New energy battery internal microcrack detection method

Abstract

The invention provides a new energy battery internal microcrack detection method, which comprises the following steps: acquiring a battery internal three-dimensional structure image by adopting a high-resolution X-ray tomography technology, extracting a pole piece area through an image segmentation algorithm, and carrying out bending deformation simulation on a pole piece to obtain a pole piece internal stress distribution cloud picture; and combining the retrained health state evaluation and residual service life prediction model with battery electrochemical performance characterization data, coupling microcrack evolution and battery performance degradation, predicting a future crack propagation trend and a corresponding performance degradation condition of the battery, and guiding quality control in a battery design and production process.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method for detecting micro-cracks inside a new energy battery. Background Art

[0002] During the bending deformation process of the pole piece of the new energy battery, the uneven stress distribution inside the electrode material leads to the generation and expansion of microcracks. The crack nucleates at the grain boundary and extends along the grain boundary, allowing the electrolyte to penetrate into the crack tip area, changing the electrochemical environment of the area. The infiltrated electrolyte forms a liquid-rich phase at the crack tip, resulting in a significant increase in the ionic conductivity of the area. At the same time, the insulating second phase precipitated at the grain boundary reduces the intracrystalline electronic conductivity. The mismatch between ionic conductivity and electronic conductivity leads to increased polarization and increased charge transfer impedance, which in turn causes the battery internal resistance to increase. In addition, the chemical reaction caused by electrolyte penetration accelerates the corrosion in the crack tip area, reducing the crack propagation resistance and accelerating crack propagation. Crack propagation and electrolyte penetration form a positive feedback, which ultimately leads to battery capacity attenuation and shortened life. In order to address the problem of electrolyte penetration caused by crack propagation and the resulting change in battery internal resistance, it is necessary to establish a multi-field coupling model to predict the electrolyte concentration distribution and internal resistance change law in the crack tip area. At the same time, it is necessary to develop in-situ characterization technology to monitor the morphological evolution and electrical performance degradation during crack propagation in real time, reveal its internal mechanism, and provide a basis for battery health status assessment and life prediction. Summary of the invention

[0003] The present invention provides a method for detecting microcracks inside a new energy battery, which mainly includes: High-resolution X-ray tomography technology is used to obtain the three-dimensional structure image of the battery's interior. The pole piece area is extracted through an image segmentation algorithm, and the bending deformation of the pole piece is simulated to obtain a cloud map of the stress distribution inside the pole piece. According to the cloud map of the stress distribution inside the pole piece, the finite element analysis method is used to simulate the crack expansion path under stress, and combined with the grain boundary structure information of the pole piece material, it is determined whether the crack will extend along the grain boundary direction. If the crack extends along the grain boundary, the molecular dynamics simulation technology is used to analyze the penetration behavior of the electrolyte molecules at the crack. Machine vision technology is used to perform high-speed imaging of the battery surface to obtain surface morphology images, and a convolutional neural network algorithm is used to extract and classify features of the image to identify areas containing microcracks. At the same time, the acoustic emission sensor is used to collect the crack acoustic signal during the crack expansion process, and the time-frequency domain characteristics of the crack acoustic signal are extracted through wavelet transform. The regional information containing microcracks obtained by machine vision detection is combined with the crack The time-frequency domain features of the acoustic signal are fused, and the severity of the microcracks inside the battery is judged by the depth, length, density and expansion rate of the cracks. According to the severity of the microcracks, the support vector machine algorithm is used to evaluate the health status and predict the remaining service life of the battery, and a health status assessment and remaining service life prediction model is obtained; in the process of microcrack detection, an active learning strategy is used to iteratively optimize the health status assessment and remaining service life prediction model, and the crack severity level and remaining service life of the battery sample are annotated according to the microcrack detection results. The annotated data is added to the training set, and the health status assessment and remaining service life prediction model is retrained; the retrained health status assessment and remaining service life prediction model is combined with the battery electrochemical performance characterization data, and the microcrack evolution and battery performance degradation are coupled to predict the future crack expansion trend of the battery and the corresponding performance degradation, so as to guide the quality control in the battery design and production process.

[0004] The technical solution provided by the embodiment of the present invention may have the following beneficial effects: The present invention discloses a method for detecting microcracks inside a new energy battery. The method obtains the internal structure of the battery through high-resolution X-ray tomography, and analyzes crack propagation and electrolyte penetration behavior by combining finite element analysis and molecular dynamics simulation. At the same time, machine vision and acoustic emission technology are used to detect surface microcracks, and multi-source data are integrated to establish a health status assessment and life prediction model. The present invention also uses an active learning strategy to iteratively optimize the model, and couples the evolution of microcracks with the battery performance degradation to achieve the prediction of future crack propagation trends and performance degradation. This method can accurately evaluate the health status of the battery, predict the remaining service life, provide guidance for quality control in the battery design and production process, and effectively improve the reliability and service life of the battery. BRIEF DESCRIPTION OF THE DRAWINGS

[0005] Figure 1The present invention is a flow chart of a new energy battery internal microcrack detection method.

[0006] Figure 2 It is a schematic diagram of a method for detecting micro-cracks inside a new energy battery according to the present invention.

[0007] Figure 3 This is another schematic diagram of a new energy battery internal microcrack detection method according to the present invention. DETAILED DESCRIPTION

[0008] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0009] like Figure 1-3 In this embodiment, a new energy battery internal microcrack detection method may specifically include: S101. Use high-resolution X-ray tomography technology to obtain the three-dimensional structure image inside the battery, extract the pole area through image segmentation algorithm, and simulate the bending deformation of the pole to obtain the internal stress distribution cloud map of the pole.

