Evaporator defect detection method and system based on deep learning

Abstract

This application provides a deep learning-based method and system for evaporator defect detection, relating to the field of anomaly detection. The method includes: acquiring multi-view image data of the evaporator; extracting multi-scale features, including the number of fins, their arrangement direction, and the spatial location of notches, through the backbone network layer of a deep learning target detection model; nonlinearly enhancing the multi-scale features through the model's feature enhancement network layer to obtain enhanced multi-scale features; and then extracting reconstructed features through a hybrid attention mechanism layer; based on the reconstructed features, obtaining the detection results of evaporator fins and notches through the model's detection layer, and comparing them with preset structural parameters to obtain the defect detection results. This application, used in evaporator defect detection, solves the technical problem of existing technologies being unable to stably identify evaporator defects under changing environmental and evaporator structural parameters.

Description

Technical Field

[0001] This application relates to the field of anomaly detection, and in particular to a method and system for detecting evaporator defects based on deep learning. Background Technology

[0002] In the actual production process of evaporators, assembly between the water chamber and the D-tube is usually required. Different models or process schemes of evaporators have differences in structural parameters such as the number of fins, the spacing between fins, and the position of the D-tube notches. If the water chamber is misinstalled or the fin structure is abnormal during assembly, it may lead to refrigerant leakage, decreased heat exchange performance, or even failure of the entire unit.

[0003] Current production lines commonly employ a method of setting physical notches at the bottom of D-shaped tubes to prevent incorrect assembly of the water chambers through structural differences. This is complemented by machine vision systems that detect the location and structural features of the notches. Traditional visual inspection solutions typically rely on rule matching or manually set thresholds, using edge extraction, line detection, and template matching to identify and determine the number, spacing, and notches of partitions. However, existing deep learning-based defect detection methods often focus on identifying and determining single appearance defects or single structural parameters, failing to jointly model the inherent relationships between multiple structural parameters from the perspective of overall structural consistency.

[0004] In evaporator assembly scenarios, the number of fins, the spacing between fins, and the location of the D-tube notch are not independent but together constitute a structural constraint combination for error-proofing assembly. Testing only one parameter, even if the test result itself is correct, may still lead to misjudgments during mixed-specification product production lines, thus failing to meet the reliability requirements for structural error-proofing on industrial production lines. Furthermore, this type of method is sensitive to changes in lighting, viewing angle deviations, and product surface reflections, and often requires readjusting numerous parameters when product specifications change, resulting in poor detection stability and versatility, making it difficult to adapt to the flexible production needs of multiple specifications and batches. Therefore, there is an urgent need for an evaporator defect detection method to solve the technical problem of existing technologies being unable to reliably identify evaporator defects under changing environmental and evaporator structural parameters. Summary of the Invention

[0005] This application provides a method and system for detecting evaporator defects based on deep learning, which solves the technical problem that existing technologies cannot stably identify evaporator defects under changing environmental and evaporator structural parameters.

[0006] To achieve the above objectives, this application adopts the following technical solution: Firstly, a deep learning-based method for evaporator defect detection is provided, comprising: acquiring first-view and second-view image data of the evaporator; wherein the first-view image is used to collect image information related to the fin arrangement structure, and the second-view image is used to collect image information related to the pipe notch structure, so that the fin structure parameters and the notch structure parameters form complementary constraints in space; based on the first-view and second-view image data, multi-scale features including fin quantity features, fin arrangement direction features, and notch spatial location information are extracted through the backbone network layer of a deep learning target detection model; the backbone network layer is a multi-scale feature extraction network constructed based on the YOLO architecture and incorporating a coordinate attention mechanism; after aligning and fusing the multi-scale detection features, nonlinear enhancement is performed through the feature enhancement network layer of the deep learning target detection model to obtain enhanced multi-scale features; the enhanced multi-scale features are extracted and reconstructed through a hybrid attention mechanism layer of the deep learning target detection model consisting of a channel attention mechanism and a self-attention mechanism; based on the reconstructed features, the detection results of the evaporator fins and notches are obtained through the detection layer of the deep learning target detection model; the detection results include the number of fins, the fin spacing, and the number of notches; the detection results are compared with preset structural parameters to obtain the defect detection results.

[0007] The technical solution of this application is not simply to detect targets such as evaporator fins or gaps. Instead, it uses multi-view image joint modeling to extract multiple structural parameters simultaneously under the same detection framework. Based on the combination relationship between the structural parameters, it achieves the determination of evaporator structural consistency. This solves the technical problem that existing technologies cannot stably identify evaporator defects under environmental and evaporator structural parameter changes, while completing defect detection for error-proof assembly scenarios.

[0008] In conjunction with the first aspect mentioned above, in one possible implementation, before extracting multi-scale features containing the number of septa, arrangement direction, and spatial location information of gaps through the backbone network layer of a deep learning object detection model based on first-view and second-view image data, the method further includes: constraining the training process of the student model through a distillation mechanism between the teacher model and the student model, wherein the teacher model is a deep learning object detection model, and the student model is an object detection model obtained by simplifying the structure of the teacher model; during the training process of the student model, first-view and second-view image data are simultaneously input into the teacher model and the student model, and the feature output and detection output of the teacher model at the corresponding positions of the backbone network layer, feature enhancement network layer, and detection layer are obtained, and the feature output and detection output are used as supervision information for the training of the student model to obtain the distilled deep learning object detection model.

[0009] In conjunction with the first aspect mentioned above, one possible implementation involves using feature output and detection output as supervisory information for training the student model. This includes: in the feature layer constraint stage, aligning the corresponding scale feature outputs of the teacher model and student model in the backbone network layer and feature enhancement network layer, and constraining the student model based on the aligned feature outputs; in the attention constraint stage, aligning the spatial weight mapping and channel weight mapping generated by the hybrid attention mechanism layer in the teacher model and student model, and constraining the student model based on the spatial weight mapping and channel weight mapping; in the output layer constraint stage, performing temperature scaling on the detection layer output of the teacher model to obtain supervisory information in the form of a probability distribution, and constraining the detection layer output of the student model based on the supervisory information; during model training, setting weight coefficients for the feature layer constraint stage, attention constraint stage, and output layer constraint stage, and adjusting each weight coefficient according to the training stage; and during model training, introducing a WIoU-based localization loss function to constrain the positional information of the detection results of the student model.

