Wire crimping quality x-ray image ai intelligent evaluation method and system

Abstract

The application provides a wire crimping quality X-ray image AI intelligent evaluation method and system, belonging to the technical field of image detection. The method comprises the following steps: acquiring an X-ray image of a wire crimping position, wire crimping related non-image modal data and external environment temperature data; preprocessing the X-ray image, and extracting spatial features in the preprocessed X-ray image by using a convolutional neural network to obtain an image feature vector; encoding the non-image modal data to obtain a non-image feature vector, and fusing the image feature vector and the non-image feature vector to obtain a multi-modal fusion feature vector; inputting the multi-modal fusion feature vector into a defect classification model based on deep learning to obtain a defect type and a severity; correcting the severity based on the external environment temperature data, and predicting a trend of the defect within a preset time window to obtain a defect development trend; and marking the defect type, the severity and the defect development trend on the X-ray image to obtain a detection report.

Description

Technical Field

[0001] This application relates to the field of image inspection technology, and in particular to an AI-powered intelligent evaluation method and system for X-ray images of conductor crimping quality. Background Technology

[0002] Wire crimping is a core connection process in fields such as power transmission and automotive wiring harnesses. Its quality determines the reliability of system operation. Once internal defects exist, it can easily lead to serious accidents such as increased contact resistance, localized overheating, or even wire breakage. Due to the concealed nature of crimping defects, X-ray inspection, which can visualize the internal structure, has become the mainstream non-destructive testing method in the industry and has been incorporated into relevant technical guidelines and accident prevention measures.

[0003] However, traditional inspection relies on manual image interpretation, which suffers from low efficiency, susceptibility to experience-based bias leading to missed or incorrect judgments, and an inability to integrate non-image information such as wire material and crimping process. Furthermore, thermal stress caused by external environmental temperatures accelerates defect evolution, and manual image interpretation struggles to quantify this impact and predict defect development trends. These issues limit the accuracy and efficiency of crimping quality inspection.

[0004] Therefore, there is an urgent need for an AI-powered intelligent image evaluation method and system for wire crimping quality X-ray images to overcome the aforementioned industry pain points. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides an AI-powered intelligent image evaluation method and system for X-ray images of wire crimping quality.

[0006] A first aspect of this application provides an AI-powered intelligent image evaluation method for X-ray images of wire crimping quality, comprising: Acquire X-ray images of the wire crimping area, non-image modal data related to wire crimping, and external ambient temperature data; The X-ray image is preprocessed, and spatial features in the preprocessed X-ray image are extracted using a convolutional neural network to obtain an image feature vector; The non-image modal data is encoded to obtain a non-image feature vector, and the non-image feature vector is fused with the image feature vector to obtain a multimodal fusion feature vector; The multimodal fusion feature vector is input into a deep learning-based defect classification model to identify the defect type and severity of wire crimping. The severity is corrected based on the external ambient temperature data, and the development trend of the defect within a preset time window is predicted to obtain the defect development trend. The defect type, the severity, and the defect development trend are visually marked on the X-ray image, and a detection report is generated, which includes defect information and processing suggestions.

[0007] A second aspect of this application provides an AI-powered intelligent image evaluation system for X-ray images of wire crimping quality, comprising: The data acquisition module is used to acquire X-ray images of the wire crimping area, non-image modal data related to wire crimping, and external ambient temperature data; The feature extraction module is used to preprocess the X-ray image and extract spatial features from the preprocessed X-ray image using a convolutional neural network to obtain an image feature vector. The modality fusion module is used to encode the non-image modality data to obtain a non-image feature vector, and to fuse the non-image feature vector with the image feature vector to obtain a multimodal fusion feature vector; The defect identification module is used to input the multimodal fusion feature vector into a deep learning-based defect classification model to identify the defect type and severity of the wire crimping. The trend prediction module is used to correct the severity based on the external ambient temperature data and predict the development trend of the defect within a preset time window to obtain the defect development trend. The report generation module is used to visually mark the defect type, the severity, and the defect development trend on the X-ray image and generate a detection report, which includes defect information and processing suggestions.

[0008] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the above-described AI intelligent image evaluation method for wire crimping quality X-ray images.

[0009] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method for intelligent evaluation of X-ray images of wire crimping quality.

[0010] The beneficial effects of the AI-powered intelligent X-ray image evaluation method and system for conductor crimping quality provided in this application are as follows: This application overcomes the limitations of traditional manual evaluation, which relies on experience and is inefficient. By extracting spatial features of the image through a convolutional neural network and fusing non-image information such as conductor material and crimping process through multimodal fusion, the accuracy of defect type identification and severity determination is effectively improved, further avoiding missed or misjudgments. Secondly, by correcting the severity and predicting the development trend of defects through external environmental temperature data, an upgrade from static detection to dynamic prediction is achieved. Furthermore, by visualizing and marking X-ray images and generating inspection reports including defect information and processing suggestions, the inspection process is simplified, work efficiency is improved, and the utilization and traceability of inspection data are enhanced, providing efficient and intelligent technical support for conductor crimping quality control. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating an AI-powered intelligent image analysis method for X-ray images of wire crimping quality, provided in an embodiment of this application. Figure 2 This is a structural block diagram of an AI-powered intelligent image interpretation system for X-ray images of wire crimping quality provided in an embodiment of this application. Figure 3 This is a schematic block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0012] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0013] To make the purpose, technical solution, and advantages of this application clearer, the following will be described in conjunction with the appendix. Figure 1-3 The following is an explanation using specific examples.

[0014] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating an AI-powered intelligent image evaluation method for X-ray images of wire crimping quality according to an embodiment of this application. The method includes: S101: Acquire X-ray images of the wire crimping area, non-image modal data related to the wire crimping, and external ambient temperature data.

[0015] In this embodiment, the goal of data acquisition is to obtain multi-source information representing the crimping state. First, X-ray images of the wire crimping part are acquired to capture key areas of the wire crimping part. Key areas include the stress section of the crimping joint, the mating surface between the wire and the terminal, and the surrounding transition area. Image clarity includes internal structural details, such as the arrangement of the wire strands, the density distribution after crimping, and whether there are hidden features such as gaps or breaks. During X-ray image acquisition, parameters such as X-ray intensity and exposure time need to be controlled.

[0016] Secondly, non-image modal data related to wire crimping provides the background support for interpreting crimping quality. This type of data relates the rationality of the crimping process to the physical properties of the wire itself. Specifically, it includes wire material information, crimping process parameters, and wire specification information. This data is extracted from production process records, wire product manuals, and crimping equipment logs. Wire material information includes copper, aluminum, or alloy type, and corresponding physical parameters such as the coefficient of thermal expansion and modulus of elasticity. Crimping process parameters include the pressure value of the crimping tool, the number of crimps, the crimping die model, and the ambient humidity during crimping. Wire specification information includes the wire cross-sectional area, the number of strands, and the insulation layer thickness.

[0017] Finally, the collected external environmental temperature data includes the current real-time temperature, the rate of temperature change per unit time (hourly temperature rise / fall), historical temperature sequences (e.g., daily high and low temperatures, number of temperature fluctuations, and duration of each fluctuation over the past 72 hours), and a temperature forecast sequence for a preset time window obtained through meteorological forecasting. Temperature data is collected in real-time via temperature sensors or obtained from a meteorological platform. The preset time window is set based on the actual service scenarios of wire crimping, the influence of temperature on the evolution of crimping defects, and the anticipated needs of engineering operation and maintenance.

[0018] The acquisition of these three types of data follows the principles of synchronicity, accuracy, and completeness. For example, X-ray images and non-image modal data need to correspond to the same crimping connector; temperature data needs to be consistent with the X-ray image acquisition time point, while excluding abnormal data, such as blurred images caused by imaging failures, incorrect process parameter recordings, and temperature sensor malfunction data.

[0019] S102: Preprocess the X-ray image and use a convolutional neural network to extract spatial features from the preprocessed X-ray image to obtain the image feature vector.

