extinction coefficient after temperature drift compensation. The maximum calibrated value for the extinction coefficient is [value to be filled in]. .
[0016] In S3, when the turbidity scalar parameter of the transformer oil is greater than or equal to the first turbidity threshold and less than the second turbidity threshold, the adjustment method is achieved through the following formula: in, The wavelength of the spectral barycenter of the synthesized light. For the spectral weights of red light, For the spectral weights of green light, For the spectral weights of blue light, The wavelength of red light It is the wavelength of green light. It is the wavelength of blue light.
[0017] In S3, when the turbidity scalar parameter of the transformer oil is greater than or equal to the second turbidity threshold, the adjustment method is achieved through the following formula: in, For the round trip time of light, Let D be the speed of light in transformer oil, and D be the target distance. The camera shutter speed is controlled at... Always open, at Always turn off Select the gate width for the time.
[0018] Environmental optical state feature vector in S1 The calculation is performed using the following formula: in, The extinction coefficient after temperature drift compensation. This is the volume scattering function after temperature drift compensation. The turbidity scalar parameter is... T It is the transpose of the vector.
[0019] The matching method in S2 is based on a pre-trained weight allocation function. The environmental optical state feature vector, target distance, and robot speed are used as input variables, while the light source power, spectral ratio, polarization angle, and time gate width are used as output variables. The weight allocation function is implemented using the following formula: in, For the first One output variable, For the first The input variable for the first... The weight coefficients of each output variable, For the first One input variable, The number of input variables. The range of values is [1- ], For the number of dimensions of the input variables, The range of values is natural numbers greater than 0. For the first The bias values of each output variable.
[0020] In S4, differential weight allocation for different target geometries is achieved through the following formula: The geometry is a cylindrical winding, and the weighting formula is: in, For the first The distance from each region to the winding axis, where σ is the standard deviation of the Gaussian function. For the first The lighting clusters are divided by the lighting LED array. Brightness weight of each region and The value of is a natural number greater than 0; The geometry is a planar box wall, and the mathematical model for weight allocation is: in, For the first The distance from each area to the edge of the box wall, The maximum half width of the box wall. For the first The lighting clusters are divided by the lighting LED array. Brightness weight of each region and The value of is a natural number greater than 0; The geometry is a sleeve interface, and the mathematical model for weight allocation is: in, For the first The distance from each area to the interface gap, The width of the interface gap. For the Dirac function, For the first The lighting clusters are divided by the lighting LED array. Brightness weight of each region and The value of is a natural number greater than 0.
[0021] S4 spatial noise reduction coefficient The calculation is performed using the following formula: in, The average pixel intensity of the target region; The intensity of the light reflected from the target.
[0022] The present invention also provides a transformer oil subsurface vision enhancement system, comprising: Oil environment optical state feature acquisition module: acquires the oil extinction coefficient, volume scattering function and turbidity scalar parameters of transformer oil in the transformer under test, and generates an environmental optical state feature vector after temperature drift compensation; Variable matching module: Based on the environmental optical state feature vector, target distance, and robot moving speed, it matches the light source power, spectral ratio, polarization angle, and time gate width of the light source mounted on the robot and the camera mounted on the robot. The graded control module: When the turbidity scalar parameter of the transformer oil is less than the first turbidity threshold, the angle between the light source and the camera polarizer is adjusted to the matched polarization angle, orthogonal polarization image and parallel polarization image are acquired, and the depolarization scattering component is suppressed based on the difference coefficient proportional to the extinction coefficient. When the turbidity scalar parameter of the transformer oil is greater than or equal to the first turbidity threshold and less than the second turbidity threshold, the spectral ratio of the light source mounted on the robot is adjusted so that the spectral centroid wavelength of the light emitted by the light source shifts towards the long wavelength direction. When the turbidity scalar parameter of transformer oil is greater than or equal to the second turbidity threshold, a nanosecond-level pulsed laser is activated as the light source. Based on the speed of light propagation in transformer oil, the round-trip time of light is calculated according to the target distance. Based on the time gating width, the camera shutter time window is controlled. Target recognition module: Based on the geometry of the target, the brightness weight of the light source on the robot in different regions is determined. The spatial noise reduction coefficient is calculated based on the average pixel intensity of different regions of the target and the reflected light intensity of the target. After adjusting the brightness weight to make the spatial noise reduction coefficient greater than a preset threshold, the target image is acquired and output.
