【摘要】 振动信号分析是机械设备状态监测与故障诊断的核心手段，但其面临着一个核心困境，即传统专家系统流程固化，而端到端的深度模型虽具自适应学习能力，却存在“黑箱”与可复现性不足等问题。本文提出一种自主信号处理的神经符号多智能体框架：以大语言模型（LLM）作为决策中枢，协同一个由可解释的、符号化的信号处理算子构成的工具库，实现自主的振动信号分析与诊断。框架采用规划-执行-审查的多智能体架构，迭代优化信号处理决策链。同时为确保LLM规划的逻辑自洽，避免出现算子的错误调用，所有算子都依据其维度变化与语义变换特性被形式化规约。具体地，将算子划分为升维、同维、降维、多信号四类，并给出供大模型理解的语义信息加以约束。轴承故障诊断数据集上的验证表明：本框架能够自主生成具有清晰物理含义的信号处理决策链，成功复现了“包络谱-峭度”等专家级可解释的诊断算法。在清华大学轴承数据集的单域测试中，Gemini 2.5 Pro版本达到97.8%的准确率；在渥太华大学变转速数据集的跨域测试中，仅用“加速”和“减速”工况训练，在未见过的工况上实现了99.3%的准确率，证明了框架的泛化能力。该研究为构建可信、可复现且可扩展的新一代智能诊断系统提供了一种具有潜力的范式。更多还原

【Abstract】 Vibration signal analysis is a cornerstone of machine condition monitoring and fault diagnosis,yet it faces a central dilemma.Traditional expert systems have rigid workflows,while end-to-end deep models,despite their adaptive learning abilities,suffer from being‘black boxes’ with insufficient reproducibility. This paper introduces a neuro-symbolic multi-agent framework for autonomous signal processing. The framework utilizes a large language model(LLM) as a central decision-maker, coordinating a toolbox of interpretable,symbolic signal processing operators to enable autonomous vibration signal analysis and diagnosis. The framework adopts a Plan-ExecuteReview multi-agent architecture to iteratively optimize the signal processing decision chain. To ensure the logical consistency of the planning and prevent incorrect operator calls,all operators are formally regulated based on their dimensional and semantic transformation properties.Specifically,they are constrained by semantic information for the LLM to comprehend. Validation on bearing fault diagnosis datasets shows that this framework can autonomously generate signal processing decision chains with clear physical meaning and has successfully reproduced expert-level, interpretable diagnostic algorithms such as ‘envelope spectrum-kurtosis’. In single-domain tests on the Tsinghua University bearing dataset,the Gemini-2.5-pro version reached an accuracy of 97.8%. In cross-domain tests on the University of Ottawa variable-speed dataset, the framework, trained solely on ‘acceleration’ and ‘deceleration’ conditions, achieved 99.3% accuracy on unseen conditions,proving its generalization ability. This research provides a promising new paradigm for building trustworthy,reproducible,and scalable nextgeneration intelligent diagnostic systems.更多还原