KutupASP DijitalEach page answers one question: the short answer first, then a worked example and the common mistakes.
A 32-bit float spans two 16-bit Modbus registers. Devices may order those registers and their bytes differently; common orders are ABCD, CDAB, BADC and DCBA. Test each order against a known value, such as 25.0 (0x41C80000), and confirm the chosen order with a second known value.
Modbus RTU runs over a serial line, usually RS-485, and ends each frame with a CRC-16. Modbus TCP carries the same register model and function codes over Ethernet, normally on port 502, with an MBAP header instead of a CRC. The data meaning is identical; a gateway bridges the two when needed.
MQTT is a lightweight publish/subscribe transport that routes messages through a broker, with free-form topic names and payloads. Sparkplug B is a specification built on top of MQTT. It defines a topic namespace, a protobuf payload, and birth and death certificates that track node state. Plain MQTT defines no data model.
OPC DA (Classic) is built on Microsoft COM and DCOM and is tied to Windows. OPC UA is platform-independent, uses its own binary TCP or HTTPS transport, and has security built in: certificates, signing and encryption. It also offers typed information models, subscriptions, method calls and historical access.
Keep the raw file first, then check the delimiter, decimal separator, encoding, header, timestamps, time zone and units. Mark sentinel codes such as -999 as missing, remove only exact duplicates, review outliers before touching them, and log every change so the cleaning can be audited.
Design a production dashboard for one decision and one audience. Show three to five metrics, each with its target and recent trend. State units, time range and data freshness clearly. If a value is missing, show it as missing, not as zero.
A data historian is a time-series store for process data. Each record holds a tag, a value, a timestamp and a quality code. It is built for high write rates and long retention, and it often compresses data, so its settings limit what a later analysis can see.
Running a model close to the machine makes sense when the application needs low latency, must keep working during connection loss, handles large raw data volumes or privacy limits, or needs predictable timing in a control loop. Model size, update management and monitoring pull the other way. The choice should follow those requirements, not a trend.
WebAssembly is a portable binary format. Code from C, C++ or Rust compiles to it and runs in a sandbox at near-native speed, in browsers and other runtimes. For analytics, data can stay local and nothing needs installing. Limits: a memory cap, no direct hardware access, and a single thread by default.
The scaling formula is PV = LRV + (I − 4) / 16 × (URV − LRV). I is the measured current in mA, LRV the lower and URV the upper range value. For percent, use (I − 4) / 16 × 100. The 16 mA span from 4 to 20 mA maps linearly onto the range.
A 4-20 mA signal carries a measured value as current, a common analog standard. The live zero at 4 mA means 0 percent of range, so a broken wire (0 mA) can be told apart from a real minimum reading. Current is also unaffected by voltage drop along the cable.
There is no fixed calibration interval. It depends on how critical the instrument is, its drift history (as-found and as-left records), the manufacturer's advice, process risk and quality requirements. Start from the manufacturer's advice, lengthen the interval only with stable records, and shorten it when results fall outside tolerance.
The FFT (Fast Fourier Transform) is an algorithm that splits a recorded vibration signal into its frequency components. The result is a spectrum: a plot of how much amplitude sits at each frequency. On a machine, this makes peaks at running speed, gear mesh or defect frequencies visible one by one.
Spectral leakage is the spreading of a peak's energy into neighboring frequency bins when the record does not hold a whole number of periods. A window function tapers the record edges, which reduces that spreading. Hann suits general vibration, flat-top suits amplitude accuracy, and the rectangular window suits whole-period or transient records.
ISO 20816 vibration severity zones classify how acceptable a machine's broadband vibration is. Zone A is typical of newly commissioned machines, zone B is acceptable for unrestricted long-term operation, zone C is usually unsatisfactory for continuous operation, and zone D is severe enough to cause damage. The boundaries depend on the machine group.
Crest factor is the peak value divided by the RMS value; for a pure sinusoid it is about 1.41 (√2). Kurtosis measures how peaked the amplitude distribution is: about 3 for Gaussian noise, and higher for impulsive signals. Both can rise in early bearing damage, but neither should be used alone to decide.
The Nyquist rule says that the sampling frequency fs must be more than twice the highest frequency you want to measure. A component above fs/2 folds back and appears at a wrong, lower frequency; this is called aliasing. For that reason, an analog anti-alias filter must sit before the converter.
