Battery alarm scene data augmentation and algorithm robustness verification method

Abstract

The application relates to the technical field of battery early warning and discloses a battery alarm scene data enhancement and algorithm robustness verification method, which acquires a real alarm sample set; a pre-trained cascade generation model is used to generate a synthetic alarm sample; the synthetic alarm sample and the real alarm sample are mixed in proportion to construct an enhanced training set; a battery alarm related target recognition algorithm is trained using the enhanced training set, and the robustness improvement degree of the battery alarm related target recognition algorithm is quantitatively verified under a plurality of preset disturbance conditions; wherein the cascade generation model comprises the following steps: a denoising diffusion probability model is used to generate a basic sample, a generative adversarial network is used to perform detail enhancement and distribution correction on the generated result, a preset strategy is used to control a random latent variable to complete sample screening, and a high-quality and diversified alarm data set is constructed. Through the method, the robustness and stability of the alarm recognition algorithm can be significantly improved, and reliable prediction and accurate early warning of a battery system under actual complex working conditions can be realized.

Description

Technical Field

[0001] This invention relates to the field of battery early warning technology, specifically to a method for data augmentation and algorithm robustness verification in battery alarm scenarios. Background Technology

[0002] In the actual operation of battery systems, alarm data usually exhibits a significant uneven distribution. There is a relatively sufficient amount of data under normal operating conditions, while the frequency of actual alarm events is low, resulting in an extremely limited number of fault or abnormal alarm samples.

[0003] Existing data augmentation methods rely on traditional interpolation, noise perturbation, or simple signal transformation. These methods lack the ability to deeply model the temporal structure and cross-dimensional correlations of alarm data, making it difficult for generated samples to effectively cover the diversity of real alarm patterns. In addition, while augmentation methods based on classic generative adversarial networks can improve the diversity of sample distribution to some extent, the training process often suffers from pattern collapse and unstable oscillations, especially under small sample conditions, which can easily lead to generated results deviating from the real alarm distribution.

[0004] In recent years, diffusion models have demonstrated significant stability in generation tasks. However, their applications in engineering are mostly based on multimodal non-Gaussian modeling, which is complex and involves a large number of parameters, making it difficult to deploy in resource-constrained environments such as energy storage and battery management systems. Furthermore, the long training paths and high sampling times of existing diffusion models in battery alarm scenarios also affect their practical efficiency.

[0005] Therefore, simplifying the model structure and improving the generation quality and computational efficiency while maintaining model stability remain key issues that existing technologies cannot effectively address. Summary of the Invention

[0006] This invention aims to provide a method for data augmentation and algorithm robustness verification in battery alarm scenarios, in order to solve the technical problems of scarce battery alarm samples, uneven data distribution, and insufficient model generalization ability.

[0007] The basic solution provided by this invention is: a method for data augmentation and algorithm robustness verification in battery alarm scenarios, comprising: Obtain a real alarm sample set; generate synthetic alarm samples using a pre-trained cascaded generative model; mix the synthetic alarm samples with the real alarm samples in proportion to construct an enhanced training set; use the enhanced training set to train a battery alarm-related target recognition algorithm, and quantitatively verify the improvement in robustness of the battery alarm-related target recognition algorithm under various preset perturbation conditions. Cascaded generative models include: A denoising diffusion probability model trained on a real alarm sample set is adopted. In the inference stage, starting from the standard noise prior, the model is gradually restored through a reverse diffusion sampling process to generate a primary synthetic alarm sample that conforms to the distribution characteristics of the real alarm sample set. The trained generative adversarial network is used. Its generator takes the primary synthetic alarm sample output by the denoised diffusion probability model and random latent variables as input, performs detail enhancement on the primary synthetic alarm sample, and outputs a refined sample that is more consistent with the distribution characteristics of the real alarm sample set after being verified by the discriminator. Among them, a preset strategy is used to control random latent variables, and the refined samples are screened by a discriminator to obtain synthetic alarm samples.

