The disclosure relates to methods and apparatus for electrical signal analysis and appliance monitoring.
For an electrical load such as a complex appliance, a detected electrical signal relating to that electrical load-such as the current to the load or the voltage across it with respect to time as power is switched on-will itself have a complicated form. This form will be representative of internal state changes within the complex appliance or other electrical load.
It would be desirable to use the form of such electrical signals to represent and analyse changes in internal state within such loads. A technique that proved effective to do this would have numerous uses, for example in appliance monitoring-a change in the form of the electrical signal may be used to determine whether a component within the complex appliance was failing, or if it was working outside its normal range of operation.
In a first aspect, the disclosure provides a method of detecting states of an electrical load from an electrical signal, comprising: providing a mathematical model for the electrical signal; using a sliding window to estimate parameters of the model, wherein a plurality of windows are determined for the electrical signal, a window function is applied for each of the windows, and parameters of the model are determined by interpolation; reconstructing the waveform from the determined parameters and subtracting the reconstructed waveform from the original signal to obtain a residual signal; determining a state transition where the residual signal exceeds a threshold; and detecting a state as existing for the time period between successive state transitions.
Using this approach, the electrical signal can be broken down logically and repeatedly into a sequence of states. Variation in this sequence of states—for example, absence of a particular state, or a variation in the time taken for specific states to appear or finish—can be used to monitor an electrical device or an electrical system for changes, and may even be used for fault diagnosis, particularly if specific states can be identified with specific real world states or criteria.
This mathematical model may be a sinusoidal model comprising a fundamental frequency and harmonics of the fundamental frequency. This is a particularly suitable approach to take in respect of an AC electrical signal.
In embodiments, the window function is a Rife-Vincent window function—this choice of window function is found to be particularly appropriate to this approach to analysis of the electrical signal. This may be a Class 1 Rife-Vincent Window function, and this Class 1 Rife-Vincent Window function is of order 6 or greater.
In embodiments, wherein the residual signal exceeding a threshold may involve determining whether the root mean square value of the residual signal exceeds the threshold value.
In embodiments, the method may further comprise sampling the electrical signal into a buffer, and applying the sliding window to the buffer. The buffer may then be emptied when a state transition is detected.
In a second aspect, the disclosure provides a method of monitoring an electrical system, comprising: detecting states of an electrical load from an electrical signal according to the method of the first aspect above; and comparing the states of the electrical load with expected states of the electrical load.
In embodiments, comparing the states of the electrical load with expected states of the electrical load may comprise determining the presence or absence of states of the electrical load. Comparing the states of the electrical load with expected states of the electrical load may also comprise determining times at which one or more states of the electrical load are present.
In embodiments, this method may further comprise determination of a fault in the electrical system from comparing the states of the electrical load with expected states of the electrical load.
In a third aspect, the disclosure provides an electrical device comprising a detection element for detecting values of an electrical signal passing through the electrical device, wherein the electrical device is adapted for use in the method of either the first or the second aspect above.
Such an electrical device may be an electrical appliance, and wherein detected states are states of the electrical appliance. Alternatively, the electrical device may be a circuit breaker.
Embodiments of the disclosure will now be described, by way of example, with reference to the following figures, in which:
As indicated above, it would be desirable to use a detected electrical signal relating to an electrical load to analyse internal state changes relating to the load.
Building 1 has a plurality of electrical circuits 2 powered from an electrical supply 3. These electrical circuits 2 will typically contain control elements such as circuit breakers 12—such circuit breakers would be a particularly logical point to contain a detection element 10. This is not the only possible point in the system where this approach could be used, however—it could be employed elsewhere, or in multiple locations through the building 1 or in the wider electrical system containing the building. Individual loads within the building may have their own detection element 10—for example appliances such as refrigeration unit 14 or furnace 15 may have their own detection element 10. These detection elements 10 may be used particularly for detection and analysis at the turning on of a load—particularly the turning on of a motor, or a set of motors. Such detection may be used in a local substation 16 supplying electrical power to a group of buildings, or even in a larger substation 17 supplying a plurality of local substations 16.
In embodiments, detection and analysis may both take place within the detecting element—for example, within a “smart” circuit breaker 12—or the analysis may take place within a computing system connected to the detection element 10 (such as a home computing hub 13) or remotely (for example, through a cloud service).
The general approach employed in embodiments of the disclosure involves automated segmentation of the electrical signal detected by a detection element. Typically, the input to be assessed from the detection element is a digitally acquired time-series representation of an alternating current and voltage source. The output of the segmentation process may then be a set of segments extracted from the time-series corresponding to data acquired during the different states of a connected load. This is achieved by analyzing the residual signal produced by a model of the electrical signal for abrupt changes or state transitions.
The data between transitions are considered as discrete segments (referred to below as “segments” or “data segments”. The benefit of this electrical signal segmentation approach is that the data segments correspond to individual states of a load such as a complex appliance, enabling these individual states to be individually analyzed and tracked. This enables more fine-individually condition monitoring of a complex appliance or a system of loads.
