Remaining useful life

Remaining useful life is the future time, cycle count or other usage until a component reaches a specified end-of-life criterion, conditional on its current condition and assumed future operation. The endpoint must be stated: crossing a performance limit, exhausting a discharge and reaching a long-term degradation threshold are different events. For an increasing damage variable $D$, threshold $D_c$ and present time $t$, a mathematical definition is the first-passage timeA prediction of this quantity supports condition-based maintenance, while requiring more than detection of existing damage.

For an illustrative deterministic model, let current normalized damage be0.60, the end threshold1.00 and future damage rate0.02 per hour. Linear extrapolation gives $R=(1-0.60)/0.02=20\mathrm{h}$. If the plausible future rate instead ranges from0.015 to0.025 per hour, the corresponding times range from about26.67 down to16h. Remaining useful life uncertainty therefore includes future operating conditions and degradation-model uncertainty, as well as measurement error.

A primary NASA battery-aging dataset makes the endpoint concrete. A custom testbed cycled electric batteries under controlled load and temperature, recording quantities including electric potential difference, electric current and temperature. The reported aging endpoint was30% loss of rated capacity, from2Ah to1.4Ah. That long-term endpoint is distinct from the terminal-potential threshold used to end an individual discharge.^[ref]

A useful validation compares predictions made before the endpoint with subsequently observed crossing times under documented operation. Reporting one predicted number without its endpoint, usage assumptions or uncertainty can conceal the main sources of disagreement between a model and a real endurance test.

An end-of-life criterion specifies the condition at which a component is treated as having exhausted its useful operation for a stated purpose. It supplies the endpoint needed to define remaining useful life. The criterion may be a performance threshold or an observed failure condition; useful operation can end before complete physical destruction. Its units, measurement procedure and operating conditions belong in the specification.

A primary NASA battery-aging test used30% loss of rated capacity, from2Ah to1.4Ah, as the aging endpoint. The apparatus separately used terminal-potential limits to finish individual discharges. Those two endpoints describe different processes and timescales.^[ref] A prediction stated only as “time until the battery ends” would therefore omit an essential physical definition.

For an illustrative linear capacity model, suppose present capacity is1.8Ah and future fade is0.01Ah per cycle. A1.4Ah end criterion gives40 remaining cycles, whereas a1.2Ah criterion gives60. The assumed degradation path is unchanged; the endpoint definition changes the reported lifetime. The example thresholds and rate are hypothetical, rather than a recommendation or a fit to the NASA data.

Competing end-of-life criteria also need a logical rule. If operation ends when either limit is first reached, the lifetime is $R=\min (R_1,R_2)$. Threshold crossings at12 and20h then give12h. A rule requiring both limits to be exceeded would define another event. Recording the criterion before evaluating a prediction avoids changing the target after observing the result, and supports an interpretable remaining-life validation error.

Competing end-of-life criteria are two or more conditions, any one of which can terminate useful operation under the stated specification. They define a remaining useful life by the earliest relevant crossing. If the prospective crossing times, measured from the present, are $T_1,\dots ,T_n$, thenThis is the event structure studied by competing risks. A condition ending one operating cycle must not inadvertently be treated as the component's aging endpoint: NASA's battery experiment separately defines a terminal-potential discharge cutoff and a30% rated-capacity-loss aging criterion.^[ref]

For an original mechanical example, suppose a component can reach either an unacceptable vibration level in12h or an unacceptable temperature in20h under a specified operating scenario. If either condition stops operation, its predicted useful life is12h. Requiring both conditions would instead define another stopping rule. An observed early crossing also does not establish when the other criterion would have been reached if the test had continued.

For uncertain crossing times, the survival function of useful life isOnly under independence does this equal the product of the individual survival functions. For an explicitly hypothetical independent pair of exponential distributions with rates $0.01{\mathrm{h}}^{-1}$ and $0.02{\mathrm{h}}^{-1}$, the product is $e^{-0.03t}$ when $t$ is measured in hours. The expected first crossing is $1/0.03\approx 33.33\mathrm{h}$, smaller than either individual mean,100h or50h.

The independence assumption is substantive. Shared temperature, load or damage can correlate crossing times, so multiplying marginal survival functions can give the wrong answer. Moreover, an intervention triggered by the first crossing can change the subsequent trajectory. The physical end-of-life criterion, dependence model and assumed future operation therefore belong with any numerical lifetime prediction.

