A Deep Learning-Based Classification and Detection Method and System for Microchannel Aluminum Flat Tubes

Abstract

This invention discloses a deep learning-based method and system for classifying and detecting microchannel aluminum flat tubes, relating to the field of microchannel aluminum flat tube production. The method includes a time-division stroboscopic image acquisition module for microchannel aluminum flat tubes, an asynchronous multi-image processing module, an asynchronous intelligent analysis module, a classification and detection collection module, and an anomaly handling module. The system includes a microchannel aluminum flat tube transmission device and a multi-image acquisition subsystem, a multi-image preprocessing and classification screening subsystem, a microchannel aluminum flat tube intelligent analysis subsystem, a microchannel aluminum flat tube classification and detection collection subsystem, an anomaly handling and control subsystem, and a real-time production analysis dashboard subsystem. This invention achieves >99% classification detection accuracy and microsecond-level response capability through time-division stroboscopic image acquisition, asynchronous processing design, and a classification and detection collection device, utilizing an improved YOLOv12 algorithm to achieve parallel detection of multiple defects.

Description

Technical Field

[0001] This invention relates to the field of microchannel aluminum flat tube production, and in particular to a method and system for producing microchannel aluminum flat tubes based on deep learning. Background Technology

[0002] Microchannel aluminum flat tubes (MAFTs) are core heat transfer elements in modern high-efficiency heat exchangers, and their quality and reliability directly determine the energy efficiency level and service life of the refrigeration system. Refrigerant leakage caused by surface defects in MAFTs can reduce system energy efficiency by up to 32%. These defects mainly include: microscopic surface defects: pits (diameter ≥ 0.3 mm), scratches (length > 2 mm), impurity embedding (diameter > 0.2 mm), etc., caused by friction during raw material rolling or conveying machinery; geometrical anomalies: interlocking deformation (angle deviation > 5°), length and width deviations (± 0.15 mm), originating from stamping die wear; coating defects: lack of zinc spraying (coverage < 95%), spray coating accumulation (area > 10 mm²), affecting corrosion resistance. Existing detection technologies face the following bottlenecks:

[0003] 1. Manual inspection mode

[0004] Low efficiency: Limited by the human eye's ability to distinguish, the detection speed is ≤3m / s, which cannot match the current 10m / s high-speed production line; High missed detection rate: The average missed detection rate of pit defects is 5.2%, and the missed detection rate of scratch defects is as high as 7.8%; Soaring costs: A single production line requires 12 inspectors (three shifts), and labor costs account for 18% of the production cost.

[0005] 2. Traditional machine vision solutions

[0006] Existing automated detection technologies face three major technological barriers:

[0007] (1) Imaging interference from highly reflective surfaces

[0008] The surface reflectivity of aluminum is >85% (wavelength 550nm). Traditional continuous light sources produce strong specular reflection, resulting in a contrast ratio of <15% between the defect area and the background and an image signal-to-noise ratio (SNR) of <20dB (especially in the 650-850nm band).

[0009] (2) High-speed motion imaging blur

[0010] When the production line speed is ≥8m / s: the effective exposure time of a single frame must be ≤125μs. The existing area array camera frame rate (1kHz) causes motion blur ≥0.1mm (exceeding the defect size tolerance limit).

[0011] (3) Real-time processing delay

[0012] Traditional serial image processing (such as the OpenCV Haar classifier) ​​takes more than 150ms to analyze a single tube. When multiple defects are detected simultaneously, the computational load increases dramatically (CPU utilization > 90%), causing pipeline blockage.

[0013] The industry urgently needs to overcome the following core challenges: at the optical level, it needs to suppress metal reflection and enhance defect features within a μs time; at the algorithm level, it needs to achieve asynchronous analysis of multi-source images at 200+ frames / second; at the system level, it needs to meet the end-to-end latency of <50ms from imaging to classification and detection.

[0014] The time-division stroboscopic image acquisition and asynchronous intelligent analysis method proposed in this application is an innovative solution designed to address the aforementioned pain points. It suppresses specular reflection by using a time-division stroboscopic light source (850nm / 650nm dual-band), achieves 0.01mm / pixel resolution imaging by combining it with a line scan camera, and utilizes an improved MAFT deep learning algorithm to achieve parallel detection of multiple defects, ultimately achieving a classification and detection accuracy of >99% and microsecond-level response capability. Summary of the Invention

[0015] To address industry challenges in the classification and inspection of microchannel aluminum flat tubes, such as imaging interference from highly reflective surfaces, imaging blurring due to high-speed motion, real-time processing delays, diverse defects, and secondary damage during classification and inspection, this application provides a deep learning-based method for classifying and inspecting microchannel aluminum flat tubes. Figure 1 As shown, the following technical solution is adopted:

[0016] S1, the microchannel aluminum flat tube time-division stroboscopic image acquisition module, realizes time-division stroboscopic multi-type image generation and high-throughput transmission and acquisition. It consists of a classification detection conveyor belt, an industrial line scan camera, a time-division stroboscopic flash source and controller, a grating trigger, and a high-throughput image acquisition card. The upper and lower line scan cameras, triggered by the grating and working in conjunction with the time-division stroboscopic flash source, acquire multiple types of images from the high-speed passing microchannel aluminum flat tube, which are then collected by the high-throughput image acquisition card and sent to the image processing server. The specific working logic is as follows: Figure 2 As shown.

