A method and system for automatically identifying and positioning a battery protection plate feeding device

Abstract

The application discloses a kind of automatic identification positioning battery protection plate feeding device method and system, relating to battery protection plate feeding intelligent visual positioning technology, first acquisition is carried the global visual image of battery protection plate tray, the collected image is adaptively anti-reflection pretreatment, eliminate metal pin and the anti-reflection noise of board surface;Respectively extract the feature of preset reference mark on tray, and the pin profile of battery protection plate, board printing mark feature, compare the reference mark feature extracted with pre-stored tray standard template, calculate tray placement offset, combined with homography transformation complete overall image deformation correction.The present application can effectively eliminate the interference of reflection, automatically correct the positioning error caused by tray offset and protection plate warping deformation, and the positioning accuracy and stability are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of visual positioning technology for battery protection plate loading, and in particular to a method and system for automatically identifying and positioning battery protection plate loading devices. Background Technology

[0002] As a core safety control component of lithium battery packs, battery protection boards are widely used in the manufacturing processes of 3C digital batteries, power lithium batteries, and energy storage lithium batteries. Their loading process is the foundation for subsequent spot welding, performance testing, and PACK assembly. In recent years, the global lithium battery industry has seen rapid capacity expansion and continuous improvement in production line automation and intelligence. This has led to increasingly higher core demands from the industry regarding the positioning accuracy, operational efficiency, multi-model compatibility, and full-process traceability of the protection board loading process.

[0003] One of the mainstream battery protection board loading and positioning solutions in the industry is the traditional machine vision grayscale template matching scheme. This scheme uses a fixed-mount area array industrial camera with direct illumination to capture images of the material tray, pre-stores grayscale feature templates for the corresponding model of protection board, and compares the similarity between the captured image and the template through a cross-correlation matching algorithm to obtain the positioning coordinates of the protection board, which are then sent to the loading robot for picking. This scheme has low deployment costs and simple core algorithm logic, and is widely used in production lines with single production models and high material tray positioning accuracy. However, this scheme has obvious shortcomings in practical applications: the metal pins of the battery protection board are prone to strong reflections when exposed to direct illumination or ambient light, resulting in pixel overexposure and loss of features, with a feature matching error rate of up to 10%; it lacks material tray offset correction and protection board warping and tilt correction mechanisms, and the positioning error can reach more than 2mm when the material tray is placed incorrectly or the protection board is deformed, which can easily lead to picking misalignment and pin bending and scrapping; when switching production models, it is necessary to manually re-capture the template and adjust the matching threshold, which can take up to 30 minutes per switch, resulting in extremely low adaptation efficiency.

[0004] Another mainstream solution is a laser contour scanning-based positioning method. This method uses a line laser displacement sensor to scan along the plane of the material tray, collecting the 3D contour point cloud of all protection boards, and obtaining the positioning coordinates through contour feature matching. This method is not affected by visible light reflection and has high positioning accuracy, and is often used in power battery protection board production lines with high precision requirements. However, this method also has inherent drawbacks: full-width scanning of a single material tray can take more than 30 seconds, and the operating efficiency is much lower than that of vision solutions, making it unsuitable for high-speed production lines with a cycle time of more than 1200 pieces / hour; the equipment purchase and maintenance costs are 3-5 times that of vision solutions; the laser lens is prone to dust accumulation and requires regular manual cleaning, resulting in high operation and maintenance costs; it can only identify contour features and cannot read the marking information printed on the board, so it cannot achieve full-process product traceability, nor can it identify different models of protection boards mixed together, resulting in insufficient adaptability.

[0005] As lithium battery production lines upgrade towards higher speeds, greater flexibility, and traceability, the two existing positioning solutions have shortcomings in balancing efficiency, accuracy, cost, and adaptability, failing to meet the actual production needs of the current lithium battery industry. The industry urgently needs new positioning technologies to address these pain points. Summary of the Invention

[0006] The present invention proposes a method and system for automatically identifying and positioning battery protection plate loading devices, in order to solve the problems mentioned in the prior art that the existing battery protection plate loading and positioning schemes cannot simultaneously achieve accuracy, efficiency, cost and traceability adaptability.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a method for an automatic identification and positioning battery protection plate loading device, comprising the following steps: S1. Using a global industrial camera installed directly above the material tray to be loaded, capture the global original image of all battery protection boards carried on the tray under diffuse lighting conditions. The resolution of the global original image is not less than 1920×1080. At the same time, use a close-range industrial camera to capture high-resolution positioning images of the AprilTag reference positioning marks preset at the four corners of the tray. The resolution of the positioning images is not less than 2592×1944. S2. First, perform grayscale conversion and noise filtering on the global original image, then perform adaptive de-reflection preprocessing to eliminate the reflective interference of the metal pins of the battery protection board. After that, perform directional feature enhancement processing on the pin edge features and printed marking features of the battery protection board to obtain an image to be identified with contrast that meets the matching requirements. S3. Call the pre-stored standard feature template of the corresponding battery protection board, use the normalized cross-correlation matching algorithm to traverse the image to be identified, and select candidate boxes with a matching degree higher than 0.8 as the initial positioning area of ​​each battery protection board. After performing homography deformation correction on the initial positioning area to eliminate warping and tilting errors, extract the pin contour features and board identification features of each battery protection board. S4. First, decode the positioning image of the reference positioning mark, calculate the three-dimensional offset between the actual pose of the material tray and the preset standard pose to complete the material tray offset compensation. Then, based on the extracted pin contour features and board marking features, combine the pixel-physical coordinate transformation coefficient to calculate the six-degree-of-freedom actual pose parameters of each battery protection board, and package them to generate the corresponding loading coordinate instructions that conform to the industrial communication protocol. S5. The loading coordinate command is sent to the controller of the four-axis SCARA loading robot through the real-time communication link. The robot adjusts the nozzle posture and picks up the corresponding battery protection board, completing the high-precision loading and positioning operation of the battery protection board.

