预测性维护
In the era of digitization the concept of Smart Factory attracts a lot of attention. Modern industry becomes connected and highly automated. Such factories need their machines to run smoothly and with minimal down times. Predictive maintenance helps to deal with breakages. It aims to identify possible failures and helps to schedule the maintenance of detected devices.
在数字化时代,智能工厂的概念吸引了很多关注。 现代工业变得互联且高度自动化。 这样的工厂需要他们的机器运行平稳,停机时间最少。 预测性维护有助于处理破损。 它旨在识别可能的故障并帮助安排检测到的设备的维护。
In current blog post I illustrate the process of building a model that detects failures of factory machines. I use an open dataset from one of the Schwan’s factories. It contains time series values that include telemetry, device errors and failures. The aim is to predict device failure 12 hours before it happens. An assumption is that half of a day is enough for a technician to react and to handle a possible issue.
在当前的博客文章中,我说明了构建检测工厂机器故障的模型的过程。 我使用Schwan一家工厂的开放数据集。 它包含时间序列值,包括遥测,设备错误和故障。 目的是在发生故障前12小时预测设备故障。 假设半天时间足以使技术人员做出React并解决可能的问题。
探索数据集 (Exploring dataset)
The first thing to do is reading the dataset and loading data into pandas data frames. This logic is quite trivial, so I do not post it here. Instead you can access the original Jupyter Notebook at this link.
首先要做的是读取数据集并将数据加载到熊猫数据框中。 这种逻辑非常琐碎,因此我不在此发布。 相反,您可以通过此链接访问原始的Jupyter Notebook。
In general there are 3 files to read: telemetry.csv, failures.csv and errors.csv. At the end we get 3 pandas data frames: telemetry_df, failures_df and errors_df.
通常,有3个文件需要读取:telemetry.csv,failures.csv和errors.csv。 最后,我们得到3个熊猫数据帧:telemetry_df,failures_df和errors_df。
![Image for post](https://img-blog.csdnimg.cn/img_convert/b576030dd774e1d8323021eb6772386e.png)
![Image for post](https://img-blog.csdnimg.cn/img_convert/c0cda151a27ce43476d014289a2701f0.png)
In total there are 876100 rows of telemetry values for 100 machines during 1 year on an hour basis. This is a lot, but most of the time devices work well. All in all the data set contains only 3919 errors and 761 failures. The latest are the values we try to predict.
在1年中,以小时为单位,总共有100台机器的876100行遥测值。 这很多,但是大多数时候设备运行良好。 所有数据集总共仅包含3919个错误和761个失败。 最新是我们尝试预测的值。
We have data for 100 machines. In real predictive maintenance cases it often makes sense to create a separate model for each machine for having best predictions. In this example we assume that one model might work for all the device