[0017] Specifically, this module classifies and detects conveyor belts with a thickness not exceeding 5 cm. It adopts a two-section design, with a gap of no more than 10 cm between the front and rear conveyor belts. This gap is used for the lower linear scanning camera and the lower time-division flash source to capture images of the lower surface of the microchannel aluminum flat tube. The angle between the camera and the time-division flash source and the conveyor belt does not exceed 45 degrees. The upper linear scanning camera and the upper time-division flash source are deployed above the rear conveyor belt to capture images of the upper surface of the microchannel aluminum flat tube. The angle between the camera and the time-division flash source and the conveyor belt does not exceed 45 degrees. The image acquisition accuracy is no less than 0.01 mm. The time-division flash source uses a dual-band wavelength of 850 nm / 650 nm, and the flash frequency is synchronized with the grating trigger. The high-throughput acquisition card uses a PCIe interface with a bandwidth of no less than 64 Gbps.

[0018] S2, the asynchronous image processing module, performs preprocessing on the multi-class images acquired by S1 through time-division strobe scanning. This preprocessing includes cropping, segmentation, binarization connected component analysis, length measurement, and width measurement. The acquired multi-class images include at least standard images, diffuse images, specular images, gloss ratio images, and shape images. Specifically, the standard images undergo ROI cropping and Gaussian filtering; the diffuse images undergo morphological opening and Blob analysis; the specular images undergo nonlocal mean denoising and edge enhancement; the gloss ratio images undergo pixel fusion with a weight ratio of 0.7:0.3 between the standard and specular images; and the shape images undergo thresholding (50-150) and gradient... Canny edge detection and Hough transform processing.

[0019] S3, the asynchronous intelligent analysis module, employs an asynchronous multi-threaded mechanism and utilizes an improved MAFT deep learning enhancement algorithm to perform asynchronous intelligent analysis on the multi-source images preprocessed by S2. The images are written into the database according to three categories: normal, defective, and abnormal. Defective types include dents, indentations, impurities, scratches, bumps, dirt, and watermarks, while abnormal types include lock-hole abnormalities, length and width abnormalities, lack of zinc spraying, and paint accumulation. Simultaneously, a classification detection control signal is generated and sent to S4 and S5 for processing. This signal consists of five variables: image frame number, timestamp, product status, defect type, and confidence level. The product status is composed of three categories: normal, defective, and abnormal. Specific defect dimensions are as follows: Defective products: dents (diameter ≥ 0.3 mm), indentations (length > 1 mm), impurities (diameter > 0.2 mm), scratches (length > 2 mm), bumps (diameter > 0.2 mm), dirt (area > 0.4 mm²), watermarks (area > 0.5 mm²); Abnormal products: lock deformation (angle deviation > 5°), no zinc spraying (coverage < 95%), length and width out of tolerance (± 0.15 mm), paint accumulation (area > 10 mm²).

[0020] Surface defects in microchannel aluminum flat tubes are characterized by low contrast with the background, small feature size, and elongated shape, making them difficult to effectively identify using traditional defect detection methods. While YOLOv12, a high-efficiency target detection algorithm, performs well in real-time industrial scenarios, its original version still has limitations in handling surface defects of microchannel aluminum flat tubes. This application improves the YOLOv12 algorithm by proposing an improved MAFT deep learning enhancement algorithm. This algorithm, tailored to the surface defect characteristics of microchannel aluminum flat tubes, introduces median-enhanced channel and spatial attention blocks (MECS), constructs an adaptive hierarchical feature fusion network (AHFIN), and employs a sliding alignment loss mechanism (SAL). This enhances the sensitivity to subtle changes on the surface of microchannel aluminum flat tubes, addresses defects of different scales on the surface, and improves bounding box localization accuracy, particularly for the identification of elongated defects. Details are as follows:

[0021] To enhance the model's sensitivity to subtle changes on the surface of the microchannel aluminum flat tube, a Median-enhanced Channel and Spatial Attention Block (MECS) was designed, such as... Figure 3 As shown, this module combines channel attention and spatial attention mechanisms to enhance the model's ability to process specific channel information in the input feature map. Specifically, this mechanism expands the receptive field by combining median enhancement with depthwise convolution, thereby generating attention weights that can effectively capture subtle defect features. This method not only improves the accuracy of key feature recognition but also enhances the model's robustness and reliability in complex background environments, thus improving the detection capability of small targets or subtle defects. A detailed explanation follows:

[0022] First, input the feature map. It is a three-dimensional tensor The size is ,in Indicates the number of channels. It is the height of the feature map, and This represents the width. This input feature map is fed into the channel attention module. In this module, the input feature map is first subjected to three different pooling operations: average pooling, median pooling, and max pooling. These pooling operations aim to capture global information from different perspectives, with median pooling being particularly helpful in suppressing noise in complex backgrounds. The pooled data is then fed into a multilayer perceptron (MLP) to generate a channel attention weight matrix. The size of this matrix is Each element in the algorithm represents an importance score for the corresponding channel, reflecting the channel's importance to the current task. This process allows the model to learn which channels are most important for the current task and enhances the information of these channels.

[0023] Then, after processing by the channel attention module, the refined feature map is obtained. Its calculation method is to convert the original feature map With channel attention weights Element-wise multiplication is performed. Then, this refined feature map is further fed into the spatial attention module. Here, our goal is to generate a spatial attention map. Its size is The purpose of spatial attention maps is to identify and highlight key regions or patterns in feature maps that are relevant to the task, thereby enabling the model to focus on important spatial locations.

[0024] Finally, the feature map after double processing is represented as follows: ,here This indicates element-wise multiplication. The purpose of this is to achieve deep fusion between spatial and channel features, thereby enhancing the model's sensitivity to subtle changes and its robustness and localization accuracy in complex scenarios.

[0025] (2) Adaptive hierarchical feature fusion network

[0026] To address the issue of defects at different scales on the surface of microchannel aluminum flat tubes, this application introduces an Adaptive Hierarchical Feature Integration Network (AHFIN). The structure of this network is as follows:

[0027] Input: Multi-scale feature map

[0028] Weighted fusion: Features from each layer are fused using a weighted approach, with the weights obtained through adaptive learning.

[0029]

[0030] in, It is the first The weights of the layer features satisfy the following conditions: , Indicates the first Deformable convolution operations on layers.

[0031] (3) Sliding alignment loss mechanism

[0032] To improve the accuracy of bounding box localization, especially for elongated defects, this application introduces a sliding alignment loss (SAL) mechanism. This mechanism comprehensively considers the differences in position, size, and overlap between the predicted and ground truth bounding boxes.

[0033] Definition: Given a prediction box and real frame SAL is defined as:

[0034]

[0035] in,

[0036] - , These are the weighting coefficients for each type of loss.

[0037] - It refers to the number of samples.

[0038] Specific definitions of each loss:

[0039] Location loss Used to measure the deviation of the center points of the predicted bounding box and the ground truth bounding box in the x and y directions:

[0040]

[0041] in, and Let represent the center coordinates of the predicted bounding box and the ground truth bounding box of the i-th sample, respectively.

[0042] Size loss The difference in size between two frames is measured by calculating the logarithmic difference between their width and height.

[0043]

[0044] in, and Let represent the width and height of the predicted bounding box and the ground truth bounding box of the i-th sample, respectively.

[0045] Cross-combination loss This is used to measure the degree of overlap between the predicted bounding box and the ground truth bounding box.

[0046]

[0047] in, This represents the area of ​​the intersection between the predicted bounding box and the ground truth bounding box. This represents the area of ​​their union.

[0048] This sliding alignment loss design, by comprehensively modeling the positional deviation, size difference, and overlapping areas of the bounding boxes, can more comprehensively optimize the alignment between the predicted and ground truth boxes. Compared to traditional IoU loss or L1 / L2 loss, SAL exhibits stronger robustness and localization accuracy when dealing with elongated, small, or orientation-sensitive targets.

[0049] Optionally, S4, the classification detection and collection module, consists of a bracket, a high-speed rodless cylinder, a flexible spring and classification control system, a lateral classification detection slide rail and a collection box. It receives the classification detection control signal sent by S3 and controls the backward movement of the equipment according to the classification collection algorithm. The defective products are ejected to the left, the abnormal products are ejected to the right, and the normal products pass through. The ejected products are pushed into the collection box by the lateral classification detection slide rail.

[0050] Specifically, the high-speed rodless cylinder has a thrust of 1.2 kN and a response time of ≤0.1 s; the flexible spring is made of beryllium copper alloy with an elastic modulus of 128 GPa and a lifespan of >1.2 × 10⁻⁶ kN. The classification control system has a cycle time of 1ms; the lateral classification detection slide rail speed is no less than 3m / s; the collection box can adjust the collection window according to the size of the microchannel aluminum flat tubes produced, and can also automatically adjust the collection height and provide a full warning. The specific working logic is as follows: Figure 4 As shown.