[0008] Preferably, the adaptive de-reflection preprocessing in step S2 specifically involves: firstly, calculating the grayscale histogram of the global original image, taking the grayscale value corresponding to the 95th percentile of the grayscale histogram as the reflection threshold, using OTSU adaptive threshold segmentation to obtain the initial reflection region, then removing isolated noise points through morphological opening operations to obtain continuous high-reflection regions, extracting the average grayscale value of non-reflective pixels within a neighborhood of 3 pixels extending outward from the high-reflection region as the filling value, and performing pixel smoothing filling on the high-reflection region, retaining the edge information of the region boundary during the filling process to avoid generating false edges that interfere with subsequent feature extraction.

[0009] Preferably, the directional feature enhancement in step S2 specifically involves: pre-setting the enhancement weight of the pin contour feature to 1.8 and the enhancement weight of the board identification feature to 1.5; firstly, traversing the entire original image through a size matching sliding window to select candidate areas whose size conforms to the preset size range of the corresponding model battery protection board as the areas to be enhanced; performing feature classification on the pixels in the areas to be enhanced; increasing the gradient magnitude of pixels belonging to the pin edge; increasing the texture contrast of pixels belonging to the board identification; and keeping the contrast of background pixels in non-areas to be enhanced unchanged to avoid background interference features being mistakenly enhanced.

[0010] Preferably, the deformation correction in step S3 specifically involves: firstly, identifying the sub-pixel level corner coordinates of the four preset square positioning marks on the surface of the battery protection board within the initial positioning area; calculating the deviation between the actual relative positions of the four corner marks and the relative positions of the four positioning marks in the preset standard template; calculating the homography transformation matrix based on the least squares method; and mapping and correcting all pixel coordinates of the entire initial positioning area through the homography transformation matrix. After correction, the relative position error of the four positioning marks is less than 0.1 pixels, completely eliminating the deformation error caused by warping and tilting of the battery protection board within the tray.

[0011] Preferably, the actual pose parameter calculation in step S4 is specifically as follows: Canny edge detection is performed on the initial positioning area after deformation correction to extract the continuous closed edge contour of the pin, calculate the minimum bounding rectangle of the contour, obtain the center pixel coordinates and long side rotation angle of the minimum bounding rectangle, combine the three-dimensional offset obtained by the material tray offset compensation and the pre-calibrated pixel-physical coordinate conversion coefficient to convert the pixel coordinates into X, Y, and Z physical coordinates in the world coordinate system, and add the rotation angle parameter and the nozzle control signal to jointly generate a loading coordinate instruction conforming to the Modbus-RTU protocol. The instruction includes a unique number corresponding to the battery protection board for subsequent traceability.

[0012] Preferably, after step S5, a positioning effect verification step is also included: when the loading robot picks up the battery protection board and moves it to the verification station, the verification camera is triggered to acquire a high-definition top-view image of the battery protection board, extract the pin outline and board body identification features in the image, calculate the deviation between the actual pose after picking up and the pre-picked pose, if the deviation is greater than the preset 0.5mm accuracy threshold, it is determined that the picking accuracy does not meet the requirements, the currently acquired qualified battery protection board image is added to the template training set, the protection board feature template is updated using an incremental learning method, and the reference offset parameters of the material tray are updated synchronously to compensate for the positioning parameters of subsequent battery protection boards in the same batch, so as to avoid batch positioning errors.

[0013] Preferably, a system for automatically identifying and positioning a battery protection plate loading device includes: The image acquisition module consists of a 2K resolution global CMOS industrial camera, a diffuse ring light, and a 5-megapixel close-range industrial camera. The global industrial camera is installed directly above the material tray to be loaded, and it acquires the global original image of all the battery protection boards carried on the tray. The diffuse ring light can automatically adjust its brightness according to the ambient light intensity. The close-range industrial camera is aimed at the AprilTag reference positioning marks at the four corners of the tray, and it acquires high-resolution positioning images of the reference positioning marks. The preprocessing module is equipped with an adaptive de-glare algorithm unit and a directional feature enhancement algorithm unit. The adaptive de-glare algorithm unit completes the identification and smooth filling of the reflective areas of the global original image, while the directional feature enhancement algorithm unit performs directional contrast enhancement on the pin contour features and board marking features of the battery protection board, and outputs the image to be identified that meets the matching requirements. The feature extraction and correction module is equipped with a pre-trained feature template library for multiple battery protection boards. It uses a normalized cross-correlation matching algorithm to match the image to be identified to obtain the initial positioning area of ​​each battery protection board. Then, it calls the homography transformation correction unit to complete the deformation correction of the initial positioning area and extracts the corrected pin contour features and board identification features. The pose calculation module is equipped with a reference mark decoding unit, an offset compensation unit, and a pixel-physical coordinate transformation unit. First, it decodes the reference positioning mark to obtain the three-dimensional offset of the material tray and completes the offset compensation. Then, it combines the transformation coefficient to calculate the actual pose parameters of the six degrees of freedom of each battery protection board and generates the corresponding loading coordinate command. The instruction output module supports multi-protocol industrial communication and sends the loading coordinate instructions to the controller of the loading actuator in real time, thereby controlling the actuator to complete the loading operation.

[0014] Preferably, the global industrial camera of the image acquisition module is installed 1.2m directly above the material tray to be loaded, with the distance between the edge of the field of view and the edge of the tray not less than 5cm, to avoid incomplete image acquisition caused by tray offset. The frame rate of the global industrial camera is not less than 30fps, supporting real-time acquisition during dynamic loading. The close-range industrial camera is installed 0.8m diagonally above the tray, with each reference positioning mark configured with an infrared supplementary light to avoid interference from visible light reflection on AprilTag recognition. The recognition accuracy of the reference positioning mark is not less than 99.99%.

[0015] Preferably, it also includes a template update module, which consists of a verification image acquisition unit, a precision comparison unit, and a template iteration unit. The verification image acquisition unit is installed at the verification station next to the moving path of the feeding actuator to acquire high-definition images of the picked-up battery protection board. The precision comparison unit compares the acquired image features with the pre-stored standard features and calculates the picking deviation. When the number of deviations exceeding the preset threshold reaches 3 times, the template iteration unit is triggered to add the features of the latest acquired qualified battery protection board to the template library, update the protection board feature template using an incremental learning method, and recalculate the reference offset parameters of the material tray and synchronize them to the pose calculation module.