Segmentation is a known data analysis approach, used for images in particular but also for other types of signal. Segmentation algorithms aim to simplify the analysis of signals or images by partitioning an image or signal into regions, which can then be analyzed separately. It is desirable that the resulting partition of an image/signal be meaningful, in the sense that each such partition captures data that shares some characteristic or encompasses some important feature (this may be an object in the case of images, or an event in the case of time-series).
Here, the intention is to partition an AC electrical signal into segments corresponding to the various states of the load(s) drawing power on the circuit under observation. States are defined as the portion of waveform data between two adjacent abrupt changes in the electrical signal.
First of all, a mathematical model for the electrical signal is provided 310. In the case of an AC waveform, this will typically be a sinusoidal model involving a fundamental frequency and a plurality of harmonics.
A sinusoidal model for use with an AC waveform can be formulated mathematically as follows,
where i is the current signal, n is the sample index, J is the number of harmonics (including the fundamental), j is the index for J, and cj, fj and θi correspond to the amplitude, frequency and phase of the Jth harmonic (including the fundamental); these are the parameters of the model.
The next step is to estimate 320 values for the parameters of the model from the detected electrical signal. This can be achieved by using a sliding window across the detected signal data-a window function is applied 322 to the window, and parameters of the model are determined by interpolation 324.
The parameters of the model may for example be estimated online in a sliding windowed manner using an Interpolated Windowed Fast Fourier Transform (IWFFT), which applies a Rife-Vincent window function to each window of data to reduce leakage of spectral components in the frequency domain, and which then identifies peaks in the frequency domain and interpolates between these peaks using derived expressions. Practical application of this approach is described in more detail below.
After this, the waveform is reconstructed 330 from the determined parameters. The difference between a reconstruction of the signal via the model and the original signal, known as the residual, can be subjected to a threshold to detect events in the signal—this approach can be used here to detect state transition events. The reconstructed waveform is subtracted from the original signal to obtain 340 a residual signal. From this residual signal, state transitions are detected 350—a state transition is identified when the residual signal exceeds a threshold value. States are then identified 360 as existing between successive state transitions. These states can then be used in a separate analysis process 370.
The electrical signal is sampled regularly at the detection element, and samples are buffered 410 up to a predetermined window size. A FIFO buffer is used, and as will be described below the same buffer is used for a whole segment of signal, representing a state. In the example shown here, sampling is carried out at 1 KHz and 1024 samples are used, giving a window size of approximately a second.
A window function is then applied 420 to the buffered data to prevent leakage-a generalized cosine window can be used here. These have the following form:
A particularly effective choice is found to be the use of a Rife-Vincent window—these minimize the high-order sidelobe amplitude—at order 1 (k=1), this is functionally equivalent to the widely used Hann window—here a higher order window (order 6 in the embodiment considered here) is used.
A Fast Fourier Transform (FFT) is then applied 430 to the resulting window of data to obtain a complex spectrum of the data. From this spectrum, there are calculated 440 frequency, amplitude and phase of the fundamental frequency, and also of all harmonics up to the Nyquist frequency (half of the sampling rate)—interpolation is used between the harmonics if these contain significant energy. Specifically, the bins with the largest amplitude and second largest amplitude in the region around the fundamental frequency are found 442—this is repeated for all harmonics. Frequency is then estimated 444 (window-by-window) using an appropriate interpolation expression derived for frequency for the specified window function (in this case, for a Class I, Order 6, Rife-Vincent window). Suitable expressions can be found in Jalen S̆tremfelj and Dusan Agrez, “Nonparametric Estimation of Power Quantities in the Frequency Domain Using Rife-Vincent Windows”, IEEE Transactions on Instrumentation and Measurement 62(8):2171-2184 August 2013. The same approach is then taken to estimating amplitude 446 and to estimating phase 448.
In this way, a set of estimated parameters for the waveform are built up. The waveform can then be reconstructed 450 according to the sinusoidal model indicated above with these estimated parameters. The reconstructed signal can then be subtracted 460 from the original window of waveform data—this produces a residual signal for the window. This residual signal is used to determine whether there is an event—particularly an event representing a state transition—during the window. If there is no event, the model should provide an accurate representation and the residual signal will be very low, but if there is an event, the model will no longer be accurate and the residual signal will be significantly larger. Consequently, a root-mean-squared value is determined 470 for the residual signal for the window, and this is compared 480 with a predetermined threshold value. If the threshold value is exceeded, indicating significant energy in the residual signal, an event is detected 490. If no event is detected, then the process goes back to the start and a new window is considered. If an event is detected, as well as the event being recorded, the segment buffer is flushed 495—the process will then continue from the start for a new window but also for a new segment.
The resulting pattern of state segments can be used in a number of different ways. Where the load is a complex appliance, the segments can be used for diagnostic purposes, as they will typically represent physical state within the appliance. The non-appearance of a particular state may indicate a component malfunction, or changes in the time before a particular state appears may indicate degradation of a component—the segments may thus be used to assist in a repair process, or to indicate that servicing is necessary. Where the load is actually an aggregate of different loads, specific states may indicate particular loads or combinations of loads being active.
The skilled person will appreciate that many further embodiments are possible within the spirit and scope of the disclosure set out here.
Filing Document | Filing Date | Country | Kind |
---|---|---|---|
PCT/EP2021/069264 | 7/12/2021 | WO |