[0051] The classification and collection algorithm is as follows: S4 receives the classification detection control signal sent by S3. Simultaneously, when the product to be inspected passes through the conveyor belt interval of S1, the grating sensor triggers the classification control system to generate a time trigger Tn, where Tn = L / V, T is the movement time of the product from the conveyor belt interval to the flexible spring, n is the number of time triggers, L is the distance between the conveyor belt interval and the flexible spring, and V is the conveyor belt speed. The classification control system receives the control signal C sent by the asynchronous intelligent analysis module of S3 and writes this signal C into a preset control signal memory queue M[0…n], where len(M) is the queue length and n is the queue number. Simultaneously, the classification control system sorts the control signal queue M according to the timestamp in the control signal of S3 in chronological order. It performs a modulo operation len(M) based on the image frame number in the control signal of S3 to obtain the modulus value N. When the time trigger Tn is triggered, the classification control system reads M[N]. The product status value in the classification and detection control signal is categorized into three types: normal, defective, and abnormal. When the value is normal, the product passes through the S4 classification and detection collection module normally. When the value is defective, the S4 classification and detection collection module ejects the defective product to the left in its running direction. When the value is abnormal, the S4 classification and detection collection module ejects the abnormal product to the right in its running direction.

[0052] Optional, S5, the anomaly handling module, receives the classification detection control signal sent by S3. When the number of consecutive occurrences of a certain type of defective product or a certain type of abnormal product exceeds the threshold, it triggers an audible and visual alarm and automatically sends a shutdown signal to the equipment.

[0053] Specifically, when the anomaly handling module receives the classification detection control signal, if the number of consecutive occurrences of the same type of defect or anomaly exceeds the set threshold (for example, the threshold is set to 5), the module will trigger the red flashing of the three-color light and the buzzer, and send an E-stop signal to the device via the Modbus TCP protocol, and send a deceleration and stop signal to the conveyor belt control module.

[0054] Secondly, this application provides a deep learning-based microchannel aluminum flat tube classification and detection system, such as... Figure 5 As shown, the following technical solution is adopted:

[0055] The system includes a microchannel aluminum flat tube transmission device and multiple image acquisition subsystems, multiple image preprocessing and classification detection subsystems, a microchannel aluminum flat tube intelligent analysis subsystem, a microchannel aluminum flat tube classification detection and collection subsystem, an anomaly handling and control subsystem, and a real-time production analysis dashboard subsystem, specifically:

[0056] Microchannel aluminum flat tube transmission device and multi-image acquisition subsystem

[201] : It realizes the functions of microchannel aluminum flat tube transmission, imaging device deployment and multi-image generation and acquisition. It consists of a transmission device, a two-section transmission belt and a multispectral imaging device. The transmission device is driven by a brushless DC motor and uses a Hall sensor to realize closed-loop speed control (accuracy ±0.05m / s). The two-section transmission belt has a V-shaped guide groove in the front section to realize self-centering and a vacuum adsorption platform in the rear section to eliminate tube vibration. The multispectral imaging device uses a line scan camera, 0.01mm pixel size, three-dimensional point cloud reconstruction accuracy ±5μm, time-frequency flash source, 850nm (penetrating oxide layer) + 650nm (surface texture enhancement), dual-band switching time ≤1μs. The triggering device has a grating trigger with 100kHz pulse / position error ±3μm and anti-electromagnetic interference design (EMC≥60dB). The tube passes through the front conveyor belt at a speed greater than 8 m / s, with the grating triggering a pulse every 0.1 mm; the lower light source completes the 650 nm → 850 nm band switching within 3 μs, and the lower camera simultaneously acquires the bottom surface image; the tube enters the rear vacuum zone, and the upper light source-camera combination acquires the top surface image; a single tube generates 2000 frames of raw images (including at least 5 types such as standard image / diffuse reflection image / specular reflection image / shape image).

[0057] Multi-class image preprocessing and classification detection screening subsystem

[202] : Includes a five-channel parallel image preprocessing engine, which creates five CUDA (Compute Unified Device Architecture) streams, each bound to an independent video memory partition, and uses a double buffering mechanism to achieve parallel operation of current frame processing and next frame preloading. Image preprocessing, such as Figure 6 As shown, the standard image is cropped with ROI to remove 10% of the edge area; the diffuse image is opened using morphological operation with a 3x3 circular kernel; the specular image is denoised using non-local mean with a 15x15 search window; the gloss ratio image is weighted and fused with a weight of 0.7 for the standard image and 0.3 for the specular image; and the shape image is detected using Canny detection with a double threshold of 50 / 150. At the same time, the length and width pixels of the pipe are measured and calculated using the diffuse image and the standard image for initial screening and classification. When the length and width of the pipe exceed the set threshold,

[202] will send an initial screening signal to

[204] and

[205] . This signal includes three variables: image frame number, timestamp, and product status. The product status is divided into three categories: normal, defective, and abnormal. Defective types include pits, indentations, impurities, scratches, bumps, dirt, and watermarks. Abnormal types include abnormal locking, abnormal length and width, no zinc spraying, and spray coating accumulation.