[0016] Preferably, the pose calculation module further includes an adaptive communication unit. The adaptive communication unit supports multiple industrial communication protocols such as Modbus-RTU, TCP / IP, and PROFINET. It can automatically identify the communication protocol type of the feeding actuator controller and establish a matching communication link. The communication delay is no higher than 10ms. It supports real-time batch issuance of feeding coordinate commands and can receive the feeding actuator's feedback information on successful pickup, failed pickup, and material shortage status. It synchronizes the status information to the production line MES system to achieve full traceability of the feeding process. The communication link supports reconnection after disconnection. It can automatically re-establish the connection within 3 seconds after disconnection to ensure the continuity of the feeding process.

[0017] Compared with existing technologies, the beneficial effects of this invention are: This invention effectively eliminates reflective interference from the metal pins of the battery protection board by combining adaptive de-reflection preprocessing with directional feature enhancement. It specifically enhances the core features required for positioning, such as pin contours and board markings. At the same time, it is combined with a material tray reference mark offset compensation and homography deformation correction mechanism to automatically correct positioning errors caused by material tray placement deviation and protection board warping and tilting. This effectively solves the problems of lost reflective features and insufficient positioning accuracy caused by the lack of correction mechanism in existing traditional visual positioning solutions, and significantly improves positioning stability.

[0018] This invention utilizes a pre-stored multi-model feature template combined with incremental learning for automatic template iteration. When switching models, the corresponding parameters can be directly called without manual annotation and debugging. At the same time, it adopts a high-speed global vision acquisition combined with parallel computing processing logic, resulting in fast positioning and recognition speed. It can adapt to the cycle time requirements of high-speed production lines and effectively solves the problems of high cost of model change and debugging of existing traditional vision solutions and slow recognition speed of laser contour scanning solutions that cannot adapt to high-speed production lines, thus greatly improving work efficiency.

[0019] This invention adopts a pure vision hardware architecture, eliminating the need for expensive laser scanning equipment and resulting in lower deployment and maintenance costs. It can simultaneously identify printed markings on the board and reference markings on the tray. Combined with a multi-protocol adaptive communication unit, it can interface with various industrial control equipment and production management systems to achieve full traceability of the material loading process. This effectively solves the problems of high cost, inability to identify markings for traceability, and insufficient adaptability of existing laser positioning solutions.

[0020] This invention can cover the material loading and positioning scenarios of all types of battery protection boards, including 3C consumer soft-pack batteries, power lithium batteries, and energy storage lithium batteries. It is compatible with protection boards of different sizes and pin specifications, and can be quickly deployed without major modifications to the existing production line structure. It has extremely high promotion and application value in the field of lithium battery production automation. Attached Figure Description

[0021] Figure 1 This is a complete flowchart of the overall operation of the lithium battery protection board safety system proposed in this invention. Figure 2 This is a flowchart of the multi-source cell parameter synchronous acquisition and dynamic frequency control proposed in this invention; Figure 3 This is a flowchart of the feature vector construction and fault center similarity matching process proposed in this invention; Figure 4 This is a flowchart illustrating the response mechanism mapping and three-level hierarchical protection action execution proposed in this invention. Figure 5 This is the fault-tolerant control watchdog monitoring and multi-board cross-bus synchronous networking control diagram proposed in this invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Reference Figures 1 to 5This invention discloses a method for an automatic identification and positioning battery protection plate loading device, comprising the following directly implementable steps: First, complete the hardware deployment: Install a 2-megapixel global CMOS industrial camera (Hikvision MV-CA020-10GC) on an adjustable aluminum profile bracket 1.2m above the material tray to be loaded, and install a 120mm diameter diffused ring light. The brightness of the light can be adaptively adjusted within the range of 0-1000 lux to avoid direct glare. Then, align four 5-megapixel close-range grayscale industrial cameras (Hikvision MV-CA050-10GM) with the 20mm×20mm AprilTag 36h11 reference positioning marks affixed to the four corners of the material tray. Each close-range camera is equipped with a 1W 850nm infrared light to avoid visible light interference with mark recognition.

[0024] S1. After the tray arrival sensor (diffuse reflection photoelectric switch) detects that the tray has been pushed to the preset loading station, it sends a trigger signal to the camera controller. The global industrial camera and four close-range industrial cameras expose simultaneously. The exposure time of the global industrial camera is set to 20ms to acquire a global original image covering the entire tray with a resolution of 1920×1080. The exposure time of the close-range industrial cameras is set to 15ms to acquire high-resolution positioning images of the four reference positioning markers with a resolution of 2592×1944. All images are transmitted to the edge computing box (NVIDIA Jetson Xavier NX) via Gigabit Ethernet for subsequent processing.

[0025] S2. The edge computing box first performs grayscale conversion on the global original image. The grayscale conversion formula is: Where R is the grayscale value of the red channel pixel, G is the grayscale value of the green channel pixel, B is the grayscale value of the blue channel pixel, and Gray is the grayscale value of the single-channel image after conversion. After grayscale conversion, a 3×3 Gaussian kernel is used to perform Gaussian filtering to remove image noise. The formula for calculating the Gaussian kernel is: Where x and y are the relative coordinates of the pixels within the kernel, and σ is the standard deviation of the Gaussian distribution. In this embodiment, σ = 1.5 is used. After filtering, adaptive de-reflection preprocessing is performed to eliminate the reflective interference of the metal pins. Then, directional feature enhancement processing is performed on the pin edge features and the printed marking features of the board to obtain the image to be identified with contrast that meets the matching requirements.

