Transformer oil pollution degree evaluation method based on lensless imaging and machine learning

Abstract

The application relates to a transformer oil pollution degree evaluation method based on lens-free imaging and machine learning, relates to the technical field of transformer oil detection, and is used for solving the problems that the existing transformer oil particle pollution detection technology cannot obtain particle three-dimensional morphology and the detection result lacks quantitative correlation with the pollution degree. The application comprises the following steps: S1, acquiring a lens-free diffraction image of particles in a transformer oil sample to be detected; S2, amplitude reconstruction of the diffraction image; S3, phase reconstruction and thickness inversion of the diffraction image; S4, particle region segmentation and extraction; S5, structured feature extraction; S6, particle classification; and S7, pollution degree grade evaluation. The application realizes synchronous acquisition of particle two-dimensional morphology and three-dimensional thickness information by acquiring particle diffraction information in a lens-free imaging mode, combining amplitude reconstruction and thickness inversion, realizes automatic identification of particle types by combining structured feature extraction with a machine learning model, and realizes intelligent and quantitative evaluation of the pollution degree grade of transformer oil.

Description

Technical Field

[0001] This invention relates to the field of transformer oil testing technology, specifically a method for assessing transformer oil contamination based on lensless imaging and machine learning. Background Technology

[0002] As oil-immersed transformers operate for longer periods, the transformer oil is inevitably affected by mechanical wear, insulation paper aging, electrical breakdown, and external contamination, resulting in various particulate contaminants, including cellulose particles, metal particles, carbon particles, and bubbles. These particles are prone to polarization, migration, and aggregation under the influence of an electric field, causing localized electric field distortion, reducing the oil's breakdown voltage, and accelerating the aging of the oil-paper insulation system. This has become a significant hidden danger leading to transformer insulation failures.

[0003] Existing detection methods for particulate contamination in transformer oil mainly include optical obscuration, microscopy, ferrography, laser spectroscopy, and microfluidic imaging. The shortcomings of the above detection methods are: (1) Optical obscuration can only provide the total number and size distribution of particles, but cannot identify the morphology, type, or three-dimensional features of particles, and cannot be effectively used for contamination source analysis or insulation fault diagnosis. (2) Microscopy requires complex sample preparation, has a limited field of view, and relies on manual experience judgment, which is highly subjective and difficult to meet the needs of high-throughput detection and rapid on-site assessment. (3) Ferrography is not suitable for non-magnetic particles such as cellulose and carbon particles, and also requires sample preparation and manual judgment. (4) Laser spectroscopy is easily affected by the background spectrum of the oil itself, and cannot accurately identify the spatial position and morphology information of small particles. It usually cannot directly construct a quantitative relationship between particle morphology and contamination degree. (5) Microfluidic imaging technology relies on lens optical systems, which have problems such as high cost, complex structure, and limited field of view, and is not suitable for deployment in substations.

[0004] In summary, the above detection methods share the following common problems: (1) They cannot simultaneously obtain the three-dimensional morphology parameters of particles. (2) They cannot perform automated classification and are highly dependent on human experience. (3) They cannot establish a quantitative correlation between the characteristics of contaminant particles and the deterioration of oil insulation performance. (4) They are difficult to achieve rapid, high-throughput, and low-cost on-site detection. Summary of the Invention

[0005] The purpose of this invention is to provide a method for assessing transformer oil contamination based on lensless imaging and machine learning, which addresses the problems of existing transformer oil particulate contamination detection technologies, such as the inability to obtain the three-dimensional morphology of particles, the lack of quantitative correlation between detection results and contamination levels, and the difficulty in achieving rapid on-site detection.

[0006] The technical solution adopted by the present invention to solve its technical problem is: a method for assessing transformer oil contamination based on lensless imaging and machine learning, including the following steps.

[0007] S1. Lensless diffraction images of particles in the transformer oil sample under test are obtained using a lensless imaging method.

[0008] S2. Amplitude reconstruction of lensless diffraction image: Amplitude reconstruction of the diffraction image obtained in step S1 is performed using the angular spectrum method.

[0009] S3. Phase reconstruction and thickness inversion of diffraction images.

[0010] use (1) Calculate the complex amplitude distribution; where, For complex amplitude The modulus; It represents exponential operations with the natural constant as the base; Represents the imaginary unit, satisfying ; The phase of the light field; determined by the formula (2) Calculate the thickness of the particles; where, The difference in refractive index between the particles and the transformer oil; Indicates the wavelength of the incident light.