[0010] An automatic scanning parameter optimization mechanism is used to obtain the radiation attenuation data of the battery sample, and a back-projection reconstruction algorithm is used to obtain a three-dimensional structural image of the battery based on the radiation attenuation data; regional growth segmentation is performed based on the grayscale characteristics of the pole piece in the three-dimensional structural image, and the segmented area is processed by Gaussian filtering to obtain the pole piece contour surface data; a tetrahedral finite element mesh is constructed for the pole piece contour surface data, and the Newton-Raphson iterative solver is used to calculate the grid node stress and strain values; continuous stress distribution data is obtained based on the node stress and strain values ​​through a cubic spline interpolation algorithm, and a discrete data interval mapping is used to generate a pole piece internal stress distribution cloud map.

[0011] Specifically, the voltage and current of the radiation source are adjusted by the automatic optimization mechanism of scanning parameters. During the scanning process, the battery sample is rotated at multiple angles and the radiation attenuation data is recorded. The back-projection reconstruction algorithm is used to reconstruct the three-dimensional structural image of the battery. Regional growth segmentation is performed according to the grayscale characteristics of the pole piece in the three-dimensional structural image. Gaussian filtering is applied to the segmented pole piece region to eliminate discrete noise points, and the complete pole piece contour surface data is obtained through regional connectivity analysis. A tetrahedral finite element mesh is constructed for the pole piece contour surface data, and displacement boundary constraints are imposed on the grid nodes. The Newton-Raphson iterative solver is used to calculate the stress and strain values ​​of each node during the deformation of the pole piece. A stress distribution data set is generated based on the node stress and strain values, and the stress data is processed continuously using the cubic spline interpolation algorithm. The corresponding relationship between stress value and numerical level is established through discrete data interval mapping to generate an internal stress distribution cloud map. The voltage of the radiation source in X-ray tomography is usually adjusted between 50 kV and 200 kV, and the current varies between 50 μA and 500 μA. The scanning process uses 0.5 degree angle stepping to complete 360-degree rotation. Each angle collects a projection image of 2048×2048 pixels, and the radiation attenuation data records 16-bit grayscale values. The cone beam geometry structure is used for back-projection reconstruction, and the projection data is reconstructed and calculated by the FDK algorithm. The reconstructed voxel resolution can reach 5 microns, and the obtained three-dimensional digital image has a resolution of 10243 voxels. The grayscale characteristics of the pole piece are manifested as obvious attenuation coefficient differences with the surrounding electrolyte, diaphragm and other materials. The grayscale value of the copper foil substrate is between 50,000 and 60,000, and the grayscale value of the active material layer is between 30,000 and 40,000. When the region is grown and segmented, the center point of the active material layer is used as the seed point, and the similar grayscale value area is expanded outward, and the search threshold is set to ±2000. Gaussian filtering uses a 3D convolution kernel of 5×5×5 with a standard deviation of 1.3 to suppress noise in the segmented pole piece area. The independent areas with an area of ​​less than 100 voxels are removed by 26-connected domain analysis to obtain a complete pole piece contour surface. The tetrahedral mesh unit constructed on the pole piece contour surface has a side length of 20 microns. The deformation boundary constraint imposes a fixed constraint of 0 displacement on the fixed end of the pole piece winding, and an axial compression displacement of 5 mm is imposed on the free end. The material constitutive model adopts an isotropic linear elastic model, the elastic modulus of the copper foil is 120 GPa, the Poisson's ratio is 0.3, the elastic modulus of the active material layer is 15 GPa, and the Poisson's ratio is 0.25. A convergence tolerance of 0.01 is used for the Newton-Raphson iterative solution, and the maximum number of iterations is set to 100 times to calculate the stress and strain field of each grid node. The cubic spline interpolation uses an uneven B-spline basis function, and the control point spacing is twice the grid size. The node stress data is processed continuously to reduce the stress jump caused by meshing. The discrete data interval is divided into 10 equal intervals between the maximum and minimum stress values, and the interval mapping adopts a rainbow color scheme, with the minimum stress corresponding to blue and the maximum stress corresponding to red.The obtained internal stress distribution cloud map of the pole piece clearly shows the stress concentration area during the bending deformation process of the pole piece. The stress peak mainly appears at the fixed end of the pole piece and the large curvature deformation area, and the stress level varies in the range of 0 to 300 MPa.

[0012] S102. Based on the stress distribution cloud map inside the electrode, the finite element analysis method is used to simulate the crack expansion path under the action of stress, and combined with the grain boundary structure information of the electrode material, it is determined whether the crack will extend along the grain boundary direction. If the crack extends along the grain boundary, the molecular dynamics simulation technology is used to analyze the penetration behavior of the electrolyte molecules at the crack.

[0013] The maximum principal stress criterion is used to mark the crack starting position on the stress distribution cloud map, and the crack extension driving force distribution data is obtained through the nonlinear fracture mechanics equation; the grain boundary position in the electron microscopic analysis image of the electrode surface is calibrated according to the crack extension driving force distribution data, the grain boundary network topology structure is extracted, and the angle between the grain boundary orientation angle and the stress field direction is calculated; the atomic scale lattice model is reconstructed using the grain boundary network topology structure, and the atomic motion equation is solved under periodic boundary conditions to obtain the atomic position data during the microevolution process; the local concentration distribution of electrolyte molecules at the crack is calculated according to the atomic position data during the microevolution process, and the electrolyte molecule penetration diffusion flux is determined through the molecule-grain boundary interaction potential energy curve.

[0014] Specifically, the maximum principal stress criterion is used to mark the crack starting position on the stress distribution cloud map, the crack extension path discriminator is constructed through the nonlinear fracture mechanics equation, and the crack extension driving force distribution data is calculated using the stress intensity factor. According to the crack extension driving force distribution data, the grain boundary position in the electron microscopic analysis diagram of the pole piece surface is calibrated, the grain boundary network topology is extracted, and the angle between the grain boundary orientation angle and the stress field direction is calculated in the crack front area. The crack extension path is determined by the grain boundary energy threshold. For the determined extension path along the grain boundary, the atomic scale lattice model is reconstructed using the grain boundary network topology, and the multi-body potential function is used to describe the interatomic force under periodic boundary conditions. The atomic motion equation is solved to obtain the microscopic evolution process.

