Deep auto-encoder network for mechanical fault diagnosis of high-voltage circuit breaker operating mechanism

3.1 Extraction of time–frequency energy distribution of vibration signals

During the on-and-off operation of the high-voltage circuit breaker, the internal components of the mechanism constantly collide and rub, resulting in multiple vibrations. The vibration signals collected by the sensor are the superposition of all the vibrations in the on-and-off process [10]. The frequency, time, and intensity of vibration may change after mechanical failure. To reflect the attenuation or enhancement of the corresponding frequency components in the circuit breaker vibration signal, m layers wavelet packet transformation is applied to the denoised vibration signal s(t), and the signal components s _i (t), i = 1,2, …, 2 ^m of different frequency bands 2 ^m are obtained. To reflect the changes in the occurrence time and duration of a certain vibration during the opening and closing operation, the signal components of each frequency band are further divided into equal parts N along the time axis [11]. After wavelet packet transformation and equal time segmentation, vibration signals are divided into 2 ^m × N time–frequency subplanes, and the energy size E _i,n in each sub-plane is calculated as follows [12]:

In the formula, E _i,n represents the energy of the time–frequency subplane in the ith frequency band and the nth time period of the vibration signal. t _n and t _n+1 are the start time and the end time of the period, respectively.

Arrange the energy of the time–frequency subplane E _i,n according to frequency and time order, and the time–frequency energy distribution matrix of the vibration signal Q is obtained as follows:

In the process, the number of wavelet packet transform layers m and the number of time segments N, respectively, determine the sensitivity of frequency and time change [13]. Too low sensitivity will not detect changes caused by small frequency and time fluctuations. If it is too high, normal frequency and time fluctuations may be misjudged as faults, and subsequent fault identification will take longer. Take an example of the closing vibration signal of LW36–126 type circuit breaker operating mechanism for a better understanding. The sampling frequency of the signal is 100 kHz. The three-layer wavelet packet transform is carried out for it, and then, the obtained signal components are equally divided into 12 parts. WP1–WP8 are signal components of different frequency bands after wavelet packet transformation, respectively, representing the eight different frequency bands of 0–6.25, 6.25–12.50, 12.50–18.75, 18.75–25.00, 25.00–31.25, 31.25–37.50, 37.50–43.75, and 43.75–50.00 kHz. Formula (2) is used to calculate the energy of each subplane and the time–frequency energy distribution of the vibration signal is obtained [14]. Compared with other states, the energy in the time–frequency subplane of the vibration signal of the operating mechanism jam significantly decreased, especially the vibration almost disappeared 50 ms after the vibration starts. If the vibration occurs in other states and the operating mechanism is blocked, the vibration is delayed by about 10 ms. Therefore, the operating mechanism jam fault can easily be distinguished from other circuit breaker conditions. The main difference between the normal state and the core jam and screw loosening is that when the circuit breaker is in the normal state, the vibration signal in 60–70 ms vibration intensity is higher and that in 70–80 ms vibration intensity is lower after the opening operation. However, when the core of the circuit breaker is jammed or the base screw is loose, the vibration intensity of the signal is low at 60–70 ms, but high at 70–80 ms. The vibration signals in the second frequency band (6.25–12.50 kHz) and the third frequency band (12.50–18.75 kHz) are obviously weakened. However, the main manifestation of the two fault states is delayed and the fault features are obviously similar. So it is difficult to distinguish them by simple methods such as the threshold method. Multidimensional scale analysis (MDS) is applied to map the high-dimensional time–frequency energy distribution into a three-dimensional plane, and the time–frequency energy distribution of multiple sets of vibration signals in different states can be visualized [15]. It can be found that except for the operating mechanism, the other three types of states overlap each other in space and are difficult to be clearly distinguished. The reason is that the vibration signal of the circuit breaker is very uncertain, and the vibration signal collected from the same circuit breaker in the same state has a certain fluctuation in frequency domain and time domain. It can be seen that although time–frequency energy distribution can comprehensively reflect the changes of vibration signal in frequency, time, and vibration amplitude, it is impossible to make an effective and reliable fault diagnosis without further mining its potential deep information. Therefore, DAEN is adopted to mine deep features in time–frequency energy distribution to complete the subsequent fault identification process.