Bearing defect frequencies follow from the geometry (ball count Nb, ball diameter d, pitch diameter D, contact angle φ) and the shaft frequency fr. For the outer race, BPFO = (Nb/2) × fr × (1 − (d/D) × cos φ). Measured peaks should land near these values, not exactly on them.
Sensor drift is a slow, systematic change in a reading while the true input stays constant. It shows up through redundant sensor comparison, checks against a known state, physical balances and long-term baseline trends. Drift must be separated from real process change, or the conclusion will be wrong.
For a sinusoid, the acceleration amplitude equals 2πf times the velocity amplitude, and velocity equals 2πf times displacement. So acceleration suits high-frequency faults such as bearings and gears, velocity suits overall machine condition, and displacement suits large, low-frequency motion. The right choice depends on the purpose of the measurement.
Envelope analysis reveals amplitude modulation in a high-frequency resonance band. The signal is band-pass filtered, its envelope (the amplitude curve) is extracted with the Hilbert transform, and the FFT of that envelope is computed. Bearing defects then show up as peaks at their defect frequencies in the envelope spectrum.
Predictive maintenance uses the measured condition of equipment to forecast a future failure or its remaining life, then schedules the work accordingly. Preventive maintenance follows a calendar or running-hour rule and ignores the actual condition. The key decision is when to act, not which model to pick.
MTBF is the mean time between failures for a repairable system: operating time divided by the number of failures. MTTR is the mean time to repair: total repair time divided by the number of repairs. Availability is A = MTBF / (MTBF + MTTR). MTBF is not a lifetime or a guarantee. Read the three numbers together.
Remaining useful life (RUL) is the time from now until a health indicator (HI) crosses its failure threshold. Estimates use a degradation trend, life data such as Weibull models, or a model learned from run-to-failure histories. A useful RUL is reported as a range, not a single number.
RUL estimates carry uncertainty from sensor noise, from an imperfectly known degradation rate, and from future operating conditions. A single number hides that uncertainty and invites false confidence. So estimates are reported as percentiles such as P10, P50 and P90, and the band widens as the forecast horizon grows.
Beta, the Weibull shape parameter, shows how failure rate changes with age. Below 1 points to early-life failures, near 1 to random failures, and above 1 to wear-out. A scheduled replacement age only makes sense when beta is above 1, because random failures are not reduced by replacing parts on time.
Cp compares the specification width with the process spread: Cp = (USL − LSL) / 6σ. Cpk also accounts for an off-centre mean: Cpk = min((USL − μ) / 3σ, (μ − LSL) / 3σ). Both assume the process is stable and roughly normal, so check that first.
Use a Shewhart chart for large, sudden shifts, where it is simple and sufficient. For small, sustained shifts of roughly 0.5 to 1.5 sigma, CUSUM and EWMA respond sooner. The choice depends on the shift size to catch and the false-alarm rate you accept.
There is no universal good OEE. The often-quoted 85 % figure comes from TPM literature and is a reference point, not a law. The right comparison is your own baseline for the same line, product and shift. Set targets from that baseline, using the same planned-stop definition throughout.
Control limits are calculated from the process's own variation and show whether it is stable. Specification limits come from design or the customer and show whether the output is acceptable. The two are independent: a process can be in control and still out of specification, or the reverse.
Overall equipment effectiveness (OEE) is the share of planned production time that turns into good parts made at the ideal cycle speed. It is the product of three factors: OEE = A × P × Q. A is availability, P is performance and Q is quality, and each is a ratio between 0 and 1.
Alarm flooding occurs when alarms arrive faster than an operator can meaningfully respond. ISA-18.2 and similar guidance treat more than ten alarms in ten minutes per operator as a flood. Main causes are chattering, stale or repeated alarms and cascades. Fix the alarm design first; AI comes after that.
A Kalman filter estimates a hidden state, such as the true level or slope of a signal, from noisy measurements. It works recursively: each step predicts the state with a model, then corrects the prediction with the new measurement. The Kalman gain sets how much the correction trusts the measurement relative to the model.
A PID loop often oscillates when the gain is too high, the integral is faster than the process lag, dead time is large, or valve stiction is present. Examine the oscillation period, controller output and measurement together. Lowering the gain is usually the first safe step.
A PID controller's output is the sum of three terms. P (proportional) responds to the error right away. I (integral) accumulates the error over time and removes the steady offset. D (derivative) responds to how fast the error changes and reduces overshoot. Each term's effect is set by its own gain.
