Lithium battery charging voltage curve reconstruction method based on double discriminator CGAN

Abstract

This invention discloses a method for reconstructing lithium battery charging voltage curves based on a dual-discriminator CGAN, relating to the field of lithium-ion battery health management and data completion. The method includes: preprocessing lithium-ion battery charging voltage data to output a standardized voltage sequence; dividing the scenario into two-sided missing sections based on actual operating conditions and generating corresponding missing masks and conditional vectors; constructing a dual-discriminator conditional generative adversarial network model containing a generator, a global voltage discriminator, and a local pattern discriminator; training the dual-discriminator conditional generative adversarial network model using a course learning strategy and a composite loss function; after training, saving the model parameters with the best reconstruction accuracy on the validation set and deploying the optimal model in the battery management system's data completion process to reconstruct two-sided missing charging voltage data appearing in actual operating conditions in real time. This invention can adapt to two-sided missing scenarios, maintain high reconstruction accuracy even with a high proportion of missing data, and has good industrial deployability.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery health management and data completion technology, and more specifically to a method for reconstructing lithium battery charging voltage curves based on dual discriminator CGAN. Background Technology

[0002] With the widespread deployment of new energy vehicles and mobile energy storage devices, the constant current charging voltage curve recorded in real time by the battery management system (BMS) has become a key data source for assessing battery state of health (SOH), remaining useful life (RUL), and safety status. However, in actual vehicle or portable applications, due to sensor failures, communication interruptions, user-initiated power outages, and transient interference, the collected voltage curve often suffers from large-scale missing segments, causing downstream battery health assessment algorithms to fail. Therefore, there is an urgent need for a battery charging voltage curve reconstruction method that can adapt to two-sided missing scenarios, maintain high reconstruction accuracy even with severe missing proportions, and possesses good industrial deployability. Summary of the Invention

[0003] In view of this, the present invention provides a method for reconstructing the charging voltage curve of a lithium battery based on a dual discriminator CGAN, which solves the problems existing in the background technology.

[0004] To achieve the above objectives, the present invention provides the following technical solution: A method for reconstructing the charging voltage curve of a lithium battery based on a dual discriminator CGAN includes the following steps: S1. Preprocess the lithium-ion battery charging voltage data and output a standardized voltage sequence; S2. Based on actual working conditions, divide the two-sided missing scenarios and generate corresponding missing masks and condition vectors; S3. Construct a dual-discriminator conditional generative adversarial network model that includes a generator G, a global voltage discriminator D1, and a local pattern discriminator D2. S4. Employ a course-based learning strategy and combine it with a composite loss function to train a dual-discriminator conditional generative adversarial network model. S5. After training, save the model parameters with the best reconstruction accuracy on the validation set and deploy the best model in the battery management system data completion process to reconstruct the bilateral missing charging voltage data that appears in actual working conditions in real time.

[0005] Optionally, in S1, when preprocessing the lithium-ion battery charging voltage data, each charging cycle is resampled into a voltage sequence of fixed length, and the constant current charging cutoff voltage is determined according to the characteristics of the battery chemical system to locate the end point of the constant current segment.

[0006] Optionally, in S2, the two-sided missing scenarios include three categories: mild, typical, and severe, each corresponding to different observed SOC intervals; for each charging cycle, a missing mask is generated according to the specified two-sided missing scenario, and a polynomial prediction is applied between the known segment and the missing segment to construct a condition vector.

[0007] Optionally, in S3, the generator G is used to receive a random noise vector, a conditional vector, and a missing mask as inputs, and outputs the reconstructed complete charging voltage curve. Specifically, the generator G adopts a U-Net encoder-decoder structure, which includes a four-level downsampling encoder, a bottleneck layer, and a three-level upsampling decoder. A skip connection is set between the four-level downsampling encoder and the three-level upsampling decoder. The last layer of the generator G uses the Tanh activation function to limit the output to the normalized voltage range.

