Plastic package material defect detection method and system in radiation environment based on machine vision

Abstract

The invention relates to the technical field of image processing, and discloses a plastic package material defect detection method and system in a radiation environment based on machine vision. The method comprises the following steps: acquiring an anti-radiation image in a gamma-ray environment through a lead-tungsten alloy shielding high-radiation tolerance type image sensor, and performing adaptive contrast enhancement. Defect features including microcracks, bubbles, deformation and chromatic aberration are extracted by using dual-tree complex wavelet transform and an HSV color space; and inputting the multi-dimensional features into an AR-DCNN network to carry out multi-task learning, identifying defect types and carrying out scoring. Based on Weibull distribution and a random forest algorithm, the radiation-resistant life of the plastic package material is predicted, and a material degradation curve and safety threshold early warning are obtained. According to the invention, standardized processing of the radiation damage quantitative index is realized, and the prediction accuracy is effectively improved.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and particularly to a method and system for detecting defects in plastic encapsulation materials under a radiation environment based on machine vision. Background Art

[0002] In modern aerospace and electronics fields, plastic encapsulation materials, as an important part of encapsulation materials, are widely used in the protection and encapsulation of electronic components, integrated circuits and other devices. Especially in high-radiation environments such as spacecraft and satellites, these materials are subjected to long-term radiation damage. Existing technologies usually rely on traditional testing methods, such as manual inspection, physical tests, etc., to evaluate the performance and radiation damage of plastic encapsulation materials. These methods often require a large amount of time and labor costs, and it is difficult to obtain the damage information of materials in complex radiation environments in real time and accurately. In addition, the traditional methods have limited processing capabilities for large-scale data and cannot effectively mine the potential relationship between damage and material properties, resulting in insufficient accuracy and efficiency.

[0003] In the analysis of radiation damage quantification indicators by existing technical methods, there is a lack of efficient data processing and intelligent evaluation means. The traditional test data processing methods mostly rely on manual experience and cannot uniformly and standardize the processing of various different radiation damage indicators. In addition, existing methods usually rely on a single data source, resulting in incomplete data dimensions and difficulty in adapting to complex radiation environment changes. Therefore, existing technologies lack effective quantitative analysis tools for accurately evaluating the electrical performance, mechanical strength, and thermal stability of plastic encapsulation materials. Especially in large-scale applications, their limitations are more obvious and cannot meet the high requirements of modern spacecraft and electronic devices for predicting the performance of plastic encapsulation materials. Summary of the Invention

[0004] This application provides a method and system for detecting defects in plastic encapsulation materials under a radiation environment based on machine vision, which is used to realize the standardized processing of radiation damage quantification indicators and effectively improve the prediction accuracy.

[0005] In a first aspect, the present application provides a method for detecting defects in plastic encapsulation materials in a radiation environment based on machine vision. The method for detecting defects in plastic encapsulation materials in a radiation environment based on machine vision includes: collecting images of the surface of the plastic encapsulation material of the spacecraft electronic component under a γ-ray environment through a lead tungstate alloy-shielded high-radiation-tolerant image sensor and performing adaptive contrast enhancement processing to obtain a radiation-resistant enhanced image; performing dual-tree complex wavelet transform feature decomposition and HSV color space defect feature extraction on the radiation-resistant enhanced image to obtain multi-dimensional feature representation data including micro-crack, bubble, deformation, and color difference features; inputting the multi-dimensional feature representation data into an AR-DCNN network with residual connection and attention mechanism for dilated convolution multi-task learning to obtain defect type recognition and severity scoring results; predicting the radiation resistance life of the plastic encapsulation materials for nuclear facilities based on the defect type recognition and severity scoring results through Weibull distribution fitting and random forest algorithm to obtain a material performance degradation curve and a safety threshold warning.

[0006] In a second aspect, the present application provides a system for detecting defects in plastic encapsulation materials in a radiation environment based on machine vision. The system for detecting defects in plastic encapsulation materials in a radiation environment based on machine vision includes:

[0007] An enhancement module for collecting images of the surface of the plastic encapsulation material of the spacecraft electronic component under a γ-ray environment through a lead tungstate alloy-shielded high-radiation-tolerant image sensor and performing adaptive contrast enhancement processing to obtain a radiation-resistant enhanced image;

[0008] An extraction module for performing dual-tree complex wavelet transform feature decomposition and HSV color space defect feature extraction on the radiation-resistant enhanced image to obtain multi-dimensional feature representation data including micro-crack, bubble, deformation, and color difference features;

[0009] An input module for inputting the multi-dimensional feature representation data into an AR-DCNN network with residual connection and attention mechanism for dilated convolution multi-task learning to obtain defect type recognition and severity scoring results;

[0010] A prediction module for predicting the radiation resistance life of the plastic encapsulation materials for nuclear facilities based on the defect type recognition and severity scoring results through Weibull distribution fitting and random forest algorithm to obtain a material performance degradation curve and a safety threshold warning.

[0011] In a third aspect, there is provided a device for detecting defects in plastic encapsulation materials in a radiation environment based on machine vision, including: a memory and at least one processor, wherein instructions are stored in the memory; the at least one processor calls the instructions in the memory so that the device for detecting defects in plastic encapsulation materials in a radiation environment based on machine vision executes the above-mentioned method for detecting defects in plastic encapsulation materials in a radiation environment based on machine vision.

[0012] In a fourth aspect, a computer-readable storage medium is provided, in which instructions are stored, and when they run on a computer, the computer is made to execute the above-mentioned method for detecting defects in encapsulation materials in a radiation environment based on machine vision.

[0013] In the technical solution provided by this application, in terms of the standardization processing of radiation damage quantification indicators, through feature normalization processing, this solution enables various indicators to be compared within a unified numerical range, providing a reliable data basis for subsequent intelligent prediction. The standardized damage index vector can eliminate the influence between different dimensions, making the data have better comparability and consistency, and providing effective input data for the training of machine learning models. Secondly, a training set is constructed based on the samples in the historical experimental database. By matching the standardized damage index vector with the samples, a comprehensive prediction of the material properties is realized. The use of the random forest algorithm for ensemble learning through multiple decision trees can effectively avoid overfitting under diverse data inputs, improving the accuracy and robustness of the prediction. The advantage of the random forest algorithm lies in its ability to handle high-dimensional data and its strong modeling ability for complex relationships between features. Therefore, it can more accurately predict multiple performance indicators such as the electrical properties, mechanical strength, and thermal stability of encapsulation materials. Especially in the prediction of electrical properties, the ensemble learning method based on the decision tree ensemble can use the average prediction value of all decision trees as the final result, thereby improving the prediction accuracy of insulation strength and volume resistivity, which is crucial for the long-term reliability of spacecraft and electronic components in high-radiation environments. Similarly, in the prediction of mechanical properties and thermal stability, the decision tree model can also effectively combine radiation damage quantification indicators to provide accurate predictions for mechanical properties such as tensile strength, bending strength, and hardness, as well as thermal stability indicators such as heat distortion temperature and thermal expansion coefficient. These indicators play a decisive role in the heat resistance reliability of spacecraft electronic components in a radiation environment. Accurate prediction can provide theoretical support and data basis for the optimization and application of materials. The application of artificial intelligence algorithms, especially in the random forest model, in this solution greatly improves the efficiency of data processing and model prediction. Compared with traditional manual testing and calculation methods, the random forest can quickly extract potential rules and patterns from complex data through parallel processing of multiple decision trees, avoiding the subjectivity and errors of manual evaluation. More importantly, as the number of samples increases, the random forest can gradually improve the prediction accuracy without much manual intervention, with a high level of automation and intelligence. This feature is particularly important for spacecraft electronic components and related fields that require large-scale data processing, as it can quickly respond to new experimental data and achieve real-time and accurate damage assessment. Description of the Drawings

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0015] Figure 1 It is a schematic diagram of an embodiment of the method for detecting defects in plastic encapsulation materials under a radiation environment based on machine vision in the embodiments of the present application;

[0016] Figure 2 It is a schematic diagram of an embodiment of the system for detecting defects in plastic encapsulation materials under a radiation environment based on machine vision in the embodiments of the present application;

[0017] Figure 3 It is a structural schematic block diagram of the device for detecting defects in plastic encapsulation materials under a radiation environment based on machine vision in the embodiments of the present invention. Specific embodiments

[0018] The embodiments of the present application provide a method for detecting defects in plastic encapsulation materials under a radiation environment based on machine vision. The terms "first", "second", "third", "fourth", etc. (if any) in the specification, claims and above-mentioned drawings of the present application are used to distinguish similar objects and do not necessarily describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments described here can be implemented in an order other than those illustrated or described here. In addition, the terms "comprising" or "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0019] For ease of understanding, the following describes the specific process of the embodiments of the present application. Please refer to Figure 1 An embodiment of the method for detecting defects in plastic encapsulation materials under a radiation environment based on machine vision in the embodiments of the present application includes:

[0020] Step S101: Collect images of the surface of the plastic encapsulation material of the spacecraft electronic component under a γ-ray environment through a lead-tungsten alloy shielded high-radiation-tolerant image sensor and perform adaptive contrast enhancement processing to obtain a radiation-resistant enhanced image;

[0021] Step S102: Perform dual-tree complex wavelet transform feature decomposition and HSV color space defect feature extraction on the radiation-resistant enhanced image to obtain multi-dimensional feature representation data including micro-crack, bubble, deformation, and color difference features;

[0022] Step S103: Input the multi-dimensional feature representation data into the AR-DCNN network with residual connection and attention mechanism for dilated convolution multi-task learning to obtain the defect type recognition and severity score results;

[0023] Step S104: Based on the defect type recognition and severity score results, perform the radiation resistance life prediction of the plastic encapsulation materials used in nuclear facilities through Weibull distribution fitting and random forest algorithm to obtain the material performance degradation curve and safety threshold warning.

