Instruments

Listening got a number.

For the first decades of rotating machinery the instrument was a person: a screwdriver pressed to the housing with the handle at the ear, a hand on the bearing cap, and the memory of how it felt last month. The judgement was real, but it could not be written down or compared between two people. In 1939 T. C. Rathbone published a chart that put amplitude against frequency and drew bands across it, from smooth to very rough, and severity had a number. The 1950s brought the velocity pickup (lesson 5) and the vibration meter: an overall reading on a dial and, with a tunable filter and patience, one frequency at a time. Slide along the timeline.

IEPE: two wires, both directions.

The piezoelectric crystal of lesson 5 gives a tiny charge and needs a special cable and a charge amplifier. In the 1970s the amplifier moved inside the sensor, and the result is the IEPE accelerometer (Integrated Electronics Piezo-Electric; ICP is one maker's name for the same thing), the standard sensor on routes and in plants ever since. It runs on two wires. The collector pushes a constant current of a few milliamps up the cable; the sensor's electronics hold their output at a bias of around 11 volts and let the vibration ride on it as a small AC signal; the collector blocks the DC and keeps the ride. A general-purpose unit gives 100 millivolts per g, measures to ±50 g, and is flat from about 0.5 Hz to 10 kHz.

Three sensitivities cover nearly everything. 100 mV/g is the everyday choice. Slow machines, below about 600 rpm, use 500 mV/g to lift tiny accelerations above the noise. Gearboxes and anything with hard impacts use 10 mV/g so that 500 g does not clip the amplifier. Mounting sets the top of the usable range (lesson 5): a stud to 10 kHz, a magnet to about 1 or 2 kHz.

Noise: the floor you cannot see below.

Every sensor and every amplifier adds a hiss of its own, quoted as a spectral density in micro-g per root hertz. What shows in the spectrum is that density times the square root of each line's width, so the floor is not a property of the sensor alone. Widen Fmax at the same number of lines and the floor rises; add lines and it falls, because the same hiss is shared among more of them. A bearing tone of half a milli-g stands clear of a 25 µg/√Hz floor and is gone against a 1000 µg/√Hz one.

Fmax: how high to look.

Fmax is the top of the spectrum you ask for. The collector samples at 2.56 times it (lesson 6) and filters away everything above it, so a frequency beyond Fmax is not off the edge of the picture, it was never recorded. Too low and the gear mesh and the bearing ring are missing. Too high and the same lines, spread over a wider range, are coarser, and the detail at the bottom blurs. Rules of thumb: ten times running speed for a general check of a motor or pump, three and a quarter times the gear mesh frequency for a gearbox, and 5 to 10 kHz when the question is early bearing wear.

LOR: how finely.

Lines of resolution, LOR, is how many bins Fmax is divided into. Each line is Fmax ÷ LOR wide, and the recording must last LOR ÷ Fmax seconds, one over the line width (lesson 7). The twin peaks of the two-pole motor and the mains, 0.7 Hz apart, need lines narrower than about 0.35 Hz: at an Fmax of 1000 Hz that is 3200 lines and three seconds of data. More lines cost time at every point and bytes in the database, and buy two things: neighbours told apart, and a lower noise floor per line.

The portable data collector.

In the mid-1980s the FFT analyser shrank into a battery-powered box that could hold a list of machines and points, and the route was born. Software on a computer keeps the database: every machine, every bearing, the settings for each point and the history behind it. In the morning the day's route is loaded into the collector. The technician walks it with a sensor on a magnet: place it, wait for the bias to settle, take the reading, move on, a minute or so a point, two hundred or so points a day. In the evening the data goes back to the database, which trends it and flags what has changed. Most of the world's machines are still monitored this way.

What a month misses.

A fault that takes a year to develop is caught comfortably by a monthly route. A bearing that goes from a whisper to a roar in ten days is not: the route is at the machine once in that time, if at all. Choose the fault and the interval, and watch where the first reading past the alert falls.

A rack in the control room.