[0008] Optionally, in S3, the global voltage discriminator D1 consists of multiple one-dimensional convolutional and pooling layers. It receives the concatenation of the complete charging voltage curve, conditional vector, and missing mask as input, and outputs a... The scalar values ​​between them are used to determine the global authenticity of the entire curve.

[0009] Optionally, in S3, the local pattern discriminator D2 is used to receive the complete charging voltage curve as input and output a position-by-position true / false probability sequence of the same length as the complete charging voltage curve. Specifically, the local pattern discriminator D2 is a one-dimensional fully convolutional network, with each one-dimensional convolutional layer followed by a spectral normalization and LeakyReLU activation function layer.

[0010] Optionally, in S4, the course learning strategy is as follows: during training rounds 1-100, only mild and typical bilateral missing scenarios are sampled; during training rounds 101-300, mild, typical, and severe missing scenarios are sampled using a probability weighted mixture of 0.4:0.4:0.2; during training rounds 301-800, mild, typical, and severe missing scenarios are sampled uniformly; at the same time, training samples are sampled evenly according to the SOH interval.

[0011] Optionally, in S4, the formula for the composite loss function is as follows:

[0012] In the formula: To combat losses in D1, To combat losses for D2, To reconstruct the loss, For smoothness loss; , These are learnable uncertainty parameters used to adaptively balance the adversarial losses D1 and D2. , For the corresponding precision weights, , To prevent , An infinitely increasing regular term; , Fixed weighting coefficients for reconstruction loss and smoothness loss; Reconstruction loss Using asymmetric head-and-tail weighted L1 loss, the expression is:

[0013] Where: Reconstruction loss weight The weight ratio for missing head and missing tail positions is 1:2, and 0 is used for observed positions. For sequence position index, For the reconstructed voltage sequence after mask fusion at position The voltage value at that location, For the true voltage sequence in Voltage value at; Smoothing loss A scale-adaptive weighting method is used to impose constraints on the second-order difference of the generated segment, and The value gradually decreased from 1.0 to 0.1 as training progressed.

[0014] Optionally, in S4, when training the dual-discriminator conditional generative adversarial network model, a label smoothing technique is used. Specifically, the labels of the real samples are changed from a fixed value of 1.0 to a value of... Uniform sampling within a given interval changes the label of the generated samples from a fixed 0.0 to... Uniform sampling is performed across intervals. The generator G, global voltage discriminator D1, and local pattern discriminator D2 are updated alternately. In each training step, the generator G is updated twice, while the global voltage discriminator D1 and the local pattern discriminator D2 are updated once each.

[0015] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a lithium battery charging voltage curve reconstruction method based on dual discriminator CGAN, which has the following beneficial effects: (1) Compared with traditional mathematical extrapolation methods (constant filling, polynomial fitting), the present invention significantly reduces the average absolute error of voltage reconstruction in the bilateral missing scenario, and the reconstruction curve can effectively retain the nonlinear phase transition characteristics in the battery charging process. In particular, the morphological recovery accuracy of the phase transition region at the tail of the constant current section is significantly better than that of the traditional method.

[0016] (2) Compared with the single-discriminator conditional generative adversarial network method without using a second discriminator, the dual-discriminator architecture of the present invention (including a global voltage discriminator D1 and a local pattern discriminator D2) reduces the average reconstruction error of the optimal validation set under three-scene mixed training from about 2.60mV to about 2.45mV, with a relative improvement of about 5.8%, while keeping all other configurations such as generator structure, training protocol, and loss function exactly the same. This improvement is due to the position-by-position local pattern supervision signal provided by the local pattern discriminator D2, which makes up for the shortcomings of a single global discriminator in constraining the authenticity of local details.

[0017] (3) The bilateral missing scenario defined in this invention is more in line with the non-fixed start and end SOC characteristics of user charging behavior under the actual working conditions of electric vehicles, thus expanding the industrial applicability of the battery voltage reconstruction method.