[0020] In this embodiment, the preprocessing stage requires constructing a multi-step purification and enhancement process based on the imaging characteristics and potential interference of X-ray images. Specifically, firstly, image normalization is performed to unify the size and pixel grayscale range (normalized to the [0,1] interval) of X-ray images from different acquisition devices and exposure parameters, in order to eliminate feature shifts caused by device differences and parameter fluctuations. Subsequently, a combination of adaptive median filtering and Gaussian filtering is used to filter out salt-and-pepper noise and Gaussian noise generated during X-ray imaging, while preserving the edge details and density variation characteristics of the crimping area (crimping defects, such as gaps and broken strands, are reflected in local grayscale differences and edge discontinuities). Next, image enhancement processing is performed, using histogram equalization or contrast-limited adaptive histogram equalization technology to improve the contrast between the key crimping area and the background, and enhance the visual recognition of latent defects (local density unevenness). Finally, threshold segmentation and morphological operations (erosion, dilation) are used to extract the effective area of ​​the crimped joint and remove irrelevant background interference (device borders, wire insulation edges) from the image.

[0021] In the feature extraction stage, based on the preprocessed image, the hierarchical structure of the convolutional neural network is used to progressively mine and abstract spatial features. The convolutional neural network extracts features from the image layer by layer through the sliding calculation of the convolutional kernels: the shallow network (2-3 convolutional layers) mainly captures the low-order spatial features of the image, including basic information such as the edge contour of the crimped area, gray-level gradient, and local texture. These features indicate whether the overall shape of the crimped joint is regular. The middle network (4-6 convolutional layers) further integrates the shallow features and extracts the intermediate semantic features, such as the arrangement pattern of the conductor strands, the uniformity of the density distribution of the crimped surface, and structural abrupt changes in local areas. These features are related to the preliminary judgment of crimping defects. The deep network (7 or more convolutional layers) forms high-order semantic features through feature combination and abstraction, which are used to characterize the specific shape of the defect (shape of the gap, location and length of the break), spatial distribution pattern, and degree of difference from the normal area. Finally, through fully connected layers or global average pooling operations, the multi-scale spatial features extracted by each layer of the convolutional neural network are integrated into a fixed-dimensional image feature vector, which includes both local detail information and overall structural features of the image.

[0022] S103: Encode the non-image modal data to obtain non-image feature vectors, and fuse the non-image feature vectors with the image feature vectors to obtain multimodal fusion feature vectors.

[0023] In this embodiment, firstly, an encoding method corresponding to the type of non-image modal data is used to convert the non-image modal data into non-image feature vectors. The non-image modal data includes conductor material information, crimping process parameters, and conductor specification information. Secondly, a multi-layer structure of a convolutional neural network is used to extract multi-scale spatial features from the preprocessed X-ray image, obtaining shallow, mid-level, and deep feature vectors. The non-image feature vectors are then fused with the shallow, mid-level, and deep feature vectors respectively, and an attention mechanism is used to obtain shallow, mid-level, and deep fused features. Finally, the shallow, mid-level, and deep fused features are interacted and aggregated across scales to form a multi-modal fused feature vector.

[0024] S104: Input the multimodal fusion feature vector into the deep learning-based defect classification model to identify the defect type and severity of wire crimping.

[0025] In this embodiment, the construction of the defect classification model requires customized design based on the characteristics of wire crimping defects. Specifically, the identification target is first determined: defect types include problems in the crimping process, including but not limited to wire strand breakage, gaps in the crimping area, material deformation caused by excessive crimping, loose and uncompacted strands, and misaligned crimping of terminals and wires; the severity needs to be converted into quantifiable level standards, such as classifying defects into three levels: minor, moderate, and severe, based on defect size (gap area, fracture length), impact on structural strength (stress concentration factor), and operational risk level (whether there is an immediate risk of failure). The defect classification model adopts a multi-task learning architecture, that is, embedding both a defect type classification head and a severity regression head into the deep learning network simultaneously, and achieving collaborative optimization of the two tasks by sharing the feature extraction capabilities of the backbone network. The type classification head uses the cross-entropy loss function; the severity regression head uses the mean squared error loss function, enabling prediction of continuous or graded values ​​of defect severity, so that the defect classification model can both distinguish defect categories and make quantitative judgments on the severity of the same type of defect.

[0026] The training set for the defect classification model consists of multimodal labeled samples including different wire specifications, crimping processes, defect types, and severity levels. The total number of samples is divided into training and testing sets in an 8:2 ratio. The training set covers materials such as copper / aluminum / alloys, various process parameters, and all defect types and severity levels (light, medium, and severe). Samples of rare defects that have been sampled are balanced, and the sample size is expanded to 3-5 times the original size through data augmentation such as image distortion and density modulation. All samples are double-labeled with defect type labels and severity labels containing measured parameters. The network is based on Vision. The Transformer backbone network was adapted and modified, and an attention mechanism module was added to enhance the recognition of key feature dimensions such as gap density distribution and fracture structure abrupt changes. Sample balancing and data augmentation strategies were integrated to solve the problem of industrial sample imbalance. The final defect classification model achieved an overall recognition accuracy of ≥98% on the training set, a single defect type recognition accuracy of ≥97% and a severity classification accuracy of ≥96%. On the independent test set, the overall recognition accuracy was ≥95%, the common defect recognition accuracy was ≥96%, the rare defect recognition accuracy was ≥90%, and the severity classification accuracy was ≥93%.

[0027] This defect classification model uses the AdamW optimizer and sets the initial learning rate to 2e. -4 Based on the cosine annealing decay and learning rate preheating strategy, the weight decay coefficient is 1e. -5 The batch size is 8-16, the number of training rounds is 50-100, the weights of the loss functions for defect type classification and severity regression are 1.0 and 0.5 respectively, and 80% is used as the confidence threshold during the inference stage.

[0028] In this embodiment, during the inference phase, after the multimodal fused feature vector is input into the defect classification model, the hierarchical feature processing of the backbone network sequentially completes further abstraction and semantic association of the features. Specifically, firstly, the bottom layer network captures the correlation information (feature differences between defective and normal regions) in the fused features; then, the middle layer network mines the mapping relationship between defect types and process parameters and image features (the correlation between over-crimping defects in aluminum wires and specific grayscale distributions and pressure parameters); finally, the classification head and regression head output the probability distribution and severity quantification results of the defect types (e.g., moderate defects correspond to a void area of ​​0.8 mm). 2To ensure the reliability of the identification results, the defect classification model also includes a post-processing mechanism. This mechanism uses probability thresholds for filtering (e.g., if the type identification probability is less than 80%, it is marked as a suspected defect and requires manual review) and logical verification of severity and defect type. For example, a complete break in the strand corresponds to a severe defect, and if the defect classification model outputs a mild defect, it is considered an abnormal result. The 80% probability threshold is determined based on the actual industrial working conditions of wire crimping defect detection, the engineering accuracy requirements for defect identification, and the measured performance of the defect classification model.

[0029] S105: Based on external ambient temperature data, the severity is corrected, and the development trend of the defect is predicted within a preset time window to obtain the defect development trend.

[0030] In this embodiment, based on the thermal expansion coefficient and elastic modulus of the conductor material, the crimping process parameters, and the conductor specifications, the thermal stress distribution generated by temperature changes in the crimping area is calculated. Based on the thermal stress distribution and severity, a temperature-stress coupled defect evolution model is constructed. Simultaneously, based on historical temperature sequences, a temperature cycle cumulative damage factor is calculated, and the severity is corrected to obtain the corrected cumulative damage severity. This corrected severity is used as the initial state. The current temperature, the rate of temperature change, and the predicted temperature sequence within a preset time window are input into the temperature-stress coupled defect evolution model to obtain a predicted defect development trend.

