What signal-level explanation means

A global ranking says which sensors or variables matter on average. It does not show which temporal structure carried a particular output: which window, which variable, which shape in the signal.

A temporal record contains more than a list of variable names. The same variable can have a different role in different windows, and its shape over time can matter to the model. Averaging influence across records answers a useful overall question but can lose the local context needed to examine one output.

This is why a low position in a global ranking should not be treated as proof that an input never matters. An anonymized virtual-sensor programme makes the distinction concrete: structural analysis identified a sensor that was marginal to overall accuracy but carried the model into a specific predictive region. The client's engineering team retained it for safety reasons; a global importance ranking would have rated it low.

The example shows the difference between an average contribution and a role in a particular region. Input reduction stayed an engineering decision, made by the customer with the structure in view.

Xpdeep localizes the time windows, variables and signal morphology that carried the model output. On time-series data, accuracy is at least equivalent to standard deep learning and frequently superior. As far as Xpdeep knows, no one else exposes attention at signal resolution.

Time windows locate when the relevant structure occurs. Variables locate which inputs carry it. Signal morphology concerns the shape in those portions of the signal. Consider these together: a variable name without its window, or a window without the signal being examined, leaves part of the temporal explanation out.

The signal figure above is illustrative. The solid and dotted traces distinguish two variables, and the dashed frame marks a window to examine. It is not customer data or a measured result.

Prediction and structural explanation share a forward pass; the explanation localizes what carried the model output. It describes a dependency in the model, not a physical law.

Model debugging, artefact localization, measurement correction and fault-attribution workflows can start from where the output was carried. The explanation describes the model's learned structure; on its own it does not establish causality or identify a root cause.

One capability shows how far this can go: a model error can be traced to a recoverable distortion in an input, and a signal correction prescribed through counterfactual inference. This describes what the technology can do, not a new customer result.

For the full output, read The integrated decision object. For the comparison with post-hoc methods, read Post-hoc explanation vs structural model control.

Begin with a specific output and the reason it needs review. Locate the attributed variables and windows in the corresponding input record. Examine the signal morphology there alongside the available measurement and operating context. This keeps the review tied to the output being investigated rather than substituting an average ranking from a different question.

For a measurement concern, the localization can focus a check of the recorded data and the measurement process. For an unexpected model output, it can focus examination of the learned behaviour and the assessment data. For a possible fault, it can identify portions of the input worth comparing with independent engineering evidence. These are uses of localization to guide investigation, not fault diagnoses supplied by explanation alone.

Keep observations and hypotheses separate. A relevant signal shape may be a genuine operating feature, a measurement issue or something whose meaning remains unresolved. Structural explanation identifies its role in the model output; the engineering team needs other evidence to assess its meaning in the system.

An investigation may lead to a measurement correction, a review of the model or no change. If targeted structural editing or controlled correction is required, Native supplies those capabilities; Retrofit reads a frozen model structurally and does not edit, optimize or correct it. Correction is targeted and validated between operating periods, not continuous learning.

For a supported intervention, prescription is a separate inference returning a minimal feasible intervention with structural justification. It is not a global optimum or a causal conclusion inferred from localization. Keep that technical output distinct from the customer's decision about whether and how to act.

A practical review ends by stating which variables and windows carried the output, what the signal examination established, what still needs investigation and which customer assessment is required before a change is used. That is more useful than simply announcing that a sensor is important, and more disciplined than treating explanation as a root-cause verdict.