Control cable anomaly detection model and method based on deep learning

Abstract

The invention discloses a deep learning-based control cable anomaly detection model and method, and the method comprises the steps: building an LSTM-VAE model, and employing a reconstruction error between VAE input and output; the TFDR result is input into LSTM-VAE frame by frame, and the abnormal score of each frame is calculated according to the logarithm reconstruction probability. And when the abnormal score exceeds 1, the system automatically gives an alarm to prompt the occurrence of the abnormality. The method has the beneficial effects that the probability distribution difference between normal data and abnormal data can be measured, and meanwhile, the cable joint abnormity which is difficult to detect by the existing method can be detected.

Description

Technical Field

[0001] The present invention relates to the field of cable detection, and in particular to a control cable anomaly detection model and method based on deep learning. Background Art

[0002] In the power system, cables are the blood vessels for transmitting power, measurement signals, and control signals. According to their functions, they can be divided into power cables, instrument cables, and control cables. Most of the above cables are in the form of shielded cables, with a common conductive layer around the conductor for electromagnetic shielding. Cables operating in harsh environments will age due to moisture intrusion or inconsistent temperatures, leading to failures. Therefore, various fields are actively researching the development of diagnostic technology.

[0003] Cable diagnostics can be divided into physical, chemical, and electrical methods. Physical methods that utilize mechanical properties such as extension, as well as methods that use chemical reactions with oxygen can accurately check the condition of the cable. However, since both are destructive methods that require cutting the cable, their use is usually limited to cross-verification of the results of other diagnostic methods. Among the various electrical diagnostic methods used for control cables, the most representative methods include dissipation factor (tanδ) measurement, partial discharge (PD) measurement, and reflection measurement.

[0004] Dissipation factor measurement is widely used to estimate the overall integrity of the cable, but it cannot detect local faults. In addition, since the dissipation factor measurement uses the relationship between the charging current and the loss current, it requires a high-power charging source and a large area of ​​test equipment. These characteristics make this method difficult to apply to long-distance transmission lines of tens of kilometers or even longer. On the other hand, PD measurement can detect local faults in the control cable, but high-voltage equipment must be used to inject the test voltage. But high-voltage equipment is a necessary condition for injecting the test voltage, and a high level of expertise is required to interpret the results. Reflectometry uses reflected waves at impedance discontinuities to overcome the limitations of the above methods. By analyzing the signal reflected from the fault, the reflectometer can detect local faults within the cable even with simple test equipment.

[0005] Traditional reflection methods can be divided into time domain reflectometry (TDR), frequency domain reflectometry (FDR), and time-frequency domain reflectometry (TFDR) according to the domain where the analysis reference signal is located. Recently, spread spectrum TDR (SSTDR) using m-sequence and step frequency waveform reflectometry (SFWR) using local sinusoidal signals have obtained high-resolution results. Among them, TFDR has been actively applied to various types of shielded cables, such as instrumentation and control cables, high-voltage DC cables, and high-temperature superconducting cables. So far, reflection methods have been used to detect faults occurring on cables rather than cable joints. However, since most faults in cable systems occur at cable joints where the electric field is highly concentrated, it is necessary to study a method to distinguish between defective states and normal states of cable joints.

[0006] Despite the urgent need, there are still the following unresolved issues. First, conventional reflectometry cannot identify the difference between the reflection signal of a faulty cable joint and the reflection signal of a normal cable joint. In the scenario of joint failure, PD measurement is mainly based on its advantage of being able to obtain high-resolution discharge through the ground wire. In contrast, due to the particularity of cable joints, it is difficult for conventional reflectometry to detect joint failures. Since the cable joint is a representative impedance discontinuity, a reflection signal already exists in the normal state. When a fault occurs, the existing reflection signal changes slightly, forming a large blind area, making anomaly detection more complicated. In addition, it is difficult to establish an abnormal threshold that can distinguish between normal and early abnormalities using conventional reflectometry. In traditional TFDR, time-frequency cross correlation (TFCC) is used as a monitoring indicator. However, since TFCC changes even when the cable system is operating healthily, it is difficult to draw a line between normal and abnormal cable conditions. Solving each problem requires different goals. By solving the first problem, abnormal scenarios occurring at impedance mismatch points (such as cable joints and terminals) can be detected. By solving the second problem, a threshold for identifying abnormalities can be automatically set even when there are multiple normal groups. To achieve anomaly detection using reflectometry, it is necessary to develop an algorithm that can automatically classify various groups of normal and abnormal states in the cable system, and machine learning methods are suitable candidates for this purpose.

