A lithium battery cycle life improvement method based on deep learning

Abstract

The present application relates to the technical field of button lithium battery assembly, and particularly discloses a lithium battery cycle life improvement method based on deep learning, which is used to solve the problems of existing monocular two-dimensional imaging collection of solid button lithium battery interface, such as lithium foil reflection, electrolyte light transmission producing shadow, three-dimensional morphology loss, low defect identification accuracy, only after removing defective products, unable to online process control, easy to form interface dry area to breed lithium dendrite, and short cycle life, etc. The method comprises the steps of collecting image data, using a deep learning module for multi-class identification, obtaining assembly process control instructions, and executing the control instructions, combining binocular depth imaging and deep learning closed-loop control of lithium battery assembly process parameters. The present application controls the interface of the lithium battery solid-state electrolyte and the electrode during assembly through binocular depth imaging and deep learning closed-loop control, optimizes the interface wetting state, suppresses lithium dendrite attenuation, and helps to improve the cycle life of the lithium battery.

Description

Technical Field

[0001] This invention relates to the field of button cell lithium battery assembly technology, and in particular to a method for improving the cycle life of lithium batteries based on deep learning. Background Technology

[0002] The wetting morphology of the solid-solid interface in the assembly stage of coin cell lithium batteries is crucial in determining the battery's cycle life. Chinese invention patent application CN117881961A discloses a method, device, equipment system, and medium for detecting the wetting state of batteries. Specifically, it discloses an image detection scheme for lithium battery wetting state, which qualitatively identifies obvious wetting defects on the cell surface through a single two-dimensional planar image and grayscale threshold segmentation, allowing for offline sorting of substandard cells to mitigate capacity decay. However, this detection method lacks depth information in two-dimensional images, failing to identify micron-level interface voids, hidden dry areas, and lithium foil wrinkles in solid-state batteries. Furthermore, traditional threshold algorithms can only determine pass or fail, lacking multi-category differentiation of complex interface morphologies and failing to output quantitative morphology parameters for process control. The detection results are only used for post-processing rejection of defective products, unable to be linked to micro-injection and closed-loop process adjustments in the stacking equipment, requiring a fixed injection volume and constant lamination pressure throughout the assembly process. When the aforementioned defects exist, the dry area at the interface causes current concentration, rapid lithium dendrite growth, localized electrolyte enrichment leading to continuous side reactions, repeated damage to the SEI film, a continuous increase in cell impedance, and a sharp drop in capacity. Although existing technologies include mechanical pressurization of solid-state batteries and electrode recurrent neural network (CNN) defect detection schemes, visual recognition is not performed, or only surface defects of bare electrodes are identified. This lack of ability to quantify the three-dimensional morphology of the solid-solid interface and adaptively adjust assembly parameters ultimately induces cycle failure in the assembled lithium battery, affecting its cycle life. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method for improving the cycle life of lithium batteries based on deep learning. By using binocular depth imaging and deep learning closed-loop control to regulate the interface between the solid electrolyte and the electrode of the lithium battery during assembly, the interface wetting state is optimized, lithium dendrite decay is suppressed, and the cycle life of lithium batteries is improved.

[0004] To achieve the above objectives, the present invention provides the following technical solution: A deep learning-based method for improving the cycle life of lithium batteries is applied to the assembly process of Li / SSE / Li symmetric batteries, lithium iron phosphate batteries, or ternary solid-state full batteries (CR2025 cells) in an argon glove box. The method involves quantitatively injecting a pre-set electrolyte at the interface between the solid electrolyte and the electrode, and includes the following steps: Step 1: Acquire image data: After the cell stacking is completed, acquire binocular depth vision images of the interface between the solid electrolyte and the electrode, acquire three-dimensional morphological depth information of the interface between the solid electrolyte and the electrode, and generate an interface immersion composite image dataset carrying actual working condition labels and depth pixel information. Step 2: Use deep learning module for multi-class recognition: Use deep learning model to perform multi-class recognition on interface wetting conformal image data, and output electrolyte wetting area, interface micron void area, lithium foil wrinkled area, electrolyte local enrichment area, and output interface porosity, interface wrinkle area, and interface wetting uniformity. Step 3: Obtain assembly process control instructions: Input the parameters obtained in Step 2 into the DDPG reinforcement learning controller and output control instructions for the micro-pipette drop position correction amount, electrolyte filling volume compensation amount, and dynamic clamping force of the stacking pre-compression mechanism. Step 4, execute the control command: Execute the control command to adaptively adjust the secondary electrolyte replenishment and lamination pressure, so that the electrolyte fills the interfacial voids and generates a stable SEI film.