[0058] Microchannel aluminum flat tube intelligent analysis subsystem

[203] : adopts the improved algorithm of deep learning YOLOv12, namely the MAFT deep learning enhancement algorithm. This algorithm introduces MECS attention blocks in the algorithm architecture, which greatly improves the detection rate of micro-dimples; AHFIN feature fusion, which fuses the feature map of the backbone, makes the minimum detection scale of defects more accurate; sliding alignment loss SAL mechanism improves the positioning accuracy of scratch-type defects to the micrometer level. The specific logic is as follows: Figure 7 As shown.

[0059] In terms of defect classification, such as Figure 8The system adopts a three-level classification system of normal, defective, and abnormal. The defective categories include dents, indentations, impurities, scratches, bumps, dirt, and watermarks. The abnormal categories include lock anomalies, length and width anomalies, lack of zinc spraying, and coating accumulation. This facilitates multimodal evidence fusion and improves the classification detection strategy. In model training, a training optimization strategy is adopted. The training data is classified according to the defect classification method of normal, defective, and abnormal. There are no less than 2,000 samples of normal and each defect type. The three common defects of dents, scratches, and lock anomalies are subjected to random occlusion + Gaussian noise, affine transformation + motion blur, and angle offset enhancement processing, respectively. The inference is implemented using an asynchronous multi-threaded architecture, achieving an inference latency of less than 20ms / frame and a throughput of 50 frames / second. The intelligent analysis subsystem sends the processing results to the

[204] and

[205] subsystems in the form of classification detection control signals. The signals consist of five variables: image frame number, timestamp, product status, defect type, and confidence level. The product status is divided into three categories: normal, defective, and abnormal.

[0060] Microchannel aluminum flat tube classification, detection, and collection subsystem

[204] : This subsystem, as the final execution unit of the classification and detection action, achieves millisecond-level classification and detection on high-speed production lines through high-precision timing control and innovative mechanical design, while ensuring zero secondary damage to the tubes. The classification and detection execution mechanism consists of a classification control system, a high-speed rodless cylinder, a beryllium copper alloy flexible spring, a lateral classification and detection slide rail, and an intelligent collection box. The rodless cylinder has a thrust of 1.2kN and a response time of ≤80ms; the flexible spring is made of beryllium copper alloy with an elastic modulus of 128GPa and a fatigue life of 1.2× The PLC control system has a cycle time of 1ms; the lateral sorting and detection slide rail speed is 3m / s; the collection box can adjust the collection window according to the size of the microchannel aluminum flat tubes produced, and can also automatically adjust the collection height and provide a full warning; to prevent secondary damage, nylon cloth is installed around the sorting and detection slide rail to buffer the falling of defective products when they are ejected. The sorting and collection algorithm determines whether the product is normal, defective, or abnormal based on the sorting and detection control signal, and the sorting control system controls the completion of the sorting operation. The specific classification and collection algorithm is as follows: When the product to be inspected passes through the microchannel aluminum flat tube transmission device, the grating sensor triggers the classification control system to generate a time trigger timing Tn, Tn=L / V, where Tn is the simulated movement time of the product to be inspected from the grating sensor to the flexible spring, n is the number of time triggers, L is the distance between the grating sensor and the flexible spring, and V is the conveyor belt speed; the classification control system receives the control signal C sent by

[203] and writes the signal C into the preset control signal memory queue M[0…n], where len(M) is the queue length and n is the queue number; at the same time, the classification control system sorts the control signal queue M according to the timestamp in the control signal in chronological order; the modulus len(M) operation is performed according to the image frame number in the control signal to obtain the modulus value N; when the time trigger Tn is triggered, the classification control system reads M[N]. The product status value in the classification detection control signal is classified into three categories: normal, defective, and abnormal. When the value is normal, the product passes the detection normally

[204] . When the value is defective,

[204] the defective product is ejected to the left in the running direction. When the value is abnormal,

[204] the abnormal product is ejected to the right in the running direction.