[0007] Anomaly detection is divided into supervised anomaly detection (using both normal and anomaly data), single-class anomaly detection (using only normal data), and unsupervised anomaly detection (using unlabeled data). The recently introduced Deep Support Vector Data Description (DeepSVDD) is the most famous single-class anomaly detection technique. In Deep Support Vector Data Description, the boundary between normal and anomaly data is drawn by mapping the normal data to the smallest hypersphere. However, since this method uses a single center, the performance of single-class anomaly detection degrades when the normal data pattern is diverse. In the unsupervised anomaly detection category, autoencoders (AEs) that recover input data are mainly used because they can handle various types of normal data. Autoencoders have been widely used for anomaly detection in various fields, including but not limited to video surveillance, intrusion detection, and time series anomaly detection. Since AE reconstruction is limited to the normal data used for training, the reconstruction error is usually used to detect anomalies. Unlike traditional AEs that map input data to latent vectors, variational AEs (VAEs) map input data to random parameters of a probability distribution, such as the mean and variance of the latent vector. Therefore, VAE is actively used in anomaly detection due to its ability to measure the difference in probability distribution between normal data and abnormal data.

[0008] Although research on anomaly detection in control cables is actively being conducted, it is still in its infancy. The only recently published study applied anomaly detection techniques based on SSTDR and VAE to control cables used in PV systems. The proposed method successfully detects disconnections regardless of whether the current waveform fluctuates over time. However, even with existing TFCCs, conventional TFDRs detect disconnections regardless of solar radiation. Summary of the invention

[0009] The purpose of this invention is to solve the above problems and design a control cable anomaly detection model and method based on deep learning. The specific design scheme is:

[0010] A control cable anomaly detection model and method based on deep learning. The LSTM-VAE model is built and the reconstruction error between the VAE input and output is used. The TFDR results are input into the LSTM-VAE frame by frame, and the anomaly score of each frame is calculated based on the logarithmic reconstruction probability. When the anomaly score exceeds 1, the system will automatically issue an alarm to indicate that an anomaly has occurred.

[0011] The VAE is a variational autoencoder;

[0012] AE is a widely used unsupervised learning method consisting of two consecutive neural networks, called encoder and decoder. To train AE, the target output of the decoder is set to the same vector as the encoder input. In addition, the dimension of the encoder output is generally set lower than the dimension of the input vector. The encoder attempts to compress the input vector into important features, and the decoder attempts to restore the input vector from the compressed features. The output vector z of the encoder is called the latent variable, and the space mapped by z is called the latent space.

[0013] VAE is an extension of AE. AE maps the input vector x to the latent variable z through the encoder, but VAE uses a probabilistic encoder to map x to the probability distribution of z. Similarly, the decoder of AE generates an approximation of x based on the latent variable z, but the probabilistic decoder of VAE generates the mean and variance of x based on the given z.

[0014] Mathematically, the encoder and decoder of VAE are represented by q φ (z|x) and p θ (x|z) represents, where φ and θ are the parameters of the encoder and decoder respectively. To find the optimal parameters of VAE through maximum likelihood estimation, the marginal likelihood p(x)=∫p θ (x|z)p(z)dz is maximized. However, since the search space of z is continuous and large, the marginal likelihood p(x) is difficult to achieve, so during VAE training, the lower bound of the marginal log-likelihood is maximized. According to Jensen's inequality, the boundary of the marginal log-likelihood is as follows:

[0015]

[0016] Among them, p θ (z) is the prior distribution of z, and KL is the Kullback–Leib ler (KL) divergence. Since the KL divergence represents the difference between two probability distributions, during training, p θ (z) becomes q φ The first term of the lower bound corresponds to the reconstruction error, and the second term regularizes the latent variable z.

