Today, manufacturing machines are continuously equipped with various sensors, whose data enable to derive a comprehensive picture of the current state of each machine. Predictive maintenance approaches make use of this data in order to predict the occurrence of possible failures before they actually occur, thereby significantly reducing production and service costs. The application of machine learning to sensor data streams is an essential part of data-driven predictive maintenance in order to find the patterns in the data that are indicators of upcoming faults. Thus, research on machine learning for predictive maintenance is a recent and very challenging field. However, there are currently no appropriate data sets available that can be used for this kind of research. In this paper we therefore propose an approach for the generation of predictive maintenance data by using a physical Fischertechnik model factory equipped with several sensors. Different ways of reproducing real failures using this model are presented as well as a general procedure for data generation.
Today, we are facing the beginning of a transformation towards the fourth
industrial revolution, called Industry 4.0, which is characterized by
manufacturing and service innovations based on cyber-physical systems, big data and an
intensive use of methods from artificial intelligence [
For PredM to work, knowledge is required about characteristic data patterns that are indicators of specific faults that have occurred or that are likely to occur in the future. Due to the large number of potential faults as well as the large variety of production machinery and components used, the manual definition of such patterns is not feasible. Instead, machine learning (ML) is required to automatically derive such patterns from available data. Patterns could describe healthy states as well as states which are in the transition phase towards an upcoming breakdown. The application of ML to PredM, however, comes along with a variety of challenges, in particular related to the complexity and heterogeneity of data, the lack of labelled data, the need for transfer learning, as well as the necessity of deriving models that enable explainable decision support. Thus, PredM is an ideal application area for research in ML due to its complexity and practical relevance. However, appropriate and realistic data is required to conduct such research. While plenty of data sets are available for research purposes in ML, there is a lack of data that can be used immediately for PredM applications. Also it is nearly impossible (at least for Universities) to get real data from industry due to the serious confidentiality issues involved.
In this paper we therefore address the issue of obtaining data appropriate for ML research in PredM. The contribution of this paper is manyfold. First, we present a brief overview of PredM and the involved research challenges for ML (Sect. 2). Then, we characterize the required data to address these challenges, analyze existing data sets as well as methods for the generation of new research data (Sect. 3). The main contribution of the paper is the presentation of an approach for the generation of PredM data based on a physical model of a specific production environment implemented based on a Fischertechnik (FT) model factory (Sect. 4). We further describe various ways for injecting faulty behaviour in a defined way into the FT model factory, in order to collect the respective data that can be used to learn the related patterns for prediction. We describe the current state of realization as well as our planned future work (Sect. 5). 2 2.1
Industrial maintenance involves all measures that are required to ensure or to re-establish the proper functioning of industrial machinery. The goal is to prevent the occurrence of failures that could lead to breakdowns or downtimes of machines or that could lead to safety concerns. For example, a common cause of failure is wear, which is a gradual damage or deformation of material due to forces, which occurs in many mechanical components, in particular in bearings, O-rings, or gears. In general, systematic maintenance procedures increase the availability of machines, reduce costs, and enable to schedule required maintenance actions. Traditional, preventive maintenance involves the systematic inspection of machines following a fixed time schedule or a fixed mileage, which is based on the simplified assumption that failures mostly occur after a certain and known operating time or effort. However, it often happens that failures occur before the scheduled maintenance activity or that maintenance actions are performed although they are not yet necessary. Thus, PredM aims at overcoming the fixed time schedule approach by introducing methods that are able to individually predict upcoming failures. The goal is to perform maintenance actions only when they are really necessary, i.e., not too early and not too late. For companies, PredM has the advantage that maintenance costs can be reduced significantly by better utilization of capacities and by avoiding downtimes in manufacturing.