[0011] S4. Particle region segmentation and extraction.

[0012] The particle regions in the distribution image are segmented and labeled to obtain the two-dimensional projection region of a single particle; and the thickness distribution map of the particle is extracted based on the two-dimensional projection region.

[0013] S5. Structured feature extraction.

[0014] Based on the two-dimensional projection region of a single particle, two-dimensional geometric features of the particle are extracted from the amplitude modulus image to characterize the particle's morphological structure in the plane. The two-dimensional geometric features include length L, width W, projected area A, and aspect ratio AR. Based on the two-dimensional projection region of the single particle, three-dimensional thickness features of the particle are extracted from the corresponding thickness distribution map to characterize the structural differences of the particle in the thickness direction. The three-dimensional thickness features include thickness T and roundness C. The two-dimensional geometric features and the three-dimensional thickness features are combined into a feature vector X={L,W,T,C,AR,A}.

[0015] S6. Particle classification based on XGBoost: Input the feature vector X into the XGBoost classification model to identify the particle type.

[0016] S7. Pollution level assessment.

[0017] Based on the particle type identification results, statistical analysis is performed on the quantity, size, and thickness characteristics of different particle types to form a set of particle statistical parameters. Based on the particle statistical parameters, a pollution contribution index for each type of particle is constructed to quantify the impact of different particle types on the degree of transformer oil pollution. The pollution contribution of each type of particle is obtained by a weighted combination of its quantity, size, and thickness parameters. By weighted fusion of the pollution contributions of different particle types, a comprehensive pollution index of the oil sample is obtained to characterize the overall pollution level of the transformer oil. The pollution level of the transformer oil is output as a basis for transformer insulation health status assessment, operation and maintenance decision-making, and fault early warning analysis.

[0018] Further, the specific content of step S1 is as follows: The lensless imaging module for implementing the lensless imaging method includes an LED light source, a micro-aperture, and a CMOS image sensor. The transformer oil sample to be tested is placed on a transparent glass slide and illuminated by the LED light source. The light source generates illumination light through the micro-aperture. The scattered light generated by the particles under the illumination light interferes with the direct reference light on the surface of the CMOS image sensor, forming a lensless diffraction image. The sample scattered light and the direct reference light are superimposed on the surface of the CMOS image sensor to form a light intensity distribution. (3); among which, The coordinates on the imaging plane are Lensless diffraction intensity distribution values ​​collected at the location; The light scattered by the sample; This is the direct reference light.

[0019] Furthermore, based on The initial complex amplitude of the light field at the imaging plane Represented as: (4); among which, Indicates the initial phase distribution. ;right Performing a two-dimensional Fourier transform yields the spectral distribution function of the initial light field in the spatial frequency domain. : (5); among which, Represents the spatial frequency variable along the x-direction; Represents the spatial frequency variable along the y-direction; This represents a two-dimensional Fourier transform operator; the angular spectrum method is used to recover the two-dimensional amplitude map of the sample from the diffraction image, propagating to a distance... Spatial frequency spectrum of the backplane light field (6); among which, This represents the propagation distance between the particle plane and the imaging plane; the propagation distance is then calculated. ,right Perform the inverse Fourier transform, i.e. (7) Obtain the complex amplitude at the propagation distance z. ;in, This is the inverse Fourier transform operator.

[0020] Further, in step S5, the length L is defined as the maximum side length of the minimum bounding rectangle of the particle in the two-dimensional projection region of a single particle, which is obtained by performing boundary detection on the binary image and calculating its minimum bounding rectangle.

[0021] Further, in step S5, the width W is defined as the length of the shorter side of the smallest bounding rectangle of the particle, and the ratio of the length L to the width W is the aspect ratio AR. (8) Describe the degree of flatness and elongation of the particle shape.

[0022] Furthermore, in step S5, the formula for calculating roundness C is: (9), where A represents the projected area of ​​the particle; P represents the perimeter of the particle's outline.

[0023] Furthermore, in step S5, the thickness T is taken as... The maximum value.

[0024] Furthermore, in step S6, the particle classification model adopts the XGBoost classification model based on gradient boosting decision tree. By training on sample data with known particle type labels, the mapping relationship between structured features and particle types is established, and the trained particle classification model is obtained.

