撰写专利说明书

Title Automotive-Grade Millimeter-Wave Radar Fusion Localization System with Temporal Filtering and Vehicle–Cloud Cooperative Processing

Technical Field The disclosure relates to vehicular localization systems. More particularly, it pertains to automotive-grade localization utilizing millimeter-wave (mmWave) radar in combination with other sensors and map priors, incorporating temporal filtering algorithms and coordinated processing between on-vehicle and cloud resources.

Background Conventional vehicle localization typically relies on Global Navigation Satellite Systems (GNSS) and inertial measurement units (IMUs), sometimes combined with camera- or LiDAR-based map matching. These approaches may degrade in urban canyons, tunnels, sensor-occluded environments, adverse weather, or feature-sparse regions. Millimeter-wave radar sensors operating, for example, in the 76–81 GHz band can provide robust perception under rain, fog, or dust. However, radar-based localization faces challenges including multipath, specular returns, lower angular resolution relative to LiDAR, and clutter.

Existing radar localization approaches often employ single-frame detections or pairwise scan matching susceptible to transient outliers and non-Gaussian noise. Processing and storage constraints on embedded automotive platforms further limit the complexity of on-vehicle algorithms. Moreover, maintaining up-to-date priors (e.g., reflectivity maps or landmark catalogs) benefits from fleet-scale aggregation and curation that are difficult to perform exclusively on-vehicle.

Accordingly, there is a need for an automotive-grade system that (i) fuses mmWave radar with other sensors and map priors, (ii) employs temporal filtering tailored to radar phenomenology and asynchronous sensor timing, and (iii) distributes computation and data management between in-vehicle and cloud components to achieve robust, real-time localization under automotive constraints.

Summary In one aspect, disclosed is a system comprising: (a) one or more automotive-grade mmWave radar sensors; (b) an IMU and optional GNSS receiver; (c) an electronic control unit (ECU) configured for real-time, on-vehicle multi-sensor fusion utilizing a temporal filtering algorithm; and (d) a cloud service configured to generate, validate, and distribute radar-centric map priors and to perform fleet-scale optimization. The system provides robust vehicle pose estimates by fusing radar detections (e.g., range–Doppler–angle points or radar power images) with inertial and optional GNSS data, subject to temporal alignment and outlier-robust filtering. The cloud service and the ECU cooperate to adapt priors and parameters, to trigger selective uplink of low-confidence segments, and to return updated models and map tiles.

In some embodiments, the temporal filtering algorithm includes a delayed-state estimator or a smoothing-based factor graph that accommodates asynchronous timestamped measurements, includes Doppler-informed motion constraints between successive radar frames, and applies outlier-robust association gating. In further embodiments, the vehicle–cloud interface uses confidence measures and change-detection heuristics to govern bandwidth usage. The system is implementable with automotive-grade hardware, communication interfaces, and safety lifecycle processes.

Advantages can include improved localization availability and accuracy during GNSS degradation, resilience to adverse weather, reduced drift during IMU-only intervals through Doppler-constrained temporal smoothing, and reduced on-vehicle computational burden through selective cloud offloading and curated priors.

Definitions

• “Automotive-grade” refers to components, design, and processes suitable for integration into road vehicles, including operating temperature ranges typically from about −40°C to about +105°C at the module level, electromagnetic compatibility per applicable standards, and functional safety lifecycle processes consistent with ISO 26262 for the intended Automotive Safety Integrity Level (ASIL).
• “mmWave radar” refers to frequency-modulated continuous-wave (FMCW) radar devices operating, for example, within 76–81 GHz bands, subject to applicable regional spectrum regulations.
• “Temporal filtering” refers to algorithms that estimate vehicle state over time using probabilistic models with explicit handling of time, including filters (e.g., Kalman variants, particle filters) and smoothers (e.g., factor-graph optimization).
• “Cloud” refers to remote computing resources accessible via a network and includes edge servers.

Brief Description of the Drawings Fig. 1 shows an example system architecture including sensors, on-vehicle fusion module, and cloud cooperation. Fig. 2 illustrates a data flow for radar signal processing to produce detections and radar power images. Fig. 3 depicts a temporal filtering pipeline, including time synchronization, measurement modeling, gating, and smoothing. Fig. 4 shows a factor graph representing vehicle states with IMU preintegration, radar Doppler constraints, GNSS factors, and map factors. Fig. 5 illustrates cloud-based prior generation and update, including map tile management and confidence feedback to vehicles. Fig. 6 outlines an example method for confidence-triggered selective uplink and cloud return of updated priors.