[0017] For anomaly detection, only data acquired from normal cables is used to train the VAE. After training, the VAE shows a high reconstruction probability for normal state data. However, since the VAE has never encountered abnormal state data, the reconstruction probability drops sharply when an anomaly occurs. The monitoring system can be designed by setting an appropriate threshold for the reconstruction probability.

[0018] The TFDR is an improved fault location technology, specifically a time-frequency domain reflectometry method;

[0019] The reference signal used has an instantaneous frequency that increases linearly with time. Different from the reference signal of traditional reflection measurement methods such as TDR and FDR, the reference signal of TFDR is located in the time domain and frequency domain through the Gaussian envelope. The Gaussian envelope linear chirp is the reference signal of TFDR, and its expression is as follows:

[0020]

[0021] Here A is the amplitude, which is proportional to the energy of the reference signal. α determines the time duration (TD), and the chirp rate β determines the frequency bandwidth (BW). ω0 and t0 are the center frequency (CF) and center time of the reference signal, respectively. Before testing, the design parameters should be optimized according to the physical structure, propagation characteristics and total length of the test cable.

[0022] The designed reference signal is injected into the test cable and propagates. When the propagating signal encounters impedance discontinuity, the signal will be multiplied by the reflection coefficient and reflected backward. In order to detect the reflected signal from the acquired time series, TFDR uses the Wigner-V il le distribution to convert the reference signal and the reflected signal into the time-frequency domain. The time-frequency distribution of the reference signal can be obtained as follows:

[0023]

[0024] Similarly, the TF distribution of the reflected signal can also be calculated, that is, W r (t, ω),

[0025] The last step of TFDR is time-frequency feature template matching. Since the instantaneous frequency of the reference signal increases linearly, TFDR can easily detect the reflected signal in the time-frequency domain using TFCC. The TFCC between the reference signal and the reflected signal is expressed as follows:

[0026]

[0027] Since TFCC is normalized based on the energy of the reference signal and the reflected signal, the derived value is always limited between 0 and 1. Since TFCC is based on the similarity with the reference signal and is insensitive to changes in the reflected signal, the application of TFCC in cable monitoring has inherent disadvantages. This paper proposes a new monitoring indicator called anomaly score and compares it with TFCC through actual experiments.

[0028] The whole process of TFDR is to form a local peak of TFCC through impedance discontinuities such as cable joints, terminals and faults. Even if the cable joints and terminals are intact after installation, there will be reflected signals. Although cable faults most often occur at joints and terminals, the pre-existing reflected signals will become blind spots, making fault detection complicated. The control cable abnormality detection method proposed in the present invention solves this problem by memorizing the normal cable state after installation as a baseline.

[0029] LSTM stands for long short-term memory network;

[0030] The LSTM-VAE is a hybrid model that combines LSTM and VAE, specifically a control cable anomaly detection network architecture.

[0031] The network architecture of the LSTM-VAE includes a data window, a normalization layer, an encoder, a latent variable sampling, a decoder, and a scoring layer.

[0032] In the scoring layer, p θ (x N |z) is the input data x N The reconstruction probability is obtained by using the decoder x N The reconstruction probability is calculated by the mean and variance of . The reconstruction probability is scaled using a logarithmic function, and the scaled probability is the same as the first term of the lower bound in (4). After training the VAE using the TFDR results under normal conditions, the reconstruction probability of each frame is calculated during monitoring. This probability value is high in normal data encountered by the VAE, but it decreases when abnormal data is input into the network. Therefore, the minimum reconstruction probability under normal conditions can be used as an abnormal threshold to divide normal and abnormal data. Finally, the anomaly score is defined by normalizing the reconstruction probability. The anomaly score of the kth frame is defined as follows:

[0033]

[0034] R(k)=log(p θ (x N (k)|z(k))) (7)

[0035] Among them, R n is the set of logarithmic reconstruction probabilities obtained under normal conditions, and R(k) is the logarithmic reconstruction probability of the kth frame. The anomaly score S(k) of the kth frame is the final output of the scoring layer and is monitored.

[0036] The encoder and decoder of VAE contain LSTM layers and dense layers to process time series data.

[0037] The frames are normalized using a fully connected network so that the chirp signal has a larger peak and the noise has a smaller peak.