[0025] Further, in step S6, the feature vector X of each particle is input into the XGBoost classification model, and several CART decision trees within the XGBoost classification model perform feature condition judgments, including whether the thickness is greater than the threshold, whether the roundness is higher than the threshold, and whether the aspect ratio exceeds the limit; the leaf nodes of each tree output the corresponding weight values, and finally the weighted summation module obtains the prediction scores for different particle categories; the final category label of the particle is obtained by softmax or maximum value selection, including metal particles, cellulose particles, carbon particles, and bubbles.

[0026] Furthermore, in step S6, XGBoost employs a gradient boosting strategy: the first tree performs initial classification using six types of features; the second tree learns the classification error of the first tree; the third tree continues to learn the feature patterns that the first two trees failed to process until the loss function converges; the objective function... (10); among which, The penalty coefficient represents the number of leaf nodes; T is the number of leaf nodes in the current tree. This represents the output weight of the j-th leaf node; It is the L2 regularization coefficient.

[0027] Furthermore, in step S6, the outputs of all trees are summed to obtain the scores for each category: (11); among which, This represents the final prediction result for the i-th sample; This represents the output of the k-th decision tree; K represents the total number of trees in the model. Indicates the first The structured feature vector corresponding to each particle sample.

[0028] The beneficial effects of the present invention are: (1) By acquiring particle diffraction information through lensless imaging, large field-of-view particle detection can be achieved without a microscope objective, reducing system complexity and hardware cost; (2) By combining amplitude reconstruction and thickness inversion, the two-dimensional morphology and three-dimensional thickness information of particles can be acquired simultaneously, improving the comprehensiveness and accuracy of particle characterization; (3) By combining structured feature extraction with machine learning models, automatic identification of particle types can be achieved, reducing manual interpretation errors; (4) By combining the particle type identification results with particle quantity, size and thickness parameters, intelligent and quantitative assessment of transformer oil contamination level can be achieved; (5) The detection process is clear and highly automated, suitable for online or on-site rapid assessment of transformer oil contamination status. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of a lensless imaging system.

[0030] Figure 2 This is a schematic diagram of the XGBoost structure;

[0031] Figure 3 The graph shows the XGBoost loss function. Detailed Implementation

[0032] Lensless imaging, a method based on light field encoding and numerical reconstruction, boasts advantages such as simple structure, large field of view, no need for lenses, and portability, and has been used for imaging research on tiny particles, cells, and suspended impurities. Lensless imaging obtains particle amplitude and phase maps by recording diffraction holograms and reconstructing them using angular spectral analysis, thus possessing the potential to extract particle size, shape, and height features. However, current publicly available technologies mainly focus on image restoration, particle visualization, or target detection using deep learning. A three-dimensional particle parameter extraction mechanism based on lensless imaging for transformer oil applications has not yet been established, nor has it been combined with machine learning models for intelligent pollution level assessment. Furthermore, existing technologies propose using support vector machines, K-means clustering algorithms, and convolutional neural network (CNN) algorithms for machine learning models. However, these methods still have significant shortcomings in practical applications. For example, support vector machines and K-means clustering algorithms rely on manually selected features, have limited ability to express high-dimensional structured parameters, and their classification boundaries are significantly affected by noise, making it difficult to maintain stability. Against this backdrop, gradient boosting trees have gradually gained attention due to their efficiency in structured data classification. Extreme Gradient Boosting (XGBoost), as a representative of ensemble learning algorithms, possesses several advantages that are difficult for other algorithms to match. XGBoost uses CART decision trees as its base learner, automatically selecting optimal features and thresholds, and is sensitive to morphological differences, making it particularly suitable for classification tasks with granular structured features. The introduction of L1 / L2 regularization terms effectively suppresses overfitting and avoids performance fluctuations caused by oil sample differences, imaging noise, or complex particle distributions, making it suitable for engineering applications involving multiple environments and oil samples. XGBoost supports parallel feature splitting and has low computational resource requirements, enabling real-time classification on ordinary industrial computing platforms or embedded devices. It does not rely on large GPUs, making it suitable for the speed and reliability requirements of online substation inspection.

