江思远 Jiang Siyuan — 资深软件架构师 / 全栈 + AI 应用工程师 Senior Software Architect / Full-Stack + AI Application Engineer

三服务解耦架构：采集服务（44 测点 × 10 秒轮询）→ 规则报警服务（固定安全阈值，要害事件一票否决）→ 智能分析服务（预测 + 异常检测 + 健康评分），三者在同一数据流上各司其职、互不阻塞。 Three-service decoupled architecture: Acquisition (44 points × 10 s polling) → Rule Alarm (fixed safety thresholds, critical events veto) → Intelligent Analysis (forecasting + anomaly detection + health scoring), sharing one data stream without blocking each other.

• 时序大模型在线预测：进程内嵌 Chronos-2（1.2 亿参数），20 个测点批量预测未来 5 步，CPU 单轮约 0.3 秒（预算 10 秒）；采用逐列降级机制——引擎不可用或预测偏离历史波动 3 倍时该列单独回退 naive 基线，大模型失效不会拖垮系统。 Online forecasting with Chronos-2 (120M params) embedded in-process; 20 channels forecast 5 steps ahead in ~0.3 s per round on CPU (10 s budget). Per-column degradation falls a single column back to a naive baseline, so a model failure never brings down the system.
• 三层异常检测（替换 Grubbs/LSCP/SRA 内核）：① 点型层（滚动中位数抗毛刺）② 组一致性层（冗余数据源互检）③ 持续漂移层（中位数 vs 历史基线，2000 轮滚动窗口 + SPOT 自适应阈值，目标误报率千分之一）；基线快照原子持久化，重启不失忆。 Three-layer anomaly detection replacing Grubbs/LSCP/SRA kernels: ① point (rolling median de-spike) ② group-consistency (redundant-source cross-check) ③ sustained-drift (median vs baseline, 2,000-round window + SPOT adaptive thresholds, 1/1000 false-alarm target); baseline snapshots persisted atomically, restarts keep memory.
• 健康评分体系：100 − 报警扣分（每类 −10，上限 30，安全一票否决）− 三层 × 三群加权异常扣分（共 70）；含 10% 免死线与“数据不足则输出未知”的保守策略。 Health score = 100 − alarm penalties (−10 per category, cap 30, safety veto) − weighted anomaly penalties (70 total); with a 10% exemption line and a conservative “output unknown when data is insufficient” policy.
• 交付物：监控大屏、MQTT 数据契约与字段字典、面向非专业读者的算法汇报口径文档（把模型行为翻译成业务语言）。 Deliverables: monitoring dashboard, MQTT data contract & field dictionary, and an algorithm reporting doc written for non-expert readers.

机理模型 + 机器学习双引擎并行：C++ 机理引擎（边缘侧）执行可配置的物理逻辑判断，Python 预测引擎负责统计学习与异常检测，两路结果融合为系统健康评分，兼顾物理可解释性与数据驱动精度。 Physics model + ML in parallel: a C++ physics engine (edge side) runs configurable physical logic checks while a Python engine handles statistical learning and anomaly detection; both fuse into a health score, combining explainability with data-driven accuracy.

• 工程交付：MQTT 发布点表与字段字典、Modbus 点表、逻辑表、故障现象与判断条件对照表、内网离线 Ubuntu 部署方案、机理与算法两套环境的生产配置分离。 Delivery: MQTT publish point table & field dictionary, Modbus register map, logic tables, fault-symptom-to-condition mapping, air-gapped Ubuntu deployment, and production config separation for the two runtimes.

44 个测点 × 52.4 万行（全年 1 分钟粒度）的系统性选型与回测。 A systematic selection & backtest over 44 channels × 524,000 rows (one year at 1-minute granularity).

• EDA：平稳性、周期性、相关性、变点检测、功率协变量；按方差增长指数 β 分三类过程分别建模，识别“看似强周期、实为随机游走持续性”的伪周期陷阱。 EDA on stationarity, periodicity, correlation, changepoints, power covariates; channels split into three process types by variance-growth exponent β, exposing the pseudo-periodicity trap (random-walk persistence masquerading as periodicity).
• 预测回测：Chronos-2 零样本、LoRA 微调、LightGBM、ETS 与 naive 横向对比。两条反向结论：LoRA 微调整体不敌零样本；MASE 在近常量列失真，改用 RMAE 并列口径。 Backtest: Chronos-2 zero-shot, LoRA, LightGBM, ETS, naive compared side by side, with two negative findings — LoRA underperformed zero-shot; MASE breaks on near-constant channels, requiring RMAE in parallel.
• 无标签异常检测选型（合成注入评测 + 三处口径修正）：Hampel、diff+MAD、组中位数一致性、PCA 共模、CUSUM、SPOT；给出三层推荐架构。 Unsupervised anomaly-detection selection (synthetic injection + three metric corrections): Hampel, diff+MAD, group-median, PCA common-mode, CUSUM, SPOT; delivering a three-layer recommended architecture.
• 方法论工程化：notebook 单源生成 → 语法与依赖校验 → 本地干跑 → 云端 GPU 执行 → 结果回传交叉分析，全流程脚本化、可复现。 Methodology engineered end to end: notebook single-source → validation → local dry run → cloud GPU execution → cross-analysis; fully scripted and reproducible.