Remaining useful life uncertainty describes uncertainty about the future crossing time of a stated end-of-life criterion. It can arise from current-condition measurements, degradation parameters, model inadequacy and future operation. The same cycle count need not imply the same condition. NASA's battery-aging dataset explicitly documents variation associated with discharge depth, rest periods and intrinsic differences, and distinguishes discharge-time prediction from long-term aging prediction.^[ref]

For an original two-scenario example, let normalized current damage be0.60 and the threshold1.00. Suppose future damage rate is0.01 per hour with probability0.25 and0.02 per hour with probability0.75. The corresponding remaining useful life values are40 and20h. Their expected lifetime isThe mean rate is0.0175 per hour, but extrapolating with that rate gives $0.40/0.0175\approx 22.86\mathrm{h}$. Averaging the inputs before a nonlinear inverse calculation therefore gives a different answer from averaging the predicted lifetimes.

A forecast can report a probability distribution or a prediction interval together with the assumed future-use scenarios. A narrow interval for the estimated mean is not automatically a narrow interval for a future component's crossing time. The reported uncertainty must identify the random quantity it covers and the model that supplies its probabilities.

Remaining-life validation error compares a forecast made at a specified prediction time with the endpoint subsequently observed. Testing predictions on independent endurance records helps assess whether stated intervals cover observed outcomes at the claimed frequency. Such evaluation must preserve the original endpoint and future-use assumptions; otherwise a numerical disagreement can mix model error with a changed prediction question.

A remaining-life validation error compares a forecast of remaining useful life made at a specified time with the subsequently observed time to the same end-of-life criterion. If the criterion is reached at $t_f$ and a prediction is issued at $t<t_f$, the reference life is $R(t)=t_f-t$. Define the signed error asWith this convention, positive error means predicting more useful life than was observed. A forecast must retain its issue time and endpoint definition for this difference to have a physical meaning.

For an original cycle-count example, suppose the specified endpoint is observed at cycle100. At cycle60, a forecast of50 remaining cycles has reference life40 and error $+10$. At cycle90, a forecast of20 has reference life10 and the same error $+10$. Their absolute relative errors are25% and100%, respectively. Equal cycle errors therefore need not have equal significance near the endpoint. A relative error divided by the reference life is undefined at the endpoint itself.

Errors should also be distinguished from aggregate scores. Two forecasts with errors $+10$ and $-10$ have zero mean signed error, but mean absolute error10 and mean squared error100 in squared cycle units. Cancellation can hide large individual discrepancies. Reporting errors by prediction time preserves information that a single average discards.

NASA's controlled battery tests supply a concrete physical setting: measured current, terminal potential and temperature accompany charge/discharge records, with a defined capacity-loss aging endpoint. The records can support testing forecasts against later observations while retaining the endpoint definition.^[ref] Their different operating histories also make indiscriminate pooling at equal cycle count inappropriate.

The remaining-life validation cutoff separates information available when issuing the forecast from information revealed later for scoring it. Using the eventual crossing time to choose the forecast, or tuning a model on the same future record used to claim validation, changes the experiment. For a prediction interval, one can additionally count how often the subsequently observed life lies inside the interval, while retaining its stated probability level and assumptions about future operation.

The remaining-life validation cutoff is the information boundary used when reconstructing a forecast of remaining useful life at a particular issue time. Observations available by that time can enter the forecast; subsequent observations can establish the eventual end-of-life criterion crossing for scoring it. Keeping these roles separate prevents data leakage from turning a retrospective calculation into an apparently successful prediction.

For an original example, suppose a test reaches its specified endpoint at cycle100 and a forecast is evaluated as if issued at cycle60. The reference remaining life is40 cycles, known only after observing the later endpoint. Constructing a predictor that returns $100-t$ because the complete test record already reveals100 produces zero remaining-life validation error by construction. That is not evidence that the endpoint was predictable from the first60 cycles.

Preprocessing has the same information constraint. A trailing three-sample average $z_t=(x_{t-2}+x_{t-1}+x_t)/3$ uses observations available at $t$. A centered average $z_t^{\prime }=(x_{t-1}+x_t+x_{t+1})/3$ needs a future observation. The centered average can be legitimate with a declared one-sample reporting delay, but then its forecast issue time is $t+1$, not $t$. Parameter fitting, normalization and selection among candidate models also belong inside the information boundary.

NASA's battery-aging records provide a physical example of the required distinction: they contain measured discharge current, terminal potential and temperature, together with a capacity-loss aging endpoint and run-to-failure histories.^[ref] A validation experiment can reveal only the prefix through its chosen cutoff to a prediction method, save the forecast, and subsequently compare it with the endpoint in the withheld suffix.

A fixed model trained on other complete histories can be evaluated on a new component's prefix. A method that updates online must instead document which new observations have become available before each forecast. In either case, preserving the cutoff, endpoint definition and assumptions about future operation makes the resulting remaining-life validation error interpretable. Pooling all prediction times into one score can conceal a violation at a particular cutoff.