[0005] As a further aspect of the present invention, in step one, a binocular depth vision image of the interface between the solid electrolyte and the electrode is acquired by a binocular imaging unit. The binocular imaging unit includes a sealed telescopic bracket, two industrial depth cameras, a coaxial ring-shaped shadowless cold light lamp, and a light sensor. The two industrial depth cameras are symmetrically fixed at the ends of the sealed telescopic bracket and extend into the argon glove box to be aligned with the center of the CR2025 cell stack. The coaxial ring-shaped shadowless cold light lamp is mounted at the front end of the lens of the industrial depth camera, and the ring-shaped shadowless cold light lamp outputs 45° oblique diffused cold light to illuminate the interface between the solid electrolyte and the electrode. The two industrial depth cameras synchronously trigger the shutter to acquire two-dimensional images and depth data through a synchronization signal line, and perform pixel-level registration and fusion of the two images. The light sensor acquires the light intensity and angle physical parameters of the ring-shaped shadowless cold light lamp in real time and synchronously associates them with the image dataset.

[0006] As a further aspect of the present invention, in step one, the binocular imaging unit further includes a micrometer-level electric telescopic slide and a grating displacement sensor; two industrial depth cameras are fixed on the sliding seat of the micrometer-level electric telescopic slide, which is arranged parallel to the solid electrolyte interface; after the cell positioning is completed, the micrometer-level electric telescopic slide drives the industrial depth camera to reciprocate three times along the direction parallel to the interface, with a single translation step length of 50μm, and the grating displacement sensor collects the coordinate physical parameters of each positioning point; depth images are collected synchronously at each point, and the industrial control equipment completes distortion correction and physical stitching of the multi-field images to fully cover the bonding interface corresponding to the lithium foil and the positive electrode sheet, and the coordinate parameters collected by the grating displacement sensor are associated with the image dataset.

[0007] As a further aspect of the present invention, in step one, the binocular imaging unit further includes a transparent alumina stage, a micro negative pressure adsorption pipeline, and a pressure closed-loop solenoid valve; the transparent alumina stage is positioned between two industrial depth cameras, the micro negative pressure adsorption pipeline is connected to the transparent alumina stage, and the micro negative pressure adsorption pipeline outputs a preset negative pressure to adsorb the electrolyte cell through the pressure closed-loop solenoid valve, eliminating the gap between the solid electrolyte and the electrode; the two industrial depth cameras are arranged vertically, with the upper industrial depth camera acquiring an image of the electrolyte surface layer and the lower industrial depth camera acquiring an image of the underlying voids through the transparent alumina stage; the two images are acquired synchronously and fused, and the pressure sensor associated with the negative pressure pipeline acquires the negative pressure value and correlates it to the image dataset.

[0008] Compared with the prior art, the technical effects of the present invention are as follows: This invention acquires the three-dimensional morphological depth information of the interface between the solid electrolyte and the electrode using a binocular depth vision acquisition method, and constructs a composite image dataset by binding it with actual measurement condition labels. Then, a deep learning model completes pixel-level identification of various micro-defects at the interface and outputs quantitative morphological parameters. Furthermore, the DDPG reinforcement learning controller links the electrolyte injection and lamination mechanisms to output process control instructions that match the morphological defects and adaptively adjusts the assembly parameters. This can accurately identify hidden micro-defects at the solid-solid interface of the solid battery, actively optimize the electrolyte wetting state before packaging, suppress lithium dendrites and irreversible interface side reactions, reduce the risk of interface impedance rise and capacity cliff decay during cell cycling, and improve the cycle life of solid coin lithium batteries. Attached Figure Description

[0009] Figure 1 This is a flowchart of the method according to the first embodiment of the present invention; Figure 2 This is a schematic diagram of the technical route of the first embodiment of the present invention; Figure 3 This is a schematic diagram of the technical route of the second embodiment of the present invention; Figure 4 This is a schematic diagram of the technical route of the third embodiment of the present invention. Detailed Implementation

[0010] 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.

[0011] Example 1

[0012] Existing monocular two-dimensional images, when acquiring images of coin-type lithium batteries, suffer from issues such as lithium foil reflection, solid electrolyte light transmission forming shadows, lack of three-dimensional morphological information of the interface, low defect identification accuracy, and the inability to remove defective cells after the fact. This results in the inability to control the assembly process in real time, leading to the growth of lithium dendrites in the dry area of ​​the solid-state battery interface and a decrease in cycle life.