[0015]

[0016] r i represents the position vector of the i-th atom, v i represents the velocity vector of the ith atom, F ij represents the interaction force between atoms i and j, m iRepresents the mass of the i-th atom, and Δt represents the time step. According to the atomic position data in the microscopic evolution process, the local concentration distribution and diffusion flux of the electrolyte molecules at the crack are calculated, and the penetration and diffusion behavior of the electrolyte molecules are determined in combination with the molecular-grain boundary interaction potential energy curve. The driving force for crack propagation in the pole piece comes from the stress field distribution. The maximum principal stress criterion determines the crack propagation tendency by calculating the stress components in different directions. The stress intensity factor reflects the stress field intensity at the crack tip, and its value is usually in the range of 0.5 to 5 MPa square root. When the maximum principal stress at the crack tip exceeds the material strength limit, the crack begins to propagate, and the stress field redistribution causes the crack to continue to propagate. In a typical lithium-ion battery pole piece, the maximum principal stress caused by bending deformation can reach 300 MPa, far exceeding the strength limit of the pole piece material of 80 MPa. The grain boundary network topology is obtained by electron microscopy analysis, and the resolution of microscopy analysis reaches 50 nanometers, which can clearly distinguish the grain boundaries. The angle between the grain boundary orientation angle and the stress field direction determines whether the crack propagates along the grain boundary. When the angle is less than 30 degrees and the grain boundary energy is less than 200 joules per square meter, the crack tends to propagate along the grain boundary. In the microstructure of the electrode, the grain size distribution ranges from 2 to 10 microns, and the grain boundary network presents a complex three-dimensional network structure, which provides a preferential path for crack propagation. The atomic scale lattice model is constructed based on crystallographic data. For typical graphite negative electrode materials, the hexagonal crystal system is used to describe the arrangement of carbon atoms, and the lattice constant a is 0.246 nanometers and c is 0.671 nanometers. The multi-body potential function uses the Tersoff potential to describe the interaction between carbon atoms, and the potential function parameters are calibrated by first-principles calculations. The atomic motion equation is solved using the velocity Verlet algorithm, with a time step of 0.5 femtoseconds and a simulation temperature controlled at 298 Kelvin. The penetration behavior of electrolyte molecules at the crack is controlled by the interaction between molecules and the grain boundary surface, which mainly includes van der Waals forces and electrostatic forces. In the crack tip area, the concentration of electrolyte molecules can reach 3 to 5 times the bulk concentration, and the molecular diffusion coefficient decays exponentially with the distance from the crack surface. The molecular-grain boundary interaction potential energy curve was obtained through molecular dynamics simulation. The potential well depth is in the range of 0.2 to 0.5 electron volts, and the corresponding adsorption energy is 20 to 50 kilojoules per mole. During the crack propagation process, the continuous penetration of electrolyte molecules accelerates the degradation of the material.

[0017] The penetration depth and amount of electrolyte at the crack are obtained, and the electrochemical impedance spectroscopy testing technology is introduced to perform molecular dynamics simulation. The change pattern of battery internal resistance with crack extension and electrolyte penetration is measured. According to the pre-generated battery internal resistance change curve, a quantitative relationship model between internal resistance and crack extension degree is established.

[0018] An ion migration flux calculation method is used to obtain the ion concentration distribution of the electrolyte in the crack channel, and the electrolyte penetration and diffusion data are obtained by calculating the gradient of the concentration distribution; an ion transfer impedance equation is established based on the electrolyte penetration and diffusion data, and after applying an AC excitation signal to the electrode, the current response versus frequency curve is recorded to obtain the battery impedance spectrum characteristics; an ion migration equivalent circuit is constructed based on the battery impedance spectrum characteristics, and a parameter optimization algorithm is used to fit and calculate the battery internal resistance change law during the crack propagation process; a deep learning predictor is trained based on the battery internal resistance change law, and the electrolyte penetration and diffusion data and crack propagation parameters are used as feature inputs to obtain the internal resistance evolution prediction equation, and a quantitative relationship model between internal resistance and crack propagation degree is established.

[0019] Specifically, the ion migration flux calculation method is used to obtain the ion concentration distribution of the electrolyte in the crack channel, the interatomic potential function in molecular dynamics is used to solve the adsorption site of the electrolyte molecule on the crack surface, and the electrolyte penetration and diffusion data are obtained by concentration gradient calculation. According to the electrolyte penetration and diffusion data, the ion transfer impedance equation is established, an AC excitation signal is applied to the electrode at a constant temperature, the current response curve with frequency is recorded, and the battery impedance spectrum characteristics are obtained by complex impedance calculation. According to the impedance spectrum characteristics, an ion migration equivalent circuit is constructed, the ion transmission path caused by crack extension is introduced into the circuit structure, and the parameter optimization algorithm is used to fit and calculate the change law of the battery internal resistance during the crack extension process. According to the battery internal resistance change law, a deep learning predictor is trained, and the electrolyte penetration and diffusion data and crack extension parameters are used as feature inputs to generate an internal resistance evolution prediction equation, and a quantitative relationship model between internal resistance and crack extension degree is established. The migration process of the electrolyte in the crack channel of the lithium-ion battery has obvious spatial distribution characteristics, and the ion concentration gradually decreases from the crack surface to the inside, forming a high concentration accumulation in the crack tip area. In a typical lithium salt electrolyte, the lithium ion concentration can reach 1.5 mol / L on the crack surface, while it drops to 0.2 mol / L deep in the crack, and the concentration gradient changes rapidly within a range of 100 microns. The interaction between the electrolyte molecules and the crack surface is mainly manifested as van der Waals force and electrostatic force, with an adsorption site density of 5 per square nanometer and an adsorption energy in the range of 20 to 50 kilojoules per mole. The impedance spectrum measurement uses an AC excitation signal of 10 Hz to 100 kHz, with an excitation voltage amplitude of 10 millivolts, and records the current response amplitude and phase at different frequencies. The electrochemical impedance spectrum of a healthy battery shows a typical semicircular feature, with the semicircle diameter corresponding to the charge transfer resistance, the high frequency band corresponding to the electrolyte resistance, and the low frequency band reflecting the ion diffusion impedance. When the crack expands and causes electrolyte penetration, the characteristic frequency of the impedance spectrum shifts, the semicircle deformation increases, and new impedance features appear in the low frequency band. The ion migration equivalent circuit consists of resistors, capacitors, and diffusion impedance elements, where the resistor element represents the electronic conductivity, the capacitor element represents the double layer effect, and the diffusion impedance element describes the ion transport process. The electrolyte penetration in the crack channel introduces a new transmission path, which appears as a parallel branch in the equivalent circuit. The resistance value in the branch decreases with the increase of crack length, from the initial 1000 ohms to 100 ohms. The parameter optimization process uses nonlinear least squares method, and the fitting accuracy is better than 0.1%. The input features of the internal resistance predictor include parameters such as crack length, penetration depth, ion concentration gradient and conductivity. The feature dimension is 10 dimensions. It adopts a 4-layer fully connected neural network structure, and the number of neurons in each layer is 64, 32, 16 and 8 respectively. The training data set contains 500 sets of measurement data under different crack states, in which the crack length varies from 50 to 500 microns and the penetration depth varies from 10 to 100 microns.The internal resistance change pattern of the predictor output shows that the battery internal resistance increases by 15 to 20 milliohms for every 100 microns the crack extends, and this change trend remains stable at different operating temperatures.