[0024] It can be understood that the execution subject of this application can be a plastic encapsulation material defect detection system based on machine vision, or a terminal or a server, and specific limitations are not made here. In this embodiment of the application, the server is used as the execution subject for illustration.

[0025] In this embodiment of the application, image acquisition and enhancement processing are performed on the surface of the plastic encapsulation materials of spacecraft electronic components. The plastic encapsulation materials of spacecraft electronic components are prone to defects such as microcracks in a radiation environment, and these defects are difficult to detect by conventional methods. This method uses a lead-tungsten alloy shielded high-radiation-tolerant image sensor. A lead-tungsten alloy composite shielding layer with a specified thickness is set outside the sensor to form a high-shielding-rate anti-radiation sensor, and the internal circuit adopts an anti-radiation reinforcement scheme with multiple redundant designs. In a γ-ray radiation environment, the parameter automatic correction unit monitors the radiation exposure level of the sensor in real time. When the radiation intensity exceeds the preset threshold, the camera sensitivity parameter and exposure time are automatically adjusted. The collected original image is decomposed into multiple frequency sub-bands through wavelet decomposition, and an adaptive threshold related to the current radiation intensity is applied to the high-frequency sub-bands for denoising. Subsequently, an adaptive contrast enhancement algorithm with a fixed pixel window is applied to the denoised image for local enhancement, and the enhancement coefficient is dynamically determined according to the local entropy value of the image. The interference types in the preliminarily enhanced image are identified through pattern matching, and the corresponding compensation strategy is applied for interference correction to finally obtain an anti-radiation enhanced image.

[0026] Feature extraction is performed on the anti-radiation enhanced image. First, the geometric deformation correction algorithm is applied, and the image distortion caused by radiation is eliminated through the bicubic B-spline transformation model with multiple control points. Subsequently, the dual-tree complex wavelet transform is applied to the corrected image, and the image is decomposed into multiple decomposition scales, each scale containing multiple directional sub-bands, to obtain multi-scale and multi-directional feature sub-bands. The dual-tree complex wavelet transform, as a unique method for analyzing the surface texture of encapsulation materials in the nuclear radiation environment, has a significant detection effect on radiation microcracks. Local second-order matrix features, Gabor texture features, and morphological features are extracted from these feature sub-bands respectively to obtain a structural feature vector. At the same time, the corrected image is converted to the HSV color space, and the statistical features of the hue, saturation, and value channels are extracted to obtain a color feature vector. This process quantifies the characteristic discoloration of the encapsulation materials caused by γ-ray radiation. The structural feature vector and the color feature vector are concatenated to form an initial feature vector, which is reduced in dimension and enhanced in separability through principal component analysis and linear discriminant analysis, and then the difference between the defect area and the normal area is enhanced through a non-linear mapping function to construct a multi-scale feature pyramid, and finally a multi-dimensional feature representation data containing microcracks, bubbles, deformation, and color difference features is obtained. The multi-dimensional feature representation is input into a specially designed deep learning network for processing. An AR-DCNN (Anti-Radiation Deep Convolutional Neural Network) network with an encoder-decoder structure is constructed. The encoder contains multiple encoding blocks, and each encoding block is set with three parallel multi-scale convolutional units. After the encoder, a radiation feature enhancement module is connected, and the multi-scale feature map is processed through a residual connection block and a channel attention mechanism to obtain an enhanced feature representation. The channel attention mechanism is specifically used to enhance the channel response of the encapsulation material features of spacecraft electronic components. The enhanced feature representation is processed through multiple dilated convolutional layers of the decoder to form a multi-scale receptive field, and upsampling is performed through transposed convolution between the decoding blocks and fused with the feature map of the corresponding encoding layer. The fused feature map is input into a multi-task learning head, including three parallel branches: a classification head, a segmentation head, and a scoring head. The classification head outputs the probability distribution of the defect type through a fully connected layer, the segmentation head generates a defect area segmentation mask through a 1×1 convolution operation, and the scoring head outputs the defect severity score value through a fully connected layer after global average pooling. The model is optimized through a combined loss function, and finally the defect type recognition and severity score results are output.

[0027] Construct a density-weighted Voronoi diagram to associate the detected defects with the surrounding areas and obtain a radiation damage pattern map. Apply the kernel density estimation algorithm to calculate the damage distribution density function on the material surface and obtain the damage distribution characteristics of the plastic encapsulated material surface. Support vector regression is used to analyze the correlation between different types of defects and radiation exposure, distinguish the defects caused by radiation from those caused by non-radiation factors, and obtain the radiation damage quantification index. After normalizing these quantification indexes, input them into the random forest algorithm to construct a performance degradation evaluation model containing multiple decision trees, and output the predicted values of key performance parameters such as electrical performance, mechanical strength, and thermal stability. Apply the Weibull distribution to fit these predicted values, establish the functional relationship between the performance degradation rate and time, and obtain the material performance degradation curve. The Weibull distribution is particularly suitable for describing the failure mode of plastic encapsulated materials in a γ-ray radiation environment. Determine the predicted time when the performance drops to a specified percentage of the initial value according to the performance degradation curve, set multiple safety thresholds, and finally obtain the early warning information of the safety threshold for the use of plastic encapsulated materials in nuclear facilities.

[0028] Taking the plastic encapsulated material of a spacecraft communication component as an example, when the cumulative dose of this material reaches 500 Gy in a γ-ray radiation environment, images are collected by a highly radiation-tolerant image sensor. After adaptive contrast enhancement processing, the dual-tree complex wavelet transform is decomposed into 6 scales, and the HSV color space features are extracted. It is found that the mean value of the H channel deviates from the initial value by 0.15, indicating that the material has obvious color change. After the multi-dimensional features are input into the AR-DCNN network, two microcracks (type confidence 0.92) and one bubble (type confidence 0.87) are identified, and the severity score is 7.5. Through the density-weighted Voronoi diagram and kernel density estimation, it is determined that the damage area is mainly concentrated on the edge of the component. Support vector regression analysis shows that the correlation coefficient between the crack and radiation is 0.85, which belongs to the typical damage caused by radiation. The random forest algorithm predicts that the insulation strength drops to 81% of the initial value, and the heat distortion temperature decreases by 12°C. The Weibull distribution fitting determines that the material performance degradation rate follows specific shape parameters and scale parameters. It is predicted that when the radiation dose accumulates by another 300 Gy, the material performance will drop below the safety threshold, and the system issues an early warning message, suggesting replacing the plastic encapsulated material of this component before the next maintenance cycle.

[0029] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0030] Set a lead-tungsten alloy composite shielding layer with a specified thickness outside the image sensor to obtain a radiation-resistant sensor with a high shielding rate, and implement a radiation-hardening solution with multiple redundancies for the internal circuit of the radiation-resistant sensor to obtain a circuit system that maintains normal functions in a high-energy ray environment;

[0031] The parameter automatic correction unit monitors the radiation exposure level of the sensor in real time. When the radiation intensity exceeds the preset threshold, it automatically adjusts the sensitivity parameter and exposure time of the camera to obtain the image acquisition parameters suitable for the current radiation environment;

[0032] Collect the surface image of the plastic packaging material according to the image acquisition parameters to obtain the material image in the original radiation environment;

[0033] Decompose the material image in the original radiation environment into multiple frequency sub-bands by wavelet decomposition, and apply an adaptive threshold related to the current radiation intensity to the high-frequency sub-bands for denoising to obtain a denoised image;

[0034] Apply an adaptive contrast enhancement algorithm with a fixed pixel window size and a set overlap rate to the denoised image for local enhancement. The enhancement coefficient is dynamically determined according to the local entropy value of the image to obtain a preliminary enhanced image;

[0035] According to the scattering, drift, and thermal noise radiation interference feature patterns included in the interference feature library, identify the interference type in the preliminary enhanced image through pattern matching to obtain the image interference features;

[0036] Apply the corresponding compensation strategy to the preliminary enhanced image according to the image interference features to correct the interference and obtain the final anti-radiation enhanced image. At the same time, record the current radiation type, radiation intensity, and cumulative dose to form radiation environment parameter data.

