Condition Monitoring

The starting point of predictive maintenance is the same problem as preventive maintenance, which includes efficiency loss due to downtimes and unexpected stops in the production facilities. On the other hand, predictive maintenance uses data analytics to analyze machine health via vibration, pressure, and temperature data in order to predict machine malfunctions before unexpected stops occur. When the AI algorithms detect the anomaly in the dataset, in other words, observe irregular behavior from the standard parameters, the system alerts technicians in the field to check the condition of the machine.

Cost reduction seems to be the key benefit of condition-based predictive maintenance. Since it does not require skilled staff and maximizes the efficiency of the monitored machine, it directly saves money by creating planned maintenance optimization. However, this cost-saving is not the most significant benefit of predictive maintenance; the bathtub explains it. This theory says that if any equipment is newly assembled, it starts working with all the risks in the early failure zone, and the risk of failure is high until it moves into the safe operating zone. If you do not carry out your planned maintenance activities when necessary, you increase the risk of failure every time. The production facilities that have the technology to predict when equipment could fail with condition monitoring techniques do not only save money through efficient production capacity also predictive maintenance directly extends the lifetime of the machines used.

Obviously, the side benefit of predictive maintenance is the continual and steady production process. Since the system predicts the machine malfunctions before it happens, required maintenance and repair actions are taken, and the production line does not stop anytime. 

Rotating equipment naturally generates vibrations and operates on these vibrations. The aim is to keep these vibrations at acceptable levels to ensure production reliability. Therefore, although various sensors can be used in predictive maintenance applications in rotating equipment, the most basic measurement unit is vibration. At this stage, what information the signals received from the accelerometers contain and the processability of this information into the machine learning model constitute the basic performance parameter in anomaly detection. Two main features can be extracted from the acceleration signals collected from the accelerometers -features extracted in the time domain and features extracted in the frequency domain-.

While metrics such as RMS, Crest, Kurtosis, Peak can be obtained from the time domain, features such as total harmonic distortion, 1X,2X,…,10X harmonic indications, spectral centroid and sideband energy are extracted from the frequency domain. At this stage, it is highly important to transfer the features in the frequency domain to the model to create a reliable machine learning model. The frequency content is the signature of the normal or abnormal state of that machine. On the other hand, it is equally important to enter the speed information into the model for harmonic analysis independent of the vibration sensor. When we include each frequency amplitude in the frequency domain from a 3-axis vibration sensor to the machine learning model as a feature, it is seen that more than 20,000 features are transferred to the model in each measurement, depending on the sensor bandwidth. So here, the machine learning model can compare many features in multidimensional mathematical space better than the human eye can distinguish between data in 2D and 3D space. In this way, even the most sensitive anomalies can be detected according to the resolution of the model.

There are hundreds of different machine types that vary according to the usage area in the industry. Many of these machines operate in cycles, power and processes arising from factors such as constantly changing production speed, raw materials and processes. E.g; A simple rolling mill will be operated at different speeds depending on the desired material quality, material thickness and production speed. Vibration data from the machine will also vary depending on the speed. In rule-based predictive maintenance applications, anomaly alarms will occur at every speed change, but this is not true. At this stage, each process parameter (speed, power, etc.) must be included in the model and first of all, the machine modes must be determined. Vibration data were collected at 4 different speeds from a test roller. When these data were inserted into the machine learning model, it was observed that there were 6 different modes. In these modes, 4 of them work at different speeds, and 1 of them is the mode in which the machine does not work. The last mode is the mode in which the anomaly occurs.

If the measurements are not colored in the chart above, we can observe that they are in 5 different groups. However, as can be seen in the graph, the model has 6 different modes. The reason for this is that the machine operating cycle in the cluster, which is divided into 2 groups, is the critical speed for that machine, creating a new abnormal operating mode by creating mechanical looseness in time in the machine assembly and coupling over time as you can see as purple measurements in Figure.2. Although the standard deviation of the data in the cluster, which is divided into 2 groups, is clearly evident compared to the other clusters, it is very difficult to distinguish that there are 2 different modes in that cluster.

Two measurements operating in the same mode are as above and can be clearly distinguished between blue and purple plot in the spectrum. However, it is very difficult to distinguish these two measurements by looking at the average Vrms values. Of course, another reason for this is that Vrms is insufficient for high frequency vibrations, although it varies from machine to machine, Vrms in some fault types and Grms static alarms in some fault types may be insufficient for anomaly detection. Therefore, transferring all spectral information to the machine learning model ensures optimum results in anomaly detection.

