A battery assembly process foreign matter defect identification method

Abstract

The present application relates to battery manufacturing quality detection technical field, specifically to a kind of battery assembly process foreign matter defect identification method, comprising: generating alternating magnetic field region by electromagnetic coil, the battery to be measured is placed in alternating magnetic field region and stays for a preset duration;After that, infrared imaging device is used to collect thermal image at preset inclination angle to the battery to be measured, and obtain infrared thermal image;Obtain standard infrared thermal image and background temperature field, local temperature difference chart, temperature gradient chart and candidate hot spot mask;Establish image recognition model, the image recognition model includes global infrared image branch, hot spot salient feature branch, hot spot attention weighting module and classification layer;The standard infrared thermal image, local temperature difference chart, temperature gradient chart and candidate hot spot mask are input into the trained image recognition model, and the abnormal probability of the battery to be measured is obtained;Foreign matter defect alarm is carried out based on abnormal probability value;The present application can improve the accuracy of battery assembly process foreign matter defect identification.

Description

Technical Field

[0001] This invention relates to the field of battery manufacturing quality inspection technology, specifically to a method for identifying foreign object defects during battery assembly. Background Technology

[0002] With the rapid development of new energy vehicles, energy storage systems, low-altitude aircraft, and portable electronic devices, the requirements for manufacturing consistency and safety reliability of energy storage and power batteries such as lithium-ion batteries, new solid-state batteries, and sodium-ion batteries are constantly increasing. During battery manufacturing, processes such as electrode cutting, stacking, winding, transfer, welding, and packaging can introduce foreign objects such as metal debris, equipment wear particles, cutting burrs, and dust agglomerates. Among these, ferromagnetic metal foreign objects, once remaining inside the battery, may puncture the separator or induce abnormal local electrochemical reactions during subsequent charging and discharging, leading to micro-short circuits, localized temperature rises, capacity decay, and in severe cases, even thermal runaway accidents. Therefore, timely, accurate, and online identification of internal foreign object defects during battery assembly is a crucial step in improving battery manufacturing quality and safety performance.

[0003] Existing methods for detecting foreign objects in batteries mainly include visual inspection, X-ray inspection, ultrasonic inspection, magnetic inspection, and electrical performance screening. These methods can only identify visible defects such as surface defects on electrodes, edge burrs, and misalignment of stacked electrodes. They are difficult to directly identify tiny metallic foreign objects already encased inside the battery. For ferromagnetic foreign objects hidden inside the battery, which are small in size and not easily imaged directly, these methods suffer from problems such as failure to identify them, insufficient sensitivity, high false positive and false negative rates, and difficulty in adapting to online inspection cycles. Visual and X-ray inspections can observe the internal structure of the battery and have certain advantages in foreign object identification, stacked electrode alignment detection, and electrode tab welding quality inspection. However, their equipment costs are high, the detection systems are complex, and they have high requirements for radiation protection and production line layout. At the same time, the identification stability is still limited for foreign objects that are small in size, have low contrast, or are obscured by complex structures. Especially after battery stacking or assembly, foreign objects may be obscured by electrodes, separators, or electrolyte layers, making it difficult to detect potential risks in a timely manner by relying solely on surface images or static structural images. Therefore, it is necessary to provide a foreign object defect identification method that can generate an observable response to internal ferromagnetic foreign objects and can be adapted to the inspection cycle of online production lines. Summary of the Invention

[0004] In view of the above problems, the present invention provides a method for identifying foreign object defects in the battery assembly process, which solves the technical problem that the identification of foreign object defects in the battery assembly process is not accurate enough in the prior art.

[0005] This invention provides a method for identifying foreign object defects during battery assembly, comprising the following steps: Step S1: Generate an alternating magnetic field region through an electromagnetic coil, place the battery under test in the alternating magnetic field region for a preset time, and raise the temperature of the ferromagnetic foreign object in the battery under test. Step S2: Remove the battery under test from the alternating magnetic field area, and use an infrared imaging device to acquire a thermal image of the battery under test at a preset tilt angle; the infrared thermal image is a multi-frame continuous image. Step S3: Preprocess and enhance the infrared thermal image to obtain a standard infrared thermal image and a background temperature field; Determining a local temperature difference map, a temperature gradient map, and a candidate hotspot mask based on the standard infrared thermal image and the background temperature field includes: subtracting the temperature value of each pixel location in the standard infrared thermal image from the corresponding background temperature value to obtain the local temperature difference map; and calculating the spatial gradient of the standard infrared thermal image to obtain the temperature gradient map. Step S4: Establish an image recognition model, which includes a global infrared image branch, a hotspot salient feature branch, a hotspot attention weighting module, and a classification layer; the image recognition model receives the standard infrared thermal image, local temperature difference map, temperature gradient map, and candidate hotspot mask, and outputs the probability of ferromagnetic foreign object anomalies; train the image recognition model to obtain a trained image recognition model; Step S5: Input the standard infrared thermal image, local temperature difference map, temperature gradient map and candidate hot spot mask into the trained image recognition model to obtain the ferromagnetic foreign object anomaly probability of the battery under test; when the ferromagnetic foreign object anomaly probability is greater than a preset threshold, a foreign object defect alarm is triggered.