Detailed Description of Embodiments

I. System Overview A vehicle includes:

• One or more mmWave FMCW radar sensors configured for azimuth and, in some embodiments, elevation estimation using multiple-input multiple-output (MIMO) arrays. The radar operates, for example, within 76–81 GHz and produces per-frame detections comprising range, radial velocity (Doppler), and angle(s), with associated signal-to-noise or covariance estimates.
• An IMU providing tri-axial accelerometer and gyroscope data at a higher rate than radar frames. Optional sensors include a GNSS receiver, wheel-speed sensors, odometry from powertrain control modules, and cameras or LiDAR for complementary data.
• An ECU that executes fusion algorithms in real-time, using a temporal filtering pipeline. The ECU is communicatively coupled to vehicle networks (e.g., CAN FD and automotive Ethernet) and to external networks (e.g., cellular) for cloud connectivity.
• A cloud service that maintains radar-centric map priors (e.g., reflectivity or occupancy tiles, landmark catalogs, and pose graphs), performs fleet-scale optimization and change detection, and provides over-the-air (OTA) updates to vehicles.

II. Radar Processing and Measurement Models A. Signal Processing Each radar sensor performs chirp-level processing (e.g., range FFT, Doppler FFT) and angle estimation (e.g., beamforming or super-resolution techniques). Post-detection processing optionally includes:

• Constant false alarm rate (CFAR) detection with adaptive noise floor estimation.
• Clustering of detections into static and dynamic subsets using Doppler and temporal consistency.
• Formation of radar power images or occupancy grids by aggregating power returns into ego-stabilized cells.

B. Ego-Motion Compensation Between successive radar frames, ego-motion is estimated using IMU preintegration and, if available, wheel odometry, to transform detections into a common frame for temporal correlation.

C. Measurement Models The fusion pipeline uses measurement functions:

• Point-level radar factors: z_r,k = h_r(x_k, m) + v_k, where x_k is vehicle pose/velocity at time k, m are map parameters (if used), and v_k is noise with possibly non-Gaussian characteristics.
• Doppler motion factors between frames: v_r,k ≈ n_k·(R(θ_k)·v_k) + ε_k, where n_k denotes radial direction and R(θ_k) is rotation into radar frame.
• Map correlation factors: correlation between ego-stabilized radar power images and a radar reflectivity prior tile yields a scalar score modeled as a likelihood.

III. Temporal Filtering A. Time Synchronization Measurements are timestamped and aligned to a common clock. The filter accommodates asynchronous arrivals with a buffering horizon ΔT allowing out-of-sequence updates.

B. Estimator Structure In one embodiment, the ECU executes a fixed-lag smoother implemented as a factor graph with sliding window length L (e.g., 2–10 s):

• States include position, orientation, linear velocity, IMU bias terms.
• Factors include: (i) IMU preintegration factors linking consecutive states; (ii) radar Doppler constraints capturing short-horizon motion; (iii) radar–map correlation factors when a prior is available; and (iv) GNSS pseudorange/velocity factors when valid.
• Nonlinear least-squares optimization (e.g., Gauss–Newton) is performed with robust loss functions (e.g., Huber or Cauchy) to mitigate outliers.

In alternative embodiments, an invariant extended Kalman filter or a Rao–Blackwellized particle filter is employed, wherein pose is represented on SE(3) and Doppler observations constrain velocity hypotheses.

C. Outlier Handling and Gating

• Statistical gating is applied using predicted innovation covariance; general likelihood ratio tests filter spurious detections.
• Multi-hypothesis association is optionally maintained for ambiguous static-scene correspondences, pruned via hypothesis scoring over lag L.

D. Confidence Quantification The ECU computes covariance metrics (e.g., marginal pose covariance) and performance indicators (e.g., normalized innovation squared statistics, correlation scores). Confidence governs both application-level use and vehicle–cloud messaging.