[0018] (4) The local pattern discriminator D2 of the present invention has only about 23,000 parameters, which is more than 80% less than the traditional discriminator using a fully connected layer (about 140,000 parameters), making it more suitable for deployment and operation in an embedded BMS.

[0019] (5) The voltage curve reconstructed by the present invention can be directly used as input for various downstream tasks such as battery health status assessment, remaining life prediction, and anomaly diagnosis, and has good methodological versatility. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0021] Figure 1 The overall flowchart of the lithium battery charging voltage curve reconstruction method based on dual discriminator CGAN provided by the present invention; Figure 2 The overall system architecture diagram of the dual discriminator generative adversarial network provided by this invention; Figure 3 This is a schematic diagram of a bilateral missing scenario provided by the present invention; Figure 4 The U-Net network structure diagram of the generator G provided by this invention; Figure 5 A schematic diagram of the one-dimensional fully convolutional network structure and receptive field of the local pattern discriminator D2 provided by the present invention; Figure 6 This is a schematic diagram of the course learning and training process provided by the present invention. Detailed Implementation

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

[0023] This invention discloses a method for reconstructing lithium battery charging voltage curves based on dual discriminator CGAN, such as... Figure 1 As shown, it includes the following steps: S1. Preprocess the lithium-ion battery charging voltage data and output a standardized voltage sequence.

[0024] Specifically, when preprocessing lithium-ion battery charging voltage data, each charging cycle is resampled into a voltage sequence of fixed length, and the constant current charging cutoff voltage is determined based on the characteristics of the battery chemical system to locate the end point of the constant current segment.

[0025] In this embodiment, lithium iron phosphate (LFP) chemistry battery data is used for illustration. The original data comes from standardized battery cycle tests. Each charging cycle includes a constant current (CC) charging segment and a constant voltage (CV) charging segment, and the voltage is sampled at non-uniform time intervals.

[0026] The preprocessing process specifically includes: removing non-charging segment data; and sorting data according to voltage upper and lower limits. The constant current segment is extracted; each charging cycle is resampled into an equally spaced voltage sequence of length 256 using linear interpolation; the voltage amplitude of each cycle is normalized according to the minimum and maximum values ​​of the entire batch of data, and then mapped to... The interval is defined. Simultaneously, the constant current segment endpoint index cc_end for each cycle is located based on the constant current cutoff voltage (3.595V) of the LFP system, which is used for physical boundary constraints in subsequent training.

[0027] S2. Based on actual working conditions, divide the two-sided missing scenarios and generate corresponding missing masks and condition vectors.

[0028] Specifically, the two-sided missing scenarios include three categories: Light, Typical, and Heavy, each corresponding to different observed SOC intervals. For each charging cycle, a missing mask is generated according to the specified two-sided missing scenario, and a polynomial prediction is applied between the known segment and the missing segment to construct a condition vector.

[0029] In this embodiment, as Figure 3 As shown, the mild scenario corresponds to the observed SOC range. This means that approximately 20% of the head is missing, while the tail is not; the typical scenario corresponds to the observed SOC range. That is, approximately 20% of the head and 20% of the tail are missing; severe scenarios correspond to the observed SOC range. That is, approximately 30% of the head and 30% of the tail are missing; during training, the specific SOC boundary points corresponding to each scene can be... Random jitter within a range is used to enhance the model's generalization ability.

[0030] For the normalized voltage sequence of each charging cycle Calculate the start and end indices of the missing segment in the sequence according to the selected scenario, and generate a binary missing mask of length 256. ,in Indicates position Voltage data at that location has been observed. Indicates position Voltage data is missing at this location. Condition vector. With the original voltage sequence Homomorphic, observed location Missing position The quadratic polynomial predictions from the observed segments are used to fill in the gaps, serving as auxiliary inputs to the generator G; in The location of the subsequent constant pressure section will be uniformly referred to as... Fill in the blanks.