[0023] Compared with the prior art, the technical solution provided by this invention has the following advantages: This invention achieves integrated acquisition of multi-dimensional parameters through a sensing module mounted on a robot, and corrects optical parameter errors with a temperature drift compensation mechanism, eliminating environmental interference such as oil temperature, providing high-precision input parameters for light field control, and ensuring accurate control decisions from the source.
[0024] The design incorporates three turbidity threshold ranges, corresponding to three differentiated light field control modes, enabling full-scene adaptation from low-turbidity to extremely turbid oil environments. This hierarchical control mechanism overcomes the limitations of traditional single illumination schemes, achieving high-quality imaging effects under different degrees of oil degradation.
[0025] For targets with different geometries, such as transformer windings, tank walls, and bushing interfaces, differentiated weight allocation is used to highlight the texture and gap features of the targets. At the same time, by calculating the spatial noise reduction coefficient and ensuring that the spatial noise reduction coefficient is greater than a preset threshold, the problems of target edge blurring and reflection interference caused by traditional uniform lighting are effectively solved, significantly improving the three-dimensional recognition of targets and the defect detection rate.
[0026] This invention also continuously optimizes the imaging quality and lighting parameters dynamically to ensure stable imaging quality throughout the movement of the inspection robot, providing reliable visual support for the accurate identification of faults in internal transformer components. Attached Figure Description
[0027] Figure 1 This is a schematic diagram of the operation of the transformer oil submersible vision enhancement system of the present invention; in the figure, 0101 is an independent ring illumination cluster, 0102 is a camera mounted on the robot, and 0103 is an infiltrated inspection robot. Detailed Implementation
[0028] To further understand the content of this invention, the invention will be described in detail with reference to the embodiments. Visual enhancement operation under transformer oil at a 500kV substation.
[0029] Operating environment parameters: Transformer type: 500kV oil-immersed power transformer; Oil condition: 8 years old; Oil is moderately turbid, oil temperature is 45℃; Detection targets: internal windings, bushing interfaces; Detection distance range: 0.5m-3.0m; Robot movement speed: 0.1m / s.
[0030] System hardware configuration: Sensing unit on the robot: Reference photoelectric pair module: 532nm wavelength at the transmitting end, 0.1m optical path at the receiving end; transmission light intensity detection accuracy ±0.1%; Lateral scattering detection module: photosensitive element sensitivity 0.01, arranged 90° laterally; Temperature compensation module: temperature sensor range -20℃ ~ 150℃, accuracy ±0.1℃.
[0031] Distance and speed measurement module: ultrasonic sensor range 0.1m - 1m, lidar range 1m - 5m, ranging accuracy ±1cm; speed measurement accuracy ±0.005m / s.
[0032] Graded adjustment unit: Polarization differential adjustment module: polarizer adjustment range 0°-90°, servo accuracy ±0.1°; Spectral centroid shift module: red (650nm), green (520nm), blue (470nm) three-color LED array, maximum power of monochromatic 5W; Ultra-high speed gating imaging module: pulsed laser wavelength 532nm, pulse width 5ns, shutter gating width 200ns.
[0033] Target recognition unit: The robot is equipped with an LED array that is divided into 4 independent lighting areas arranged in a ring. The brightness adjustment range is 0-100%, and the adjustment accuracy is ±1%.
[0034] Control and processing unit: Embedded processor: ARM Cortex-A53, 1.2GHz.
[0035] Image processing unit: Supports real-time YOLOv8 algorithm for target recognition, with an accuracy rate of ≥95%.
[0036] Example 1 A method for enhancing visual perception under transformer oil includes the following steps: S1: Obtain the extinction coefficient, volume scattering function, and turbidity scalar parameters of the transformer oil in the transformer to be tested, and generate an environmental optical state feature vector after temperature drift compensation.