[0008] The working principle and advantages of this invention are as follows: This invention constructs a denoising diffusion probabilistic model (DDPM) to generate basic samples, and utilizes generative adversarial networks (GANs) to enhance details and correct distribution in the generated results. A preset strategy is employed to control random latent variables for sample selection, ultimately constructing a high-quality, diverse alarm dataset. This method significantly improves the robustness and stability of alarm recognition algorithms, enabling reliable prediction and accurate early warning of battery systems under complex real-world operating conditions.

[0009] First, it offers high training stability, overcoming the inherent limitations of single generative models. In existing technologies, pure GAN models are prone to mode collapse and gradient instability, making training difficult in small-sample alarm scenarios; while pure diffusion models are stable during training, their generation efficiency is low and details are blurred. This invention employs a cascaded architecture combining DDPM and GAN, where DDPM acts as a pre-generator providing a stable global structural foundation, while GAN serves only as a post-enhancer responsible for local detail compensation. This decoupled design makes the overall training process more controllable, avoiding the training collapse risk of GANs under weak supervision and compensating for the insufficient detail generation capability of diffusion models, making it particularly suitable for battery alarm data enhancement scenarios where labeled samples are scarce.

[0010] Secondly, the generated data is of high quality, simultaneously capturing both global structure and local details. Pure diffusion models generate samples that match the real distribution in global statistical characteristics, but suffer from blurred high-frequency details and insufficient edge and texture resolution. Pure GAN-generated samples offer sharp details but are prone to artifacts and exhibit poor distribution diversity. This invention employs a two-stage cascaded generation process: the first stage utilizes DDPM's progressive denoising mechanism to ensure the generated samples are highly consistent with real alarm data in terms of global structure, such as temporal trends and amplitude ranges; the second stage leverages GAN's adversarial learning to enhance local details, further improving signal texture and feature resolution. The resulting refined samples outperform single models or simple concatenation schemes in both overall distribution fit and fidelity in capturing subtle high-frequency signal features, more accurately reflecting the real alarm distribution.

[0011] Third, the generation efficiency is controllable, meeting the requirements of engineering deployment. Pure diffusion models require tens to thousands of iterative sampling steps, making them difficult to deploy in battery management systems or edge devices with high real-time requirements. This invention adopts a cascaded strategy combining coarse diffusion generation with fast GAN enhancement, which can significantly shorten the generation time by reducing the number of DDPM backsampling steps and optimizing noise scheduling, transferring the complex and cumbersome iterative denoising burden to a single-step forward propagation GAN generator. While ensuring generation quality, it significantly reduces inference latency, is easy to integrate into embedded platforms, supports online or offline alarm sample synthesis and continuous verification, and meets the real-time and feasibility requirements of engineering applications.

[0012] Fourth, it significantly improves the robustness and generalization ability of downstream alarm algorithms. Traditional data augmentation methods (such as noise addition, pruning, and time warping) struggle to generate out-of-distribution samples, limiting their improvement on algorithm generalization. This invention, through a conditional sampling strategy, can proactively generate high-quality synthetic samples covering multiple operating conditions, disturbance levels, and alarm types, effectively expanding the distribution boundary of training data. Alarm recognition algorithms trained using this augmented dataset maintain stable recognition performance under complex conditions such as sensor noise interference, operating condition drift, and individual device differences, significantly reducing false alarm and false negative rates, enhancing adaptability to rare alarm types and unseen operating conditions, and ultimately improving the operational safety of battery systems under real-world complex conditions. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating a method for data augmentation and algorithm robustness verification in a battery alarm scenario, as provided in an embodiment of the present invention. Detailed Implementation