[0020] S103. Use machine vision technology to perform high-speed imaging of the battery surface to obtain surface morphology images, and use convolutional neural network algorithms to extract and classify features of the images to identify areas containing microcracks. At the same time, use acoustic emission sensors to collect crack acoustic signals during crack propagation, and extract the time-frequency domain characteristics of crack acoustic signals through wavelet transform.

[0021] A high-speed camera is used to scan the battery surface to obtain an original image, and a median filter algorithm is used to perform denoising and sharpening processing on the original image to obtain an enhanced surface image; a deep learning classifier is used to extract features from the enhanced surface image, and microcrack features are separated by a sobel operator, and the coordinates of the microcrack area are obtained according to the feature map; the position of the acoustic emission sensor is arranged according to the microcrack area coordinates, and the stress wave signal is recorded by a high-frequency signal acquisition circuit, and a pure crack acoustic emission waveform is obtained by processing with a Butterworth filter; discrete wavelet decomposition is performed on the pure crack acoustic emission waveform, and the time-frequency domain characteristics of the crack acoustic signal during the crack propagation process are obtained by calculating the energy distribution of the frequency band coefficient and the time series characteristics.

[0022] Specifically, a high-speed camera is used to continuously scan the battery surface to obtain the original image, the image brightness and contrast are adjusted according to the light intensity and camera parameters, and the median filter algorithm is used to denoise and sharpen the original image to obtain an enhanced surface image. A deep learning classifier is constructed for the enhanced surface image, the image gradient features are extracted by the sobel operator, the microcrack features are separated by multi-layer feature extraction operations, and the coordinates of the microcrack area are marked according to the feature map. The acoustic emission sensor layout position is determined according to the coordinates of the microcrack area, the stress wave signal is recorded in real time by a high-frequency signal acquisition circuit, and the low-frequency vibration noise is eliminated by a Butterworth filter to obtain a pure crack acoustic emission waveform. Discrete wavelet decomposition is performed on the pure crack acoustic emission waveform, the energy distribution and time series characteristics of each frequency band coefficient are calculated, and the time-frequency domain characteristics of the crack acoustic signal during the crack extension process are restored by wavelet reconstruction. The high-speed camera uses a planar array CCD sensor to realize image acquisition, the pixel resolution reaches 4096×3072, the frame rate is set to 1000 frames per second, and the exposure time is controlled at 100 microseconds. The light source uses a circular LED array to provide uniform illumination, with a light intensity of 10,000 lux and a color temperature of 5,500 Kelvin. The median filter uses a 5×5 sliding window to eliminate salt and pepper noise, and the image sharpening uses the Laplace operator to enhance the edge contrast. The signal-to-noise ratio of the processed image is increased to 40 decibels. The deep learning classifier uses an 8-layer convolution structure, the convolution kernel size increases from 3×3 to 11×11, and the number of feature map channels increases from 32 to 256. The Sobel operator extracts image gradients in the horizontal and vertical directions, and areas with gradient amplitudes exceeding 3 times the local mean are marked as potential cracks. The classifier training samples contain 10,000 annotated images, including 5,000 microcrack samples, with crack widths ranging from 5 to 50 microns and lengths ranging from 100 to 1,000 microns. The acoustic emission sensor is made of piezoelectric ceramic material, with a resonant frequency of 150 kHz and a bandwidth covering 100 to 200 kHz. The sensor array is arranged in a rectangular shape with a spacing of 50 mm. A total of 16 measuring points are arranged to form a sound source positioning network. The sampling rate of the signal acquisition circuit is set to 2 MHz, the quantization accuracy is 16 bits, and the acquisition time window is 1 millisecond. The cutoff frequency of the Butterworth filter is set at 80 kHz and 220 kHz, with an attenuation slope of 24 dB per octave to filter out mechanical vibration and electromagnetic interference. Discrete wavelet decomposition uses orthogonal wavelet basis functions to decompose the acoustic emission signal into 6 layers, and the frequency bands are divided into 0 to 31.25 kHz, 31.25 to 62.5 kHz, 62.5 to 125 kHz, 125 to 250 kHz, 250 to 500 kHz, and 500 to 1000 kHz. The energy distribution characteristics include the root mean square value, peak factor, and waveform factor of each frequency band coefficient, and the time series characteristics include rise time, duration, and ring count. Wavelet reconstruction uses a threshold denoising method, with the threshold being 2.5 times the coefficient standard deviation, and the reconstructed signal retains more than 95% of the effective information.The time-frequency characteristics of the acoustic emission signal show that the crack mainly excites the 200 to 400 kHz frequency band in the initial stage of propagation, and shifts to the 500 to 800 kHz frequency band in the acceleration stage of propagation.