[0037] Specifically, setting a lead-tungsten alloy composite shielding layer outside the image sensor is a key step in realizing a radiation-resistant sensor. The lead-tungsten alloy composite shielding layer is composed of two materials. Lead has good gamma-ray shielding ability, while tungsten has excellent shielding effect on neutron rays. In the nuclear facility environment, the shielding layer thickness is determined according to the radiation intensity. For the gamma-ray radiation environment, the high-shielding-rate radiation-resistant sensor formed by the lead-tungsten alloy composite shielding layer can effectively isolate more than 90% of the radiation sources, ensuring that the sensor chip is protected from radiation damage. Implementing a radiation-hardening scheme with multiple redundant designs for the internal circuit of the radiation-resistant sensor is the key for the sensor to maintain normal functions in a high-energy ray environment. The multiple redundant design adopts a triple-module redundant architecture, that is, the same function is executed in parallel by three independent circuit modules, and the majority result is selected as the output through a voting mechanism, thus eliminating the error codes caused by single-event upsets. This circuit system also uses radiation-hardened electronic components, including transistors with special doping processes and integrated circuits with thickened gate oxide layers. The parameter automatic correction unit monitors the radiation exposure level of the sensor in real time through a built-in radiation detection module. The radiation detection module consists of a PIN diode array. When the radiation intensity exceeds the preset threshold, it automatically adjusts the sensitivity parameter and exposure time of the camera. During the adjustment process, the sensitivity parameter increases or decreases according to the logarithmic ratio of the radiation intensity, while the exposure time is adjusted according to the inverse square root relationship of the radiation intensity. The combination of the two forms image acquisition parameters suitable for the current radiation environment, ensuring that sufficiently clear images can be obtained under different radiation intensities.

[0038] Acquiring the surface image of the plastic packaging material according to the image acquisition parameters is the process of obtaining the material image under the original radiation environment. During acquisition, the camera maintains a fixed distance from the surface of the plastic packaging material, and the illumination angle is set to 45 degrees to avoid strong reflections. The acquired original image is a 16-bit grayscale image with a resolution of 2048×1536 pixels, and the pixel depth contains sufficient information for subsequent detection of subtle defects.

[0039] The material image under the original radiation environment usually contains a large amount of radiation noise and needs to be denoised through wavelet decomposition. Wavelet decomposition decomposes the image into multiple frequency sub-bands, including a low-frequency sub-band and three high-frequency sub-bands (horizontal, vertical, and diagonal directions). For each decomposition level, the low-frequency sub-band is decomposed again to form a multi-level wavelet decomposition tree. An adaptive threshold related to the current radiation intensity is applied to the high-frequency sub-bands for denoising. The threshold calculation formula is: Threshold value = Base threshold + Radiation intensity coefficient × Current radiation intensity. Soft threshold processing is performed on the wavelet coefficients exceeding the threshold, that is, the coefficients exceeding the threshold are reduced by the threshold value, retaining the continuity of the signal. After wavelet inverse transformation, the denoised image is obtained, which effectively removes the high-frequency noise introduced by the radiation environment.

[0040] Applying an adaptive contrast enhancement algorithm to the denoised image is a key step in enhancing the defect features of the plastic-sealed material. This algorithm processes the image block by block using a fixed pixel window size (usually 16×16 pixels) and a sliding window with a set overlap rate of 50%. The enhancement coefficient is dynamically determined based on the local entropy value of the image, and the calculation of the local entropy value takes into account the probability distribution of the pixel gray values within the window. For regions with high entropy values (indicating a large amount of information and possibly containing defects), a larger enhancement coefficient is used; for regions with low entropy values (indicating a small amount of information and possibly being the background), a smaller enhancement coefficient is used. After processing, a preliminary enhanced image is formed, and the defect features in this image are effectively highlighted.

[0041] The interference feature library is a unique image processing resource in a radiation environment, containing typical radiation interference feature patterns such as scattering, drift, and thermal noise. Scattering appears as star-like noise in the image, drift appears as a gradual change in the overall brightness of the image, and thermal noise appears as randomly distributed bright spots. By calculating the similarity between the preliminary enhanced image and each pattern in the interference feature library through template matching, the main interference types in the current image are identified, and an image interference feature description is formed. Applying corresponding compensation strategies according to the image interference features is the key to finally obtaining a high-quality anti-radiation enhanced image. For scattering interference, median filtering is applied to remove point noise; for drift interference, bilateral filtering is applied to smooth the brightness gradient while preserving the edges; for thermal noise, morphological opening operation is applied to remove random bright spots. After the compensation processing is completed, the final anti-radiation enhanced image is obtained, and at the same time, the current radiation type (such as γ-ray, X-ray, etc.), radiation intensity (Gy / h), and cumulative dose (Gy) are recorded to form radiation environment parameter data, which are transmitted to the subsequent processing module together with the image.

[0042] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0043] Applying a geometric deformation correction algorithm to the anti-radiation enhanced image, and eliminating the image distortion caused by radiation through a bicubic B-spline transformation model with multiple control points to obtain a corrected image with sub-pixel accuracy;

[0044] Applying a dual-tree complex wavelet transform to the corrected image, decomposing the image into multiple decomposition scales, and each scale contains multiple directional sub-bands to obtain multi-scale multi-directional feature sub-bands. Among them, the dual-tree complex wavelet transform is a unique method for analyzing the surface texture of plastic-sealed materials in a nuclear radiation environment and has a significant detection effect on radiation micro-cracks;

[0045] Extracting local second-order matrix features, Gabor texture features, and morphological features from the multi-scale multi-directional feature sub-bands respectively to obtain the structural feature vector of the spacecraft electronic component packaging material;

[0046] Convert the corrected image to the HSV color space, and extract the statistical features of the hue, saturation, and value channels respectively, including mean, standard deviation, skewness, kurtosis, energy, and entropy, to obtain a color feature vector. Among them, the HSV color space feature extraction quantifies the characteristic discoloration of the plastic encapsulation material caused by gamma-ray radiation;

[0047] Concatenate the structural feature vector and the color feature vector to form an initial feature vector, and reduce the feature dimension to a lower dimension through principal component analysis, retaining the main information content to obtain a dimensionality-reduced feature vector;

[0048] Further reduce the dimensionality of the dimensionality-reduced feature vector through linear discriminant analysis, while maximizing the separability between different defect categories to obtain an optimized feature vector;

[0049] Apply the defect feature enhancement algorithm to the optimized feature vector, enhance the difference between the defect area and the normal area in the feature space through a non-linear mapping function, and construct a multi-scale feature pyramid to obtain multi-dimensional feature representation data containing typical micro-cracks, bubbles, deformations, and color difference features in the nuclear facility environment.

[0050] Specifically, apply the geometric deformation correction algorithm to eliminate the image distortion caused by radiation. In a high-radiation environment, images often produce radial distortion and tangential distortion, manifested as straight lines becoming curved or distorted. The geometric deformation correction algorithm is corrected through a bicubic B-spline transformation model with multiple control points. The specific operation is to identify feature points in the image, establish a correspondence between these feature points and the ideal positions, and construct a transformation matrix. The bicubic B-spline transformation model uses 16 control points to affect the mapping relationship of each pixel, ensuring smooth and continuous transformation. The correction process iteratively minimizes the Euclidean distance between the actual positions and the ideal positions of the control points until sub-pixel accuracy (error less than 0.5 pixel) is achieved. The geometric shape of the corrected image after transformation is restored to a radiation-free interference state, laying a foundation for subsequent feature extraction. Applying the dual-tree complex wavelet transform to the corrected image is a key step in extracting multi-scale and multi-directional features. The dual-tree complex wavelet transform is a method for analyzing the surface texture of plastic encapsulation materials unique to the nuclear radiation environment. Compared with the traditional wavelet transform, it increases direction selectivity and is particularly suitable for detecting radiation micro-cracks. This transform uses two sets of orthogonal wavelet filters, one set of filters is responsible for the real part, and the other set is responsible for the imaginary part, jointly constituting a complex wavelet. In actual calculations, the image is decomposed at multiple levels. Each level of decomposition produces a low-frequency sub-band and high-frequency sub-bands, and each high-frequency sub-band is further divided into multiple directions (usually 6 directions: ±15°, ±45°, ±75°), thus obtaining multi-scale and multi-directional feature sub-bands. For the plastic encapsulation materials in the nuclear facility environment, the dual-tree complex wavelet transform has a significant detection effect on micro-cracks because it can capture edge and texture features in different directions, especially those fine linear cracks caused by radiation.