[0033] The purpose of this invention is to overcome the shortcomings of existing transformer oil particulate contamination detection technologies, such as the inability to obtain three-dimensional morphology of particles, the inability to automatically classify them, the lack of quantitative correlation between detection results and contamination level, and the difficulty in achieving rapid on-site detection. This invention proposes an intelligent assessment method for transformer oil contamination based on lensless imaging and machine learning. By obtaining multi-dimensional features of particles in the oil through lensless imaging, and using the XGBoost classification model to identify particle types and assess contamination levels, a rapid, accurate, and quantifiable assessment of the degree of transformer oil contamination can be achieved.

[0034] This invention mainly addresses the following problems: (1) How to accurately acquire the three-dimensional morphological features of particle length, width, and thickness without the need for a microscope lens. (2) How to automatically extract the geometric features and morphological information of particles from reconstructed images, avoiding the subjectivity and inefficiency of human interpretation. (3) How to utilize structured features to achieve intelligent classification of particle types, improving recognition accuracy and generalization ability. (4) How to establish a quantitative relationship between the three-dimensional features of particles and the degree of transformer oil contamination, achieving automatic assessment of contamination levels. (5) How to construct a portable, fast, and environmentally adaptable method for detecting contamination levels. The transformer oil contamination assessment method based on lensless imaging and machine learning of this invention includes the following steps.

[0035] S1. Obtain lensless diffraction images of particles in the transformer oil sample to be tested.

[0036] This invention employs a lensless imaging method, such as... Figure 1 As shown, the lensless imaging module includes an LED light source, a microaperture, and a CMOS image sensor. The LED light source is a partially coherent source, and its coherence is insufficient to form a clear diffraction pattern. After spatial filtering by the microaperture, the lateral coherence of the illumination light is effectively improved, enabling the light wave to generate a stable diffraction field on the sample surface, forming a lensless hologram for numerical reconstruction. The sample scattered light and the direct reference light are superimposed on the CMOS surface to form the light intensity distribution. (1). Among them, The coordinates on the imaging plane are The lensless diffraction intensity distribution value collected at the location is the measurement result of the square of the complex amplitude mode of the diffraction light field; The object light scattered by the sample, i.e., the scattered light; The reference light is direct. At this point, the interference fringes contain both amplitude and phase information, which forms the basis for subsequent numerical reconstruction.

[0037] The transformer oil sample to be tested was placed on a transparent glass slide and illuminated by a 550nm LED light source, with the light source passing through a micro-aperture to generate illumination light with high spatial coherence. The scattered light from the particles interfered with the direct reference light on the surface of the CMOS image sensor, forming a lensless diffraction image. This diffraction image served as the input for subsequent reconstruction. The acquired diffraction image is shown below. Figure 1 As shown in Figure a.

[0038] S2. Amplitude reconstruction of lensless diffraction images.

[0039] Amplitude reconstruction is performed on the diffraction image obtained in step S1 using the angular spectrum method.

[0040] Specifically, based on coordinates on the imaging plane Lensless diffraction intensity distribution values ​​collected at [location] The initial complex amplitude of the light field at the imaging plane Represented as: (2). Among them, To represent the initial phase distribution, this invention uses an initial phase of 0, i.e. ; Represents the imaginary unit, satisfying ; This indicates exponential operations with the natural constant as the base. Performing a two-dimensional Fourier transform yields the spectral distribution function of the initial light field in the spatial frequency domain. : (3). Among them, Represents the spatial frequency variable along the x-direction; Represents the spatial frequency variable along the y-direction; This represents the two-dimensional Fourier transform operator.

[0041] The two-dimensional amplitude pattern of the sample was recovered from the diffraction pattern using the angular spectral method. The propagation of the diffraction field in the frequency domain conforms to the free-space transfer function, and the calculation formula is as follows: (4). Among them, To propagate to a distance Spatial frequency spectrum of the backplane optical field; Indicates the wavelength of the incident light used by the imaging system; This represents the propagation distance between the particle plane and the imaging plane. The propagation distance is obtained using formula (4). ,right Perform the inverse Fourier transform, i.e. (5) Obtain the complex amplitude at the propagation distance z. .in, This is the inverse Fourier transform operator. For complex amplitudes... The amplitude modulus value is obtained by taking the modulus value. It is used to characterize the two-dimensional contour structure of particles.

[0042] S3. Phase reconstruction and thickness inversion of diffraction images.

