Arc fault detection in DC power supply

Abstract

The disclosure provides a method of detecting arc faults in a DC power distribution system. The method comprises: sampling, using one or more sensors, a voltage signal on a conductor forming a current path of the power distribution system to generate a voltage signal sample vector; sampling, using the one or more sensors, a current signal in the conductor to generate a current signal sample vector; generating one or more feature vectors based on the voltage signal sample vector and the current signal sample vector and having fewer elements than the current signal sample vector or the voltage signal sample vector; providing the one or more feature vectors to an arc fault classification model; and receiving as an output of the arc fault classification model an indication of whether an arc fault is occurring in the power distribution system.

Description

[0001] Arc Fault Detection in DC Power Supply

[0002] Field

[0003] This disclosure relates to the detection of arc fault currents in a DC power supply. In particular, the disclosure relates to the detection of currents in a high voltage DC power supply.

[0004] Background

[0005] Known methods of arc fault detection rely on the detection of characteristic features of power line signals that occur during arcing. The characteristics of features of signals during arcing are highly dependent on the system in which they occur. For example, arc faults in AC power supplies have very different signal characteristics to arc faults in DC power supplies. Within AC and DC power supplies, the characteristics of power signals at are highly dependent on the voltage at which the arc fault takes place. Arc fault detection in AC systems is often simpler to achieve in comparison to DC systems as the zero-crossing of the voltage and current waveforms allows signal analysis to be confined to a narrow frequency band. As such, known methods of arc fault detection that are effective in certain systems are not usually applicable to other systems. There is a need, therefore, for improved methods of arc fault detection that are applicable to regimes in which reliable methods arc fault detection are not currently known. In particular, there is need for methods of arc fault detection that can be used in DC power supply networks operating at potentials up to 700V and currents greater than 7 A, up to 150 A.

[0006] Summary

[0007] The disclosure provides a method of detecting arc faults in a DC power distribution system. The method comprises: sampling, using one or more sensors, a voltage signal on a conductor forming a current path of the power distribution system to generate a voltage signal sample vector; sampling, using the one or more sensors, a current signal in the conductor to generate a current signal sample vector; generating one or more feature vectors based on the voltage signal sample vector and the current signal sample vector, the one or more feature vectors having fewer elements than the current signal sample vector or the voltage signal sample vector; providing the one or more feature vectors to an arc fault classification model; and receiving as an output of the arc fault classification model an indication of whether an arc fault is occurring in the power distribution system. In some examples, generating the one or feature vectors comprises: transforming the voltage signal sample vector using a first time-frequency transformation to generate a transformed voltage coefficient vector; transforming at least part of the current signal sample vector using a second time-frequency transformation, which may be the same as or different to the first time-frequency transformation, to generate a transformed current coefficients vector; and generating the one or more feature vectors based on the transformed voltage coefficient vector and the transformed current coefficient vector, wherein the one or more feature vectors has fewer elements than the transformed voltage coefficient vector or the transformed current coefficient vector.

[0008] In some examples, the first and second time-frequency transformations are discrete wavelet transforms, DWTs, and, preferably, Haar DWTs.

[0009] In some examples, each element of the feature vector is generated by a method comprising one of: performing a summation of the values of the transformed voltage coefficient vector corresponding to wavelet coefficients of a given DWT scale; or performing a summation of the absolute values of the transformed voltage coefficient vector corresponding to wavelet coefficients of a given DWT scale; performing a summation of the values of the transformed current coefficient vector corresponding to wavelet coefficients of a given DWT scale; or performing a summation of the absolute values of the transformed current coefficient vector corresponding to wavelet coefficients of a given DWT scale.

[0010] In some examples, the elements of the one or more feature vectors are generated by a method comprising: performing a summation of the values of the transformed voltage coefficient vector corresponding to wavelet coefficients for each DWT scale; performing a summation of the absolute values of the transformed voltage coefficient vector corresponding to wavelet coefficients for each given DWT scale; performing a summation of the values of the transformed current coefficient vector corresponding to wavelet coefficients for each DWT scale; and performing a summation of the absolute values of the transformed current coefficient vector corresponding to wavelet coefficients for each DWT scale.

[0011] In some examples, the method further comprises: generating a plurality of further voltage signal sample vectors and current signal sample vectors from measurements of the conductor for different time periods according to a sliding signal window method, and determining whether an arc fault current is occurring in any of the time periods.