[0031] The external environmental temperature data includes the current temperature, temperature change rate, historical temperature series, and temperature prediction series within a preset time window. The historical temperature series includes the daily maximum temperature, daily minimum temperature, number of temperature fluctuations, and temperature duration within a preset time period.

[0032] S106: Visualize and mark the defect type, severity, and defect development trend on X-ray images, and generate an inspection report that includes defect information and handling suggestions.

[0033] In this embodiment, firstly, differentiating visual identifiers are used to distinguish defect types. For example, a red rectangle marks a broken strand, a blue circle marks a gap in the crimping, and a yellow triangle marks excessive deformation in the crimping. The line type of the frame (solid or dashed) can further distinguish the confirmation status of the defect (solid line indicates a confirmed defect, dashed line indicates a suspected defect); at the same time, the defect type name ("gap defect") is marked next to the frame. Secondly, the severity is presented through color depth, marking symbols, or numerical quantification. For example, a light-colored frame marks a mild defect, a medium-colored frame marks a moderate defect, and a dark-colored frame marks a severe defect, or a specific quantitative indicator (gap area 1.2mm) is marked. 2(Medium-level defects). Dynamic arrows, gradient colors, or text annotations are used to supplement the defect development trend. For example, a red arrow points to the direction of defect extension, indicating an expected expansion of 0.3mm in the next 7 days; a gradient color from light to dark fills the defect area, representing the risk escalation gradient; and flashing warnings or emergency handling alerts are overlaid on high-risk defects with severe fractures. Furthermore, the visual markers support interactive functions; technicians can click on the marker box to view detailed information, such as the multimodal data corresponding to the defect, key points of feature extraction, and temperature-stress coupling analysis results.

[0034] The generation of the inspection report in this embodiment follows the principles of logic and guidance. Structurally, it includes core defect information, supplementary explanations, and handling suggestions. Core defect information includes key parameters for each defect, such as a unique defect number, its specific location in the X-ray image (e.g., the crimped joint is slightly to the left of the center, coordinates X:235, Y:189), defect type and judgment criteria, severity, and defect development trend. Supplementary explanations integrate background information, including basic information about the wire crimping and data acquisition conditions. Basic information includes the wire material, specifications, and crimping process parameters; data acquisition conditions include X-ray imaging parameters, temperature acquisition time, and environmental conditions. Handling suggestions provide differentiated and feasible solutions based on the defect type, severity, and development trend. For example, for minor defects, continuous monitoring and optimization of crimping pressure parameters in subsequent batches are recommended; for moderate defects, rework and re-crimping are recommended, followed by re-inspection before use; for severe defects, direct scrapping is recommended, along with investigation of crimping equipment malfunctions and the rationality of process parameters.

[0035] As can be seen from the above, this application breaks through the limitations of traditional manual image evaluation, which relies on experience and is inefficient. By extracting spatial features of images through convolutional neural networks and fusing non-image information such as wire material and crimping process through multimodal fusion, it effectively improves the accuracy of defect type identification and severity determination, further avoiding missed or misjudgment. Secondly, by correcting the severity and predicting the development trend of defects through external environmental temperature data, it achieves an upgrade from static detection to dynamic prediction. Furthermore, by visualizing and marking X-ray images and generating inspection reports including defect information and processing suggestions, it simplifies the inspection process, improves work efficiency, and enhances the utilization and traceability of inspection data, providing efficient and intelligent technical support for wire crimping quality control.

[0036] In one embodiment of this application, before preprocessing the X-ray image and extracting spatial features from the preprocessed X-ray image using a convolutional neural network to obtain an image feature vector, the method further includes: Extract the dataset of unlabeled wire-crushed X-ray images from the X-ray images to obtain unlabeled X-ray images; Based on the physical mechanism of compression defects, simulated defect samples are synthesized on unlabeled X-ray images through geometric deformation and density modulation to construct self-supervised training sample pairs. The encoder is trained using a contrastive learning framework with self-supervised training sample pairs to obtain pre-trained weights. Pre-trained weights are used as initialization parameters for a convolutional neural network to extract spatial features from pre-processed X-ray images.

[0037] In this embodiment, firstly, a dataset of clean, unlabeled wire crimping X-ray images is selected and extracted from the X-ray images to obtain unlabeled X-ray images. These unlabeled X-ray images include normal crimping structures or natural minor defects, without manually labeled information such as defect category, size grade, etc.

[0038] Secondly, adhering to the physical formation mechanisms of real defects such as wire crimping cracks, gaps, strand misalignment, and insufficient compaction, this method abandons random and irregular image manipulation and employs controllable geometric deformation and grayscale density modulation techniques to synthesize simulated defect samples on compliant, unlabeled X-ray images. Geometric deformation simulates local stretching, compression, displacement, and fracture deformation, restoring structural distortions caused by improper compression stress; density modulation, by adjusting local X-ray grayscale values, replicates the differences in X-ray penetration caused by internal voids, looseness, and gaps. Based on these methods, pairs of training samples are constructed for each original image, forming standardized self-supervised training sample pairs.

[0039] Furthermore, self-supervised training samples are used to specifically train the feature encoder through a contrastive learning framework. During training, the encoder is guided to distinguish the feature differences between normal crimped structures and synthetic defect structures, deeply exploring the underlying patterns of inherent texture, contour distribution, and grayscale changes in the crimped area. After iterative optimization, pre-trained weights adapted to the specific features of wire crimped X-ray images are solidified. These pre-trained weights possess the basic ability to identify common structural anomalies, density variations, and contour deviations. Finally, the pre-trained weights are used as initialization parameters for subsequent convolutional neural networks. These pre-trained weights are adapted to the imaging characteristics and defect features of wire crimped X-ray images.

[0040] From the above, it can be concluded that this embodiment synthesizes simulated defect samples that fit the actual working conditions by using the real physical mechanism of crimping defects and employing controllable geometric deformation and density modulation methods. It also constructs self-supervised training sample pairs, and then completes encoder training to obtain pre-trained weights based on the contrastive learning framework. Finally, it initializes the convolutional neural network with the pre-trained weights, which effectively solves the pain points of scarce labeled samples of wire crimping defects, high cost of manual labeling, and uneven distribution of real defect samples in industrial scenarios. This enhances the feature capture capability of hidden defects such as internal strand breakage and density anomalies.

[0041] In one embodiment of this application, based on the physical mechanism of crimping defects, simulated defect samples are synthesized on unlabeled X-ray images through image deformation, local stretching and compression, and density modulation to construct self-supervised training sample pairs, including: Extract the geometric contour and density distribution features of the indented region in unlabeled X-ray images; Based on the physical formation mechanism of actual defects during wire crimping, a set of defect simulation strategies is constructed. Based on the defect simulation strategy set, a first enhanced view and a second enhanced view are generated for the same unlabeled X-ray image. The first enhanced view is the base view obtained after size normalization and contrast normalization of the unlabeled X-ray image, and the second enhanced view is the synthetic defect view with defect simulation strategy superimposed on the base view. The first and second augmented views corresponding to the same unlabeled X-ray image are used as positive sample pairs, and the first augmented views corresponding to different unlabeled X-ray images are used as negative sample pairs. The positive and negative sample pairs are used as self-supervised training sample pairs.

[0042] In this embodiment, the underlying information of the key crimping area in the unlabeled X-ray image is extracted, including the outer contour shape of the crimping joint, the geometric boundary of the internal strand arrangement, and the density distribution characteristics such as material compactness and penetration difference corresponding to the X-ray grayscale, so as to anchor the benchmark parameters of the normal crimping structure.

[0043] Based on this, and relying on the physical mechanisms of real defects such as conductor compression under pressure, wire deformation, uneven compaction, and stress concentration, a standardized set of defect simulation strategies is established. These strategies include: applying a nonlinear deformation field to the crimped area through thin-plate spline interpolation to simulate local voids and contour distortion; simulating non-uniform crimping offset of steel cores or aluminum strands by radially stretching or compressing local pixels; and simulating density anomalies caused by loose crimping, over-pressure, or under-pressure by density modulation of pixel grayscale values ​​in local areas.