[0061] Anomaly Handling and Control Subsystem

[205] :

[203] Sends classification detection control signals to this subsystem. The system stores classification detection control signals in sequence and makes anomaly judgments based on continuous product status and defect type. The system adopts a three-level interlocking protection mechanism. Level I: When the same defect exceeds the threshold continuously, an audible and visual alarm is triggered and the production speed is reduced; Level II: When the critical dimension deviation rate is greater than the threshold, the machine is stopped for calibration and the mold is locked; Level III: When the system hardware fails, an emergency power cut-off and air pressure maintenance are performed. For example, when the indentation defect occurs 5 times in a row, the following is triggered: the three-color light flashes red (frequency 2Hz), sends an E-stop signal to the equipment, and the conveyor belt decelerates linearly to a stop. At the same time, the system also includes the switching and operation status monitoring functions of the conveyor belt device, classification detection and collection device, and image acquisition device.

[0062] Real-time production analysis dashboard subsystem

[206] : Real-time monitoring of production speed, product yield, and image acquisition status of upper and lower cameras. It also realizes defect distribution heat map analysis, defect cluster analysis, report statistical analysis, reserves MES interface and supports work order traceability.

[0063] In summary, this application includes at least one of the following beneficial technical effects:

[0064] 1. A deep learning-based classification and detection method for microchannel aluminum flat tubes, characterized by comprising a microchannel aluminum flat tube time-division stroboscopic image acquisition module, a multi-class image asynchronous processing module, an asynchronous intelligent analysis module, a classification detection and collection module, and an anomaly handling module.

[0065] 2. A deep learning-based microchannel aluminum flat tube classification and detection system, comprising a microchannel aluminum flat tube transmission device and a multi-image acquisition subsystem, a multi-image preprocessing and classification detection screening subsystem, a microchannel aluminum flat tube intelligent analysis subsystem, a microchannel aluminum flat tube classification detection and collection subsystem, an anomaly handling and control subsystem, and a real-time production analysis dashboard subsystem.

[0066] 3. This application proposes a deep learning-based microchannel aluminum flat tube classification and detection method and system, which is an innovative solution designed to address industry pain points. Through time-division stroboscopic image acquisition, asynchronous processing design, classification and detection collection device, and anomaly handling mechanism, and by utilizing an improved deep learning algorithm to achieve parallel detection of multiple defects, it achieves a classification and detection accuracy of >99% and microsecond-level response capability. The overall performance meets the requirements of the production line. Attached Figure Description

[0067] Figure 1 This is a flowchart of a deep learning-based microchannel aluminum flat tube classification and detection method in an embodiment of this application.

[0068] Figure 2 This is the working logic diagram of the time-division flash image acquisition in this application.

[0069] Figure 3 This is a diagram of the median enhancement channel and spatial attention module in this application.

[0070] Figure 4 This is the logic diagram for the classification, detection, and collection of data in this application.

[0071] Figure 5 This is a structural block diagram of a microchannel aluminum flat tube classification and detection system based on deep learning, as described in an embodiment of this application.

[0072] Figure 6 This is the logic diagram of the five-channel parallel image preprocessing in this application.

[0073] Figure 7 This is a logical diagram of the deep learning optimization algorithm framework of this application.

[0074] Figure 8 This is a classification diagram of surface defects in some microchannel aluminum flat tubes according to embodiments of this application.

[0075] Figure reference numerals: 201, Microchannel aluminum flat tube transmission device and multi-image acquisition subsystem; 202, Multi-image preprocessing and classification detection primary screening subsystem; 203, Microchannel aluminum flat tube intelligent analysis subsystem; 204, Microchannel aluminum flat tube classification detection and collection subsystem; 205, Anomaly handling and control subsystem; 206, Real-time production analysis dashboard subsystem. Detailed Implementation

[0076] This embodiment discloses a deep learning-based classification and detection method for microchannel aluminum flat tubes, including a time-division stroboscopic image acquisition module for microchannel aluminum flat tubes, a multi-class image asynchronous processing module, an asynchronous intelligent analysis module, a classification and detection collection module, and an anomaly handling module. The specific implementation is as follows:

[0077] 1. Microchannel aluminum flat tube time-division stroboscopic image acquisition module

[0078] This module adopts a time-division stroboscopic scheme based on a line scan industrial camera. Its core components include:

[0079] (1) Classification and inspection conveyor belt system

[0080] Modular design, with an 8cm interval between the front and rear conveyor belts to form a detection isolation zone, made of PVC material with a thickness of 4.5cm, and the surface is engraved with positioning lines with a precision of 50μm. It adopts a mainstream frequency converter to achieve closed-loop speed control of 8-12m / s and acceleration of 0.5g.

[0081] (2) Line scan camera array

[0082] Industrial high-sensitivity line scan cameras are deployed at both the top and bottom, with a resolution of 8192 pixels; they use 1.4 / 23mm lenses and are angled at 38° with the conveyor belt; the line frequency is 120kHz and the image acquisition accuracy is 0.008mm / pixel.