[0043] Complex amplitude distribution is expressed as (6). Among them, This represents the phase of the light field. To extract the three-dimensional thickness information of the particles, [the following is used]: Phase mapping of the optical field reflects the modulation of the optical path by the sample, providing a basis for geometric thickness inversion. The propagation of light within the sample... With geometric thickness satisfy: (7). Among them, This represents the difference in refractive index between the particles and the transformer oil. (Using the formula...) (8) Obtain the thickness of the particles, which reflects the differences in the three-dimensional morphology of different particles.

[0044] S4. Particle region segmentation and extraction.

[0045] After obtaining the amplitude modulus and thickness distribution in steps S2 and S3, morphological processing is performed to segment and label the particle regions in the distribution image, thereby obtaining the two-dimensional projection region of a single particle; and based on the two-dimensional projection region, the thickness distribution map of the particle is extracted accordingly, such as... Figure 1 As shown in b, it is used for subsequent structured feature extraction.

[0046] S5. Structured feature extraction.

[0047] Based on the two-dimensional projection region of a single particle, two-dimensional geometric features of the particle are extracted from the amplitude modulus image to characterize the particle's morphological structure in the plane. The two-dimensional geometric features include length L, width W, projected area A, and aspect ratio AR.

[0048] The length L is defined as the maximum side length of the minimum bounding rectangle of the particle's two-dimensional projection region, reflecting the particle's extension along its principal scale direction. It is obtained by performing boundary detection on the binary image and calculating its minimum bounding rectangle. The length L has significant distinguishing power for identifying cellulose-like strip-shaped particles.

[0049] The width W is defined as the length of the short side of the smallest bounding rectangle of the particle, and is the scale information of the particle in the vertical principal direction.

[0050] The aspect ratio (AR) is defined as the ratio of length (L) to width (W), describing the degree of elongation or flatness of the particle shape and helping to distinguish between long fibrous and blocky particles. (9).

[0051] Based on the two-dimensional projection region of a single particle, the three-dimensional thickness features of the particle are extracted from the corresponding thickness distribution map to characterize the structural differences of the particle in the thickness direction. The three-dimensional thickness features include thickness T and roundness C.

[0052] Thickness T is taken The maximum value.

[0053] The formula for calculating roundness C is: (10), where A represents the projected area of ​​the particle, and P represents the pixel area occupied by the particle on the reconstructed amplitude map; P represents the perimeter of the particle outline, and the closer the roundness is to 1, the closer the particle is to a circle. Bubbles usually have high roundness, while cellulose and metal fragments are irregular in shape and have low roundness. Therefore, roundness is an important classification feature.

[0054] The two-dimensional geometric features and three-dimensional thickness features are combined into a feature vector X={L,W,T,C,AR,A}, which is used as the input to the particle classification model.

[0055] S6. Particle classification based on XGBoost.

[0056] In step S5, a structured feature vector X containing two-dimensional geometric features and three-dimensional thickness features has been constructed for each particle. Based on the feature vector X, a machine learning classification model is used to identify the particle type.

[0057] In this invention, the particle classification model adopts the XGBoost classification model based on gradient boosting decision trees. By training on sample data with known particle type labels, a mapping relationship between structured features and particle types is established, resulting in a trained particle classification model.

[0058] The feature vector of each particle is input into the XGBoost classification model, where several CART decision trees within the XGBoost model perform feature condition judgments. For example, they assess whether the thickness exceeds a threshold, whether the roundness exceeds a threshold, and whether the aspect ratio exceeds a limit. The leaf nodes of each tree output corresponding weight values, and a weighted summation module ultimately obtains the predicted score for different particle categories. The final category label for the particles is obtained through softmax or maximum value selection, including metal particles, cellulose particles, carbon particles, and bubbles. This method has the advantages of strong interpretability, high adaptability to structured features, and small sample requirements, making it more suitable for practical applications compared to deep neural networks.