[0013] like Figure 1 As shown, this invention proposes a method for improving the cycle life of lithium batteries based on deep learning, applied to the assembly process of Li / SSE / Li symmetric batteries, lithium iron phosphate batteries, or ternary solid-state full batteries (CR2025 cells) in an argon glove box. The method involves quantitatively injecting a pre-defined electrolyte into the interface between the solid electrolyte and the electrode, and includes the following steps: Step 1: Acquire image data: After the cell stacking is completed, acquire binocular depth vision images of the interface between the solid electrolyte and the electrode, acquire three-dimensional morphological depth information of the interface between the solid electrolyte and the electrode, and generate an interface immersion composite image dataset carrying actual working condition labels and depth pixel information. Step 2: Use deep learning module for multi-class recognition: Use deep learning model to perform multi-class recognition on interface wetting conformal image data, and output electrolyte wetting area, interface micron void area, lithium foil wrinkled area, electrolyte local enrichment area, and output interface porosity, interface wrinkle area, and interface wetting uniformity. Step 3: Obtain assembly process control instructions: Input the parameters obtained in Step 2 into the DDPG reinforcement learning controller and output control instructions for the micro-pipette drop position correction amount, electrolyte filling volume compensation amount, and dynamic clamping force of the stacking pre-compression mechanism. Step 4, execute the control command: Execute the control command to adaptively adjust the secondary electrolyte replenishment and lamination pressure, so that the electrolyte fills the interfacial voids and generates a stable SEI film.

[0014] In specific implementation, such as Figure 2As shown, in step one, a binocular depth vision image of the interface between the solid electrolyte and the electrode is acquired through a binocular imaging unit. The binocular imaging unit includes a sealed telescopic bracket, two industrial depth cameras, a coaxial ring shadowless cold light lamp, and a light sensor. The two industrial depth cameras are symmetrically fixed at the ends of the sealed telescopic bracket and extend into the argon glove box to be aligned with the center of the CR2025 cell stack. The coaxial ring shadowless cold light lamp is mounted at the front end of the lens of the industrial depth camera. The ring shadowless cold light lamp outputs 45° oblique diffused cold light to illuminate the interface between the solid electrolyte and the electrode. The two industrial depth cameras synchronously trigger the shutter to acquire two-dimensional images and depth data through a synchronization signal line, and perform pixel-level registration and fusion of the two images. The light sensor acquires the light intensity and angle physical parameters of the ring shadowless cold light lamp in real time and synchronously associates them with the image dataset. This solves the problem that fixed monocular imaging has problems such as lithium foil mirror reflection and solid electrolyte light transmission forming shadows, which cause interface imaging distortion and the obstruction of micro-gaps in the bottom layer, making it impossible to completely acquire the three-dimensional morphology information of the interface.

[0015] In specific implementation, the deep learning model in step two is the illumination adaptive correction U-Net model. The illumination adaptive correction U-Net model incorporates the illumination physical parameters collected by the illumination sensor, performs grayscale correction of reflected and shadow areas on the binocular fused image based on these parameters, sets up a reflection occlusion recognition channel, calculates the true value of interface gaps based on pixel compensation of reflected areas, and outputs the interface porosity, interface wrinkle area, and interface wetting uniformity morphology parameters after reflection compensation. Building upon the imaging in step one, by adding an illumination correction U-Net preprocessing branch to the deep learning model, illumination parameters are introduced to compensate for image grayscale, and a reflection occlusion compensation channel is set up to avoid distortion in gap measurement values ​​caused by differences in ring light illumination and to correct the interface morphology information recognition error caused by light and shadow occlusion.

[0016] In specific implementation, in step three, the DDPG reinforcement learning controller is electrically connected to the micro-pipette stepper motor and the stacked pre-compression cylinder via a synchronization signal line. The control commands of the DDPG reinforcement learning controller are correlated and matched with the interface porosity parameters after reflection compensation. When the compensated interface porosity exceeds a preset threshold, the DDPG reinforcement learning controller outputs a first control command, driving the stacked pre-compression cylinder to increase the clamping force and driving the micro-pipette stepper motor to increase the electrolyte filling volume. When the wetting uniformity is lower than a preset threshold, the DDPG reinforcement learning controller outputs a second control command, driving the micro-pipette triaxial displacement motor to adjust the dripping position and replenish the electrolyte to the shaded area. By establishing a linkage mapping relationship between the lithium battery assembly process production equipment through the porosity parameters after illumination compensation in step three, the matching distortion of the control command compensation amplitude caused by reflection interference is avoided. By synchronously linking the cylinder and pipetting motor with the porosity compensation value to differentiate pressure adjustment and replenishment, the interface porosity hidden by light and shadow is accurately filled, the lithium dendrites caused by local current concentration are suppressed, the SEI film of the solid-state battery interface is stabilized, and the capacity decay over long cycles is effectively alleviated.