[0051] Extracting features from multi-scale and multi-directional feature sub-bands is an important step in describing the surface structure of materials. First, local second-order matrix features are extracted. By calculating the covariance matrix of the local area of each feature sub-band, its eigenvalues and eigenvectors are extracted to describe the directionality and anisotropy of the region. Gabor texture feature extraction uses Gabor filter banks with different scales and directions to convolve the image, capturing texture information of different frequencies and directions, which is particularly suitable for describing the periodic texture and irregular micro-cracks on the surface of plastic encapsulated materials. Morphological feature extraction calculates topological properties such as regional connectivity and the number of branch points by applying morphological operations such as opening and closing operations and skeleton extraction to the feature sub-bands, effectively describing the shape features of defects such as bubbles and deformations. The combination of these three types of features forms a structural feature vector, comprehensively characterizing the surface structure characteristics of spacecraft electronic component packaging materials. The corrected image is converted from the RGB color space to the HSV color space for feature extraction. The HSV color space decomposes colors into three channels: hue, saturation, and value. This decomposition method is closer to the way humans perceive colors and is insensitive to light changes. For the characteristic color change of plastic encapsulated materials caused by γ-ray radiation, the HSV color space is particularly suitable for quantitative analysis. In the specific extraction process, statistical features are calculated from the three channels respectively: the mean reflects the average level of the channel; the standard deviation describes the degree of dispersion of the color distribution; the skewness measures the asymmetry of the color distribution; the kurtosis represents the sharpness of the distribution; and the energy and entropy describe the concentration and chaos of the color distribution respectively. These statistics form a color feature vector, precisely quantifying the color change characteristics of the material caused by radiation. After concatenating the structural feature vector and the color feature vector to form an initial feature vector, the feature dimension is usually high, containing a large amount of redundant information. Dimensionality reduction is performed through principal component analysis (PCA), retaining the main information content. PCA calculates the covariance matrix of the feature vectors, solves its eigenvalues and eigenvectors, and selects the first few eigenvectors corresponding to the largest eigenvalues as the projection basis, retaining the main variation information of the data while reducing the dimension, and projecting the high-dimensional feature vector into a low-dimensional space to obtain a reduced-dimensional feature vector.

[0052] The linear discriminant analysis (LDA) is further applied to the dimensionality-reduced feature vectors. Different from PCA, LDA is a supervised dimensionality reduction method that takes into account class information. By maximizing the ratio of the between-class scatter matrix to the within-class scatter matrix, it finds the projection direction that can best distinguish different classes. In the detection of plastic package material defects, LDA maximizes the separability of different types of defects (microcracks, air bubbles, deformation, color difference) in the feature space, thereby obtaining optimized feature vectors. The defect feature enhancement algorithm is applied to the optimized feature vectors to enhance the difference between the defect region and the normal region in the feature space through a non-linear mapping function. The non-linear mapping function usually adopts the Sigmoid form, stretching the samples near the boundary of the feature space to increase the distance between the defect region and the normal region in the feature space. On this basis, a multi-scale feature pyramid is constructed to retain the main features at different scales to capture defect information of different sizes, and finally a multi-dimensional feature representation data containing typical microcrack, air bubble, deformation, and color difference features in the nuclear facility environment is obtained.

[0053] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0054] Construct an AR-DCNN network with an encoder-decoder structure. The encoder contains multiple encoding blocks, and each encoding block is set with three parallel multi-scale convolutional units to obtain multi-scale feature maps;

[0055] A radiation feature enhancement module is connected after the encoder. The multi-scale feature maps are processed through residual connection blocks and a channel attention mechanism to obtain enhanced feature representations. Among them, the channel attention mechanism is used to enhance the channel response of the plastic package material features of spacecraft electronic components;

[0056] The enhanced feature representations are processed through multiple dilated convolutional layers of the decoder to form multi-scale receptive fields, and the feature parsing results at different scales are obtained. Upsampling is performed through transposed convolution between the decoding blocks and fused with the feature maps of the corresponding encoding layers to obtain fused feature maps;

[0057] The fused feature maps are input into a multi-task learning head, including three parallel branches: a classification head, a segmentation head, and a scoring head, to obtain an initial defect prediction result;

[0058] The initial defect prediction result is optimized by applying a combined loss function. The combined loss function includes the weighted cross-entropy loss for defect classification, the Dice loss for segmentation, and the mean squared error loss for scoring, to obtain an optimized defect detection model;

[0059] The optimized defect detection model is used to process newly input plastic package material images in the nuclear facility environment, and the defect type recognition and severity scoring results are output.

[0060] Specifically, based on the obtained multi-dimensional feature representation data, an AR-DCNN (Anti-Radiation Deep Convolutional Neural Network) network with an encoder-decoder structure is constructed for defect detection and evaluation. The AR-DCNN network is designed specifically for defect detection of plastic encapsulated materials in a radiation environment and has special anti-radiation noise capabilities. The encoder part contains 5 encoding blocks, and each encoding block is set with three parallel multi-scale convolutional units, using convolutional kernel sizes of 3×3, 5×5, and 7×7 respectively. This parallel design enables the network to capture features of different scales simultaneously. After each convolutional operation, a batch normalization layer and a LeakyReLU activation function are connected. Batch normalization accelerates the training convergence process by normalizing the input of each layer, and LeakyReLU sets a slope of 0.2 on the negative half-axis to avoid the problem of neuron death. The number of convolutional kernels increases layer by layer from 64 in the first layer to 512 in the fifth layer. Max pooling is used for downsampling between layers to reduce the size of the feature map while retaining significant features, and finally a multi-scale feature map is obtained.

[0061] A radiation feature enhancement module is connected after the encoder, which is the key part that differentiates the AR-DCNN network from ordinary deep convolutional networks. This module contains a residual connection block and a channel attention mechanism. The residual connection block adopts a pre-activation design, that is, batch normalization and activation are performed first, and then convolutional operations are carried out, effectively alleviating the problem of gradient disappearance in deep networks. The channel attention mechanism is specifically used to enhance the channel response of the plastic encapsulated materials of spacecraft electronic components. Its working principle is to weight according to the importance of different feature channels. First, global average pooling and max pooling are performed on the input feature map to generate two channel descriptors respectively, then these two descriptors are processed by a shared multi-layer perceptron, and finally the processing results are added and passed through a Sigmoid activation function to obtain weights between 0 and 1. These weights are multiplied by each channel of the original feature map to enhance the response intensity of important feature channels, suppress irrelevant channels, thereby highlighting the unique defect features in a radiation environment and obtaining an enhanced feature representation.

[0062] The enhanced feature representation is processed through multiple dilated convolutional layers of the decoder. Dilated convolution, also known as atrous convolution, expands the receptive field by inserting holes in the convolutional kernel while keeping the number of parameters and computational complexity unchanged. The decoder contains 4 decoding blocks, and each decoding block uses 3×3 convolutional layers with different dilation rates (1, 2, 4, and 8 respectively) to form multi-scale receptive fields, which can detect defect features of different sizes simultaneously. Upsampling is performed between the decoding blocks through transposed convolution. The transposed convolution operation restores the low-resolution feature map to a high-resolution feature map through learned parameters, with a stride of 2, achieving 2-fold upsampling. The upsampled feature map is fused with the feature map of the corresponding encoding layer through skip connections. The fusion method uses weighted summation, and the weights are automatically learned through network training. The skip connections directly transfer the high-resolution detailed information in the encoding stage to the decoding stage, compensating for the lost spatial details during the upsampling process, and finally obtaining the fused feature map.

[0063] The fused feature map is input into the multi-task learning head, which is the output part of the AR-DCNN network and includes three parallel branches: a classification head, a segmentation head, and a scoring head. The classification head converts the feature vector into a probability distribution of defect types through a fully connected layer, and the output includes confidence scores for five categories: four main types of defects, namely microcracks, bubbles, deformation, and color difference, and the normal region. The segmentation head maps the feature map into a binary mask through 1×1 convolution to identify the exact location and shape of the defect region. The scoring head compresses the feature map into a one-dimensional vector through global average pooling and then maps it into a severity score value within the range of 0-10 through a fully connected layer. The outputs of these three branches together constitute the initial defect prediction result, providing comprehensive information on the type, location, and severity of the defects.

[0064] The initial defect prediction result is optimized using a combined loss function. The design of the combined loss function is the key to multi-task learning. The weighted cross-entropy loss is used for the defect classification task, and the weights are set as crack:bubble:deformation:color difference:normal = 3:2:2:2:1, reflecting the importance differences of different types of defects. The segmentation task uses a combination of Dice loss and weighted cross-entropy. Dice loss focuses on improving the overlap degree of the segmentation mask and is particularly effective in dealing with class imbalance problems. The scoring task uses mean squared error loss to directly measure the gap between the predicted score and the true score. The losses of the three tasks are weighted and summed according to the weight ratio classification:segmentation:scoring = 0.3:0.5:0.2 to form the final combined loss. The network parameters are updated through the backpropagation algorithm to optimize the loss function, and finally, an optimized defect detection model is obtained.

[0065] When processing a newly input image of plastic encapsulated material in the nuclear facility environment using the optimized defect detection model, first, the image is subjected to the same preprocessing steps as in the training stage. Then, the image is segmented into overlapping local regions by means of a sliding window and fed into the AR-DCNN network. Each window generates a set of prediction results, and the results of all windows are combined through an integrated decision-making strategy (such as weighted voting), and finally, the defect type recognition and severity score results are output.