[0044] Subsequently, based on the defect simulation strategy set, two sets of differentiated views are generated for a single unlabeled original image. Specifically, the first enhanced view serves as a clean baseline sample, undergoing only standardization and normalization processes such as size unification and contrast calibration, while preserving the geometric and density characteristics of the original normal pressing. The second enhanced view, on the other hand, overlays simulated defects matching the physical mechanism onto this baseline view, generating a synthetic defect view with simulated flaws.

[0045] Finally, based on the training logic of comparative learning, self-supervised sample pairs are constructed. The defect-free baseline view and the defective composite view generated from the same original image are set as positive sample pairs; and the baseline views corresponding to different original images are combined in pairs to set as negative sample pairs.

[0046] From the above, it can be concluded that this embodiment extracts the geometric contour and density distribution features of the crimped area in the unlabeled X-ray image, and constructs a standardized defect simulation strategy set based on the formation mechanism of real defects in wire crimping. Then, it uses image deformation, local stretching and compression, and density modulation to generate homogeneous normalized basic views and controllable synthetic defect views, and standardizes the construction of positive and negative sample pairs of the same image for self-supervised comparative learning. This not only generates highly realistic defect samples that fit engineering practice and avoids false and invalid features interfering with model training, but also reduces the cost and error of manual annotation. Furthermore, it allows the encoder to learn the subtle differences in morphology, texture, and gray-scale density between normal crimping structures and various defects, enhancing the feature capture ability of difficult-to-identify defects such as hidden gaps, internal broken strands, and uneven compaction.

[0047] In one embodiment of this application, non-image modal data is encoded to obtain a non-image feature vector, and the non-image feature vector is fused with an image feature vector to obtain a multi-modal fused feature vector, including: Non-image modal data includes conductor material information, crimping process parameters, and conductor specification information; The non-image modal data is converted into a non-image feature vector by adopting an encoding method corresponding to the type of non-image modal data; Multi-scale spatial features in preprocessed X-ray images are extracted using the multi-layer structure of a convolutional neural network, resulting in shallow, mid-layer, and deep feature vectors. Non-image feature vectors are fused with shallow, mid-level, and deep feature vectors respectively, and shallow fused features, mid-level fused features, and deep fused features are obtained through an attention mechanism. Shallow fusion features, mid-level fusion features, and deep fusion features are interacted and aggregated across scales to form a multimodal fusion feature vector.

[0048] In this embodiment, the non-image modal data includes three key engineering data categories: conductor material information, crimping process parameters, and conductor specification information. Conductor material information includes the type of copper, aluminum, or alloy, as well as the corresponding coefficient of thermal expansion, modulus of elasticity, mechanical parameters, and thermal properties. Crimping process parameters include the pressure value of the crimping tool, the number of crimping operations, the crimping die model, and the ambient humidity during crimping. Conductor specification information includes the conductor cross-sectional area, standard wire diameter, number of strands, and insulation layer thickness.

[0049] Subsequently, for the attribute differences of different types of non-image data, a differentiated adaptive encoding method is used to complete the structured transformation. For example, embedding encoding is used for categorical text data (material type, mold specification), and normalization and floating-point encoding are performed for continuous numerical process parameters (pressure, size). Finally, a unified mapping is performed to non-image feature vectors with regular dimensions and semantic parsing. At the same time, based on the multi-level feature extraction capability of convolutional neural networks, the preprocessed X-ray images are decomposed layer by layer into multi-scale spatial features: the shallow network includes basic visual features such as edge contours, local textures, and subtle grayscale changes; the middle network captures intermediate semantic features such as strand arrangement, pressing density, and local structural differences; and the deep network extracts high-order abstract features such as the overall shape of defects, large-scale density anomalies, and key structural mutations, forming a three-level image feature vector system from fine-grained to global.

[0050] Subsequently, cross-modal precise fusion is achieved through an attention mechanism, whereby the encoded non-image feature vectors are sequentially and deeply associated with the shallow, middle, and deep image feature vectors. Specifically, the attention weights are adjusted based on the physical correlation of the current material, process, and specifications, strengthening key effective features and weakening irrelevant and redundant information. For example, the weights for density anomaly features are amplified for cemented carbide conductors, while the focus for high-voltage connection processes is on strengthening the attention to structural deformation features.

[0051] Finally, the information correlation between shallow, medium and deep fusion features is established to overcome the limitations of single-scale features, integrate the laws of visual imaging with the physical laws of material processing, and obtain a multimodal fusion feature vector.

[0052] As can be seen from the above, this embodiment integrates non-image modal data and adapts differential coding to generate regular non-image feature vectors. At the same time, it extracts multi-scale spatial features of X-ray images hierarchically based on convolutional neural networks. Then, it uses an attention mechanism to fuse non-image features with image features at each level, strengthens effective correlation, and suppresses redundant information. Finally, it forms a unified multi-modal fusion feature vector through cross-scale interactive aggregation. This not only makes up for the shortcomings of single X-ray visual recognition that easily overlooks material properties and process conditions, and avoids the risk of misjudgment based solely on images, but also leverages the complementary advantages of image visual features and engineering prior data, thereby improving the accuracy of identifying hidden pressing defects caused by material, specifications, and processes.

[0053] In one embodiment of this application, shallow fusion features, mid-level fusion features, and deep fusion features are interacted and aggregated across scales to form a multimodal fusion feature vector, including: Based on the conductor material information, crimping process parameters and conductor specification information, a physical prior constraint matrix is ​​constructed. The physical prior constraint matrix includes the stress distribution prior and density change prior of the crimping area. Spatial location encoding is performed on shallow fusion features, mid-layer fusion features and deep fusion features respectively; The physical prior constraint matrix is ​​matched with the spatial location code to calculate the physical relevance weight of each spatial location at different scales. The physical correlation weights are weighted element-wise with the fusion features of the corresponding scale to obtain the shallow, medium and deep physical enhancement features. The physically enhanced shallow features, physically enhanced mid-level features, and physically enhanced deep features are adaptively aggregated to form a multimodal fusion feature vector.

[0054] In this embodiment, a physical prior constraint matrix is ​​first constructed based on the conductor material, crimping process parameters, and conductor specifications. This physical prior constraint matrix integrates mechanical and imaging laws. On the one hand, it embeds prior stress distribution in the crimping area under different working conditions to determine the stress concentration location, force transmission path, and vulnerable areas prone to cracking. On the other hand, it incorporates prior density changes corresponding to X-ray imaging to define the grayscale and structural thresholds corresponding to the normal compaction range and the porous range. This ensures that the entire feature aggregation process no longer relies solely on data-driven approaches but rather aligns with the actual physical mechanism of conductor crimping.

[0055] Building upon this foundation, spatial location encoding was performed on each of the three types of fused features obtained earlier—shallow, mid-layer, and deep—marking the pixel coordinates, pressing area location, and local structural range corresponding to each feature, thus binding feature information with physical spatial location. Subsequently, the physical prior constraint matrix was matched and linked with the spatial location encoding of each feature. Based on the perception range of features at different scales, the physical relevance weights of each spatial location in the shallow detail, mid-layer structure, and deep global dimensions were calculated point-by-point. The magnitude of the physical relevance weights was aligned with the degree of stress concentration and density anomaly sensitivity, increasing the weights of high-risk areas prone to voids, breaks, and deformations, while appropriately decreasing the weights of areas with normal structural features.