• A 档 15 个免训练算法（CPU 直接可跑）：指针表读数、液位/油位读数、数字表 OCR、明火烟雾、裂缝分割、跑冒滴漏粗筛、区域入侵、劳保穿戴、行为识别、零样本分割、异物检测、可见光测距、欧拉视频放大、红外热斑判级。 Tier A — 15 training-free algorithms (run on CPU): pointer-gauge reading, level/oil reading, digital OCR, flame & smoke, crack segmentation, leak screening, region intrusion, PPE compliance, behaviour recognition, zero-shot segmentation, foreign-object detection, monocular ranging, Eulerian video magnification, infrared hot-spot grading.
• B 档自训模型：红外洒漏 YOLOv8s mAP50=0.968（P 0.953/R 0.885，可交付）；可见光蒸汽分割 Mask mAP50=0.853；可见光洒漏 mAP50=0.686。 Tier B — self-trained: infrared leak YOLOv8s mAP50=0.968 (P 0.953/R 0.885, production-ready); visible steam-seg Mask mAP50=0.853; visible leak mAP50=0.686.
• 视觉多模态微调：Qwen3-VL-2B LoRA 微调（RPM-10K），500 图测试集 Acc@5%量程=78%。 Vision-language fine-tuning: Qwen3-VL-2B LoRA (RPM-10K), Acc@5% of range = 78% on a 500-image test set.
• 工程化：真实样例一键验证脚手架（ROI 限域 + 工程量标定 sidecar）、OCR 依赖独立环境隔离（uv + 版本钉死）、云端 T4 训练 → 本地 CPU 推理交付形态。 Engineering: one-command verification scaffold (ROI + calibration sidecar), isolated OCR env (uv + pinned versions), cloud T4 training → local CPU inference.

• 仪表读表（指针/数显）、安全帽检测（YOLOv8）、跌倒检测、阈值越限判定、人员位置上报；识别结果按类型分别写入巡检记录、异常台账、人员告警、位置表。 Gauge reading (analog/digital), hard-hat detection (YOLOv8), fall detection, threshold-breach judgement, personnel location reporting; results written by type into inspection records, anomaly registers, alerts, and location tables.
• 越限结果直接触发平台已有的自动化工作流生成审批待办——打通“算法 → 业务系统 → 流程闭环”，算法侧与业务侧职责边界清晰。 Breach results trigger the platform's existing automation workflows to create approval to-dos — closing the algorithm → business system → workflow loop with a clear responsibility boundary.

• Go（Fiber）后端 + 静态前端（ECharts）：Modbus TCP 轮询（100 ms 周期）与 TCP 二进制帧高频频谱双链路，自定义帧协议支持多客户端、多通道、分片重组。 Go (Fiber) backend + static ECharts frontend: dual links of Modbus TCP polling (100 ms) and TCP binary-frame spectra; a custom frame protocol with multi-client, multi-channel, fragment reassembly.
• 存储分层：TDengine 存时序数据，SQLite 存配置与账户；WebSocket 按需推送频谱（幅度谱/时域/时频/瀑布图/谱质心/谱滚降/倒频谱），另接 SFTP 远程录波。 Storage layering: TDengine for time-series, SQLite for config & accounts; WebSocket on-demand spectrum push (amplitude/time/time-frequency/waterfall/centroid/rolloff/cepstrum), plus SFTP waveform recording.
• 生产交付：esbuild 打包 + 静态资源嵌入二进制 + -ldflags "-s -w" -trimpath，单可执行文件部署。 Delivery: esbuild bundling + assets embedded in the binary + -ldflags "-s -w" -trimpath, deployed as a single executable.

• 面向 ARM64 服务器（国产 OS）与 K8s 的容器化微服务：采集器、信号处理（FFT 频谱特征）、故障诊断（规则引擎 + 知识图谱推理）、存储、API 服务各自独立。 Containerized microservice for ARM64 servers (Chinese Linux distros) and K8s: collector, signal processing (FFT features), diagnosis (rule engine + knowledge-graph reasoning), storage, and API each independent.
• Kafka 解耦数据流；边缘侧先做 FFT 降低传输带宽；NPU 加速推理；规则与知识图谱以 YAML 外部化，业务逻辑与代码分离。 Kafka decouples the data flow; FFT computed on the edge to cut bandwidth; NPU-accelerated inference; rules & knowledge graphs externalized as YAML, separating logic from code.
• Rust workspace 多 crate 工程组织，配置驱动。 Rust workspace with multiple crates, configuration-driven.

