Battery state of charge detection method and system based on closed-loop adaptive ultra-low field MRI

Abstract

The present application relates to the technical field of ultra-low field magnetic resonance system, and particularly relates to a battery state of charge detection method and system based on closed-loop adaptive ultra-low field MRI, comprising: collecting the phase change of water molecule H signal caused by the magnetic field disturbance of the battery to be detected to reconstruct a first signal phase image; extracting a high-information-area mask, then resampling to obtain a second signal phase image; inputting the second signal phase image into a state of charge detection model to predict the state of charge value of the battery to be detected. The battery magnetic resonance detection scheme in the prior art needs to use a high-field magnetic resonance system, which is not conducive to the use of the production line. The detection equipment is replaced with an ultra-low field magnetic resonance system, so that the shielding room and other structures in the traditional magnetic resonance system are omitted, and the imaging accuracy is improved by first locating the high-information-area and then resampling the area, thereby improving the accuracy of the detection model in estimating the state of charge.

Description

Technical Field

[0001] This invention relates to the field of ultra-low field magnetic resonance imaging (MRI) systems, specifically to a method and system for detecting the state of charge of a battery based on closed-loop adaptive ultra-low field MRI. Background Technology

[0002] Magnetic resonance battery testing is a technical solution that uses a magnetic resonance system to detect magnetic field disturbances caused by electrochemical reactions inside lithium-ion batteries during various testing processes. It helps to analyze the phase disturbance of the ¹H signal in the external water medium of the battery to invert the internal magnetic susceptibility changes.

[0003] There are already many existing technical solutions for measuring electrochemical reactions based on magnetic resonance technology.

[0004] For example, patent application CN202311358572.2 discloses an in-situ nuclear magnetic resonance (NMR) detection device for early thermal runaway of lithium batteries. This device includes a sealed enclosure for holding the lithium battery to be tested, a gas pretreatment device for cooling and filtering thermal runaway gas, a mass flow meter for measuring the volume of the thermal runaway gas, an NMR spectrometer for in-situ detection of the thermal runaway gas, and a gas post-treatment device for collecting and emitting the thermal runaway gas. The sealed enclosure is equipped with components that can be heated to induce thermal runaway in the lithium battery. The inlet of the gas pretreatment device is connected to the sealed enclosure, and the outlet of the gas pretreatment device is equipped with a temperature sensor. The inlet of the mass flow meter is connected to the outlet of the gas pretreatment device, and the inlet of the NMR spectrometer is connected to the outlet of the mass flow meter. The inlet of the gas post-treatment device is connected to both the mass flow meter and the outlet of the NMR spectrometer. This device enables in-situ analysis of the gas generation patterns and composition during the early thermal runaway process of lithium batteries.

[0005] For example, patent application CN202411392709.0 discloses a method for real-time detection of ion concentration in aqueous ion batteries using magnetic resonance imaging, comprising the following steps: S1, assembling the aqueous ion battery; S2, placing the aqueous ion battery in the detection area of ​​a magnetic resonance imaging (MRI) instrument and connecting it to an electrochemical workstation; S3, determining the region where the electrode-electrolyte interface is located using the MRI instrument; S4, applying power and acquiring two-dimensional image data of the electrode-electrolyte interface at equal time intervals to obtain a characterization spectrum of the electrolyte ion concentration. This method is suitable for in-situ detection of the battery using an MRI instrument, and can obtain ion concentration information at the electrode interface of the aqueous ion battery, thereby analyzing the deposition and stripping of metal at the electrode interface. It has high control sensitivity, stable and reliable detection results, and has the potential for widespread application.

[0006] However, in practical implementation, the inventors found that the non-destructive testing of the internal states (such as state of charge, lithium dendrites, and interface defects) of current commercial lithium-ion batteries (such as LCO and NMC) mainly relies on high-field magnetic resonance imaging (≥7 T) or synchrotron X-ray CT. Although high-field MRI can detect the phase perturbation of the ¹H signal in the external water medium of the battery through an "inside-out" strategy to invert the internal magnetic susceptibility change, its equipment is expensive, bulky, requires strict radio frequency and magnetic shielding, and is extremely sensitive to small metal parts in the battery (such as connectors and debris), easily producing serious artifacts or even system failures, making it difficult to deploy in industrial production lines. Summary of the Invention

[0007] To address the aforementioned problems in the existing technology, a method for detecting the state of charge of a battery based on closed-loop adaptive ultra-low field MRI is provided; furthermore, a system for implementing this method is also provided.