[0066] In a specific embodiment, the process of performing the step of inputting the fused feature map into the multi-task learning head, including three parallel branches of a classification head, a segmentation head, and a scoring head, may specifically include the following steps:

[0067] Input the fused feature map into the classification head, perform feature transformation through a fully connected layer to obtain the probability distribution of the defect types of the plastic encapsulated material, and the defect types include microcracks, air bubbles, deformation, color difference, and normal regions that are unique to the radiation environment;

[0068] Apply the 1×1 convolution operation in the segmentation head to the fused feature map to perform pixel-level feature mapping and obtain a binary defect region segmentation mask;

[0069] Input the fused feature map into the scoring head after global average pooling, and obtain the defect severity score value through mapping by a fully connected layer;

[0070] Perform an argmax operation on the probability distribution of the defect types to extract the defect type label corresponding to the highest probability and obtain the defect type prediction result;

[0071] Apply morphological post-processing to the defect region segmentation mask, including opening and closing operations and area filtering, to obtain a refined defect region boundary;

[0072] Fuse the defect type prediction result, the refined defect region boundary, and the defect severity score value to form a complete initial defect prediction result for the plastic encapsulated material of the spacecraft electronic component.

[0073] Specifically, the fused feature map is input into the classification head for feature transformation. The classification head consists of multiple fully-connected layers, and each fully-connected layer contains a linear transformation and a non-linear activation function. Specifically, the fused feature map first compresses the spatial dimension through global average pooling and is converted into a one-dimensional feature vector. Then, it passes through two fully-connected layers. The first layer uses the ReLU activation function, and the second layer does not use an activation function and directly outputs the raw scores. Finally, the raw scores are converted into a normalized probability distribution through the Softmax function to obtain the probability distribution of the encapsulation material defect types. These defect types include microcracks, bubbles, deformations, color differences, and normal regions that are unique to the radiation environment. Each type corresponds to a probability value, and the sum of all probability values is 1. Applying the 1×1 convolution operation of the segmentation head to the fused feature map is a process of pixel-level feature mapping. The 1×1 convolution operation refers to a convolutional layer with a convolution kernel size of 1×1, which is used to change the number of channels and perform information interaction between channels without changing the spatial size of the feature map. In the segmentation head, the 1×1 convolution maps the number of channels of the fused feature map from the original multi-channels (such as 256 or 512) to the number of target categories (usually set to 2 in the radiation environment, representing the defect region and the non-defect region). The output at each pixel position is the probability value that the pixel belongs to the defect region, and the value range is from 0 to 1. The probability map is processed through a threshold (usually the threshold is set to 0.5), and the pixels greater than the threshold are marked as 1 (defect region), and the pixels less than the threshold are marked as 0 (non-defect region) to obtain a binary segmentation mask of the defect region.

[0074] Feeding the fused feature map into the scoring head after global average pooling is a process for evaluating the severity of the defect. Global average pooling is a special pooling operation that calculates the global average value for each feature channel, reducing the entire feature map to a one-dimensional vector, where each element represents the average activation intensity of a channel. This operation not only significantly reduces the number of parameters but also retains the global information in the feature map. The scoring head receives this one-dimensional vector and maps it to a scalar value representing the severity score of the defect through one or more fully-connected layers. The scoring range is usually set to 0 - 10, where 0 indicates no defect and 10 indicates the most severe defect.

[0075] Performing the argmax operation on the defect type probability distribution is a step to extract the final predicted category from the probability distribution. The argmax operation finds the index with the largest probability value in the probability distribution, that is, the most likely defect type. For example, if the probability distribution is [0.05, 0.82, 0.06, 0.04, 0.03], corresponding to the defect types [microcrack, bubble, deformation, color difference, normal], then the argmax operation returns the index 1, indicating that the predicted defect type is a bubble. This hard decision method converts the probability distribution into a clear defect type label to obtain the defect type prediction result.

[0076] Applying morphological post - processing to the defect region segmentation mask is a crucial step in improving the quality of the segmentation results. Morphological post - processing includes opening and closing operations and area filtering. The opening operation is a combined operation of erosion followed by dilation, which is used to remove small noise and burrs; the closing operation is a combined operation of dilation followed by erosion, which is used to fill small holes and connect disconnected regions. In actual processing, the closing operation is first applied to fill the small holes in the defect region, and then the opening operation is applied to remove the noise and burrs on the edges. Area filtering removes overly small connected regions, which are usually false positive results caused by noise. Through these morphological operations, a refined defect region boundary with smooth edges and coherent interior is obtained.

[0077] Fusing the defect type prediction results, the refined defect region boundary, and the defect severity score value is the final step in forming the complete defect prediction result. The fusion process first assigns a unique identifier to each detected defect region, and then associates the defect type label and severity score with the corresponding defect region. For the case where there are multiple separated defect regions, connected component analysis is needed to divide the segmentation mask into independent defect instances, each instance containing attribute information such as location, area, type, and severity. The final generated complete initial defect prediction result of the spacecraft electronic component encapsulation material contains all the key information of the defects, providing a basis for subsequent performance evaluation and life prediction.

[0078] Taking the plastic encapsulation material for aerospace electronic components as an example, after the material works in a high-radiation environment for a period of time, defect detection is required. The fused feature map with a size of 128×128×256 obtained from the AR-DCNN network is input into the classification head of the multi-task learning head. A 256-dimensional feature vector is obtained through global average pooling, and then a five-dimensional vector [-2.1, 4.3, 0.9, -0.5, -3.2] is obtained through two fully connected layers (the first layer is 256×128, and the second layer is 128×5), representing the original scores of five types of defects. After applying the Softmax function, a probability distribution [0.01, 0.87, 0.08, 0.03, 0.01] is obtained, indicating that there is an 87% probability of bubble-like defects in this area. At the same time, the segmentation head applies a 1×1 convolution to the fused feature map, maps 256 channels to 2 channels, and then obtains the defect probability map at each pixel position through the Sigmoid function. After thresholding (threshold 0.5), an initial binary mask is obtained, where the continuous white area represents the potential defect positions. A closing operation is applied using a structuring element with a radius of 2 to fill several small holes inside the defects; then an opening operation is applied using a structuring element with a radius of 1 to smooth the defect edges; finally, connected regions with an area less than 10 pixels are removed to obtain the refined defect region boundary. The scoring head receives the feature vector after global average pooling and maps it to a defect severity score value of 7.8 through a fully connected layer, indicating that this is a relatively serious defect. Through connected component analysis, the segmentation mask is divided into 3 independent defect instances, which are located at different positions of the material. After each instance is associated with its predicted type and score value, a complete initial defect prediction result is formed, successfully detecting and evaluating the defects caused by radiation on the plastic encapsulation material of spacecraft electronic components.

[0079] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0080] Construct a density-weighted Voronoi diagram based on the defect type recognition and severity score results, associate the detected defects with the surrounding areas, and obtain a radiation damage pattern map;

[0081] Apply the kernel density estimation algorithm to the radiation damage pattern map, use the Gaussian kernel function to calculate the damage distribution density function on the material surface, and obtain the damage distribution characteristics on the surface of the plastic encapsulation material;

[0082] According to the damage distribution characteristics on the surface of the plastic encapsulation material, apply support vector regression to analyze the correlation between different types of defects and radiation exposure, distinguish the defects caused by radiation from those caused by non-radiation factors, and obtain a radiation damage quantification index;

[0083] Input the radiation damage quantification index into the random forest algorithm to construct a performance degradation evaluation model for plastic encapsulation materials, and output the predicted values of the key performance parameters of plastic encapsulation materials for nuclear facilities, including electrical performance, mechanical strength, and thermal stability indicators;

[0084] Apply the Weibull distribution to fit the predicted values of the key performance parameters, establish the functional relationship between the performance degradation rate and time, and obtain the material performance degradation curve. Among them, the Weibull distribution is particularly suitable for describing the failure mode of plastic encapsulation materials in the γ-ray radiation environment;

[0085] Determine the predicted time when the performance drops to a specified percentage of the initial value according to the material performance degradation curve, set multi-level safety thresholds, and obtain the early warning information of the use safety threshold of plastic encapsulation materials for nuclear facilities.

[0086] Specifically, the density-weighted Voronoi diagram is a spatial segmentation method that divides the plane into multiple regions, and each region corresponds to a seed point (which is the detected defect point in this method). Different from the traditional Voronoi diagram, the density-weighted Voronoi diagram considers the weights of the seed points, and the region corresponding to the seed point with a larger weight is larger. In the actual construction process, first, each detected defect point is used as a seed point, and the weight calculation formula is W = S×(0.7A + 0.3G), where S is the defect severity score value (range 0 - 10), A is the normalized defect area (a value between 0 and 1, obtained by dividing the defect area by the maximum possible area), and G is the normalized gradient value (a value between 0 and 1, indicating the clarity of the defect edge). Use the Fortune algorithm to generate the Voronoi diagram. This algorithm is based on the scan line principle, scans the seed points on the plane from left to right, and gradually constructs the Voronoi boundary. For each pair of adjacent seed points, calculate the position of the Voronoi boundary according to their positions and weights to form a complete Voronoi diagram. The resulting density-weighted Voronoi diagram intuitively shows the distribution pattern of radiation damage on the material surface, and each defect point and its influence area form a radiation damage pattern map. Applying the kernel density estimation algorithm to the radiation damage pattern map is an effective method for quantifying the damage distribution density. Kernel density estimation is a non-parametric density estimation technique that places kernel functions on each data point and then sums all the kernel functions to obtain a smooth density distribution. In this method, a Gaussian kernel function is used for kernel density estimation, and the Gaussian kernel function has good smoothness and mathematical properties. For each position point on the material surface, calculate the damage density value at this position according to its distance from each defect point and the defect weight. The bandwidth parameter h of kernel density estimation is automatically determined by the Silverman rule: h = 0.9×min(σ, R / 1.34)×n (-1 / 5), where σ is the standard deviation of the seed points, R is the interquartile range, and n is the number of seed points. By performing kernel density estimation on the entire material surface, a continuous damage distribution density function is obtained, forming the surface damage distribution characteristics of the encapsulation material. This distribution characteristic can intuitively reflect which areas are severely damaged and which areas are less damaged, facilitating subsequent correlation analysis.