[0010] In conjunction with the first aspect mentioned above, in one possible implementation, based on first-view and second-view image data, multi-scale features containing the number of evaporator septa, their arrangement orientation, and the spatial location of the gaps are extracted through the backbone network layer of a deep learning object detection model. This includes: inputting the first-view and second-view image data into the multi-layer convolutional units of the backbone network layer to obtain basic feature maps at different receptive field scales; jointly modeling the spatial and channel dimensions of the basic feature maps using a coordinate attention mechanism to obtain weighted intermediate feature maps; and outputting feature representations at corresponding scales at different network layers based on the intermediate feature maps to obtain multi-scale features containing fine-grained structural features and overall structural features. These multi-scale features are used to characterize the number distribution, arrangement orientation, and spatial location information of the gaps in the image.

[0011] In conjunction with the first aspect mentioned above, in one possible implementation, the intermediate feature map... Satisfy the following formula:

[0012] in, Here is the basic feature map of channel c, where h is the height coordinate and w is the width coordinate; Use the Sigmoid activation function; This indicates summing the channel dimensions after compression, where C is the original number of channels in the input feature map, and r is the channel compression ratio. Indicates the number of intermediate channels after compression; The trainable weight parameters generated for height-oriented attention represent the degree of influence of the k-th channel feature on the c-th channel attention weight; This represents the feature value of the k-th channel at the h-th position in the height direction; The offset parameter represents the height direction; The trainable weight parameters generated for the width-direction attention represent the degree of influence of the k-th channel feature on the c-th channel attention weight; This represents the feature value of the k-th channel at the w-th position in the width direction; This represents the offset parameter in the width direction.

[0013] In conjunction with the first aspect mentioned above, in one possible implementation, after aligning and fusing the multi-scale detection features, nonlinear enhancement is performed through the feature enhancement network layer of the deep learning object detection model to obtain enhanced multi-scale features. This includes: performing size alignment processing on the multi-scale detection features so that each scale feature satisfies the fusion condition in terms of spatial resolution and channel dimension, thereby obtaining aligned multi-scale detection features; combining the aligned multi-scale detection features to obtain a fused feature representation; and inputting the fused features into the feature enhancement network layer, performing nonlinear mapping on the fused features through a combination of linear transformation and nonlinear activation function operations to obtain enhanced multi-scale features.

[0014] In conjunction with the first aspect mentioned above, in one possible implementation, the enhanced multi-scale features are processed through a hybrid attention mechanism layer of a deep learning object detection model consisting of channel attention and self-attention mechanisms to extract reconstructed features. This includes: inputting the enhanced multi-scale features into a channel attention sub-layer, performing statistical operations on the enhanced multi-scale features along the channel dimension to generate corresponding channel weight vectors, and performing channel weighting processing on the enhanced multi-scale features based on the channel weight vectors to obtain first attention features; inputting the first attention features into a self-attention sub-layer, performing channel correlation modeling on the first attention features in multiple feature subspaces to generate self-attention weights, and performing weighted modulation on the first attention features based on the self-attention weights to obtain reconstructed features.

[0015] In conjunction with the first aspect mentioned above, in one possible implementation, based on recombined features, the detection results of evaporator fins and notches are obtained through the detection layer of a deep learning target detection model. This includes: inputting the recombined features into the detection layer to generate target prediction features; outputting a first detection result for fin targets and a second detection result for notch targets based on the target prediction features, wherein the first detection result for fin targets includes category information, location coordinate information, and confidence information of the fin targets; the second detection result for notch targets includes category information, location coordinate information, and confidence information of the notch targets; spatially locating and counting the fin targets based on the first detection result for fin targets to obtain a fin quantity parameter, and calculating a fin spacing parameter based on the location coordinate information of adjacent fin targets; spatially locating and counting the notch targets based on the second detection result for notch targets to obtain a notch quantity parameter; and using the fin quantity parameter, fin spacing parameter, and notch quantity parameter as the detection results for evaporator fins and notches.

[0016] In conjunction with the first aspect mentioned above, in one possible implementation, the defect detection result is obtained by comparing the detection result with the preset structural parameters. This includes: acquiring the preset structural parameters corresponding to the evaporator to be tested, including preset parameters for the number of partitions, preset parameters for the spacing between partitions, and preset parameters for the number of notches; comparing the number of partitions with the preset parameters for the number of partitions to generate a partition number comparison result; comparing the spacing between partitions with the preset spacing parameters to generate a spacing comparison result; comparing the number of notches with the preset parameters for the number of notches to generate a notch number comparison result; and determining the defect detection result of the evaporator based on the partition number comparison result, the spacing between partitions comparison result, and the notch number comparison result.

[0017] Secondly, a deep learning-based evaporator defect detection system is provided, comprising: an image acquisition module, a feature extraction module, a feature enhancement module, a hybrid attention module, a target detection module, and a defect determination module; The system comprises the following modules: an image acquisition module for acquiring first-view and second-view image data of the evaporator; a feature extraction module for extracting multi-scale features, including the number of septa, their arrangement direction, and the spatial location of notches, based on the first-view and second-view image data and through the backbone network layer of a deep learning target detection model; a feature enhancement module for aligning and fusing the multi-scale features, and then nonlinearly enhancing the fused features through the feature enhancement network layer of the deep learning target detection model to obtain enhanced multi-scale features; a hybrid attention module for inputting the enhanced multi-scale features into a hybrid attention mechanism layer consisting of channel attention and self-attention mechanisms, and weighting the enhanced multi-scale features to obtain reconstructed features; a target detection module for outputting the detection results of septa targets and notch targets based on the reconstructed features and through the detection layer of the deep learning target detection model, and obtaining septa number parameters, septa spacing parameters, and notch number parameters based on the detection results; and a defect determination module for comparing the septa number parameters, septa spacing parameters, and notch number parameters with preset structural parameters to generate defect detection results for the evaporator.