[0012] In some examples, the method further comprises, based on detecting that an arc fault has occurred, providing a trigger signal to a switching device to cause a current path in the DC power distribution system to be interrupted.

[0013] In some examples, the trigger signal is provided to the switching based on a moving average of the output of the arc fault classification model over a series of consecutive inputs

[0014] An arc detection device is provided. The arc detection device comprises: a conductive path forming part of a DC power distribution network; one or more sensors configured to measure a voltage and a current in a conductive path; an AD converter configured to sample current measurements and voltage measurements to generate a voltage signal sample vector and a current signal sample vector; and a processor configured to: generate one or more feature vectors based on the voltage signal sample vector and the current signal sample vector, the one or more feature vectors having fewer elements than the current signal sample vector or the voltage signal sample vector; provide the one or more feature vectors to an arc fault classification model; and receive as an output of the arc fault classification model an indication of whether an arc fault is occurring in the power distribution system.

[0015] In some examples, the processor is further configured to: transform the voltage signal sample vector using a first time-frequency transformation to generate a transformed voltage coefficient vector; transform at least part of the current signal sample vector using a second time-frequency transformation, which may be the same as or different to the first time-frequency transformation, to generate a transformed current coefficients vector; and generate the one or more feature vectors based on the transformed voltage coefficient vector the transformed current coefficient vector, wherein the one or more feature vectors has fewer elements than the transformed voltage coefficient vector or the transformed current coefficient vector.

[0016] In some examples, the arc detection device further comprises a switching device configured to interrupt a current path forming part of the DC power distribution network when an arc fault is detected. Brief Description of the Figures

[0017] Fig. 1 is a schematic illustration of stages of an arc fault detection process;

[0018] Fig. 2 is a schematic illustration of a sliding window sampling process in accordance with examples of the disclosure;

[0019] Fig. 3 is an illustration of the relationship between a sampled signal and transformed coefficients in examples of the disclosure;

[0020] Fig. 4 is a schematic illustration of a classification model in accordance with examples of the disclosure;

[0021] Fig. 5 is a flowchart illustrating a method in accordance with examples of the disclosure

[0022] Fig. 6 is a schematic illustration of a system in accordance with examples of the disclosure.

[0023] Detailed Description

[0024] The present disclosure relates to arc fault detection in DC power distribution systems. The methods and devices of the present disclosure allow arc fault detection in DC power distribution systems up to at least 750V and 150A.

[0025] A voltage and current signal on a signal line are sampled over a period of time. The voltage signal and current signal are both separately transformed using a timefrequency transformation. The transformed voltage and current signals are then reduced in dimension via a feature extraction process. The extracted features are provided to a binary classification model trained to classify whether or not the extracted features correspond to arc fault conditions.

[0026] The method of arc fault detection of the present disclosure is particularly advantageous as it does not favour a particular frequency band and may analyse signals of frequency up to 4MHz or greater. The wide frequency band allows features corresponding to a large range of arc fault characteristics to be extracted and classified. As such, the method is applicable to a wide range of use-cases having disparate arc fault signal characteristics. By extracting features from both voltage and current signals for input to the classification model, a robust classification method is provided that is able to correctly identify arc fault characteristics in input signals without overfitting to training data. With reference to Figs. 1 to 5, a method of arc fault detection is described in more detail. The embodiment described in detail below includes specific advantageous processes for transforming samples signals into inputs for a classification model for illustrative purposes. The skilled person would understand that alternatives to these processes can be used in different examples arc fault detection methods within the scope of the disclosure.

[0027] A power line signal is monitored by one or more sensors to measure a voltage signal and a current signal. The measurement of the voltage signal and current signal can be performed at any point of a power distribution system and may, for example, be performed at a gearbox of the power distribution system. An analogue-to-digital converter is used to sample the voltage signal and the current signal to generate a series of discrete measurements. The continuous signal of both the current and source voltage may be filtered by hardware antialiasing filters prior to sampling. The sampled values of the voltage signal may be stored as a voltage signal sample vector (or array), , with each element of the voltage signal sample vector corresponding to a sampled measurement of the voltage signal. The sampled values of the current signal may be stored as a current signal sample vector (or array), 5^ , with each element of the current signal sample vector corresponding to a sampled measurement of the current signal. The treatment of the current measurements and voltage measurements in the present disclosure are substantially the same, and references to a signal in the disclosure can be considered to apply equally to both the current signals and the voltage signal.

