Can frequencies help us predict the future?

From repeating patterns to measurable forecasts

Hello, frequency information can help predict some events. This is possible because a measured system exhibits repeating behaviour or signs of change, rather than because its future is hidden inside a frequency. Changing machine vibration can inform maintenance decisions; daily consumption cycles can help estimate tomorrow’s energy demand. However, frequency alone is not a forecasting tool. Useful predictions require measurements, a model and validation.

Editor’s note: Detecting a condition, recognising an emerging risk and predicting when an event will happen are different tasks. Finding an unusual vibration component does not reveal exactly how many hours remain before a machine stops. The difference between a warning sign and a certain outcome is central to this topic.

What does frequency analysis actually reveal?

A signal is recorded over time, and Fourier analysis examines its components at different frequencies. An FFT is an efficient method of calculating the discrete Fourier transform; it is not a forecasting algorithm by itself. A frequency spectrum can reveal dominant components. Predicting what comes next also requires understanding why those components occur and under what conditions they will persist.

A simple example is x(t) = A sin(2πft + φ), where A is amplitude, f is frequency, φ is phase and t is time. If these parameters remain constant, the next peak can be calculated. Frequency alone does not determine when that peak occurs; phase information is also needed. If the system speeds up, slows down or experiences an external disturbance, the constant-parameter prediction becomes inaccurate.

Industrial use: recognising developing faults

Vibration data from sensors can establish a baseline for healthy machine operation. Later increases in particular frequency bands or new components can indicate changes that deserve investigation. Condition monitoring assesses the current state; predictive maintenance uses that information to plan maintenance. Estimating remaining useful life numerically requires an additional, suitable degradation model and validation.

A shaft rotating at 1500 rpm has a rotational frequency of 1500 / 60 = 25 Hz. This is different from the motor’s electrical supply frequency. Increasing vibration near 25 Hz may offer a clue, but a single peak does not establish a fault diagnosis. Speed, load, installation and other symptoms must be considered together. Spectral peaks can move when speed changes; operating conditions such as low-frequency operation with a drive therefore affect comparisons.

A PLC can provide operating states and alarms, while SCADA can display time-stamped trends. When setting up the data exchange between PLC and SCADA, remember that fast vibration signals require suitable acquisition hardware. Not every PLC analogue input or SCADA logging interval supports high-frequency vibration analysis. Components lost through slow acquisition cannot be recovered afterwards.

Energy demand and repeating time series

Factory consumption may repeat with shifts, days of the week and seasons. Fourier terms can represent these cycles alongside production volume, outdoor temperature and operating schedules. Here, daily or weekly periodicity describes repeated behaviour over time, rather than the carrier frequency of a radio wave. A new shift, production shutdown or equipment change can invalidate the previous pattern.

Weather forecasting and earthquake prediction are different

Doppler radar uses Doppler information in returned signals to measure precipitation motion along the radar beam. This is one part of atmospheric observation; weather forecasts combine observations with physical models and other data. A carrier frequency used in communication should not be confused with the meteorological variables describing the weather. Because small differences in initial conditions can grow, ensemble forecasts explore multiple possible scenarios.

Earthquakes present a different problem. Seismic frequency content helps analyse an event and its ground motion. It does not provide a validated method for reliably identifying a future earthquake’s location, time and magnitude in advance. Long-term hazard assessment and aftershock probabilities differ from specifying an exact date. Earthquake early warning operates after an earthquake has started, attempting to alert a location before strong shaking arrives; it does not predict an earthquake before it begins.

How should medical and financial examples be interpreted?

MRI combines magnetic fields and radiofrequency pulses to obtain tissue signals and construct images. This supplies information for diagnosis and monitoring; a frequency does not directly reveal a person’s future health. A model forecasting disease progression requires separate validation using suitable patient data.

Frequency components can also be investigated in financial time series. However, an apparent historical cycle does not guarantee the next price movement. News, policy changes and participant behaviour can disrupt it. A curve that fits a historical chart and a successful forecast on previously unseen data are different achievements.

Measurement errors can undermine a forecast before modelling begins

When measuring frequency, sampling rate, sensor bandwidth and recording duration matter. For a band-limited signal, the basic sampling requirement is f_s > 2f_max; practical acquisition also requires suitable margin and anti-alias filtering before sampling. Otherwise, high-frequency components can appear as false lower frequencies. A model may then learn an acquisition artefact rather than a real system behaviour.

Windowing and spectral leakage also affect the result. The FFT treats the record as periodic, so a record containing an incomplete number of cycles can spread energy into neighbouring frequencies. Moreover, a single magnitude spectrum does not fully describe the sequence of events over time. Changing processes may require time-frequency analysis or analysis referenced to rotational speed.

What can AI improve, and what does it not guarantee?

Artificial intelligence and machine learning can combine spectral features with load, temperature and historical fault records. However, a machine’s ability to predict the future remains limited by the conditions under which its data and model are valid. More computing power does not automatically create missing information. Claiming that a quantum computer was used is not, by itself, evidence of accuracy.

The relationship between frequency and energy or frequency-dependent quantum phenomena does not mean every measured signal encodes future events. The limitations of the machine and time-series examples here do not require a direct appeal to Heisenberg’s uncertainty principle. Noise, incomplete observations, changing conditions and model error are usually the practical issues to investigate first.

How do we establish whether a forecast is useful?

Fit the model using historical data and test it on later periods excluded from training. Processing steps that leak future information can create misleading results. Compare performance against simple approaches such as “the last value continues” or “the same hour last week repeats”. Report the forecast horizon, error and prediction interval together; uncertainty commonly increases further into the future.

When assessing a frequency-based forecast, ask: What was measured, which conditions were assumed to persist, how broadly was it tested, and how large were its errors? Clear answers can make frequency analysis useful for maintenance and planning. An impressive spectrum or a claim of certainty alone does not demonstrate forecasting skill.