[0056] Based on the perception range of different scale features, the physical relevance weights of each spatial location in the shallow detail, mid-layer structure, and deep global dimensions are calculated point by point. Specifically, the first step is to map the encoded information of each spatial location (pixel coordinates, overlay area location, local structural range) to the physical prior constraint matrix, matching the stress concentration coefficient (mechanical dimension) and density anomaly sensitivity (imaging dimension) corresponding to that location to obtain the physical basis score S(x,y) (x,y are spatial coordinates) of that location. The second step is to assign a perception coefficient K to each scale according to the perception range characteristics of shallow / mid-layer / deep features: shallow layers target local details, K... 浅 Weight amplification is applied to microscale structures (fine voids); the middle layer targets region structures, K 中Weighted adaptation is applied to the overall density / stress distribution of the press-fit surface; the deeper layer targets global semantics, K 深 The weight of the weak areas of the crimped joint is amplified to form the perception correction coefficients at each scale. In the third step, for each spatial location, the physical basis score is multiplied by the perception correction coefficient of the corresponding scale, and then normalized (to the [0,1] interval) to obtain the physical relevance weight of the point in the shallow detail, middle structure and deep global dimensions.

[0057] Based on physical correlation weights, element-wise weighted processing is performed on the fusion features at the corresponding scales to generate physically enhanced shallow, medium, and deep features, respectively. This amplifies effective features related to mechanical defects and density anomalies, while effectively suppressing irrelevant noise and invalid redundant features.

[0058] Finally, the three physically enhanced features are fused and aggregated to connect the information interaction between fine-grained local details, mesoscale structural correlations and large-scale global semantics, and to fully take into account the hidden defect details and the overall pressure-bearing stress law, so as to generate a multimodal fused feature vector that combines image features, process information, material properties and physical and mechanical constraints.

[0059] From the above, it can be concluded that this embodiment constructs a physical prior constraint matrix based on wire material, crimping process, and wire specifications. It then uses spatial location coding matching based on multi-scale fusion features to calculate the physical correlation weights of different locations. Through element-wise weighting, it achieves physical enhancement of features at each level. Finally, through adaptive aggregation, it generates a multi-modal fusion feature vector. This overcomes the shortcomings of purely data-driven models that lack engineering mechanism support. It can effectively enhance the features of high-stress and density-abnormal-prone areas based on the laws of crimping mechanics and the principles of ray imaging. This improves the accuracy of amplifying subtle hidden defect information and suppressing irrelevant noise interference. It also establishes cross-scale correlations between shallow, medium, and deep features, ensuring that the fused features simultaneously conform to visual imaging laws, material properties, and crimping stress mechanisms. This effectively reduces the probability of misjudgment and missed judgment, thereby improving the professionalism, accuracy, and reliability of subsequent defect identification and grading in engineering implementation.

[0060] In one embodiment of this application, the severity is corrected based on external environmental temperature data, and the development trend of the defect within a preset time window is predicted to obtain the defect development trend, including: External environmental temperature data includes current temperature, temperature change rate, historical temperature series, and temperature prediction series within a preset time window. Based on the thermal expansion coefficient and elastic modulus of the conductor material, the crimping process parameters and conductor specification information, the thermal stress distribution generated by temperature change in the crimping area is calculated. Based on the distribution and severity of thermal stress, a temperature-stress coupled defect evolution model is constructed; Calculate the cumulative damage factor from temperature cycling based on historical temperature sequences; The severity is corrected based on the cumulative damage factor from temperature cycling to obtain the severity of cumulative damage. Using the severity of cumulative damage as the initial state, the current temperature, the rate of temperature change, and the predicted temperature sequence within a preset time window are input into the temperature-stress coupled defect evolution model to obtain a prediction of the defect development trend.

[0061] In this embodiment, the current temperature, temperature change rate, historical temperature sequence, and predicted temperature sequence within a preset time window are considered. The historical temperature sequence includes the daily maximum temperature, daily minimum temperature, number of temperature fluctuations, and duration of temperature changes within a preset time period. Based on this, and using the inherent physical parameters of the conductor material, such as the coefficient of thermal expansion and modulus of elasticity, along with the predetermined crimping process parameters and conductor dimensions, the thermal expansion and contraction effects caused by temperature rises and falls and alternating hot and cold temperatures within the crimped joint are deduced. The thermal stress distribution map of the entire crimping area is calculated, determining the location, magnitude, and transmission range of stress concentration caused by temperature changes.

[0062] Subsequently, based on the severity of the defects and the distribution of thermal stress, a temperature-stress coupled defect evolution model was constructed, which deeply binds the environmental temperature fluctuations, internal stress accumulation and defect propagation laws, and establishes a dynamic correlation mechanism in which temperature changes induce stress superposition and stress superposition accelerates defect deterioration. The temperature-stress coupled defect evolution model is used to describe the propagation rate and direction of defects under thermal stress.

[0063] Specifically, based on long-term accumulated historical temperature sequences, the following parameters are statistically analyzed: frequency of thermal cycling N (number of temperature fluctuations, i.e., the number of cycles exceeding the normal temperature range), duration of extreme temperatures ti (duration of a single high / low temperature event), extreme temperature deviation ΔTi (difference between a single extreme temperature and the reference temperature for wire crimping), and rate of temperature change Ti (rate of rise / fall of a single temperature fluctuation). Based on the inherent physical properties of the conductor material and crimping process parameters, basic correction coefficients are determined: conductor thermal expansion coefficient α, elastic modulus E, material temperature fatigue threshold T0 (critical temperature at which fatigue damage begins at the crimping point), and stress fatigue coefficient k of the crimped structure (characterizing the structural fatigue sensitivity caused by the crimping process). The formula for calculating the damage value of a single temperature cycle is: Di = k·α·E·ΔTi·ti·Ti / T0, where Di = 0 if the temperature does not reach the fatigue threshold T0, indicating no damage. The damage values ​​of all single temperature cycles in history are accumulated and then normalized (the larger the value, the more severe the cumulative damage), resulting in the temperature cycle cumulative damage factor. Temperature cycling cumulative damage factor is used to characterize the cumulative effect of temperature alternation on press-fit defects, thereby correcting the severity of defects, making up for the shortcomings of static X-ray images in reflecting long-term environmental aging, and obtaining the true defect severity level that is close to actual service wear.

[0064] Finally, the actual severity after cumulative damage correction is used as the initial state of evolution. The current temperature, the rate of temperature change, and the temperature prediction sequence for the future preset time period are then input into the temperature-stress coupled defect evolution model to deduce the expansion speed, size increase, and risk escalation of defects under different temperature conditions, and output the defect development trend within the preset time window.

[0065] As can be seen from the above, this embodiment calculates the thermal stress distribution in the crimping area by using external ambient temperature data, the thermal expansion coefficient of the conductor, the elastic modulus, the crimping process and specification parameters, builds a temperature-stress coupled defect evolution model, and calculates the temperature cycle cumulative damage factor based on historical temperature sequences to correct the severity of the original defect. Then, using the corrected real damage state as the initial condition, it inputs real-time and future temperature data to deduce the dynamic development trend of the defect, effectively making up for the shortcomings of relying solely on static X-ray images to determine defects.

[0066] In one embodiment of this application, the severity is corrected based on a temperature cycling cumulative damage factor to obtain a cumulative damage-corrected severity, including: Determine the corresponding temperature sensitivity coefficient based on the defect type; Multiply the temperature cycling cumulative damage factor by the temperature sensitivity coefficient to obtain the type-adaptive cumulative damage correction factor. The severity is obtained by superimposing or multiplying the cumulative damage correction factors that are appropriate for the type.

[0067] In this embodiment, recognizing the varying degrees to which different defects are affected by temperature, the defects are first differentiated and matched based on their identified types, with a temperature sensitivity coefficient individually assigned to each type of crimping defect. For example, different defects such as internal voids, broken strands, loose crimping, and joint deformation have different tolerances to thermal cycles and temperature expansion / contraction due to variations in structural morphology, stress location, and material bonding conditions; therefore, their corresponding temperature sensitivity coefficients are also set differently. The temperature sensitivity coefficient characterizes the degree to which a defect type is sensitive to cumulative damage from temperature cycles.