[0028] As illustrated in Fig. 2, the current signals and voltage signals are processed in blocks containing a fixed number of sampled measurements. The sampling rate is constant and may be, preferably, at least 8MHz. Other sampling rates can be used depending on the application. Each block comprises the signal sample vector corresponding to a series of samples taking over a fixed time period. Each signal sample vector may be a column vector containing M = 216=65,536 samples, though other block sizes can be used. The signal may be processed using a sliding window, where consecutive blocks overlap each over. For example, when the overlap of two consecutive blocks is M / 2 samples, each of the samples appears in two different blocks. Each block may be separately analysed sequentially in accordance with the present disclosure in order to determine whether an arc fault is occurring within the given block. For both the voltage signal and the current signal, the signal sample vector is processed according to a time-frequency transformation 101 to generate transformed coefficients. Preferably, the time-frequency transformation is a discrete wavelet transformation, DWT, such as a Haar wavelet transformation. Other transformations, such as a Fourier transform may be used in other embodiments. Preferably, both the voltage signal sample vector and the current signal sample vector are transformed according to the same time-frequency transformation, though they may be different to each other in examples.

[0029] The result of the time-frequency transformation of the voltage sample signal vector FJj is to generate a transformed voltage coefficients vector F^, wherein each element of the transformed voltage coefficients vector F^ corresponds to a coefficient of the transformed voltage sample signal vector F^. Similarly, each element of the transformed current signal vector ^ corresponds to a coefficient of the transformed current sample signal vector

[0030] Fig. 3 illustrates the relationship between the original sample signal and the Haar wavelet coefficients generated by a Haar wavelet transformation. For simplicity, the illustrated example is of a block comprising 8 signal samples, rather than 216signal samples. The Haar wavelet coefficients for a given wavelet correspond to discrete frequency ranges (described here as the "scale") and discrete time periods.

[0031] As the scale becomes coarser (using long wavelets), the spread of the wavelets in frequency decreases and their time spread increases. The finest scale consists of M / 2 coefficients. At each level of the scale, the number of coefficients is half that of the previous level of the scale, resulting in an "inverse pyramid" containing M-l coefficients. The last remaining coefficient is the DC component, c_3,l show in Fig. 3.

[0032] Large signal blocks produce very long vectors of coefficients F^. In order to improve the performance of the classifier 103 by reducing overfitting of the data, it is advantageous to generate an input based on the transformed signal vectors F„ with a reduced number of elements.

[0033] Based on the transformed signal coefficient vectors F^, one or more feature vectors (or xn) are generated. A current feature vector may be generated based on the transformed current coefficient vector and a voltage feature vector may be generated based on the transformed voltage coefficient vector. The elements of the current feature vector and the voltage feature vector may be combined into a single feature vector having twice the number of elements as each individual feature vector.

[0034] The feature vectors q are generated by applying a statistical measure to the transformed signal coefficients. Each element of the feature vector may comprise information relating to several elements of the transformed signal coefficient vectors. Therefore, the feature vectors have fewer elements than the transformed signal coefficient vectors. When more than one feature vector is generated, the total number of elements in all the feature vectors may be fewer than the number of elements in transformed voltage coefficient vector or the transformed current coefficient vector.

[0035] When the elements transformed signal coefficient vectors comprise DWT coefficients, the elements of a combined feature vector can be generated according to the following rules:

[0036] For the voltage signal: The sum of wavelet coefficients from each DWT scale n = 0, 1, ... , 15. where cn fcare DWT coefficients from the nth scale, see Figure 3. (Oth scale corresponds to the scale with the shortest wavelet, 15th to the longest wavelet.)

[0037] For the voltage signal: The Z1norm (obtained as the sum of absolute values) of DWT coefficients from each DWT scale

[0038] For the current signal: The sum of wavelet coefficients from each DWT scale where ck nare DWT coefficients from the nth scale. For the current signal: The Z1norm (obtained as the sum of absolute values) of DWT coefficients from each DWT scale

[0039] These above measures greatly reduce the number of features. There are log2(M) = 16 wavelet scales. For each scale we compute two statistical measures listed above and this is repeated for both signals (source voltage and arc current). The process yields 2 - 2 - log2(M) = 64 measures in total. DWT coefficients can be analysed in groups with the same scale, thereby considerably reducing the number of parameters of a classifier (the number of weights w . In this case the reduction is from M = 2 • 216= 131,072 parameters to only 64 parameters (which is a reduction by more than three orders of magnitude).