In AI-based predictive maintenance applications, in the absence of historically labeled data, supervised learning is not possible, so anomaly detection using unsupervised learning algorithms will be the best start for the first step. In order to implement this, meaningful features from the data received from the sensors should be included in the models. After the anomaly detections are approved by the maintenance personnel inspection, the semi-supervised learning model can be started by labeling the detected anomalies with their root causes. With the labeling of the data, the root-causes of the anomalies of the equipment can be estimated after the anomaly. Therefore, a generalized AI model based on data is very difficult due to variations such as transmission types and working conditions that change according to industry. For this reason, starting the applications of generalized anomaly detection algorithms using machine-specific transfer learning methods will be the first step in the process. Then, the detected anomalies are labeled by the field workers on the system and root-cause analysis is made. By all these efforts, the first step of the AI-Based Predictive Maintenance journey has been taken.

Although condition monitoring systems are generally applied to machines that operate continuously with high downtime costs, when criticality analysis is made, condition monitoring systems can be applied to machines working in shifts or machines that make stop-starts if they are more critical. In this type of machine, the measurements taken at certain periods are insufficient. While it causes the measurement to be missed, on the one hand, it also negatively affects battery power consumption in wireless sensors, as mentioned above. With the integration of PLC-SCADA systems into the condition monitoring systems, optimum results can be obtained with the SCADA system triggering the status monitoring system while the machine is running.

Integration of condition monitoring systems into PLC-SCADA systems is not only at the point of triggering and optimizing measurement periods, but also at the point of sharing all collected data with condition monitoring systems, along with current process parameters and predictive analytics applications, general equipment efficiency, and production and maintenance outputs such as key performance indicators. It is also possible to monitor and control online.

Considering that machine tools are the most productive equipment of manufacture, it is expected to make them operate at their maximum capacity and avoid production losses caused by unplanned repair costs, downtime, or quality problems. Thus, keeping them in optimum condition is a priority in highly efficient production systems. This situation can be achieved through a preventive maintenance strategy for critical components of the machine tool. Thanks to this strategy, the maintenance task can be pre-planned according to the actual maintenance need of the equipment. For this reason, a decent maintenance plan which includes condition monitoring for spindles becomes important to reduce downtime and total cycle cost, and to increase the service life of the machine tool.

In condition-based monitoring, there are several monitoring techniques to monitor possible failure conditions of components. Among these, oil analysis, thermography, ultrasonics, and vibration are the most common. Each of these techniques has different abilities to predict startup failure. As we mentioned in the P-F curve in our previous blog on condition monitoring, malfunctioning machines can signal failure with increased vibration levels before contaminant particles on the oil are detected. Thanks to its versatility and reliability, vibration analysis as machine tools has become one of the most accepted and used for rotary equipment. Vibration monitoring detects deteriorations such as wear, imbalance, misalignment, and fatigue in parts that are under rotational or reciprocating motion.

Machine tool spindles are complex delicate assemblies with a considerable number of components. Each of them plays a vital role in the functionality of the spindle. The individual performance of the components contributes to a certain extent to the overall performance of the spindle. Bearings in machine tool spindles are the most sensitive component in practice. Consequently, any major issues are usually reflected and detected through the bearings. Most of the time, other issues affect the performance and working life of bearings. Therefore, bearing issues are the most common malfunctions encountered in the spindle.

Roller bearing life (L₁₀) is a nominal calculation at the design stage that does not take all factors, which may affect the bearing life, into account. Due to various static and dynamic loads such as working at variable speeds, lubrication conditions, vibrations generated during machining, unbalance and assembly tolerances, roller bearing life ends before the calculated cycle. Therefore, periodic maintenance of the spindles is of utmost significance in order to prevent crucial stoppages and high damage to the machine.

As we have just explained above, monitoring the condition of the spindles has great importance for CNC machines. In the industry, such studies are usually carried out every three or four months, or the equipment health is monitored with the built-in temperature sensors in the grinding wheel spindle. Nevertheless, such maintenance strategies can have costly consequences for the long haul. Therefore, condition-based maintenance is the most effective method for achieving maximum performance. At this point, applications in the industry monitor systems by taking measurements at certain time intervals. This technique does not constitute a very applicable process as it is not known whether the CNC machine is working during those time intervals or not. Besides, parameters that vary according to the application such as fly cutters, machining surface, and rotation speeds also make the situation even more chaotic. Sensemore solves this problem with the Trigger device. Trigger sends the measurement order to the sensor by transmitting the 5V signals to the receiver with a code to be returned at the end of each process. Thus, it can monitor the vibrations of a CNC lathe in the grinding wheel spindle at the end of each machining and in a way that the parameters remain the same in every measurement.

In the maintenance operations, downtime and spare part costs are the most widely known issues. But the other costs are hidden in the processes. If you concentrate on downtime and spare costs, you probably miss the costs of processes and operations. But this is a trap so many companies fall. When the costs of operations increase, most companies reduce training cost, the number of auditors, maintenance frequency and so on. In short term, they save money and survive, but in long-term this is not sustainable. The short-term plans provide unsustainable improvement policies which cause a loss of money in long-term.  In order to gain sustainable advantage in long-term, these improvement processes must be long-term. This issue not only requires decisions, but corporate policy changes.

