State monitoring of switching power supply in power system

Voltage deviation calculation

The deviation is calculated between the voltage waveform of the monitored switching power supply and the waveform of the normal switching power supply under the same load. The greater the deviation, the worse the health status of the switching power supply.

To obtain the voltage waveform of the normal switching power supply, the waveform from a new switching power supply under different loads is adopted as a reference. Figure 1 shows the voltage and current waveforms of switching power supplies in different stages of life cycle, where the red is the voltage and the blue is the current. The waveform of the switching power supply is similar to the reference waveform under normal conditions. When the switching power supply degrades, the voltage waveform gradually deviates, and the degradation of the power supply can be reflected by the degree of waveform deviation. For the faulty ones, there will be no voltage output.

Fig. 1Open in figure viewerPowerPointVoltage and current waveforms under different stages of life cycle.

Voltage waveform image analysis

The deviation degree of the voltage waveform mainly reflects the numerical variation of the voltage waveform. But the shape of the output waveform is not involved. Thus this method has difficulty to sense the degradation when the switching power supply is only mildly degraded. On the other hand, mild degradation can be seen in the voltage waveform image.

At present, convolutional neural network is normally used for image data analysis. Convolutional neural network is based on neural network. This paper adopts the ResNet18 convolutional neural network structure. The ResNet18 network structure is shown in Figure 2. It consists of 17 convolutional layers and a fully connected layer. The convolutional layer is designed as a residual structure. With such a structure, the ResNet18 convolutional neural network obtains the state classification probability that the switching power supply belongs to. Finally, the degradation state of the maximum probability of the switching power supply is calculated through the softmax function as the network output result.

Fig. 2Open in figure viewerPowerPointResNet18 structure diagram.

Experiment and result analysis

The experiment uses the CUS75E series switching power supply that is widely applied on site. The parameters of this type of switching power supply are: standard output voltage 5 V; maximum output current 12 A; maximum power 60 W; nominal maximum ripple voltage 120 mV. In total, 35 switching power supplies in different stages of life cycle were collected for experiments. According to the device status, the switching power supply can be divided into four different states: normal, mildly degraded, severely degraded, and faulty. The distribution of the collected switching power supplies is shown in Table 1:

The data is divided into a training set and a test set with the ratio of 8:2. The training set contains 4 power supplies in normal state, 12 power supplies with mild degradation, 10 power supplies with serious degradation, and 1 power supply with fault. The test set includes one normal power supply, three mildly degraded power supplies, three severely degraded power supplies, and one faulty power supply.

Data augmentation

The switching power supplies are tested with five different loads. Each test is carried out three times and different measured data fragments are intercepted. Finally, 391 pieces of output voltage data were obtained in the training set, and 106 pieces of voltage data were obtained in the test set.

To have more data for neural network training, image enhancement is performed on the training set data to increase the number of pictures. Image enhancement schemes include coordinate axis scaling, image cropping, rotation, etc. In the end, a total of 8000 training images were obtained, which were divided into training set with 7000 images and verification set with 1000 ones. For the data of the test set, no image augmentation is performed.

Neural network training

The test is based on python3.6, and the ResNet18 neural network model is built using the pytorch framework. To speed up the training speed of the model, the pre-trained model parameters based on the ImageNet dataset are used. The pre-trained model is trained on a large amount of data, which is conducive for model convergence. The pre-trained model parameters do not include the last fully connected layer.

The cross entropy is selected as the loss function. And the learning rate is 0.001. The experiment is carried out on the RTX3090, and the number of training generations is 30. The loss and test accuracy of the training set and validation set is recorded after each generation of training is completed. As shown in Figure 3, the accuracy of the training set and the test set are 99.24% and 96.40%, indicating that the neural network model based on the voltage waveform image can accurately distinguish the state of switching power supply.

Fig. 3Open in figure viewerPowerPointChange chart of detection accuracy.

Switching power supply prediction result

Fuzzy clustering algorithm is applied for the voltage waveform deviation as shown in Figure 4. Triangle membership function is utilized. For a given voltage deviation, four output results, namely normal, mild degradation, severe degradation, and fault, can be derived with certain probability.

Fig. 4Open in figure viewerPowerPointMembership of fuzzy algorithm.

For the voltage waveform image analysis, the model output is also divided into four categories: normal, mildly degraded, severely degraded, and fault by the softmax algorithm.

Therefore, it is only necessary to know the mass function of independent evidence to achieve information combination. In this paper, two evidence are available, i.e., assessment result from voltage deviation and waveform image analysis. The value α for the mass function of voltage deviation is derived from the probability for each state with fuzzy clustering algorithm, and the value for mass function of waveform image analysis is taken as the maximum probability before softmax treatment.

For performance comparison, the classification based on voltage waveform deviation, voltage waveform image analysis, and the method of this paper are conducted. The LSTM and DBN algorithms are based on references [6, 7]. The results are shown in Table 2 where the number indicates the percentage of correct classification. All algorithms have good results for the normal and fault power supplies since the characteristics of these switching power supplies are more obvious. But for mildly degraded and severely degraded switching power supplies, the algorithm proposed in this paper can do better.