[0083] (3) Time-division stroboscopic source system

[0084] It adopts a dual-band strobe controller with a wavelength combination of 850nm + 650nm; the light source switching time is ≤800ns; it is equipped with a custom grating trigger (pulse frequency 120kHz); it is installed at a 45° angle; and the Fresnel lens achieves uniform light spot distribution with light intensity uniformity of ±3%.

[0085] (4) Image acquisition and processing system

[0086] Four image acquisition cards with a total bandwidth of 256Gbps; the preprocessing server adopts an x86 architecture and has four 24G GPUs.

[0087] The timing sequence of the time-division stroboscopic control is synchronized by an encoder: the machine encoder outputs 20,000 pulses / revolution, and the frequency is multiplied by 4 to obtain an 80kHz reference frequency; the light source control: after receiving the trigger signal, the light source controller switches the band according to the preset timing sequence, with a delay gradient of 250ns / step; the camera exposure: the exposure time is precisely controlled to 780ns, which completely overlaps with the light source lighting time.

[0088] 2. Asynchronous processing module for multiple image types

[0089] A multi-source image preprocessing pipeline technique was employed. Standard image processing involved ROI cropping: dynamic thresholding to remove 12% of edge regions; Gaussian filtering: a 3×3 adaptive filter with σ=1.2. Diffuse image processing employed morphological operations: 5×5 opening operation → 7×7 closing operation; Blob analysis: area threshold of 15μm², aspect ratio of 1:3 to 3:1. Specular image processing used nonlocal mean denoising: a 21×21 search window, similarity threshold of 0.3; Sobel operator gradient amplification by 1.8 times. Gloss ratio image processing employed weighted fusion: 0.65 for standard image, 0.35 for specular image; CLAHE algorithm, clip_limit=2.0.

[0090] The asynchronous processing architecture uses Kafka message queues to implement multi-threaded processing.

[0091]

[0092] 3. Asynchronous Intelligent Analysis Module

[0093] An improved deep learning algorithm, YOLOv12, is employed for asynchronous multi-threaded processing. The resulting data sends control signals to the classification and detection module and the anomaly handling module. These control signals include five variable parameters: image frame number, timestamp, product status, defect type, and confidence score. The improved YOLOv12 algorithm, with its Median-enhanced Channel and Spatial Attention Block (MECS), divides the feature map into a grid of regions, and each region independently calculates its attention weights, using the following method:

[0094]

[0095] The Sliding Alignment Loss (SAL) uses a dynamic bounding box adjustment coefficient α = 0.85, and the loss function is: L_sal = αL_cls + (1-α)L_iou. The Residual Efficient Layer Aggregation Network (R-ELAN) is implemented as follows:

[0096]

[0097] 4. Classification, detection, and collection module

[0098] (1) Mechanical structure design

[0099] The classification and detection actuator uses a servo valve to control the cylinder, with a push rod speed of 1.5 m / s and an acceleration of 8 m / s.

[0100] Positioning accuracy ±0.2mm, equipped with photoelectric switch limit protection.

[0101] The flexible spring design is made of 304 stainless steel with a thickness of 0.5mm; the preload is 2.5N and the elastic coefficient is 120N / mm; it has a Teflon coating and a coefficient of friction of 0.04.

[0102] (2) Motion control logic

[0103] Example of a PLC control program:

[0104]

[0105] 5. Exception Handling Module

[0106] (1) Three-level interlocking protection mechanism

[0107] Level I protection (five consecutive pieces with the same defect) triggers a red flashing tri-color light (frequency 2Hz); sends an E-stop signal to the cutting machine, and the conveyor belt decelerates linearly to a stop.

[0108] Level II protection (critical dimension out-of-tolerance rate > 2%): shutdown for calibration, mold locking, and generation of maintenance work order; alarm information is pushed to mobile terminal via MES system.

[0109] Level III protection (system hardware failure): emergency power failure, gas line pressure maintenance, safety door locking; fault codes are automatically recorded to the database and a diagnostic report is generated.

[0110] (2) Exception handling process

[0111] Signal reception: Receives classification detection control signals sent by the intelligent analysis subsystem; Anomaly judgment: Stores signals in sequence, triggering Level I protection when 5 consecutive items with the same defect are detected; Alarm mechanism: Three-level alarm system including audible and visual alarm, SMS notification, and system pop-up window; Recovery strategy: Automatically generates maintenance work orders and allows access to remote expert systems.

[0112] This implementation method constructs a complete intelligent classification and inspection solution for microchannel aluminum flat tubes through precise hardware system design, innovative optimization of algorithm modules, and precise timing coordination of the control system. The system has been deployed on the production line of a microchannel aluminum flat tube manufacturer. Actual operation data shows that the product yield has increased by 27%, the classification and inspection efficiency has increased by 3.2 times, and the overall equipment efficiency (OEE) has reached 95.3%.

