Predictive maintenance (PdM) is a prominent anomaly prediction strategy in the manufacturing system given the increasing need to minimize downtime and economic losses. It is available for PdM to monitor industrial equipment continuously with smart electrical sensors and predict the health condition with machine learning algorithms. However, the performance of previous algorithms is often limited by lacking consideration of both attribute contribution to final results and temporal dependence. To solve the problem, this article introduces a general PdM framework based on Internet-of-Things technology, cloud computing, and total productive maintenance. In this framework, an attribute attentioned long short-term memory network ($A^2$-LSTM) is proposed. The $A^2$-LSTM takes a sequence of electrical records as input to extract attributes. Afterwards, different attributes are fused into the attribute attention network, which can adjust the importance of each attribute automatically. Next, the reweighted attributes are fed into the health prediction component to establish temporal dependence for the manufacturing system. Finally, the output of $A^2$-LSTM, i.e., remaining useful life, can support workers to carry out equipment maintenance. The effectiveness of the method is verified by real-world cases and the comparison results show that the $A^2$-LSTM is promising.

鉴于对减少停机时间和经济损失的需求日益增加，预测性维护 (PdM) 是制造系统中一种突出的异常预测策略。PdM 可以使用智能电气传感器持续监控工业设备，并使用机器学习算法预测健康状况。然而，由于缺乏对最终结果的属性贡献和时间依赖性的考虑，先前算法的性能通常受到限制。为了解决这个问题，本文介绍了一个基于物联网技术、云计算、全生产性维护的通用PdM框架。在这个框架中，一个属性注意的长短期记忆网络（${\mathrm{一个}}^2$-LSTM) 被提出。这${\mathrm{一个}}^2$-LSTM 将一系列电子记录作为输入来提取属性。然后，将不同的属性融合到属性注意网络中，可以自动调整每个属性的重要性。接下来，重新加权的属性被输入到健康预测组件中，以建立制造系统的时间依赖性。最后，输出${\mathrm{一个}}^2$-LSTM，即剩余使用寿命，可以支持工人进行设备维护。通过实际案例验证了该方法的有效性，对比结果表明：${\mathrm{一个}}^2$-LSTM 很有希望。