毕业生知识库

      随着制造技术的迅猛发展，日渐复杂的工业生产过程对智能化监测技术的 需求日益迫切。得益于传感器技术与计算机技术在生产过程监测中的大规模应 用，能够反映生产设备运行状态的监测数据得以被大量收集与存储。大量的历 史数据客观上推动了基于数据驱动的故障预测与健康管理技术的发展。借助故 障预测与健康管理相关技术，可以建立复杂工业过程模型，实现生产过程健康 状态的实时监测、故障诊断、以及工业系统剩余使用寿命的预测，并以此为基 础辅助制定合理的维护计划及决策，为人员、设备安全，以及生产活动的顺利 进行提供保障。         随着生产过程的复杂化，许多生产过程具备多种运行工况，使得监测数据 呈现多模态分布；同时，生产环境的变化和生产需求的变更，也会使工业生产 系统的运行出现新的特征。这些问题的出现显著增加了监测任务的复杂程度。 目前，将故障预测与健康管理技术应用于复杂多模态生产过程时仍存在许多挑 战。本文围绕多模态生产过程的健康监测任务，在异常检测、故障诊断以及剩 余使用寿命预测等方面开展研究工作。论文的主要工作如下：

       1) 针对传统基于单模型的异常检测方法对多模态过程检测精度偏低的问 题，提出了一种基于双自编码模型的异常检测方法。首先，构建基于双自编码 器的正常过程重建模型，采用两个子模型分别完成宏观数据与微观数据的重建 工作，并通过两个模型重建结果的融合，实现对多尺度数据的高精度重建。其 次，构建了一个预测模块，估计当前监测数据的正常重建误差，并对重建模型 输出的实际重建误差进行校正，通过减少不同模态正常监测数据重建误差分布 差异的方式，减少因模态误判导致的误检。最后，计算校正后的监测数据重建 误差相对于其正常分布范围的偏移量，来衡量系统运行状态的异常程度。测试 结果表明，所提出的方法能够有效提高多模态生产过程中的异常状态的检测精 度。

       2) 针对实际生产过程中由于新出现的故障样本数目稀少，难以保证诊断模 型的充分训练，从而影响故障识别精度的问题，提出了一种基于特征聚类的小 样本故障诊断方法。该方法借助历史数据集训练了一个具备特征聚类功能的基 础模型，将不同的故障样本投影到不同的特征簇中，以各故障的统计中心作为 对应故障的表征，并在训练过程中，采用类平衡的数据增广策略扩充训练数据， 提升模型的泛化性能；然后在固定基础模型特征提取模块的前提下，利用新故 障样本与历史故障样本微调模型，将诊断模型迁移到扩展后的故障诊断任务中， 并利用少量新故障样本计算新故障的表征，在此基础上依据监测样本与各故障 面向多模态生产过程的健康监测方法研究 II 表征的相似程度完成小样本条件下的故障诊断。测试结果表明，该方法能够有 效缓解小样本条件带来的模型训练过拟合问题，对新故障具有较好的学习与识 别能力。

       3) 针对传统网络扩展类增量式故障诊断方法容易造成诊断模型扩增过快的 问题，提出了一种新的增量式故障诊断方法。在诊断模型的更新过程中，该方 法将网络扩展划分为故障特征提取模块的扩展与故障识别模块的扩展两部分。 当出现新增故障需要进行模型更新时，如果能够依据现有特征进行故障分类， 则维持特征提取模块不变，通过调整故障识别模块的方式完成新故障的识别； 如果现有特征提取模块无法有效提取新故障特征，则对其进行扩展，学习新故 障的特征提取能力，然后再完成故障识别模块的更新。在诊断模型的更新过程 中，该方法通过减少特征提取模块不必要的调整，抑制整个诊断模型规模的增 长速度，并在新特征提取模块的学习过程中，借助知识蒸馏策略，促使该模块 更加有效地提取新故障特征。测试结果表明，相较于已有的模型扩展类增量式 故障诊断方法，该方法能够以更加缓慢的网络扩展速度获得更高的故障识别准 确度。

       4) 针对传统剩余使用寿命预测方法对多工况系统的预测精度偏低的问题， 提出了一种基于卷积神经网络的多模型预测方法。首先，在数据预处理阶段， 采用最近邻方法识别系统的当前工况，并通过计算当前监测数据相对于正常工 况数据相对偏差的方式，简化监测数据分布，并抑制多模态数据对剩余使用寿 命预测任务的影响。在此基础上，为提高对存在个体差异的系统的剩余使用寿 命预测精度，构建多模型剩余使用寿命预测模型，采用多个功能不同的子特征 提取器提取监测数据的特征，并依据系统的运行状态选择合适的特征进行特征 融合，计算系统当前时刻的剩余使用寿命。测试结果表明，该方法能够有效提 高多工况系统剩余使用寿命预测精度。

       最后，对本文工作进行了总结，并对未来的研究工作进行了分析与展望。 

    With the rapid development of manufacturing techniques, increasingly complex industrial production processes demand intelligent monitoring methods for their monitoring tasks. Since the wide application of sensor and computer technologies in industrial production process monitoring, the data reflecting the operation states of these processes can be extensively collected and stored. The massive monitoring data have objectively driven the development of data-driven prognostic and health management techniques. The models for complex industrial processes can be established with these techniques to achieve the monitoring tasks of these processes, such as real-time health monitoring, fault diagnosis, and remaining useful life prediction. Then, reasonable maintenance schedules can be designed based on the results of these tasks, which guarantee stable production processes and the safety of personnel and equipment.