[0038] The amplitude of the reflected signal generated by the abnormal point in the test cable is usually smaller than the chirp signal that exists in the normal state, such as the reference signal and the reflected signal from the cable joint and terminal. Therefore, if the windowed TFDR result is directly used as the input of the VAE without normalization, even if the VAE cannot recover the signal, the reconstruction error of the abnormal reflected signal is small due to the smaller reflection. On the contrary, if all input frames are normalized to have the same peak value, the VAE cannot recover the added random noise, and the reconstruction error will increase even in the normal state,

[0039] As mentioned above, the purpose of the normalization layer is to set the peak value high if the data frame contains a lot of chirp signals, and to set the peak value low if it does not contain chirp signals. Figure 2 As shown in Figure 1, the normalization layer includes a fully connected network to estimate the proportion of the chirp signal in the input frame. Before entering the network, the peak values ​​of all frames are initialized to 1. The k-th output frame x of the normalization layer N (k) is the k-th input frame x(k) multiplied by the normalized weight W in the network k The training set of the normalized network is constructed by sweeping the center time t0 and phase of the reference signal, and the target variable of the kth training frame is Determine as follows:

[0040]

[0041] Where Ck is the length of the chirp signal in the kth training frame, TD is the time length of the reference signal,

[0042] In order to make the normalization layer achieve the expected effect in the actual monitoring environment, some Gaussian noise is included in the training set of the normalization layer. The target variable of the kth training frame is Determined as (5), but the target variable of the Gaussian noise frame is fixed to 0. In addition, from the perspective that the normalization layer is a matched filter of the chirp signal, the window width that determines the frame size is most often used as the TD of the reference signal. If the window width is reduced, the spatial resolution will increase. However, due to the degradation of the performance of the matched filter, the anomaly detection performance will also decrease.

[0043] The structure of the encoder and decoder is determined according to the type of input data. If the input is an image, convolutional layers and deconvolutional layers are usually used for feature extraction processing. Using the proposed method, even if the signals look the same, it is possible to determine whether the test cable is normal or abnormal based on the position of the chirp signal. Therefore, the design of the encoder and decoder combines LSTM and dense layers. The LSTM layer can affect the anomaly detection result of the next input frame through the hidden state, while the dense layer is used to convert the hidden state into the desired output, such as the mean and variance of z and xN.

[0044] The control cable anomaly detection model and method based on deep learning obtained by the above technical solution of the present invention have the following beneficial effects:

[0045] It can measure the difference in probability distribution between normal data and abnormal data, and detect cable joint anomalies that are difficult to detect with existing methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 The present invention is a control cable anomaly detection network architecture;

[0047] Among them, each frame used as the input of the normalization layer is denoted as x(1), x(2), .., x(M);

[0048] M represents the total number of frames;

[0049] Figure 2 is the normalized layer structure of the present invention;

[0050] Figure 3 This is a flow chart of control cable abnormality detection according to the present invention. DETAILED DESCRIPTION

[0051] The present invention will be described in detail below in conjunction with the accompanying drawings.

[0052] A control cable anomaly detection model and method based on deep learning. The LSTM-VAE model is built and the reconstruction error between the VAE input and output is used. The TFDR results are input into the LSTM-VAE frame by frame, and the anomaly score of each frame is calculated based on the logarithmic reconstruction probability. When the anomaly score exceeds 1, the system will automatically issue an alarm to indicate that an anomaly has occurred.

[0053] The VAE is a variational autoencoder;

[0054] AE is a widely used unsupervised learning method consisting of two consecutive neural networks, called encoder and decoder. To train AE, the target output of the decoder is set to the same vector as the encoder input. In addition, the dimension of the encoder output is generally set lower than the dimension of the input vector. The encoder attempts to compress the input vector into important features, and the decoder attempts to restore the input vector from the compressed features. The output vector z of the encoder is called the latent variable, and the space mapped by z is called the latent space.

[0055] VAE is an extension of AE. AE maps the input vector x to the latent variable z through the encoder, but VAE uses a probabilistic encoder to map x to the probability distribution of z. Similarly, the decoder of AE generates an approximation of x based on the latent variable z, but the probabilistic decoder of VAE generates the mean and variance of x based on the given z.