PredM is based on forecasting failures based on the current state captured by various sensors, such as vibration, temperature, humidity, or acoustic sensors. In addition, parameters characterizing the current state in the production process (e.g. position sensors or switches as well as the activity state of actuators) are relevant. Machines in real production environments may have hundreds of various sensors producing data streams with high frequency. As there is a wide variety of factories, plants, and other systems that depend on reliability, it is not economically feasible to build a customized solution for each plant. Thus, research is focusing on universal parts of machines, such as bearings, gearboxes, motors, valves, pumps, compressors, and so on to find universal solutions. 2.2
The increasing number of sensor data streams makes manual monitoring and
analysis impossible, which is why ML and especially deep learning are suitable
for PredM data processing [
Research Data for Machine Learning in Predictive Maintenance
For conducting ML research for PredM (as well as for other kinds of Industry 4.0 applications of ML), it is necessary to have data available that is to some degree comparable to the data in industrial settings. Thus, data sets are necessary which are composed of various data streams with different characteristics (according to the type of sensors used in industry) together with related data about the current status of the production component or process. For learning to predict failures, the data streams must contain patterns, similar to those that occur in real situations, for example as a result of wear. Also the data sets must be at least partially labelled with the respective fault to be predicted. Ideally, we need large data sets describing several instances of the same fault and data sets describing the same fault in various different but similar components (e.g. wear in the bearings of different motors) to investigate transfer learning approaches. 3.2
Eker et al. [
The available data sets known to us have a variety of issues which makes them not suitable for ML research on PredM. The data sets are either too small for ML, several of them have anonymized features, and nearly all sets are insufficiently labelled for the mentioned PredM tasks. Further, the data sets usually only consider individual components or working station cells, but do not provide a comprehensive snapshot of a factory’s sensor data. 3.3
Since there is no sufficient or adequate data from real industrial factories publicly available for research purposes, it is desirable to collect or generate them. Sensor data generation without the real production environment at hand can be categorized into four groups: 1. fully synthetical, 2. synthetical based on previous data, 3. synthetical based on a virtual simulation model, and finally 4. based on a simplified physical model.
Fully Synthetic Data Generation Fully synthetic data generation means
that sensor data is generated by an algorithm based on given parameters. The
resulting streams are based on a statistical structure and can contain concept
drifts (changing of underlying statistical properties over time). Typical
parameters are the data generating distribution (e.g. Gaussian), noise rate, data
dimensionality, and generation periodicity. For instance, Hahsler et al. [
generate sensor data is to learn the underlying properties of an existing data
distribution in order to generate new data. This can be done by training a
generative and discriminative neural model by learning either explicitly the
parameters of the distribution [
Synthetic Data Generation Based on a Virtual Simulation Model A
further approach is the creation of a virtual simulation model with the properties
of the real model and use this for data generation. This approach, for example,
has been applied to aircraft gas turbines [
a virtual model of a factory or machine, there is also the possibility of using a real but simplified physical model. Regarding the level of abstraction and the constituents of the model, such models can be divided into two categories.
The first category are models which are equipped with real industrial
components leading to minor abstractions. Examples of such factories are Learning
Factories [
The second category consists of models with a higher level of abstraction,
which are build using non-industrial components. The advantage of this
approach is the significantly low cost involved in building such a model. There
are several platforms which enable the simple cost-efficient construction of such
models. Among them the most popular are Lego Mindstorms4 and
Fischertechnik (FT)5. Examples are the Smart-LEGO Factory6 at DFKI, the FT plant
model implemented for applicability validation of Industry 4.0 components [
Generating Research Data for Predictive Maintenance by FT Model Factory As publicly available data sets for research on ML for PredM are rare and limited to single components or working cells, we now describe an approach for the generation of such data based on a simplified physical model. The aim is to provide a cost-effective way to generate data according to the requirements sketched in Sect. 3.1. This requires constructing a physical model of a factory, attaching appropriate sensors and related data collection hard- and software, as well as developing means for simulation faults. 4.1
Our Industry 4.0 factory model is built based upon the FT Factory Simulation1 as shown in Fig. 1. It has been selected due to its superior robustness compared to Lego Mindstorms. As FT is also used in University education for automation engineers, the available FT components are already closer to real industrial components than those of Lego. The FT factory model that we use consists of four modules: a sorting line with color detection, a multi-processing station with oven and milling machine, a high-bay warehouse, and a vacuum gripper 3 http://www.smartfactory.de/ 4 https://www.lego.com/en-us/mindstorms 5 https://www.fischertechnik.de/en 6 https://www.dfki.de/web/aktuelles/dfki-cebit-2016/smart-lego 1 https://www.fischertechnik.de/en/service/elearning/simulating/fabrik-simulation24v robot. Each module is operated by its own controller based on an ARM Cortex A8 CPU with various analog and digital input/output ports running under a LINUX kernel. Overall, the model consists of nine light barriers, ten switches, twelve motors and three compressors. Moreover, we enhanced the model with six three-axis acceleration sensors that are mounted on motors and compressors for vibration measuring and four differential pressure sensors are measuring the pressure generated from the three compressors. These sensors are connected to a separate Raspberry Pi controller. Furthermore, two micro-electro-mechanical systems (MEMS) each with a gyroscope, an accelerometer, and a geomagnetic sensor will be installed on the Robotic Vacuum Gripper and the High-Bay’s storage and dispensing machine. RFID-Tags for product identification and a reading device are also planned. Further, we will extend the oven model with a heating pad in order to change the color of thermo-colored product materials. This process will be monitored by a thermal imaging camera.