[0014] The following detailed explanation illustrates the specific implementation methods: The basic implementation examples are as follows: Figure 1 As shown: A method for data augmentation and algorithm robustness verification in battery alarm scenarios, comprising: Obtain a real alarm sample set; generate synthetic alarm samples using a pre-trained cascaded generative model; mix the synthetic alarm samples with the real alarm samples in proportion to construct an enhanced training set; use the enhanced training set to train a battery alarm-related target recognition algorithm, and quantitatively verify the improvement in robustness of the battery alarm-related target recognition algorithm under various preset perturbation conditions. Cascaded generative models include: A denoising diffusion probability model trained on a real alarm sample set is adopted. In the inference stage, starting from the standard noise prior, the model is gradually restored through a reverse diffusion sampling process to generate a primary synthetic alarm sample that conforms to the distribution characteristics of the real alarm sample set. The trained generative adversarial network is used. Its generator takes the primary synthetic alarm sample output by the denoised diffusion probability model and random latent variables as input, performs detail enhancement on the primary synthetic alarm sample, and outputs a refined sample that is more consistent with the distribution characteristics of the real alarm sample set after being verified by the discriminator. In this process, a preset strategy is used to control random latent variables, and a discriminator is used to evaluate the quality of refined samples to obtain synthetic alarm samples.

[0015] The specific process includes: S1. The real alarm samples are multi-dimensional time series data recorded when the battery system experiences an alarm event. Each sample contains sensor features such as voltage sequence, current sequence, and temperature sequence within a fixed-length time window, as well as the corresponding alarm category label and operating parameters.

[0016] The construction and training of the denoising diffusion probability model includes: a forward noise scheduling module, used to determine the maximum number of time steps for training and the Gaussian noise intensity scheduling strategy for each step; a time encoding module, used to encode the time steps as conditional inputs to the noise prediction network; and a noise prediction network, using a U-Net structure, used to receive noisy samples and time encodings, and predict the added noise components. The training objective is to minimize the reconstruction error between the noise predicted by the noise prediction network and the actual noise. Specifically, this includes: S11. Construct and initialize the forward diffusion noise scheduling module and time encoding module of the diffusion model (DDPM, Denoising Diffusion Probabilistic Model). Determine the maximum number of time steps for training and the Gaussian noise intensity scheduling strategy for each step, design the encoding method for time steps (e.g., sinusoidal position encoding or embedding vector), and initialize the U-Net network structure parameters, optimizer, and model saving strategy for noise prediction.

[0017] First, construct the forward diffusion noise scheduling function required for DDPM. Let the actual battery alarm sample be... The total number of diffusion steps is Forward diffusion uses a fixed set of Gaussian noise variance sequences:

[0018] in, This represents a real battery alarm sample, specifically a multi-dimensional time-series feature vector collected by the battery system when an alarm event occurs, containing time-series data of sensor signals such as voltage, current, and temperature within a sliding window; Indicates the time step index during the diffusion process ( =1,2,…,T), which are the stages in a battery alarm scenario where Gaussian noise is gradually applied to the original alarm signal. The larger the value, the deeper the degree of noise pollution; it is generated from the reverse direction. =T is gradually denoised and restored to t=0 to reconstruct the alarm signal.

[0019] Its cumulative retention factor is:

[0020] The formula for the forward diffusion process is:

[0021] The time step encoding uses sinusoidal position encoding:

[0022] in, For the current diffusion time step, For encoding dimension index, The total dimension of the encoded vector. and The first The second step of the time-coded vector and the 2nd With +1 component (sine for even positions, cosine for odd positions), this encoding will discrete time steps Mapped to continuous A dimensional vector is used as a conditional input to U-Net to distinguish processing stages with different noise intensities.

[0023] Initialize U-Net parameters, optimizer, and EMA (Exponential Moving Average) smoothing factor.

[0024] S12. Prepare the data sampling and time-series noise pair generation process for diffusion training. Samples are drawn from the real alarm sample set according to the training batch. Based on the noise scheduling of S1, random sampling time step t:

[0025] Resampling standard Gaussian noise:

[0026] And based on the forward noise strategy, the real samples are transformed into corresponding noisy samples. That is, to directly construct training samples according to the forward diffusion formula:

[0027] At the same time, prepare the corresponding time-coded information as an auxiliary condition for network input.

[0028] S13, U-Net inverse noise prediction training.

