Cable defect positioning method and system based on convolutional neural network, and electronic device

Abstract

A cable defect positioning method and system based on a convolutional neural network, and an electronic device. The method comprises: establishing a defective cable transmission model for a cable to be tested having n defects, and on the basis of the defective cable transmission model, acquiring a total reflection signal after an initial incident signal is reflected by the n defects, wherein n is a positive integer (S11); taking the total reflection signal as an input of a trained first convolutional neural network, and extracting time domain features corresponding to the n defects by the first convolutional neural network (S12); and establishing a probability distribution model for the time domain features, and on the basis of the probability distribution model, respectively obtaining first defect positions in one-to-one correspondence with the n defects (S13). By using the method, the accuracy of cable defect positioning can be improved.

Description

Cable defect location method, system and electronic equipment based on convolutional neural network Technical Field

[0001] The present invention relates to the technical field of cable defect detection, and in particular to a cable defect location method, system and electronic equipment based on a convolutional neural network. Background Art

[0002] With the rapid development of the national economy, the cable industry is booming, and cables are increasingly used in power systems. Due to the complex environment in which submarine cables are laid, they are susceptible to factors such as seabed structures, natural disasters, and ship anchors, leading to faults that degrade the cable's overall insulation performance. If not discovered and addressed promptly, these faults can lead to breakdown accidents. Therefore, submarine cable insulation condition assessment and defect / fault location technology have become key areas of focus for power operation and maintenance departments.

[0003] Time-frequency domain reflectometry (TFDR) is a defect location method based on time-frequency analysis, which aims to provide complete information about signal variations in both the time and frequency domains. However, due to the inherent characteristics of signals, it is impossible to completely eliminate ambiguity. Specifically, when a signal has a relatively high time resolution, its frequency resolution may be low. This means that it is impossible to simultaneously obtain complete information about the signal in both the time and frequency domains. Therefore, time-frequency analysis has a certain degree of uncertainty. Due to this uncertainty, the characteristics of the reflected signal obtained will vary when different time-frequency analysis windows are selected, which affects the accuracy of the location results.

[0004] Since the uncertainty of time-frequency analysis is inevitable, the traditional TFDR method uses the Wigner distribution time-frequency analysis method to process the measured input-reflection superposition signal to obtain relevant information about the signal. However, the time-frequency energy distribution diagram of the signal obtained after processing is affected by cross-terms. Although existing technologies include methods such as smoothing windowing or modal decomposition to suppress cross-terms, they will cause certain distortions in the final signal. In addition, traditional defect location methods such as FDR (Frequency domain reflectometry) or TDR (Time domain reflectometry) rely on human experience and subjective judgment to determine whether there is a defect, the defect location, and the severity of the defect in the final location curve. The resulting location results are highly random and have low applicability. From the above analysis, it can be seen that the current TFDR method not only has the problem of cross-term interference when used for cable defect location, but also relies on human experience to judge the defect location.

[0005] Summary of the Invention

[0006] The purpose of the present invention is to address the deficiencies of the above-mentioned prior art and propose a cable defect location method, system and electronic equipment based on convolutional neural network, which can improve the accuracy of cable defect location.

[0007] In a first aspect, the present invention provides a cable defect location method based on a convolutional neural network, comprising:

[0008] Establishing a defective cable transmission model for a cable to be tested having n defects, and obtaining a total reflected signal of an initial incident signal after reflection from the n defects according to the defective cable transmission model; wherein n is a positive integer;

[0009] Using the total reflection signal as input to a trained first convolutional neural network, the first convolutional neural network extracts time domain features corresponding to the n defects;

[0010] A probability distribution model is established for the time domain feature, and first defect positions corresponding to n defects are obtained according to the probability distribution model.

[0011] The present invention obtains the total reflection signal of the cable under test after reflection from multiple defects. Since the obtained total reflection signal contains the reflection signals after reflection from all defects, the convolutional neural network performs a unified convolution operation on the reflection signals of multiple defect positions, and can automatically obtain all corresponding time domain features for each reflection signal at one time, without relying on human experience, thereby improving the accuracy of cable defect positioning and improving the efficiency of defect positioning. Moreover, the convolutional neural network obtains the defect position according to the corresponding probability distribution model of the input total reflection signal, and converts the traditional two-classification problem of whether there is a defect at each position into a probability problem, thereby further improving the accuracy of cable defect positioning.

[0012] Furthermore, the first convolutional neural network is obtained, including:

[0013] According to the multi-defect cable simulation model, first time domain signals at multiple defect locations are obtained respectively by time-frequency domain reflectometry, and Gaussian distributions are added as training labels at the multiple defect locations respectively;

[0014] The obtained multiple first time domain signals are respectively used as inputs of the one-dimensional initial second convolutional neural network, and the corresponding second defect positions are output in sequence. Training is performed according to the second defect positions and the training labels to obtain a trained first convolutional neural network.