  As industrial production processes become more complex, many processes have various operation settings. As a result, the monitoring data of these processes often have multimodal distributions. Meanwhile, changes in environments and production requirements also give rise to new characteristics in these processes. These issues significantly increase the difficulties of these monitoring tasks. As a result, the application of current monitoring techniques on multimode industrial processes still faces many challenges. This dissertation focuses on the health monitoring tasks of multimode industrial processes. A series of studies have been carried out in the areas of anomaly detection, fault diagnosis, and remaining useful life prediction. The main works of this dissertation are summarized as follows:

  1) To improve the performance of traditional single-model-based methods for anomaly detection tasks on multimode industrial processes, a novel bi-auto-encoder-based anomaly detection method is proposed. Firstly, a bi-auto-encoder model is constructed for the data reconstruction of normal data, which uses two sub-auto-encoders to reconstruct the macro and micro data, respectively. This model fuses the reconstruction results of both sub-auto-encoders to achieve high-precision reconstruction for multiple-scale data. Secondly, a prediction module is used to estimate the normal reconstruction error of the current monitoring sample and compensate for the actual reconstruction error output from the reconstruction model. The false detection caused by mode misidentification is suppressed by reducing the difference in reconstruction error distribution of normal monitoring data from different modes. Finally, the deviation of the compensated reconstruction error relative to its normal distribution is calculated to measure the anomaly degree of the monitored industrial process. The testing results show that the proposed method can effectively improve the detection accuracy of anomalies in multimode industrial processes.

    2) Due to the sparse sample number of newly emerged faults in real industrial production processes, traditional diagnosis methods usually have insufficient ability to identify new faults in the absence of sufficient new fault data for model training. To address this problem, a novel feature-clustering-based few-shot diagnosis method is proposed. Firstly, a feature clustering-based diagnosis model is trained on a historical data set, which transforms different fault samples into different feature clusters. The fault representations are calculated as the statistical centers of corresponding feature clusters. During the training procedure of the model, a class-rebalanced data augmentation strategy is used to expand the training data to improve the generalizability of the model. To meet the expanded fault diagnosis tasks, the diagnosis model is fine-tuned using new fault samples and historical samples with its feature extraction module fixed. The representation of each new fault can be calculated with the features of a small number of corresponding samples. Then, the few-shot fault diagnosis can be achieved based on the similarity between the feature of the monitoring samples and each fault representation. The test results show that the proposed method effectively alleviates the overfitting problem of model training caused by the few-shot conditions, and has good generalizability to new faults.

  3) Traditional network-expansion-based incremental fault diagnosis models usually have rapid expansion speed for their network structures during their model update processes. To address this problem, a novel incremental fault diagnosis method is proposed. During the model update processes for new tasks, the network expansion is divided into the expansion of the feature extraction modules and the expansion of the fault identification module. If new faults can be identified based on existing features, the feature extraction modules remain unchanged. The recognition of new faults is accomplished by adjusting the fault identification module. If the feature extraction modules fail to extract valid new fault features, they are extended to learn new feature extraction capabilities for new faults. Then, the fault identification module is updated. During the model update processes, the network expansion speed of the diagnosis model is suppressed by reducing the unnecessary adjustment of its feature extraction modules. During the training procedure of new feature extraction modules, a knowledge distillation strategy is employed to improve the modules’ feature extraction capability for new fault features. The test results show that the proposed method can achieve higher diagnosis accuracy with slower network expansion speed than existing network-extension-based methods.

    4) Traditional remaining useful life prediction methods have relatively low prediction accuracy for multimode industrial systems. To handle this problem, a novel multiple-model prediction method based on convolutional neural networks is proposed. Firstly, in the data preprocessing step, the nearest neighbor algorithm is employed to identify the current working condition of the monitored system. The distribution of monitoring data is simplified by calculating the relative deviations of the current monitoring data from the normal condition data, which alleviates the influence of multimode data for the remaining useful life prediction tasks. Secondly, a multiple-model-based prediction model is constructed to improve the prediction performance for the systems with individual differences. Multiple sub-feature extractors with different functions are used to extract the sub-features of the monitoring sample, and the proper sub-features are selected and merged based on the system’s working status. Finally, the remaining useful life of the system is predicted by the merged feature. The test results show that the proposed method can effectively enhance the remaining useful life prediction precision for multimode industrial systems.

    Finally, the results of the research are summarized, and future work is analyzed and prospected.