IV. Vehicle–Cloud Cooperative Processing A. Priors and Tiles The cloud maintains geographically indexed tiles containing radar-centric priors, such as reflectivity grids and landmark probability distributions. Tiles are versioned and include quality metadata.

B. Fleet-Scale Optimization The cloud aggregates uploaded segments tagged by low confidence or change indicators. Batch optimization refines map priors and resolves inter-vehicle alignment using robust pose graph optimization with loop closures. Dynamic objects are suppressed via temporal consistency filters.

C. Selective Uplink and OTA Downlink

• Trigger policy: The ECU uploads compressed radar power images or lightweight descriptors, IMU summaries, and provisional poses when (i) confidence drops below a threshold, (ii) correlation against current tile is poor, or (iii) geofenced updates are requested.
• Compression: Descriptors may include low-rank features or learned embeddings to reduce bandwidth.
• The cloud returns updated priors or parameter sets. ECU verifies authenticity and integrity and applies updates via an OTA mechanism conforming to vehicle cybersecurity policies.

V. Automotive Considerations

• Environmental: Components are selected and designed for automotive temperature ranges and vibration environments, with electromagnetic compatibility per applicable standards.
• Functional safety: The system is developed under a functional safety process consistent with ISO 26262 appropriate to the function’s ASIL allocation. The safety concept may include degraded modes (e.g., IMU-only with conservative vehicle functions) upon confidence loss.
• Networking: On-vehicle communications utilize, for example, CAN FD for control and automotive Ethernet (optionally with time-sensitive networking) for high-throughput sensor and map data.
• Security: Uplink and downlink communications are authenticated and encrypted; data is access-controlled per security policies.

VI. Example Method (On-Vehicle)

1. Acquire IMU data at a first rate and radar frames at a lower rate; timestamp all data.
2. Preintegrate IMU between filter updates; predict vehicle state.
3. Process radar frames to obtain detections and, optionally, a radar power image; classify static vs dynamic returns.
4. Associate detections or power images to current map priors when available; compute correlation scores.
5. Update the temporal filter with radar Doppler constraints, map correlation factors, and GNSS factors if valid; perform fixed-lag smoothing with robust losses.
6. Compute pose estimate and covariance; output to vehicle subsystems and record confidence metrics.
7. If confidence below threshold or change detected, compress and upload a segment to the cloud; otherwise, continue local operation.
8. Receive updated map tiles or parameters from the cloud; validate and apply; purge superseded tiles.

VII. Example Cloud Workflow

1. Receive uploaded segments with provisional poses and confidence tags.
2. Perform quality control, reject corrupted or low-information data.
3. Optimize a regional pose graph aggregating multiple vehicles; update reflectivity/occupancy priors while suppressing dynamic artifacts.
4. Validate updated tiles; increment version and sign.
5. Notify vehicles in the affected region to fetch updated tiles; provide delta updates where feasible.

VIII. Implementation Variants

• Sensor configurations: Single front radar; multiple corner and long-range radars; combinations thereof.
• Map priors: Radar-only; radar plus lane-level semantics; probabilistic landmarks derived from persistent radar scatterers.
• Temporal filters: Fixed-lag smoothers; multi-rate Kalman filters; particle filters; hybrid schemes.
• Communication policies: Periodic background synchronization vs strictly event-triggered uploads; integration with V2X roadside units for local edge processing.

IX. Non-Limiting Examples of Parameters

• Radar band: approximately 76–81 GHz, subject to regional regulation.
• Sliding window: approximately 2–10 s depending on compute budget.
• Upload trigger: covariance trace exceeding a threshold; correlation score below a threshold; GNSS dilution of precision above a threshold; or map version mismatch.
• Robust loss: Huber with a tuning constant selected based on empirical radar noise characteristics.

X. Industrial Applicability The disclosed system is applicable to passenger vehicles, commercial trucks, and autonomous platforms requiring robust localization under diverse environmental conditions, with the ability to scale accuracy through fleet-cooperative map maintenance and to manage computational resources via vehicle–cloud partitioning.

Legal Notes and Claim Construction

• The terms “comprise,” “comprising,” “include,” and “including” are used in an open-ended, non-limiting sense.
• The order of steps in method claims (if any) is not limiting unless expressly stated.
• Features described in connection with one embodiment are combinable with other embodiments unless context dictates otherwise.
• Ranges recited herein are intended to include all subranges and individual values within the stated endpoints.