[0026] S3. The standard feature template of the corresponding battery protection board, pre-stored within the edge computing box, is invoked. The Normalized Cross-Correlation (NCC) matching algorithm is used to traverse the image to be identified. The NCC calculation formula is as follows: Where I is the sliding window region of the image to be recognized, and T is the standard feature template. The average pixel value of the sliding window region. The average pixel value of the template is (x,y), the pixel coordinates within the sliding window are (x',y'), and the pixel coordinates within the template are (x',y'). After traversal, candidate boxes with NCC values ​​higher than 0.8 are selected as the initial positioning areas for each battery protection board. After homography deformation correction is performed on the initial positioning areas to eliminate warping and tilting errors, the HOG feature extraction algorithm is used to extract the pin contour features and board identification features of each battery protection board.

[0027] S4. First, the AprilTag recognition library is called to decode the positioning images of the four reference positioning tags, calculate the X, Y, and Z three-dimensional offsets between the actual pose of the material tray and the preset standard pose, and complete the material tray offset compensation. Then, based on the extracted pin contour features and board marking features, combined with the pixel-physical coordinate transformation coefficients obtained in advance by Zhang Zhengyou calibration method, the actual pose parameters of the six degrees of freedom of each battery protection board are calculated. The pose parameters are packaged to generate loading coordinate instructions that conform to the Modbus-RTU industrial communication protocol. Each instruction is accompanied by a unique ID of the corresponding battery protection board, which is convenient for subsequent full-process traceability.

[0028] The S5 edge computing box sends the loading coordinate instructions to the controller of the four-axis SCARA loading robot (Yamaha YK400XR) in real time via RS485 industrial bus. After receiving the instructions, the controller adjusts the X, Y, Z coordinates and rotation angle of the suction nozzle, controls the anti-static rubber suction nozzle to generate a negative pressure of -70kPa to pick up the corresponding battery protection board, moves it to the subsequent processing station, and then disconnects the negative pressure to complete the placement, realizing the high-precision loading and positioning operation of the battery protection board.

[0029] This invention also discloses a specific implementation method for adaptive deglare preprocessing: First, a gray-level histogram is calculated for the global original image after gray-level filtering. The gray-level value corresponding to the 95th percentile of the gray-level histogram is taken as the initial glare threshold. The OTSU adaptive threshold segmentation algorithm is used to perform binarization segmentation on the image to obtain the initial glare region. The inter-class variance calculation formula of the OTSU algorithm is: ;in The percentage of background pixels. The average gray level of the background pixels. The percentage of pixels in the foreground (reflective area). The average gray level of the foreground pixels. The average gray level of the entire image, when The grayscale value corresponding to the maximum value is the optimal segmentation threshold. After segmentation, morphological opening operations are performed using 3×3 rectangular structuring elements to remove isolated noise pixels, resulting in continuous highly reflective regions. The average grayscale value of non-reflective pixels within a 3-pixel neighborhood extending outward from the highly reflective region is extracted as the fill value. Bilinear interpolation is used to smoothly fill the pixels within the highly reflective region. Before filling, the Canny edge detection algorithm is used to extract the edge pixels of the highly reflective region. During the filling process, the edge pixels retain their original values ​​to avoid generating false edges that interfere with subsequent feature extraction. This step can be implemented using the following core Python code: import cv2 import numpy as np # Read the original image and convert it to grayscale img_gray = cv2.cvtColor(img_raw, cv2.COLOR_BGR2GRAY) # Calculate the 95th percentile reflectivity threshold reflect_thresh = np.percentile(img_gray, 95) # OTSU divides reflective areas _, binary_reflect = cv2.threshold(img_gray, reflect_thresh, 255,cv2.THRESH_BINARY+cv2.THRESH_OTSU) # Morphological opening operation denoising kernel = np.ones((3,3), np.uint8) binary_reflect = cv2.morphologyEx(binary_reflect, cv2.MORPH_OPEN,kernel) # Calculate neighborhood mean fill mask = binary_reflect == 255 neighborhood = cv2.dilate(mask.astype(np.uint8), kernel, iterations=3) neighbor_pixels = img_gray[(neighborhood == 1)&(~mask)] fill_value = np.mean(neighbor_pixels) img_gray[mask] = fill_value This invention also discloses a specific implementation method for targeted feature enhancement: enhancement weights are pre-set according to the feature attributes of the battery protection board, wherein the enhancement weight of the metal pin contour feature is set to 1.8 and the enhancement weight of the printed marking feature on the board is set to 1.5; firstly, a size-matching sliding window is used to traverse the entire original image, the sliding window size is consistent with the pixel size of the corresponding model of battery protection board, the sliding window movement step is set to 2 pixels, and candidate areas whose size conforms to the preset size of the protection board within ±5% range are selected as the areas to be enhanced, and the background areas that are not to be enhanced are not enhanced; The pixels within the enhancement area are classified using a pre-trained lightweight MobileNetV2 binary classification model. The output categories are pin area, board marking area, and other areas. Contrast enhancement is applied to pixels of different categories using corresponding weights. The enhancement formula is as follows: ,in is the original pixel grayscale value, and w is the enhancement weight for the corresponding category. To enhance pixel grayscale values, grayscale values ​​exceeding 255 are truncated to 255, and grayscale values ​​below 0 are set to 0 to avoid pixel overflow.

[0030] This invention also discloses a specific implementation method for deformation correction of the initial positioning area: First, the cv2.goodFeaturesToTrack corner detection algorithm is called to identify the corner coordinates of the four preset 2mm×2mm square positioning marks on the surface of the battery protection board within the initial positioning area. Then, the cv2.cornerSubPix algorithm is called to perform sub-pixel optimization on the corner coordinates, with an optimization accuracy of up to 0.01 pixels. The deviation between the actual relative positions of the four sub-pixel corners and the relative positions of the four positioning mark corners in the preset standard template is calculated. Based on the least squares method combined with the RANSAC robust estimation algorithm, the 3×3 homography transformation matrix H is calculated. The mapping formula for homography transformation is: Where x and y are the pixel coordinates before correction, x' and y' are the pixel coordinates after correction, and H is the homography transformation matrix, which includes rotation, translation, scaling, and deformation parameters. The obtained homography transformation matrix is ​​used to map and correct all pixel coordinates in the entire initial positioning area. After correction, the reprojection error is calculated. When the reprojection error is less than 0.1 pixels, the correction is considered qualified, which can completely eliminate the deformation error caused by warping and tilting of the battery protection board in the tray. If the reprojection error exceeds the threshold, template matching is re-executed to avoid correction errors.

