基于跨运行稳定统计特征与类别感知自适应融合的流程工业故障诊断方法研究

Process Industry Fault Diagnosis Based on Cross-Run Stable Statistical Features and Class-Aware Adaptive Fusion

摘要: 工业时序文本化有助于利用过程变量的语义信息，但会压缩原始信号中的频域结构和局部变化特征；与此同时，辅助数值信息对不同故障类别及不同样本的补偿作用并不一致。针对上述问题，本文构建基于跨运行稳定统计特征的数值分支，并提出类别感知自适应融合方法(Class-Aware Adaptive Fusion, CAAF)。该方法首先根据故障效应方向与幅度在两个独立开发运行间的一致性筛选主频幅值占比、主频位置和单调性三类统计特征，通过变量身份映射与轻量编码器生成数值类别得分；随后，将传统样本级融合权重扩展为样本–类别级权重，采用类别系数刻画数值证据对不同候选类别的修正作用，并利用文本置信度门控调节其样本介入程度。TEP 21类故障诊断结果表明，CAAF的准确率达到94.51%，较纯文本分支、静态融合和QMF分别提高1.83、1.14和0.75个百分点，在IDV-16至IDV-20五类困难故障上的平均准确率达到89.35%。消融实验、多种子重复训练和数值信息扰动实验进一步表明，性能增益主要来自稳定统计信息对文本分支类别相关错判的选择性补偿。

Abstract: Converting industrial time series into structured text facilitates the use of process semantics but compresses frequency-domain structures and local variations in the original signals. Meanwhile, the complementary value of auxiliary numeric information varies across fault classes and samples. To address these issues, this paper constructs a numeric branch based on cross-run stable statistical features and proposes class-aware adaptive fusion (CAAF). Dominant-frequency amplitude ratio, dominant-frequency location, and monotonicity are first selected according to the consistency of fault-effect directions and magnitudes across independent development runs. Variable-identity mapping and a lightweight encoder then transform these statistics into numeric class scores. CAAF further extends conventional sample-wise fusion to sample-class-specific weighting: class-wise coefficients characterize class-dependent numeric corrections, while a text-confidence gate controls sample-dependent intervention. Experiments on the 21-class Tennessee Eastman Process benchmark show that CAAF achieves an accuracy of 94.51%, exceeding the text-only branch, static fusion, and QMF by 1.83, 1.14, and 0.75 percentage points, respectively. Its average accuracy on the five difficult classes from IDV-16 to IDV-20 reaches 89.35%. Ablation, multi-seed repeated training, and numeric-information perturbation experiments further indicate that the improvement mainly arises from selective compensation for class-dependent errors of the text branch.