While specific embodiments have been described to enable practice by a person of ordinary skill in the art, variations and modifications within the scope of the appended claims (if presented) will be apparent. The disclosure is intended to satisfy enablement and written description requirements by providing sufficient structure, examples, and alternatives for the recited subject matter.

标题：可调谐微型光谱传感器及其信号处理与校准方法

技术领域 本发明涉及光谱测量技术领域，具体涉及一种采用可调谐窄带光学元件与单像素/小型探测器实现的微型光谱传感器，以及与之配套的信号处理与校准方法。该发明适用于便携式、多传感器集成及嵌入式光谱分析场景，包括但不限于环境监测、食品与农业检测、过程分析与质量控制等。

背景技术 现有微型光谱仪通常采用衍射光栅与线阵探测器的分光结构，器件体积与功耗难以进一步缩减，且在强烈机械振动、温度漂移及散射光干扰条件下，稳定性受限。另一方面，基于可调谐滤光器（例如微机电系统MEMS法布里-珀罗干涉腔、液晶可调谐滤光器LCTF或声光可调谐滤光器AOTF）的单像素扫描式光谱传感器，具有结构紧凑、成本低、可集成度高等优势，但在实际应用中面临以下问题：

1. 波长-驱动映射的非线性与温度依赖性造成谱位偏移与分辨率退化；
2. 探测器暗电流、增益漂移与非均匀性导致辐射定标不稳定；
3. 调谐过程中通带形状随驱动与环境变化发生变化，谱仪线形（Instrument Line Shape, ILS）不恒定；
4. 由于器件尺寸与光学布局限制，杂散光与多级反射引入的光谱串扰提升了重构谱的偏差；
5. 工厂标定与现场复核成本较高，长期稳定性与可追溯性欠佳。

因此，亟需一种在结构、驱动控制、信号处理与校准策略上协同设计的可调谐微型光谱传感器系统，以实现体积小、功耗低、抗漂移强、易标定且可溯源的光谱测量能力。

发明内容 本发明提供一种可调谐微型光谱传感器及其信号处理与校准方法，包括：

1. 光学-机电一体化结构：包括入射光学、准直与限光组件、可调谐窄带滤光器（优选MEMS法布里-珀罗干涉腔，亦可为LCTF或AOTF）、次级阶次抑制滤光片、光电探测器（单像素或小型阵列）、参考光路与温度传感器。
2. 驱动与控制：采用闭环或半闭环的调谐控制策略，通过容值/位移读出或光学参考信号实现谱位稳定；利用微振幅抖动与同步解调提升信噪比。
3. 信号处理：包括同步采样、暗电流与本底扣除、增益与非线性校正、波长映射线性化、按波长栅格的采样重排、基于仪器响应矩阵的去卷积与杂散光校正、正则化光谱重构与噪声估计、辐射定标与动态范围扩展。
4. 标定体系：提出工厂标定、现场快速复核与自校准三层级策略；通过标准光源/单色仪/气体吸收谱标尺建立波长与辐射可追溯性；采用正则化的逆问题求解获得稳定的仪器响应逆矩阵，并存储于非易失存储器中；通过内置参考发光器件与温度探头进行实时漂移补偿。
5. 数据完整性与可追溯性：在存储介质记录标定版本、系数有效期、环境条件与不确定度预算；通过校验码防篡改。

与现有技术相比，本发明至少具备如下有益效果：

• 在小体积与低功耗约束下，实现高稳定谱位控制与可溯源的波长/辐射定标；
• 引入仪器线形建模、杂散光矩阵校正及正则化重构，提高在复杂光场与低信噪比条件下的光谱准确性；
• 通过温度补偿、闭环调谐与参考通道，实现长期漂移抑制；
• 实施三层级标定流程，降低维护成本并提升现场可用性。

附图说明 图1为本发明可调谐微型光谱传感器的总体结构示意图。 图2为可调谐滤光器（以MEMS法布里-珀罗为例）的结构与驱动示意图。 图3为调谐扫描与同步采样的时序示意图。 图4为信号处理与光谱重构算法流程图。 图5为工厂标定流程示意图，包括波长标定、辐射定标与杂散光建模。 图6为杂散光响应矩阵建模方法示意图。 图7为温度漂移补偿与波长映射线性化示意图。 图8为典型重构光谱与参考光谱的对比示意图。