[0031] S3. Construct a dual-discriminator conditional generative adversarial network model containing a generator G, a global voltage discriminator D1, and a local pattern discriminator D2. For example... Figure 2 As shown, the specific design of this model is as follows: (1) Generator G (total number of parameters is about 5.8 million) is used to receive random noise vector, condition vector and missing mask as input, and output the reconstructed complete charging voltage curve.

[0032] In this embodiment, the generator G adopts a U-Net encoder-decoder structure, such as... Figure 4 As shown, it includes a four-level downsampling encoder, a bottleneck layer, and a three-level upsampling decoder; a skip connection is set between the four-level downsampling encoder and the three-level upsampling decoder to preserve multi-scale features, and the last layer of the generator G uses the Tanh activation function to limit the output to the normalized voltage range.

[0033] The encoder consists of four levels of one-dimensional convolutional blocks, with the number of output channels for each level being 48, 96, 192, and 384, respectively. Each convolutional block contains a Conv1d, a BatchNorm1d, and a LeakyReLU activation function.

[0034] The bottleneck layer is a Conv1d residual block that keeps the number of channels constant at 384.

[0035] The decoder consists of three levels of transposed convolutional blocks, with 192, 96, and 48 output channels for each level, respectively. Each level of the decoder has a skip connection to the corresponding layer of the encoder, allowing low-level detailed features to be directly passed to the corresponding decoding layer. The final layer of the decoder uses the Tanh activation function to constrain the output to a normalized voltage range.

[0036] (2) The global voltage discriminator D1 is composed of multiple one-dimensional convolution and pooling layers, and is used to receive the complete charging voltage curve. Conditional vector and missing mask Taking concatenation along the channel dimension as input, the output is a... The scalar values ​​between them are used to determine the global authenticity of the entire curve.

[0037] In this embodiment, the global voltage discriminator D1 consists of four levels of one-dimensional convolutional layers, each followed by a LeakyReLU activation function and spectral normalization. Finally, adaptive average pooling compresses the features into a scalar, which is then passed through a fully connected layer and a sigmoid function before being output. The true and false probabilities between them represent the judgment result of the global authenticity of the entire voltage curve.

[0038] (3) The local pattern discriminator D2 (with a total of approximately 23,000 parameters) is used to receive the complete charging voltage curve as input and output a position-by-position true / false probability sequence with the same length as the complete charging voltage curve.

[0039] Specifically, the local pattern discriminator D2 is a one-dimensional fully convolutional network. Each one-dimensional convolutional layer is followed by a spectral normalization and LeakyReLU activation function layer, without any fully connected layers or global pooling layers. D2 directly receives the complete charging voltage curve as input without needing to perform differentiation or other derivative transformations. The output of D2 is a one-dimensional probability sequence of the same length as the complete charging voltage curve, where each element represents the probability of the local pattern being genuine or false at the corresponding location. The final receptive field of D2 is designed to be 15 to 25 sampling points, meaning that each output location only discriminates the authenticity of the pattern within its surrounding local window.

[0040] In this embodiment, as Figure 5As shown, the local mode discriminator D2 is formed by cascading five layers of one-dimensional convolution. The first four layers all adopt spectral normalized convolution plus LeakyReLU activation function. The convolution kernel sizes are 7, 5, 5, 5, and 3 in sequence, and the number of output channels is 16, 32, 64, 32, and 1 in sequence. The effective receptive field of the entire network is 21 sampling points, accounting for about 8.2% of the total length of the voltage sequence. The last layer Conv1d is followed by a Sigmoid function, which outputs a probability sequence of the same length as the input voltage curve , wherein the output at each position represents the authenticity probability of the local mode of 21 sampling points around the position.

[0041] S4. A curriculum learning strategy is adopted, and a dual-discriminator conditional generative adversarial network model is trained with a composite loss function, which gradually transitions from light scenarios to mixed weighted scenarios and then to full scenarios according to the training stage.