[0023] S104. The regional information containing microcracks obtained by machine vision detection is integrated with the time-frequency domain characteristics of the crack acoustic signal, and the severity of the microcracks inside the battery is determined by the depth, length, density and expansion rate of the cracks. According to the severity of the microcracks, a support vector machine algorithm is used to perform health status assessment and remaining service life prediction on the battery to obtain a health status assessment and remaining service life prediction model.

[0024] The image data of the microcrack area and the spectrum diagram of the acoustic emission signal are obtained, and the fused crack state data is obtained through feature normalization operation; the feature weight is calculated by a multi-objective optimization method based on the crack state data, and the microcrack damage level is obtained through feature combination discrimination rules; the classification boundary parameters are calibrated according to the microcrack damage level using the damage evolution records in the historical operation data, and the health status assessment model is obtained by optimizing the kernel function parameters through the cross-validation method; the performance attenuation law is extracted from the output result of the health status assessment model using the decay curve fitting method, and the battery life prediction model is obtained through probability density estimation.

[0025] Specifically, crack morphological characteristic parameters are extracted according to the image data of the microcrack area, crack extension parameters are calculated in combination with the spectrum of the acoustic emission signal, crack feature vectors are constructed through feature normalization operations, and the fused crack state data is obtained. A microcrack evaluation index system is established for the crack state data, and the characteristic weights of crack depth, length, density, and extension rate are calculated by a multi-objective optimization method, and the microcrack damage level is obtained according to the feature combination discrimination rule. A support vector machine state evaluator is constructed according to the microcrack damage level, and the classification boundary parameters are calibrated using the damage evolution records in the historical operation data. The kernel function parameters are optimized by the cross-validation method to obtain a battery health state assessment model. A life prediction function is established for the output results of the health state assessment model, and the performance attenuation law is extracted by the decay curve fitting method. The remaining life distribution is calculated by probability density estimation to construct a battery life prediction model. The crack morphological characteristic parameters contain multiple dimensional information. The crack length extracted from the image data is distributed in the range of 50 to 500 microns, the crack width is 2 to 20 microns, and the depth direction is estimated by the acoustic emission signal amplitude in the range of 10 to 100 microns. The spectrum of the acoustic emission signal shows that the crack propagation process mainly excites the 100 to 500 kHz band, the energy is concentrated in the 200 to 300 kHz range, and the signal duration is 200 to 500 microseconds. The feature standardization uses the maximum and minimum method to map parameters of different dimensions to the range of 0 to 1. The microcrack evaluation index system divides crack features into two categories: geometric features and dynamic features. The geometric features include three indicators: depth, length, and density, and the dynamic features include three indicators: expansion rate, acoustic emission event rate, and energy release rate. The analytical hierarchy method is used to determine the index weights. The weight ratio of geometric features to dynamic features is 3:2, and the weight of each sub-indicator is between 0.15 and 0.25. The damage level is divided into three levels: slight damage, moderate damage, and severe damage, and the corresponding damage index thresholds are 0.3 and 0.7, respectively. The support vector machine state evaluator uses the radial basis kernel function, and the kernel function parameters are optimized by the grid search method. The search range covers 0.01 to 100, and the step size is divided according to the logarithmic scale. The historical operation data contains 1000 sets of complete cycle condition records, each set of data contains 500 sampling points, and the sampling interval is 1 hour. The cross-validation adopts the 5-fold method, and the classification accuracy of the validation set reaches 92%. The classification boundary parameter optimization adopts the sequence minimum optimization algorithm, and the number of iterations is limited to 1000 times. The life prediction function is constructed based on the Weibull distribution, and the prediction time span is from 0 to 1000 hours. The performance attenuation law is manifested as a dual decay of capacity and power. The capacity attenuation rate is between 0.02% and 0.05% per cycle, and the power attenuation rate is between 0.01% and 0.03% per cycle. The probability density estimation adopts the kernel density method, and the bandwidth parameter is determined by cross-validation to be in the range of 0.1 to 0.5.The remaining life distribution is right-skewed, with the ratio of the mean to the median between 1.2 and 1.5, and the 90% confidence interval covering 30% above and below the mean. In practical applications, the end of life is determined when the battery capacity decays to 80% of the rated value or the power decays to 70% of the rated value.

[0026] S105. During the microcrack detection process, an active learning strategy is used to iteratively optimize the health status assessment and remaining service life prediction model. The crack severity level and remaining service life of the battery sample are labeled according to the microcrack detection results, and the labeled data is added to the training set to retrain the health status assessment and remaining service life prediction model.

[0027] The sample information entropy is calculated according to the probability distribution output by the health status assessment model, and the sequence of samples to be labeled is obtained through the maximum information gain criterion; microcrack detection is performed on the sample sequence to be labeled, and the image feature clustering method is used to divide the severity of the cracks, and the remaining life curve is fitted in combination with the historical life data to obtain a new training sample set; the kernel function adaptive update method is used to process the new training sample set, and the distribution characteristics of the new and old data are fused through the sample importance weighting method to obtain the status assessment model; according to the health index sequence output by the status assessment model, the Bayesian parameter estimation method is used to update the life prediction function, and the life prediction model is obtained through the residual optimization criterion.