[0031] This invention also discloses a specific implementation method for calculating the actual pose parameters of the battery protection board: Canny edge detection is performed on the initial positioning area after deformation correction. The low threshold of the Canny algorithm is set to 50, and the high threshold is set to 150. Continuous closed edge contours of the pins are extracted. The cv2.findContours algorithm is called to filter contours whose perimeter conforms to the preset range of the corresponding model pin contour. Then, the cv2.minAreaRect algorithm is called to calculate the minimum bounding rectangle of the contour, obtaining the center pixel coordinates, long side rotation angle, and length and width parameters of the minimum bounding rectangle. Combining the three-dimensional offset obtained from the tray offset compensation with the pre-calibrated pixel-physical coordinate transformation coefficient, the following formula is used to convert it into physical pose parameters in the world coordinate system: ; ; ; ;in , , Pick coordinates in the world coordinate system. , The center pixel coordinates of the smallest bounding rectangle. , The coordinates of the principal points of the camera intrinsic parameter matrix. , This refers to the physical size corresponding to a single pixel (0.1 mm / pixel in this embodiment). , , This is the three-dimensional offset obtained from the material tray offset compensation. The preset standard height of the tray. The rotation angle of the long side of the minimum bounding rectangle; the above pose parameters are packaged into a Modbus-RTU protocol frame. The frame structure is as follows: 1 byte address code, 1 byte function code (0x10), 2 bytes register start address, 2 bytes register quantity, 1 byte data length, 16 bytes pose data (4 bytes each of X, Y, and Z floating-point type, Rz is 2 bytes integer type, unit is 0.1°, 2 bytes protection board ID), 2 bytes CRC16 check code. After packaging, it is ready to be sent.

[0032] In this invention, a specific implementation method for verifying the positioning effect is also disclosed: after the loading robot picks up the battery protection board, when it moves to the preset verification station, the verification camera is triggered to collect a high-definition image of the battery protection board from above, extract the pin outline and board body identification features in the image, calculate the deviation between the actual pose after picking up and the pre-picked pose, and if the deviation is greater than the preset 0.5mm accuracy threshold, it is determined that the picking accuracy does not meet the requirements. When the picking accuracy fails to meet the requirements three times in a row, the HOG features of the currently acquired qualified battery protection board images are added to the template training set. The protection board feature template is updated using an incremental learning method. During the update, the weight of the new features is set to 0.3, and the weight of the original template features is set to 0.7 to avoid template drift. At the same time, the average offset of the last 10 tray offset compensations is recalculated, the tray reference offset parameters are updated, and the positioning parameters of subsequent battery protection boards in the same batch are automatically compensated to avoid batch positioning errors.

[0033] This invention discloses a system for automatically identifying and positioning battery protection plate loading devices, specifically including the following hardware and software units: The image acquisition module consists of one 2-megapixel global CMOS industrial camera, one diffuse ring light, four 5-megapixel close-range grayscale industrial cameras, and four 850nm infrared lights. The global industrial camera is mounted on an adjustable bracket 1.2m above the material tray to be loaded, covering the entire tray and used to acquire the global raw image of all battery protection boards carried on the tray. The diffuse ring light is linked with an ambient light sensor and can automatically adjust the brightness of the light according to the ambient light intensity to avoid direct glare. The four close-range industrial cameras are respectively aimed at the AprilTag reference positioning markers at the four corners of the tray to acquire high-resolution positioning images of the reference positioning markers. All cameras are connected to the edge computing box via gigabit Ethernet to receive trigger signals and transmit the acquired images back.

[0034] The preprocessing module, running within the edge computing box, is equipped with an adaptive de-glare algorithm unit and a directional feature enhancement algorithm unit. The adaptive de-glare algorithm unit performs reflective area identification, segmentation, and smooth filling of the global original image to eliminate reflective interference from metal pins. The directional feature enhancement algorithm unit performs directional contrast enhancement on the pin contour features and board marking features of the battery protection board to suppress background noise and output an image to be identified that meets the matching requirements.

[0035] The feature extraction and correction module runs within the edge computing box and is equipped with a pre-trained feature template library for multiple battery protection boards. It supports the storage and retrieval of up to 100 different battery protection board templates. It uses a normalized cross-correlation matching algorithm to match the image to be identified to obtain the initial positioning area of ​​each battery protection board. Then, it calls the homography transformation correction unit to complete the deformation correction of the initial positioning area, eliminating warping and tilting errors. The HOG feature extraction algorithm is used to extract the corrected pin contour features and board identification features.

[0036] The pose calculation module runs within the edge computing box and includes a reference marker decoding unit, an offset compensation unit, and a pixel-to-physical coordinate transformation unit. The reference marker decoding unit calls the AprilTag recognition library to decode the positioning image of the reference positioning marker and obtain the three-dimensional offset of the material tray. The offset compensation unit combines the offset to compensate for the positioning coordinates, eliminating the error caused by the material tray placement offset. The pixel-to-physical coordinate transformation unit converts the pixel coordinates into physical pose parameters in the world coordinate system based on the pre-calibrated camera intrinsic and extrinsic parameters, and packages them to generate loading coordinate instructions that conform to the industrial communication protocol.

[0037] The instruction output module adopts an industrial-grade RS485 and Ethernet interface, supports multi-protocol industrial communication, and sends the loading coordinate instructions to the controller of the loading actuator in real time, so as to control the actuator to complete the loading operation.