[0006] Preferably, in step S3, the preprocessing step of the infrared thermal image includes: performing temperature correction, background temperature estimation, temperature normalization, noise suppression, and invalid region removal on the infrared thermal image. The background temperature estimation step includes: statistically analyzing the normal temperature distribution on the surface of the battery under test, and using the normal temperature distribution on the surface of the battery under test as the background temperature field. In step S3, the step of performing local image enhancement on the infrared thermal image includes at least one of local contrast enhancement, local temperature difference amplification, edge enhancement, and multi-scale feature enhancement.

[0007] Preferably, in step S3, the specific extraction process of the candidate hotspot mask includes: (1) Mark the areas in the local temperature difference map where the temperature rise exceeds the first preset threshold as initial candidate hotspot areas; (2) Based on the temperature gradient map, the initial candidate hotspot regions are screened, and regions with temperature gradients exceeding the second preset threshold are retained to obtain candidate hotspot regions; (3) Morphological processing is performed on the candidate hotspot region to eliminate internal voids and smooth the boundaries to obtain the candidate hotspot mask.

[0008] Preferably, step S3 further includes: Capture multiple consecutive frames of infrared thermal images, extract candidate hotspot regions for each frame, calculate the positional consistency coefficient, temperature decay trend, and duration of the candidate hotspots across multiple frames, and establish a temporal stability feature map.

[0009] Preferably, in step S4, the global infrared image branch receives a standard infrared thermal image and outputs global features; The hotspot salient feature branch receives local temperature difference map and temperature gradient map, or receives local temperature difference map, temperature gradient map and time-series stability feature map, and outputs local hotspot features; The hotspot attention weighting module obtains attention weighting enhancement features based on candidate hotspot masks; The classification layer obtains the probability of ferromagnetic foreign objects based on global features, local hotspot features, and attention-weighted enhancement features.

[0010] Preferably, the global infrared image branch and the hotspot salient feature branch include multiple convolutional layers, pooling layers, and activation layers; The hotspot attention weighting module performs element-wise multiplication on the candidate hotspot mask and the intermediate feature map to obtain attention-weighted enhanced features. The classification layer includes one or more fully connected layers, and a Softmax activation function or a Sigmoid activation function.

[0011] Preferably, in step S4, the step of training the image recognition model includes constructing a training dataset; the step of constructing the training dataset specifically includes: Obtain multiple normal battery samples and abnormal battery samples, and combine the normal battery samples and abnormal battery samples to form a training dataset; The normal battery sample is an infrared thermal image of a battery without ferromagnetic foreign matter after excitation by an alternating magnetic field, along with its corresponding local temperature difference map, temperature gradient map, and candidate hot spot mask. The abnormal battery sample includes an abnormal hotspot image and its corresponding local temperature difference map, temperature gradient map and candidate hotspot mask; The abnormal hotspot images include measured abnormal samples, artificially constructed defect samples, and simulated hotspot samples.

[0012] Compared with the prior art, the present invention has at least the following beneficial effects: (1) This invention can convert ferromagnetic foreign matter defects inside the battery, which are difficult to observe directly, into thermal anomaly signals that can be identified by infrared imaging, thereby realizing non-contact online detection during battery assembly. By exciting the battery under test with an alternating magnetic field, foreign matter such as internal ferromagnetic metal particles, metal debris, or equipment wear particles generates induced eddy currents and further forms local Joule heating. This invention can convert hidden, small, and low-visibility foreign matter defects into local hot spot features in infrared thermal images, significantly improving the ability to identify internal foreign matter defects. At the same time, this invention can be directly integrated into the battery assembly line without disassembly, contact detection, or damage to the battery structure, meeting the requirements of continuous online detection and facilitating early screening before subsequent packaging, liquid injection, formation, or shipment.

[0013] (2) This invention improves the adaptability of testing batteries of different specifications through an adjustable alternating magnetic field excitation method. This invention uses a Helmholtz coil, a energized solenoid, or an equivalent electromagnetic excitation device to generate an alternating magnetic field. The coil current, frequency, and duty cycle are adjusted according to the size, thickness, structure, and testing cycle of the battery under test, enabling ferromagnetic foreign objects to generate a recognizable temperature rise without damaging the battery itself. This method is applicable to the testing needs of batteries of different sizes, thicknesses, and assembly states, avoiding the risks of insufficient heating, insufficient detection sensitivity, or localized overheating caused by fixed excitation conditions.

