Access Official Technical PDF

Share Summary:LinkedInEmail

Time series anomaly detection is an important part of preparing industrial data for analysis. Automated instruments produce continuous measurements such as temperature, wind speed, solar radiation, water levels, and other process variables. Errors in measurement or transmission, as well as system changes, can introduce unusual observations that distort model estimates and contribute to forecasting errors. Before engineers use these data for monitoring, forecasting, clustering, or other analysis, they need a practical way to identify irregularities.

Climate time series illustrate why this task is not always straightforward. Data may have multiple seasonal patterns, serial dependence, and many measurement locations. An observation that looks unusual in isolation may need to be understood in the context of the signal around it. Two approaches developed for climate data offer different ways to address the problem: autoregressive cost updates (ACU) and wavelet-based anomaly detection.

Why time series anomaly detection matters

Data preprocessing can take substantial effort in an analysis workflow. If irregular observations are left unexamined, they can affect model specification and parameter estimates, and ultimately weaken forecasts. The challenge is not limited to obvious spikes. Some problems appear as patterns that depart from expected behavior without presenting as a single, unmistakable outlier.

For industrial automation teams, this has a direct practical implication: analysis quality depends partly on the quality checks applied to historical and incoming measurements. A preprocessing method should be selected with the data’s characteristics and the likely form of disruption in mind. No single approach performs best across every scenario and evaluation criterion.

Two approaches for detecting anomalies in time series

Autoregressive cost updates for isolated outliers

The ACU approach is presented as a relatively simple method for anomaly detection. It is designed to be straightforward to implement and explain, and it detects anomalies without requiring model calibration. These qualities may be useful when teams need a method that is easier to communicate as part of a data-quality workflow.

In tests using synthetic time series, ACU was recommended when a series was expected to contain mostly solitary outliers. This makes the approach a candidate for data where irregularities are isolated rather than sustained across several consecutive observations. The source describes its intended use and comparative results but does not prescribe a specific deployment configuration or threshold-setting procedure.

Wavelet analysis for consecutive irregularities

The second approach uses a wavelet transform, a signal-processing technique that represents a time series as a combination of different frequencies. The method examines the highest-frequency components to identify irregularities. In the reported comparison, the wavelet-based approach was the preferred alternative when multiple consecutive outliers were present.

This distinction matters because a run of abnormal observations can behave differently from a single isolated point. A method suited to one unusual reading may not be the best choice for a sustained irregular pattern. Wavelet analysis also has potential beyond temperature data: the approach is described as adaptable to other climate series, including humidity, precipitation, and wind speed, as well as time series from other application areas.

Match the method to the data problem

The reported evaluation tested the proposed methods on synthetic time series and compared them with existing preprocessing algorithms. Its central practical finding is not that one algorithm dominates, but that performance depends on the scenario and the quality criteria used. The recommendations can be summarized as follows:

  • Mostly solitary outliers: consider ACU as a comparatively simple detection approach.
  • Multiple consecutive outliers: consider the wavelet-based method among the alternatives tested.
  • Unclear irregular patterns: avoid relying only on obvious spikes; examine whether the series contains less apparent departures from its expected pattern.
  • Different data types: assess whether the chosen method is appropriate for the signal and application rather than assuming that a result for temperature automatically applies unchanged elsewhere.

These are selection guidelines, not a substitute for validation on the data being analyzed. The comparison does not identify a universal best method, and teams should treat anomaly detection as part of a broader preprocessing and data-quality process.

Applying the approaches to larger data sets

Both algorithms were also applied to a multivariate temperature data set collected at multiple sites in Novi Sad, Serbia. The application focused in particular on irregular patterns that were not clear-cut outliers and could therefore be difficult to identify. The authors reported that both approaches handled large data sets effectively, supporting their use as a basis for subsequent analysis such as clustering or modeling.

For plant and systems teams, the relevance is the workflow: detection can help prepare measurements before downstream methods are applied. In a multi-sensor environment, preprocessing may be especially important because irregularities can complicate comparisons across locations or variables. The reported application concerns temperature measurements; it should not be read as a performance guarantee for every industrial data set or operating condition.

Practical implications for automation teams

When designing a time-series preprocessing workflow, engineers and integrators can begin by characterizing the suspected data-quality issue. Is the concern a small number of isolated readings, or are irregular values likely to continue over several time steps? That distinction can guide whether ACU or wavelet analysis is a more suitable candidate for evaluation.

It is also important to consider the broader purpose of the data. Measurements may feed forecasting, clustering, or other models, and anomaly handling takes place before those methods are applied. Documenting the chosen approach and assessing its behavior on representative data can make the preprocessing decision clearer to analysts and operations stakeholders.

Conclusion

Reliable analysis of sensor time series begins with attention to data quality. Autoregressive cost updates and wavelet-based detection provide two different options: ACU is recommended for series with mostly solitary outliers, while the wavelet approach is favored for multiple consecutive outliers in the reported tests. Because no method performed best in every scenario, industrial teams should match detection techniques to the irregularities they expect and validate preprocessing before using cleaned data for forecasting or further analysis.

Access Official Technical PDF

About The Author