Читать книгу Is My Machine OK? - Robert Perez X. - Страница 12
Оглавление“In God we trust; all others must bring data.” W. Edwards Deming,American statistician, professor, author, lecturer and consultant
W. Edward Deming’s quote above illustrates how important he thought data was when attempting to analyze processes. Decision makers always want to know if their processes are steady or changing, and if they are within required tolerances. Similarly, machinery data is vital when trying to state your case about what should be done to correct an undesirable situation. How can you possibly convince a manager to shut down and repair a critical machine without a reliable and convincing set of data? Remember, however, that raw data alone is not information; it is only by carefully analyzing data that we can hope to convert raw data into helpful information. Data analysis can allow you to paint a picture of a machine’s present condition and predict what might happen in the future. The goal in this chapter is to present a common sense approach to data analysis through the use of examples.
Humans are excellent at recognizing and interpreting patterns. (Some would say we are a little too good at seeing celebrity faces, animals, secret symbols, etc., in grilled cheese sandwiches, potato chips, and water stains. Also, there are those who read all kinds of secret messages in magazines, revered documents, and books.) We can quickly spot if the data are steady, trending upward, erratic, or cyclical, to name a few common patterns. The first challenge is to objectively study critical data sets and determine if they are normal or not. Below we will discuss the most commonly encountered data relationships and what they might mean.
Data sets are any group of measurements from a monitoring system—such as a distributed control system (DCS) or a supervisory control and digital acquisition system (SCADA)—which has been archived in a digital format using a data acquisition system, along with a time variable that allows variables to be analyzed and compared. A lion’s share of process and machine-related data that we analyze is static data. This type of data has only amplitude. Typically, static data is analyzed in one of three ways:
1.A single data set is plotted versus time to spot trends.
2.Several data sets are plotted together to see how they compare to one another.
3.Two data sets are plotted one versus the other to see if there is a relationship, i.e. correlation, between the two.
Let’s look at a few commonly encountered data patterns. (In the following analyses, we will assume the data is reliable and accurate, i.e., the appropriate sensor is installed in the right machine location and has been properly calibrated. In the real world, we should always question whether data is dependable.) In theory, there are countless data patterns that can be generated from machine data. Some may have no rhyme or reason, whereas others, luckily, yield patterns that tell clear tales about what is happening. Here we will concentrate on the patterns that tell a story about a machine’s condition.
A data series is any collection of values of interest. The data value may represent machine load, machine vibration, a bearing temperature, a process flow, or any other variable of interest. A time series or trend is when a data series is plotted or analyzed with respect to time. Trend plots are popular because they are easy to create and interpret. They can be used to see what is happening to a machine or process variable over time and predict what will happen in the future. The example trends in Figure 5.1 show the satisfaction rating of two product brands that are plotted with respect to time. Here we can make a few observations:
1.The satisfaction ratings of both brands are improving over time and
2.Brand B is fairing slightly better in rating trials.
Figure 5.1 Example trends
Flat Trends
The simplest types of trend are the flat trends as seen in Figure 5.2. The “low” trend is the type of trend we would all like to see. The amplitude is steady and about 40% of the alarm value. We can conclude that this is a well-applied, healthy machine with no indications of an impending problem.
The “high” trend in Figure 5.2 shows a flat data trend that exceeds the recommended alarm level. The first question that should be asked is: Are these high levels a recent occurrence or have levels been high forever, or at least since anyone can remember? If the answer is high levels are a recent occurrence, the reason for the sudden change must be investigated. If the answer is high levels have been around for a while, then one of the following reasons for the consistently high trend must be considered:
Figure 5.2 Two examples of steady state trends
1.Design problem
2.Assembly problem
3.Installation problem
4.The machine is being operated in a way it was not intended to be operated.
5.The machine has been misapplied.
In order to determine which of these might be the cause of the high readings, it is important to have start-up data. If amplitudes were high immediately after start or commissioning and have remained flat, then one of these root causes may apply.
Trends with Step Changes
Another version of a flat trend is a trend with a step change, as seen in Figure 5.3. Step change plots can either illustrate a step change up or a step change down. Whenever a step change is seen, we can deduce that something has changed suddenly inside the machine or around it.
Figure 5.3 An example of an upward step change and downward step change in data levels
Here are a few machinery or process-related reasons for an upward step change:
•Something has become lodged in an impeller, resulting in a sudden imbalance.
•An oil cooler has become suddenly plugged or fouled, leading to higher bearing temperatures.
•A downstream process strainer partially plugs, leading to a higher pump discharge pressure.
•A reciprocating compressor valve fails, causing a sudden rise in compressor discharge temperature.
Here are a few machinery or process-related reasons for a downward step change:
•An upstream process strainer partially plugs, leading to a lower pump flow.
•A cold front blows in, causing sudden oil cooling and subsequently lowers bearing temperatures.
•A reciprocating compressor valves fails due to gas stream particulates, causing a sudden drop in compressor flow.
•A sudden drop in centrifugal compressor speed reduces vibration levels as it falls below a critical speed.
