Physically Verifiable Evidence and LLM-Based Reporting for Bearing Fault Diagnosis 基于物理可验证证据和LLM报告的风机轴承故障诊断

TL;DR

• The paper addresses the lack of physically verifiable evidence in AI-based fault diagnosis by proposing a Diagnostic Evidence Network (DENet) that outputs classification, predicted characteristic frequency, and temporal localization of transient impulses.
• The deviation between predicted and theoretical frequency serves as a label-free validation signal to detect misclassifications with high AUROC values (0.970 and 0.871), even in high-confidence regimes where traditional confidence scores fail.
• A QLoRA-adapted language model is constrained to translate diagnostic content without generating new information, reducing unsupported claims from 10-12% to 2% and eliminating fabricated quantities.
• The framework is encoder-agnostic and tested across four encoders and three public datasets, showing no statistically significant accuracy cost while enhancing trustworthiness in safety-critical applications.

Why It Matters

This work is highly relevant for AI practitioners and researchers focusing on trustworthy deployment of AI systems in safety-critical domains like industrial maintenance. By introducing physically verifiable evidence and constraining generative models to avoid hallucinations, it directly addresses two major barriers to real-world adoption: the inability to validate predictions against physical reality and the risk of unreliable automated reporting. These advancements are essential for building robust, interpretable, and deployable AI solutions in engineering contexts where errors can have severe consequences.

Technical Details

• Diagnostic Evidence Network (DENet): An encoder-agnostic multi-task framework extending standard classifier outputs to include structured evidence records comprising classification labels, predicted characteristic frequencies (comparable against theoretical values derived from bearing geometry and shaft speed), and temporal localization of transient impulses inspectable on raw waveforms.
• Validation Signal: The difference between predicted and theoretical frequency acts as an inference-time, label-free detector for misclassifications, achieving AUROC scores of 0.970 and 0.871 across datasets, outperforming softmax-based confidence detectors especially in high-confidence regions.
• Frequency Estimation Accuracy: Achieves approximately 6 Hz error on 1,024-point segments where spectral estimation methods are structurally inapplicable, demonstrating practical utility even under challenging conditions.
• Language Model Constraint: Utilizes QLoRA adaptation to constrain a pre-trained language model strictly to translation tasks—converting diagnostic findings into natural language reports without generating new content—which reduces unsupported claim rates from 10–12% down to ~2% and completely eliminates fabricated numerical or qualitative assertions.
• Experimental Setup: Evaluated using multiple neural network encoders (including CNNs and transformers) applied to three publicly available bearing fault vibration datasets; results show consistent improvement in verifiability without compromising primary diagnostic accuracy metrics.

Industry Insight

Industrial AI teams should prioritize integrating physical consistency checks into their diagnostic pipelines—not just relying on internal model statistics—to ensure predictions remain grounded in domain-specific knowledge before triggering maintenance actions. Adopting frameworks like DENet enables proactive detection of potential failures during operation rather than post-hoc auditing, significantly reducing downtime risks. Additionally, when deploying LLMs for automated reporting, organizations must implement strict constraints preventing generative freedom unless absolutely necessary; fine-tuning via parameter-efficient methods such as QLoRA offers a viable path toward maintaining fidelity while minimizing hallucination risks in technical documentation workflows.

TL;DR

• 提出诊断证据网络（DENet），通过输出结构化证据记录实现故障预测的物理可验证性。
• 利用特征频率偏差作为无标签的推理时验证信号，在高置信度区域仍能有效检测错误分类。
• 采用QLoRA约束语言模型仅翻译不生成内容，将报告中 unsupported claim 率从10-12%降至2%。
• 在四个编码器和三个公共数据集上验证了方法的有效性，且未造成显著准确率损失。
• 解决了安全关键系统中AI诊断缺乏物理现实校验及生成式报告幻觉两大信任问题。

为什么值得看

该工作针对工业AI落地中“黑盒决策不可信”的核心痛点，提供了一种无需修改底层模型即可增强可解释性与可靠性的工程化方案，对推动高安全性场景下的AI部署具有直接指导意义。其提出的物理验证机制与受限生成范式，为其他领域建立类似可信诊断框架提供了可复用的方法论参考。

技术解析

• DENet框架解耦编码器与输出层，支持任意主流编码器接入，扩展输出包含类别标签、理论对比频率值及瞬态脉冲时间定位三重证据结构。
• 特征频率预测误差被设计为独立于softmax置信度的异常检测指标，在1024点短段频谱估计失效场景下仍保持约6Hz精度，AUROC达0.970/0.871。
• 验证信号不依赖真实标签，在模型高置信但实际错误的“盲区”区域表现优于传统置信度阈值法，填补现有验证体系空白。
• 报告生成模块基于QLoRA微调大语言模型，强制其仅执行诊断内容的语义映射与格式化转译，禁止任何事实性内容创作，从源头杜绝虚构数据。
• 实验覆盖三类公开轴承数据集与四种不同编码器架构，证明方法具备强泛化能力且不影响原始诊断性能。

行业启示

• 工业AI系统应强制引入“物理一致性检查”环节，将数学预测结果映射到可测量的实体参数进行交叉验证，避免纯统计驱动的风险累积。
• 在维护报告等决策支撑场景中，生成式AI应采用“约束式翻译”而非自由生成功能，通过架构限制消除 hallucination 隐患，保障信息可靠性。
• 未来智能运维平台需构建多层级信任评估体系：底层用物理规律校验中间层模型输出，顶层用受限生成确保人机交互内容真实可控，形成端到端可信链条。