[0008] The specific technical solution is as follows: A method for detecting the state of charge (SOC) of a battery based on closed-loop adaptive ultra-low field MRI, comprising: Step S1: Immersing the battery under test in an inorganic salt solution and charging and discharging the battery under test, while using an ultra-low field magnetic resonance system to acquire the phase change of the water molecule H signal caused by the magnetic field disturbance of the battery under test to reconstruct a first signal phase image; Step S2: Using a SOC detection model, feature extraction is performed on the first signal phase image to obtain a high-information-content region mask, and then the magnetic pole region indicated by the high-information-content region mask is resampled using the ultra-low field magnetic resonance system to obtain a second signal phase image; Step S3: Inputting the second signal phase image into the SOC detection model to predict the SOC value of the battery under test; During the sampling process, a dual-echo gradient echo sequence is used to acquire and reconstruct the phase of the water molecule H signal to obtain the first signal phase image and the first signal phase image.

[0009] On the other hand, the state of charge detection model includes a spatial feature learning module and a state of charge detection module with shared connectivity; the first signal phase image is input to the spatial feature learning module and segmented by a codec to obtain the high-information region mask; the second signal phase image is input to the spatial feature learning module and the state of charge detection module and the state of charge value is predicted.

[0010] On the other hand, the spatial feature learning module includes: a phase image mask layer, which converts the input signal phase image into a mask phase image; a first encoder, which is connected to the phase image mask layer; the first encoder extracts features from the mask phase image through a series of convolutional layers to obtain a convolutional image; the first encoder sends the convolutional image to the shared connection; and a first decoder, which is connected to the first encoder; the first decoder upsamples the convolutional image through a series of upsampling layers to reconstruct the high-information-content region mask.

[0011] On the other hand, the state of charge detection module includes: a phase image layer that receives the signal phase image; a second encoder connected to the phase image layer and the shared connection; the second encoder extracts high-dimensional features based on the signal phase image and the convolutional image; a linear processing layer that performs linear processing on the high-dimensional features to obtain linear features; a ReLU activation function layer connected to the linear processing layer; the ReLU activation function layer performs non-linear processing on the linear features using the ReLU activation function to obtain non-linear features; a temporary de-energization layer connected to the ReLU activation function layer; the temporary de-energization layer performs regularization processing on the non-linear features to obtain temporary de-energization features; and a regression layer connected to the temporary de-energization layer; the regression layer predicts the state of charge value based on the temporary de-energization features.

[0012] On the other hand, before performing step S1, the method further includes: step A01: acquiring cavity phase signal image of water tank; then, during the acquisition of the first signal phase image and the second signal phase image, the cavity phase signal image is subtracted in advance.

[0013] On the other hand, the acquisition process of acquiring the first signal phase image or the second signal phase image in steps S1 and S2 includes: step B01: acquiring a signal sequence using the ultra-low field magnetic resonance system; step B02: denoising the signal sequence in the frequency domain using the Hamming method to obtain a denoised sequence; step B03: reconstructing the image from the denoised sequence, and then performing rigid body registration to obtain a signal phase image; the signal phase image is the first signal phase image or the second signal phase image.

[0014] On the other hand, the state of charge detection module is trained based on MSE regression loss and contrastive learning loss.

[0015] On the other hand, during the execution of steps S1 and S2, the double echo time of the double echo gradient echo sequence is adjusted according to the type of battery under test.

[0016] On the other hand, thermal drift compensation is also performed based on the cavity phase map during the detection process.

[0017] A detection system for implementing the above-described battery state-of-charge detection method.

[0018] The above technical solution has the following advantages or beneficial effects: Addressing the issue that existing battery magnetic resonance testing solutions require high-field magnetic resonance systems, which are unsuitable for production line use, this embodiment combines ultra-low field MRI with small-sample deep learning for the first time. This solves four major industrial pain points in commercial battery non-destructive diagnostics: cost, compatibility, portability, and data efficiency. It provides a feasible technical path for battery production quality control, condition monitoring, and health assessment. By designing a customized dual-echo gradient echo sequence and combining it with resampling of high-information regions, it overcomes key challenges such as weak signals, high noise interference, and scarce labeled data under ultra-low field conditions. This provides a new, scalable path for non-destructive quality control, internal defect screening, and online SOC monitoring in battery production. Attached Figure Description

[0019] Embodiments of the invention will be described more fully with reference to the accompanying drawings. However, the drawings are for illustration and explanation only and do not constitute a limitation on the scope of the invention.