3.2 Fault diagnosis model based on DAEN

3.2.1 Auto-encoder

The DAEN is formed by stacking the input layer and hidden layer of multiple auto-encoder (AE). The AE is divided into two processes, encoding and decoding. It is an unsupervised network with the purpose of self-interpretation, and its structure is shown in Figure 2. During the coding process, the c-dimension input data are mapped to obtain the d-dimension hidden layer data. In the decoding process, the hidden layer data are used to reconstruct the input data and obtain the same c-dimension output data as the input data. Therefore, both the input and output layers of the AE contain c neural nodes, and the hidden layer contains d neural nodes. The training process is presented as follows [16,17].

Figure 2

Structure of AE.

X _k is any sample data in the AE training sample set, where k = 1, 2., S. S is the number of samples. The hidden layer data H is obtained after X _k coding, which is shown in formula (3) [18].

(3) H = f ( W X k + b ) .

In the aforementioned formula, W is the weight matrix between the input layer and the hidden layer of the AE. b is the hidden layer bias. f is the activation function sigmoid.

The hidden layer data H are decoded to obtain the reconstructed data X k ^ of input data, and the decoding process is shown in Eq. (4) [19].

(4) X k ^ = g ( W ′ H + b ′ ) .

In the aforementioned formula, W ′ is the weight matrix between the hidden layer and the output layer of the AE. b′ is the output layer bias. g is the activation function sigmoid.

When the input data X _k is close enough to the reconstructed data X k ^ , the AE training is completed. At this point, the hidden layer data H in the AE can be regarded as the feature expression of the input data X _k .

To minimize the error between the input data X _k and the reconstructed data X k ^ , the mean square error of the two is defined as the reconstructed error function. The redundant information in the hidden layer data is further removed, and the sparse penalty term is introduced into the reconstruction error function. The reconstruction error function is set as formula (5).

(5) J ( θ ) = 1 S ∑ k = 1 S ∥ X k − X k ^ ∥ 2 + β ∑ l = 1 d ρ ln ρ ρ l + ( 1 − ρ ) ln 1 − ρ 1 − ρ l .

In the formula, θ is the AE network parameters, including weight matrix W and W ′, bias term b′ and b. β is the sparse penalty term coefficient. ρ is the sparse parameter set. ρ _l is the average activation degree of the lth neural node in the hidden layer.

The gradient descent algorithm is used to reduce the reconstruction error function J(θ). When it drops to the specified value or reaches the specified number of iterations, the AE training is completed.

3.2.2 Training of DAENs

The DAEN with a hidden layers can be regarded as formed by stacking AEs. Each hidden layer and its previous networks are the hidden layer and the input layer of the AE, respectively. The hidden layer of the AE serves as the input layer of the next AE [20], as shown in Figure 3.

Figure 3

DAEN structure.

The training of a AEs is completed one by one from the bottom layer to the top layer. The weight matrix and bias between the hidden layer and the input layer of these AEs are the initial parameters of the DAEN to complete the unsupervised pretraining process of the DAEN. Then the parameters of DAEN are fine-tuned. At the last layer of the network, the BP neural network is selected for classification and the whole network is taken as a whole. The labeled data are used to adjust the network parameters of each layer through the error back propagation algorithm, and the training of DAEN is completed.

3.2.3 High-voltage circuit breaker diagnosis method based on DAEN

Figure 4 shows the fault diagnosis process of the high-voltage circuit breaker using DAEN, which is divided into four steps as follows [21]:

1. The time–frequency energy distribution of vibration signals is extracted as the input of DAEN. The method described in Section 1 is used to extract the time–frequency energy distribution of circuit breaker vibration signals. Due to the large difference in energy magnitude in time–frequency energy distribution, direct input as feature is not conducive to network training. Therefore, Z-score standardized processing is carried out to enhance the comparability of data.

2. The pretraining process of DAEN is carried out by using time–frequency energy distribution without labels. The layer number of the hidden layers of the DAEN is set and then the number of auto-encoders in the network is determined. The learning rate, iteration times, and training target of each layer of auto-encoder are set, and the weight matrix and bias term of auto-encoder are randomly initialized. The time–frequency energy distribution features without labels are input into the DAEN, and each auto-encoder is trained layer by layer. The trained auto-encoder weight matrix and bias term are used to initialize the DAEN.

3. The time–frequency energy distribution with labels is used to fine-tune the DAEN. The learning rate, iteration times, and training objectives of the DAEN are set. The time–frequency energy distribution with labels is input into the DAEN, and the label value is taken as the expected output of the network. The parameters of the initialized DAEN are adjusted through the error back propagation algorithm until the training goal or iteration number of the network is met.