[0014] (3) This invention effectively improves the identification ability of weak hot spots and small hot spots by performing background correction, temperature normalization, noise suppression, local image enhancement, and hot spot feature extraction on infrared thermal images. This processing can reduce the impact of ambient temperature fluctuations, residual heat of equipment, thermal interference of the conveying mechanism, and overall battery temperature rise on the detection results, and further constructs standardized infrared images, local temperature difference maps, temperature gradient maps, and candidate hot spot masks, enabling the model to focus more on hot spot areas induced by suspected foreign objects. By including abnormal hot spot samples containing thermal response characteristics of ferromagnetic foreign objects into the training set, and combining measured abnormal samples, artificially constructed defect samples, and simulated hot spot samples, this invention further improves the model's identification stability for different battery sizes, different imaging conditions, different hot spot intensities, and different defect locations.

[0015] (4) This invention reduces the risk of false positives and false negatives by fusing global infrared images with salient hotspot features. The image recognition model of this invention can adopt a combination of global infrared image branch and hotspot salient feature branch. On the one hand, it learns the overall temperature field distribution and background thermal field pattern of the battery under test. On the other hand, it focuses on learning the local temperature rise intensity, area, boundary morphology, temperature gradient and thermal diffusion characteristics of abnormal hotspots. Furthermore, it uses candidate hotspot masks to perform hotspot attention weighting on the model features, which enhances the model's response to suspected abnormal areas and suppresses interference from non-hotspot background areas. This approach can avoid the feature dilution problem caused by relying solely on whole-image classification and improve the reliability of foreign object defect determination.

[0016] (5) This invention can effectively reduce the risk of foreign object defects in the battery manufacturing process. This invention can promptly identify ferromagnetic foreign object defects in key processes after stacking, before packaging, or after assembly, and output a foreign object defect alarm signal after detecting abnormal hot spots, triggering production line actions such as marking, sorting, rejection, or shutdown inspection. Thus, this invention can achieve early screening before battery defects further expand or enter subsequent processes, which is beneficial to reducing the defect level in the battery manufacturing process and improving the consistency and safety reliability of battery products. Attached Figure Description

[0017] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.

[0018] Figure 1 This is a schematic diagram of infrared thermal image acquisition provided by the present invention.

[0019] Figure 2 This is a schematic diagram illustrating the thermal anomaly identification results of a battery containing foreign matter under the action of an alternating magnetic field, as provided by the present invention.

[0020] Figure 3 This is a schematic diagram of the image recognition model process provided by the present invention.

[0021] Figure 4 The flowchart of the foreign object defect identification method in the battery assembly process provided by the present invention. Detailed Implementation

[0022] To better understand the above-described objectives, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other. Furthermore, the present invention can be implemented in other ways different from those described herein; therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0023] To illustrate the effectiveness of the method proposed in this invention, a specific embodiment is provided below to describe the above technical solution in detail. This embodiment uses a pouch lithium-ion battery after lamination as the test object, but the scope of application of this invention is not limited to this. It can also be applied to square cells, solid-state battery lamination units, or other battery structures that have completed lamination, winding, or preliminary assembly. Figure 4 As shown, a method for identifying foreign object defects during battery assembly is disclosed, and the specific implementation steps are as follows: Step S1: Generate an alternating magnetic field region through an electromagnetic coil, place the battery under test in the alternating magnetic field region for a preset time, and raise the temperature of the ferromagnetic foreign object in the battery under test. In this step, the present invention generates an alternating magnetic field region through an electromagnetic coil, and places the battery under test in the alternating magnetic field region for a preset time. This causes ferromagnetic foreign objects that may exist inside the battery under test to generate induced eddy currents under the action of the alternating magnetic field, thereby generating Joule heating and forming a local heating region, realizing the transformation of internal ferromagnetic foreign object defects into observable thermal anomaly signals.

[0024] Specifically, an alternating magnetic field excitation station is set up in the battery assembly line, and the foreign object detection platform of the alternating magnetic field excitation station is as follows: Figure 1 , Figure 2 As shown, the foreign object detection platform has a coil assembly and a conveying and positioning module; the coil assembly is fixedly installed on the outer periphery of the conveying and positioning module; the battery to be tested is placed on the conveying and positioning module, and the conveying and positioning module can drive the battery to be tested to move and pass through the coil assembly.

[0025] In some embodiments, the coil assembly may be a Helmholtz coil, a energized solenoid, or other equivalent electromagnetic excitation device capable of generating an alternating magnetic field. The Helmholtz coil can generate a relatively uniform magnetic field distribution in the central region of the coil, ensuring that the magnetic field strength experienced by the battery under test is basically consistent across different parts of the testing station, thereby improving the uniformity of the thermal response to foreign objects and the reliability of the detection.

[0026] In some embodiments, the present invention controls the alternating magnetic field generated by the coil assembly. Specifically, the current signal of the coil assembly can be controlled by a computer. The coil current, alternating frequency, and duty cycle of the coil assembly can be set according to the size, thickness, and moving speed of the transport and positioning module of the battery under test. The coil current is used to adjust the intensity of the alternating magnetic field; a larger coil current results in a stronger magnetic field, a stronger induced eddy current response from the ferromagnetic foreign object, and a more significant local temperature rise. The alternating frequency is used to adjust the induced eddy current response characteristics in the ferromagnetic metal foreign object. When the alternating frequency is within a suitable range, the eddy current loss inside the ferromagnetic foreign object reaches a high level, resulting in high heating efficiency. The duty cycle is used to adjust the effective duration of the alternating magnetic field and the degree of heat accumulation. By adjusting the duty cycle, overheating of the battery under test can be avoided while ensuring detection sensitivity.