[0020] Figure 1 This is an overall schematic diagram of an embodiment of the present invention;

[0021] Figure 2 This is a schematic diagram of the state of charge detection model in an embodiment of the present invention;

[0022] Figure 3 This is a schematic diagram of step S01 in an embodiment of the present invention;

[0023] Figure 4 This is a schematic diagram of the data acquisition process in an embodiment of the present invention;

[0024] Figure 5 This is a schematic diagram of phase diagram analysis in an embodiment of the present invention;

[0025] Figure 6 This is a schematic diagram comparing the detection performance of the SOCNet model in an embodiment of the present invention;

[0026] Figure 7 This is a schematic diagram illustrating the comparative analysis of MAE with different confusion ratios and prediction models in an embodiment of the present invention;

[0027] Figure 8This is a schematic diagram of LCO battery performance evaluation, which is not seen in the embodiments of the present invention. Detailed Implementation

[0028] 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.

[0029] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0030] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0031] This invention includes: a method for detecting the state of charge (SOC) of a battery based on closed-loop adaptive ultra-low field MRI, comprising: Step S1: immersing the battery under test in an inorganic salt solution and charging and discharging the battery, while simultaneously using an ultra-low field magnetic resonance system to acquire the phase change of the water molecule H signal induced by the magnetic field disturbance of the battery under test to reconstruct a first signal phase image; Step S2: using a SOC detection model to extract features from the first signal phase image to obtain a high-information-content region mask, and then using an ultra-low field magnetic resonance system to resample the magnetic pole region indicated by the high-information-content region mask to obtain a second signal phase image; Step S3: inputting the second signal phase image into the SOC detection model to predict the SOC value of the battery under test; during the sampling process, a dual-echo gradient echo sequence is used to acquire and reconstruct the phase of the water molecule H signal to obtain the first signal phase image and the first signal phase image.

[0032] Specifically, addressing the issues of existing battery magnetic resonance detection schemes requiring high-field magnetic resonance systems, high MRI costs, electromagnetic shielding requirements, and sensitivity to small metal components within the battery, making them unsuitable for industrial applications, this embodiment replaces the detection equipment with an ultra-low-field magnetic resonance system. This eliminates the need for shielding chambers and other structures found in traditional magnetic resonance systems. Furthermore, by first locating high-information-content regions and then resampling those regions, imaging accuracy is improved, enhancing the accuracy of the detection model's state of charge estimation. It requires no radio frequency or magnetic shielding, is insensitive to small metal components, can be imaged in open environments, and has low hardware costs.

[0033] Specifically, the above technical solution is mainly implemented as a software example in an ultra-low field magnetic resonance system for estimating the state of charge of a battery. The ultra-low field magnetic resonance system typically has a pair of permanent magnets to form a directional magnetic field, the detection area of ​​which is located. The field strength is typically below 0.5T.

[0034] It should be noted that the above technical solution is also applicable to low-field magnetic resonance systems with a field strength of less than 1.5T. However, since low-field magnetic resonance systems have larger magnet components and shielding structures, it can be regarded as a modified technical solution. Nevertheless, the above detection method is also applicable to this type of low-field magnetic resonance system.

[0035] The principle behind using magnetic resonance imaging (MRI) to detect the electrochemical processes of a battery lies in the fact that when an electrochemical reaction occurs, such as the extraction or insertion of active materials, it induces a change in local magnetic susceptibility within the battery, perturbing the external main magnetic field—that is, the main magnetic field of the ultra-low field magnetic resonance system. This magnetic field perturbation modulates the phase of the H signal of water molecules, which can then be detected by the ultra-low field MRI system. Furthermore, the degree of phase perturbation changes with the battery's state of charge (SOC), allowing the current SOC value of the battery to be deduced based on this principle.