[0038] This invention also discloses the specific parameter configuration of the image acquisition module: a global industrial camera is installed 1.2m directly above the material tray to be loaded, with a field of view of 700mm×500mm. The standard tray size is 600mm×400mm, and the distance between the edge of the field of view and the edge of the tray is not less than 5cm to avoid incomplete image acquisition caused by tray offset; the global industrial camera has a frame rate of not less than 30fps, supporting real-time acquisition during dynamic loading, and clear images can be acquired even if the tray moves slightly; four close-range industrial cameras are installed 0.8m above the four opposite corners of the tray, and each reference positioning mark is equipped with an 850nm infrared supplement light with adjustable brightness to avoid interference from visible light reflection on AprilTag recognition. The recognition accuracy of the reference positioning mark is not less than 99.99%, and the single recognition time is less than 1ms.

[0039] This invention also discloses a specific implementation of the template update module: The template update module consists of a verification image acquisition unit, a precision comparison unit, and a template iteration unit. The verification image acquisition unit is a 2-megapixel industrial camera installed at the verification station next to the moving path of the feeding actuator. When the nozzle picking up the protection board passes through the verification station, the acquisition is triggered to obtain a high-definition top-view image of the battery protection board. The precision comparison unit compares the acquired image features with the pre-stored standard features and calculates the picking deviation. When the number of deviations exceeding the preset threshold reaches 3 times, the template iteration unit is triggered to perform weighted fusion of the HOG features of the newly acquired qualified battery protection board with the features of the original template. The updated template needs to be verified by the test set, and it will be officially used only if the matching accuracy is not less than 99%. At the same time, the reference offset parameters of the material tray are recalculated and synchronized to the pose calculation module. The update process takes less than 100ms and will not affect the normal operation of the production line.

[0040] This invention also discloses a specific implementation of the adaptive communication unit: the adaptive communication unit mounted on the pose calculation module supports three mainstream industrial communication protocols: Modbus-RTU, TCP / IP, and PROFINET. It can automatically identify the communication protocol type of the feeding actuator controller. The identification logic is to send the corresponding protocol probe frames in sequence. After receiving a valid reply, a matching communication link can be established. The average communication delay is no higher than 8ms, and the maximum delay is no more than 10ms. It supports real-time batch issuance of feeding coordinate commands, and can issue feeding commands for up to 20 battery protection boards at a time, meeting the batch feeding requirements of multiple protection boards on a single tray. At the same time, it can receive status information such as successful pickup, failed pickup, material shortage, and empty tray returned by the feeding actuator. The status information is synchronized to the production line MES system via the MQTT protocol, and a feeding process log is automatically generated. Each log includes the unique ID of the corresponding battery protection board, pickup time, pose parameters, and feeding result information, realizing full traceability of the feeding process. The communication link supports an automatic reconnection mechanism after disconnection. After disconnection, a probe frame is sent every 500ms, and the communication connection can be automatically re-established within 3 seconds. After the connection is restored, the unsent feeding instructions are automatically resent, so there will be no loss of instructions, ensuring the continuity of the feeding process and restoring normal operation without manual intervention.

[0041] Scenario Example 1: Mass production of protection boards for consumer-grade soft-pack batteries This embodiment is applied to the spot welding and feeding station of soft-pack battery protection boards in a 3C digital battery production workshop. A single tray carries 100 soft-pack battery protection boards with a size of 50mm×20mm. The tray is made of black ABS material. The nickel pins of the protection boards are easily affected by direct sunlight from the fluorescent lights on the top of the workshop, which can cause strong reflections. The tray is manually pushed into place, and the maximum offset in a single placement can reach 5mm. The original traditional template matching and positioning scheme has an error rate of 8% due to the interference of reflections. When the tray is offset, it is necessary to manually adjust the position repeatedly. The feeding accuracy is only 92%, and only 1200 pieces can be fed per hour. It is necessary to arrange a dedicated person to be on duty to replenish the material, resulting in low production efficiency.

[0042] When deploying the system of this invention, a global industrial camera is installed 1.2m above the material tray. The brightness of the diffused ring light is set to 300 lux to counteract ambient direct light. Four close-range industrial cameras are respectively aimed at the AprilTag markings at the four corners of the material tray, each equipped with an 850nm infrared light. The system pre-stores the feature template of this model of soft-pack battery protection board. The reflection threshold of the adaptive de-reflection module is set to the 95th percentile of the grayscale histogram. The orientation enhancement module sets the pin feature weight to 1.8 and the board QR code marking weight to 1.5. The reprojection error threshold of the homography correction module is set to 0.1 pixels. The offset update frequency of the material tray offset compensation is once per material tray. The adaptive communication unit automatically identifies the Modbus-RTU protocol of the existing SCARA robot and establishes a link. The command issuance delay is controlled within 8ms.

[0043] The system operation flow in this scenario is as follows: the tray arrival sensor triggers data acquisition → the preprocessing module completes de-glare removal and feature enhancement → the feature extraction and correction module completes template matching and deformation correction → the pose calculation module completes tray offset compensation and pose calculation → the instruction output module issues the loading instruction → the loading robot completes the pickup → the verification module completes accuracy verification and template update. The feature recognition error rate caused by glare in the original technology is reduced to below 0.1%, tray offset no longer requires manual adjustment, positioning accuracy is improved to over 99.9%, and loading efficiency is increased to 2400 pieces per hour. No manual supervision is required, completely solving the shortcomings of the original technology such as high glare interference, poor tray offset adaptability, and low positioning accuracy. The core glare area recognition logic in this scenario can reuse the code snippet corresponding to claim 2; only the glare threshold quantile value needs to be adjusted to 95 for direct operation.

[0044] Scenario Example 2: Multi-model switching and material feeding scenario for power lithium battery protection boards This embodiment is applied to the protection board loading station of the power lithium battery PACK production line. It is necessary to switch 8-10 different sizes and pin specifications of power battery protection boards every day. A single material tray can carry 20 large-sized protection boards with a thickness of 2mm. Due to the stress of packaging and storage, the warping of the protection board can reach up to 1mm. The original positioning system required manual re-marking of feature templates and debugging of recognition parameters every time the model was changed. The debugging time was up to 30 minutes. When the warping deformation of the protection board was large, the positioning error exceeded 2mm. The nozzle often deviated, resulting in pin bending and scrapping. The scrap rate of protection boards reached 1.2%.