[0027] In this embodiment, the maximum input current of the coil is set to 5A, the coil constant is 0.116mT / A, the coil cold resistance is 0.310Ω, the coil inductance is 42.5μH, the alternating frequency is set to 100kHz, and the safe operating temperature range of the coil is room temperature to 65℃. These parameter settings allow ferromagnetic metallic foreign objects to generate a localized temperature rise that can be detected by an infrared imaging device without damaging the battery body. The detection station is located after stacking, before packaging, or after assembly, enabling online detection.

[0028] The battery under test is placed in the alternating magnetic field region and stays there for a preset time. In this embodiment, the preset time for the battery under test to stay in the alternating magnetic field region is 20s to 30s, preferably 20s.

[0029] The dwell time can be achieved by pausing the conveyor positioning module, segmenting the conveyor, or using the inspection station, and can be adjusted according to the production line cycle time, battery size, foreign object detection sensitivity requirements, and thermal response intensity.

[0030] When there are no ferromagnetic metal foreign objects inside the battery under test, the overall temperature field of the battery surface changes little or uniformly. When there are ferromagnetic metal particles, metal debris, equipment wear particles, or cutting burrs inside the battery under test, these foreign objects generate induced eddy currents under the action of an alternating magnetic field. These induced eddy currents generate Joule heating inside the foreign object, causing a more significant local temperature rise in the area where the foreign object is located compared to the surrounding normal area. Through this step, the present invention transforms internal ferromagnetic foreign object defects that are difficult to observe directly into information that can be captured by an infrared imaging device.

[0031] Step S2: Remove the battery under test from the alternating magnetic field area, and use an infrared imaging device to acquire a thermal image of the battery under test at a preset tilt angle; the infrared thermal image is a multi-frame continuous image. In this step, after the battery under test is excited by the alternating magnetic field, the delivery and positioning module moves the battery out of the alternating magnetic field region, such as... Figure 1 As shown.

[0032] An infrared imaging device is used to acquire thermal images of the battery under test. Specifically, to reduce the diffusion or attenuation of the thermal response of ferromagnetic metal foreign objects during the transfer process, the distance between the alternating magnetic field region and the infrared thermal image acquisition region is set to 10cm~30cm, preferably 20cm.

[0033] Furthermore, after the battery under test is removed from the alternating magnetic field region, the time it takes for the battery to move to the infrared thermal image acquisition area should not exceed 5 seconds, preferably 2 seconds. By shortening the transfer distance and transfer time, it is ensured that the local hotspots generated by the foreign object can still maintain sufficient temperature rise contrast during infrared imaging acquisition.

[0034] like Figure 1 As shown, the infrared imaging device is positioned laterally on one side or above the battery under test to acquire infrared thermal images of the battery's surface. A preset tilt angle is formed between the optical axis of the infrared imaging device and the normal direction of the battery's surface. This preset tilt angle is 15° to 45°, preferably 30°. The distance between the infrared imaging device and the battery under test is set to 15cm to 30cm, preferably 20cm. This tilted arrangement effectively reduces false hotspots caused by reflections from the battery surface film, specular reflections from metal components, fixture reflections, external heat source reflections, or environmental thermal radiation reflections, avoiding misidentification of reflective areas as foreign object hotspots and improving the contrast and identifiability of true local temperature rise areas in the infrared thermal image.

[0035] In some embodiments, infrared thermal image acquisition can be performed using a single-frame acquisition method, where a single-frame infrared thermal image of the battery under test is acquired after the battery under test arrives at the infrared thermal image acquisition area.

[0036] In some embodiments, infrared thermal image acquisition employs a continuous multi-frame acquisition method. After the battery under test arrives at the infrared thermal image acquisition area, several frames of infrared thermal images are continuously acquired, and each frame of the infrared thermal image records the temperature distribution, hot spot location, and hot spot intensity changes.

[0037] The number of frames acquired in the continuous multi-frame image acquisition can be 3 to 10, and the acquisition interval can be 0.1s to 1s. By comparing the spatial consistency, temperature change trend, and duration of hotspot areas in the multi-frame images, real foreign object hotspots can be distinguished from instantaneous noise points, occasional reflective points, or random thermal disturbances, reducing the false judgment rate. Preferably, the infrared thermal image acquisition cycle is consistent with the production cycle of the battery production line to meet the needs of online inspection.

[0038] Step S3: Preprocess and enhance the infrared thermal image to obtain a standard infrared thermal image and a background temperature field; Based on the standard infrared thermal image and the background temperature field, a local temperature difference map, a temperature gradient map, and a candidate hotspot mask are determined. In this step, the acquired infrared thermal image is first preprocessed. This preprocessing includes temperature correction, background temperature estimation, temperature normalization, noise suppression, and invalid region removal. A detailed description follows: The temperature correction steps include: correcting the temperature values ​​in the image according to the calibration parameters of the infrared thermal imaging device, eliminating the temperature measurement error and temperature drift of the infrared imaging device, and making the temperature data between images acquired in different batches and at different times comparable.