[0018] This application provides a deep learning-based method and system for evaporator defect detection. By acquiring first- and second-view image data of the evaporator and performing multi-view joint modeling, it can comprehensively characterize key structural information such as the number of evaporator fins, fin spacing, and spatial location of D-tube notches. Joint modeling of spatial location and channel information is achieved in the backbone network layer based on a coordinate attention mechanism, effectively strengthening the feature representation directly related to the fin arrangement and notch location. Through the alignment and fusion of multi-scale detection features and the nonlinear enhancement of the feature enhancement network layer, local and overall features at different structural scales are collaboratively expressed in a unified feature space. Furthermore, a hybrid attention mechanism layer composed of channel attention and self-attention mechanisms is used to weighted recombine structurally relevant features. This method highlights features highly relevant to the number, spacing, and structural height of the baffles. In the detection layer, baffle and notch targets are detected and their parameters calculated separately to avoid confusion between different structural information. By comparing the detected baffle number, spacing, and notch number parameters with preset structural parameters, automated judgment of water chamber misinstallation and baffle structural anomalies is achieved. Simultaneously, through distillation training constraints between the teacher and student models, the detection model effectively reduces model size and inference complexity while maintaining detection accuracy. This meets the application requirements for high-precision, real-time error-proof detection of multi-specification products on evaporator production lines, solving the technical problem that existing technologies cannot stably identify evaporator defects under changing environmental and evaporator structural parameters.

[0019] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0020] Figure 1 A system architecture diagram of an evaporator defect detection system based on deep learning provided for embodiments of this application; Figure 2 A flowchart illustrating a deep learning-based evaporator defect detection method provided in this application embodiment; Figure 3 A flowchart illustrating another deep learning-based evaporator defect detection method provided in this application embodiment; Figure 4 A flowchart illustrating another deep learning-based evaporator defect detection method provided in this application embodiment; Figure 5 This is a schematic diagram of the structure of an evaporator defect detection system based on deep learning, provided in an embodiment of this application. Detailed Implementation

[0021] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0022] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0023] The deep learning-based evaporator defect detection method provided in this application can be applied to, for example... Figure 1 The deep learning-based evaporator defect detection system shown is as follows: Figure 1 As shown, the system includes: an image acquisition module 101, a feature extraction module 102, a feature enhancement module 103, a hybrid attention module 104, a target detection module 105, and a defect determination module 106. The image acquisition module 101 is used to acquire first-view image data and second-view image data of the evaporator. Feature extraction module 102 is used to extract multi-scale features, including the number of septa, the arrangement direction, and the spatial location of gaps, based on first-view image data and second-view image data, through the backbone network layer of a deep learning target detection model. The feature enhancement module 103 is used to align and fuse multi-scale features, and to perform nonlinear enhancement on the fused features through the feature enhancement network layer of the deep learning object detection model to obtain enhanced multi-scale features. The hybrid attention module 104 is used to input the enhanced multi-scale features into the hybrid attention mechanism layer composed of channel attention mechanism and self-attention mechanism, and to perform weighted processing on the enhanced multi-scale features to obtain reconstructed features; The target detection module 105 is used to output the detection results of the septum target and the notch target respectively through the detection layer of the deep learning target detection model based on the recombination features, and to obtain the septum number parameter, septum spacing parameter and notch number parameter based on the detection results; The defect determination module 106 is used to compare the parameters of the number of partitions, the spacing between partitions, and the number of notches with the preset structural parameters to generate the defect detection results of the evaporator.

[0024] To address the technical problem that existing technologies cannot reliably identify evaporator defects under varying environmental and evaporator structural parameters, this application provides a deep learning-based evaporator defect detection method. The method includes: acquiring first-view and second-view image data of the evaporator; extracting multi-scale features, including the number of fins, their arrangement direction, and the spatial location of notches, through the backbone network layer of a deep learning target detection model based on the first-view and second-view image data; the backbone network layer is a multi-scale feature extraction network constructed using a coordinate attention mechanism based on the YOLO architecture; aligning and fusing the multi-scale detection features, then performing nonlinear enhancement through the feature enhancement network layer of the deep learning target detection model to obtain enhanced multi-scale features; extracting reconstructed features through a hybrid attention mechanism layer of the deep learning target detection model, consisting of channel attention and self-attention mechanisms; obtaining detection results for evaporator fins and notches through the detection layer of the deep learning target detection model based on the reconstructed features; the detection results include the number of fins, the spacing between fins, and the number of notches; and comparing the detection results with preset structural parameters to obtain the defect detection results.

[0025] Figure 2 A flowchart illustrating the deep learning-based evaporator defect detection method provided in this application embodiment is shown below. Figure 2 As shown, the method includes: S201. Acquire first-view and second-view image data of the evaporator.

[0026] The first-view image data and the second-view image data refer to visible light images acquired from different shooting angles of the evaporator, respectively, used to characterize the appearance information of the fin structure and D-shaped tube notches in different spatial directions within the evaporator. The first-view image is used to acquire image information related to the fin arrangement structure, while the second-view image is used to acquire image information related to the tube notch structure, ensuring that the fin structure parameters and notch structure parameters form complementary constraints in space.

[0027] In one possible implementation, at least two industrial cameras with different imaging directions are set up at the inspection station of the evaporator production line to simultaneously acquire first-view and second-view image data of the evaporator in the assembled state for subsequent processing.

[0028] It should be noted that the first and second perspectives can be set according to the structural characteristics of the evaporator, and there are differences between the two perspectives to avoid the loss of key information due to shading or structural overlap under a single perspective. In the evaporator structure, the arrangement direction of the baffles, the spacing between the baffles, and the D-shaped tube notch are spatially non-coplanar. Under a single perspective, they are easily affected by shading, perspective distortion, or structural overlap, resulting in the inability to stably extract some structural parameters. This application sets non-overlapping first and second perspectives and jointly models the structural features under different perspectives in the same model, so that the baffle structural parameters and the notch structural parameters form complementary constraints in space, thereby providing a reliable data foundation for subsequent structural consistency determination.

[0029] As an example, the first viewing angle can be set to a frontal view perpendicular to the evaporator fin arrangement direction, and the second viewing angle can be set to a side view at a preset angle to the frontal view, so as to simultaneously cover the fin arrangement relationship and the location characteristics of the D-shaped tube notch.

[0030] This step introduces a multi-view image data acquisition method, which enables the spatial distribution characteristics of the key structures of the evaporator to be fully recorded, providing sufficient information for subsequent feature extraction and structure determination.

[0031] S202. Based on first-view and second-view image data, multi-scale features including the number of septa, the arrangement direction, and the spatial location of gaps are extracted through the backbone network layer of a deep learning target detection model.