[0087] Applying support vector regression analysis based on the surface damage distribution characteristics of the encapsulation material to the correlation between different types of defects and radiation exposure is the key to identifying the causes of defects. Support vector regression is a machine learning method that finds a hyperplane that can minimize the prediction error by mapping data into a high-dimensional feature space. In this method, the input features include defect geometric features, texture features, and spatial distribution features (a total of 58 dimensions), and the output is the correlation coefficient between the defect and radiation exposure (between 0 and 1). Support vector regression uses an RBF kernel function. The kernel parameter γ determines the width of the kernel, the penalty coefficient C controls the balance between the model complexity and the fitting error, and ε defines the range of errors not to be included. By analyzing the relationship between defect features and historical radiation exposure data, the support vector regression model can learn the patterns of radiation-specific defects. Defects with a correlation coefficient greater than 0.7 are determined to be caused by radiation, defects less than 0.3 are determined to be caused by non-radiation factors, and defects between the two are caused by mixed factors. Through this analysis, radiation damage quantification indicators are obtained, including the proportion of the total area of radiation-induced defects, the average depth of radiation defects, and the spatial distribution characteristics of radiation defects.

[0088] Inputting the radiation damage quantification indicators into the random forest algorithm to construct an evaluation model for the performance degradation of the encapsulation material is the core step in predicting the material performance. Random forest is an ensemble learning method composed of multiple decision trees. Each tree is independently trained, and the average value or the majority voting result is taken during prediction. During the construction process, first, the feature normalization process is performed on the radiation damage quantification indicators to unify the numerical ranges of each indicator. Then, the standardized damage indicator vector is paired with the material performance parameters in the historical experimental data to form a training data set. When training the random forest, each decision tree is trained using a randomly sampled data subset (the sampling ratio is usually 70%) and a feature subset (the number of features is usually the square root of the total number of features), enhancing the generalization ability of the model. Each decision tree is constructed through recursive binary splitting, and the best splitting feature and threshold are selected at each node to maximize the purity of the child nodes. After training, for the newly input radiation damage quantification indicators, the final prediction result is obtained by averaging the prediction values of all decision trees, and the predicted values of the key performance parameters of the encapsulation material for nuclear facilities are output, including electrical performance (insulation strength, volume resistivity), mechanical strength (tensile strength, flexural strength, hardness), and thermal stability (heat distortion temperature, coefficient of thermal expansion) indicators.

[0089] Applying the Weibull distribution to fit the predicted values of key performance parameters is an important step in establishing a performance degradation model. The Weibull distribution is a continuous probability distribution widely used in reliability engineering and life analysis. Its mathematical form is flexible and can adapt to various failure data. The Weibull distribution is particularly suitable for describing the failure mode of plastic encapsulation materials in a gamma-ray radiation environment because the performance degradation of materials caused by radiation usually follows a non-linear change law. In the fitting process, first, the performance parameter values at different cumulative radiation doses are normalized to the ratio relative to the initial value, and then the maximum likelihood estimation method is used to determine the shape parameter and scale parameter of the Weibull distribution. The shape parameter describes the change trend of the failure rate over time, and the scale parameter is related to the characteristic life of the material. Through the Weibull distribution parameters obtained by fitting, a functional relationship between the performance degradation rate and the cumulative radiation dose (or time) is established, thereby obtaining the material performance degradation curve. For different performance parameters (such as insulation strength, mechanical strength, etc.), the corresponding degradation curves are established respectively to comprehensively evaluate the performance change of the material in the radiation environment.

[0090] Determining the predicted time when the performance drops to a specified percentage of the initial value according to the material performance degradation curve is the last step in generating warning information. The performance degradation curve describes the change law of material performance with the cumulative radiation dose. Through these curves, the performance value of the material at any future radiation dose can be predicted. For plastic encapsulation materials in nuclear facilities, multiple safety thresholds are usually set, such as 90%, 80%, and 70% of the initial value as the three-level warning thresholds. By solving the degradation curve equation, the cumulative radiation dose values when the performance reaches these thresholds are found, and then according to the radiation intensity (dose rate) of the facility, the corresponding time points are converted. These time points constitute the warning information for the safety threshold of the use of plastic encapsulation materials, guiding the nuclear facility operator to carry out material replacement or maintenance plans.

[0091] In a specific embodiment, the process of performing the step of inputting the radiation damage quantification index into the random forest algorithm to construct a performance degradation evaluation model for plastic encapsulation materials may specifically include the following steps:

[0092] Perform feature normalization processing on the radiation damage quantification index to unify the numerical ranges of each index and obtain a standardized damage index vector;

[0093] Match the standardized damage index vector with the samples in the historical experimental database to construct a training data set and obtain a sample set containing input features and output performance parameters;

[0094] Apply the random forest algorithm to the sample set to construct multiple decision trees. Each decision tree is trained using a randomly sampled subset of features to obtain a set of decision trees;

[0095] Perform ensemble learning on the decision tree ensemble. For each performance index of the encapsulation material of nuclear facilities, use the average predicted value of all decision trees as the final prediction result to obtain the predicted value of electrical performance, including insulation strength and volume resistivity;

[0096] Predict the mechanical strength index based on the decision tree ensemble, calculate the tensile strength, flexural strength and hardness parameters, and obtain the predicted value of mechanical performance;

[0097] Use the decision tree ensemble to predict the heat distortion temperature and coefficient of thermal expansion to obtain the predicted value of thermal stability. Among them, the predicted value of thermal stability is directly related to the heat resistance reliability of spacecraft electronic components in the radiation environment.

[0098] Specifically, the radiation damage quantification index needs to be subjected to feature normalization processing. Radiation damage usually manifests as different damages suffered by the encapsulation material in the radiation environment, and these damages can be quantified by several physical or chemical performance indexes, such as insulation strength, volume resistivity, heat distortion temperature, tensile strength, etc. However, the numerical ranges of these indexes vary greatly, and direct use will affect the training effect of subsequent machine learning models. Therefore, feature normalization processing must be carried out, that is, the data of each index is transformed into a unified numerical range so that the subsequent model can effectively learn the relationship between different features.

[0099] One of the common methods of normalization is to subtract the minimum value of each feature data and then divide it by the range (the difference between the maximum value and the minimum value) of the feature, so as to obtain a standardized data between 0 and 1. Through this process, all the damage index data will be within the same numerical range, thus avoiding the situation that the weights of some features are too large or too small in the subsequent machine learning model. The standardized damage index vector will be matched with the samples in the historical experiment database. Historical experimental data refers to the performance data of various encapsulation materials measured through past experiments under different radiation environments. These data include a series of physical parameters such as the insulation strength, volume resistivity, tensile strength, flexural strength of the material under different radiation environments, and these are used as the output parameters of the sample set.

[0100] By matching the standardized damage index with the samples in the historical experiment database, a training data set is constructed, which includes input features (standardized damage index) and output performance parameters (such as insulation strength, volume resistivity, etc.). This training set is used to train a machine learning model, especially the random forest algorithm. The random forest algorithm constructs multiple decision trees to perform classification or regression tasks. Each decision tree is constructed by randomly sampling the training data and selecting a subset of features. Through this method, the random forest can effectively avoid the overfitting problem and improve the generalization ability of the model.

[0101] The working principle of the random forest is to predict the input data through multiple decision trees, and then average or vote on the prediction results of each tree to obtain the final prediction result. For example, assume that multiple decision trees are trained on the training set, and each tree predicts the electrical performance value of the material (such as insulation strength, volume resistivity, etc.) based on the input vector of radiation damage standardized indicators. The final result is the average of the prediction results of all trees.

[0102] For the prediction of electrical performance, such as insulation strength and volume resistivity, the final prediction value can be obtained by taking the mean of the output results of each decision tree. In addition to electrical performance, the random forest can also be used to predict parameters such as mechanical properties and thermal stability. For the prediction of mechanical properties, including indicators such as tensile strength, flexural strength, and hardness, it is also predicted through the ensemble learning method of the decision tree set. Assume that for the tensile strength index, after training, the prediction values of multiple decision trees are obtained (such as the prediction values given by each tree are 500 MPa, 510 MPa, 495 MPa, etc.). Then the final predicted value of the tensile strength is also the average of these values. For thermal stability, the decision tree set is also used to predict the heat distortion temperature and coefficient of thermal expansion. Thermal stability directly affects the heat resistance reliability of electronic components in the plastic encapsulation materials in spacecraft, so this prediction is crucial. For example, if the prediction results of multiple decision trees for the heat distortion temperature are 300 °C, 305 °C, 290 °C, 295 °C, etc., the final prediction result is also the average of these values.