[0045] The system of this invention pre-stores feature templates for all production models of power battery protection boards. When switching models, simply select the corresponding model on the touch interface to automatically call up the corresponding template and match the corresponding pins and identification feature enhancement weights. In response to the large warpage of the power battery protection board, the RANSAC iteration count of the homography correction module is increased from the default 100 times to 200 times, while the reprojection error threshold remains unchanged at 0.1 pixels. The accuracy threshold of the template update module is adjusted to 0.3mm, and the template is automatically updated after three consecutive deviations exceeding the limit. The adaptive communication unit automatically recognizes the PROFINET interface of the existing six-axis industrial robot and can establish a link without manually configuring the IP address and communication parameters. The loading status is automatically synchronized to the factory MES system.

[0046] The system operation flow in this scenario is as follows: model switching command issuance → automatic invocation of corresponding model template and parameters → material tray arrival acquisition → preprocessing enhancement → template matching → deformation correction → offset compensation → pose calculation → command issuance → material loading verification → automatic template iteration. Model switching time is reduced from 30 minutes to less than 10 seconds, positioning error caused by warping is reduced from 2mm to below 0.3mm, and the scrap rate of protection boards is reduced from 1.2% to below 0.05%, completely solving the shortcomings of the original technology, such as poor adaptability to multiple models, insufficient warping correction capability, and high debugging costs. The core logic of monofilament transformation correction in this scenario can reuse the algorithm formula corresponding to claim 4; only the iteration number parameter needs to be adjusted to adapt to the correction requirements of large warping amounts, without modifying the core code logic.

[0047] refer to Figure 1 This diagram illustrates the closed-loop workflow of the lithium battery protection board safety system throughout its entire lifecycle from a macro-control perspective. The core control link relies on the efficient acquisition of multi-source parameters and subsequent hierarchical linkage response. First, the acquisition module, under a common clock source and dynamic frequency regulation, achieves time-aligned acquisition of multi-dimensional physical parameters of the battery cell. Subsequently, the multi-feature fault prediction module uses normalization operators and vector matching algorithms to calculate the specific fault type and hazard level. Following this, the graded protection execution module matches and outputs differentiated actions such as power limiting, two-level disconnection, or multi-channel alarms based on the hazard determination results.

[0048] Meanwhile, the entire main control chain is monitored around the clock by a fault-tolerant control module with independent power supply. In the event of a software crash, the fault-tolerant module can directly trigger a forced hardware shutdown and lock the uploaded critical logs. Each system-level module is horizontally connected to a dedicated fault-tolerant bus via a high-speed communication bus, forming a highly fault-tolerant protection system that balances predictive accuracy with inherent equipment safety.

[0049] refer to Figure 2This diagram details the synchronization and coordination mechanism of each subunit in the front-end multi-source cell parameter acquisition module, as well as the dynamic sampling rate adaptive tuning control logic that fluctuates with the charging and discharging current intensity. The acquisition system hardware consists of multiple parallel detection subunits for individual cell voltage, NTC surface temperature, AC internal resistance, and the total voltage and current of the main circuit. To eliminate the time difference caused by distributed sampling, the system physically cascades the clock synchronization pins of all subunits to the clock output of the main controller. A common clock source triggers the synchronous operation of each channel, thereby injecting a unified timestamp into each frame of acquired data.

[0050] During operation, the main circuit sampling subunit continuously monitors the changing trend of the main charging and discharging current through the manganese copper shunt. When the charging and discharging current rises sharply and exceeds the set high threshold, the system immediately controls the clock output to switch to the high frequency range to improve sampling accuracy and prevent missed instantaneous overcurrent. When the current falls back below the low threshold, it adaptively drops to the low frequency long-time range, significantly reducing the system's own operating power consumption.

[0051] refer to Figure 3 This diagram specifically analyzes the internal software computation logic of the multi-feature fault prediction module after receiving standardized data, including high-dimensional feature transformation, multi-model similarity verification, and algorithm library self-evolution iteration. Due to the significant differences in physical dimensions such as voltage, temperature, and internal resistance, the system first calls the feature normalization unit, using the interval extremum scaling operator to linearly convert all input features into normalized standardized feature values, and then concatenates them in a specific space to synthesize a multi-dimensional feature vector with a fixed dimension.

[0052] Subsequently, the fault feature matching unit intervenes, dynamically loading specific feature dimension weight coefficients based on the fault type to be verified, such as short circuit, overcharge, and cell degradation, and calculating the weighted cosine similarity between this real-time feature vector and the pre-stored fault feature set center vector. To avoid major accidents, the similarity threshold for higher-risk faults is set lower. If the final judgment result is true, this typical fault instance will be added to the incremental database, and the feature set center vector will be periodically recalculated to achieve fine-tuning and updating of the algorithm's accuracy.

[0053] refer to Figure 4 This figure illustrates the control details of how the graded protection execution module implements gradient power control and physical disconnection through a hardware array based on the received signals of different fault hazard levels.

[0054] When the system determines that the danger signal is a level one minor fault, the power limiting unit intervenes and controls the power transistor array with low on-resistance to perform linear high-speed adjustment of the on-duty cycle. Under the premise of not interrupting the main power supply circuit and ensuring the continued operation of the load, the charging and discharging power is limited in a step-by-step linear manner. When the danger signal is upgraded to a level two medium fault, the system skips the power transistor fine-tuning stage, immediately issues a full disconnect command, cuts off the control power transistor of the charging and discharging main circuit, and simultaneously pushes an emergency alarm containing the fault cell number to the host computer through a multi-channel link. When faced with a level 3 severe safety risk such as fire or explosion, the control program will take hardening cascade intervention measures. While instantly shutting down the power transistor, it will immediately output a high level to the hardware protection trigger unit, directly blowing the one-time fuse in the charging and discharging circuit, achieving a physical-level permanent circuit cutoff, and ensuring inherent safety on site.