[0032] The backbone network is a multi-scale feature extraction network built on the YOLO architecture with the introduction of a coordinate attention (CA) mechanism. It employs a hierarchical convolutional structure and incorporates a coordinate attention mechanism during feature extraction to fuse spatial location information while modeling channel semantics. Specifically, the backbone network layer sequentially includes input convolutional units and multi-level feature extraction convolutional units. Each level of convolutional unit encodes features of the evaporator image at different receptive field scales. A coordinate attention mechanism module is embedded after the convolutional feature output, encoding one-dimensional global information in the height and width directions of the feature map to generate channel attention weights containing spatial location information, thereby weighting and modulating the original convolutional features. Simultaneously, the backbone network layer outputs feature maps at different network depths. Shallow features represent fine-grained structural information such as the edges of the septa and the contours of the gaps, while deeper features represent overall structural information such as the overall distribution and arrangement direction of the septa and the relative position of the gaps. This ultimately forms a multi-scale feature output for subsequent feature fusion and object detection.

[0033] In one possible implementation, first-view image data and second-view image data are respectively input into the backbone network layer. Basic feature maps under different receptive fields are obtained through multi-layer convolution operations. A coordinate attention mechanism is introduced into each layer feature map to encode the spatial information in the height and width directions before feeding it back into the channel features.

[0034] It should be noted that by introducing a coordinate attention mechanism into the YOLO architecture, the network can focus more on spatial regions directly related to the arrangement direction of the partitions and the position of the gaps during the feature extraction stage, while maintaining the compactness of the network structure.

[0035] Based on the above steps, this step can simultaneously take into account local structural information and overall arrangement information during the feature extraction stage, so that features such as the number of septa, arrangement direction, and spatial location of gaps can be fully expressed in the feature space.

[0036] S203. After aligning and fusing the multi-scale detection features, nonlinear enhancement is performed through the feature enhancement network layer of the deep learning object detection model to obtain enhanced multi-scale features.

[0037] Feature alignment fusion refers to the process of unifying the spatial resolution and channel dimension of features from feature layers at different scales to meet the fusion operation conditions. This includes: scaling up the feature maps using interpolation upsampling and unifying the number of channels for features at different scales through channel mapping; or compressing the feature maps with higher spatial resolution using downsampling to ensure that features at all scales maintain consistency in spatial resolution and channel dimension. The feature enhancement network layer is used to perform further nonlinear mapping on the fused features.

[0038] In one possible implementation, the multi-scale detection features are size aligned so that each scale feature satisfies the fusion condition in terms of spatial resolution and channel dimension, resulting in aligned multi-scale detection features. The aligned multi-scale detection features are then combined to obtain a fused feature representation. The fused features are then input into a feature enhancement network layer, and through a combination of linear transformation and nonlinear activation function, the fused features are nonlinearly mapped to obtain enhanced multi-scale features.

[0039] It should be noted that by aligning and fusing multi-scale features, information fragmentation caused by inconsistent resolution between features of different scales can be avoided, thereby improving the efficiency of feature utilization.

[0040] As an example, the fused features are input into a feature enhancement network layer containing convolution operations and nonlinear activation functions, and the fused features are nonlinearly transformed to enhance the discriminative features related to the septum structure and the location of the gap.

[0041] Based on the above steps, this step achieves complementary representation of structural information from different scales within the same feature space through the unified fusion and enhancement of multi-scale features, thereby improving the ability of features to characterize the differences in complex structures.

[0042] S204. Enhance multi-scale features by extracting and recombining features through a hybrid attention mechanism layer of a deep learning object detection model composed of channel attention and self-attention mechanisms.

[0043] The channel attention mechanism is used to model the importance of different channel features to the detection task, while the self-attention mechanism is used to model the global correlation between features. The hybrid attention mechanism layer introduces both channel attention and self-attention mechanisms sequentially within the same feature processing unit to jointly model the input features along both the channel and spatial correlation dimensions. The hybrid attention mechanism layer adopts a sequential structure: first, the enhanced multi-scale features are weighted along the channel dimension using a channel attention sub-layer; then, the weighted features are input to the self-attention sub-layer to model the global correlation between features, thereby obtaining the reconstructed features.

[0044] In one possible implementation, the enhanced multi-scale features are first input into the channel attention sub-layer. By performing global statistics on each channel feature and generating a channel weight vector, the enhanced multi-scale features are processed with channel weighting to strengthen the discriminative channel information related to the number of evaporator fins, the fin arrangement direction, and the notch structure, thus obtaining the first attention feature. Subsequently, the first attention feature is input into the self-attention sub-layer to model the correlation between different spatial locations and different feature subspaces. Through self-attention weights, feature interaction and fusion across locations and scales are achieved, and then the first attention feature is rearranged to output the recombined feature.

[0045] It should be noted that the channel attention mechanism can effectively highlight channel features that are strongly correlated with the number of septa and the gap structure, while suppressing the interference of redundant information.

[0046] Based on the above steps, this step uses a hybrid attention mechanism to redistribute the weights of the enhanced features, so that the structure-related features are enhanced in expression globally.

[0047] S205. Based on the recombination features, the detection results of the evaporator fins and notches are obtained through the detection layer of the deep learning target detection model.

[0048] The detection layer refers to the prediction network at the end of the deep learning target detection model, which is used to classify the target and regress its spatial location based on the input features. The prediction network includes multiple prediction branches set in parallel, each corresponding to a feature layer with a different spatial resolution, to adapt to the scale differences presented by the evaporator fins and notches in the image. Each prediction branch includes several convolutional operation units, which sequentially perform feature compression and feature mapping on the input features and output the target prediction features. The detection results include the number of fins and the spacing between fins for fin targets, and the number of notches for notch targets. The detection results are directly output by the detection layer or calculated through post-processing.

[0049] In one possible implementation, the recombined features are input into the detection layer to generate target prediction features. Based on the target prediction features, a first detection result for the partition targets and a second detection result for the notch targets are output. The first detection result for the partition targets includes the category information, location coordinate information, and confidence information of the partition targets. The second detection result for the notch targets includes the category information, location coordinate information, and confidence information of the notch targets. Based on the first detection result for the partition targets, the partition targets are spatially located and counted to obtain the number of partitions, and the partition spacing parameter is calculated based on the location coordinate information of adjacent partition targets. Based on the second detection result for the notch targets, the notch targets are spatially located and counted to obtain the number of notches. The number of partitions, the partition spacing parameter, and the number of notches are used as the detection results for the evaporator partitions and notches.