[0059] XGBoost is based on a forward-distributed additive model. It iteratively builds decision trees, with each tree learning the residuals from the previous iteration, thus gradually approaching the true classification boundary. When constructing each CART decision tree, XGBoost automatically selects the feature that maximizes the decrease in the loss function and a threshold from six features as the split point. XGBoost introduces a second-order gradient for Taylor expansion of the objective function, making the loss function optimization more accurate and improving classification performance. Figure 2 As shown, XGBoost employs a gradient boosting strategy: the first tree performs initial classification using six types of features; the second tree learns the classification error of the first tree; the third tree continues to learn feature patterns that the first two trees failed to handle until the loss function converges. The objective function includes a regularization term: (11). Among them, The penalty coefficient for the number of leaf nodes is represented by a hyperparameter; T is the number of leaf nodes in the current tree. It is the L2 regularization coefficient, a hyperparameter used to control the penalty intensity; represents the output weight of the j-th leaf node, i.e., the score given by the model when the sample falls on that leaf. By introducing the regularization constraint in formula (11), the depth of the tree and the output value of the leaf nodes are restricted, thereby preventing the model from overfitting and improving the overall generalization ability. In the stepwise fitting process of the gradient boosting tree, different features play different roles at different stages: roundness and aspect ratio are suitable for the early stage of model training, used to quickly distinguish between spherical particles and fibrous particles. Thickness, length and area have a higher participation in the later stage of model training, used to finely correct complex samples located near the class discrimination boundary. At the same time, width and area features can be used to supplement the discrimination of irregular particle shapes. Finally, through the ensemble learning of multiple decision trees, a strong classifier composed of multiple weak classifiers is formed, realizing the joint modeling of multi-dimensional particle features, so that the model's learning ability in complex particle recognition tasks is significantly better than that of a single decision tree or a discrimination method based on manual rules.

[0060] Each decision path eventually reaches a leaf node, and the leaf node outputs the weights. This reflects the confidence level that the feature combination points to a certain type of particle. The outputs of all trees are summed to obtain the scores for each category: (12). Among them, This represents the final prediction result for the i-th sample; This represents the output of the k-th decision tree, i.e., the leaf weight corresponding to the sample falling into a certain leaf node of the tree; K represents the total number of trees in the model. Indicates the first The structured feature vector corresponding to each particle sample.

[0061] In building the XGBoost-based particle classification model, to ensure good generalization ability and avoid overfitting, the particle samples after lensless imaging reconstruction and feature extraction were divided into training and validation sets in a 4:1 ratio. The training set, comprising 80% of all samples, was used for gradient boosting training, iteratively learning the mapping relationship between particle length, width, thickness, roundness, aspect ratio, and their respective categories. The validation set, comprising 20%, was used to monitor model performance in real-time during training, assess the loss function's decreasing trend, and adjust the learning rate, tree depth, and regularization hyperparameters to prevent the model from overfitting the training data and thus reducing its generalization ability. This 4:1 data partitioning ratio ensured sufficient sample size for building a stable classification tree model during training while allowing for objective evaluation of model performance through the validation set. The training process is as follows: Figure 3 As shown, under this partitioning strategy, the model's loss function can converge stably, and both classification accuracy and robustness reach the optimal range, further improving the reliability of this invention in practical transformer oil detection applications. Based on the prediction results of formula (12)... It characterizes the type and category of particles and serves as the basis for subsequent pollution level assessments.

[0062] S7. Pollution level assessment.

[0063] In step S6, the type identification results of each particle in the oil sample have been obtained. Based on the particle type identification results, statistical analysis is performed on the quantity characteristics, size characteristics, and thickness characteristics of different types of particles to form a set of particle statistical parameters.

[0064] The differences in quantity, size, and three-dimensional structural characteristics of different particle types are used to reflect the quantity level and potential hazard of contaminants in the oil sample, thereby comprehensively assessing the contamination level of transformer oil. Based on particle statistical parameters, a contamination contribution index for each type of particle is constructed to quantify the impact of different particle types on the contamination level of transformer oil. The contamination contribution of each type of particle is obtained by a weighted combination of its quantity, size, and thickness parameters.

[0065] By weighted fusion of the pollution contributions of different particle types, a comprehensive pollution index is obtained for the oil sample, which characterizes the overall pollution level of the transformer oil. The weighting coefficients for different particle types reflect the differences in their impact on insulation performance and operational safety. The output pollution level of the transformer oil serves as the basis for assessing the insulation health status of the transformer, making operation and maintenance decisions, and analyzing fault warnings.

[0066] In summary, this invention achieves intelligent and quantifiable assessment of transformer oil particulate contamination through lensless imaging, phase reconstruction, 3D feature extraction, and machine learning classification techniques. This method eliminates the need for a microscope, facilitating on-site deployment and offering advantages such as high throughput, low cost, and high robustness. It is suitable for applications involving power equipment condition monitoring and oil insulation performance evaluation.

[0067] The pollution level assessment process of the present invention will be described below with reference to specific embodiments.