[0039] The background temperature estimation steps include: establishing a background temperature field model based on the temperature distribution of normal samples, the baseline temperature distribution of batteries in the same batch, historical background temperature fields, or the temperature distribution of non-abnormal areas in the current image. The background temperature field is used to characterize the normal temperature distribution features of the battery under test in the absence of foreign object defects. By establishing the background temperature field, the normal temperature distribution on the surface of the battery under test can be used as a reference benchmark, highlighting the localized abnormal temperature rise caused by eddy current heating of ferromagnetic metal foreign objects.

[0040] The temperature normalization step includes: normalizing the infrared thermal image to reduce the impact of ambient temperature fluctuations, infrared imaging equipment temperature drift, residual heat in the equipment, thermal interference from the conveying mechanism, and overall battery temperature rise on subsequent defect identification results. Temperature normalization ensures that infrared thermal images from different environmental conditions and different batches of batteries have a unified temperature characterization scale.

[0041] The noise suppression steps include: using methods such as median filtering, Gaussian filtering, bilateral filtering or morphological filtering to reduce random noise, sensor noise and quantization noise in infrared images, while preserving hotspot edge information.

[0042] The invalid region removal process includes: based on the battery contour detection results, the detection area boundary setting, or a preset mask, removing non-battery surface areas such as the conveying mechanism, fixtures, and background areas, retaining only the valid detection area on the surface of the battery to be tested. By removing invalid regions, temperature interference from non-battery areas can be avoided from affecting anomaly detection.

[0043] After the above preprocessing, a standard infrared thermal image and background temperature field are obtained.

[0044] Subsequently, the present invention performs local image enhancement on the preprocessed infrared thermal image. The local image enhancement includes at least one of local contrast enhancement, local temperature difference amplification, edge enhancement, and multi-scale feature enhancement, used to enhance the difference between weak temperature rise, small area, and low contrast hotspot regions and the surrounding background region.

[0045] The steps for enhancing local contrast include: histogram equalization, adaptive histogram equalization, or contrast-limited adaptive histogram equalization.

[0046] The edge enhancement can be achieved using the Laplacian operator, Sobel operator, or Canny edge detection method. By enhancing the local image, the grayscale and temperature differences between the anomalous hotspot and the surrounding background area are improved, enabling subsequent image recognition models to more accurately identify local hotspots caused by tiny ferromagnetic metallic foreign objects.

[0047] This invention calculates a local temperature difference map based on a preprocessed standard infrared thermal image and a background temperature field. Specifically, the temperature value at each pixel location in the standard infrared thermal image is subtracted from the corresponding background temperature value to obtain the local temperature difference map. This local temperature difference map characterizes the local temperature rise relative to the background temperature at various locations on the surface of the battery under test, highlighting the intensity of temperature rise of abnormal hotspots relative to the surrounding background area.

[0048] This invention calculates a temperature gradient map based on a standard infrared thermal image. The temperature gradient map can be obtained by calculating the spatial gradient of the standard infrared thermal image, specifically using the Sobel operator, Prewitt operator, Laplacian operator, or other gradient calculation methods. The temperature gradient map characterizes the degree of local temperature change, hotspot boundary characteristics, and heat diffusion morphology. In the heating region of a ferromagnetic foreign object, the temperature gradient is large due to the higher temperature at the center and the lower temperature around it; while in normal areas or areas with uniform temperature rise, the temperature gradient is smaller. Therefore, the temperature gradient map can effectively characterize the boundary clarity and heat diffusion characteristics of hotspots.

[0049] Furthermore, the present invention extracts candidate hotspot regions based on the local temperature difference map and temperature gradient map, and generates a candidate hotspot mask. The specific extraction process of the candidate hotspot mask is as follows: (1) Mark the area in the local temperature difference map where the temperature rise exceeds the first preset threshold as the initial candidate hot spot area; the first preset threshold can be determined according to the temperature distribution statistical characteristics of a normal battery, for example, it can be set to 2 to 5 times the standard deviation of the background temperature; (2) Based on the temperature gradient map, the initial candidate hot spot area is screened and the area with temperature gradient exceeding the second preset threshold is retained to obtain the candidate hot spot area; in this way, the present invention can eliminate areas with blurred boundaries and small temperature gradients, because the real foreign object heating point usually has a relatively clear temperature boundary. (3) Morphological processing is performed on the candidate hotspot regions, including dilation, erosion, opening or closing operations, to eliminate internal holes and smooth the boundaries, thereby obtaining a candidate hotspot mask. The candidate hotspot mask is a binary image, wherein the pixel value of the candidate hotspot region is 1, and the pixel value of the non-candidate hotspot region is 0.