[0028] Specifically, an active learning sampling method is used to select annotation objects from the newly added battery samples, and the sample information entropy is calculated according to the probability distribution output by the health status assessment model. The samples with rich information are selected by the maximum information gain criterion to generate a sequence of samples to be annotated. The annotation discrimination criterion is constructed according to the microcrack detection data of the sample sequence to be annotated, and the image feature clustering method is used to divide the severity of the cracks. The remaining life curve is fitted in combination with the historical life data to form a new training sample set. For the new training sample set, the kernel function adaptive update method is used to expand the support vector machine feature space, and the distribution characteristics of new and old data are fused by the sample importance weighting method to construct an iteratively optimized state assessment model. According to the health index sequence output by the state assessment model, the Bayesian parameter estimation method is used to update the life prediction function, and the prediction deviation is corrected by the residual optimization criterion to obtain the iteratively updated life prediction model. The active learning sampling method is selected based on the sample information entropy. For each new battery sample, its predicted probability distribution under the current assessment model is calculated. The information entropy value of the sample reflects the uncertainty of the model's prediction of the sample. The typical information entropy distribution shows that for samples with a predicted probability close to 0.5, the information entropy value is the highest, reaching 0.693, while for samples with a predicted probability close to 0 or 1, the information entropy value is lower, below 0.1. The samples with the top 20% information entropy values ​​are selected for annotation based on the maximum information gain criterion. The annotation discrimination criterion combines the two dimensions of image features and life features. The image features are automatically divided into crack levels using a clustering method. The number of cluster centers is set to 3, corresponding to the three damage levels of slight, moderate, and severe. The image features include three indicators: crack length, width, and density. After standardization, the Euclidean distance between samples is calculated. The clustering analysis results show that the crack length of the slight damage cluster is below 200 microns, the moderate damage is between 200 and 500 microns, and the severe damage is more than 500 microns. The remaining life is based on the fitting of the Weibull distribution based on historical data, with a shape parameter between 2.5 and 3.5 and a scale parameter between 800 and 1200. The kernel function adaptive update uses a Gaussian kernel function, and the kernel width parameter is dynamically adjusted with the number of samples. The initial value is set to the median of the distance between the nearest neighbor samples in the feature space. The sample importance weighting method is based on the exponential decay of the sample age, and the decay coefficient is set to 0.95, so that newer samples get higher weights in model training. The updated support vector machine classifier has an accuracy increase of 5 to 8 percentage points in cross-validation, especially the ability to recognize newly emerging damage patterns is significantly enhanced. The Bayesian parameter estimation uses the Markov chain Monte Carlo method to sample the posterior distribution of the parameters of the life prediction function, the number of sampling is set to 1000, and the burn-in period is set to 200. The residual optimization is based on the root mean square error criterion, and the parameters are iteratively optimized by the gradient descent method. The initial value of the learning rate is set to 0.01, and the cosine annealing strategy is used for dynamic adjustment.The prediction error of the optimized life prediction model within the 90% confidence interval is reduced to within ±15%, and the time span covers 50 to 1000 hours. The prediction results show that for slightly damaged samples, the remaining life prediction value is in the range of 600 to 800 hours, for moderately damaged samples it is in the range of 300 to 500 hours, and for severely damaged samples it is in the range of 100 to 200 hours.

[0029] S106. Combine the retrained health status assessment and remaining service life prediction model with the battery electrochemical performance characterization data, couple the microcrack evolution with the battery performance degradation, predict the future crack propagation trend of the battery and the corresponding performance degradation, and guide the quality control in the battery design and production process.

[0030] A data fusion method is used to process the microcrack damage index and electrochemical impedance spectrum data output by the health status assessment model to obtain a correspondence matrix between crack status and electrochemical parameters; a mapping function of microcrack extension and performance degradation is extracted according to the correspondence matrix, and the influence intensity value of crack extension rate on electrochemical parameters is calculated by the least squares method; a deep learning predictor is trained for the influence intensity value, a time series prediction structure is used to calculate the crack extension path, and a battery performance degradation prediction curve is obtained through the electrochemical parameter evolution equation; a production quality evaluation database is established according to the performance degradation prediction curve, and the standard thresholds of crack extension and performance degradation are determined by statistical analysis methods to obtain the electrode manufacturing process parameter control indicators.

[0031] Specifically, according to the microcrack damage index and electrochemical impedance spectrum data output by the health status assessment model, the data fusion method is used to establish the corresponding relationship between the crack state and the electrochemical parameters, and the coupling characteristic matrix of the battery performance parameters and crack evolution is calculated by the dynamic estimation algorithm. The mapping function of microcrack extension and performance degradation is extracted for the coupling characteristic matrix, the least squares method is used to calculate the influence intensity of the crack extension rate on the electrochemical parameters, and the change law of capacity attenuation and internal resistance growth is obtained by the characteristic decomposition method. According to the change law, the deep learning predictor is trained, the crack extension path in the future working cycle is calculated by the time series prediction structure, and the battery performance degradation prediction curve is obtained by combining the electrochemical parameter evolution equation. A production quality evaluation system is established for the performance degradation prediction curve, the standard threshold of crack extension and performance degradation is determined by the statistical analysis method, and the parameter optimization method is used to generate the pole piece manufacturing process parameter control index. The microcrack damage index reflects the degree of deterioration of the internal structure of the battery, usually distributed between 0 and 1. When the index exceeds 0.6, it indicates that the crack extension has entered a rapid development stage. The electrochemical impedance spectroscopy data contains impedance responses with frequencies ranging from 0.1 Hz to 100 kHz. The low frequency band reflects the charge transfer impedance, the medium frequency band corresponds to the double layer effect, and the high frequency band characterizes the ohmic impedance. The data fusion adopts the weighted average method, the weight coefficient is determined by maximum likelihood estimation, the coupling characteristic matrix dimension is 10×10, and it contains the complete correspondence between the crack state quantity and the electrochemical parameters. The mapping function uses piecewise linear fitting to describe the effect of crack extension on battery performance. In the mild damage stage (damage index less than 0.3), the capacity decay rate is 0.02% per cycle, and the internal resistance growth rate is 0.01% per cycle. In the medium damage stage (damage index 0.3 to 0.6), the capacity decay accelerates to 0.05% per cycle, and the internal resistance growth reaches 0.03% per cycle. Characteristic decomposition shows that the capacity decay is mainly affected by the crack length, with a contribution rate of 60%, while the internal resistance growth is mainly related to the crack density, with a contribution rate of more than 70%. The deep learning predictor adopts a long short-term memory network structure. The input features include the crack evolution data and electrochemical parameter change trends of the past 50 cycles. The number of hidden nodes is 128, and the output predicts the performance changes of the next 20 cycles. The time series prediction shows that the crack growth rate increases exponentially with the number of cycles. The first 10 cycles grow slowly, maintaining at about 5 microns per cycle, and the next 10 cycles accelerate to 20 microns per cycle. The electrochemical parameter evolution equation is established based on Butler-Volmer kinetics, taking into account the mass transfer and charge transfer process at the electrode-electrolyte interface. The production quality evaluation system sets three levels of warning thresholds, corresponding to different crack growth states. The first warning threshold is a crack length of 300 microns, a capacity decay of 5%, and an increase in internal resistance of 10%. The second warning threshold is a crack length of 500 microns, a capacity decay of 10%, and an increase in internal resistance of 20%. The third warning threshold is a crack length of 800 microns, a capacity decay of 15%, and an increase in internal resistance of 35%.The electrode manufacturing process parameters include three key variables: rolling pressure, coating thickness, and drying temperature. Through orthogonal experiments, the optimal parameter combination was determined to be a rolling pressure of 2 to 3 MPa, a coating thickness of 60 to 80 microns, and a drying temperature of 120 to 140 degrees Celsius.