A PLC is a real-time controller that reads inputs, runs logic and drives outputs. SCADA is a supervisory layer for many controllers, with alarms, trends, logging and remote commands. A PLC runs without SCADA, and SCADA does not replace control logic. An HMI is the operator screen for one machine or area.
PLC scan time is the time one cycle takes to read inputs, run the program and write outputs, measured in milliseconds. It determines the controller's response time, its deterministic behaviour and what a data collector can see. The value is usually read from the CPU diagnostics screen.
To turn an SOP into a flowchart, write one action per line and state every decision, loop and parallel step with words such as if, otherwise, repeat until and at the same time. Name the role for each step. Flow Maker accepts text in this form directly. Have an operator verify the chart.
A process flowchart shows the steps and decision points of a task with standard symbols. Start and end use ovals, actions use rectangles, decisions use diamonds, and arrows show the order. Each decision branch should be labelled, and the chart should have one start point and no dead ends.
Anomaly detection learns what normal behaviour looks like in a system’s historical data, then flags observations that deviate clearly from it. On a sensor table it marks rows that look unusual. A flag is not a diagnosis: it says the process has changed, not why. Finding the cause still needs operator knowledge and maintenance records.
Reduce false alarms by putting decision rules after the raw model: require several consecutive out-of-range samples, use hysteresis between entry and exit thresholds, keep separate models per operating mode, and suppress known events such as start-up. Tune these rules on a validation period, not the training data, and check that real events are still caught.
Yes, with a narrow claim. Label-free methods learn normal behaviour and flag departures from it. They report that a pattern is unusual, not which fault it is. Validation means replaying past incidents and reviewing flagged periods with operators. Used this way, the output is a review list for maintenance, not a diagnosis.
Hotelling T2 measures how extreme a point is within the normal variation captured by a principal component analysis (PCA) model plane. SPE, also called Q, measures how far the point lies off that plane, which signals that the correlation structure has broken. The two statistics catch different faults, so both are monitored together.
Isolation Forest uses random trees that split data along single features at random cuts. Rare, unusual points are isolated after fewer splits, giving them a score near 1. Typical points score near 0.5, meaning no clear distinction. The method needs no distance measure or scaling.
There is no fixed row count. The number of recorded failure events and the coverage of operating modes matter more than the number of rows. Start with condition thresholds and anomaly detection. Remaining useful life (RUL) models need run-to-failure histories, along with load, speed and dated maintenance logs.
Precision is the share of alarms that match a real event, and recall is the share of real events that trigger an alarm. Precision = TP/(TP+FP); recall = TP/(TP+FN). Raising the threshold usually raises precision and lowers recall. Choose the threshold where both values, and their combined F1 score, are acceptable for the cost of a missed event.
Split time series by time, never by random shuffling: the training period comes before the test period. Fit scalers, feature windows and other preprocessing only on the training data. For evaluation, use walk-forward splits, so each test window lies in the future relative to the data the model was trained on. Random splits make results look better than they are.
IEC 62443 is a series of standards for the cybersecurity of industrial automation and control systems (IACS). It defines roles for the asset owner, the system integrator and the product supplier. It splits a system into zones and conduits and sets a target security level, from SL 1 to SL 4, for each zone.
IT security usually prioritises confidentiality and data integrity. OT security puts safety and availability first, because an attack or a wrong patch can have physical consequences. OT assets run for many years, patch windows are short, and many protocols lack authentication. Common responses are segmentation, passive monitoring and change management.
The Purdue model splits industrial networks into functional levels: 0 field devices, 1 basic control (PLCs, RTUs), 2 area supervision (HMI, SCADA), 3 site operations (MES, historian). Level 3.5, an industrial DMZ, is often added. Levels 4 and 5 are business and enterprise. The model helps plan network segmentation.
A hallucination is an answer the model produces that has no support in its sources but reads as fluent and certain. To reduce them, ground answers in retrieved documents, require citations, and make the system decline when no evidence is found. Then measure the result on a fixed question set. No method removes errors completely.
RAG (retrieval-augmented generation) means a language model first searches a document store for passages related to the question, then writes its answer from those passages. The retrieved text is added to the prompt alongside the question. When documents change, the index is updated and the model does not need to be retrained.
No results. Try another term or clear the topic filter.
NQ answers from the library’s sources and shows which ones it used.