Effective chiller fault diagnosis is of great importance for maintaining a better service and energy efficiency. Deep learning proficiently solves some problems challenging Artificial Intelligence and becomes one of the excellent candidates for fault diagnosis recently. This study proposes a novel fault diagnosis strategy for a chiller, which merges simulated annealing (SA) into a deep neural network (DNN) to obtain effective, efficient, and stable performance. The proposed SA-DNN strategy is carefully compared with DNN and back-propagation (BP) network. The results show that SA-DNN enhances the diagnostic accuracy, shortens the running time, and greatly improves the model stability. The optimal network structure has 2 hidden layers (HL) with each layer 64 nodes, and the overall diagnostic accuracy for seven typical faults attains 99.30%. The nodes in the first HL are proved to be dominant over those in the second or behind because the mapping of the second can hardly make corrections if that of the first is deformed already. The global faults are hard to be diagnosed due to the global effect, but the proposed strategy achieves satisfactory results with the highest individual accuracy reaching 99.79% for excess oil and the lowest 97.52% for refrigerant leakage. The features used for diagnosis have an influence on the accuracy of the proposed method.

有效的冷水机故障诊断对于维持更好的服务和能源效率至关重要。深度学习可以有效地解决一些挑战人工智能的问题，并且最近成为故障诊断的极佳候选者之一。这项研究提出了一种新的冷水机故障诊断策略，该策略将模拟退火（SA）合并到深度神经网络（DNN）中以获得有效，高效和稳定的性能。拟议的SA-DNN策略与DNN和反向传播（BP）网络进行了仔细比较。结果表明，SA-DNN提高了诊断的准确性，缩短了运行时间，大大提高了模型的稳定性。最优的网络结构具有2个隐藏层（HL），每层包含64个节点，对七个典型故障的总体诊断准确性达到99.30％。事实证明，第一个HL的节点比第二个或后面的节点占主导地位，因为如果第一个HL的节点已经变形，则第二个的映射很难进行校正。全局故障由于全局影响而难以诊断，但是所提出的策略取得了令人满意的结果，对于过量的机油，单个精度最高，达到99.79％；对于制冷剂泄漏，最小精度为97.52％。用于诊断的特征会影响所提出方法的准确性。但是所提出的策略取得了令人满意的结果，对于多余的机油，单个精度最高，达到99.79％，对于制冷剂泄漏，最小精度为97.52％。用于诊断的特征会影响所提出方法的准确性。但是所提出的策略取得了令人满意的结果，对于多余的机油，单个精度最高，达到99.79％，对于制冷剂泄漏，最小精度为97.52％。用于诊断的特征会影响所提出方法的准确性。