[0037] The process of obtaining the extinction coefficient of transformer oil is as follows: The reference photoelectric pair module in the sensing unit of the robot is controlled, with the output wavelength of the transmitter adjustable within the range of 400nm to 900nm, and the initial light intensity is... A collimated light beam propagates along a fixed optical path of L = 0.1m in transformer oil, and the transmitted light intensity is detected in real time at the receiving end. The extinction coefficient of the oil was calculated based on the Beer-Lambert law. The formula for calculating the extinction coefficient can be obtained as follows: in, The wavelength of the light source, For real-time oil temperature, Optical path length The unit is This reflects the total absorption and scattering effect of light by the oil. To improve the accuracy of extinction coefficient measurement, the system selects 10 characteristic wavelength points within the 400nm-900nm wavelength range for segmented measurement, namely 400nm, 450nm, 500nm, 532nm, 550nm, 600nm, 650nm, 700nm, 800nm, and 900nm. Multi-wavelength data fusion reduces the measurement error of a single wavelength. The fusion formula is: in, Let be the weight value at the z-th wavelength. The weight value is positively correlated with the detection sensitivity of the photodiode at that wavelength; the higher the sensitivity, the larger the weight value. The value range is [0.05, 0.15], and satisfies the following condition: .
[0038] Extinction coefficient calculation: Initial light intensity at the emitter The receiver measured , .
[0039] The process of obtaining the volume scattering function of transformer oil is as follows: The intensity of lateral scattered light generated by suspended particles in the oil is captured by the high-sensitivity photosensitive element of the lateral scattering detection module. Volume scattering function Based on the initial light intensity at the transmitter, the solid angle of the detector, and the effective scattering volume, the scattered light intensity in the 90° direction is calculated using the following formula: in, The initial light intensity at the transmitting end, The intensity of scattered light detected at a 90° angle. These are the system calibration coefficients. , Effective scattering volume (unit: m) 3 ), Solid angle received by the detector (unit: ), Obtained in the laboratory by standard scattering plate calibration.
[0040] Volume scattering function calculation: Initial light intensity at the emitter Lateral probe acquisition : .
[0041] The temperature drift compensation process is as follows: Because the extinction coefficient and volume scattering function of transformer oil drift with temperature, a preset temperature drift compensation curve is retrieved. This curve was obtained through experimental calibration. Temperature drift compensation for the extinction coefficient and volume scattering function was achieved using the following formula: in, The extinction coefficient after temperature drift compensation. This is the volume scattering function after temperature drift compensation. The temperature drift compensation factor for the extinction coefficient. The temperature drift compensation factor for the volume scattering function. The current oil temperature, This is the standard reference temperature.
[0042] oil temperature Call the fitted curve, extinction coefficient, and temperature drift compensation factor. Volume scattering function temperature drift compensation factor Standard reference temperature The extinction coefficient and volume scattering function of the oil after temperature drift compensation are: Scalar parameters of turbidity in transformer oil The acquisition process is as follows: in, This is a weighting coefficient, calibrated according to the transformer oil type, with a value range of [0.4, 0.6]. For new oil... Take 0.4, Take 0.6 for aged oil that has been in operation for more than 10 years. Take 0.6, The weighting factor is set to 0.4; for oils with an operating life of 1-10 years, the weighting factor is calculated using linear interpolation, and the interpolation formula is: in, This refers to the service life of the transformer oil, expressed in years.
[0043] scalar parameters of turbidity in transformer oil The specific calculation is as follows: For oil ages of 8 years, falling between 1 and 10 years, the weighting coefficient is calculated using linear interpolation. =0.5, =0.5 , .
[0044] After temperature drift compensation and turbidity Generate environmental optical state feature vectors ,Right now .
[0045] S2: Based on the environmental optical state feature vector, target distance, and robot moving speed, the light source power, spectral ratio, polarization angle, and time gate width of the light source on the robot are matched.
[0046] The input vector is composed of the environmental optical state feature vector, the target distance, and the robot's moving speed as input variables. The output vector is composed of the light source power, spectral ratio, and the polarization angle and time gating width of the camera mounted on the robot as output variables. A multidimensional decision matrix M is constructed. The multidimensional decision matrix M is generated by training based on experimental data covering different oil qualities, different target distances, and different movement speeds to ensure its generalization ability and decision accuracy in diverse oilfield operating environments.
[0047] The input vector is calculated as follows: The target distance is D = 2.0m, and the robot's moving speed is v = 0.1m / s; the output variable dimension is m = 4, and the output vector... Each parameter corresponds to the light source power, polarization angle, spectral ratio, and time gate width, respectively. The weight coefficient matrix W (dimension 4×5) and bias vector b are obtained by training using the gradient descent method. The training objective is to minimize the mean square error between the predicted value and the actual optimal value.