When dealing with step changes, it is useful to talk about percent changes in value observed. Let’s start with the simple equation below to define the term “alarm margin” or AM.
| (4.1) |
where a is the alarm level, c is the current measurement, and n is the normal level. The alarm margin provides a quick and easy way to determine how much margin there is between the present measurement and level where action will be required. Let’s look at a few examples to clarify the use of this equation. Let’s first assume we are looking at vibration. The normal level is 0.15 i.p.s., the alarm level is 0.5 i.p.s., and the current vibration reading is 0.35 ips. Plugging these terms into the alarm margin equation, we get:
This result means 42.9% of the alarm margin remains before reaching the alarm point. If we now assume c=0.45 ips, then the remaining life is:
Here, only 14.3% of the “normal” alarm margin remains. An AM of 100% means you are operating at a normal level and you have 100% of the alarm margin remaining. Working with these two examples, suppose the remaining life changes from 42.9% to 14.3% in just one week. In this case, it’s easy to see you probably should be planning a repair very soon. However, if this same change in condition takes a year; you probably have a week or two to put a repair plan together before having to shut down the machine for a repair. A good rule of thumb for step changes is to investigate any step change that represents a 25% change in the alarm margin.
Upward and Downward Trends
The next commonly encountered trends are upward and downward trends as seen in Figure 5.4. These types of trends provide clear indications that something is changing. The main difference between step changes and upward and downward trends is that upward and downward trends occur over weeks, months, or years whereas step changes are sudden events that occur over minutes, hours, or days.
Here are a few machinery or process-related reasons for an upward trend:
•A pump is gradually eroding, fouling, or corroding, which leads to a change in rotor balance.
•Internal centrifugal compressor clearances are wearing over time and causing a gradual increase of steam consumption to maintain the required flow. This results in an upward trend in driver speed over time.
•A gradual fouling heat exchanger is causing high oil temperature and hence higher bearing temperatures.
•Gradual wear in a sleeve bearing leads to a gradual increase in shaft vibration.
Figure 5.4 An example of an upward trend and downward trend in data levels
Here are a few machinery or process-related reasons for a downward trend:
•Internal pump wear is causing a gradual drop in flow.
•Gradual plugging in a downstream reactor bed is generating higher pump backpressure and, hence, lower pump flows.
•An expected gradual drop in ambient temperatures due to change in the season leads to a predictable drop in bearing temperatures.
•Vibration levels on a variable speed drive gradually drop as pump flow demands drop.
Data Sampling
A commonly asked question is: How frequently should data be sampled to see what is going on? The first question to ask is: What is the sampling frequency? DCS and SCADA system have the capability of sampling and storing data every few seconds or less. However, as time passes, the stored values are either lost completely or averaged so that they may be archived. Data acquisition systems can average older data over longer and longer intervals in an attempt to compress it into more compact forms, i.e., first hourly then daily, then weekly, then monthly, etc. Therefore, the sooner you analyze the data, the better your chances are of getting high time resolution data for analysis.
Manually collected data are usually acquired on a set collection interval. For example, vibration data are typically collected quarterly, monthly, or weekly whereas pump efficiency data are collected annually. This means you are often stuck with a paucity of data to work with. If these values are critical to the plant, then a continuous means of measuring and storing this data should be considered.
For long-term trends—such as those caused by gradual internal wear—weekly or monthly data is probably fine for analysis purposes. However, for rapid events that may occur during start-ups or shutdowns, the shortest data collection interval is preferred. This means that you need to get to the DCS or SCADA data and extract the pertinent data before archiving occurs.
Cyclic trends
Cyclic trends are data trends that appear to oscillate repeatedly over some time interval. The cyclic example shown in Figure 5.5 is a three-year temperature trend of noon air temperatures in a California vineyard. It is easy to see that temperatures oscillate from about 40°F in the winters to almost 110°F in the summers. The trick with cyclic trend is to analyze the data over a time period that significantly exceeds the characteristic cycle period so that multiple cycles may be observed.
Two common causes of cyclic trends are 1) seasonal or daily changes in ambient temperatures and 2) load changes due to seasonal process demands. Another special example of a cyclical trend is a process that operates in batch processes that last over a prescribed time interval, such as a Coker Unit where cycles last between about 24 hours. Here you should see obvious cyclical trends with a minimum of a 24–hour period.
Is it the Machine or the Process?
When a sudden change or trend in data is spotted, your first question should be: Is it the machine or the process? The real questions should be: Is the machine simply responding to a normal change in process conditions?
Figure 5.5 Daily noon air temperatures at a California vineyard over time
Is the process upset or deteriorating? Or, is the machine deteriorating? Before focusing in on a machine problem, interview the process operators to see if either flows, pressures, temperatures, product compositions, etc., have changed or if there is a known process issue, such as plugged reactor bed or fouled cooler. Once you have determined process conditions are normal and steady, you can begin analyzing machine data.
Another popular and frequently used data analysis method is correlation analysis. There is a correlation between two variables if a statistical relationship exists between them. The easiest way to see if a potential correlation exists is to plot one variable versus the other in Excel® or other graphing software. In Figure 5.6, gas turbine output power is plotted versus the ambient air temperature. This plot was generated using the “X-Y Scatter” function of Excel®.
Figure 5.6 illustrates a negative correlation between gas turbine power output and ambient air temperature. It is easy to see that there seems to be a clear relationship between gas turbine output power and the ambient temperature. Note that this correlation is a negative one—as the air temperature rises, the output power falls. Most readers probably already know this correlation exists. The point of this example is to show what a correlation looks like and how they are generated.
Figure 5.6 An example of a data set with a negative correlation, m = −0.2012 and R2 = 0.938
Table 5.1 What R2 Values Mean