[0029] Noisy samples The time-encoded input to the U-Net is used to predict the corresponding noise component or the cleaned-up sample representation. This step trains the U-Net to estimate the noise. The objective is to minimize: Noisy samples With time-encoded input to U-Net,

[0030] in, To improve U-Net's noise prediction, the Adam optimizer is used, and an EMA update is introduced:

[0031] in, Here, θ is the exponential moving average parameter (used to generate more stable samples during inference), θ is the U-Net parameter for the current training step, and γ∈(0,1) is the decay coefficient (usually taken as 0.999~0.9999), which controls the proportion of historical parameters retained. In θ(xt,t), θ is the set of learnable parameters of the U-Net. for Sample of a battery alarm that is constantly producing noise. The current time step is used. The reconstruction error between the real noise and the predicted noise is used as the training objective (using mean square or similar stability loss). Forward propagation, backpropagation and parameter updates are performed alternately according to the conventional training process, and training stability measures (such as learning rate scheduling, gradient clipping, etc.) are applied to ensure the generalization and stability of U-Net at different time steps.

[0032] S14. Validation and optimization of diffusion model training.

[0033] Periodically backsample from the standard noise distribution using the U-Net obtained during training to generate candidate samples, and evaluate the generation quality using a sample similarity index. Adjust the time step, noise scheduling, U-Net width / depth, or training hyperparameters based on the validation results until the backsampling achieves the expected trade-off between quality and efficiency, and then fix the model version.

[0034] The backsampling distribution is:

[0035] The mean term is:

[0036] The clarity, structural fidelity, and distribution consistency of the generated results are used to adjust parameters such as time steps, noise sequences, or network depth until backsampling achieves the desired trade-off between quality and efficiency and fixes the model version.

[0037] S2, the construction and adversarial training of the Generative Adversarial Network (GAN) include: the generator adopts an encoder-decoder structure; the receptive field and multi-layer discriminant output strategy of the discriminator are designed according to the time-series characteristics of the alarm signal; adversarial training adopts an alternating update strategy: fixing the generator to update the discriminator to minimize the classification loss; fixing the discriminator to update the generator to maximize the probability that the discriminator classifies its output as true; training stability measures include label smoothing, gradient pruning, and discriminator step adjustment. Specifically, this includes: S21. Construct GAN.

[0038] Design a generator based on an encoder-decoder structure. Used to receive the initial generated samples of the diffusion model. and random latent variables The diffusion samples are enhanced with details using convolutional or deconvolutional networks, and the enhanced alarm samples are output; a discriminator is designed. This is used to distinguish between generator-enhanced samples and real alarm samples, and to select the discriminator's receptive field and multi-layer discriminative output strategy based on the task. In specific applications such as battery alarm scenarios: The primary battery alarm signal samples generated for DDPM (containing multi-dimensional time-series features such as voltage, current, and temperature) are used, where z is a random latent variable to introduce sample diversity under the same operating conditions (corresponding to individual differences or measurement noise). The generator G is specifically designed for battery time-series signals, and its convolutional encoder is responsible for extracting the local time-series features of the alarm signal (such as voltage drops, abnormal temperature peaks, etc.). The decoder recovers the time-domain resolution and amplitude range of the signal through transposed convolution. The discriminator... The receptive field is designed based on the typical duration of battery alarm signals (e.g., several seconds to several minutes). A multi-scale discrimination strategy is adopted to simultaneously evaluate the authenticity of short-term local features and the consistency of long-term overall distribution of the signal. This enables the adversarial training of GAN to be effectively optimized for the special temporal structure of battery alarm signals, thereby improving the fidelity of generated samples in terms of alarm features.

[0039] S22, GAN adversarial training.

[0040] Following the classic Goodfellow adversarial training sequence: in each training iteration, the generator is fixed first, and the discriminator is trained using a batch of real samples and a batch of pseudo-samples initially generated by the diffusion model and processed by the generator. The discriminator's optimization objective is to minimize the classification loss of D.

[0041] Then, with the discriminator fixed, the generator is trained to increase the probability that the discriminator classifies its output as "real" (using non-saturating adversarial loss or other classic loss forms). The generator optimization objective is:

[0042] To improve training stability, strategies such as small-amplitude label smoothing, frequent sampling of real samples, and adjusting the discriminator step count when necessary can be employed during discriminator training. WGAN-GP loss can be used as an example.