[0017] Here is a specific example: A standard CR2025 solid-state Li / SSE symmetrical battery cell was selected, and 3 μL of electrolyte was added to the interface, accounting for 0.002% of the total mass of the solid electrolyte. The ring-shaped coaxial binocular acquisition module of this embodiment was used, with the ring-shaped cold light set to an illumination intensity of 320 lx and a diffusion angle of 45°. A 640×640 resolution RGB-D depth image was simultaneously acquired, and the illumination sensor was used to generate a dataset in real time by binding the illumination parameters. After illumination adaptive correction using U-Net, the error in calculating the average porosity of the interface decreased from 11.7% in the traditional monocular scheme to 2.3%, and the accuracy of wrinkle area recognition increased from 78.2% to 96.5%. When the compensated interface porosity reached 8.2% (threshold 5%), the pressure of the stacked cylinder driven by the DDPG reinforcement learning controller was increased from 0.35 MPa to 0.52 MPa, and 0.4 μL of electrolyte was added per replenishment. When the wetting uniformity of the shaded area was less than 62%, the pipette was offset laterally by 0.6 mm for targeted replenishment. In a batch test of 100 battery cells, this solution reduced the interface dry area and lithium dendrite defect rate from 28% in the traditional solution to 3%, and improved the capacity retention rate of the cells after 1000 cycles from 67.1% to 89.4%. Compared with the existing monocular two-dimensional imaging, which can only sort out defective products after the fact, this solution relies on hardware imaging, correction network, and closed-loop linkage of equipment to repair interface defects in advance, which significantly reduces process losses and significantly extends cycle life.

[0018] Example 2

[0019] Example 1 relies on fixed-point binocular camera imaging, resulting in a limited field of view coverage in a single shot. This leads to severe imaging distortion at the cell edges, and significant errors in identifying and calculating gaps and wrinkles at the edge interfaces. Ring-shaped supplementary lighting alone cannot eliminate field-of-view boundary distortion, output globally uniform morphology parameters, or distinguish between differential defects at the cell center and edges. Furthermore, targeted electrolyte replenishment and voltage adjustment at the cell edges are not possible during control, and edge wetting defects still easily induce battery cycle degradation, indicating room for improvement. Example 2 differs from Example 1 in that it specifically describes further improvements to address the aforementioned shortcomings of Example 1.

[0020] like Figure 3 As shown, in step one, the binocular imaging unit also includes a micrometer-level motorized telescopic slide and a grating displacement sensor. Two industrial depth cameras are fixed on the sliding base of the micrometer-level motorized telescopic slide, which is arranged parallel to the solid electrolyte interface. After the cell is positioned, the micrometer-level motorized telescopic slide drives the industrial depth camera to reciprocate three times along the direction parallel to the interface, with a single translation step of 50μm. The grating displacement sensor collects the coordinate physical parameters of each positioning point. Depth images are collected synchronously at each point. The industrial control equipment performs distortion correction and physical stitching on the multi-field images, completely covering the bonding interface corresponding to the lithium foil and positive electrode sheet. The coordinate parameters collected by the grating displacement sensor are associated with the image dataset. In this embodiment, step one adds a micrometer-level translation slide and a grating displacement sensor, stitching together multiple points to collect complete cell interface images and binding displacement coordinate labels. This avoids the problem of missing edge gaps and wrinkles due to the narrow field of view of the binocular camera at fixed points and severe image distortion at the cell edges, achieving full coverage acquisition of the three-dimensional morphology of the cell center and edge interface. In step one of this embodiment, a micron-level translational slide and a grating displacement sensor are added. Multi-point stitching is used to acquire complete images of the battery cell interface and bind displacement coordinate labels. This can solve the problems of narrow field of view of fixed point binoculars, severe image distortion at the edge of the battery cell, and missed detection of edge gaps and wrinkles. It can achieve full coverage acquisition of the three-dimensional morphology of the battery cell center and edge interface.

[0021] In step two, the deep learning model is a multi-view fusion U-Net model. The multi-view fusion U-Net model incorporates coordinate physical parameters collected by a grating displacement sensor, sets regional weighted correction logic for the image stitching seam area, and establishes an image edge distortion recognition channel. It then calculates the interface porosity and interface wrinkle area of ​​the cell's central and edge regions, respectively, and outputs regional topography parameters. Since step one uses multi-view stitched images and coordinate labels, step two adds a coordinate weighted correction branch to the multi-view fusion U-Net. This solves the problem of distortion at the sliding table stitching seam and image edges caused by the improvements made in step one, resulting in averaged topography calculations across the entire domain and an inability to distinguish between center and edge defects. It enables the calculation of independent topography parameters for each region.