[0056] Mathematically, the encoder and decoder of VAE are represented by q φ (z|x) and p θ (x|z) represents, where φ and θ are the parameters of the encoder and decoder respectively. To find the optimal parameters of VAE through maximum likelihood estimation, the marginal likelihood p(x)=∫p θ (x|z)p(z)dz is maximized. However, since the search space of z is continuous and large, the marginal likelihood p(x) is difficult to achieve, so during VAE training, the lower bound of the marginal log-likelihood is maximized. According to Jensen's inequality, the boundary of the marginal log-likelihood is as follows:

[0057]

[0058] Among them, p θ (z) is the prior distribution of z, and KL is the Kullback–Leibler (KL) divergence. Since the KL divergence represents the difference between two probability distributions, during training, p θ (z) becomes q φ The first term of the lower bound corresponds to the reconstruction error, and the second term regularizes the latent variable z.

[0059] For anomaly detection, only data acquired from normal cables is used to train the VAE. After training, the VAE shows a high reconstruction probability for normal state data. However, since the VAE has never encountered abnormal state data, the reconstruction probability drops sharply when an anomaly occurs. The monitoring system can be designed by setting an appropriate threshold for the reconstruction probability.

[0060] The TFDR is an improved fault location technology, specifically a time-frequency domain reflectometry method;

[0061] The reference signal used has an instantaneous frequency that increases linearly with time. Different from the reference signal of traditional reflection measurement methods such as TDR and FDR, the reference signal of TFDR is located in the time domain and frequency domain through the Gaussian envelope. The Gaussian envelope linear chirp is the reference signal of TFDR, and its expression is as follows:

[0062]

[0063] Here A is the amplitude, which is proportional to the energy of the reference signal. α determines the time duration (TD), and the chirp rate β determines the frequency bandwidth (BW). ω0 and t0 are the center frequency (CF) and center time of the reference signal, respectively. Before testing, the design parameters should be optimized according to the physical structure, propagation characteristics and total length of the test cable.

[0064] The designed reference signal is injected into the test cable and propagates. When the propagating signal encounters impedance discontinuity, the signal will be multiplied by the reflection coefficient and reflected backward. In order to detect the reflected signal from the acquired time series, TFDR uses Wigner-Ville distribution to convert the reference signal and the reflected signal into the time-frequency domain. The time-frequency distribution of the reference signal can be obtained as follows:

[0065]

[0066] Similarly, the TF distribution of the reflected signal can also be calculated, that is, W r (t, ω),

[0067] The last step of TFDR is time-frequency feature template matching. Since the instantaneous frequency of the reference signal increases linearly, TFDR can easily detect the reflected signal in the time-frequency domain using TFCC. The TFCC between the reference signal and the reflected signal is expressed as follows:

[0068]

[0069] Since TFCC is normalized based on the energy of the reference signal and the reflected signal, the derived value is always limited between 0 and 1. Since TFCC is based on the similarity with the reference signal and is insensitive to changes in the reflected signal, the application of TFCC in cable monitoring has inherent disadvantages. This paper proposes a new monitoring indicator called anomaly score and compares it with TFCC through actual experiments.

[0070] The whole process of TFDR is to form a local peak of TFCC through impedance discontinuities such as cable joints, terminals and faults. Even if the cable joints and terminals are intact after installation, there will be reflected signals. Although cable faults most often occur at joints and terminals, the pre-existing reflected signals will become blind spots, making fault detection complicated. The control cable abnormality detection method proposed in the present invention solves this problem by memorizing the normal cable state after installation as a baseline.

[0071] LSTM stands for long short-term memory network;

[0072] The LSTM-VAE is a hybrid model that combines LSTM and VAE, specifically a control cable anomaly detection network architecture.

[0073] The network architecture of the LSTM-VAE includes a data window, a normalization layer, an encoder, a latent variable sampling, a decoder, and a scoring layer.