All controllers are connected via an Ethernet network and communicate via
remote procedure calls. The overall control software for the entire production
process is distributed over the controllers, each of which is in charge of a
certain module of the factory. For processing the generated data, we selected the
SMACK stack [
The overall manufacturing process is designed as a cycle, meaning that data can be generated without manual interference. The process starts from the High Bay where workpieces are dispensed and transported to the Multi Processing Station. After processing, they are sorted by color, transported by the Robotic Vacuum Gripper and finally stored in the High Bay where the process repeats. 4.2
By using the FT model along with the developed software, the manufacturing process is executed in a continuous loop. As FT blocks are quite robust and all physical connections are very stable, problems occur quite rarely and hence the model is able to run properly over a very long period of time. However, in order to be able to produce data for predictive maintenance, faults must occur such that the resulting data can be collected. As such faults do not occur naturally (within an acceptable time limit) realistic faults must be artificially infused into the model.
Figure 2 describes the interplay between reality, our FT model, the creation of faults, and finally the data generation. In general, reality defines which failure types are measurable and reasonable. Our FT model is a smaller and simplified representation of reality and due to this it certainly restricts our ability to reproduce realistic defects. Also life expectancy for components in real machines is months to years and degradation processes are very slow. Thus we have to compress the time dimension, i.e., we have to significantly shorten the time during which a certain type of fault causes its typical effects. Based on these limitations we define plausible defects that can be simulated by our physical model such that data is generated that can be used for learning and evaluating prognostic models on predictive maintenance.
In general, there are several ways in which behaviour can be generated similar to a failure in reality.
abnormal behaviour of the physical model, additional actuators can be integrated whose activities cause certain disturbances. For example, workpieces can be pushed from the conveyor, a pressure line can be virtually broken by inserting a pressure valve, or an additional motor can be inserted to produce an additional mechanical load on a drive shaft.
Adapting the Controller Software for Actuator Based on knowledge about how certain failures (e.g. motor problems due to wear) have an impact on an actuator (reduced or unstable revolution speed), the controller software can be designed such that it controls the actuator in a way that it behaves as if it would exhibit the failure. For example, the motor supply voltage can be reduced following a certain pattern or the frequency for the pulse-width modulation of motor power supply can be lowered to increase vibration.
Simulating Defective Sensors Faults related to a defect of any kind of sensor are also quite likely. They could also have a significant impact on the production process, in case the sensor is used within the control procedure of the machine. For example, a defective position switch might cause problems, as a gripper is not able to adjust itself to the correct position. Defective sensors can be easily simulated as part of the control software by manipulating the value they produce. 4.3
The just described ways of generating faulty behaviour have to be embedded into an overall generation process for maintenance data. Therefore, the following data generation process has been developed, allowing to generate a large number of labelled maintenance data sets automatically. This process runs in a loop consisting of four steps (see Fig. 2): 1. Selection of the particular error (e.g. motor failure due to wearing) to be produced in the current run, including the relevant parameters (which motor, degree of wearing, failure pattern curve, time horizon of wear process, etc.). 2. Configuration of the controller software to run the factory in a mode, in which the respective fault reproduction is enabled. 3. Start of the controller software to run the production process. During the run of the factory, all data is collected and stored in an Apache Cassandra data base and labelled with the respective failure being produced. 4. After the failure has occurred, the factory model is reset to a defined initial state compensating for any inconsistencies that might have resulted from the insertion of the failure. 4.4
In reality, bearing faults are the biggest failure source with almost 40% to 50%
of electrical motors. They result in vibration signatures of higher amplitudes,
increased noise, and also in a reduced motor torque and thus motor speed [
In this paper, we address the problem of data generation to enable ML research for PredM. We surveyed currently available data sets and present several approaches for data generation. We then present an new approach for PredM data generation based on a FT factory model. As of today, the mechanical and electrical side of the model is nearly completely realized, the sensor data is collected, processed, and stored using the SMACK-Stack as described. First failures, as the example case just described, are implemented.
Future work will first of all address the implementation of additional failure
scenarios based on the approaches described in Sect. 4.2. This work is quite
difficult as it requires at least a basic understanding of typical mechanical faults
and their consequences in order to be able to reproduce them on the model.
In general it would be desirable to find ways of validating to which degree the
data we are collecting is realistic compared to real production data. However,
this will be clearly difficult as data from comparable real components, such as
provided in [