具体实施方式 一、装置总体结构 在一个优选实施例中，装置包括：

• 入射光学与限光：入瞳与狭缝/光阑用于定义光束角度与通光量；微透镜实现准直或近轴准直。
• 可调谐滤光器：优选采用MEMS法布里-珀罗干涉腔（上下高反射膜间距可变），其有效通带宽度在若干纳米量级（具体数值依设计确定）。为抑制高阶透射，设置阶次抑制滤光片。
• 光电探测器：硅或铟镓砷等材料的光电二极管；可为单像素器件，亦可为小型阵列以实现并行测量或参考通道构建。
• 参考光路：可包括内置参考发光二极管（中心波长稳定、具有温度补偿）及光路耦合，用于现场快速复核波长与响应。
• 温度与环境传感：温度传感器布设于可调谐滤光器与探测器附近；必要时增设湿度与加速度传感，用于环境扰动补偿。
• 模拟前端与数模转换：跨阻放大器（含可编程增益）、低噪声滤波与高分辨模数转换；为调谐器件提供高分辨数模转换与驱动波形。
• 控制与存储：微控制器或SoC管理扫描、采集与算法处理；非易失存储器存储标定系数、版本与校验码；接口包括I2C/SPI/UART等。

二、调谐与扫描控制

• 波长驱动：通过控制MEMS腔距（如静电驱动）改变谐振波长，采用分段线性或多项式电压-波长映射，配合实时温度项补偿。
• 闭环/半闭环稳定：可利用电容位移读出或参考光谱线获得反馈，减少温度与老化引起的漂移。
• 同步抖动与锁相：在目标腔距叠加微小正弦抖动，使用同步解调提升有效信噪比并估计局部斜率以辅助闭环。
• 扫描策略：采用往返扫描与重采样策略降低滞后与迟滞误差；支持多增益/多积分时间的高动态范围（HDR）采样。

三、信号处理流程 在不限制本发明范围的前提下，信号处理可包括下列步骤：

1. 原始信号预处理

• 同步采样与去抖：按调谐时序的相位点采集电信号，进行数字锁相放大或等效带通滤波；对每个波长驻留时间内的样本进行稳健平均。
• 暗电流与本底扣除：在光路遮挡或离线测得暗电流模型下扣除暗信号；考虑温度与积分时间依赖性。
• 增益与非线性校正：对跨阻放大器与ADC的增益、偏置及非线性进行校正；必要时采用分段查表。

2. 波长映射与重采样

• 波长映射：建立驱动变量（电压/位移/相位）至波长的函数λ=f(u,T)，其中T为温度。该函数在工厂标定中得到初始估计，并在现场通过参考光源微调。
• 格点化与插值：将不等间隔的采样点重采样至等间隔波长栅格，采用保持谱线形的插值方法（如基于仪器线形感知的重采样）。

3. 仪器线形与杂散光校正

• 线形建模：用窄线宽单色光源测得不同波长处的系统线形，得到随波长变化的ILS核h(λ; λ0)。
• 杂散光矩阵：采用窄带扫频与遮挡试验构建杂散光响应矩阵S，表征非理想透射、反射和散射引起的非对角耦合。

4. 光谱重构（正则化逆问题）

• 响应模型：测量向量y≈H x + n，其中H为卷积ILS与杂散光矩阵叠加形成的仪器响应矩阵，x为待求真谱，n为噪声。
• 正则化求解：采用Tikhonov或总变差（TV）正则，求解 x̂ = argmin_x ||H x − y||_2^2 + α||L x||_2^2， 其中L为平滑或梯度算子；α通过广义交叉验证（GCV）或L曲线选取。
• 约束条件：可引入非负约束与谱稀疏先验，提高稳定性并抑制物理不合理解。

5. 辐射定标与不确定度估计

• 相对与绝对定标：使用可追溯标准光源建立响应函数R(λ)并进行平场/非均匀性校正；必要时用积分球确保入瞳均匀照明。
• HDR合成：对不同增益/积分时间的谱段进行曝光补偿与稳健融合，扩展动态范围。
• 噪声与不确定度：基于残差统计与系统噪声模型估算标准不确定度u(λ)；在输出元数据中记录。