[0074] In the scoring layer, p θ (x N |z) is the input data x N The reconstruction probability is obtained by using the decoder xN The reconstruction probability is calculated by the mean and variance of . The reconstruction probability is scaled using a logarithmic function, and the scaled probability is the same as the first term of the lower bound in (4). After training the VAE using the TFDR results under normal conditions, the reconstruction probability of each frame is calculated during monitoring. This probability value is high in normal data encountered by the VAE, but it decreases when abnormal data is input into the network. Therefore, the minimum reconstruction probability under normal conditions can be used as an abnormal threshold to divide normal and abnormal data. Finally, the anomaly score is defined by normalizing the reconstruction probability. The anomaly score of the kth frame is defined as follows:

[0075]

[0076] R(k)=log(p θ (x N (k)|z(k))) (7)

[0077] Among them, R n is the set of logarithmic reconstruction probabilities obtained under normal conditions, and R(k) is the logarithmic reconstruction probability of the kth frame. The anomaly score S(k) of the kth frame is the final output of the scoring layer and is monitored.

[0078] The encoder and decoder of VAE contain LSTM layers and dense layers to process time series data.

[0079] The frames are normalized using a fully connected network so that the chirp signal has a larger peak and the noise has a smaller peak.

[0080] The amplitude of the reflected signal generated by the abnormal point in the test cable is usually smaller than the chirp signal that exists in the normal state, such as the reference signal and the reflected signal from the cable joint and terminal. Therefore, if the windowed TFDR result is directly used as the input of the VAE without normalization, even if the VAE cannot recover the signal, the reconstruction error of the abnormal reflected signal is small due to the smaller reflection. On the contrary, if all input frames are normalized to have the same peak value, the VAE cannot recover the added random noise, and the reconstruction error will increase even in the normal state,

[0081] As mentioned above, the purpose of the normalization layer is to set the peak value high if the data frame contains a lot of chirp signals, and to set the peak value low if it does not contain chirp signals. Figure 2 As shown in Figure 1, the normalization layer includes a fully connected network to estimate the proportion of the chirp signal in the input frame. Before entering the network, the peak values ​​of all frames are initialized to 1. The k-th output frame x of the normalization layer N (k) is the k-th input frame x(k) multiplied by the normalized weight W in the network kThe training set of the normalized network is constructed by sweeping the center time t0 and phase of the reference signal, and the target variable of the kth training frame is Determine as follows:

[0082]

[0083] Where Ck is the length of the chirp signal in the kth training frame, TD is the time length of the reference signal,

[0084] In order to make the normalization layer achieve the expected effect in the actual monitoring environment, some Gaussian noise is included in the training set of the normalization layer. The target variable of the kth training frame is Determined as (5), but the target variable of the Gaussian noise frame is fixed to 0. In addition, from the perspective that the normalization layer is a matched filter of the chirp signal, the window width that determines the frame size is most often used as the TD of the reference signal. If the window width is reduced, the spatial resolution will increase. However, due to the degradation of the performance of the matched filter, the anomaly detection performance will also decrease.

[0085] The structure of the encoder and decoder is determined according to the type of input data. If the input is an image, convolutional layers and deconvolutional layers are usually used for feature extraction processing. Using the proposed method, even if the signals look the same, it is possible to determine whether the test cable is normal or abnormal based on the position of the chirp signal. Therefore, the design of the encoder and decoder combines LSTM and dense layers. The LSTM layer can affect the anomaly detection result of the next input frame through the hidden state, while the dense layer is used to transform the hidden state into the desired output, such as the mean and variance of z and xN

[0086] Example 1 Abnormal Detection Process

[0087] Figure 3 The control cable anomaly detection flow chart is shown. Once the cable to be monitored is determined, the reference signal for TFDR can be optimally designed. TFDR is performed by injecting the designed reference signal into the cable, and the TFDR results are continuously input into the LSTM-based VAE (LSTM-VAE). If the LSTM-VAE has not been trained before, the network should be initially trained with newly measured data under the assumption that the test cable is normal. After training, Rn is created using the reconstruction probability of normal data, and the maximum and minimum values ​​of Rn are found to calculate the anomaly score. When monitoring the cable in real time, inputting the measured data into the trained LSTM-VAE will output the anomaly score for each frame. If the score exceeds 1, an alarm is immediately issued, indicating that an anomaly has occurred.