[0050] It should be noted that the septum targets and the notch targets are distinguished by independent category labels in the detection layer. The septum number parameter, septum spacing parameter, and notch number parameter correspond to the detection output of different targets, avoiding mixed calculation between different structural features, thereby ensuring the structural consistency and interpretability of the detection results.

[0051] As an example, in this embodiment of the application, the detection layer can set up multi-scale prediction branches based on the YOLO series target detection structure, predict the septum target and the gap target by anchor box or no anchor box, and combine the non-maximum suppression algorithm to filter overlapping candidate boxes to obtain stable and reliable detection results.

[0052] Based on the above steps, this step realizes unified target detection and structural parameter extraction of evaporator fins and notch structures, enabling the model to output key detection information such as the number of fins, the spacing between fins, and the number of notches in the same inference process, thereby improving the integration and real-time performance of the detection process.

[0053] S206. Compare the test results with the preset structural parameters to obtain the defect detection results.

[0054] In one possible implementation, preset structural parameters corresponding to the evaporator to be tested are obtained. These preset structural parameters include a preset number of partitions, a preset partition spacing, and a preset number of notches. The number of partitions is compared with the preset number of partitions to generate a partition number comparison result. The partition spacing is compared with the preset partition spacing to generate a partition spacing comparison result. The number of notches is compared with the preset number of notches to generate a notch comparison result. Based on the partition number comparison result, the partition spacing comparison result, and the notch comparison result, the defect detection result of the evaporator is determined. Specifically, during parameter comparison, a consistency judgment rule is used for the number of partitions and the number of notches, and a threshold judgment rule is used for the partition spacing. When the difference between the detected structural parameters and the preset structural parameters exceeds a preset allowable range, a structural defect is determined to exist.

[0055] It should be noted that allowable error ranges or judgment rules can be set for the comparison of the number of baffles, the comparison of the baffle spacing, and the comparison of the number of notches, to accommodate structural fluctuations within the manufacturing tolerance range of different evaporator models, while ensuring that the defect judgment results match the actual assembly requirements. Preset structural parameters can be set according to the parameter range in the actual application scenario.

[0056] As an example, in this embodiment of the application, when the number of spacers or the number of notches is inconsistent with the corresponding preset structural parameters, or when the spacer spacing parameter exceeds the preset allowable range, the current evaporator can be determined to have an assembly defect or structural abnormality, and the defect detection result can be output to the production line control system to trigger the corresponding processing flow.

[0057] Based on the above steps, this step compares the model detection results with explicit structural standards to achieve automated judgment of the evaporator assembly status and structural integrity, making the defect identification process have clear criteria and traceability, and improving the reliability and consistency of evaporator error prevention detection.

[0058] The method provided in this application embodiment can simultaneously complete the detection, counting, and structural parameter extraction of spacers and notches under a unified model framework. It utilizes multi-view images to complete the joint modeling and detection of spacer and notch structures in the same deep learning target detection model, and realizes the unified extraction and consistency determination of key structural parameters such as the number of spacers, the spacing between spacers, and the number of notches. This process employs end-to-end feature learning and target prediction, avoiding reliance on manual rules and fixed thresholds. This allows the detection process to adapt to changes in lighting conditions, shooting angles, and evaporator specifications, thereby improving the stability and versatility of evaporator error-proof assembly detection while ensuring accuracy. It is suitable for automated production inspection scenarios involving multiple evaporator specifications. Compared to existing defect detection methods that rely solely on single structural parameter thresholds, this application simultaneously acquires the number of partitions, partition spacing, and the number of notches within the same inspection process. By jointly comparing multiple structural parameters based on preset structural parameters, it achieves the determination of the overall consistency of the evaporator assembly structure. This effectively avoids misjudgments caused by single-parameter compliance in scenarios involving switching between multiple product specifications or mixed production lines of structurally similar products. It solves the technical problem of existing technologies being unable to stably identify evaporator defects under changing environmental and evaporator structural parameters.

[0059] In one possible implementation, combining Figure 2 ,like Figure 3 As shown, prior to S202, the deep learning-based evaporator defect detection method provided in this application embodiment further includes the following S301 and S302: S301. The training process of the student model is constrained through a distillation mechanism between the teacher model and the student model.

[0060] The teacher model is a deep learning object detection model, and the student model is an object detection model obtained by simplifying the structure of the teacher model.

[0061] Among them, the teacher model refers to a deep learning object detection model with a large parameter scale and strong feature expression ability, which is used to provide high-quality feature representation and prediction results. The student model refers to a lightweight object detection model obtained by channel pruning, layer compression or structural simplification while maintaining the consistency of the overall network function.

[0062] In one possible implementation, the teacher model and the student model use the same input format and detection task settings, enabling them to perform synchronous forward inference based on the same batch of samples during training, thereby providing one-to-one reference information for subsequent distillation constraints.

[0063] It should be noted that the distillation mechanism is only introduced during the model training phase. In the actual online detection or inference phase, only the student model is retained to participate in defect detection, which does not affect the deployment complexity of the system.

[0064] As an example, the teacher model can employ a full-scale, multi-layered feature extraction and detection structure, while the student model can be constructed by reducing the number of feature channels or merging some network layers to meet the real-time detection requirements of the production line.

[0065] Based on the above steps, by introducing constraint information from the teacher model during the training phase, the student model can still learn a more comprehensive structural representation ability even with limited parameter size.

[0066] S302. During the training of the student model, first-view image data and second-view image data are simultaneously input into the teacher model and the student model. The feature output and detection output of the teacher model at the corresponding positions of the backbone network layer, feature enhancement network layer and detection layer are obtained. The feature output and detection output are used as supervision information for the training of the student model to obtain the distilled deep learning target detection model.

[0067] In one possible implementation, first-view and second-view image data are simultaneously input into the teacher and student models. Feature outputs and detection outputs of the teacher model at corresponding positions in the backbone, feature enhancement, and detection layers are obtained. In the feature layer constraint phase, the feature outputs of the teacher and student models at corresponding scales in the backbone and feature enhancement layers are size-aligned, and the student model is constrained based on the aligned feature outputs. In the attention constraint phase, the spatial weight mappings and channel weight mappings generated by the hybrid attention mechanism layers in the teacher and student models are aligned, and the student model is constrained based on these mappings. In the output layer constraint phase, the detection layer output of the teacher model undergoes temperature scaling to obtain supervisory information in the form of a probability distribution, and the detection layer output of the student model is constrained based on this supervisory information. During model training, weight coefficients are set for the feature layer constraint phase, attention constraint phase, and output layer constraint phase, and these weight coefficients are adjusted according to the training phase. During model training, a WIoU-based localization loss function is introduced to constrain the positional information of the detection results of the student model.