4. Fault diagnosis of vibration signal of the diagnostic circuit breaker is performed. The features of time–frequency energy distribution in the data to be diagnosed are extracted by the same method. It is input into the DAEN, the label value is output, and the results of the corresponding DAEN fault diagnosis are obtained.

Figure 4

Fault diagnosis process.

4.1 Acquisition of vibration signal of high-voltage circuit breaker

In this study, LW36-126 SF6 high-voltage circuit breaker of a company was taken as the experimental object, and the operation state of the circuit breaker was simulated when the mechanical fault occurred. Vibration signals were collected by LC0102T accelerometer, whose frequency range (±10%) was 0.5–13,000 Hz, resonance frequency was 50 kHz, and sensitivity was 5 mV/g. The acceleration sensor was fixed above the base of the brake spring of the operating mechanism by screws, which was near the center of the operating mechanism and adjacent to the brake spring pull rod. In the LW36−126 circuit breaker, the energy required for opening and closing was transmitted to the output connecting rod of the operating mechanism through the opening spring rod and finally reached the arc extinguishing chamber. So the vibration signal collected at this position could reflect the mechanical features of the circuit breaker operating mechanism more comprehensively. The axis of the mounting screw hole of the acceleration sensor was consistent with the vibration direction. When the circuit breaker was closed, vibration signals were collected at a sampling frequency of 100 kHz. The collected vibration signal was recorded by the signal conditioning input waveform monitoring recorder. In addition to the normal state, three typical mechanical faults of circuit breakers, such as operating mechanism jam, iron core jam, and loose base screw, were simulated. At the beginning of the experiment, each fault state of the circuit breaker was simulated, and the opening and closing operation was carried out continuously to collect the vibration signal of the operating mechanism closing. The obtained vibration signals had the same recording starting point and sampling time. On the premise of not damaging the experimental circuit breaker, the fault simulation was realized by the following three methods.

1. The operating mechanism was jammed, and the closing energy was weakened, leading to the decline of the opening and closing speed and even the occurrence of misoperation and rejection. The closing energy of the experimental circuit breaker was stored in two closing springs, so the small spring was taken out to reduce the closing energy. The operation mechanism of the circuit breaker was simulated.

2. The core fault was a delay fault in nature, which could lead to the delay of the electromagnet dynamic and static core collision and the closing retaining switch hitting the latch in the opening and closing processes, but it still contained all the vibration laws in the normal opening and closing operation. This fault was caused by excessive resistance of iron core caused by the dirt of iron core and unqualified fit of iron core. Therefore, in the experiment, the foreign matter was mixed into the iron core to simulate the failure of iron core jam.

3. The looseness of the base screw was easy to lead to increased wear of the mechanism, poor line contact, and other faults, which contained all the information in the normal opening and closing operation similar to the failure of the iron core jam. Due to screw looseness, vibration acceleration and vibration time collected by the sensor would change. In this experiment, the base screw was screwed out 5 mm to simulate the looseness fault of the base screw.

In the fault simulation experiment, the fault point of the operating mechanism was located in the closing spring. Iron core jam acted on the closing electromagnet. A loose base screw was located above the base of the brake spring. In practical engineering, sample data in different states were often uneven and the sample data in the normal state was far more than that in other states, namely, there was the problem of sample imbalance. Therefore, in this experiment, 120 sets of vibration signals of the circuit breaker in the normal state, 60 sets of vibration signals of operating mechanism jam fault, 60 sets of vibration signals of iron core jam fault, and 80 sets of vibration signals of base screw looseness fault were collected.