[0043] S3 also includes joint fine-tuning, employing at least one of the following strategies after the generative adversarial network (GAN) has completed adversarial training to achieve a balance between generating quality and maintaining the original alarm signal structure: Strategy 1, using a fixed output from the denoising diffusion probability model as a condition, jointly train the generator of the GAN and the U-Net back-end parameters of the denoising diffusion probability model; Strategy 2, add time-step conditions to the generator input of the GAN. Specifically, this includes: After completing the basic adversarial training, the discriminator is used to evaluate the discriminative distribution of the generated samples, the diversity of the generator output, and the coverage of potential patterns. If further consistency needs to be improved, a limited number of joint fine-tunings can be performed at this stage. That is, with a fixed diffusion model output as a condition, the back-end parameters of the generator and U-Net are jointly trained (or time step conditions are added at the generator input) to achieve a better match between the two in terms of generation quality and maintaining the original signal structure.

[0044] The joint fine-tuning adopts a cascaded structure:

[0045] Simultaneously, some parameters in the diffusion model related to the alignment with input features can be fine-tuned to ensure that the high-frequency details output by the generator are consistent with the low-frequency structure generated by the diffusion model.

[0046] S4, Obtain a real alarm sample set; generate synthetic alarm samples using a pre-trained cascaded generative model; mix the synthetic alarm samples with the real alarm samples in a certain proportion to construct an enhanced training set; use the enhanced training set to train the battery alarm-related target recognition algorithm, and quantitatively verify the improved robustness of the battery alarm-related target recognition algorithm under various preset perturbation conditions. Specifically, this includes: S41. Diffusion backsampling generates primary synthetic alarm samples. Starting from the standard noise prior, the sample space is gradually restored according to the set number of backsampling steps and the U-Net noise predictor, generating several candidate primary synthetic alarm samples; during sampling, the random seed, time step interpolation, or temperature parameters can be controlled to adjust the sample diversity.

[0047] Diffusion backsampling generates primary synthetic alarm samples. Starting with a standard noise prior:

[0048] By progressively reconstructing the sample space using the set number of reverse sampling steps and the U-Net noise predictor, several candidate primary synthetic alarm samples are generated:

[0049] Until a sample is obtained .

[0050] S42. GAN generator is enhanced with details. A random latent variable z (and optional conditional label) is added to each primary sample. The trained generator is used to fine-tune the sample at least once, and the output is a refined sample that is closer to the true distribution after being verified by the discriminator. The generator output can be filtered or resampled using the discriminator score threshold to remove low-quality samples.

[0051] The GAN generator is enhanced with details. A random latent variable z (and optional conditional labels) is appended to each primary sample. The generator trained in step S6 is then used to perform the following:

[0052] It achieves texture enhancement, detail completion, and distribution alignment, and can be scored using a discriminator:

[0053] As a selection criterion, the generator output can be filtered using a discriminator score threshold or resampled to remove low-quality samples.

[0054] S43. Diversity Control, Conditional Sampling, and Sample Selection. To meet the enhancement needs of different alarm scenarios, a conditional sampling strategy is provided, and a multiple sampling-selection mechanism is used (i.e., multiple different random latent variables z are added in parallel to the same primary synthetic alarm sample, multiple candidate refined samples are generated by the GAN generator, and then the candidate samples are scored and ranked according to the discriminator confidence score, the cosine similarity between the feature space and the real sample, or statistical distribution indicators (such as mean, variance consistency), and the samples that meet the preset threshold or are in the preset top K positions are retained as the final enhanced data) to ensure that the results are both representative and diverse.

[0055] Pre-defined strategies for controlling random latent variables include: employing conditional sampling strategies based on category labels, operating parameters, or severity vectors to control latent variables during the generation process; and adjusting the diversity of generated samples by modifying the random seed, time step interpolation, or temperature parameters in the backdiffusion sampling process. Adjusting latent variables:

[0056] or category condition vector It can achieve:

[0057] The methods for selecting and retaining refined samples during the screening process include: generating multiple candidate refined samples in parallel for the same conditions, and filtering them based on discriminator confidence, feature space similarity, or statistical indicators, retaining samples that meet preset thresholds or preset rankings. Diverse data is generated for different alarm categories or different operating conditions; multiple candidate samples can be sorted and filtered.