[0040] In general, the number of elements of the feature vectors q (or ^n) is preferably at least 100 times less than the number of elements in either the voltage signal sample vector or the current signal sample vector, and more preferably, at least 1000 times less than the signal sample vectors.

[0041] Each element of the feature vector q may be further transformed according to the following scaling equation for individual elements of these vectors: , 2, ..., 64, 65, 66.

[0042] Here, the absolute value is used to secure that the natural logarithm yields real-valued output. The offset +1 guarantees that the minimal value stored in features xlrx2, ... , x66is non-negative.

[0043] The feature vector q (or xn) is provided as an input to a logistic regression classifier that is trained to classify feature vectors as arc-fault conditions or non-arc fault conditions. The classifier provides a binary output depending on the classification.

[0044] An ensemble of classifiers may be used to improve the accuracy of the classification, as shown in Fig. 4. The ensemble of classifiers may be implemented as an artificial neural network.

[0045] The output of the classifier may be provided to a control module configured to open a switch in the event that an arc fault is detected. The system may sequentially analyse a series of sample blocks according to a moving window in order to detect arc fault currents in real time.

[0046] In some examples, the control module is configured to provide an output signal to trigger a switch based on the output of the classifier over several sample blocks. In some example, the mean of outputs from the classifier from 5 consecutive windows is used to determine whether to open a switch to interrupt a current in the power distribution system. Preferably, the outputs of fewer than ten consecutive blocks are used to determine whether to trigger opening of the switch, as time required for the output increases the greater the number of blocks that is used in the moving average.

[0047] Fig. 5 illustrates a flowchart describing a method in accordance with the disclosure.

[0048] In operation S10, a voltage signal on a conductor forming a current path of the power distribution system is sampled to generate a voltage data series vector. A current signal is also sampled to generate a current data series vector.

[0049] In operation S20, one or more feature vectors based on the voltage signal sample vector and the current signal sample vector are generated. The one or more feature vectors has fewer elements than the current signal sample vector or the voltage signal sample vector. When more than one feature vector is generated, the total number of elements in all the feature vectors may be fewer than the number of elements in transformed voltage coefficient vector or the transformed current coefficient vector.

[0050] Preferably, generating the one or more feature vectors based on the voltage signal sample vector and the current signal sample vector comprises a first step of transforming the signal sample vectors to generate transformed coefficient vectors, and a second step of calculating the feature vectors based on the elements of the transformed coefficient vectors.

[0051] In operation S30, the system provides the one or more feature vectors to an arc fault classification model.

[0052] In operation S40, an indication of whether an arc fault is occurring in the power distribution system is received as an output of the arc fault classification model. Based on the receiving an indication that an arc fault is occurring, a switch may be opened to prevent current flow in the DC power distribution network.

[0053] In some examples, the opening of the switch may be performed based on a moving average of outputs of the classification model over a series of input blocks consecutive blocks. Preferably, fewer than ten input consecutive blocks are used to generate the moving average of the outputs of the classification model. Consecutive input blocks may overlap with each other such that a fraction of the signal sample values of a given input block are shared with the preceding input block. Performing a moving average of the outputs of the classification model improves robustness of the detection of short circuits during short transient disturbances.

[0054] Fig. 6 schematically illustrates a system in accordance with the invention. One or more measuring devices / sensors / detectors 601 are used to measure a voltage and a current in a current path of a DC power supply system. The voltages and currents may be measured simultaneously and sampled using an AD converter. The sampled measurements are provided to one or more processors 602. The processors 602 process the voltage signals and current signals to generate one or more feature vectors. The one or more feature vectors are provided by the processors 602 as an input to a classifier 603, which may be a machine learning model such as an artificial neural network. The classifier 603 provides an output indicating that an arc fault is occurring or is not occurring in a given signal block. When the classifier provides an indication that an arc fault is occurring, the processor 603 provides a trigger signal to a switching device 604 that interrupts the current supply in at least part of the DC power distribution network. The switching device 604 may, for example, comprise an electromechanical relay, a semiconductor switch, or any other form of switching device suitable for interrupting a current supply in a conductive path.