[0068] Subsequently, the temperature cycling cumulative damage factor is multiplied by the temperature sensitivity coefficient of the corresponding defect type to generate a type-adaptive cumulative damage correction factor that conforms to the actual aging law of this type of defect. Finally, a quantitative calculation method of superposition or weighted multiplication is used to link the determined defect severity with the type-adaptive correction factor to complete the dynamic correction of the initial rating, and obtain the final corrected severity that incorporates the long-term temperature fatigue effect and is more in line with the field service conditions.

[0069] From the above, it can be concluded that this embodiment addresses the differences in sensitivity of different pressing defects to temperature fatigue by matching temperature sensitivity coefficients according to defect types and obtaining type-appropriate correction factors by multiplying the cumulative damage factor based on temperature cycles. Then, it dynamically corrects the severity of the original defects through superposition or multiplication, abandoning the uniform and coarse correction mode. This achieves differentiated and refined temperature damage correction, which not only improves the fit with the true expansion law of various defects affected by thermal stress and effectively corrects the shortcomings of single static image rating that ignores long-term temperature fatigue damage, but also makes the corrected severity level more consistent with the actual service aging state on site.

[0070] In one embodiment of this application, the temperature sensitivity coefficient includes a single temperature sensitivity coefficient and a synergistic temperature sensitivity coefficient; Based on the defect type, determine the corresponding temperature sensitivity coefficient, including: Based on the number of defect types in the crimping area and the relative positional relationship between each defect type, it is determined whether there are multiple defect types. When a single defect type exists, the corresponding single temperature sensitivity coefficient is determined based on that defect type. When multiple defect types exist, the synergistic temperature sensitivity coefficient is calculated based on the individual temperature sensitivity coefficient of each defect type and the coupling factor between the defect types. The synergistic temperature sensitivity coefficient is used as the temperature sensitivity coefficient.

[0071] In this embodiment, the temperature sensitivity coefficient is divided into two categories: single temperature sensitivity coefficient and synergistic temperature sensitivity coefficient. This is specifically designed to distinguish the differentiated temperature change response characteristics when there is a single defect and when multiple defects coexist, so that damage correction is no longer limited to the basic judgment of a single defect.

[0072] When determining a specific temperature sensitivity coefficient, the first step is to assess the actual number and types of defects present in the crimping area. Simultaneously, the spatial relative positions of different defects are analyzed, and it is determined whether multiple defect types exist. If only one type of independent defect exists, without any other adjacent defects superimposed, a single temperature sensitivity coefficient can be directly matched based on the inherent properties of that defect. This coefficient represents the inherent expansion sensitivity of that type of defect when subjected to thermal stress and temperature cycling alone.

[0073] When two or more defects coexist at the press-fit joint, a single coefficient is no longer used; instead, a joint calculation is performed based on the defect coupling factor. Specifically, the individual temperature sensitivity coefficient corresponding to each type of defect is first retrieved. Then, the coupling factor characterizes the synergistic enhancement effect generated when multiple defects coexist, i.e., the superimposed effects of stress concentration and accelerated temperature expansion caused by the coexistence of multiple adjacent defects are amplified. Through weighted calculation, a synergistic temperature sensitivity coefficient adapted to the current composite defect state is obtained.

[0074] From the above, it can be concluded that this embodiment divides the temperature sensitivity coefficient into two types: single type and synergistic type. The values ​​are differentiated according to the number of defects and their spatial relationship in the pressing area. A single defect directly matches the basic sensitivity coefficient. When multiple defects coexist, the synergistic enhancement effect of mutual superposition is quantified according to the coupling factor and the synergistic temperature sensitivity coefficient is calculated. This restores the real mechanism of stress concentration superposition and accelerated temperature aging when multiple defects coexist. It eliminates the problem of underestimating the risk caused by using a single coefficient to evaluate composite defects. It makes temperature damage correction more in line with complex working conditions and further improves the accuracy, rationality and engineering practicality of defect severity calibration and subsequent development trend prediction in multi-defect scenarios.

[0075] In one embodiment of this application, an AI-powered intelligent image evaluation method for X-ray images of wire crimping quality further includes: Obtain the relative positional relationships between various defect types, including defect spacing and defect overlap. When the defect spacing is greater than the preset spacing threshold, the coupling factor is 1, indicating that there is no synergistic enhancement effect between the defects and the degree of defect overlap is 0. When the defect spacing is less than or equal to the spacing threshold, the coupling factor is determined based on the degree of defect overlap, where the coupling factor is positively correlated with the degree of defect overlap and the coupling factor is greater than 1.

[0076] In this embodiment, during the defect spacing measurement stage, based on the spatial coordinate system of X-ray images, the straight-line distance between different defects is calculated with the geometric center of each defect region as the reference, so that the spacing data can represent the distribution relationship of defects in physical space. The determination of the degree of defect overlap is achieved through pixel-level region intersection analysis. First, the contour boundary and pixel coverage of each defect are extracted, and then the specific degree of defect overlap is quantified by calculating the ratio of the intersection area to the union area of ​​different defect regions (0% indicates no overlap, 50% indicates partial overlap, and 100% indicates complete overlap).

[0077] The straight-line distance between defects is calculated based on the spatial coordinate system of the X-ray image. First, the geometric center of each defect region is determined and assigned corresponding two-dimensional coordinates (x1, y1), (x2, y2), ..., (xn, yn). Then, the straight-line distance between any two defect geometric centers is calculated using the formula for the straight-line distance between two points in a Cartesian coordinate system. The formula is: Where d is the straight-line distance between the two defects, and (x1,y1) and (x2,y2) are the coordinates of the geometric centers of the two defect regions, respectively.

[0078] Based on this, a preset spacing threshold is set, which is dynamically adapted based on factors such as the characteristics of the conductor material, crimping process parameters, and defect type. For example, the spacing threshold can be appropriately reduced for copper conductors, while it needs to be increased accordingly for aluminum conductors. When the defect spacing is greater than the preset distance threshold, it indicates that each defect is relatively independent in physical space, their stress fields will not interfere with each other, and the effects of temperature changes on various defects will not have a cumulative effect. Therefore, the coupling factor is set to 1, and the defect overlap is 0.

[0079] When the defect spacing is less than or equal to the spacing threshold, it indicates that each defect is within the effective range of mutual influence. At this time, the degree of defect overlap becomes the key variable that determines the size of the coupling factor. The coupling factor is strictly positively correlated with the degree of defect overlap. Meanwhile, the value of the coupling factor is always greater than 1, thus highlighting the enhanced effect of temperature sensitivity under the synergistic effect of multiple defects. Specifically, when the defect spacing is within the threshold range but there is no overlap (overlap rate of 0%), the coupling factor is greater than 1 (e.g., 1.1-1.3), indicating that although the defects do not directly overlap, the adjacent stress fields will penetrate each other, causing the thermal stress caused by temperature changes to be superimposed around the defects, indirectly accelerating the expansion of the defects. When the defects partially overlap (overlap rate of 1%-50%), the coupling factor gradually increases with the increase of the overlap rate, because the stress concentration effect in the overlapping area will be effectively enhanced, and the thermal expansion and contraction caused by temperature fluctuations will generate a greater stress impact at the overlapping part, making the overall sensitivity of the defects to temperature increase. When the defects completely overlap, the coupling factor reaches its maximum value. At this time, multiple defects merge to form a composite defect body, whose stress concentration coefficient is greater than that of a single defect. The fatigue damage caused by temperature changes will accumulate exponentially, posing a serious threat to the stability of the press-fit structure.

[0080] From the above, it can be concluded that this embodiment accurately extracts two types of spatial correlation features between multiple defects: defect spacing and defect overlap. It sets a spacing threshold to delineate the influence boundary. When the defect spacing is greater than the spacing threshold, the coupling factor is set to 1 to determine that there is no synergistic enhancement effect. When the defect spacing is within the effective influence range, a coupling factor greater than 1 is positively matched according to the defect overlap to quantify the temperature-sensitive synergistic amplification effect caused by the superposition of stress fields and the overlap of weak structural areas of adjacent defects. This not only realizes the refined and quantifiable determination of the multi-defect coupling effect, but also avoids the underestimation of defect hazards and evaluation deviation caused by the previous uniform value or neglect of spatial position relationship. It also conforms to the real physical laws of thermal stress transmission and defect propagation at the wire crimping part.