[0103] The above described the method for detecting defects in plastic encapsulation materials under a radiation environment based on machine vision in the embodiments of the present application. Next, the system for detecting defects in plastic encapsulation materials under a radiation environment based on machine vision in the embodiments of the present application will be described. Please refer to Figure 2 , an embodiment of the system for detecting defects in plastic encapsulation materials under a radiation environment based on machine vision in the embodiments of the present application includes:

[0104] An enhancement module, configured to perform image acquisition and adaptive contrast enhancement processing on the surface of the plastic encapsulation material of the spacecraft electronic component through a lead-tungsten alloy shielded high-radiation tolerance image sensor under a γ-ray environment to obtain an anti-radiation enhanced image;

[0105] An extraction module, configured to perform dual-tree complex wavelet transform feature decomposition and HSV color space defect feature extraction on the anti-radiation enhanced image to obtain multi-dimensional feature representation data including micro-crack, bubble, deformation, and color difference features;

[0106] An input module, configured to input the multi-dimensional feature representation data into an AR-DCNN network with residual connection and attention mechanism for dilated convolution multi-task learning to obtain a defect type recognition and severity scoring result;

[0107] A prediction module, which is used to predict the radiation resistance life of the plastic-sealed materials for nuclear facilities through Weibull distribution fitting and random forest algorithm based on the defect type identification and severity scoring results, and obtain the material performance degradation curve and safety threshold warning.

[0108] Through the collaborative cooperation of the above-mentioned various components, in terms of the standardized processing of radiation damage quantification indicators, this solution performs feature normalization processing, enabling various indicators to be compared within a unified numerical range, providing a reliable data basis for subsequent intelligent prediction. The standardized damage index vector can eliminate the influence between different dimensions, making the data have better comparability and consistency, and providing effective input data for the training of machine learning models. Secondly, a training set is constructed based on the samples in the historical experimental database. By matching the standardized damage index vector with the samples, a comprehensive prediction of the material performance is realized. The random forest algorithm performs ensemble learning through multiple decision trees. Under diverse data inputs, it can effectively avoid overfitting and improve the accuracy and robustness of prediction. The advantage of the random forest algorithm lies in its ability to handle high-dimensional data and its strong modeling ability for complex relationships between features. Therefore, it can more accurately predict multiple performance indicators such as the electrical properties, mechanical strength, and thermal stability of plastic-sealed materials. Especially in the prediction of electrical properties, the ensemble learning method based on the decision tree ensemble can use the average prediction value of all decision trees as the final result, thereby improving the prediction accuracy of insulation strength and volume resistivity, which is crucial for the long-term reliability of electronic components in spacecraft and high-radiation environments. Similarly, in the prediction of mechanical properties and thermal stability, the decision tree model can also effectively combine radiation damage quantification indicators to provide accurate predictions for mechanical properties such as tensile strength, bending strength, and hardness, as well as thermal stability indicators such as heat distortion temperature and thermal expansion coefficient. These indicators play a decisive role in the heat resistance reliability of spacecraft electronic components in a radiation environment. Accurate prediction can provide theoretical support and data basis for the optimization and application of materials. The application of artificial intelligence algorithms, especially in the random forest model, in this solution greatly improves the efficiency of data processing and model prediction. Compared with traditional manual testing and calculation methods, the random forest can quickly extract potential rules and patterns from complex data by processing multiple decision trees in parallel, avoiding the subjectivity and errors of manual evaluation. More importantly, as the number of samples increases, the random forest can gradually improve the prediction accuracy without too much manual intervention, with a high level of automation and intelligence. This feature is particularly important for spacecraft electronic components and related fields that require large-scale data processing, as it can quickly respond to new experimental data and achieve real-time and accurate damage assessment.

[0109] Above Figure 2The defect detection system of plastic encapsulation materials under a radiation environment based on machine vision in the embodiments of the present invention will be described in detail from the perspective of modular functional entities. Next, the defect detection device of plastic encapsulation materials under a radiation environment based on machine vision in the embodiments of the present invention will be described in detail from the perspective of hardware processing.

[0110] Figure 3 FIG. 4 is a schematic structural diagram of a defect detection device of plastic encapsulation materials under a radiation environment based on machine vision provided by an embodiment of the present invention. The defect detection device 300 of plastic encapsulation materials under a radiation environment based on machine vision may vary greatly due to different configurations or performances, and may include one or more processors (central processing units, CPUs) 310 (for example, one or more processors) and a memory 320, and one or more storage media 330 (for example, one or more mass storage device terminals) for storing application programs 333 or data 332. Among them, the memory 320 and the storage media 330 may be transient storage or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations on the defect detection device 300 of plastic encapsulation materials under a radiation environment based on machine vision. Further, the processor 310 may be configured to communicate with the storage media 330 and execute a series of instruction operations in the storage media 330 on the defect detection device 300 of plastic encapsulation materials under a radiation environment based on machine vision to implement the steps of the above-mentioned defect detection method of plastic encapsulation materials under a radiation environment based on machine vision.

[0111] The defect detection device 300 of plastic encapsulation materials under a radiation environment based on machine vision may further include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating systems 331, such as Windows Serve, Mac OS X, Unix, Linux, FreeBSD, and so on. Those skilled in the art can understand that Figure 3 the shown structural diagram of the defect detection device of plastic encapsulation materials under a radiation environment based on machine vision does not limit the defect detection device of plastic encapsulation materials under a radiation environment based on machine vision provided by the present invention, and may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0112] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions run on a computer, the computer is caused to execute the steps of the method for detecting defects in plastic encapsulation materials in a radiation environment based on machine vision.

[0113] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described systems, systems, and units can refer to the corresponding processes in the foregoing method embodiments and will not be described herein again.

[0114] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions to cause a device for detecting defects in plastic encapsulation materials in a radiation environment based on machine vision (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0115] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of various embodiments of the present invention.

Claims

1. A method for detecting defects in plastic encapsulation materials under a radiation environment based on machine vision, characterized in that, The method includes: Performing image acquisition and adaptive contrast enhancement processing on the surface of the plastic encapsulation material of spacecraft electronic components through a lead-tungsten alloy shielded high-radiation-tolerant image sensor under a γ-ray environment to obtain a radiation-resistant enhanced image; Performing dual-tree complex wavelet transform feature decomposition and HSV color space defect feature extraction on the radiation-resistant enhanced image to obtain multi-dimensional feature representation data including micro-crack, bubble, deformation, and color difference features; Inputting the multi-dimensional feature representation data into an AR-DCNN network with residual connection and attention mechanism for dilated convolution multi-task learning to obtain defect type recognition and severity scoring results; Based on the defect type recognition and severity scoring results, performing radiation-resistant life prediction of the plastic encapsulation material for nuclear facilities through Weibull distribution fitting and random forest algorithm to obtain a material performance degradation curve and a safety threshold warning; 2. The method for detecting defects of plastic encapsulation materials in a radiation environment based on machine vision according to claim 1, wherein The performing image acquisition and adaptive contrast enhancement processing on the surface of the plastic encapsulation material of spacecraft electronic components through a lead-tungsten alloy shielded high-radiation-tolerant image sensor under a γ-ray environment to obtain a radiation-resistant enhanced image includes: Setting a lead-tungsten alloy composite shielding layer with a specified thickness outside the image sensor to obtain a high-shielding-rate radiation-resistant sensor, and implementing a radiation-hardening scheme with multiple redundancy designs for the internal circuit of the radiation-resistant sensor to obtain a circuit system that maintains normal functions under a high-energy ray environment; Real-time monitoring the radiation exposure level of the sensor through a parameter automatic correction unit, and when the radiation intensity exceeds a preset threshold, automatically adjusting the sensitivity parameter and exposure time of the camera to obtain image acquisition parameters suitable for the current radiation environment; Collecting the surface image of the plastic encapsulation material according to the image acquisition parameters to obtain a material image under the original radiation environment; Decomposing the material image under the original radiation environment into multiple frequency sub-bands through wavelet decomposition, and applying an adaptive threshold denoising related to the current radiation intensity to the high-frequency sub-bands to obtain a denoised image; Applying an adaptive contrast enhancement algorithm with a fixed pixel window size and a set overlap rate to the denoised image for local enhancement, and dynamically determining the enhancement coefficient according to the local entropy value of the image to obtain a preliminary enhanced image; Identifying the interference type in the preliminary enhanced image through pattern matching according to the scattering, drift, and thermal noise radiation interference feature patterns included in the interference feature library to obtain image interference features; Applying a corresponding compensation strategy to the preliminary enhanced image according to the image interference features for interference correction to obtain a final radiation-resistant enhanced image, and recording the current radiation type, radiation intensity, and cumulative dose to form radiation environment parameter data; 3. The method for detecting defects in plastic encapsulation materials in a radiation environment based on machine vision according to claim 1, characterized in that, The performing dual-tree complex wavelet transform feature decomposition and HSV color space defect feature extraction on the radiation-resistant enhanced image to obtain multi-dimensional feature representation data including micro-crack, bubble, deformation, and color difference features includes: Applying a geometric deformation correction algorithm to the radiation-resistant enhanced image, and eliminating the image distortion caused by radiation through a bicubic B-spline transformation model with multiple control points to obtain a corrected image with sub-pixel accuracy; Apply the dual-tree complex wavelet transform to the corrected image, decompose the image into multiple decomposition scales, each scale containing multiple directional sub-bands, to obtain multi-scale and multi-directional feature sub-bands. Among them, the dual-tree complex wavelet transform is a surface texture analysis method unique to the plastic encapsulation material in the nuclear radiation environment, and has a significant detection effect on radiation micro-cracks; Extract the local second-order matrix feature, Gabor texture feature and morphological feature from the multi-scale and multi-directional feature sub-bands respectively to obtain the structural feature vector of the spacecraft electronic component packaging material; Convert the corrected image to the HSV color space, and extract the statistical features of the three channels of hue, saturation and lightness respectively, including mean, standard deviation, skewness, kurtosis, energy and entropy, to obtain the color feature vector. Among them, the HSV color space feature extraction is used to quantify the characteristic color change of the plastic encapsulation material caused by γ-ray radiation; Concatenate the structural feature vector and the color feature vector to form an initial feature vector, and reduce the feature dimension to a lower dimension through principal component analysis to retain the main information content, and obtain a dimensionality-reduced feature vector; Further reduce the dimension of the dimensionality-reduced feature vector through linear discriminant analysis, and at the same time maximize the separability between different defect categories to obtain an optimized feature vector; Apply the defect feature enhancement algorithm to the optimized feature vector, enhance the difference between the defect area and the normal area in the feature space through a non-linear mapping function, and construct a multi-scale feature pyramid to obtain multi-dimensional feature representation data including typical micro-cracks, bubbles, deformations, and color difference features in the nuclear facility environment.