4.2 Influence of the number of hidden layers on the fault recognition effect

DAEN can dig deeply into the input layer data through multiple mappings between hidden layers to obtain deeper features. The more hidden layers there are, the closer the features are to the essence, which is more conducive to fault identification. However, with the increasing number of hidden layers, the network training time is longer. The unique features of the training sample itself may be regarded as the universal features of all samples, falling into the overfitting situation. To investigate the influence of the number of hidden layers on the fault recognition effect, 75% of the state data samples in the simulation experiment were randomly selected as the training set and the remaining 25% as the prediction set. In the process of feature extraction, considering the effect of fault diagnosis and computation, Daubechies10 wavelet base was selected to carry out a three-layer wavelet packet transformation on the denoised vibration signal after several experiments, and eight signal components with the equal broadband band were obtained. Then, each signal component was divided into 12 parts along the time axis, and the time–frequency energy distribution of vibration signal was obtained, which was input into the DAEN after standardized processing. The DAEN pretraining and fine-tuning process were set as 500 iterations with a learning rate of 0.1. The network output was the label value corresponding to the status of the circuit breaker. The label value of normal status, operation mechanism jam, iron core jam, and base screw looseness were 1,000, 0100, 0010, and 0001, respectively. The number of neural nodes in the hidden layer of each layer was temporarily the same as that in the input layer, and each layer contained 96 nodes. When the network included one to four hidden layers, the fault identification accuracy after multiple experiments is presented in Table 1.

As shown in Table 1, when the DAEN contained one hidden layer, the fault diagnosis accuracy was low, only 92.47%. When there were two hidden layers, the accuracy of fault diagnosis increased to 96.08%. As the number of hidden layers increased, the fault diagnosis accuracy did not change significantly. Therefore, the number of hidden layers was set to 2.

4.3 Influence of the distribution of hidden layer nodes on fault diagnosis effect

According to the network parameter setting method and conclusion presented in Section 4.2, the influence of the distribution of hidden layer nodes on the fault diagnosis effect after several experiments is presented in Table 2. “96-72-54-4” in Table 2 is taken as an example. This distribution represented that the input layer had 96 nodes, the hidden layer of layer 1 and layer 2 contained 72 and 54 nodes, respectively, and the output layer contained 4 nodes. Therefore, it could be seen from Table 2 that when the distribution of nodes in the hidden layer was incremental, the fault diagnosis accuracy was low. The reason was that the incremental distribution was essentially a way of processing the input data in the ascending dimension. Even if the data were deeply mined, redundant information may be introduced to interfere with the normal classification of the network. Only the decreasing distribution could suppress the redundant information in the process of data mining, but the rapid reduction of the number of nodes in the hidden layer would also inhibit the learning of the deep information, thus reducing the accuracy of fault diagnosis. Therefore, 96-48-24-4 type distribution of the diminishing hidden layer distribution was selected in the research.

4.4 Results and analysis of fault diagnosis

According to the aforementioned research results, it was determined that the DAEN contained two hidden layers with the learning rate of 0.1, and the number of pretraining and fine-tuning iterations was 500. The input layer contained 96 nodes, the same number as the time–frequency subplane. The number of hidden layer nodes was set to 50% of that of the previous layer network. The hidden layer of layer 1 and layer 2 contained 48 and 24 neural nodes, respectively. The output layer contained four neural nodes, equal to the length of the label value. From 120 groups of normal state data, 60 groups of operating mechanism jam data, 60 groups of iron core jam data, and 80 groups of state base screw looseness data, 75% data were randomly selected as the training set and the remaining 25% data were selected as the prediction set. After standardizing the time–frequency energy distribution features of the training set, they were input into the DAEN as original features. To further verify the performance of DAEN, the same data samples and feature extraction methods were used in the research to conduct fault diagnosis of circuit breakers through DAEN, BP neural network, and support vector machine. Figure 5 shows that the three identification methods in fault diagnosis of high-voltage circuit breaker could effectively identify the normal state and operating mechanism jamming. However, for the fault states with similar features such as iron core jam and base screw looseness, the BP neural network and support vector machine (SVM) could not mine the deep information in the input data, so the fault diagnosis effect was poor. DAEN could identify the two states reliably and effectively by learning its deep features from the two-layer network. Therefore, compared with the traditional shallow network, deep auto-coding network had a better diagnosis effect in high-voltage circuit breaker fault diagnosis.

Figure 5

Fault diagnosis accuracy rate.

The three identification methods in the fault diagnosis of high-voltage circuit breakers can effectively identify the normal state and the jamming of the operating mechanism. However, for the two characteristics of iron core jamming and base screw loose, which show similar fault states, the two shallow networks, BP neural network and support vector machine, cannot mine the deep information in the input data, and the fault diagnosis effect is poor. The deep self-encoding network obtains its deep features through the learning of the two-layer network, which can reliably and effectively identify these two states. Therefore, compared with the traditional shallow network, the deep self-encoding network has a better diagnostic effect in the fault diagnosis of high-voltage circuit breakers.