[0055] refer to Figure 5 This diagram details the independent hardware monitoring and protection logic of the fault-tolerant control module, as well as the high-precision timing synchronization and collaborative protection mechanism based on the fieldbus when multiple protection boards are connected in parallel. The system has a built-in watchdog unit independently powered by a button battery, which continuously polls and reads the heartbeat packets sent by each functional module via a dedicated bus. If a continuous interruption of the heartbeat or a software crash is detected, the watchdog unit bypasses all software control layers and directly drives the hardware protection circuit to blow the independent fuse, cutting off the main circuit. It also forcibly locks and stores all historical operating frames before and after the anomaly in non-volatile flash memory. Furthermore, under high-load conditions with multiple batteries connected in parallel, multiple protection boards automatically complete configuration-free networking via a standard network bus and autonomously assign unique node identification codes. When any protection board triggers a high-level protection disconnection action, it broadcasts a protection command carrying a high-precision timestamp via the bus. All nodes in the same circuit synchronously execute the disconnection based on this timestamp, strictly suppressing the synchronization error within the specified range and physically eliminating the safety hazards of parallel circulating current and instantaneous overcurrent caused by asynchronous protection actions.

[0056] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for an automatic identification and positioning battery protection plate loading device, characterized in that, Includes the following steps: S1. Acquire the global original image of the multiple battery protection boards carried by the material tray to be loaded, and at the same time acquire the positioning image of the reference positioning mark of the material tray; S2. Perform adaptive de-reflection preprocessing and feature enhancement processing on the global original image to obtain the image to be identified. The feature enhancement processing is used to perform targeted enhancement on the pin contour features and preset board marking features of the battery protection board. S3. Based on the preset protection board feature template, match the image to be identified to obtain the initial positioning area of ​​each battery protection board. After performing deformation correction on the initial positioning area, extract the contour features and identification features of each battery protection board. S4. Combine the positioning image of the reference positioning mark of the material tray to perform material tray offset compensation, calculate the actual pose parameters of each battery protection board based on the extracted contour features and mark features, and generate corresponding loading coordinate instructions. S5. Send the loading coordinate command to the loading actuator to complete the loading and positioning operation of the battery protection board.

2. The method for the automatic identification and positioning battery protection plate loading device according to claim 1, characterized in that, The adaptive de-reflection preprocessing in step S2 specifically involves: identifying highly reflective areas in the global original image, and performing pixel filling repair on the highly reflective areas based on the average grayscale value of adjacent pixels. The highly reflective areas are continuous pixel areas with grayscale values ​​greater than a preset reflectivity threshold.

3. The method for the automatic identification and positioning battery protection plate loading device according to claim 1, characterized in that, The targeted enhancement in step S2 specifically involves: based on preset pin feature weights and board identification feature weights, correspondingly increasing the pixel contrast of the area to be enhanced, wherein the area to be enhanced is a candidate area in the global original image that matches the preset size of the battery protection board.

4. The method for the automatic identification and positioning battery protection plate loading device according to claim 1, characterized in that, The deformation correction in step S3 specifically involves: identifying the relative positions of multiple preset plate marking features within the initial positioning area, performing homography transformation correction on the initial positioning area based on the preset standard marking positional relationship, and eliminating the positioning error caused by the warping deformation of the battery protection board.

5. The method for the automatic identification and positioning battery protection plate loading device according to claim 1, characterized in that, The actual pose parameter calculation in step S4 is specifically as follows: extract the minimum bounding rectangle of the pin outline of the battery protection board, combine it with the actual pose of the material tray obtained by the material tray offset compensation, calculate the center coordinates, rotation angle and picking height parameters of the minimum bounding rectangle, and generate the loading coordinate command.

6. The method for the automatic identification and positioning battery protection plate loading device according to claim 1, characterized in that, The step S5 is followed by a positioning effect verification step: real-time image of the feeding actuator picking up the battery protection board is collected to verify the picking accuracy. If the accuracy does not meet the preset requirements, the feature template of the protection board and the reference offset parameters of the material tray are updated.

7. A system for automatically identifying and positioning a battery protection plate loading device, used to implement the method for automatically identifying and positioning a battery protection plate loading device according to any one of claims 1-6, characterized in that, include: The image acquisition module is used to acquire the global original image of the multiple battery protection boards carried by the material tray to be loaded, and at the same time acquire the positioning image of the reference positioning mark of the material tray; The preprocessing module is used to perform adaptive anti-reflection preprocessing and feature enhancement processing on the global original image to obtain the image to be identified. The feature enhancement processing is used to perform targeted enhancement on the pin contour features and preset board marking features of the battery protection board. The feature extraction and correction module is used to match the image to be identified based on the preset protection board feature template to obtain the initial positioning area of ​​each battery protection board. After performing deformation correction on the initial positioning area, the contour features and identification features of each battery protection board are extracted. The pose calculation module is used to compensate for the offset of the material tray by combining the positioning image of the reference positioning mark of the material tray. Based on the extracted contour features and mark features, it calculates the actual pose parameters of each battery protection board and generates the corresponding loading coordinate instructions. The instruction output module is used to send the loading coordinate instructions to the loading actuator to complete the loading and positioning operation of the battery protection board.

8. The system for the automatic identification and positioning battery protection plate loading device according to claim 7, characterized in that, The image acquisition module includes a global industrial camera and a reference mark acquisition camera. The field of view of the global industrial camera covers the entire material tray to be loaded, and the field of view of the reference mark acquisition camera covers all reference positioning marks on the tray.

9. The system for the automatic identification and positioning battery protection plate loading device according to claim 7, characterized in that, It also includes a template update module, which is used to collect real-time images of the battery protection board after the feeding actuator picks up the battery protection board, verify the picking accuracy, and update the protection board feature template and the material tray reference offset parameters if the accuracy does not meet the preset requirements.

10. The system for the automatic identification and positioning battery protection plate loading device according to claim 7, characterized in that, The pose calculation module also includes a communication unit, which establishes a real-time communication link with the controller of the feeding actuator to support the real-time issuance of feeding coordinate commands and the real-time feedback of feeding execution status.

Images