[0022] In step three, the DDPG reinforcement learning controller is electrically connected to the segmented stacked pressure cylinder and the triaxial micro-pipette, with control commands corresponding one-to-one with the zonal morphology parameters. When the interface porosity in the cell edge region exceeds the standard, the DDPG reinforcement learning controller drives the triaxial micro-pipette to replenish electrolyte at a fixed point on the cell edge and controls the segmented stacked pressure cylinder at the edge to increase the clamping force. When there is localized electrolyte accumulation in the cell center region, the DDPG reinforcement learning controller reduces the volume of electrolyte added by the micro-pipette and controls the segmented stacked pressure cylinder to output a clamping pressure of low pressure at the center and high pressure at the edge. Step three, with its segmented cylinder and triaxial pipetting mechanism, achieves zonal control, avoiding the problem that uniform pressure adjustment and replenishment cannot adapt to the high-incidence interface defects at the edge. Based on the zonal quantification parameters, it specifically pressurizes and replenishes electrolyte at the cell edge and reduces the load and electrolyte at the center, eliminating hidden wetting defects at the edge, reducing the interface impedance difference across the entire cell, and further reducing the capacity decay rate during cycling.

[0023] Here is a specific example: Using CR2025 solid-state Li / SSE symmetrical cells of the same specifications, 3μL of electrolyte was added to the interface, accounting for 0.002% of the total mass of the solid electrolyte. Based on the ring-shaped binocular hardware of Example 1, a 50μm stepping micrometer slide and a grating displacement sensor were added. Complete cell interface images were acquired through three translational stitching operations, and coordinate condition labels were simultaneously bound and sent to the multi-field fusion U-Net. The traditional fixed-point binocular solution had a cell edge gap measurement error of 9.6%, while this solution, after zonal correction, reduced the edge measurement error to 1.9%; the accuracy of cell edge wrinkle recognition increased from 81.3% to 97.2%. When the cell edge zonal gap rate was detected to be 7.8% (threshold 5%), the segmented cylinder increased the edge clamping force from 0.36MPa to 0.53MPa, and the pipette added 0.35μL of electrolyte to the cell edge at a fixed point; when the electrolyte was concentrated in the central area, the central pressure was reduced to 0.28MPa and the electrolyte addition volume was reduced by 0.2μL. In a batch test of 100 cells, the edge defect rate in Example 1 was 12%, while the edge defect rate in this solution was reduced to 2%; the capacity retention rate after 1000 cycles increased from 89.4% to 91.7%. Compared to Example 1, which cannot distinguish between center and edge defects due to fixed field-of-view imaging, this solution can eliminate hidden wetting defects at the edge of the cell and further reduce the difference in interface impedance across the entire cell.

[0024] Example 3

[0025] Example 2 only uses a single-sided camera to capture the surface morphology of the interface, which cannot penetrate the solid electrolyte to capture the gaps between the electrode and the electrolyte layer, or the wrinkles in the underlying lithium foil. The cell is not properly secured and is prone to slight tilting, resulting in blind spots at the underlying interface and making it impossible to detect microscopic defects. Relying solely on surface morphology parameter control cannot eliminate the dry areas at the bottom layer, leading to concentrated current at the bottom layer during charging and discharging, which continuously promotes lithium dendrite formation. The battery's long-term cycle stability remains poor and requires further improvement. Unlike Examples 1 and 2, this example further improves upon the aforementioned shortcomings of Example 2.

[0026] like Figure 4 As shown, in step one, the binocular imaging unit also includes a transparent alumina stage, a miniature negative pressure adsorption pipeline, and a pressure closed-loop solenoid valve. The transparent alumina stage is positioned between two industrial depth cameras. The miniature negative pressure adsorption pipeline is connected to the transparent alumina stage. The miniature negative pressure adsorption pipeline outputs a preset negative pressure to adsorb the electrolyte cell through the pressure closed-loop solenoid valve, eliminating the gap between the solid electrolyte and the electrode. The two industrial depth cameras are arranged vertically. The upper industrial depth camera acquires images of the electrolyte surface, while the lower industrial depth camera acquires images of the underlying voids through the transparent alumina stage. The two images are acquired synchronously and fused. The pressure sensor associated with the negative pressure pipeline acquires the negative pressure value and correlates it to the image dataset. In step one of this embodiment, a negative pressure adsorption transparent stage and dual cameras with upper and lower alignment are added to acquire surface and bottom layer depth images and record negative pressure operating conditions simultaneously. This solves the problem that single-sided imaging cannot capture the gap between the electrolyte and the bottom layer of the electrode, and the blind spot caused by slight lifting of the cell. It also eliminates the imaging loss caused by cell displacement and complete acquisition of surface and bottom layer three-dimensional morphology information.