Claims

Claims1. Method of detecting arc faults in a DC power distribution system, the method comprising: sampling (S10), using one or more sensors (601), a voltage signal on a conductor forming a current path of the power distribution system to generate a voltage signal sample vector; sampling (S10), using the one or more sensors (601), a current signal in the conductor to generate a current signal sample vector; generating (S20) one or more feature vectors based on the voltage signal sample vector and the current signal sample vector, the one or more feature vectors having fewer elements than the current signal sample vector or the voltage signal sample vector; providing (S30) the one or more feature vectors to an arc fault classification model (603); and receiving (S40) as an output of the arc fault classification model an indication of whether an arc fault is occurring in the power distribution system.

2. The method of claim 1, wherein generating the one or feature vectors comprises: transforming the voltage signal sample vector using a first timefrequency transformation (101) to generate a transformed voltage coefficient vector; transforming (101) at least part of the current signal sample vector using a second time-frequency transformation, which may be the same as or different to the first time-frequency transformation, to generate a transformed current coefficients vector; and generating (102) the one or more feature vectors based on the transformed voltage coefficient vector and the transformed current coefficient vector, wherein the one or more feature vectors has fewer elements than the transformed voltage coefficient vector or the transformed current coefficient vector.

3. The method of claim 2, wherein the first and second time-frequency transformations are discrete wavelet transforms, DWTs, and, preferably, Haar DWTs.

4. The method of claim 3, wherein each element of the one or more feature vectors is generated by a method comprising one of: performing a summation of the values of the transformed voltage coefficient vector corresponding to wavelet coefficients of a given DWT scale; performing a summation of the absolute values of the transformed voltage coefficient vector corresponding to wavelet coefficients of a given DWT scale; performing a summation of the values of the transformed current coefficient vector corresponding to wavelet coefficients of a given DWT scale; or performing a summation of the absolute values of the transformed current coefficient vector corresponding to wavelet coefficients of a given DWT scale.

5. The method of claim 3 or claim 4, wherein the elements of the one or more feature vectors are generated by a method comprising: performing a summation of the values of the transformed voltage coefficient vector corresponding to wavelet coefficients for each DWT scale; performing a summation of the absolute values of the transformed voltage coefficient vector corresponding to wavelet coefficients for each given DWT scale; performing a summation of the values of the transformed current coefficient vector corresponding to wavelet coefficients for each DWT scale; and performing a summation of the absolute values of the transformed current coefficient vector corresponding to wavelet coefficients for each DWT scale.

6. The method of any preceding claim, further comprising: generating a plurality of further voltage signal sample vectors and current signal sample vectors from measurements of the conductor for different time periods according to a sliding signal window method, and determining whether an arc fault current is occurring in any of the time periods.

7. The method of any preceding claim further comprising, based on detecting that an arc fault has occurred, providing a trigger signal to a switching device (604) to cause a current path in the DC power distribution system to be interrupted.

8. The method of claim 7, wherein the trigger signal is provided to the switching based on a moving average of the output of the arc fault classification model over a series of consecutive inputs.

9. An arc detection device comprising: a conductive path forming part of a DC power distribution network; one or more sensors (601) configured to measure a voltage and a current in a conductive path; an AD converter configured to sample current measurements and voltage measurements to generate a voltage signal sample vector and a current signal sample vector; and a processor (602) configured to: generate one or more feature vectors based on the voltage signal sample vector and the current signal sample vector, the one or more feature vectors having fewer elements than the current signal sample vector or the voltage signal sample vector; provide the one or more feature vectors to an arc fault classification model (603); and receive as an output of the arc fault classification model an indication of whether an arc fault is occurring in the power distribution system.

10. The arc detection device of claim 9, wherein the processor is further configured to: transform the voltage signal sample vector using a first time-frequency transformation to generate a transformed voltage coefficient vector; transform at least part of the current signal sample vector using a second time-frequency transformation, which may be the same as or different to the first time-frequency transformation, to generate a transformed current coefficients vector; and generate the one or more feature vectors based on the transformed voltage coefficient vector and the transformed current coefficient vector, wherein the one or more feature vectors has fewer elements than the transformed voltage coefficient vector or the transformed current coefficient vector.

11. The arc detection device of claim 9 or claim 10, further comprising a switching device configured to interrupt a current path forming part of the DC power distribution network when an arc fault is detected.

Images