6. 漂移与温度补偿

• 温度项：在λ=f(u,T)与R(λ,T)中引入温度项与交叉项；利用内置温度传感器实时修正。
• 参考通道：周期性测量参考发光器件或外部标准，估计并修正慢变漂移。

四、校准方法 A. 工厂校准（可追溯性建立）

• 波长标定：使用单色仪或已知谱线的标准灯/气体吸收池，在多温度点采样，拟合λ=f(u,T)；建立残差界限与有效温度范围。
• 仪器线形与杂散光：在选定波长处，用近似δ函数的窄带源测定ILS；通过遮挡、阶次抑制过滤片与角度扫描构建S矩阵。
• 辐射定标：采用可追溯标准辐射源（经校准的积分球、黑体或光谱辐照度标准灯），测得响应R(λ)并进行平场校正。
• 逆算子构建：计算稳定的广义逆K，用于在线重构x̂≈K y；K可通过预设α与L求得，并按温度分区存储。
• 数据打包：将λ映射参数、R(λ)、ILS核、S矩阵、K、温度模型与不确定度预算写入只读或受保护区，并附带版本与校验码。 B. 现场快速复核
• 参考LED/激光线定位：在不同温度下测量参考线位置，更新λ映射中的零点与一阶项。
• 响应漂移校正：测量参考源强度，更新增益与偏置；对R(λ)做比例微调。
• 自检：执行暗电流、噪声底与线形一致性检查；若超出阈值，标记需要工厂复标。 C. 自校准（可选）
• 通过内置参考源与封装内反射标准进行周期性自校，维持短期稳定性。
• 利用环境传感数据，更新温度/湿度补偿模型的参数。

五、软件与数据完整性

• 标定数据管理：包括版本控制、适用温度/湿度范围、有效期与撤销机制；更新过程采用原子写与CRC校验。
• 处理流水线：在固件或上位机实现可配置的算法链路，并提供只读审计日志，记录每次测量的参数集。

六、实施例 实施例1（MEMS法布里-珀罗单像素扫描）

• 光学：入瞳直径数毫米量级，准直微透镜，MEMS FP腔配合阶次抑制滤光片，硅光电二极管。
• 扫描：在目标波段内分段线性电压扫描，驻留时间依据信噪比自适应分配；叠加小幅抖动并进行锁相解调。
• 标定：使用单色仪建立λ=f(u,T)；以积分球标准灯建立R(λ)；基于窄带源与遮挡构建S；求得K并写入存储。
• 在线：根据温度与参考LED反馈修正λ与R，采用K进行谱重构并输出不确定度。 实施例2（小阵列探测器与并行参考）
• 在主通道旁设置窄带参考通道与温度校正通道，用于快速估计漂移并闭环调整主扫描。
• 在光谱变化平滑场景采用子采样与预测插值，降低功耗。 实施例3（AOTF/LCTF替代）
• 用AOTF或LCTF替代MEMS FP，在响应模型与校准流程中相应调整ILS与温度项；其余流程等同。

七、术语说明

• 仪器线形（ILS）：系统对窄线宽输入的响应函数，反映分辨率与卷积核形状。
• 杂散光矩阵（S）：描述理想带通之外的能量耦合，含由散射、反射与阶次泄漏引入的非对角项。
• 正则化算子（L）：实施平滑或先验约束的线性算子，例如一阶差分矩阵。

八、工业实用性 本发明的装置与方法能够在紧凑封装和低功耗条件下提供稳定、可溯源且可维护的光谱测量能力，适用于批量制造与多场景部署。通过分层校准与在线自检机制，降低全寿命周期的维护成本，具备显著的工业应用价值。

以上说明书实施方式用于阐释本发明的技术原理与实现路径，所属领域技术人员可在不背离本发明精神与权利要求保护范围的情况下，对结构形式、算法细节与参数选择作出等效替换或变型。处理与校准中的具体数值、阈值与步长可根据目标波段、器件材料与应用需求进行合理设定。输出的不确定度与可追溯性信息应按照适用的计量与质量管理要求进行记录与维护。