[0032] The above description is only a preferred embodiment of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the present application is not limited to the technical solution formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the concept of the present application. For example, the above features are replaced with the technical features with similar functions disclosed in the present application (but not limited to) to form a technical solution.

Claims

1. A method for detecting microcracks inside a new energy battery, characterized in that: The method comprises: High-resolution X-ray tomography technology is used to obtain the three-dimensional structure image of the battery's interior. The pole piece area is extracted through an image segmentation algorithm, and the bending deformation of the pole piece is simulated to obtain a cloud map of the stress distribution inside the pole piece. According to the cloud map of the stress distribution inside the pole piece, the finite element analysis method is used to simulate the crack expansion path under stress, and combined with the grain boundary structure information of the pole piece material, it is determined whether the crack will extend along the grain boundary direction. If the crack extends along the grain boundary, the molecular dynamics simulation technology is used to analyze the penetration behavior of the electrolyte molecules at the crack. Machine vision technology is used to perform high-speed imaging of the battery surface to obtain surface morphology images, and a convolutional neural network algorithm is used to extract and classify features of the image to identify areas containing microcracks. At the same time, the acoustic emission sensor is used to collect the crack acoustic signal during the crack expansion process, and the time-frequency domain characteristics of the crack acoustic signal are extracted through wavelet transform. The regional information containing microcracks obtained by machine vision detection is combined with the crack The time-frequency domain features of the acoustic signal are fused, and the severity of the microcracks inside the battery is judged by the depth, length, density and expansion rate of the cracks. According to the severity of the microcracks, the support vector machine algorithm is used to evaluate the health status and predict the remaining service life of the battery, and a health status assessment and remaining service life prediction model is obtained; in the process of microcrack detection, an active learning strategy is used to iteratively optimize the health status assessment and remaining service life prediction model, and the crack severity level and remaining service life of the battery sample are annotated according to the microcrack detection results. The annotated data is added to the training set, and the health status assessment and remaining service life prediction model is retrained; the retrained health status assessment and remaining service life prediction model is combined with the battery electrochemical performance characterization data, and the microcrack evolution and battery performance degradation are coupled to predict the future crack expansion trend of the battery and the corresponding performance degradation, so as to guide the quality control in the battery design and production process.

2. The method according to claim 1, characterized in that The high-resolution X-ray tomography technology is used to obtain the three-dimensional structure image of the battery, the pole piece area is extracted by the image segmentation algorithm, and the bending deformation of the pole piece is simulated to obtain the internal stress distribution cloud map of the pole piece, including: The scanning parameter automatic optimization mechanism is used to obtain the radiation attenuation data of the battery sample, and the three-dimensional structure image of the battery interior is obtained through the back-projection reconstruction algorithm according to the radiation attenuation data; Performing region growing segmentation according to the grayscale characteristics of the pole piece in the three-dimensional structure image, and processing the segmented region by Gaussian filtering to obtain the pole piece contour surface data; Constructing a tetrahedral finite element mesh for the pole piece contour surface data, and using a Newton-Raphson iterative solver to calculate mesh node stress and strain values; According to the node stress and strain values, continuous stress distribution data is obtained by using a cubic spline interpolation algorithm, and a cloud diagram of the internal stress distribution of the pole piece is generated by using discrete data interval mapping.

3. The method according to claim 1, characterized in that: According to the stress distribution cloud diagram inside the pole piece, the finite element analysis method is used to simulate the expansion path of the crack under the action of stress, and combined with the grain boundary structure information of the pole piece material, it is judged whether the crack will extend along the grain boundary direction. If the crack extends along the grain boundary, the molecular dynamics simulation technology is used to analyze the penetration behavior of the electrolyte molecules at the crack, including: The maximum principal stress criterion is used to mark the crack initiation position on the stress distribution cloud map, and the crack extension driving force distribution data is obtained through the nonlinear fracture mechanics equation; According to the crack propagation driving force distribution data, the grain boundary position in the electron microscopic analysis image of the pole piece surface is calibrated, the grain boundary network topology structure is extracted, and the angle between the grain boundary orientation angle and the stress field direction is calculated; The atomic scale lattice model is reconstructed by using the grain boundary network topology structure, and the atomic motion equation is solved under periodic boundary conditions to obtain the atomic position data in the microscopic evolution process; The local concentration distribution of electrolyte molecules at the crack is calculated based on the atomic position data during the microscopic evolution process, and the electrolyte molecule penetration diffusion flux is determined through the molecule-grain boundary interaction potential energy curve; it also includes: obtaining the penetration depth and penetration amount of the electrolyte at the crack, and introducing electrochemical impedance spectroscopy testing technology to perform molecular dynamics simulation, measure the change law of battery internal resistance with crack expansion and electrolyte penetration, and establish a quantitative relationship model between internal resistance and crack expansion degree based on the pre-generated battery internal resistance change curve.