[0036] Meanwhile, considering the actual production line needs, the finished battery is usually packaged in a metal casing. In this case, the metal casing will shield the gradient sequence of the magnetic resonance system. Therefore, in this embodiment, the phase change of the H signal of water molecules caused by magnetic field disturbance is indirectly measured to indirectly reflect the magnetic susceptibility disturbance caused by the lithiumization state inside the battery, which is beneficial for production line use.

[0037] To achieve this process, a set of detection tanks filled with inorganic salt solutions, typically a mixture of NiCl2 and NaCl, is added to the detection area of ​​the ultra-low field magnetic resonance system. The battery is immersed in the inorganic salt solution. In actual production line applications, specific tests are usually required for various states of charge (SOCs) of the battery, such as AC impedance testing. This necessitates connecting the positive and negative terminals of the battery to external testing instruments to perform specific charge-discharge and testing processes. During this process, the battery under test experiences magnetic field disturbances, which in turn cause a phase change in the H signal of the water molecules in the inorganic salt solution.

[0038] The battery under test can be a lithium-ion battery or a sodium-ion battery, etc.

[0039] Based on the above system, at the start of the test, the battery under test is immersed in an inorganic salt solution to initiate an electrochemical process, which is achieved through an external testing instrument. During this process, the ultra-low field magnetic resonance system measures the H signal in the water tank according to a pre-configured GRE sequence to form an echo sequence. This echo sequence characterizes the phase change of the H signal. Subsequently, the ultra-low field magnetic resonance system performs image reconstruction on the echo sequence to obtain the first signal phase image.

[0040] During the imaging process, a customized dual-echo GRE sequence was designed, such as TE1=12.9 ms and TE2=35.78 ms, to stably capture magnetically sensitive phase signals under ultra-low field conditions, enabling high-sensitivity capture of weak magnetic field disturbances caused by battery electrochemical activity.

[0041] For the first signal phase image, a pre-trained state of charge detection model is used for processing. This state of charge detection model can extract the part of the image with high phase perturbation signal to obtain a high information area mask. This part of the information usually corresponds to the area with large local magnetic susceptibility changes inside the battery.

[0042] For the magnetic pole regions indicated by the high-information-content mask, the imaging clarity of these regions is improved by adjusting the sequence parameters. Subsequently, local high-resolution resampling is performed on key regions such as the magnetic poles to obtain the second signal phase image, thereby enhancing the defect detection capability.

[0043] Subsequently, the phase image of the second signal is fed into the state of charge detection model to estimate the state of charge value.

[0044] In one embodiment, the state of charge detection model includes a spatial feature learning module 1 and a state of charge detection module 2 with shared connectivity; a first signal phase image is input to the spatial feature learning module 1 and segmented by a codec to obtain a high-information region mask; a second signal phase image is input to the spatial feature learning module 1 and the state of charge detection module 2 and the state of charge value is predicted.

[0045] Specifically, in order to achieve better recognition of the phase image of the H signal, in this embodiment, a spatial feature learning module 1 and a state of charge detection module 2 with shared connectivity are set in the state of charge detection model, which form a fusion architecture of Few-shot Learning + Self-supervised + Contrastive Learning.

[0046] Specifically, the spatial feature learning module 1 can segment the magnetic pole regions with relatively dense phase signals from the input H signal phase image and output them as high-information region masks. During the segmentation process, the spatial feature learning module 1 can extract high-dimensional features from the phase image and pass them to the charge state detection module 2 through shared connections to avoid repeated feature extraction. Furthermore, the spatial feature learning module 1 extracts general magnetic perturbation features from the unlabeled phase image through self-supervised mask reconstruction pre-training.

[0047] The State of Charge (SOC) detection module 2 can extract signal intensity-related features from the input H signal phase image, thereby deducing the local magnetic susceptibility changes in the battery and predicting the current SOC value based on the pre-trained process. By freezing the pre-trained encoder and combining comparative learning—including bringing similar SOC samples closer together and pushing different SOC samples further apart—the module achieves high-precision SOC estimation with only a very small amount of labeled data, such as 19 phase images.

[0048] Based on the above settings, SOC can be accurately predicted on new, unseen batteries without retraining, reducing MAE by more than 50% compared to traditional models such as CNN and SVR.