• 3 台行业无人机 + 3 座自动机巢 + 32 km 管线一屏监控：实时地图（位置/状态/航向）、可见光 + 红外双窗口视频、平台算法报警实时联动与地图定位。 One screen over 3 drones + 3 automated nests + a 32 km pipeline: live map (position/status/heading), visible + infrared dual video, real-time alarm integration with map positioning.
• 前端 Svelte + TypeScript + Vite + Leaflet + flv.js；后端 Python FastAPI + aiomqtt + httpx；WebSocket 实时推送 + Webhook 接收平台报警。 Frontend Svelte + TypeScript + Vite + Leaflet + flv.js; backend Python FastAPI + aiomqtt + httpx; WebSocket push + Webhook alarm intake.
• 支持 iframe 嵌入客户其他平台并提供 postMessage 双向通信协议；深色大屏 UI 适配 1080p/2K/4K；Nginx + systemd/Docker 部署，本地离线地图瓦片。 Supports iframe embedding with a bidirectional postMessage protocol; dark large-screen UI for 1080p/2K/4K; Nginx + systemd/Docker with offline map tiles.

• 诊断规则库、目标设定与跟踪、数据字典与因子库、负荷预测（训练 + 敏感性分析）、预测驱动的 MILP 调度（含强化学习版本）。 Diagnostic rule library, target setting & tracking, data dictionary & emission-factor library, load forecasting (training + sensitivity analysis), forecast-driven MILP scheduling (incl. an RL variant).
• 交付：评审稿、汇报稿、开发计划与数据字典；核心口径为单位能耗碳排放（tCO₂/tce），达标口径 = 范围 1 + 2 − 抵消。 Deliverables: design review, presentation, dev plan & data dictionary; core metric is carbon per unit energy (tCO₂/tce), compliance = Scope 1 + 2 − offsets.

• 40 张数据表 / 45 个页面 / 10 条自动化工作流，覆盖经营驾驶舱、CRM、进销存、人事、项目、财务、OA、售后工单、合同管理、审批中心，以及 AI 识别子系统。 40 data tables / 45 pages / 10 automated workflows, covering executive dashboards, CRM, inventory & procurement, HR, projects, finance, OA, after-sales tickets, contracts, approval center, and an AI recognition subsystem.

• C++ 单机程序：Modbus RTU 采集 + 模型计算 + Redis 存储 + MQTT 发布 + 日志模块；软件按模块解耦（采集/模型/存储/发布各自独立）。 Standalone C++ program: Modbus RTU acquisition + model computation + Redis storage + MQTT publishing + logging, strictly modularized (each independent).
• 部署于 ARM64 控制器：Redis Stack + webdis + Nginx 提供 Web 访问面，systemd 托管，安装脚本经 shc 加密分发；设备名/波特率/从站 ID/任务周期/计算更新周期全部外部化为 .env 配置，现场可改不改代码。 On an ARM64 controller: Redis Stack + webdis + Nginx web layer, systemd-managed, installers encrypted via shc; device name/baud/slave ID/cycles all externalized to .env — field changes without touching code.

• C++ 后端（Crow + standalone asio 异步 HTTP + nlohmann/json + SQLite）内建前端页面，单进程即提供完整的“设备 + 网站”能力，无需依赖运行时框架。 C++ backend (Crow + standalone asio async HTTP + nlohmann/json + SQLite) with built-in frontend — a single process delivers the full "device + website" capability with no runtime framework.
• Windows 原生 gcc-arm-10.2 交叉编译直出 aarch64 可执行文件（自建 45 MB 精简 sysroot），彻底脱离 WSL 依赖；含硬件交互（旋钮拨码触发密码重置）与 PLC 寄存器读写适配。 aarch64 executables cross-compiled directly from native Windows gcc-arm-10.2 (custom 45 MB minimal sysroot), removing the WSL dependency; includes hardware interaction and PLC register read/write.

• Go CLI + CODESYS SDK / IronPython 脚本，把“库工程创建 → POU 生成 → 编译属性配置 → 运行时系统文件生成 → 编译库导出”固化为一两条命令，替代原先的手工配置流程。 Go CLI + CODESYS SDK / IronPython scripts collapse "library project creation → POU generation → compile options → runtime system file → compiled library export" into one or two commands, replacing the manual process.
• Windows 原生交叉工具链直出 Linux aarch64 动态库；以安装包形式分发给工程师使用，使用方无需任何额外环境。 Native Windows cross-toolchain directly produces Linux aarch64 shared libraries, distributed as an installer — recipients need no extra environment.