[0027] In step two, the deep learning model is a dual-layer hierarchical parsing U-Net model. This model has two parallel image input branches, receiving surface images from the upper industrial depth camera and bottom-layer images from the lower industrial depth camera, respectively, along with negative pressure values ​​from a pressure sensor. When the negative pressure value is below the standard value or the cell exhibits micro-warping, the recognition branch corresponding to the bottom-layer image increases the recognition weight of bottom-layer void pixels. The surface image input branch identifies electrolyte surface enrichment areas and surface wetting blind areas, while the bottom image input branch identifies bottom-layer micron-sized voids and lithium foil bottom-layer wrinkles, outputting layered morphology parameters such as surface wetting uniformity, bottom-layer interface porosity, and bottom-layer wrinkle area. Based on the dual-layer hierarchical images and negative pressure parameters from step one, step two constructs a dual-parallel encoding and decoding hierarchical parsing U-Net and sets a negative pressure weight adjustment branch because insufficient negative pressure weakens the bottom-layer defect features due to cell warping. Based on this improvement, the deep learning model can layer-wise identify surface enrichment and bottom-layer voids and output two sets of independent quantization parameters.

[0028] In step three, the DDPG reinforcement learning controller is electrically connected to the micropipette stepper motor and the stacking pre-compression cylinder, respectively, and the control commands are matched with the layer morphology parameters. If the porosity of the bottom interface exceeds the standard, the DDPG reinforcement learning controller first drives the stacking pre-compression cylinder to increase the clamping pressure, and then controls the micropipette to complete the electrolyte replenishment. If there is local enrichment of electrolyte on the surface, the DDPG reinforcement learning controller reduces the electrolyte filling volume and reduces the stacking clamping force in the corresponding area. If the surface wetting uniformity is low and the porosity of the bottom interface meets the standard, the DDPG reinforcement learning controller only adjusts the drop point of the micropipette. After the control is completed, the industrial depth camera re-acquires the double-layer image and sends it to the double-layer layered analysis U-Net model for verification. After all the layer morphology parameters meet the standards, the cell delivery mechanism delivers the cell to the packaging station. In this embodiment, the design of the hierarchical hardware control logic in step three is because the surface and bottom layer defects have completely different requirements for electrolyte and pressure adaptation. Based on the layer parameters, the bottom layer void defects are treated first, and the surface layer wetting point is finely adjusted separately. The solid-solid interface wetting state is optimized in both directions, and the growth of lithium dendrites is suppressed from the bottom layer, which can maximize the extension of the stable cycle life of solid-state coin cells.

[0029] Here is a specific example: Using CR2025 solid-state Li / SSE symmetrical cells of the same specifications, 3μL of electrolyte was added to the interface, with the electrolyte accounting for 0.002% of the total mass of the solid electrolyte. Based on the translational binocular acquisition structure of Example 2, a -0.02MPa negative pressure transparent stage and dual cameras were added to the upper and lower layers. Layered acquisition of surface and bottom layer images was performed, and negative pressure parameters were recorded. These images were then fed into a dual-layer layered analysis U-Net. When the negative pressure was insufficient, the weight of bottom layer feature recognition was amplified by 1.8 times. Example 2 had a 10.2% missed detection rate for bottom layer voids when only surface imaging was used; this scheme reduced the missed detection rate to 1.6% for layered recognition of bottom layer voids; the accuracy of bottom layer wrinkle recognition increased from 83.5% to 97.8%. When detecting a bottom layer porosity of 8.5%, the lamination pressure was first increased from 0.35MPa to 0.55MPa, and then 0.4μL of electrolyte was added. When only the surface layer was unevenly wetted and the bottom layer was well-adhered, only the droplet position was shifted by 0.6mm without changing the pressure. After adjustment, secondary imaging verification showed that the defect rate of the bottom dry area and lithium dendrites in a batch of 100 cells decreased from 2% in Example 2 to 0.8%; the capacity retention rate after 1000 cycles increased from 91.7% to 93.5%. Compared with Example 2, which only collected surface morphology and could not identify hidden defects in the bottom layer, this solution uses layered imaging, layered identification, and layered adjustment to suppress lithium dendrites from the bottom interface, thereby achieving full-thickness interface wetting optimization of the cell.