[0042] In this embodiment, the training process is as Figure 6 shown and is divided into the following three stages: (1) The first stage is the basic reconstruction stage (the 1st to 100th rounds of training). Only light scenarios (Light) and typical scenarios (Typical) are sampled, each accounting for 50% of the probability, so that the generator G first learns the basic reconstruction capability on simple tasks.

[0043] (2) The second stage is the weighted mixing stage (the 101st to 300th rounds of training). Heavy scenarios (Heavy) are introduced, and three types of missing scenarios, namely light, typical and heavy, are weighted and mixed for sampling according to the probability ratio of 0.4:0.4:0.2, so that the generator G gradually adapts to the missing task.

[0044] (3) The third stage is the full-scenario training stage (the 301st to 800th rounds of training). Uniform sampling is performed on three types of missing scenarios, namely light, typical and heavy, so that the generator G achieves stable convergence in all difficulty scenarios.

[0045] In all training stages, the training samples of each mini-batch are balanced and sampled according to SOH intervals. SOH is divided into three intervals: new battery (SOH>0.9), medium aging (0.7<SOH≤0.9) and deep aging (SOH≤0.7). Each interval contributes an equal proportion of training samples, preventing the model from biasing towards a certain aging interval with more samples in the training data.

[0046] Further, the formula of the composite loss function is as follows:

[0047] In the formula: , are learnable uncertainty parameters, used to adaptively balance the two adversarial losses of D1 and D2; , For the corresponding precision weights, , To prevent , An infinitely increasing regular term; , These are the fixed weighting coefficients for reconstruction loss and smoothness loss.

[0048] To avoid the difficulties in parameter tuning and insufficient scenario adaptability caused by manually fixing the adversarial loss weights of the two discriminators, this embodiment adopts an adaptive weighting mechanism based on homoscedastic uncertainty to dynamically balance the two adversarial losses of the global voltage discriminator D1 and the local pattern discriminator D2. Specifically, two learnable parameters are introduced. and , representing the uncertainty of tasks D1 and D2 respectively; the more difficult a discriminator task is, the higher its corresponding precision weight. The smaller the value, the more automatically the discriminator influences the generator gradient. Simultaneously, the regularization term in the loss... and prevent The adversarial term is completely ignored as the scale increases indefinitely. This mechanism allows the model to automatically adjust its weights during training based on the relative magnitudes of the losses of the two discriminators, without requiring manual grid search; parameters , The network parameters of the generator G, discriminators D1 and D2 are added to the optimizer for synchronous updates. The reconstruction loss and smoothness loss, as the dominant supervision signals, retain their fixed weights and are not included in the adaptive weighting.

[0049] The D1 adversarial loss is defined as the binary cross-entropy. This is used to guide the generator G to deceive the global voltage discriminator D1.

[0050] To counteract the loss in D2, since the output of D2 is of length... The probability sequence is given by taking the average of BCE at all positions. This is used to guide the generator G to fool the local pattern discriminator D2.

[0051] Reconstruction loss Asymmetric head-tail weighted L1 loss is used to enhance the reconstruction accuracy of the tail phase transition region, and the expression is:

[0052] Where: Reconstruction loss weight The weight ratio for missing head and missing tail positions is 1:2, and 0 is used for observed positions. For sequence position index, For the reconstructed voltage sequence after mask fusion at position The voltage value at that location, For the true voltage sequence in The voltage value at that point. In the above scheme, the tail weight is higher than the head weight because the tail of the battery charging curve corresponds to the critical phase transition region where the battery is close to full charge, and the reconstruction error in this region has a greater impact on the downstream battery health status assessment.

[0053] Smoothing loss A scale-aware weighting method is used to impose constraints on the second-order difference of the generated segment, and The value gradually decreases from 1.0 to 0.1 as training progresses. This ensures the basic smoothness of the generated curve in the early stages of training, while allowing the model to recover local details of the voltage curve in the later stages.