[0068] It should be noted that feature layer constraints focus on guiding the student model to learn the structural representation ability of the teacher model, rather than directly participating in the calculation of the final target prediction result. Attention constraints do not change the original attention calculation method of the model, but rather guide the student model to learn the attention allocation rules of the teacher model through a supervised mechanism. Constraints are based on the differences between the output features of the teacher and student models in corresponding network layers to guide the student model to learn the feature representation method of the teacher model; constraints are based on the differences between the attention weight distributions of the teacher and student models to guide the student model to learn the attention allocation rules of the teacher model; and constraints are based on the differences between the detection outputs of the teacher and student models.

[0069] This application embodiment introduces a distillation training mechanism, enabling the student model to simultaneously inherit the behavioral characteristics of the teacher model in terms of feature representation, attention allocation, and detection output during the training phase. While maintaining the simplification of the model structure, it achieves stable learning of the evaporator fin and notch structure information, thereby meeting the comprehensive requirements for detection accuracy and real-time performance in production scenarios.

[0070] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 4 As shown, the above S202 can be specifically implemented through the following S401, S402 and S403, which are explained in detail below: S401. Input the first-view image data and the second-view image data into the multi-layer convolutional units of the backbone network layer respectively to obtain the basic feature maps at different receptive field scales.

[0071] Among them, a multi-layer convolutional unit refers to a structure of multiple convolutional layers arranged sequentially according to the network depth in the backbone network layer. Different convolutional layers correspond to different receptive field scales and are used to extract image feature information from local to global.

[0072] In one possible implementation, first-view image data and second-view image data are respectively input into multi-layer convolutional units with the same structure but shared parameters in the backbone network layer, and basic feature maps at different network depths are obtained through layer-by-layer convolution operations.

[0073] It should be noted that the basic feature maps at different receptive field scales focus on characterizing the local edge features, arrangement orientation features, and overall structural contour information of the evaporator fins. The basic feature map output by shallow convolution can be used to reflect the edge texture and local spacing changes of the fins, while the basic feature map output by deep convolution can be used to reflect the overall distribution morphology of the fins and the spatial positional relationship of the gaps.

[0074] Based on the above steps, feature extraction is performed on the dual-view images through multi-layer convolutional units, so that subsequent attention modeling and multi-scale output have sufficient basic feature input.

[0075] S402. By using the coordinate attention mechanism, the spatial dimension and channel dimension of the basic feature map are jointly modeled to obtain a weighted intermediate feature map.

[0076] Among them, the coordinate attention mechanism is an attention structure that simultaneously encodes spatial location information and channel importance information, which is used to enhance the response strength of key feature channels while maintaining location information.

[0077] In one possible implementation, the basic feature map is aggregated along the horizontal and vertical directions respectively, and spatially aware weight vectors are generated by combining channel dimension information, thereby weighting the basic feature map to obtain an intermediate feature map.

[0078] It should be noted that the coordinate attention mechanism explicitly preserves spatial location information during the feature weighting process, enabling the model to distinguish the importance of septa and gap structures at different spatial locations.

[0079] Preferred, intermediate feature map Satisfy the following formula:

[0080] in, Here is the basic feature map of channel c, where h is the height coordinate and w is the width coordinate; Use the Sigmoid activation function; This indicates summing the channel dimensions after compression, where C is the original number of channels in the input feature map, and r is the channel compression ratio. Indicates the number of intermediate channels after compression; The trainable weight parameters generated for height-oriented attention represent the degree of influence of the k-th channel feature on the c-th channel attention weight; This represents the feature value of the k-th channel at the h-th position in the height direction; The offset parameter represents the height direction; The trainable weight parameters generated for the width-direction attention represent the degree of influence of the k-th channel feature on the c-th channel attention weight; This represents the feature value of the k-th channel at the w-th position in the width direction; This represents the offset parameter in the width direction.

[0081] Based on the above steps, this step enhances the spatial positioning and channel discrimination capabilities of the intermediate feature map, thereby improving the accuracy of perception of key structural features of the evaporator.

[0082] S403. Based on the intermediate feature map, output the feature representations of the corresponding scale at different network layers to obtain multi-scale features that include fine-grained structural features and overall structural features.

[0083] Among them, multi-scale features are used to characterize the number distribution of evaporator fins, their arrangement direction relationship, and the spatial location information of the gaps in the image.

[0084] In one possible implementation, intermediate feature maps are output at shallow, mid, and deep network locations to form feature representations at multiple scales, which are then used for subsequent feature fusion and enhancement.

[0085] It should be noted that features at different scales correspond to different levels of structural information, with high-resolution features focusing on detailed structural representation and low-resolution features focusing on modeling global structural relationships.

[0086] Based on the above steps, this step enables the model to comprehensively characterize the evaporator fins and notch structure by simultaneously utilizing both fine-grained and overall structural information within the same detection framework.

[0087] This application embodiment completes the entire process from basic feature extraction, spatial and channel joint modeling to multi-scale feature output in the backbone network layer using dual-view evaporator images. Under the premise of ensuring that spatial position information is not lost, it realizes multi-level characterization of the distribution of the number of septa, the relationship of their arrangement direction, and the spatial position of the gaps, thereby improving the stability and adaptability of subsequent detection and defect judgment.

[0088] In one possible implementation of the embodiments of this application, combined with Figure 2 ,like Figure 5 As shown, the above S204 can be implemented through the following S501 and S502, which are explained in detail below: S501. Input the enhanced multi-scale features into the channel attention sub-layer, perform statistical operations on the enhanced multi-scale features in the channel dimension to generate the corresponding channel weight vector, and perform channel weighting processing on the enhanced multi-scale features based on the channel weight vector to obtain the first attention feature.

[0089] The channel attention sublayer is used to characterize the relative importance of different feature channels in the current task. By performing statistical analysis on the channel dimension, corresponding weight coefficients are assigned to each channel.

[0090] In one possible implementation, global average pooling and global max pooling operations are performed on the enhanced multi-scale features in the spatial dimension to obtain channel-level statistical descriptions, and channel weight vectors are generated through a shared nonlinear mapping.