[0088] Example 2 Experimental device

[0089] A test setup was built to verify the performance of the proposed anomaly detection technique. The test setup consists of an arbitrary waveform generator (AWG) to generate the TFDR reference signal, a digital phosphor oscilloscope (DPO) to acquire the reflected signal, and a signal processing system to run the anomaly detection algorithm in real time. All instruments transmit data and control signals to each other through the general purpose interface bus (GPIB) to monitor the test cable in real time. The test cable is a seven-core I&C control cable commonly used in nuclear power plants. Compared with other types of control cables, I&C cables are selected for their convenience in installation, connection through connectors, and fault simulation. Since the signal interference caused by the multiple cores of I&C cables makes it difficult to analyze the reflected signal, it is suitable for verifying the performance of the proposed method. Note that the reflection measurement method can be used for all other types of control cables, so the proposed anomaly detection technique is not limited to I&C cables.

[0090] Three cable states, namely normal, abnormal A and abnormal B, were evaluated in the experiment.

[0091] For example, under normal conditions, 20-meter and 15-meter cables are connected through terminal blocks.

[0092] The abnormal A state corresponds to a cable defect. A reflected signal can be generated by connecting a resistor between the cores at 7.5 meters of the 15-meter cable. If the test cable is physically damaged to simulate a defect, the cable cannot be restored to a normal state. In the present invention, since the experiment involves the case where the cable returns to a normal state after an intermittent fault, it is reasonable to simulate the defect by a resistor.

[0093] For abnormal state B, a defect is simulated by connecting a resistor between the core wires of the terminal block. State B corresponds to the situation of insulation failure of the power cable joint or poor connection of the I&C cable junction box.

[0094] Both cable abnormal states A and B correspond to changes in the local impedance of the test cable. However, state A (the appearance of a new reflection signal) is easier to detect than state B (the existing reflection signal changes due to a fault). For these two abnormal conditions, the scenario is configured such that the cable system remains in a normal state within the first 20 seconds, switches to an abnormal state within 20 to 30 seconds, and returns to a normal state after 30 seconds. It is known that the detection delay required for abnormality detection varies depending on the application of the control cable. Therefore, the present invention sets the abnormal state to 10 seconds, which is sufficient to determine the time required for abnormality detection.

[0095] In the experiment, a parallel fault is simulated by connecting a resistor between the cable cores. When the resistance of the parallel fault decreases, the amplitude of the reflected signal increases due to the increase in the reflection coefficient. In the proposed method, the peak values ​​of all frames are set to 1 before entering the normalization layer. Therefore, the magnitude of the reflected signal does not affect the output of the LSTM-VAE. Since the main purpose of the proposed method is to detect anomalies, any resistor can be used to create a parallel fault.

[0096] Example 3 Initial network training under normal conditions

[0097] Since there is no trained LSTM-VAE network in the initial measurement, the network should be trained using the TFDR results measured in the normal state. The present invention trains the network based on 20 TFDR results measured in the normal state. Since the normal state of the cable / connector remains unchanged in the scene, all TFDR results are similar except for the measurement noise. In order to make the network robust in the actual monitoring environment, we add various Gaussian noises to the measured TFDR results, thereby increasing the training data. With the diversification of training data, the robustness of the network to measurement noise will also be enhanced.

[0098] However, in some cases, the cable system may have various normal states: cooling superconducting cables, changing the network topology of transmission lines, changing the I&C cable terminal mode, etc. Since the proposed anomaly detection technique is based on VAE, its performance will improve with the occurrence of more than one normal state. If the user manually monitors the cable system installed in a place where the surrounding environment changes, the current TFDR results must be compared with various normal state data. In contrast, the proposed method can automatically estimate the state of the cable using LSTM-VAE.

[0099] In the proposed LSTM-VAE model, the user only needs to determine two parameters: the number of hidden units of the LSTM layer and the dimensionality of the latent space. The number of hidden units determines the maximum amount of information that the LSTM layer can remember between time steps, while the dimensionality of the latent space limits the amount of information that can be stored in the bottleneck. What these two parameters have in common is that the efficiency of the proposed network decreases as the number of parameters increases. As the number of hidden units increases, the possibility of overfitting also increases. Similarly, as the dimensionality of the latent space increases, the network loses its dimensionality reduction ability. Therefore, when determining these two parameters, you can start with a small value and then gradually increase it until you can successfully reconstruct the normal data. In addition, if there is enough normal data, you can also use a validation set to determine the model parameters.