[0068] Example 1: A transformer oil sample containing cellulose aging debris was placed in a lensless imaging device. A diffraction image was acquired by illumination with an LED light source, and the amplitude modulus image and phase image were obtained through the reconstruction module of this invention. Thickness inversion was performed based on the phase image to obtain the particle thickness distribution. After processing by the structured feature extraction module, the feature vector X of a single particle was obtained as follows: The feature vector is input into the XGBoost particle classification model, which identifies the particle as a cellulose particle. Combining the number and type distribution of particles in the oil sample, a comprehensive pollution index is calculated, and based on this, the pollution level of the oil sample is determined to be Level 2, indicating that the oil sample has a certain degree of aging pollution, but has not yet reached a serious risk level.

[0069] Example 2: Thick particles exceeding 5 μm in thickness (T) were detected in the oil sample, exhibiting large size and regular shape. Identified as metal particles using the XGBoost particle classification model. Based on the high risk weight of metal particles in pollution assessment, combined with their quantity and thickness parameters, the calculated comprehensive pollution index corresponds to a pollution level of 3, indicating potential discharge or wear hazards in the transformer oil, requiring further testing or maintenance measures.

[0070] This invention acquires particle diffraction information through lensless imaging, enabling large-field particle detection without a microscope objective, thus reducing system complexity and hardware costs. By combining amplitude reconstruction and thickness inversion, it achieves simultaneous acquisition of two-dimensional particle morphology and three-dimensional thickness information, improving the comprehensiveness and accuracy of particle characterization. Through the combination of structured feature extraction and machine learning models, it achieves automatic particle type identification, reducing human interpretation errors. By combining particle type identification results with particle quantity, size, and thickness parameters, it enables intelligent and quantitative assessment of transformer oil contamination levels. The method of this invention has a clear detection process and a high degree of automation, making it suitable for rapid online or on-site assessment of transformer oil contamination status.

Claims

1. A method for transformer oil pollution degree assessment based on lensless imaging and machine learning, characterized in that, Includes the following steps: S1. Lensless diffraction images of particles in the transformer oil sample under test are obtained using a lensless imaging method. S2. Amplitude reconstruction of lensless diffraction image: Amplitude reconstruction of the diffraction image obtained in step S1 is performed using the angular spectrum method. S3. Phase reconstruction and thickness inversion of diffraction images; use (1) Calculate the complex amplitude distribution; where, For complex amplitude The modulus; It represents exponential operations with the natural constant as the base; Represents the imaginary unit, satisfying ; The phase of the light field; determined by the formula (2) Calculate the thickness of the particles; where, The difference in refractive index between the particles and the transformer oil; Indicates the wavelength of the incident light; S4. Particle region segmentation and extraction; The particle regions in the distribution image are segmented and labeled to obtain the two-dimensional projection region of a single particle; and the thickness distribution map of the particle is extracted based on the two-dimensional projection region. S5. Structured feature extraction; Based on the two-dimensional projection region of a single particle, two-dimensional geometric features of the particle are extracted from the amplitude modulus image to characterize the particle's morphological structure in the plane. The two-dimensional geometric features include length L, width W, projected area A, and aspect ratio AR. Based on the two-dimensional projection region of the single particle, three-dimensional thickness features of the particle are extracted from the corresponding thickness distribution map to characterize the structural differences of the particle in the thickness direction. The three-dimensional thickness features include thickness T and roundness C. The two-dimensional geometric features and the three-dimensional thickness features are combined into a feature vector X={L,W,T,C,AR,A}. S6. Particle classification based on XGBoost: Input the feature vector X into the XGBoost classification model to identify the particle type; S7. Pollution Level Assessment; Based on the particle type identification results, statistical analysis is performed on the quantity, size, and thickness characteristics of different particle types to form a set of particle statistical parameters. Based on the particle statistical parameters, a pollution contribution index for each type of particle is constructed to quantify the impact of different particle types on the degree of transformer oil pollution. The pollution contribution of each type of particle is obtained by a weighted combination of its quantity, size, and thickness parameters. By weighted fusion of the pollution contributions of different particle types, a comprehensive pollution index of the oil sample is obtained to characterize the overall pollution level of the transformer oil. The pollution level of the transformer oil is output as a basis for transformer insulation health status assessment, operation and maintenance decision-making, and fault early warning analysis.