[0049] In one embodiment, the spatial feature learning module 1 includes: a phase image mask layer 11, which converts the input signal phase image into a mask phase image; a first encoder 12, which is connected to the phase image mask layer 11; the first encoder 12 extracts features from the mask phase image through a series of convolutional layers to obtain a convolutional image; the first encoder 12 sends the convolutional image to a shared connection; and a first decoder 13, which is connected to the first encoder 12; the first decoder 13 upsamples the convolutional image through a series of upsampling layers to restore a high-information-content region mask.

[0050] Specifically, in order to achieve better segmentation of spatial information, this embodiment introduces a set of encoder-decoder structures. When the signal phase image is acquired, the phase image mask layer 11 converts the input signal phase image into a mask phase image. Then, the first encoder 12 and the first decoder 13 obtain the high information region mask from the image through the network structure of convolutional network-upsampling layer.

[0051] At the same time, the first encoder 12 also sends the extracted high-dimensional features into the shared connection for feature transfer.

[0052] The aforementioned spatial feature learning module 1 is a self-supervised spatial feature learning module that learns robust spatial representations from unlabeled phase maps through a random mask-reconstruction strategy.

[0053] In one embodiment, the state of charge detection module 2 includes: a phase image layer 21, which receives a signal phase image; a second encoder 22, which is connected to the phase image layer 21 and a shared connection; the second encoder 22 extracts high-dimensional features based on the signal phase image and the convolutional image; a linear processing layer 23, which performs linear processing on the high-dimensional features to obtain linear features; a ReLU activation function layer 24, which is connected to the linear processing layer 23; the ReLU activation function layer 24 performs nonlinear processing on the linear features using the ReLU activation function to obtain nonlinear features; a temporary de-energization layer 25, which is connected to the ReLU activation function layer 24; the temporary de-energization layer 25 performs regularization processing on the nonlinear features to obtain temporary de-energization features; and a regression layer 26, which is connected to the temporary de-energization layer 25; the regression layer 26 predicts the state of charge value based on the temporary de-energization features.

[0054] Specifically, to achieve better prediction of the state of charge, in this embodiment, the signal phase image is first received through the phase image layer 21, and then the second encoder 22 extracts high-dimensional features based on the signal phase image and the convolutional image passed from the shared connection. After extracting the high-dimensional features, the linear processing layer 23 performs linear processing on the high-dimensional features to obtain linear features, the ReLU activation function layer 24 performs nonlinear processing on the linear features using the ReLU activation function to obtain nonlinear features, and the deregulation layer 25 performs regularization processing on the nonlinear features to obtain deregulation features.

[0055] Finally, the regression layer classifies the extracted temporary features to determine the state of charge value corresponding to the current signal features.

[0056] In one embodiment, the state of charge detection module 2 is trained based on MSE regression loss and contrastive learning loss.

[0057] Specifically, in order to achieve better training results for the state of charge detection model, in this embodiment, a projection layer is also provided in the state of charge detection module 2, which is connected to the temporary de-energization layer 25, and is used to predict the migration of active material of the electrode based on the magnetic susceptibility change information processed by the temporary de-energization layer 25.

[0058] Among them, the MES regression loss is used to measure the regression layer 26, and the contrastive learning loss is used to measure the temporary regression layer 25. The combination of the two achieves a better training effect for the model.

[0059] In one embodiment, before performing step S1, the method further includes: step A01: acquiring cavity phase signal images of the water tank; then, during the acquisition of the first signal phase image and the second signal phase image, the cavity phase signal image is subtracted in advance.

[0060] Specifically, in order to remove environmental noise, in this embodiment, before the test begins, the magnetic resonance system is pre-controlled to use the same sequence parameters as the actual inspection to collect cavity phase signal images in the water tank as baseline images.

[0061] Subsequently, during each acquisition of the first signal phase image and the second signal phase image, the cavity phase signal image is subtracted from the acquired and reconstructed signal phase image in advance, and the subtracted image is then output as the actual first signal phase image or the second signal phase image, in order to highlight the local disturbance caused by the battery and eliminate the inherent non-uniformity of the permanent magnet.

[0062] In one embodiment, the acquisition process of acquiring the first signal phase image or the second signal phase image in steps S1 and S2 includes: step B01: acquiring the signal sequence using an ultra-low field magnetic resonance system; step B02: denoising the signal sequence in the frequency domain using the Hamming method to obtain a denoised sequence; step B03: reconstructing the image from the denoised sequence, and then performing rigid body registration to obtain the signal phase image; the signal phase image is the first signal phase image or the second signal phase image.