Claims

1. A deep learning-based microchannel aluminum flat tube classification and detection system, characterized in that, Includes the following modules: The microchannel aluminum flat tube time-division stroboscopic image acquisition module realizes the generation and high-throughput transmission and acquisition of multiple types of time-division stroboscopic images; the multi-type image asynchronous processing module performs preprocessing on the multiple types of images generated by the time-division stroboscopic image acquisition module, including cropping, segmentation, binarization connected component analysis, length measurement, and width measurement. The asynchronous intelligent analysis module employs an asynchronous multi-threaded mechanism and utilizes an improved MAFT deep learning enhancement algorithm to perform asynchronous intelligent analysis on multi-source images preprocessed by the asynchronous image processing module. Simultaneously, it generates a classification detection control signal, which is sent to the classification detection collection module and the anomaly handling module. This signal consists of five variables: image frame number, timestamp, product status, defect type, and confidence level. The classification detection collection module receives the classification detection control signal from the asynchronous intelligent analysis module and controls the movement of the backward equipment according to the classification collection algorithm. Specifically, it controls the movement of defective products to the left, abnormal products to the right, and normal products to pass. The specific classification collection algorithm is as follows: the classification detection collection module receives the classification detection control signal from the asynchronous intelligent analysis module, and simultaneously, when the product to be inspected passes through the microchannel aluminum flat tube time-division stroboscopic image acquisition module... When the conveyor belt intervals, the grating sensor triggers the classification control system to generate a time trigger Tn, where Tn = L / V, T is the movement time of the product to be inspected from the conveyor belt interval to the flexible spring, n is the number of time triggers, L is the distance between the conveyor belt interval and the flexible spring, and V is the conveyor belt speed. The classification control system receives the control signal C sent by the asynchronous intelligent analysis module and writes the signal C into the preset control signal memory queue M[0…n], where len(M) is the queue length and n is the queue number. At the same time, the classification control system sorts the control signal queue M in chronological order according to the timestamp in the control signal from the asynchronous intelligent analysis module. It performs a modulo operation len(M) on the image frame number in the control signal from the asynchronous intelligent analysis module to obtain the modulus value N. When the time trigger Tn is triggered, the classification control system reads the product status value in the classification detection control signal M[N]. This value is classified into three categories: normal, defective, and abnormal. When the value is normal, the inspected product passes through the classification detection collection module normally. When the value is defective, the classification detection collection module ejects the defective product to the left in the direction of movement. When this value is abnormal, the classification detection and collection module will pop the abnormal product to the right of its running direction; the abnormality handling module receives the classification detection control signal sent by the asynchronous intelligent analysis module. When the number of consecutive occurrences of a certain type of defective product and a certain type of abnormal product exceeds the threshold, it will trigger an audible and visual alarm and automatically send a shutdown signal to the equipment.

2. The microchannel aluminum flat tube classification and detection system based on deep learning according to claim 1, wherein the asynchronous intelligent analysis module has the following characteristics: The asynchronous intelligent analysis module employs an asynchronous multi-threaded mechanism and utilizes an improved MAFT deep learning enhancement algorithm to perform asynchronous intelligent analysis on multi-source images preprocessed by the asynchronous image processing module. Images are then written into a database according to three categories: normal, defective, and abnormal. Defective types include dents, indentations, impurities, scratches, bumps, dirt, and watermarks; abnormal types include lock-edge abnormalities, length and width abnormalities, lack of zinc spraying, and coating buildup. Simultaneously, a classification detection control signal is generated and sent to the classification detection collection module and the abnormality processing module. The improved MAFT deep learning augmentation algorithm introduces median-enhanced channel and spatial attention blocks (MECS), constructs an adaptive hierarchical feature fusion network (AHFIN), and employs a sliding alignment loss mechanism (SAL), as follows: First, the input feature map... It is a three-dimensional tensor The size is ,in Indicates the number of channels. It is the height of the feature map, and Representing the width, this input feature map is fed into the channel attention module; In this module, the input feature map is first subjected to three different pooling operations: average pooling, median pooling, and max pooling. The pooled data is then fed into a multilayer perceptron (MLP) to generate a channel attention weight matrix. The size of the matrix is Each element in the graph represents the importance score of the corresponding channel; then, after processing by the channel attention module, a refined feature map is obtained. Its calculation method is to convert the original feature map With channel attention weights After element-wise multiplication, the refined feature map is further fed into the spatial attention module to generate a spatial attention map. Its size is Finally, the feature map after double processing is represented as follows: ,here This represents element-wise multiplication; the structure of the adaptive hierarchical feature fusion network is as follows: Input: Multi-scale feature map Weighted fusion: Features from each layer are fused using a weighted approach, with the weights obtained through adaptive learning. in, It is the first The weights of the layer features satisfy the following conditions: , Indicates the first Deformable convolution operations on layers.

Images