In order to improve operational reliability in long-term, modern predictive maintenance (PdM) technologies are really important. PdM is actually a very cost-effective approach, which saves 8% to %12 more than Preventive maintenance and 40% than fix-when-broken approach. At this point, we need to mention about Condition-Based PdM and Traditional PdM. The biggest difference between them is time. Both monitor the health and condition of machines such as pumps, motors, reductors, fans etc.  However, the traditional one consists of periodical measurements performed once in 60 or 90 days. On the other hand, certain problems can occur between the measurements due changing working conditions, temperature or load and cause catastrophic problems. Condition-based provides continuous monitoring and provides opportunity to maintenance teams react faster and plan their schedule.

As I mentioned above, the visible costs are widely known. But hidden costs can cost more than those. The vibration, misalignment, assembly and other issues issues can cause defects on bearings, couplings and gaskets. Due to this simple issue, machine consumes more energy. Let me give you an example. Consider there is an imbalance of 0.11 Watt/g.mm. If there is 50 g imbalance in a 500 mm shaft, there would be 2.75 kW energy waste. Moreover, with the help of resonance and increased damping capacity of oiled interfaces, the imbalance will be bigger, so the waste. Generally this simple imbalance issues cost 5% more energy consumption. In rotating machinery working in continuous manufacturing environments, even 1% can mean millions of dollars. In addition to this, also carbon emission will be larger.

Downtime was very underestimated in the past. These days people get familiar with the issue. But downtime is not just the time the broken machine does not work. It is also the other machines cannot work, the personnel cannot work and the personnel work for overtime to fix the issue and produce ordered amount. A skipped maintenance work can pave the way for other elements to stop and unless the root cause is well analyzed, the frequency of downtime will increase. In addition to these, a manufacturing line after a fix lose 10% of its efficiency in re-start period.

The other hidden cost is human resources and training. Let us start with an example. Human Resources Department of a factory hires a technician for PdM and instead of Condition-Based systems, they give this technician a hand-held vibration analyzer. They train the technician, and the technician starts getting measurements point by point, one by one. After 2 years of work, the technician quits the work. The factory needs a new technician and information the technician collected these 2 years is lost. They need find a new technician and collect the data and face the same problems in maintenance as happens before him/her. Or COVID-19 breaks out and due to the limitations, the technician cannot enter the plant as commonly occurs in these days. Since there is no condition monitoring system is established or any one else in the factory can fill up the technician’s position, the conditions of all machinery go dark and malfunctions happen.

To sum up, without ignoring the visible costs, you should also consider hidden costs, a manufacturing plant has and build up maintenance operations. Since the short-term plans will always cause long-term loses, the long-term plans should be made, and corporate culture should be reshaped. Condition-Based is a strong tool to increase efficiency and reduce cost in long-term. Beyond the advantages of the traditional PdM, it saves both money, workforce and energy. With the increased reliability, maintenance teams can plan their schedule, define priorities and save workforce.

There are various methods applied in industry for this. The methods vary from machine to machine but generally vibration, temperature and current are monitored. At this point, choosing the most suitable method makes the job easier. For example, a current sensor can be applied for an electric motor to detect failures. On the other hand, a current sensor cannot provide you a meaningful data for pumps, compressors and fans. For these kinds of machines, it is better to use a vibration sensor. Due to this reason, the equipment on the manufacturing line should be classified and determined about how critical they are. This will lead to choose the right sensor and method. If a critical equipment will be measured and continuous data should be kept, online monitoring must be used.

In Condition Monitoring systems, the time between failure and detection of a fault is a critical parameter for choosing the right sensor. On the graph above, we can check P-F curve. This curve represents the relation between potential failure (P) and functional failure (F). Potential failures are the point where the equipment starting to show defects. For example, from the moment when G_RMS value of a bearing exceeds 3 G, till the moment when this bearing is unserviceable; there are 15 days. On P-F curve, you can measure time from horizontal axis and the state of the equipment can be seen on vertical axis. The bearing in this example should be kept under monitoring for potential failures.

Condition Monitoring is a critical tool to detect the time interval on P-F curve. By using these methods, you can maximize this time interval. As you can see on the curve, oil analyses, acoustic emission, thermography and other methods exist. But the important point is maximizing the lifetime of an equipment instead of changing it regularly. This increases the efficiency and decreases the costs.

Rotating equipment is a term defining motors, reducers, pistons and centrifugal machinery. In industrial applications, condition monitoring is generally applied to rotating equipment. The most common method for these items is vibration analysis. Vibration analysis is the technique to measure vibration severity and spectrum. The data gathered are used to define the condition of machine and detect potential faults. By this method, problems such as imbalance, bearing faults, looseness, crooked shaft and even cavitation can be observed. Vibration allows you to detect potential faults up to 3 months before faults become failures. If you are interest in vibration analysis, you can reach our blog post by clicking the link above.