Is this just ordinary random noise?

No; a persistent fault has temporal structure that affects subsequent observations.

Can the model include repair?

Yes, by specifying recovery transitions or repair events.

Does detecting a faulty sensor reveal the true system state?

Not necessarily; losing a reliable measurement can increase uncertainty.

AI Glossary / Statistics and probability / Intermediate

Persistent failure model

Definition

A persistent failure model represents a fault as a state that continues across time, rather than treating each faulty observation as an unrelated accident. In AI state estimation, it can introduce a hidden sensor-health variable whose present state depends on its previous state. The model then distinguishes changes in the monitored system from changes in the reliability of its measurements.

What it means

A sensor that repeatedly reports zero may indicate an empty resource, a broken sensor, or another condition. A model that assumes every reading is normally reliable may overreact to faulty measurements. A model that treats every fault as independent may also misread a long sequence caused by one continuing problem.
Persistence makes the history relevant. Once evidence supports a continuing fault, later suspicious readings are interpreted in light of that possibility rather than as completely fresh independent failures.

How it works

Represent the underlying system state, such as charge or temperature, separately from the sensor's health state. The health state may be healthy or failed, or it may include several failure modes.
Specify transition probabilities for the physical state and the health state. In a simple permanent-failure model, a healthy sensor has some chance of becoming failed during each interval, and a failed sensor remains failed until an explicitly modeled repair. A more general persistent model can allow recovery while making continued failure likely.
Specify observation probabilities conditional on both states. A healthy sensor might produce noisy readings near the physical value. A failed sensor might be stuck at zero, biased, or uninformative. The exact failure behavior must reflect the device and task.
Use filtering to alternate prediction and updating. Prediction propagates beliefs through the transition model. Updating uses the new observation to revise the joint probabilities of physical condition and sensor health.
Additional sensors, diagnostic checks, known operating conditions, or maintenance records can help separate explanations. Without such information, low readings caused by a genuine state change and low readings caused by failure may be difficult to distinguish.
A dynamic Bayesian network provides one representation: physical-state and health variables occur at successive times, connected by transition dependencies, and readings depend on the relevant current variables. The representation makes assumptions explicit but does not establish that those assumptions are correct.
Calibrate failure rates, noise distributions, and repair behavior from appropriate evidence. Test the model against transient glitches, persistent faults, genuine changes, and combinations of those events. Report uncertainty instead of converting every posterior probability into an unquestioned diagnosis.
If the sensor is believed failed, uncertainty about the physical state may grow as the model relies more heavily on dynamics or other evidence. A fault model should not create artificial certainty about a system it can no longer observe reliably.

Three practical examples

Robotics: a robot tracks battery charge while allowing for a sensor that becomes stuck.
Industrial monitoring: an estimator distinguishes a persistent measurement fault from a genuine process shift.
AI operations: a monitoring model represents an ongoing data-feed fault rather than counting every corrupted update as an independent event.

Think of it this way

Think of a clock that has stopped. Seeing the same time repeatedly is not a series of independent clues that the world has stopped changing. A persistent-failure model allows one continuing clock fault to explain the repeated observations.

Worked example

Consider an invented simplified model with a device known to be operating normally, so only sensor health is uncertain. Before the first reading, the probability the sensor is failed is 0.01. A failed sensor always reads zero; a healthy sensor reads zero with probability 0.05.
After a zero, Bayes' rule gives failure probability 0.01/(0.01 + 0.99 × 0.05) = 20/119, approximately 0.1681.
Before the next reading, failed sensors stay failed and a healthy sensor has a 0.01 chance of failing. The predicted failure probability becomes 20/119 + (99/119) × 0.01 = 2099/11900, approximately 0.1764.
If the next reading is also zero, the updated probability is 0.1763866/[0.1763866 + (1 minus 0.1763866) × 0.05], approximately 0.8107. Exact fraction arithmetic gives 41980/51781.
An independent-failure model resetting the prior to 0.01 each time would return approximately 0.1681 after each isolated zero. The difference illustrates temporal persistence, not evidence that either model fits a real sensor.

Illustrative case study

Fictional case: Elm Robotics receives repeated zero readings from a mobile unit's charge sensor. Its estimator retains separate beliefs about charge and sensor health and uses a second diagnostic signal before recommending maintenance. Engineers later compare the prediction with inspection records, including cases in which the charge really was low, to evaluate the model's ability to distinguish explanations.

Common misconceptions and mistakes

Do not equate persistence with certainty or assume every persistent fault is irreversible. Recovery behavior is a modeling choice.
Do not treat repeated faulty readings as independent evidence without examining their common cause.
A hidden fault variable does not guarantee identifiability. Several state combinations may produce similar observations.
Avoid copying a failure probability from an unrelated device or using a model's confident output as its own validation.

Business use

Persistent failure modeling can improve monitoring, reduce inappropriate reactions to defective signals, and guide maintenance decisions. Business value depends on realistic transition and observation models, useful independent evidence, and decision thresholds that account for the consequences of missed faults and unnecessary interventions.

Check your understanding

Why does the second zero raise the failure probability substantially?

Answer: The model carries forward the possibility of an already failed sensor.

What simplifies the worked example?

Answer: The physical device state is assumed known, so only sensor health is estimated.

Related terms

Dynamic Bayesian network
Hidden Markov model
Filtering
Transient failure
sensor model
fault diagnosis
state estimation
Bayesian updating

Learn next

Learn next how joint state-and-health inference works, how additional observations improve identifiability, and how recovery transitions distinguish permanent from persistent but repairable faults.

Sources

Sources: Russell and Norvig, Artificial Intelligence: A Modern Approach, third-edition index and sensor-failure discussion, https://aima.eecs.berkeley.edu/3rd-ed/aima-index.html and https://people.engr.tamu.edu/guni/csce642/files/AI_Russell_Norvig.pdf . The official AIMA exercise at https://aimacode.github.io/aima-exercises/dbn-exercises/ex_19/ and primary textbook search excerpts identify the battery-sensor persistent-failure model; the full mirrored PDF was not retrievable in this check. The simplified Bayesian calculation and fictional case are original.

Previous: Persistent contrastive divergence Next: Persistent homology group