Long before the route, the biggest machines had sensors that never came off. From the 1960s, turbines and compressors on sleeve bearings carried proximity probes (lesson 5) wired to racks that watched shaft position and vibration continuously and could trip the machine; API 670 standardised the arrangement in 1976. Casing accelerometers joined them, and from the 1990s the racks fed condition-monitoring computers that kept spectra and trends around the clock. The monitoring is as good as it gets: every second, every start-up, every transient. It costs a cable run, a conduit and a rack channel per point, so it was justified only where an hour of downtime paid for all three. A plant of two hundred machines might have five of them wired.

MEMS: a proof mass etched in silicon.

A MEMS accelerometer (micro-electro-mechanical system) is a mass a few hundred micrometres across, hung on silicon springs between the plates of a capacitor. Acceleration moves the mass, the capacitance changes, and electronics on the same chip turn the change into a number. It measures steady acceleration as well as vibration, so it reads gravity and knows which way it is tilted, which a piezo sensor cannot. It costs a few dollars, sleeps on microamps and wakes in milliseconds. For a decade its noise and bandwidth kept it out of serious machine work: typical parts of the 2010s rolled off at 2 or 3 kHz and hissed at a hundred micro-g per root hertz or more, which buried early bearing wear. Current parts reach 10 kHz and tens of micro-g per root hertz, close to a general-purpose IEPE, and the gap is still closing.

Analog or digital.

An IEPE sensor is analog. A voltage travels down the cable and the collector's converter turns it into samples: any collector can read any sensor, at whatever Fmax and lines it likes, and the cable is part of the measurement, with its hum, its crackle and its length. A MEMS sensor is digital at the source. The samples are made inside the sensor, at a rate and a resolution the chip fixes, and what travels is numbers, over a bus or a radio, immune to anything the path does to them. The choice moves from the collector to the sensor: you pick the measurement when you pick the part.

The route goes wireless.

The first use of the new sensor was the old job. A battery-powered triaxial sensor on a magnet, placed by hand, talks to a phone or tablet; the phone is the collector, the route is an app, and the readings go to the database as you walk. Three axes come from one placement instead of three, there is no cable to catch on a guard or wear at the connector, and the cost of entry is a sensor and a phone. What does not change is the person on the route and the month between visits. Flip the switch and watch the day go faster.

Online for every machine.

Leave the sensor on the machine, add a gateway within radio range, and the walk disappears. The sensor wakes on a schedule, takes its triaxial reading, sends it to the gateway and goes back to sleep; the gateway forwards it to the server, which trends it and raises the alarms. A reading every hour or every few, on a pump that used to be visited monthly, for the price of a sensor rather than a cable run and a rack channel. The middle of the plant, which the rack never reached and the route reached once a month, is covered. The route continues for what radio does not reach and for the eyes and ears; the rack continues for the machines that must trip in milliseconds. The same picture as before, with the hourly and daily intervals added.

Battery life is a budget.

A wireless sensor spends almost nothing asleep, a little to wake and sample, and most of its energy on the radio: a triaxial waveform is tens of kilobytes, and sending kilobytes costs far more than measuring them. Life is the battery divided by the daily spend, and the daily spend is set by three choices: how often it reads, how long each reading is, and how far the radio has to reach. Hourly readings of a routine length run for years on one cell. Readings every few minutes with long waveforms run for months.

Start from the machine.

The order of decisions is fixed. First, what must be seen: the lowest and highest frequencies that matter, from the machine's speed, its gears and its bearings (lesson 4). That sets Fmax and the sensor's frequency range. Second, how small: a slow machine's accelerations are tiny, so it wants 500 mV/g or a low-noise MEMS; a gearbox's impacts are huge, so it wants 10 mV/g. Third, where it goes: a stud or an adhesive pad for anything above a kilohertz, a magnet only for a route, and the temperature and the area classification. Last, who collects: a rack for the machine that must trip, a wireless sensor for the machine that must not surprise you, the route for the rest and for the human senses. Pick a machine and read the card.

A working plant uses all of it at once: a rack on the turbine, wireless sensors on the pumps and fans that matter, a route for the hundred small machines and for the walk itself, and a technician who still carries a screwdriver.

Back to Monitoring & diagnostics, or to the 101 lessons.