[0048] The training convergence parameters used in this embodiment are as follows: The system receives input variables collected in real time. The optimal control vector is output from the multidimensional decision matrix using a lookup table algorithm or a weight allocation function; the formula for calculating the weight allocation function is: in, For the first One output variable, For the first The input variable for the first... The weight coefficients of each output variable, For the first One input variable, The number of input variables. The range of values is [1- ], For the number of dimensions of the input variables, The range of values is natural numbers greater than 0. For the first The bias values of each output variable.
[0049] The power of the light source P (f=1) is calculated as follows: The polarization angle θ (f=2) is calculated as follows: Spectral ratio (f=3) The specific calculation is as follows: The spectral composition is as follows: weighted proportions of red light (R), green light (G), and blue light (B). The baseline values are calculated first, followed by normalized allocation. Combining long-wave migration strategies for medium-turbid environments, the final determination was made. The time-selection gate width τ (f=4) is calculated as follows: After calculation and engineering normalization adjustment, the optimal control vector is output as follows: Y=[4.6 W, 120 ∘ , 0.7:0.2:0.1, 200 ns].
[0050] S3: When the turbidity scalar parameter of the transformer oil is less than the first turbidity threshold, adjust the angle between the light source and the camera polarizer to the polarization angle matched in step S2, acquire orthogonal polarization images and parallel polarization images, and suppress the depolarization scattering component based on the difference coefficient proportional to the extinction coefficient.
[0051] When the turbidity scalar parameter of the transformer oil is less than the first turbidity threshold, the adjustment method is achieved through the following formula: in, It is an orthogonal polarization image. This is a parallel polarization image. The difference coefficients are determined using the following formula: in, The extinction coefficient after temperature drift compensation. The maximum calibrated value for the extinction coefficient is [value to be filled in]. This mode highlights the specular reflection or specific polarization characteristics of the target object by suppressing the depolarization scattering component generated by suspended particles in the oil.
[0052] When the turbidity scalar parameter of the transformer oil is greater than or equal to the first turbidity threshold and less than the second turbidity threshold, the spectral ratio of the light source mounted on the robot is adjusted so that the spectral centroid wavelength of the light emitted by the light source shifts towards the longer wavelength direction.
[0053] When the turbidity scalar parameter of the transformer oil is greater than or equal to the first turbidity threshold and less than the second turbidity threshold, the adjustment method is achieved through the following formula: in, The wavelength of the spectral barycenter of the synthesized light. For the spectral weights of red light, For the spectral weights of green light, For the spectral weights of blue light, The wavelength of red light It is the wavelength of green light. It is the wavelength of blue light.
[0054] The scalar parameter of turbidity of the transformer oil, obtained from the above calculation, is C=2.72. The first turbidity threshold C1=1.5 and the second turbidity threshold C2=3.0 are set. C1 < C < C2 represents a medium turbidity environment. The adjustment method is determined to shift the spectral centroid wavelength of the light emitted by the light source towards longer wavelengths. The final spectral ratio is then determined. .
[0055] When the turbidity scalar parameter of the transformer oil is greater than or equal to the second turbidity threshold, the adjustment method is achieved through the following formula: in, For the round trip time of light, Let D be the speed of light in transformer oil, and D be the target distance. The camera shutter speed is controlled at... Always open, at Always turn off Select the gate width for the time.
[0056] By utilizing a time window to physically shield near-field echo scattering noise, only the target-reflected light is allowed to enter the camera for imaging; the signal-to-noise ratio improvement formula for gating imaging is: In the formula, For traditional imaging signal-to-noise ratio, Select gate width for traditional exposure. The pulse width of a pulsed laser. To select the imaging signal-to-noise ratio.
[0057] S4: Based on the geometry of the target, determine the brightness weight of the light source on the robot in different regions. Calculate the spatial noise reduction coefficient based on the average pixel intensity of different regions of the target and the reflected light intensity of the target. After adjusting the brightness weight to make the spatial noise reduction coefficient greater than a preset threshold, acquire and output the target image.
[0058] like Figure 1 As shown, the LED array mounted on the robot is divided into N independent and controllable lighting clusters, where N ranges from [4, 16], and is preferably 4 in this embodiment. The clusters are evenly distributed in a ring around the camera lens. Each lighting cluster is equipped with an independent driving circuit to achieve individual brightness control. The target geometry is identified by a deep learning-based visual preprocessing algorithm, and the training dataset contains images of typical components such as transformer windings, bushings, and tank walls.