4. The method for detecting defects of plastic encapsulation materials in a radiation environment based on machine vision according to claim 1, characterized in that, Input the multi-dimensional feature representation data into the AR-DCNN network with residual connection and attention mechanism for dilated convolution multi-task learning to obtain the defect type recognition and severity score results, including: Construct an AR-DCNN network with an encoder-decoder structure. The encoder contains multiple encoding blocks, and each encoding block is set with three parallel multi-scale convolutional units to obtain a multi-scale feature map; Connect a radiation feature enhancement module after the encoder, and process the multi-scale feature map through a residual connection block and a channel attention mechanism to obtain an enhanced feature representation. Among them, the channel attention mechanism is used to enhance the channel response of the plastic encapsulation material features of the spacecraft electronic components; Process the enhanced feature representation through multiple dilated convolutional layers of the decoder to form multi-scale receptive fields, obtain the feature parsing results at different scales, and perform upsampling through transposed convolution between the decoding blocks and fuse with the feature maps of the corresponding encoding layers to obtain a fused feature map; Input the fused feature map into the multi-task learning head, including three parallel branches of a classification head, a segmentation head and a scoring head, to obtain an initial defect prediction result; Optimize the initial defect prediction result by applying a combined loss function. The combined loss function includes the weighted cross-entropy loss for defect classification, the Dice loss for segmentation and the mean square error loss for scoring, to obtain an optimized defect detection model; Use the optimized defect detection model to process the newly input plastic encapsulation material image in the nuclear facility environment, and output the defect type recognition and severity score results.

5. The method for detecting defects in plastic packaging materials under a radiation environment based on machine vision according to claim 4, wherein Inputting the fused feature map into a multi-task learning head, which includes three parallel branches: a classification head, a segmentation head, and a scoring head, to obtain an initial defect prediction result, including: Inputting the fused feature map into the classification head, performing feature transformation through a fully connected layer, to obtain the probability distribution of the encapsulation material defect types, where the defect types include microcracks, bubbles, deformation, color difference, and normal areas that are unique to the radiation environment; Applying a 1×1 convolution operation in the segmentation head to the fused feature map for pixel-level feature mapping, to obtain a binary defect area segmentation mask; Inputting the fused feature map into the scoring head after global average pooling, and performing mapping through a fully connected layer to obtain a defect severity score value; Performing an argmax operation on the defect type probability distribution, and extracting the defect type label corresponding to the highest probability to obtain a defect type prediction result; Applying morphological post-processing to the defect area segmentation mask, including opening and closing operations and area filtering, to obtain a refined defect area boundary; Fusing the defect type prediction result, the refined defect area boundary, and the defect severity score value to form a complete initial defect prediction result for the encapsulation material of spacecraft electronic components.

6. The method for detecting defects in plastic encapsulation materials under a radiation environment based on machine vision according to claim 1, wherein Based on the defect type identification and severity scoring results, predicting the radiation resistance life of the encapsulation material for nuclear facilities through Weibull distribution fitting and random forest algorithm, to obtain a material performance degradation curve and a safety threshold warning, including: Constructing a density-weighted Voronoi diagram based on the defect type identification and severity scoring results, associating the detected defects with the surrounding areas, to obtain a radiation damage pattern map; Applying a kernel density estimation algorithm to the radiation damage pattern map, and using a Gaussian kernel function to calculate the damage distribution density function on the material surface, to obtain the surface damage distribution characteristics of the encapsulation material; Applying support vector regression according to the surface damage distribution characteristics of the encapsulation material to analyze the correlation between different types of defects and radiation exposure, distinguishing defects caused by radiation from those caused by non-radiation factors, to obtain a radiation damage quantification index; Inputting the radiation damage quantification index into a random forest algorithm to construct a performance degradation evaluation model for the encapsulation material, and outputting predicted values of the key performance parameters of the encapsulation material for nuclear facilities, including electrical performance, mechanical strength, and thermal stability indicators; Applying Weibull distribution fitting to the predicted values of the key performance parameters to establish a functional relationship between the performance degradation rate and time, to obtain a material performance degradation curve, where the Weibull distribution is particularly suitable for describing the failure mode of the encapsulation material in a γ-ray radiation environment; Determining the predicted time when the performance drops to a specified percentage of the initial value according to the material performance degradation curve, setting multiple levels of safety thresholds, to obtain safety threshold warning information for the use of the encapsulation material in nuclear facilities.

7. The method for detecting defects in plastic encapsulation materials in a radiation environment based on machine vision according to claim 6, characterized in that, Inputting the radiation damage quantification index into a random forest algorithm to construct a performance degradation evaluation model for the encapsulation material, and outputting predicted values of the key performance parameters of the encapsulation material for nuclear facilities, including electrical performance, mechanical strength, and thermal stability indicators, including: Perform feature normalization on the radiation damage quantification indicators to unify the numerical ranges of the indicators and obtain a standardized damage indicator vector; Match the standardized damage indicator vector with the samples in the historical experimental database to construct a training data set and obtain a sample set containing input features and output performance parameters; Apply the random forest algorithm to the sample set to construct multiple decision trees, and each decision tree is trained using a randomly sampled feature subset to obtain a decision tree set; Perform ensemble learning on the decision tree set. For each performance indicator of the nuclear facility encapsulation material, use the average prediction value of all decision trees as the final prediction result to obtain electrical performance prediction values, including insulation strength and volume resistivity; Predict the mechanical strength indicators based on the decision tree set, calculate the tensile strength, flexural strength, and hardness parameters, and obtain mechanical performance prediction values; Use the decision tree set to predict the heat distortion temperature and coefficient of thermal expansion to obtain the thermal stability prediction value, where the thermal stability prediction value is directly related to the heat resistance reliability of spacecraft electronic components in a radiation environment.

8. A defect detection system for plastic encapsulation materials in a radiation environment based on machine vision, characterized in that, For implementing the method for detecting defects of encapsulation materials under a radiation environment based on machine vision as described in any one of claims 1-7, the system for detecting defects of encapsulation materials under a radiation environment based on machine vision includes: An enhancement module for performing image acquisition and adaptive contrast enhancement processing on the surface of the encapsulation material of the spacecraft electronic component through a lead tungstate shielded high-radiation-tolerant image sensor under a γ-ray environment to obtain an anti-radiation enhanced image; An extraction module for performing dual-tree complex wavelet transform feature decomposition and HSV color space defect feature extraction on the anti-radiation enhanced image to obtain multi-dimensional feature representation data containing micro-crack, bubble, deformation, and color difference features; An input module for inputting the multi-dimensional feature representation data into an AR-DCNN network with residual connection and attention mechanism for dilated convolution multi-task learning to obtain defect type recognition and severity scoring results; A prediction module for predicting the radiation resistance life of the encapsulation material for nuclear facilities based on the defect type recognition and severity scoring results through Weibull distribution fitting and random forest algorithm to obtain a material performance degradation curve and a safety threshold warning.

9. A defect detection device for plastic packaging materials in a radiation environment based on machine vision, characterized in that Including a memory and a processor, the memory stores a computer program that can run on the processor, and when the processor executes the computer program, it implements the method for detecting defects of encapsulation materials under a radiation environment based on machine vision as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, the processor is caused to execute the method for detecting defects of encapsulation materials under a radiation environment based on machine vision as described in any one of claims 1 to 7.

Images