[0091] It should be noted that the channel weight vector is used to reflect the differences in the contribution of different channels to the representation of the septum structure and the gap information, thereby achieving adaptive enhancement of key information channels. Feature channels related to septum edges, arrangement direction, or gap regions can obtain higher weights in the channel attention sublayer, while the response intensity of redundant or noisy channels is suppressed accordingly.

[0092] Based on the above steps, this step enables the enhanced multi-scale features to undergo a selective weighting process in the channel dimension, resulting in the first attention feature that can highlight key structural information.

[0093] S502. Input the first attention feature into the self-attention sub-layer, perform channel correlation modeling on the first attention feature in multiple feature subspaces, generate self-attention weights, and perform weighted modulation on the first attention feature based on the self-attention weights to obtain the recombined feature.

[0094] The self-attention sublayer is used to model the global correlation between features. By calculating the inter-channel dependencies in multiple feature subspaces, it enables the adaptive reorganization of feature information.

[0095] In one possible implementation, the first attention features are mapped to a query vector, key vector, and value vector space, and self-attention weights are calculated based on the similarity between the query vector and the key vector. The first attention features are then weighted and modulated using these self-attention weights to obtain the reconstructed features.

[0096] It should be noted that self-attention weights are used to characterize the correlation strength between different feature channels, enabling the model to perceive the global dependency between septum structure features and gap spatial features. For example, when septum arrangement features and gap position features are significantly correlated in spatial structure, the self-attention sublayer can enhance their collaborative response in feature representation.

[0097] Based on the above steps, this step uses self-attention weights to weight and modulate the first attention features, thereby reorganizing the feature information globally and finally obtaining the recombined features for subsequent detection layer input.

[0098] This application's embodiments introduce channel attention sublayers and self-attention sublayers sequentially, enabling enhanced multi-scale features to be progressively screened and reorganized at both the channel importance modeling and global relevance modeling levels. This not only strengthens the consistency of expression of septum and gap-related features but also suppresses the interference of irrelevant information on the detection results, thereby providing subsequent detection layers with reorganized feature representations that are more structurally clear and semantically more focused.

[0099] In another possible implementation, to more comprehensively illustrate the method provided in the embodiments of this application, the training process and parameters of the model are described in detail. The model is trained using a supervised training method, constructing a training sample set containing first-view and second-view image data of the evaporator, and uniformly labeling the septum target and the notch target. Then, the target detection model parameters are initialized, wherein the backbone network layer, feature enhancement network, hybrid attention mechanism layer, and detection layer are initialized using pre-trained weights or a random method. During training, multi-view images are simultaneously input into the model with a fixed batch size. Forward computation yields predictions of septa and notches, and the training loss is calculated based on the predictions and ground truth annotations. The training loss consists of classification loss, target confidence loss, and WIoU-based bounding box regression loss, with each loss term weighted and summed according to preset weights. Model parameters are updated using a gradient descent-based optimization algorithm. The initial learning rate is set to 0.001 and gradually reduced using piecewise decay or cosine annealing strategies. The batch size is set to 8 (different batches can be selected between 8 and 32 depending on task requirements, and the optimal batch size is chosen). The training epochs are set to 100 epochs based on the data scale. When introducing distillation training, the student model introduces additional distillation losses for the feature layer, attention layer, and output layer in addition to the detection loss mentioned above. The weights of each distillation loss are set to relatively large values ​​in the early stages of training and gradually decrease as training progresses, ultimately resulting in an evaporator defect detection model that balances detection accuracy and inference efficiency.

[0100] The foregoing mainly describes the solutions of the embodiments of this application from the perspective of device implementation. It is understood that each device, such as a deep learning-based evaporator defect detection system, includes at least one of the hardware structures and software modules corresponding to each function in order to achieve the above-mentioned functions. Those skilled in the art should readily recognize that, in conjunction with the units and algorithm steps of the various examples described in the embodiments disclosed herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is implemented in hardware or by computer software driving hardware 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 implementation should not be considered beyond the scope of this application.

Claims

1. A method for detecting evaporator defects based on deep learning, characterized in that, include: Acquire first and second view image data of the evaporator; the first and second view are two non-overlapping viewpoints; wherein, the first viewpoint is used to acquire image information related to the fin arrangement structure, and the second viewpoint is used to acquire image information related to the pipe notch structure, so that the fin structure parameters and the notch structure parameters form complementary constraints in space. Based on the first and second viewpoint image data, multi-scale features including the number of septa, the arrangement direction, and the spatial location of the gaps are extracted through the backbone network layer of a deep learning object detection model; the backbone network layer is a multi-scale feature extraction network built on the YOLO architecture with an embedded coordinate attention mechanism. After aligning and fusing the multi-scale detection features, nonlinear enhancement is performed through the feature enhancement network layer of the deep learning target detection model to obtain enhanced multi-scale features. The enhanced multi-scale features are processed through a hybrid attention mechanism layer of a deep learning target detection model consisting of channel attention and self-attention mechanisms to extract reconstructed features; these reconstructed features are used to highlight the key structural features for error prevention in evaporator assembly. Based on the recombined features, the detection results of evaporator fins and notches are obtained through the detection layer of a deep learning target detection model; the detection results include the number of fins, the spacing between fins, and the number of notches. The test results are compared with the preset structural parameters to obtain the defect detection results.

2. The method according to claim 1, characterized in that, Before extracting multi-scale features containing the number of septa, arrangement direction, and spatial location information of gaps through the backbone network layer of a deep learning object detection model based on the first and second viewpoint image data, the method further includes: The training process of the student model is constrained by a distillation mechanism between the teacher model and the student model, where the teacher model is a deep learning object detection model and the student model is an object detection model obtained by simplifying the structure of the teacher model. During the training of the student model, first-view image data and second-view image data are simultaneously input into the teacher model and the student model. The feature output and detection output of the teacher model at the corresponding positions of the backbone network layer, feature enhancement network layer and detection layer are obtained. The feature output and detection output are used as supervision information for the training of the student model to obtain the distilled deep learning target detection model.