[0059] Define the brightness weight matrix of the illumination cluster Matrix elements Indicates the first The lighting cluster is paired with the first The brightness weights of each region, and the differentiated weight allocation for different targets, are achieved through the following formula: The geometry is a cylindrical winding, and the weighting formula is: in, For the first The distance from each region to the winding axis, where σ is the standard deviation of the Gaussian function. For the first The lighting clusters are divided by the lighting LED array. Brightness weight of each region and The value of is a natural number greater than 0.
[0060] The geometry is a planar box wall, and the mathematical model for weight allocation is: in, For the first The distance from each area to the edge of the box wall, The maximum half width of the box wall. For the first The lighting clusters are divided by the lighting LED array. Brightness weight of each region and The value of is a natural number greater than 0.
[0061] The geometry is a sleeve interface, and the mathematical model for weight allocation is: in, For the first The distance from each area to the interface gap, The width of the interface gap. For the Dirac function, For the first The lighting clusters are divided by the lighting LED array. Brightness weight of each region and The value of is a natural number greater than 0.
[0062] The target geometry was identified as a "cylindrical winding".
[0063] For cylindrical windings, a "edge-filling illumination, center-reducing illumination" strategy is implemented, with edge illumination clusters (azimuth angle) Winding radius Gaussian function standard deviation Each lighting cluster illuminates an area that is considered a sector, resulting in a total of four sectors. The distance from the center of each sector to the winding axis is... Approximately (Lighting ring radius); The calculation process for weight allocation is as follows: Because this weight value is too low (close to 0.1), it does not meet the requirement of 0.7–0.9 for edge weights in the "edge illumination" strategy. Therefore, the system normalizes this weight and maps it to a preset weight range. Among them, the theoretical minimum weight Theoretical maximum weight (correspond ), preset edge weight range The final edge lighting cluster weights are obtained. Since the lighting clusters in this system are all arranged at the edge and there is no central lighting cluster, the "center dimming" strategy is not implemented. The weight of all four lighting clusters is set to 0.73 to highlight the radial texture features of the winding.
[0064] Calculate the spatial noise reduction coefficient Evaluation of backscattering suppression effect: in, The average pixel intensity of the target region; The reflected light intensity of the target is calculated using the average pixel value of the target area; this is achieved by adjusting the weight matrix. This reduces the noise reduction coefficient. Greater than the preset threshold This enhances the three-dimensional recognition of the target. Substitute the measured data: average pixel intensity of the target area Average pixel intensity in the background (backscattered) region , Preset threshold , The optimization is effective.
[0065] Backscattering suppression rate calculation: At this point, images are acquired and output. The preferred visual preprocessing algorithm for this deep learning approach is YOLOv8.
[0066] Example 2 This embodiment is based on embodiment 1, but adds a closed-loop evaluation operation for image quality.
[0067] S5: Extract three quality indicators of the target image in real time: signal-to-noise ratio, information entropy, and Tenengrad gradient value. Establish a negative feedback adjustment loop. When any quality indicator fails to reach the preset threshold, trigger the light source power adjustment, multi-frame image fusion, and robot slow-down exposure compensation strategy in sequence to update control parameters, re-acquire images, and output the target image.
[0068] Three quality indicators of the target image are extracted in real time.
[0069] Signal-to-noise ratio (SNR): in, The average pixel value of the target region. The standard deviation of the background region pixels is used as the preset signal-to-noise ratio threshold. If the image quality is below this threshold, it is considered that the image quality is substandard.
[0070] Mean pixel value of the target region: Standard deviation of background region pixels Substitute the values for specific calculations: SNR = 23.6dB (>20 dB threshold), the measured result meets the standard.
[0071] Information entropy ( The calculation formula is: in, grayscale value The pixel ratio, information entropy reflects the amount of information in the image, and a preset threshold. If the value is below this threshold, it is considered that the image details are lost.
[0072] Measured data: The target image has a uniform grayscale distribution, and the percentage of pixels at each grayscale level is statistically analyzed. Consistent with the distribution characteristics of natural images, substituting them into the calculation yields: Preset threshold , The measured results met the standards.