[0030] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0031] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for improving the cycle life of lithium batteries based on deep learning, characterized in that, This method is applied to the assembly process of Li / SSE / Li symmetric batteries, lithium iron phosphate batteries, or ternary solid-state full batteries (CR2025 cells) within an argon-filled glove box. It involves the quantitative injection of a pre-defined electrolyte into the solid electrolyte at the electrode interface, and includes the following steps: Step 1: Acquire image data: After the cell stacking is completed, acquire binocular depth vision images of the interface between the solid electrolyte and the electrode, acquire three-dimensional morphological depth information of the interface between the solid electrolyte and the electrode, and generate an interface immersion composite image dataset carrying actual working condition labels and depth pixel information. Step 2: Use deep learning module for multi-class recognition: Use deep learning model to perform multi-class recognition on interface wetting conformal image data, and output electrolyte wetting area, interface micron void area, lithium foil wrinkled area, electrolyte local enrichment area, and output interface porosity, interface wrinkle area, and interface wetting uniformity. Step 3: Obtain assembly process control instructions: Input the parameters obtained in Step 2 into the DDPG reinforcement learning controller and output control instructions for the micro-pipette drop position correction amount, electrolyte filling volume compensation amount, and dynamic clamping force of the stacking pre-compression mechanism. Step 4, execute the control command: Execute the control command to adaptively adjust the secondary electrolyte replenishment and lamination pressure, so that the electrolyte fills the interfacial voids and generates a stable SEI film.

2. The method for improving the cycle life of lithium batteries based on deep learning according to claim 1, characterized in that: In step one, a binocular depth vision image of the interface between the solid electrolyte and the electrode is acquired through a binocular imaging unit. The binocular imaging unit includes a sealed telescopic bracket, two industrial depth cameras, a coaxial ring-shaped shadowless cold light lamp, and a light sensor. The two industrial depth cameras are symmetrically fixed at the ends of the sealed telescopic bracket and extend into the argon glove box to be aligned with the center of the CR2025 cell stack. The coaxial ring-shaped shadowless cold light lamp is mounted at the front end of the lens of the industrial depth camera. The ring-shaped shadowless cold light lamp outputs 45° oblique diffused cold light to illuminate the interface between the solid electrolyte and the electrode. The two industrial depth cameras synchronously trigger the shutter to acquire two-dimensional images and depth data through a synchronization signal line, and perform pixel-level registration and fusion of the two images. The light sensor acquires the light intensity and angle physical parameters of the ring-shaped shadowless cold light lamp in real time and synchronously correlates them to the image dataset.

3. The method for improving the cycle life of lithium batteries based on deep learning according to claim 2, characterized in that: In step two, the deep learning model is the illumination adaptive correction U-Net model. The illumination adaptive correction U-Net model accesses the illumination physical parameters collected by the illumination sensor, performs grayscale correction of the reflected and shadow areas of the binocular fusion image based on the illumination physical parameters, sets up a reflection occlusion recognition channel, calculates the true value of the interface gap based on the pixel compensation of the reflected area, and outputs the interface porosity, interface wrinkle area, and interface wetting uniformity morphology parameters after reflection compensation.

4. The method for improving the cycle life of lithium batteries based on deep learning according to claim 3, characterized in that: In step three, the DDPG reinforcement learning controller is electrically connected to the micropipette stepper motor and the stacked pre-compression cylinder via a synchronization signal line. The control commands of the DDPG reinforcement learning controller are associated and matched with the interface porosity parameter after reflection compensation. When the compensated interface porosity exceeds a preset threshold, the DDPG reinforcement learning controller outputs a first control command to drive the stacked pre-compression cylinder to increase the clamping force and drive the micropipette stepper motor to increase the electrolyte filling volume. When the wetting uniformity is lower than a preset threshold, the DDPG reinforcement learning controller outputs a second control command to drive the micropipette three-axis displacement motor to adjust the drip position and replenish the liquid to the shaded area.