[0100] Example 4 Fault scenario A: Cable defect

[0101] In fault scenario A, after the cable is monitored for a period of time, a defect appears in the cable. After a period of time, the defect is removed and the cable returns to normal. Before monitoring, the reference signal of TFDR is designed in a normal state taking into account the required resolution, propagation characteristics and total cable length. The present invention uses a reference signal with a CF of 40 MHz, a frequency BW of 80 MHz and a TD of 80 ns in all cases.

[0102] In order to confirm the reconstruction error under normal and abnormal conditions, two representative TFDR results can be observed in detail. The TFDR results measured under normal and abnormal conditions A are input into LSTM-VAE and then compared with the input data x N , reconstruct data and reconstruction error. Reconstructed data is one of the sampling results output by the decoder, and the reconstruction error is defined as the original data x N and reconstructing data LSTM-VAE can be normalized from x N Refactoring In x N At larger points, the reconstruction error is slightly larger, while the error caused by dimensionality reduction is evenly distributed. On the contrary, since the reflected signal is newly detected at the fault of the test cable in the abnormal A state, it cannot be fully reconstructed, but the error is increased.

[0103] In the anomaly detection results of fault scenario A, the TFDR time series results measured under normal and abnormal A can be described. Under abnormal A, the reflection signal at the cable anomaly can be clearly observed, showing the proposed anomaly score, with a threshold of 1, and all calculated scores less than 1 are mapped to 0.

[0104] Example 5 Joint Defects

[0105] Fault scenario B deals with joint defects. As with scenario A, the performance of LSTM-VAE is observed through the TFDR results measured in normal and abnormal B. The results in the normal state are the same as those in scenario A. In the abnormal case, the reflected signal from the terminal board is slightly different because the resistor connected to the terminal board changes the amplitude and phase of the reflected signal, and LSTM-VAE cannot fully reconstruct the reflected signal from the terminal board, so a higher reconstruction error occurs at the location where the reflected signal from the terminal board appears. TFDR time series results measured in normal and abnormal B. It is obvious that due to the connection of the resistor to the terminal board, the amplitude and phase of the reflected signal have changed. The proposed anomaly score with a threshold of 1 is displayed.

[0106] The above technical solutions only reflect the preferred technical solutions of the technical solutions of the present invention. Some changes that may be made to certain parts thereof by technicians in this technical field all reflect the principles of the present invention and fall within the protection scope of the present invention.

Claims

1. A control cable anomaly detection model and method based on deep learning, characterized in that: Build the LSTM-VAE model and use the reconstruction error between VAE input and output. Input the TFDR results into LSTM-VAE frame by frame, and calculate the anomaly score of each frame based on the logarithmic reconstruction probability. When the anomaly score exceeds 1, the system will automatically issue an alarm to indicate that an anomaly has occurred. The VAE is a variational autoencoder; The TFDR is an improved fault location technology; LSTM stands for long short-term memory network; The LSTM-VAE is a hybrid model that combines LSTM and VAE, specifically a control cable anomaly detection network architecture.

2. According to the control cable anomaly detection model and method based on deep learning in claim 1, it is characterized in that: The network architecture of the LSTM-VAE includes a data window, a normalization layer, an encoder, a latent variable sampling, a decoder, and a scoring layer.

3. According to the control cable anomaly detection model and method based on deep learning in claim 1, it is characterized in that: The encoder and decoder of VAE contain LSTM layers and dense layers to process time series data.

4. According to the control cable anomaly detection model and method based on deep learning in claim 2, it is characterized in that: The frames are normalized using a fully connected network.

5. According to the control cable anomaly detection model and method based on deep learning in claim 2, it is characterized in that: The structure of the encoder and decoder is determined according to the type of input data. If the input is an image, convolutional layers and deconvolutional layers are usually used for feature extraction processing. Using the proposed method, even if the signals look the same, it is possible to determine whether the test cable is normal or abnormal based on the position of the chirp signal. Therefore, the design of the encoder and decoder combines LSTM and dense layers. The LSTM layer can affect the anomaly detection result of the next input frame through the hidden state, while the dense layer is used to convert the hidden state into the desired output, such as the mean and variance of z and xN.

Images