[0054] Both employ standard adversarial loss plus labeled smoothing techniques:

[0055] In the formula: In order to be in Label matrix of uniformly sampled intervals. In order to be in Label matrix of uniformly sampled intervals. To apply a stopping gradient operation to the generator output, i.e., to cut off the gradient backpropagation to the generator when updating the discriminator, the discriminator loss will not be backpropagated to update the generator parameters; the label smoothing mechanism can prevent the discriminator from being overconfident in real and fake samples and enhance the stability of adversarial training. For D2, label smoothing acts independently at each position.

[0056] Discriminator loss here , The adversarial loss in the generator composite loss described above is one of two different sets of losses, specifically: , The optimization objective of generator G is to drive G to generate voltage curves that can deceive the discriminator. During calculation, the discriminator parameters are fixed and the gradient is only propagated back to G; while , It is the training objective of the discriminators D1 and D2 themselves, driving the discriminators to distinguish between real samples and generated samples.

[0057] Furthermore, when training the dual-discriminator conditional generative adversarial network model, a label smoothing technique is employed to enhance the stability of adversarial training. Specifically, the labels of the real samples are changed from a fixed 1.0 to... Uniform sampling within a given interval changes the label of the generated samples from a fixed 0.0 to... Uniform sampling is used in intervals; the generator G and the global voltage discriminator D1 and local pattern discriminator D2 are updated alternately. In each training step, the generator G is updated twice while the global voltage discriminator D1 and local pattern discriminator D2 are updated once each, in order to alleviate the problem of the generator G's gradient vanishing due to the global voltage discriminator D1 and local pattern discriminator D2 being too strong.

[0058] In this embodiment, the optimizer uses Adam, and the generator G has a learning rate of [missing information]. The learning rates of discriminators D1 and D2 are respectively , In each training step, the generator G is updated twice, and the discriminator is updated once each, to balance the training speed of G and D. The learning rate decays to 0.5 times its original value every 200 epochs using the StepLR scheduler. Training is performed on a single GPU, with each batch size of 32, and the total training time is approximately 4 to 8 hours. Two learnable parameters are adaptively weighted. , The network parameters of the optimizer are updated synchronously with those of the generator G and the discriminators D1 and D2. The learner rate is 0.1 times that of the generator to ensure that the weights evolve smoothly.

[0059] S5. After training, save the model parameters with the best reconstruction accuracy on the validation set and deploy the best model in the battery management system data completion process to reconstruct the bilateral missing charging voltage data that appears in actual working conditions in real time.

[0060] In this embodiment, after every 5 rounds of training, the average missing segment reconstruction error of generator G for the three scenes is evaluated from the validation set. The model parameters that minimize this metric are recorded as the final model. During the inference phase, the incomplete voltage curve to be completed, the condition vector, and the missing mask are input into the generator G to obtain the reconstructed curve. This does not rely on any SOH labels or other auxiliary information used during training, so it can be directly deployed in the online or offline data completion process of the battery management system.

[0061] To verify the effectiveness of the method proposed in this embodiment, simulation experiments were conducted using the MATR LFP battery public dataset. Forty batteries were selected as the training set, and one battery as the independent validation set. All charging cycles for each battery were processed according to the preprocessing procedure described above. This embodiment is compared with the following comparative methods in terms of accuracy: (1) Comparison Method 1: Constant Padding, which is to directly fill the missing segment with the voltage values ​​of the adjacent endpoints of the observed segment.

[0062] (2) Comparison Method 2: Single-Discriminator Conditional Generative Adversarial Network (Single-D) method, which, while maintaining the same generator architecture, training protocol (including course learning and SOH equalization sampling), loss function (including head-tail asymmetric reconstruction loss and scale-adaptive smoothing loss) as described in this invention, removes the local pattern discriminator D2 and its corresponding adversarial loss, retaining only the global voltage discriminator D1. This comparison method is essentially an ablation experiment of the method described in this invention, used to directly verify the contribution of the local pattern discriminator D2 to the reconstruction accuracy.

