Apparatus and process for real-time detection of high-impedance faults in power lines
Receive analog signal representing real-time electrical current of electrical power transmission/distribution line; sample and digitise analog signal to generate power line current data; generate data frames of subsets of the power line current data for successive overlapping time slices, process data frames to generate frequency domain data representing frequency components for different frequency ranges; process frequency domain data to generate energy data representing average energies of frequency ranges over corresponding time slice; generate two-dimensional feature for machine learning, first dimension represents plurality of the successive time slices, second dimension represents different frequency ranges, values of the two-dimensional feature represent logarithms of corresponding average energies of frequency components of electrical current over corresponding time slices; apply deep learning classifier to two-dimensional feature to detect high-impedance fault in electrical power transmission/distribution line, and if detected: send fault signal to relay or circuit breaker to cut power to the electrical power transmission/distribution line.
1. A process for real-time detection of high-impedance faults in power lines, the process including the steps, performed in real-time, of:
receiving an analog signal representing a real-time electrical current of an electrical power transmission or distribution line as a function of time;
sampling and digitizing the received analog signal to generate corresponding power line current data representing the electrical current as a function of time;
processing the power line current data to generate data frames corresponding to respective subsets of the power line current data for respective time slices, and processing each of the data frames to generate corresponding frequency domain data representing frequency components of the electrical current over the corresponding time slice for each of a plurality of respective different frequency ranges;
processing the frequency domain data to generate energy data representing, for each of the frequency ranges and for each of the time slices, a corresponding average energy of the corresponding frequency components of the electrical current over the corresponding time slice; and
generating a two-dimensional feature for machine learning, wherein a first dimension of the feature represents a plurality of the time slices, a second dimension of the feature represents the respective different frequency ranges, and each of the values of the two-dimensional feature represents a logarithm of the corresponding average energy of the corresponding frequency components of the electrical current in the corresponding frequency range and over the corresponding time slice;
applying a deep learning classifier to the two-dimensional feature in order to generate an output to determine whether the two-dimensional feature is indicative of a high-impedance fault in the electrical power transmission or distribution line; and
if the output of the deep learning classifier is indicative of a high-impedance fault, then sending a fault signal to a corresponding relay or circuit breaker to cause the relay or circuit breaker to cut power to the electrical power transmission or distribution line,
wherein the deep learning classifier is a lightweight convolutional neural network (CNN) for classifying high impedance faults in real-time on a resource-constrained edge device, the CNN being configured with the following successive layers:
a first convolution layer with sliding filters;
a first batch normalization layer;
a first rectified linear unit (ReLU) activation layer;
a max pooling 2D layer;
a second convolution layer with sliding filters;
a second batch normalization layer;
a second ReLU activation layer;
first and second fully connected layers to process feature maps generated by the previous layers; and
a softmax activation layer to process the fully connected layer output to calculate a probability of the analog signal being indicative of an HIF.
2. The process of claim 1 , wherein the step of generating frequency domain data is executed in parallel with the steps of sampling and generating data frames.
3. The process of claim 1 , wherein the step of generating frequency domain data and the step of processing the frequency domain data are executed in parallel.
4. The process of claim 1 , wherein the received analog signal is sampled at a frequency of at least 20 kHz so that the corresponding frequency domain data represents frequency components up to at least 10 kHz.
5. The process of claim 1 , wherein each of the respective different frequency ranges has a corresponding minimum frequency and a corresponding maximum frequency, the maximum frequency being a fixed multiple of the minimum frequency.
6. The process of claim 5 , wherein, in each of the respective different frequency ranges, the corresponding maximum frequency is about twice the corresponding minimum frequency.
7. The process of claim 1 , wherein the subsets of the power line current data are successive but overlapping subsets of the power line current data for successive but overlapping time slices with about sampled 512 data points, and the data frames are generated by applying a Hann function to each of the subsets of the power line current data.
8. The process of claim 7 , wherein each successive pair of time slices overlaps by about 50%.
9. The process of claim 1 , wherein a first dimension of the feature represents about six successive but overlapping time slices, and a second dimension of the feature represents about eight different frequency ranges.