2. The method for evaluating the degree of transformer oil contamination based on lensless imaging and machine learning according to claim 1, characterized in that, Step S1 specifically involves the following: The lensless imaging module for implementing the lensless imaging method includes an LED light source, a micro-aperture, and a CMOS image sensor. The transformer oil sample to be tested is placed on a transparent glass slide and illuminated by the LED light source. The light source generates illumination light through the micro-aperture. The scattered light generated by the particles under the illumination light interferes with the direct reference light on the surface of the CMOS image sensor, forming a lensless diffraction image. The sample scattered light and the direct reference light are superimposed on the surface of the CMOS image sensor to form a light intensity distribution. (3); among which, The coordinates on the imaging plane are Lensless diffraction intensity distribution values ​​collected at the location; The light scattered by the sample; This is the direct reference light.

3. The method for assessing transformer oil contamination based on lensless imaging and machine learning according to claim 2, characterized in that, based on The initial complex amplitude of the light field at the imaging plane Represented as: (4); among which, Indicates the initial phase distribution. ;right Performing a two-dimensional Fourier transform yields the spectral distribution function of the initial light field in the spatial frequency domain. : (5); among which, Represents the spatial frequency variable along the x-direction; Represents the spatial frequency variable along the y-direction; This represents a two-dimensional Fourier transform operator; the angular spectrum method is used to recover the two-dimensional amplitude map of the sample from the diffraction image, propagating to a distance... Spatial frequency spectrum of the backplane light field (6); among which, This represents the propagation distance between the particle plane and the imaging plane; the propagation distance is then calculated. ,right Perform the inverse Fourier transform, i.e. (7) Obtain the complex amplitude at the propagation distance z. ;in, This is the inverse Fourier transform operator.

4. The lensless imaging and machine learning based transformer oil pollution degree evaluation method according to claim 3, characterized in that, In step S5, length L is defined as the maximum side length of the minimum bounding rectangle of the particle in the two-dimensional projection region of a single particle, obtained by performing boundary detection on the binary image and calculating its minimum bounding rectangle; width W is defined as the short side length of the minimum bounding rectangle of the particle, and the ratio of length L to width W is the aspect ratio AR. (8) Describe the degree of flatness and elongation of the particle shape.

5. The lensless imaging and machine learning based transformer oil pollution degree assessment method according to claim 4, characterized in that, In step S5, the formula for calculating roundness C is: (9), where A represents the projected area of ​​the particle; P represents the perimeter of the particle's outline.

6. The lensless imaging and machine learning based transformer oil pollution degree assessment method according to claim 5, characterized in that, In step S5, the thickness T takes its maximum value.

7. The lensless imaging and machine learning based transformer oil pollution degree assessment method according to claim 6, characterized in that, In step S6, the particle classification model adopts the XGBoost classification model based on gradient boosting decision tree. By training on sample data with known particle type labels, the mapping relationship between structured features and particle types is established, and the trained particle classification model is obtained.

8. The lensless imaging and machine learning based transformer oil pollution degree assessment method according to claim 7, characterized in that, In step S6, the feature vector X of each particle is input into the XGBoost classification model. Several CART decision trees within the XGBoost classification model perform feature condition judgments, including whether the thickness is greater than the threshold, whether the roundness is higher than the threshold, and whether the aspect ratio exceeds the limit. The leaf nodes of each tree output the corresponding weight values, and finally the weighted summation module obtains the prediction scores for different particle categories. The final category label of the particle is obtained by softmax or maximum value selection, including metal particles, cellulose particles, carbon particles, and bubbles.

9. The method for assessing transformer oil contamination based on lensless imaging and machine learning according to claim 8, characterized in that, In step S6, XGBoost employs a gradient boosting strategy: the first tree performs initial classification using six types of features; the second tree learns the classification error of the first tree; the third tree continues to learn the feature patterns that the first two trees failed to handle until the loss function converges; the objective function... (10); among which, The penalty coefficient represents the number of leaf nodes; T is the number of leaf nodes in the current tree. This represents the output weight of the j-th leaf node; It is the L2 regularization coefficient.

10. The method for assessing transformer oil contamination based on lensless imaging and machine learning according to claim 9, characterized in that, In step S6, the outputs of all trees are summed to obtain the scores for each category: (11); among which, This represents the final prediction result for the i-th sample; This represents the output of the k-th decision tree; K represents the total number of trees in the model. Indicates the first The structured feature vector corresponding to each particle sample.

Images