撰写专利摘要

Abstract

Disclosed are systems, methods, and non-transitory computer-readable media for privacy-preserving visual search that partition computation between an edge device and a cloud service while enforcing formal privacy guarantees and strict resource constraints. An on-device feature extractor comprises a lightweight Transformer configured with homomorphic-encryption-friendly operators (restricted to additions and multiplications, including polynomial or piecewise-linear approximations of nonlinearities) and a chip-aware quantization pipeline that selects mixed precision per layer based on a hardware energy–latency model to maintain less than 200 mW power and less than 50 ms end-to-end query latency on supported edge hardware. The feature extractor outputs anonymized embeddings produced by a differential privacy (DP) mechanism with a per-identity, per-session budget controller that accounts for composition, enforces norm clipping, calibrates noise to a target (ε, δ), and prevents budget overrun. The embeddings are encrypted on-device and transmitted to a cloud service that maintains a secure approximate nearest neighbor (ANN) index over encrypted vectors. The server computes an encrypted similarity metric—such as a polynomially realizable inner product or cosine similarity with pre-normalization—directly over ciphertexts to rank candidates without decrypting individual embeddings. The system supports zero-shot extensibility by admitting new concepts at runtime via encrypted prototypes or prompt-conditioned descriptors, without retraining the edge model. Compliance-ready application programming interfaces provide consent hooks, opt-in/opt-out controls, DP budget audit and export, key rotation, revocation, and data-subject request handling, without exposing raw images or plaintext embeddings. The integrated architecture enables privacy-preserving, open-vocabulary visual retrieval with enforceable DP budget control, at-edge anonymization, encrypted similarity search in the cloud, and hardware-aware quantization, collectively improving security, legal compliance readiness, and performance within strict power and latency constraints.

摘要：本发明涉及电池健康状态预测与联邦学习技术领域，公开一种跨站点电池健康预测的联邦学习框架。所述框架在不共享原始数据的前提下，面向多厂商、异构BMS环境，通过异构BMS适配与时间序列退化模型实现协作训练并输出可解释健康指标。系统包括：其一，跨站点异构BMS适配层，对充放电、电压、电流、温度、内阻等核心变量实施特征对齐与尺度统一，建立面向告警码与工况标签的统一术语规则，以支持批量评审与检索；其二，基于时间序列退化机理的建模子系统，采用面向退化轨迹的序列模型生成SOH与剩余寿命（RUL）预测，并输出与退化斜率、工况敏感性及不确定度相关的可解释指标；其三，联邦协作训练与自适应加权聚合模块，各站点仅上传本地模型参数/梯度及必要统计量，通过依据数据有效性、分布差异、历史稳定性与时间新鲜度等要素计算的自适应聚合权重实现全局模型更新；其四，局部漂移检测与校正机制，基于分布偏移度量对本地模型进行漂移校正并减少跨域偏差。为实现早期风险管理，本发明通过与不确定度及退化速率关联的动态阈值设计提供早期故障预警，并支持跨站点阈值校准。与现有技术相比，本发明提升了跨站点泛化能力、早期预警准确性与隐私合规性。可保护点包括：用于异构BMS的特征对齐方法、用于联邦退化建模的聚合权重计算方法、基于不确定度耦合退化速率的早期故障预警阈值设定方法，以及用于数据与告警归一化的统一术语规则。