[0058] S44. Construct an enhanced dataset and verify the robustness of the algorithm.

[0059] The synthetic samples generated and screened in steps S41–S43 are combined with real samples to create an enhanced training set according to the designed ratio:

[0060] The augmentation set was used to train and evaluate target alarm discrimination or fault identification algorithms. The model performance was tested under various preset perturbation conditions (including sensor noise perturbation, operating condition drift perturbation, and equipment difference perturbation). The quantitatively verified performance indicators included accuracy, recall, false alarm rate, discriminator confidence score (i.e., the true probability score of the GAN discriminator D(x) output to the generated samples, used to evaluate the quality of the synthesized samples) and F1 score (i.e., the harmonic mean of accuracy and recall, F1 = 2 × accuracy × recall ÷ (accuracy + recall), used to comprehensively measure the classification performance of the alarm identification algorithm) and the generalization ability on unseen operating conditions to verify the degree of improvement in the robustness of the algorithm after introducing the diffusion model and GAN augmentation.

[0061] It also includes S5, model version management, and the implementation of generation strategies.

[0062] Save the trained diffusion model (including its parameters). GAN generator (generator parameters) ) and discriminator (discriminator parameters) Versioned storage is implemented, and a reproducible generation process including time step configuration, sampling strategy and screening threshold is established so that it can be called in the operation of the battery management system. In actual deployment, an on-demand generation interface is provided to support online and / or offline alarm sample synthesis and continuous verification.

[0063] Furthermore, offline batch generation of primary synthetic alarm samples is adopted; online reading of primary synthetic alarm samples and addition of random latent variables are adopted to generate diverse and refined samples in real time for hybrid online and offline deployment.

[0064] Offline batch generation refers to backsampling a large number of real alarm samples using DDPM before model deployment, and persistently storing the generated primary synthetic alarm samples in a local database or file system for direct use in subsequent training phases, thereby avoiding real-time inference overhead during training. Online reading refers to loading primary synthetic alarm samples from the above storage medium in real time during algorithm operation or continuous verification phases, and dynamically attaching random latent variables z by the GAN generator to generate diverse and refined samples, supporting continuous expansion and enhancement of data without restarting the model, and meeting the real-time response requirements of edge devices or battery management systems for alarm sample synthesis.

[0065] This embodiment provides a method for data augmentation and algorithm robustness verification in battery alarm scenarios. It constructs a denoised diffusion probability model (DDPM) to generate basic samples, and utilizes a generative adversarial network (GAN) to enhance details and correct distribution in the generated results. A preset strategy is used to control random latent variables to complete sample selection, ultimately constructing a high-quality, diverse alarm dataset. This method significantly improves the robustness and stability of alarm recognition algorithms, enabling reliable prediction and accurate early warning of battery systems under complex real-world operating conditions.

[0066] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics of the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. A method for data augmentation and algorithm robustness verification in battery alarm scenarios, characterized in that, include: Obtain a real alarm sample set; generate synthetic alarm samples using a pre-trained cascaded generative model; mix the synthetic alarm samples with the real alarm samples in a proportional manner to construct an enhanced training set; The battery alarm-related target recognition algorithm was trained using an enhanced training set, and the improvement in robustness of the algorithm was verified by quantification under various preset perturbation conditions. Cascaded generative models include: A denoising diffusion probability model trained on a real alarm sample set is adopted. In the inference stage, starting from the standard noise prior, the model is gradually restored through a reverse diffusion sampling process to generate a primary synthetic alarm sample that conforms to the distribution characteristics of the real alarm sample set. The trained generative adversarial network is used. Its generator takes the primary synthetic alarm sample output by the denoised diffusion probability model and random latent variables as input, performs detail enhancement on the primary synthetic alarm sample, and outputs a refined sample that is more consistent with the distribution characteristics of the real alarm sample set after being verified by the discriminator. Among them, a preset strategy is used to control random latent variables, and the refined samples are screened by a discriminator to obtain synthetic alarm samples.