[0081] In one embodiment of this application, an AI-powered intelligent image evaluation method for X-ray images of wire crimping quality further includes: Based on the defect spacing and defect overlap between different defect types, the equivalent defect morphology after multi-defect coupling is determined. The equivalent defect morphology includes the equivalent defect size, equivalent defect shape, and equivalent stress concentration factor. Based on the equivalent defect morphology, the stress distribution prior and density change prior in the physical prior constraint matrix are updated to obtain the updated physical prior constraint matrix. The updated physical prior constraint matrix is ​​fed back into the cross-scale interaction and aggregation process to update the calculation of physical correlation weights in subsequent detection batches.

[0082] In this embodiment, the equivalent defect morphology after multi-defect coupling is determined based on the data of the spacing and overlap of each defect. The equivalent defect morphology includes the equivalent defect size, the equivalent defect shape, and the equivalent stress concentration factor. Specifically, the calculation of the equivalent defect size integrates the original size, spacing distribution, and overlap area of ​​each defect. For example, when the distance between two adjacent void defects is less than a preset distance threshold and they partially overlap, the equivalent size is based on the merged outer contour, and the amplification effect of the overlapping area is superimposed to obtain the equivalent length, area, or volume parameters representing the weak range of the overall structure. The equivalent defect shape is modeled by fusing the original shape (circular void, linear fracture) and relative positional relationship (parallel arrangement, cross distribution) of each defect. If the fracture defect is adjacent to the void defect and partially overlaps, a composite equivalent shape of linear extension-local expansion is formed. The calculation of the equivalent stress concentration coefficient is based on the single stress concentration coefficient, coupling factor, and geometric characteristics of the equivalent shape of each defect. Through mechanical simulation, the superposition effect of stress concentration after multi-defect coupling is quantified. For example, the equivalent stress concentration coefficient of overlapping defects is greater than that of single defects, and the value is related to the degree of overlap and the combination of defect types. Finally, equivalent defect shape data including equivalent size, shape, and stress concentration coefficient are formed.

[0083] Subsequently, the physical prior constraint matrix is ​​dynamically updated using the calculated equivalent defect morphology data. The stress distribution prior and density change prior in the physical prior constraint matrix are primarily constructed based on the mechanical and imaging laws of a single defect type and a normal press-fit structure. The update process incorporates the mechanical properties and imaging features corresponding to the equivalent defect morphology. Specifically, regarding the stress distribution prior, the original stress field distribution model is modified based on the size, shape, and equivalent stress concentration coefficient of the equivalent defect to determine the stress concentration location, transmission range, and intensity level of the multi-defect coupling region. For example, the stress value at the edge of the equivalent defect is increased, and the stress influence radius is expanded. Regarding the density change prior, the range of grayscale thresholds in the X-ray image is adjusted based on the size and shape of the equivalent defect, allowing the physical prior constraint matrix to match the composite density anomaly features presented in the image after multi-defect coupling, such as grayscale gradient changes and edge blurring in overlapping defect regions. Finally, a post-physical prior constraint matrix representing the multi-defect coupling characteristics is obtained.

[0084] Finally, the updated physical prior constraint matrix is ​​fed back in real time to the cross-scale interaction and aggregation process, forming a dynamic closed loop of detection-modeling-update-optimization. In the feature fusion stage of subsequent detection batches, the updated matrix replaces the initially constructed matrix and is used to recalculate the physical relevance weights.

[0085] As can be seen from the above, this embodiment calculates the coupled equivalent defect morphology by fusing the actual spacing and overlap between multiple defects, and then updates the physical prior constraint matrix based on this. The updated matrix is ​​then fed back to the cross-scale feature interaction aggregation stage, which effectively makes up for the shortcomings of not being able to adapt to complex working conditions with multiple defects coexisting. This reduces the probability of missed or incorrect identification in multi-defect scenarios, thereby enhancing the adaptability, generalization performance and long-term detection accuracy of the entire AI evaluation method under complex on-site working conditions.

[0086] Corresponding to the AI-powered intelligent image evaluation method for X-ray images of wire crimping quality in the above embodiment, Figure 2 This is a structural block diagram of an AI-powered intelligent image interpretation system for X-ray images of wire crimping quality, provided in one embodiment of this application. For ease of explanation, only the parts relevant to the embodiment of this application are shown. References Figure 2 The AI-powered intelligent evaluation system 20 for X-ray images of wire crimping quality includes: a data acquisition module 21, a feature extraction module 22, a modal fusion module 23, a defect identification module 24, a trend prediction module 25, and a report generation module 26.

[0087] Among them, the data acquisition module 21 is used to acquire X-ray images of the wire crimping part, non-image modal data related to wire crimping, and external ambient temperature data; The feature extraction module 22 is used to preprocess the X-ray image and extract the spatial features in the preprocessed X-ray image using a convolutional neural network to obtain the image feature vector. The modality fusion module 23 is used to encode non-image modal data to obtain non-image feature vectors, and fuse the non-image feature vectors with image feature vectors to obtain multimodal fusion feature vectors; The defect identification module 24 is used to input the multimodal fusion feature vector into the deep learning-based defect classification model to identify the defect type and severity of the wire crimping. The trend prediction module 25 is used to correct the severity based on external ambient temperature data and predict the development trend of the defect within a preset time window to obtain the defect development trend. The report generation module 26 is used to visually mark the defect type, severity and defect development trend on the X-ray image and generate an inspection report, which includes defect information and processing suggestions.

[0088] See Figure 3 , Figure 3 This is a schematic block diagram of an electronic device provided according to an embodiment of this application. Figure 3The electronic device 300 in this embodiment may include one or more processors 301, one or more input devices 302, one or more output devices 303, and one or more memories 304. The processors 301, input devices 302, output devices 303, and memories 304 communicate with each other via a communication bus 305. The memories 304 store computer programs, including program instructions. The processors 301 execute the program instructions stored in the memories 304. Specifically, the processors 301 are configured to invoke the program instructions to perform the functions of the modules in the aforementioned device embodiments, for example... Figure 2 The functions of the data acquisition module 21, feature extraction module 22, modality fusion module 23, defect identification module 24, trend prediction module 25, and report generation module 26 are shown.

[0089] It should be understood that, in the embodiments of this application, the processor 301 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0090] Input device 302 may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint orientation information), a microphone, etc., and output device 303 may include a display (LCD, etc.), a speaker, etc.

[0091] The memory 304 may include read-only memory and random access memory, and provides instructions and data to the processor 301. A portion of the memory 304 may also include non-volatile random access memory. For example, the memory 304 may also store device type information.

[0092] In specific implementations, the processor 301, input device 302, and output device 303 described in the embodiments of this application can execute the implementation methods described in any embodiment of the AI ​​intelligent evaluation method for X-ray images of wire crimping quality provided in the embodiments of this application, or they can execute the implementation methods of the electronic devices described in the embodiments of this application, which will not be repeated here.

[0093] In another embodiment of this application, a computer-readable storage medium is provided. This computer-readable storage medium stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the methods described above. Alternatively, the computer program can instruct related hardware to complete the process. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0094] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0095] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0096] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0097] In the several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces or units, or it may be an electrical, mechanical, or other form of connection.