[0050] In some embodiments, when using consecutive multi-frame infrared thermal images, temporal features of candidate hotspot regions can also be extracted. Specifically, candidate hotspot regions are extracted for each frame of the image, and the spatial location, center coordinates, area size, and temperature intensity of the same candidate hotspot are compared across different frames. A genuine foreign object hotspot should maintain a relatively consistent spatial location across multiple frames, with its temperature gradually decreasing over time and its positional drift being minimal. In contrast, instantaneous noise points, occasional reflective points, or random thermal disturbances exhibit significant changes in position and intensity across multiple frames, and their duration is relatively short. By calculating the positional consistency coefficient, temperature decay trend, and duration (number of frames) of candidate hotspots across multiple frames, a temporal stability feature map can be constructed to assist in subsequent anomaly detection.

[0051] Through the above processing, standard infrared thermal images, local temperature difference maps, temperature gradient maps, and candidate hotspot masks are obtained, as well as optional temporal stability feature maps, providing multi-dimensional input features for subsequent image recognition models.

[0052] Step S4: Establish an image recognition model, which includes a global infrared image branch, a hotspot salient feature branch, a hotspot attention weighting module, and a classification layer; The global infrared image branch receives standard infrared thermal images and outputs global features; The hotspot salient feature branch receives local temperature difference maps and temperature gradient maps, and outputs local hotspot features; The hotspot attention weighting module obtains attention weighting enhancement features based on candidate hotspot masks; The classification layer obtains the probability of ferromagnetic foreign objects based on global features, local hotspot features, and attention-weighted enhancement features. The image recognition model is trained to obtain a trained image recognition model.

[0053] like Figure 3 As shown, the image recognition model of the present invention adopts a dual-branch neural network structure to achieve accurate identification of ferromagnetic foreign objects by fusing global temperature field features and local hot spot saliency features.

[0054] The global infrared image branch receives a standard infrared thermal image as input and outputs global features.

[0055] The global infrared image branch can employ a convolutional neural network structure, including multiple convolutional layers, pooling layers, and activation layers. Convolutional operations extract multi-scale features from the image, pooling operations reduce feature dimensionality, and activation functions introduce nonlinear transformation capabilities. This global infrared image branch can learn the overall temperature distribution pattern of the battery under test, background thermal field characteristics, large-scale temperature rise anomaly characteristics, and normal temperature benchmarks at different locations. Global features provide overall background information for anomaly detection, avoiding misjudging localized normal temperature rises as foreign object defects.

[0056] The hotspot salient feature branch receives local temperature difference map and temperature gradient map as input and outputs local hotspot features.

[0057] In practical implementation, local temperature difference maps and temperature gradient maps can be stitched together along the channel dimension to form a multi-channel input image. The hotspot salient feature branch can also employ a convolutional neural network structure, but its network design focuses more on capturing local detail features, edge features, and small target features. Through multi-scale convolutional kernels or dilated convolutions, the temperature rise intensity, area size, boundary morphology, temperature gradient distribution, and heat diffusion features of hotspots can be extracted within different receptive fields. Local hotspot features can provide fine-grained discriminative information for anomaly detection, improving the ability to identify weak hotspots, small-sized hotspots, and low-contrast hotspots.

[0058] In some embodiments, when using consecutive multi-frame infrared thermal images, the local temperature difference map, temperature gradient map, and temporal stability feature map can be stitched together according to the channel dimension to form a multi-channel input image.

[0059] The hotspot attention weighting module, based on the candidate hotspot mask, performs weighted processing on the intermediate feature maps extracted from the global infrared image branch and the hotspot salient feature branch to obtain attention-weighted enhanced features. Specifically, the candidate hotspot mask and the intermediate feature map are weighted by element-wise multiplication, which enhances the feature response of the candidate hotspot region and suppresses the feature response of the non-hotspot background region.

[0060] By using hotspot attention weighting, the model can allocate more computing resources to suspected abnormal regions, reduce the impact of background regions, conveyor regions, and other irrelevant thermal interference regions on the judgment results, and output attention-weighted enhanced features.

[0061] The present invention then fuses global features, local hotspot features, and attention-weighted enhanced features to obtain fused features. This feature fusion can be achieved through feature concatenation, feature addition, weighted feature summation, or multilayer perceptron fusion. The fused features simultaneously contain the overall temperature field information and local hotspot detail information of the battery under test, resulting in stronger discriminative capabilities.

[0062] The classification layer receives the fused features and outputs the probability that the battery under test has a ferromagnetic foreign object anomaly. In some embodiments, the classification layer may include one or more fully connected layers, and a Softmax activation function or a Sigmoid activation function. The output of the classification layer is a probability value between 0 and 1, representing the confidence level that the battery under test has a ferromagnetic foreign object defect. When the probability value is close to 1, it indicates that the model judges that the battery under test is very likely to have a ferromagnetic foreign object; when the probability value is close to 0, it indicates that the model judges that the battery under test does not have a ferromagnetic foreign object anomaly.