3. The method according to claim 2, characterized in that, The step of using the feature output and detection output as supervised information for training the student model includes: In the feature layer constraint stage, the feature outputs of the teacher model and the student model at corresponding scales in the backbone network layer and feature enhancement network layer are size aligned, and the student model is constrained based on the aligned feature outputs. During the attention constraint phase, the spatial weight mapping and channel weight mapping generated by the hybrid attention mechanism layer in the teacher model and student model are aligned, and the student model is constrained based on the spatial weight mapping and channel weight mapping. In the output layer constraint stage, the detection layer output of the teacher model is subjected to temperature scaling processing to obtain supervision information in the form of probability distribution, and the detection layer output of the student model is constrained based on the supervision information. During model training, weight coefficients are set for the feature layer constraint stage, attention constraint stage, and output layer constraint stage, and the weight coefficients are adjusted according to the training stage. During model training, a WIoU-based localization loss function is introduced to constrain the location information of the student model's detection results.

4. The method according to claim 1, characterized in that, Based on the image data from the first and second perspectives, the backbone network layer of the deep learning object detection model extracts multi-scale features including the number of septa, their arrangement direction, and the spatial location of the gaps, including: The first-view image data and the second-view image data are respectively input into the multi-layer convolutional units of the backbone network layer to obtain basic feature maps at different receptive field scales; By using a coordinate attention mechanism, the spatial dimension and channel dimension of the basic feature map are jointly modeled to obtain a weighted intermediate feature map; Based on the intermediate feature map, feature representations of corresponding scales are output at different network layers to obtain multi-scale features that include fine-grained structural features and overall structural features. The multi-scale features are used to characterize the number distribution, arrangement direction relationship of evaporator fins, and spatial location information of the notches in the image.

5. The method according to claim 4, characterized in that, The intermediate feature map Satisfy the following formula: in, Here is the basic feature map of channel c, where h is the height coordinate and w is the width coordinate; Use the Sigmoid activation function; This indicates summing the channel dimensions after compression, where C is the original number of channels in the input feature map, and r is the channel compression ratio. Indicates the number of intermediate channels after compression; The trainable weight parameters generated for height-oriented attention represent the degree of influence of the k-th channel feature on the c-th channel attention weight; This represents the feature value of the k-th channel at the h-th position in the height direction; The offset parameter represents the height direction; The trainable weight parameters generated for the width-direction attention represent the degree of influence of the k-th channel feature on the c-th channel attention weight; This represents the feature value of the k-th channel at the w-th position in the width direction; This represents the offset parameter in the width direction.

6. The method according to claim 1, characterized in that, The step of aligning and fusing the multi-scale detection features, followed by nonlinear enhancement through the feature enhancement network layer of a deep learning object detection model, yields enhanced multi-scale features, including: The multi-scale detection features are size aligned so that each scale feature meets the fusion condition in terms of spatial resolution and channel dimension, resulting in aligned multi-scale detection features. The aligned multi-scale detection features are combined to obtain a fused feature representation; The fused features are input into the feature enhancement network layer, and the fused features are nonlinearly mapped by a combination of linear transformation and nonlinear activation function to obtain enhanced multi-scale features.

7. The method according to claim 1, characterized in that, The step of extracting and recombining features by passing the enhanced multi-scale features through a hybrid attention mechanism layer of a deep learning object detection model composed of channel attention and self-attention mechanisms includes: The enhanced multi-scale features are input into the channel attention sub-layer. Statistical operations are performed on the enhanced multi-scale features in the channel dimension to generate the corresponding channel weight vector. Based on the channel weight vector, the enhanced multi-scale features are subjected to channel weighting to obtain the first attention feature. The first attention feature is input into the self-attention sub-layer, and the channel correlation of the first attention feature is modeled in multiple feature subspaces to generate self-attention weights. The first attention feature is then weighted and modulated based on the self-attention weights to obtain the reconstructed feature.

8. The method according to claim 1, characterized in that, The step of obtaining the detection results of the evaporator fins and notches based on the recombined features and through the detection layer of the deep learning target detection model includes: The recombined features are input into the detection layer to generate target prediction features; Based on the target prediction features, a first detection result for the septum target and a second detection result for the notch target are output. The first detection result for the septum target includes the category information, location coordinate information, and confidence information of the septum target; the second detection result for the notch target includes the category information, location coordinate information, and confidence information of the notch target. Based on the first detection result of the partition target, the partition target is spatially located and counted to obtain the partition quantity parameter, and the partition spacing parameter is calculated based on the position coordinate information of adjacent partition targets; Based on the second detection result of the gap target, the gap target is spatially located and counted to obtain the gap quantity parameter; The parameters of the number of septa, the spacing between septa, and the number of notches are used as the detection results of the evaporator septa and notches.

9. The method according to claim 1, characterized in that, The step of comparing the detection results with preset structural parameters to obtain defect detection results includes: Obtain the preset structural parameters corresponding to the evaporator to be tested. The preset structural parameters include the preset number of partitions, the preset spacing between partitions, and the preset number of notches. The number of spacers is compared with the preset number of spacers to generate a spacer number comparison result. The spacer spacing parameter is compared with the preset spacer spacing parameter to generate a spacer spacing comparison result; The gap quantity parameter is compared with the preset gap quantity parameter to generate a gap quantity comparison result; Based on the comparison results of the number of partitions, the comparison results of the spacing between partitions, and the comparison results of the number of notches, the defect detection results of the evaporator are determined.

10. A deep learning-based evaporator defect detection system, characterized in that, The system includes: an image acquisition module, a feature extraction module, a feature enhancement module, a hybrid attention module, a target detection module, and a defect determination module; The image acquisition module is used to acquire first-view image data and second-view image data of the evaporator. The feature extraction module is used to extract multi-scale features, including the number of septa, the arrangement direction, and the spatial location of the gaps, based on the first-view image data and the second-view image data, through the backbone network layer of the deep learning target detection model. The feature enhancement module is used to align and fuse the multi-scale features, and to perform nonlinear enhancement on the fused features through the feature enhancement network layer of the deep learning object detection model to obtain enhanced multi-scale features. The hybrid attention module is used to input the enhanced multi-scale features into a hybrid attention mechanism layer composed of channel attention mechanism and self-attention mechanism, and to perform weighted processing on the enhanced multi-scale features to obtain reconstructed features; The target detection module is used to output the detection results of the septum target and the notch target respectively through the detection layer of the deep learning target detection model based on the recombination features, and to obtain the septum number parameter, septum spacing parameter and notch number parameter based on the detection results; The defect determination module is used to compare the parameters of the number of partitions, the spacing between partitions, and the number of notches with preset structural parameters to generate defect detection results for the evaporator.

Images