[0063] Specifically, to achieve better image denoising, in this embodiment, after acquiring the original signal sequence, the signal sequence is first denoised in the frequency domain using the Hamming method to obtain a denoised sequence, thereby improving the signal-to-noise ratio. Subsequently, image reconstruction is performed on the denoised sequence. During this process, the magnetic field disturbance information can be highlighted by subtracting the baseline map, followed by rigid body registration to obtain the signal phase image, compensating for the translation error caused by magnetic field drift.

[0064] The combination of the above processing methods effectively removes noise and magnetic field inhomogeneity in the production environment, thereby enabling the accurate extraction of high-quality phase maps related to the battery's state of charge (SOC).

[0065] In one embodiment, the ultra-low field magnetic resonance system is pre-configured with multiple TE sequences, and during the execution of step S1, the TE sequences are adjusted according to the type of battery under test.

[0066] Specifically, to achieve better imaging results, this embodiment designs a customized dual-echo GRE sequence for imaging of the magnetic resonance system, for example, TE1=12.9 ms, TE2=35.78 ms, to stably capture the magnetically sensitive phase signal under ultra-low field conditions. Meanwhile, considering the different rates of change of local magnetic fields generated by different battery types, the dual-echo time is automatically optimized according to the battery type (LCO / NMC) to maximize the magnetic disturbance signal-to-noise ratio.

[0067] In one embodiment, thermal drift compensation is also performed based on the cavity phase map during the detection process.

[0068] Specifically, in order to achieve better imaging results, this embodiment also performs thermal drift compensation based on the cavity phase map during the detection process to improve long-term stability.

[0069] A detection system for implementing the above-described battery state-of-charge detection method.

[0070] Figure 5 Phase diagram analysis was performed. During the charging and subsequent discharging of the NMC battery, phase diagram sequences were acquired at fixed intervals, with each phase diagram labeled with the remaining capacity. All phase diagrams were normalized using a baseline diagram with no battery. a) Imaging in the battery-free state. b) Imaging in the battery-equipped state. c) The phase signal shows a decreasing trend as the discharge process progresses. d) The phase signal shows an increasing trend as the discharge process progresses. e) Phase distribution histogram during charging (zoom function recommended for viewing). f) Phase distribution histogram during discharging (zoom function recommended for viewing). Principal component analysis of the continuous phase diagrams, diagrams g and h, also demonstrate the dynamic relationship between SOC and phase signal. g) During charging, the phase signal gradually decreases as SOC increases. h) During discharging, the phase signal gradually increases as SOC decreases.

[0071] PCA analysis and statistical distribution confirmed that NMC and LCO batteries exhibit reversible and material-specific phase change patterns during charge and discharge.

[0072] Figure 6 Performance comparison of the proposed SOCNet model: a) Comparison of SOC values ​​between predicted and actual values ​​for LCO batteries; b) Comparison of SOC values ​​between predicted and actual values ​​for NMC batteries; c) Absolute error distribution of each sample in 25 random test datasets for LCO batteries; d) Absolute error distribution of each sample in 25 random test datasets for NMC batteries; e) Statistical analysis of MAE, RMSE, and R2 for 25 random test datasets for LCO and NMC batteries; f) Diagonal affinity matrices of LCO and NMC battery blocks generated by our proposed tensor multi-view clustering method, where elements represent the similarity between battery phase diagrams.

[0073] Figure 7 This section presents a comparative analysis of MAE (Magnitude of Effect) across different obfuscation ratios and prediction models. Figures a and b show the MAE results for LCO and NMC battery packs with varying training sample sizes (5, 10, 15, and 19). Figures c and d illustrate the MAE performance of LCO and NMC battery packs using four different obfuscation ratios during data augmentation. Figures e and f compare the MAE results of different prediction models on LCO and NMC battery packs.

[0074] Figure 8This section presents the performance evaluation of unseen LCO cells. Phase diagram data were collected for two newly acquired LCO cells (labeled LCO#1 and LCO#2) at different states of charge (SOC), with the order of data randomly shuffled. Figures a and b show the predicted SOC curves for LCO#1 and LCO#2, respectively. The mean and standard deviation of 25 repeated predictions are provided for each real data point. Figures c and d use violin plots to show the predicted MAE, RMSE, and R² scores for LCO#1 and LCO#2, respectively. Figures e and f present the MAE results for LCO#1 and LCO#2 under different prediction models.