[0063] This invention provides steps for training an image recognition model, described in detail below: A training dataset is constructed, comprising normal battery samples and abnormal battery samples. Normal battery samples are infrared thermal images of batteries without ferromagnetic foreign objects acquired after excitation by an alternating magnetic field, along with their corresponding local temperature difference maps, temperature gradient maps, and candidate hotspot masks.

[0064] The method for obtaining abnormal battery samples specifically includes: using abnormal hotspot images and their corresponding local temperature difference maps, temperature gradient maps, and candidate hotspot masks as abnormal battery samples. The abnormal hotspot images include measured abnormal samples, artificially constructed defect samples, and simulated hotspot samples.

[0065] Specifically, the measured abnormal samples are battery samples containing real ferromagnetic foreign objects collected in actual production lines or laboratories; the artificially constructed defect samples are ferromagnetic metal particles or debris of different sizes and positions that are artificially embedded during the normal battery stacking process to form controllable abnormal samples; and the simulated hot spot samples are simulated local hot spots that are superimposed on the infrared thermal image of a normal battery based on the heating mechanism and heat conduction model of ferromagnetic foreign objects.

[0066] In some embodiments, data augmentation processing such as rotation, flipping, scaling, brightness adjustment, and contrast adjustment can be performed on the samples in the training set to improve the model's generalization ability and recognition robustness for foreign object hotspots of different sizes, locations, and temperature rise intensities.

[0067] During model training, cross-entropy loss, binary cross-entropy loss, or focus loss is used as the optimization objective, and stochastic gradient descent, Adam optimizer, or other optimization algorithms are employed for parameter updates. Through backpropagation, the convolutional kernel weights, bias parameters, and fully connected layer parameters are continuously adjusted to gradually reduce the error between the model's output anomaly probability and the true label. During training, the training dataset is divided into training and validation sets. The validation set is used to evaluate the model's generalization performance, and early stopping or learning rate decay strategies are employed to prevent overfitting. After a predetermined number of iterations, a trained image recognition model is obtained.

[0068] Step S5: Input the standard infrared thermal image, local temperature difference map, temperature gradient map and candidate hot spot mask into the trained image recognition model to obtain the ferromagnetic foreign object anomaly probability of the battery under test; when the ferromagnetic foreign object anomaly probability is greater than a preset threshold, a foreign object defect alarm is triggered.

[0069] Specifically, in the actual testing process, the battery under test passes through the alternating magnetic field excitation station and the infrared thermal image acquisition station in sequence. The acquired infrared thermal images are preprocessed, local image enhancement and feature extraction are performed to obtain a standard infrared thermal image, a local temperature difference map, a temperature gradient map and a candidate hot spot mask.

[0070] The standard infrared thermal image is input into the global infrared image branch of the trained image recognition model to extract global features; the standard infrared thermal image, local temperature difference map, and temperature gradient map are input into the hotspot saliency feature branch to extract local hotspot features; the intermediate features are weighted using a candidate hotspot mask and a hotspot attention weighting module to obtain attention-weighted enhanced features; the global features, local hotspot features, and attention-weighted enhanced features are fused and input into the classification layer to obtain the probability that the battery under test has ferromagnetic foreign matter anomalies.

[0071] The ferromagnetic foreign object anomaly probability is a value between 0 and 1, representing the confidence level of the model in determining the presence of ferromagnetic foreign object defects in the battery under test. The anomaly probability is compared with a preset threshold to perform anomaly determination. The preset threshold can be set according to actual production needs, false positive rate requirements, and false negative rate requirements, with a typical value range of 0.5 to 0.9, preferably 0.7. When the anomaly probability is greater than the preset threshold, the battery under test is determined to have a ferromagnetic foreign object anomaly, triggering a foreign object defect alarm; when the anomaly probability is less than or equal to the preset threshold, the battery under test is determined not to have a ferromagnetic foreign object anomaly and is considered a qualified battery.

[0072] like Figure 2 As shown, the standard infrared thermal image and the corresponding thermal anomaly identification results of the battery containing foreign matter are presented.

[0073] In this embodiment, when a battery under test is determined to be abnormal, the detection system sends a foreign object defect alarm signal to the production line control system. The production line control system then performs at least one of the following actions based on the alarm signal: audible and visual alarm, defect marking, abnormal battery sorting, abnormal battery rejection, or production line shutdown for inspection. This method allows for abnormal screening to be completed before defective batteries enter the packaging, electrolyte filling, formation, or subsequent factory inspection stages, which helps reduce the level of foreign object defects during battery manufacturing.

[0074] As can be seen from the above embodiments, the present invention utilizes an alternating magnetic field to actively excite ferromagnetic metallic foreign objects, causing them to generate eddy current heating responses. Local abnormal hot spots are then identified using an image recognition model guided by infrared thermal imaging and salient hot spot features. This method can transform ferromagnetic foreign object defects inside the battery, which are difficult to observe directly, into collectable, enhanced, and identifiable infrared thermal image features, making it suitable for non-contact online detection during battery assembly.