[0063] The evaluation metric is the mean absolute error of the missing segment reconstruction ( ). ), the unit is millivolt (mV).

[0064] In a complete control experiment spanning 800 training epochs, the dual-discriminator conditional generative adversarial network method proposed in this invention outperformed the average reconstruction error of the validation set under mixed training in three missing scenarios. Furthermore, for the dual-discriminator method of this invention, the model repeatedly approached the optimal level during subsequent training after the optimal point was reached, indicating that this accuracy improvement is repeatable and not due to training randomness.

[0065] In the comparison of reconstruction results under three missing scenarios, the constant filling method showed a stepped discontinuous reconstruction curve in the severe scenario, which could not retain the nonlinear phase transition characteristics of the battery charging process at all; although the single discriminator method could recover the overall voltage shape, local detail distortion appeared in the tail phase transition region in the severe scenario; while the method of this invention (dual discriminator architecture + local mode discriminator D2) reconstructed voltage curves in all three scenarios that were highly consistent with the real curves, especially in the severe scenario, it could accurately recover the shape of the tail phase transition region.

[0066] In a quantitative comparison of reconstruction accuracy, the method of this invention significantly outperforms the constant filling method (accuracy improvement of over 70%) in average reconstruction error across all missing scenarios, and consistently outperforms the single discriminator control method. The relative advantage of the method of this invention is particularly significant in scenarios with severe missing data, meeting the core needs of battery management systems in real-world operating conditions when faced with large amounts of missing data, and providing high-quality input data for downstream battery health assessment.

[0067] In summary, to address the challenge of high-precision reconstruction of lithium-ion battery charging voltage curves in scenarios with bilateral missing parameters, this embodiment proposes a reconstruction method based on a dual-discriminator conditional generative adversarial network. By introducing a local pattern discriminator D2 that directly takes the voltage curve as input, a position-by-position supervision signal is provided to ensure the authenticity of the local patterns in the generated curve. This effectively compensates for the shortcomings of methods relying solely on a global discriminator in constraining the authenticity of local details. By employing a course learning strategy and a SOH equalized sampling training protocol, high reconstruction accuracy is maintained even in scenarios with severe missing parameters. Furthermore, by defining a bilateral missing parameter scenario that conforms to actual operating conditions, the applicability of the method in the practical deployment of electric vehicle BMS is expanded. The reconstructed curve provided by this invention can serve as a high-quality input for various downstream tasks such as battery health status assessment, remaining life prediction, and anomaly diagnosis.

[0068] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0069] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for reconstructing the charging voltage curve of a lithium battery based on a dual discriminator CGAN, characterized in that, Includes the following steps: S1. Preprocess the lithium-ion battery charging voltage data and output a standardized voltage sequence; S2. Based on actual working conditions, divide the two-sided missing scenarios and generate corresponding missing masks and condition vectors; S3. Construct a dual-discriminator conditional generative adversarial network model that includes a generator G, a global voltage discriminator D1, and a local pattern discriminator D2. S4. Employ a course-based learning strategy and combine it with a composite loss function to train a dual-discriminator conditional generative adversarial network model. S5. After training, save the model parameters with the best reconstruction accuracy on the validation set and deploy the best model in the battery management system data completion process to reconstruct the bilateral missing charging voltage data that appears in actual working conditions in real time.

2. The lithium battery charging voltage curve reconstruction method based on dual discriminator CGAN according to claim 1, characterized in that, In S1, when preprocessing the lithium-ion battery charging voltage data, each charging cycle is resampled into a voltage sequence of fixed length. The constant current charging cutoff voltage is determined based on the characteristics of the battery chemical system, and the end point of the constant current segment is located.