[0075] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made based on the description and illustrations of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting the state of charge of a battery based on closed-loop adaptive ultra-low field MRI, characterized in that, include: Step S1: Immerse the battery under test in an inorganic salt solution and charge and discharge the battery under test. At the same time, use an ultra-low field magnetic resonance system to collect the phase change of the water molecule H signal caused by the magnetic field disturbance of the battery under test to reconstruct the first signal phase image. Step S2: Use the state of charge detection model to extract features from the first signal phase image to obtain a high-information-content region mask. Then, use the ultra-low field magnetic resonance system to resample the magnetic pole region indicated by the high-information-content region mask to obtain the second signal phase image. Step S3: Input the second signal phase image into the state of charge detection model to predict the state of charge value of the battery under test; During the sampling process, a dual-echo gradient echo sequence is used to acquire and reconstruct the phase of the H signal of water molecules to obtain the first signal phase image and the second signal phase image.

2. The battery state-of-charge detection method according to claim 1, characterized in that, The state of charge detection model includes a spatial feature learning module with shared connectivity and a state of charge detection module. The first signal phase image is input into the spatial feature learning module and segmented by a codec to obtain the high-information region mask; The second signal phase image is input into the spatial feature learning module and the state of charge detection module to predict the state of charge value.

3. The battery state-of-charge detection method according to claim 2, characterized in that, The spatial feature learning module includes: A phase image mask layer that converts the input signal phase image into a mask phase image; A first encoder, the first encoder being connected to the phase image mask layer; The first encoder extracts features from the mask phase image through a series of convolutional layers to obtain a convolutional image; The first encoder feeds the convolutional image into the shared connection; A first decoder, which is connected to the first encoder; The first decoder upsamples the convolutional image through a series of upsampling layers to restore the high-information-content region mask.

4. The battery state-of-charge detection method according to claim 3, characterized in that, The state of charge detection module includes: A phase image layer that receives the phase image of the signal; A second encoder, which connects the phase image layer and the shared connection; The second encoder extracts high-dimensional features based on the signal phase image and the convolutional image; A linear processing layer performs linear processing on the high-dimensional features to obtain linear features; A ReLU activation function layer, wherein the ReLU activation function layer is connected to the linear processing layer; The ReLU activation function layer applies the ReLU activation function to the linear features to perform nonlinear processing, resulting in nonlinear features. Temporary fallback layer, which is connected to the ReLU activation function layer; The temporary de-extension layer performs regularization on the nonlinear features to obtain the temporary de-extension features; The regression layer is connected to the temporary retreat layer; The regression layer predicts the state of charge value based on the temporary de-energization feature.

5. The battery state-of-charge detection method according to claim 1, characterized in that, Before performing step S1, the method further includes: Step A01: Acquire the cavity phase signal image of the water tank; During the acquisition of the first signal phase image and the second signal phase image, the cavity phase signal image is subtracted in advance.

6. The battery state-of-charge detection method according to claim 5, characterized in that, The acquisition process of acquiring the first signal phase image or the second signal phase image in steps S1 and S2 includes: Step B01: Acquire signal sequences using the ultra-low field magnetic resonance system; Step B02: Denoise the signal sequence in the frequency domain using the Hamming method to obtain a denoised sequence; Step B03: Perform image reconstruction on the denoised sequence, followed by rigid body registration to obtain the signal phase image; The signal phase image is either the first signal phase image or the second signal phase image.

7. The battery state-of-charge detection method according to claim 2, characterized in that, The state of charge detection module is trained based on MSE regression loss and contrastive learning loss.

8. The battery state-of-charge detection method according to claim 1, characterized in that, During the execution of steps S1 and S2, the double echo time of the double echo gradient echo sequence is adjusted according to the type of battery under test.

9. The battery state-of-charge detection method according to claim 6, characterized in that, During the detection process, thermal drift compensation is also performed based on the cavity phase signal image.

10. A detection system, characterized in that, Used to implement the battery state of charge detection method as described in any one of claims 1-9.

Images