[0075] While the specific embodiments of the present invention depict actions or steps in a particular order, this should be understood as requiring such actions or steps to be performed in the specific order shown or in sequential order, or requiring all illustrated actions or steps to be performed to achieve the desired result. In certain environments, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation may also be implemented individually or in any suitable sub-combination in multiple implementations.

[0076] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying foreign object defects during battery assembly, characterized in that, Includes the following steps: Step S1: Generate an alternating magnetic field region through an electromagnetic coil, place the battery under test in the alternating magnetic field region for a preset time, and raise the temperature of the ferromagnetic foreign object in the battery under test. Step S2: Remove the battery under test from the alternating magnetic field area, and use an infrared imaging device to acquire a thermal image of the battery under test at a preset tilt angle; the infrared thermal image is a multi-frame continuous image. Step S3: Preprocess and enhance the infrared thermal image to obtain a standard infrared thermal image and a background temperature field; Determining a local temperature difference map, a temperature gradient map, and a candidate hotspot mask based on the standard infrared thermal image and the background temperature field includes: subtracting the temperature value of each pixel location in the standard infrared thermal image from the corresponding background temperature value to obtain the local temperature difference map; and calculating the spatial gradient of the standard infrared thermal image to obtain the temperature gradient map. Step S4: Establish an image recognition model, which includes a global infrared image branch, a hotspot salient feature branch, a hotspot attention weighting module, and a classification layer; the image recognition model receives the standard infrared thermal image, local temperature difference map, temperature gradient map, and candidate hotspot mask, and outputs the probability of ferromagnetic foreign object anomalies; train the image recognition model to obtain a trained image recognition model; Step S5: Input the standard infrared thermal image, local temperature difference map, temperature gradient map and candidate hot spot mask into the trained image recognition model to obtain the ferromagnetic foreign object anomaly probability of the battery under test; when the ferromagnetic foreign object anomaly probability is greater than a preset threshold, a foreign object defect alarm is triggered.

2. The battery assembly process foreign matter defect recognition method according to claim 1, characterized by, In step S3, the preprocessing steps for the infrared thermal image include: temperature correction, background temperature estimation, temperature normalization, noise suppression, and invalid region removal. The background temperature estimation step includes: statistically analyzing the normal temperature distribution on the surface of the battery under test, and using the normal temperature distribution on the surface of the battery under test as the background temperature field. In step S3, the step of performing local image enhancement on the infrared thermal image includes at least one of local contrast enhancement, local temperature difference amplification, edge enhancement, and multi-scale feature enhancement.

3. The battery assembly process foreign matter defect recognition method according to claim 2, characterized by, In step S3, the specific extraction process of the candidate hotspot mask includes: (1) Mark the areas in the local temperature difference map where the temperature rise exceeds the first preset threshold as initial candidate hotspot areas; (2) Based on the temperature gradient map, the initial candidate hotspot regions are screened, and regions with temperature gradients exceeding the second preset threshold are retained to obtain candidate hotspot regions; (3) Morphological processing is performed on the candidate hotspot region to eliminate internal voids and smooth the boundaries to obtain the candidate hotspot mask.

4. The battery assembly process foreign matter defect recognition method according to claim 3, characterized by, Step S3 also includes: Capture multiple consecutive frames of infrared thermal images, extract candidate hotspot regions for each frame, calculate the positional consistency coefficient, temperature decay trend, and duration of the candidate hotspots across multiple frames, and establish a temporal stability feature map.

5. The battery assembly process foreign matter defect recognition method according to claim 4, characterized by, In step S4, the global infrared image branch receives a standard infrared thermal image and outputs global features; The hotspot salient feature branch receives local temperature difference map and temperature gradient map, or receives local temperature difference map, temperature gradient map and time-series stability feature map, and outputs local hotspot features; The hotspot attention weighting module obtains attention weighting enhancement features based on candidate hotspot masks; The classification layer obtains the probability of ferromagnetic foreign objects based on global features, local hotspot features, and attention-weighted enhancement features.

6. The method for identifying foreign object defects in the battery assembly process according to claim 5, characterized in that, The global infrared image branch and the hotspot salient feature branch include multiple convolutional layers, pooling layers, and activation layers; The hotspot attention weighting module performs element-wise multiplication on the candidate hotspot mask and the intermediate feature map to obtain attention-weighted enhanced features. The classification layer includes one or more fully connected layers, and a Softmax activation function or a Sigmoid activation function.

7. The battery assembly process foreign object defect recognition method of claim 6, wherein In step S4, the step of training the image recognition model includes constructing a training dataset; The specific steps for constructing the training dataset include: Obtain multiple normal battery samples and abnormal battery samples, and combine the normal battery samples and abnormal battery samples to form a training dataset; The normal battery sample is an infrared thermal image of a battery without ferromagnetic foreign matter after excitation by an alternating magnetic field, along with its corresponding local temperature difference map, temperature gradient map, and candidate hot spot mask. The abnormal battery sample includes an abnormal hotspot image and its corresponding local temperature difference map, temperature gradient map and candidate hotspot mask; The abnormal hotspot images include measured abnormal samples, artificially constructed defect samples, and simulated hotspot samples.

Images