[0073] Tenengrad gradient value ( The calculation formula is: in, , These are the Sobel gradient operator response values in the x and y directions, respectively; obtained by performing Sobel convolution on the image, where M×N is the total number of pixels in the image, and a preset threshold is used. To improve the real-time performance of indicator calculations, GPU parallel computing technology is used to accelerate the feature extraction process, dividing the image into multiple sub-blocks for parallel processing. Actual test data: After Sobel operator convolution calculation, the gradient response values are summed and averaged to obtain: Preset threshold: The actual test results met the standards.
[0074] All three image quality metrics meet the preset threshold requirements, so there is no need to trigger a compensation mechanism, and the system maintains the current optimal control vector. Y=[4.6 W, 120∘, 0.7:0.2:0.1, 200 ns].
[0075] If any quality indicator fails to reach the preset threshold, the following compensation steps are triggered sequentially: Level 1: Light source power adjustment; Level 2: Multi-frame image fusion; Level 3: Robot slows down exposure, updates control parameters, re-acquires images, and outputs the target image.
[0076] Level 1 compensation: Gradually increase the peak power of the light source pulse by 10%-50% in increments of 10%, until the maximum monochromatic light power of 5W is reached, in order to increase the intensity of the target reflected signal. The constraint condition for power increase is: in This is the power boost factor. This represents the current peak power of the light source pulse. To enhance the peak power of the light source pulse, This represents the maximum power of monochromatic light.
[0077] Secondary compensation: If the image quality still does not meet the standard after primary compensation, a multi-frame image fusion algorithm is activated. This involves Gaussian-weighted fusion of 3 to 10 consecutively acquired frames, significantly improving the image signal-to-noise ratio and stability. The calculation formula is: Where k is the number of fused frames; Let be the weight of the k-th frame image. The weight value is positively correlated with the signal-to-noise ratio of the frame, and the calculation formula is: ; Let be the grayscale value of the k-th frame image at pixel (x,y).
[0078] Level 3 compensation: If the image quality still does not meet the standard after Level 2 compensation, reduce the inspection robot's moving speed by 20%-50% in 10% increments, while simultaneously extending the camera exposure time to increase the photon integration per frame and improve overall image brightness and detail. The formula relating speed and exposure time is: In the formula This is the speed reduction factor. The moving speed of the inspection robot before secondary compensation. The moving speed of the inspection robot after secondary compensation. To compensate for the extended exposure time of the front camera To compensate for the extended camera exposure time, for example, from 10μs to 50μs, the maximum extension of the exposure time is 100μs, to avoid excessive exposure causing motion blur in the image.
[0079] When the inspection reached the middle of the winding, the local oil temperature rose to 52℃, and the oil turbidity C instantly increased to 3.1, causing the image SNR to drop to 18.2dB (below the threshold). The system triggered a three-level compensation strategy: Level 1 compensation: the peak power of the light source pulse was increased from 3.5W to 4.5W (an increase of 28.6%, which did not reach the 5W safety limit); Level 2 compensation: Gaussian weighted fusion of 5 frames of images was enabled, with the weights positively correlated with the signal-to-noise ratio of each frame; Level 3 compensation: the robot's moving speed was reduced from 0.1m / s to 0.06m / s (a decrease of 40%), and the camera exposure time was extended from 50μs to 80μs (not reaching the 100μs limit). After compensation, the image quality was restored, the SNR rebounded to 25.1dB, all indicators met the standards again, and the system continued steady-state inspection.
[0080] The above image quality assessment and parameter compensation process is executed cyclically with a fixed period of 10ms, forming a fast closed-loop control. When all image quality indicators are stable above the threshold for several consecutive cycles, the system enters the "steady-state inspection mode" to maintain the current optimal parameter combination. Once the environmental optical state or operating conditions change and cause the image quality to decline, the system immediately restarts the iterative optimization process to ensure that high-quality visual images that meet the defect identification requirements are continuously output throughout the entire inspection task.
[0081] The present invention also provides a transformer oil subsurface vision enhancement system, comprising: Oil environment optical state feature acquisition module: acquires the oil extinction coefficient, volume scattering function and turbidity scalar parameters of transformer oil in the transformer under test, and generates an environmental optical state feature vector after temperature drift compensation.
[0082] Variable matching module: Based on the environmental optical state feature vector, target distance, and robot moving speed, it matches the light source power, spectral ratio of the light source on the robot, and the polarization angle and time gate width of the camera on the robot.