10. The process of claim 1 , wherein each convolution layer applies a convolution operation by moving the filters along the layer's input vertically and horizontally by a step size or ‘stride’ (S), and calculating a dot product of weights (W) and the layer's input (I), and then adding a bias (B) term, according to:
O
[
f
]
[
x
]
[
y
]
=
∑
k
=
0
C
-
1
∑
i
=
0
X
-
1
∑
j
=
0
Y
-
1
I
[
k
]
[
S
*
x
+
i
]
[
S
*
y
+
j
]
×
W
[
f
]
[
k
]
[
i
]
[
j
]
+
B
(
f
)
where O, f, C, X and Y respectively denote the output matrix, the number of filter groups in a convolutional layer, the number of channels in the input, the filter height and the filter width.
11. The process of claim 1 , wherein each batch normalization layer normalizes its input as mini-batches according to:
y
i
=
γ
(
x
i
-
μ
B
σ
B
2
+
ε
)
+
β
where xi, μB, σB 2 , γ, and β are input, mini-batch mean, mini-batch variance, scale factor and offset, and ε is introduced to improve numerical stability as mini-batch variance approaches 0.
12. The process of claim 1 , wherein the CNN is trained to determine a set of weights using a cross entropy loss function of the form:
E
(
θ
)
=
-
∑
i
=
1
n
∑
j
=
1
k
t
ij
ln
y
j
(
x
j
,
θ
)
where θ is a parameter vector, t ij denotes the i th sample belongs to the j th class, and yj(xi,θ) is the output for the j th sample.
13. A non-volatile storage medium having stored thereon FPGA configuration data and/or processor executable instructions that, when executed by at least one processor, cause the at least one processor and/or an FPGA configured in accordance with the FPGA configuration data to execute the steps of claim 1 .
14. An apparatus for real-time detection of high-impedance faults in power lines, the apparatus including:
an input to receive an analog signal representing a real-time electrical current of an electrical power transmission or distribution line as a function of time;
a memory component; and
at least one non-volatile memory component storing FPGA configuration data to configure the FPGA;
a field-programmable gate array (FPGA) component that, when configured in accordance with the configuration data, causes the FPGA to:
sample and digitize the received analog signal to generate corresponding power line current data representing the electrical current as a function of time;
select successive subsets of the power line current data; and
process the power line current data to generate data frames representing successive subsets of the power line current data for respective successive time slices, and processing each of the data frames to generate corresponding frequency domain data representing frequency components of the electrical current over the corresponding time slice for each of a plurality of respective different frequency ranges;
wherein the non-volatile memory component stores processor-executable instructions that, when executed by the microprocessor cause it to:
process the frequency domain data to generate energy data representing, for each of the frequency ranges and for each of the time slices, a corresponding average energy of the corresponding frequency components of the electrical current over the corresponding time slice; and
generate a two-dimensional feature for machine learning, wherein a first dimension of the feature represents a plurality of the time slices, a second dimension of the feature represents the respective different frequency ranges, and each of the values of the two-dimensional feature represents a logarithm of the corresponding average energy of the corresponding frequency components of the electrical current in the corresponding frequency range and over the corresponding time slices; and
apply a deep learning classifier to the two-dimensional feature in order to generate an output to determine whether the two-dimensional feature is indicative of a high-impedance fault in the electrical power transmission or distribution line; and
if the output of the deep learning classifier is indicative of a high-impedance fault, then send a fault signal to a corresponding relay or circuit breaker to cause the relay or circuit breaker to cut power to the electrical power transmission or distribution line,
wherein the deep learning classifier is a lightweight convolutional neural network (CNN) for classifying high impedance faults in real-time on a resource-constrained edge device, the CNN being configured with the following successive layers:
a first convolution layer with sliding filters;
a first batch normalization layer;
a first rectified linear unit (ReLU) activation layer;
a max pooling 2D layer;
a second convolution layer with sliding filters;
a second batch normalization layer;
a second ReLU activation layer;
first and second fully connected layers to process feature maps generated by the previous layers; and
a softmax activation layer to process the fully connected layer output to calculate a probability of the analog signal being indicative of an HIF.
15. The apparatus of claim 14 , wherein the FPGA is configured such that the steps of sampling and selecting are executed by a first slice of the FPGA, and the step of generating frequency domain data is performed by a second slice of the FPGA different to the first slice, the first and second slices operating concurrently.
16. The apparatus of claim 14 , wherein the FPGA and microprocessor are configured for pipelined parallel execution of the steps of generating frequency domain data and processing the frequency domain data.
17. The apparatus of claim 14 , wherein the apparatus is in the form of an embedded device attached to a power line support pole and operating as an edge computing device.