Given the current emphasis on sustainability, there is a growing interest in monitoring various facets of building resource consumption. Building monitoring and automation systems most commonly exist as closed-loop systems for security, fire safety, water, electrical, and HVAC (Heating, Ventilation, and Air-Conditioning). One of the fundamental requirements of a mature engineering process model is the comparison of planned versus actual results for cost, risk, and quality control. For facility engineering, this implies that there must be a structured specification of building space requirements and a mechanism for detecting divergent system and occupant behaviors. Such a mechanism would have broad applicability for commissioning, energy efficiency, sustainability, diagnostics, maintenance, and a variety of other problems.

Organization of the different components of the Processing Workflow where various noise and variations are filtered, and data is categorized into patterns. The blocks in the upper portion operate on intensity noise while the blocks in the lower portion operate on frequency and intensity noise.
This work focused on lifecycle building control through a model of data exchanges across the entire facility lifecycle, and a computational approach for comparing expected and actual resource utilization. Application of telemetry noise filtering and a clustering algorithm can provide accurate results for comparing planned and actual daily resource utilization in the context of human-interpretable schedules.

Experiments were conducted to test this hypothesis on an algorithm using simulated and real data, and ultimately to determine the algorithm’s expected accuracy, sensitivity to noise, and its general applicability. While there is extensive research in all aspects of data mining, this work focuses on the topics of noise reduction and clustering (relevant to multiple data mining phases).

Interpolation (e.g., cubic splines) is a noise reduction method for supplying missing data in the data set. Essentially, such an interpolation technique is used to smooth out jumps in the original data set. Discrete Fourier Transformation (DFT) is used to transform a time-domain input signal into its frequency-domain components. Using this transformation, one can detect and eliminate unwanted frequency components.

Clustering is used throughout the data mining process to detect similarities between data points. In clustering, a score function measures similarity between two data points. Fuzzy clustering is a portioning method whereby a single data point may belong to more than one cluster. Each data point is assigned a set of membership level values. These values indicate the strength of the membership of the data point with each of the clusters.

A lifecycle information exchange was developed for the entire life of a facility. This lifecycle information exchange model is based on the Industry Foundation Class (IFC) and the related Construction Operators Building information exchange (COBie) international building information model standards. Given the lifecycle information exchange integration between facility requirements and sensor data, it is possible to compare expected and actual utilization of facility resources. To this end, an algorithm that categorizes daily utilization data and may be compared to expected resource utilization was developed and evaluated that schedules at an appropriate resolution. Algorithm design decisions were based on considerations about resource utilization data and a general representation of expected/actual data.

An initial set of reference patterns was created that was used during development testing and in later experiments. When compared with actual resource waveforms, these patterns represent the usage characteristics of long-duration-use electrical devices such as a lighting system, desktop computers, or exhaust fans. Typically, these devices stay operational (on) for long intervals during facility occupancy and remain inoperative (off) otherwise.

The approach implements the first two phases of data mining (data dimension/noise reduction and data set partitioning) to extract patterns in observed telemetry comparable to the reference patterns. The intensity noise reduction algorithm consists of three steps: (1) Fast Fourier Transformation (FFT), (2) Spectral Subtraction, and (3) Inverse Fast Fourier Transformation. The Fast Fourier Transformation is a popular algorithm that implements discrete Fourier transformation.

Results from the experiments reveal that accuracies of 90% and greater may be achieved through various combinations of noise where frequency noise is between 0-30%, intensity noise is 0-40%, and shift noise is 0-5%. Given a minimum to moderate amount of noise, the accuracy typically increases as the cluster threshold increases. However, this is not true for shift noise.

The approximation of accuracy between artificial and real data provides a prediction of noise levels in the real data as well as an approximation of expected accuracy. While the algorithm was evaluated for a multitude of artificial noise conditions, the artificial noise seems unrealistically disruptive when compared to the relatively clean recurring patterns observed in the real sensor data. The proposed approach is beneficial because it requires low resolution, unit-neutral data that is not likely to place additional constraints on the sampling programs of installed building automation and monitoring systems – reprogramming of data logging equipment may be cost-prohibitive in some circumstances.

This work was done by A. Christopher Bogen, Mahbubur Rashid, E. William East, and James Ross of the Army Corps of Engineers. ARL-0155

This Brief includes a Technical Support Package (TSP).
Evaluating a Data Clustering Approach for Lifecycle Facility Control

(reference ARL-0155) is currently available for download from the TSP library.

Don't have an account? Sign up here.