[0083] The graded control module: When the turbidity scalar parameter of the transformer oil is less than the first turbidity threshold, the angle between the light source and the camera polarizer is adjusted to the matched polarization angle, orthogonal polarization images and parallel polarization images are acquired, and the depolarization scattering component is suppressed based on the difference coefficient that is proportional to the extinction coefficient.
[0084] When the turbidity scalar parameter of the transformer oil is greater than or equal to the first turbidity threshold and less than the second turbidity threshold, the spectral ratio of the light source mounted on the robot is adjusted so that the spectral centroid wavelength of the light emitted by the light source shifts towards the longer wavelength direction.
[0085] When the turbidity scalar parameter of the transformer oil is greater than or equal to the second turbidity threshold, a nanosecond-level pulsed laser is activated as the light source. Based on the propagation speed of light in the transformer oil, the round-trip time of the light is calculated according to the target distance in step S2. Based on the time gating width, the camera shutter time window is controlled.
[0086] Target recognition module: Based on the geometry of the target, the brightness weight of the light source on the robot in different regions is determined. The spatial noise reduction coefficient is calculated based on the average pixel intensity of different regions of the target and the reflected light intensity of the target. After adjusting the brightness weight to make the spatial noise reduction coefficient greater than a preset threshold, the target image is acquired and output.
Claims
1. A method for visual enhancement under transformer oil, characterized in that, Includes the following steps: S1: Obtain the extinction coefficient, volume scattering function, and turbidity scalar parameters of the transformer oil in the transformer to be tested, and generate an environmental optical state feature vector after temperature drift compensation; S2: Based on the environmental optical state feature vector, target distance, and robot moving speed, the light source power, spectral ratio, polarization angle, and time gate width of the light source mounted on the robot are matched. S3: When the turbidity scalar parameter of the transformer oil is less than the first turbidity threshold, adjust the angle between the light source and the camera polarizer to the polarization angle matched in step S2, acquire orthogonal polarization images and parallel polarization images, and suppress the depolarization scattering component based on the difference coefficient that is proportional to the extinction coefficient. When the turbidity scalar parameter of the transformer oil is greater than or equal to the first turbidity threshold and less than the second turbidity threshold, the spectral ratio of the light source mounted on the robot is adjusted so that the spectral centroid wavelength of the light emitted by the light source shifts towards the long wavelength direction. When the turbidity scalar parameter of the transformer oil is greater than or equal to the second turbidity threshold, a nanosecond-level pulsed laser is activated as the light source. Based on the propagation speed of light in the transformer oil, the round-trip time of light is calculated according to the target distance in step S2. Based on the time gating width, the camera shutter time window is controlled. S4: Based on the geometry of the target, determine the brightness weight of the light source on the robot in different regions. Calculate the spatial noise reduction coefficient based on the average pixel intensity of different regions of the target and the reflected light intensity of the target. After adjusting the brightness weight to make the spatial noise reduction coefficient greater than a preset threshold, acquire and output the target image.
2. The method for enhancing visual perception under transformer oil according to claim 1, characterized in that, S4 is followed by the following steps: S5: Extract three quality indicators of the target image in real time: signal-to-noise ratio, information entropy, and Tenengrad gradient value. Establish a negative feedback adjustment loop. When any quality indicator fails to reach the preset threshold, trigger the light source power adjustment, multi-frame image fusion, and robot slow-down exposure compensation strategy in sequence to update control parameters, re-acquire images, and output the target image.
3. The method for enhancing visual perception under transformer oil according to claim 1, characterized in that, In S3, when the turbidity scalar parameter of the transformer oil is less than the first turbidity threshold, the adjustment method is achieved through the following formula: in, It is an orthogonal polarization image. This is a parallel polarization image. The difference coefficients are determined using the following formula: in, The extinction coefficient after temperature drift compensation. The maximum calibrated value for the extinction coefficient is [value to be filled in]. .
4. The method for enhancing visual perception under transformer oil according to claim 1, characterized in that, In S3, when the turbidity scalar parameter of the transformer oil is greater than or equal to the first turbidity threshold and less than the second turbidity threshold, the adjustment method is achieved through the following formula: in, The wavelength of the spectral barycenter of the synthesized light. For the spectral weights of red light, For the spectral weights of green light, For the spectral weights of blue light, The wavelength of red light It is the wavelength of green light. It is the wavelength of blue light.