[0098] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0099] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0100] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent evaluation of X-ray images of wire crimping quality using AI, characterized in that, include: Acquire X-ray images of the wire crimping area, non-image modal data related to the wire crimping, and external ambient temperature data; The X-ray image is preprocessed, and spatial features in the preprocessed X-ray image are extracted using a convolutional neural network to obtain an image feature vector; The non-image modal data is encoded to obtain a non-image feature vector, and the non-image feature vector is fused with the image feature vector to obtain a multimodal fusion feature vector; The multimodal fusion feature vector is input into a deep learning-based defect classification model to identify the defect type and severity of wire crimping. The severity is corrected based on the external ambient temperature data, and the development trend of the defect within a preset time window is predicted to obtain the defect development trend. The defect type, the severity, and the defect development trend are visually marked on the X-ray image, and a detection report is generated, which includes defect information and processing suggestions.

2. The AI-powered intelligent image evaluation method for X-ray images of conductor crimping quality according to claim 1, characterized in that, Before preprocessing the X-ray image and extracting spatial features from the preprocessed X-ray image using a convolutional neural network to obtain the image feature vector, the method further includes: Extract the dataset of unlabeled wire-crimped X-ray images from the X-ray images to obtain unlabeled X-ray images; Based on the physical mechanism of compression defects, simulated defect samples are synthesized on the unlabeled X-ray image through geometric deformation and density modulation to construct self-supervised training sample pairs. Using the self-supervised training sample pairs, the encoder is trained using a contrastive learning framework to obtain pre-trained weights; The pre-trained weights are used as the initialization parameters of the convolutional neural network to extract spatial features from the preprocessed X-ray image.

3. The method for intelligent evaluation of X-ray images of conductor crimping quality according to claim 2, characterized in that, Based on the physical mechanism of compression defects, simulated defect samples are synthesized on the unlabeled X-ray image through image deformation, local stretching and compression, and density modulation. Self-supervised training sample pairs are then constructed, including: Extract the geometric contour and density distribution features of the indented region in unlabeled X-ray images; Based on the physical formation mechanism of actual defects during wire crimping, a set of defect simulation strategies is constructed; According to the defect simulation strategy set, a first enhanced view and a second enhanced view are generated for the same unlabeled X-ray image, wherein the first enhanced view is a base view obtained after size normalization and contrast normalization of the unlabeled X-ray image, and the second enhanced view is a synthetic defect view that superimposes the defect simulation strategy on the base view. The first enhanced view and the second enhanced view corresponding to the same unlabeled X-ray image are used as positive sample pairs, and the first enhanced views corresponding to different unlabeled X-ray images are used as negative sample pairs. The positive sample pairs and the negative sample pairs are used as the self-supervised training sample pairs.

4. The method for intelligent evaluation of X-ray images of conductor crimping quality according to claim 1, characterized in that, The process of encoding the non-image modal data to obtain a non-image feature vector, and fusing the non-image feature vector with the image feature vector to obtain a multimodal fusion feature vector, includes: The non-image modal data includes conductor material information, crimping process parameters, and conductor specification information; The non-image modal data is converted into a non-image feature vector using an encoding method corresponding to the type of the non-image modal data. Multi-scale spatial features in preprocessed X-ray images are extracted using the multi-layer structure of a convolutional neural network, resulting in shallow, mid-layer, and deep feature vectors. The non-image feature vector is fused with the shallow feature vector, the middle feature vector, and the deep feature vector respectively, and shallow fused features, middle fused features, and deep fused features are obtained through an attention mechanism; The shallow, medium, and deep fusion features are interacted and aggregated across scales to form the multimodal fusion feature vector.

5. The method for intelligent evaluation of X-ray images of wire crimping quality according to claim 4, characterized in that, The step of performing cross-scale interaction and aggregation of the shallow, mid-level, and deep fusion features to form the multimodal fusion feature vector includes: Based on the conductor material information, crimping process parameters and conductor specification information, a physical prior constraint matrix is ​​constructed, which includes the stress distribution prior and density change prior of the crimping area. Spatial location encoding is performed on the shallow fusion features, mid-layer fusion features, and deep fusion features, respectively; The physical prior constraint matrix is ​​matched with the spatial location code to calculate the physical relevance weights of each spatial location at different scales. The physical correlation weights are weighted element-wise with the fusion features of the corresponding scale to obtain the shallow, medium and deep physical enhancement features. The physically enhanced shallow features, physically enhanced mid-level features, and physically enhanced deep features are adaptively aggregated to form the multimodal fusion feature vector.

6. The method for intelligent evaluation of X-ray images of wire crimping quality according to claim 5, characterized in that, The step of correcting the severity based on the external environmental temperature data and predicting the development trend of the defect within a preset time window to obtain the defect development trend includes: The external environmental temperature data includes the current temperature, the rate of temperature change, historical temperature sequences, and a temperature prediction sequence within a preset time window. Based on the thermal expansion coefficient and elastic modulus in the conductor material information, the crimping process parameters, and the conductor specification information, the thermal stress distribution generated by temperature change in the crimping area is calculated. Based on the thermal stress distribution and the severity, a temperature-stress coupled defect evolution model is constructed; Calculate the temperature cycle cumulative damage factor based on the historical temperature sequence; The severity is corrected based on the temperature cycle cumulative damage factor to obtain the cumulative damage corrected severity. Using the severity of the cumulative damage correction as the initial state, the current temperature, the temperature change rate, and the temperature prediction sequence within a preset time window are input into the temperature-stress coupled defect evolution model to obtain a defect development trend prediction.

7. The method for intelligent evaluation of X-ray images of conductor crimping quality according to claim 6, characterized in that, The step of correcting the severity based on the temperature cycle cumulative damage factor to obtain the cumulative damage corrected severity includes: Determine the corresponding temperature sensitivity coefficient based on the defect type; Multiply the temperature cycling cumulative damage factor by the temperature sensitivity coefficient to obtain a type-adaptive cumulative damage correction factor. The severity level is superimposed or multiplied by the cumulative damage correction factor adapted to the type to obtain the severity level after cumulative damage correction.

8. The method for intelligent evaluation of X-ray images of wire crimping quality according to claim 7, characterized in that, The temperature sensitivity coefficient includes a single temperature sensitivity coefficient and a synergistic temperature sensitivity coefficient; The step of determining the corresponding temperature sensitivity coefficient based on the defect type includes: Based on the number of defect types in the crimping area and the relative positional relationship between each defect type, it is determined whether there are multiple defect types. When a single defect type exists, the corresponding single temperature sensitivity coefficient is determined based on that defect type. When multiple defect types exist, the synergistic temperature sensitivity coefficient is calculated based on the individual temperature sensitivity coefficient of each defect type and the coupling factor between the defect types. The synergistic temperature sensitivity coefficient is used as the temperature sensitivity coefficient.

9. The method for intelligent evaluation of X-ray images of conductor crimping quality according to claim 8, characterized in that, Also includes: Obtain the relative positional relationship between each defect type, including the defect spacing and defect overlap. When the defect spacing is greater than the preset spacing threshold, the coupling factor is 1, indicating that there is no synergistic enhancement effect between the defects and the degree of defect overlap is 0. When the defect spacing is less than or equal to the spacing threshold, the coupling factor is determined based on the degree of defect overlap, wherein the coupling factor is positively correlated with the degree of defect overlap, and the coupling factor is greater than 1.

10. A wire crimping quality X-ray image AI intelligent evaluation system, characterized in that, include: The data acquisition module is used to acquire X-ray images of the wire crimping area, non-image modal data related to wire crimping, and external ambient temperature data; The feature extraction module is used to preprocess the X-ray image and extract spatial features from the preprocessed X-ray image using a convolutional neural network to obtain an image feature vector. The modality fusion module is used to encode the non-image modality data to obtain a non-image feature vector, and to fuse the non-image feature vector with the image feature vector to obtain a multimodal fusion feature vector; The defect identification module is used to input the multimodal fusion feature vector into a deep learning-based defect classification model to identify the defect type and severity of the wire crimping. The trend prediction module is used to correct the severity based on the external ambient temperature data and predict the development trend of the defect within a preset time window to obtain the defect development trend. The report generation module is used to visually mark the defect type, the severity, and the defect development trend on the X-ray image and generate a detection report, which includes defect information and processing suggestions.

Images