4. The method according to claim 3, characterized in that The method comprises obtaining the penetration depth and amount of the electrolyte at the crack, introducing the electrochemical impedance spectroscopy test technology, performing molecular dynamics simulation, measuring the change law of the battery internal resistance with crack extension and electrolyte penetration, and establishing a quantitative relationship model between the internal resistance and the degree of crack extension based on the pre-generated battery internal resistance change curve, including: An ion migration flux calculation method is used to obtain the ion concentration distribution of the electrolyte in the crack channel, and the electrolyte penetration and diffusion data are obtained by calculating the gradient of the concentration distribution; an ion transfer impedance equation is established based on the electrolyte penetration and diffusion data, and after applying an AC excitation signal to the electrode, the current response versus frequency curve is recorded to obtain the battery impedance spectrum characteristics; an ion migration equivalent circuit is constructed based on the battery impedance spectrum characteristics, and a parameter optimization algorithm is used to fit and calculate the battery internal resistance change law during the crack propagation process; a deep learning predictor is trained based on the battery internal resistance change law, and the electrolyte penetration and diffusion data and crack propagation parameters are used as feature inputs to obtain the internal resistance evolution prediction equation, and a quantitative relationship model between internal resistance and crack propagation degree is established.

5. The method according to claim 1, characterized in that The method uses machine vision technology to perform high-speed imaging on the battery surface to obtain a surface morphology image, and uses a convolutional neural network algorithm to extract and classify features of the image to identify areas containing microcracks. At the same time, an acoustic emission sensor is used to collect crack acoustic signals during crack propagation, and the time-frequency domain features of the crack acoustic signals are extracted through wavelet transform, including: A high-speed camera is used to scan the battery surface to obtain an original image, and a median filter algorithm is used to perform denoising and sharpening processing on the original image to obtain an enhanced surface image; A deep learning classifier is used to extract features from the enhanced surface image, microcrack features are separated by a Sobel operator, and the coordinates of the microcrack region are obtained according to the feature map; Arranging the position of the acoustic emission sensor according to the coordinates of the microcrack area, recording the stress wave signal through a high-frequency signal acquisition circuit, and processing with a Butterworth filter to obtain a pure crack acoustic emission waveform; The pure crack acoustic emission waveform is subjected to discrete wavelet decomposition, and the time-frequency domain characteristics of the crack acoustic signal during the crack propagation process are obtained by calculating the energy distribution of the frequency band coefficient and the time series characteristics.

6. The method according to claim 1, characterized in that The regional information containing microcracks obtained by machine vision detection is fused with the time-frequency domain characteristics of the crack acoustic signal, the severity of the microcracks inside the battery is judged by the depth, length, density and expansion rate of the cracks, and the health status assessment and remaining service life prediction of the battery are performed according to the severity of the microcracks using the support vector machine algorithm to obtain a health status assessment and remaining service life prediction model, including: Obtain image data of microcrack area and spectrum of acoustic emission signal, and obtain fused crack state data through feature normalization operation; Calculating feature weights using a multi-objective optimization method based on the crack state data, and obtaining microcrack damage levels using feature combination discrimination rules; According to the microcrack damage level, the classification boundary parameters are calibrated using the damage evolution records in the historical operation data, and the health status assessment model is obtained by optimizing the kernel function parameters through the cross-validation method; The performance decay law is extracted by using the decay curve fitting method for the output result of the health status assessment model, and the battery life prediction model is obtained by probability density estimation.

7. The method according to claim 1, characterized in that In the process of microcrack detection, an active learning strategy is used to iteratively optimize the health status assessment and remaining service life prediction model, the crack severity level and remaining service life of the battery sample are marked according to the microcrack detection results, the marked data is added to the training set, and the health status assessment and remaining service life prediction model is retrained, including: The sample information entropy is calculated based on the probability distribution output by the health status assessment model, and the sample sequence to be labeled is obtained through the maximum information gain criterion; Perform microcrack detection on the sample sequence to be labeled, use image feature clustering method to classify the severity of cracks, and fit the remaining life curve in combination with historical life data to obtain a new training sample set; The newly added training sample set is processed by using a kernel function adaptive update method, and the distribution characteristics of new and old data are integrated by a sample importance weighting method to obtain a state assessment model; According to the health index sequence output by the state assessment model, the life prediction function is updated by using the Bayesian parameter estimation method, and the life prediction model is obtained by using the residual optimization criterion.

8. The method according to claim 1, characterized in that: The health status assessment and remaining service life prediction model obtained by retraining is combined with the battery electrochemical performance characterization data, coupling microcrack evolution and battery performance degradation, predicting the future crack expansion trend of the battery and the corresponding performance degradation, and guiding the quality control in the battery design and production process, including: The data fusion method is used to process the microcrack damage index and electrochemical impedance spectroscopy data output by the health status assessment model to obtain the corresponding relationship matrix between the crack state and the electrochemical parameters. Extracting a mapping function between microcrack extension and performance degradation according to the corresponding relationship matrix, and calculating the impact intensity value of the crack extension rate on the electrochemical parameters by the least square method; A deep learning predictor is trained for the impact intensity value, a time series prediction structure is used to calculate the crack propagation path, and a battery performance degradation prediction curve is obtained through an electrochemical parameter evolution equation; A production quality evaluation database is established based on the performance degradation prediction curve, and the standard thresholds of crack extension and performance degradation are determined by statistical analysis methods to obtain the pole piece manufacturing process parameter control index.

Images