5. The method for enhancing visual perception under transformer oil according to claim 1, characterized in that, In S3, when the turbidity scalar parameter of the transformer oil is greater than or equal to the second turbidity threshold, the adjustment method is achieved through the following formula: in, For the round trip time of light, Let D be the speed of light in transformer oil, and D be the target distance. The camera shutter speed is controlled at... Always open, at Always turn off Select the gate width for the time.
6. The method for enhancing visual perception under transformer oil according to claim 1, characterized in that, Environmental optical state feature vector in S1 The calculation is performed using the following formula: in, The extinction coefficient after temperature drift compensation. This is the volume scattering function after temperature drift compensation. The turbidity scalar parameter is... T It is the transpose of the vector.
7. The method for enhancing visual perception under transformer oil according to claim 1, characterized in that, The matching method in S2 is based on a pre-trained weight allocation function. The environmental optical state feature vector, target distance, and robot speed are used as input variables, while the light source power, spectral ratio, polarization angle, and time gate width are used as output variables. The weight allocation function is implemented using the following formula: in, For the first One output variable, For the first The input variable for the first... The weight coefficients of each output variable, For the first One input variable, The number of input variables. The range of values is [1- ], For the number of dimensions of the input variables, The range of values is natural numbers greater than 0. For the first The bias values of each output variable.
8. The method for enhancing visual perception under transformer oil according to claim 1, characterized in that, In S4, differential weight allocation for different target geometries is achieved through the following formula: The geometry is a cylindrical winding, and the weighting formula is: in, For the first The distance from each region to the winding axis, where σ is the standard deviation of the Gaussian function. For the first The lighting clusters are divided by the lighting LED array. Brightness weight of each region and The value of is a natural number greater than 0; The geometry is a planar box wall, and the mathematical model for weight allocation is: in, For the first The distance from each area to the edge of the box wall, The maximum half width of the box wall. For the first The lighting clusters are divided by the lighting LED array. Brightness weight of each region and The value of is a natural number greater than 0; The geometry is a sleeve interface, and the mathematical model for weight allocation is: in, For the first The distance from each area to the interface gap. The width of the interface gap. For the Dirac function, For the first The lighting clusters are divided by the lighting LED array. Brightness weight of each region and The value of is a natural number greater than 0.
9. The method for enhancing visual perception under transformer oil according to claim 1, characterized in that, S4 spatial noise reduction coefficient The calculation is performed using the following formula: in, The average pixel intensity of the target region; The intensity of the light reflected from the target.
10. A transformer oil subsurface vision enhancement system, characterized in that, Includes the following modules: Oil environment optical state feature acquisition module: acquires the oil extinction coefficient, volume scattering function and turbidity scalar parameters of transformer oil in the transformer under test, and generates an environmental optical state feature vector after temperature drift compensation; Variable matching module: Based on the environmental optical state feature vector, target distance, and robot moving speed, it matches the light source power, spectral ratio, polarization angle, and time gate width of the light source mounted on the robot and the camera mounted on the robot. The graded control module: When the turbidity scalar parameter of the transformer oil is less than the first turbidity threshold, the angle between the light source and the camera polarizer is adjusted to the matched polarization angle, orthogonal polarization image and parallel polarization image are acquired, and the depolarization scattering component is suppressed based on the difference coefficient proportional to the extinction coefficient. When the turbidity scalar parameter of the transformer oil is greater than or equal to the first turbidity threshold and less than the second turbidity threshold, the spectral ratio of the light source mounted on the robot is adjusted so that the spectral centroid wavelength of the light emitted by the light source shifts towards the long wavelength direction. When the turbidity scalar parameter of transformer oil is greater than or equal to the second turbidity threshold, a nanosecond-level pulsed laser is activated as the light source. Based on the speed of light propagation in transformer oil, the round-trip time of light is calculated according to the target distance. Based on the time gating width, the camera shutter time window is controlled. Target recognition module: Based on the geometry of the target, the brightness weight of the light source on the robot in different regions is determined. The spatial noise reduction coefficient is calculated based on the average pixel intensity of different regions of the target and the reflected light intensity of the target. After adjusting the brightness weight to make the spatial noise reduction coefficient greater than a preset threshold, the target image is acquired and output.