3. The lithium battery charging voltage curve reconstruction method based on dual discriminator CGAN according to claim 1, characterized in that, In S2, bilateral missing scenarios include three categories: mild, typical, and severe, each corresponding to different observed SOC intervals; For each charging cycle, a missing mask is generated according to the specified bilateral missing scenario, and a polynomial prediction is applied between the known segment and the missing segment to construct a condition vector.

4. The lithium battery charging voltage curve reconstruction method based on dual discriminator CGAN according to claim 1, characterized in that, In S3, generator G is used to receive random noise vector, condition vector, and missing mask as input, and output the reconstructed complete charging voltage curve. Specifically, the generator G adopts a U-Net encoder-decoder structure, which includes a four-level downsampling encoder, a bottleneck layer, and a three-level upsampling decoder. A skip connection is set between the four-level downsampling encoder and the three-level upsampling decoder. The last layer of the generator G uses the Tanh activation function to limit the output to the normalized voltage range.

5. The lithium battery charging voltage curve reconstruction method based on dual discriminator CGAN according to claim 4, characterized in that, In S3, the global voltage discriminator D1 consists of multiple one-dimensional convolutional and pooling layers. It receives the concatenation of the complete charging voltage curve, conditional vector, and missing mask as input, and outputs a... The scalar values ​​between them are used to determine the global authenticity of the entire curve.

6. The lithium battery charging voltage curve reconstruction method based on dual discriminator CGAN according to claim 4, characterized in that, In S3, the local pattern discriminator D2 is used to receive the complete charging voltage curve as input and output a position-by-position true / false probability sequence with the same length as the complete charging voltage curve. Specifically, the local pattern discriminator D2 is a one-dimensional fully convolutional network, with each one-dimensional convolutional layer followed by a spectral normalization and LeakyReLU activation function layer.

7. The lithium battery charging voltage curve reconstruction method based on dual discriminator CGAN according to claim 3, characterized in that, In S4, the course learning strategy is as follows: during training rounds 1-100, only mild and typical bilateral missing scenarios are sampled; during training rounds 101-300, mild, typical, and severe missing scenarios are sampled using a probability weighted mixture of 0.4:0.4:0.2; during training rounds 301-800, mild, typical, and severe missing scenarios are sampled uniformly; at the same time, training samples are sampled evenly according to the SOH interval.

8. The lithium battery charging voltage curve reconstruction method based on dual discriminator CGAN according to claim 1, characterized in that, In S4, the formula for the composite loss function is as follows: In the formula: To combat losses in D1, To combat losses for D2, To reconstruct the loss, For smoothness loss; , These are learnable uncertainty parameters used to adaptively balance the adversarial losses D1 and D2. , For the corresponding precision weights, , To prevent , An infinitely increasing regular term; , Fixed weighting coefficients for reconstruction loss and smoothness loss; Reconstruction loss Using asymmetric head-and-tail weighted L1 loss, the expression is: Where: Reconstruction loss weight The weight ratio for missing head and missing tail positions is 1:2, and 0 is used for observed positions. For sequence position index, For the reconstructed voltage sequence after mask fusion at position Voltage value at that location, For the true voltage sequence in Voltage value at; Smoothing loss A scale-adaptive weighting method is used to impose constraints on the second-order difference of the generated segment, and The value gradually decreased from 1.0 to 0.1 as training progressed.

9. The lithium battery charging voltage curve reconstruction method based on dual discriminator CGAN according to claim 1, characterized in that, In S4, when training the dual-discriminator conditional generative adversarial network model, a label smoothing technique is used. Specifically, the labels of the real samples are changed from a fixed 1.0 to... Uniform sampling within intervals, the label of the generated samples is changed from a fixed 0.0 to... Uniform sampling within a given interval; The generator G, global voltage discriminator D1, and local pattern discriminator D2 are updated alternately. In each training step, the generator G is updated twice, while the global voltage discriminator D1 and the local pattern discriminator D2 are updated once each.

Images