2. The method for data augmentation and algorithm robustness verification in a battery alarm scenario according to claim 1, characterized in that, The construction and training of the denoising diffusion probability model includes: The forward noise scheduling module is used to determine the maximum number of training time steps and the Gaussian noise intensity scheduling strategy for each step; The time encoding module is used to encode time steps, which serve as conditional inputs to the noise prediction network. The noise prediction network, using a U-Net structure, is used to receive noisy samples and time-coded data, and to predict the added noise components. Training objective: Minimize the reconstruction error between the noise predicted by the noise prediction network and the actual noise.

3. The method for data augmentation and algorithm robustness verification in a battery alarm scenario according to claim 1, characterized in that, The construction and adversarial training of generative adversarial networks include: The generator uses an encoder-decoder structure; The discriminator's receptive field and multi-layer discriminative output strategy are designed based on the time-series characteristics of the alarm signal; Adversarial training employs an alternating update strategy: the generator is fixed while the discriminator is updated to minimize the classification loss; the discriminator is fixed while the generator is updated to maximize the probability that the discriminator classifies its output as true. Training stability techniques include label smoothing, gradient clipping, and discriminator step adjustment.

4. The method for data augmentation and algorithm robustness verification in a battery alarm scenario according to claim 1, characterized in that, This also includes joint fine-tuning, employing at least one of the following strategies after the generative adversarial network has completed adversarial training to achieve a balance between generating quality and maintaining the original alarm signal structure: Strategy 1: With a fixed output from the denoising diffusion probability model, jointly train the generator of the generative adversarial network and the back-end parameters of the U-Net of the denoising diffusion probability model. Strategy 2 involves adding time step conditions to the generator input of the generative adversarial network.

5. The method for data augmentation and algorithm robustness verification in a battery alarm scenario according to claim 1, characterized in that, The methods for selecting and retaining refined samples during the screening process include: generating multiple candidate refined samples in parallel for the same conditions, and filtering them by discriminator confidence, feature space similarity, or statistical indicators, retaining samples that meet the preset threshold or preset ranking.

6. The method for data augmentation and algorithm robustness verification in a battery alarm scenario according to claim 1, characterized in that, Preset strategies for controlling random latent variables include: using conditional sampling strategies based on category labels, operating parameters, or severity vectors to control latent variables in the generation process; and adjusting the diversity of generated samples by adjusting the random seed, time step interpolation, or temperature parameters in the backdiffusion sampling process.

7. The method for data augmentation and algorithm robustness verification in a battery alarm scenario according to claim 1, characterized in that, Multiple preset disturbance conditions include: sensor noise disturbance, operating condition drift disturbance, and equipment difference disturbance.

8. The method for data augmentation and algorithm robustness verification in a battery alarm scenario according to claim 1, characterized in that, The performance metrics for quantitative validation include accuracy, recall, false alarm rate, and discriminator confidence score. The improvement in robustness is quantified by comparing the performance metrics of the same fault identification algorithm trained under perturbation conditions using a training set consisting only of real alarm samples and an enhanced training set.

9. The method for data augmentation and algorithm robustness verification in a battery alarm scenario according to claim 1, characterized in that, It also includes model version management; saving the trained denoised diffusion probability model, generator and discriminator of the generative adversarial network in a versioned manner; and establishing a reproducible generation process that includes time step configuration, sampling strategy and screening threshold; In actual deployment, it provides an on-demand generation interface to support the synthesis and continuous verification of alarm samples online and / or offline.

10. The method for data augmentation and algorithm robustness verification in a battery alarm scenario according to claim 1, characterized in that, The system employs offline batch generation of primary synthetic alarm samples; it also uses online reading of primary synthetic alarm samples and the addition of random latent variables to generate diverse and refined samples in real time for hybrid online and offline deployment.

Images