5. The method for improving the cycle life of lithium batteries based on deep learning according to claim 2, characterized in that: In step one, the binocular imaging unit also includes a micrometer-level motorized telescopic slide and a grating displacement sensor; two industrial depth cameras are fixed on the sliding base of the micrometer-level motorized telescopic slide, which is arranged parallel to the solid electrolyte interface; after the cell is positioned, the micrometer-level motorized telescopic slide drives the industrial depth camera to reciprocate three times along the direction parallel to the interface, with a single translation step of 50μm; the grating displacement sensor collects the coordinate physical parameters of each positioning point; depth images are collected synchronously at each point, and the industrial control equipment performs distortion correction and physical stitching on the multi-field images to fully cover the bonding interface corresponding to the lithium foil and the positive electrode sheet; the coordinate parameters collected by the grating displacement sensor are associated with the image dataset.

6. The method for improving the cycle life of lithium batteries based on deep learning according to claim 5, characterized in that: In step two, the deep learning model is a multi-view fusion U-Net model. The multi-view fusion U-Net model accesses the coordinate physical parameters collected by the grating displacement sensor, sets regional weight correction logic for the image stitching seam area, sets an image edge distortion recognition channel, and calculates the interface porosity and interface wrinkle area of ​​the cell center area and cell edge area respectively, and outputs the partition morphology parameters.

7. The method for improving the cycle life of lithium batteries based on deep learning according to claim 6, characterized in that: In step three, the DDPG reinforcement learning controller is electrically connected to the segmented stacked pressure cylinder and the triaxial micro-pipette, and the control commands correspond one-to-one with the morphological parameters of the zones. When the interface porosity of the cell edge region exceeds the standard, the DDPG reinforcement learning controller drives the triaxial micro-pipette to replenish liquid at a fixed point on the edge of the cell and controls the segmented stacked pressure cylinder at the edge to increase the clamping force. When there is local enrichment of electrolyte in the center region of the cell, the DDPG reinforcement learning controller reduces the volume of the micro-pipette and controls the segmented stacked pressure cylinder to output a clamping pressure of low pressure at the center and high pressure at the edge.

8. The method for improving the cycle life of lithium batteries based on deep learning according to claim 5, characterized in that: In step one, the binocular imaging unit also includes a transparent alumina stage, a miniature negative pressure adsorption pipeline, and a pressure closed-loop solenoid valve. The transparent alumina stage is positioned between two industrial depth cameras. The miniature negative pressure adsorption pipeline is connected to the transparent alumina stage. The miniature negative pressure adsorption pipeline outputs a preset negative pressure to adsorb the electrolyte cell through the pressure closed-loop solenoid valve, eliminating the gap between the solid electrolyte and the electrode. The two industrial depth cameras are arranged vertically. The upper industrial depth camera acquires images of the electrolyte surface, while the lower industrial depth camera acquires images of the underlying voids through the transparent alumina stage. The two images are acquired synchronously and fused. The pressure sensor associated with the negative pressure pipeline acquires the negative pressure value and correlates it with the image dataset.

9. The method for improving the cycle life of lithium batteries based on deep learning according to claim 8, characterized in that: In step two, the deep learning model is a two-layer hierarchical parsing U-Net model. The two-layer hierarchical parsing U-Net model sets up two parallel image input branches, which respectively receive the surface image captured by the upper industrial depth camera and the bottom layer image captured by the lower industrial depth camera, and at the same time receive the negative pressure value collected by the pressure sensor. When the negative pressure value is lower than the standard value and the battery cell has slight warping, the recognition branch corresponding to the bottom image increases the recognition weight of the bottom gap pixels. The surface image input branch identifies the electrolyte surface enrichment area and surface wetting blind area, while the bottom image input branch identifies the bottom micron-sized voids and lithium foil bottom wrinkles. The layered output includes surface wetting uniformity, bottom interface porosity, bottom wrinkle area, and layered morphology parameters.

10. The method for improving the cycle life of lithium batteries based on deep learning according to claim 9, characterized in that: In step three, the DDPG reinforcement learning controller is electrically connected to the micropipette stepper motor and the stacking pre-compression cylinder, respectively, and the control commands are matched with the layer morphology parameters. If the porosity of the bottom interface exceeds the standard, the DDPG reinforcement learning controller first drives the stacking pre-compression cylinder to increase the clamping pressure, and then controls the micropipette to complete the electrolyte replenishment. If there is local enrichment of electrolyte on the surface, the DDPG reinforcement learning controller reduces the electrolyte filling volume and reduces the stacking clamping force in the corresponding area. If the surface wetting uniformity is low and the porosity of the bottom interface meets the standard, the DDPG reinforcement learning controller only adjusts the drop point of the micropipette; after the adjustment is completed, the industrial depth camera re-acquires the double-layer image and sends it to the double-layer hierarchical analysis U-Net model for verification. After all the layer morphology parameters meet the standard, the cell conveying mechanism conveys the cell to the packaging station.

Images