[0015] The present invention obtains training data through the traditional time-frequency domain reflection method, and supervises the initial convolutional neural network for each defect position. After learning effective features based on the traditional time-frequency domain reflection method, the cable to be tested is tested. Therefore, the present invention combines the time-frequency domain reflection method with the traditional method in the training stage to obtain a time domain signal with defect position characteristics, and uses the signal as the input of the convolutional neural network for training, which not only ensures the accuracy of defect location, but also improves the efficiency of defect location.

[0016] Furthermore, establishing a probability distribution model for the time domain feature and obtaining first defect positions corresponding to n defects according to the probability distribution model include:

[0017] The first convolutional neural network extracts the first time domain feature corresponding to each reflected signal, outputs several consecutive first probabilities corresponding to the cable segment to be tested, and takes the cable sub-segment to be tested corresponding to the largest second probability among the several first probabilities as the first defect location, thereby obtaining the first defect locations corresponding to n defects one by one.

[0018] The present invention uses a convolutional neural network to process the input total reflection signal and obtain the defect position according to the corresponding probability distribution model. In fact, the probability of a defect occurring at each position can be obtained, and the traditional binary classification problem of whether a defect exists at each position is converted into a probability problem of a defect occurring at each position. The position with the highest probability of a defect is located as the defect position, thereby improving the accuracy of cable defect positioning; in addition, the position where a defect is about to occur can also be obtained, except that the sub-segment of the cable to be tested with the highest probability in the cable segment to be tested is used as the defect position.

[0019] Furthermore, obtaining a total reflected signal of an initial incident signal after being reflected by the n defects according to the defective cable transmission model includes:

[0020] According to the defective cable transmission model, n first reflected signals of an initial incident signal reflected by n defect heads and second reflected signals of the n first reflected signals reflected by the end of the cable to be tested are obtained;

[0021] The n first reflected signals and the second reflected signal are accumulated to obtain a total reflected signal, so that the transfer function of the total reflected signal obtains a second time domain signal, and the second time domain signal is used as the input of the first convolutional neural network.

[0022] Further, obtaining the incident signal includes:

[0023] When the length of the cable to be tested is less than a first length threshold or greater than a second length threshold, the center frequency of the initial first reference signal is adjusted to a preset center frequency to obtain a second reference signal, and the second reference signal is used as the incident signal.

[0024] Furthermore, the step of establishing a defective cable transmission model for a cable to be tested having n defects includes:

[0025] For a cable to be tested with n defects, a cable equivalent distributed parameter circuit model is established, the cable characteristic impedance is obtained based on the cable equivalent distributed parameter circuit model, and a defective cable transmission model is established based on the cable characteristic impedance.

[0026] Furthermore, for a cable to be tested having n defects, establishing a cable equivalent distributed parameter circuit model includes:

[0027] For a cable to be tested with n defects, the cable unit length parameters are obtained, and a cable equivalent distributed parameter circuit model is established based on the cable unit length parameters; wherein the cable unit length parameters include: resistance, inductance, conductance and capacitance per unit length of the cable.

[0028] Furthermore, obtaining the cable characteristic impedance according to the cable equivalent distributed parameter circuit model includes:

[0029] According to Kirchhoff's law, the voltage and current at the cable head end in the cable equivalent distributed parameter circuit model are obtained, and based on the voltage and current at the cable head end, the reflection coefficient and input impedance of the cable segment to be measured from the cable end are obtained.

[0030] In a second aspect, the present invention provides a cable defect location system based on a convolutional neural network, comprising: a reflection module, a time domain feature acquisition module and a positioning module; wherein,

[0031] The reflection module is used to establish a defective cable transmission model for a cable to be tested having n defects, and obtain a total reflected signal of an initial incident signal after being reflected by the n defects according to the defective cable transmission model; wherein n is a positive integer;

[0032] The time domain feature module is configured to use the total reflection signal as input to a trained first convolutional neural network, wherein the first convolutional neural network extracts time domain features corresponding to the n defects;

[0033] The positioning module is used to establish a probability distribution model for the time domain features, and obtain the first defect positions of n defects in a one-to-one correspondence according to the probability distribution model.

[0034] In a third aspect, the present invention provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the cable defect location method based on convolutional neural network as described in the first aspect is implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] FIG1 is a schematic diagram of a flow chart of a cable defect location method based on a convolutional neural network according to an embodiment of the present invention;

[0036] FIG2 is a diagram of a cable equivalent distributed parameter circuit model provided by an embodiment of the present invention;

[0037] FIG3 is a diagram of a cable equivalent distributed parameter circuit model provided by an embodiment of the present invention;

[0038] FIG4 is a schematic diagram of signal transmission of a defect-free cable transmission model provided by an embodiment of the present invention;

[0039] 5 is a schematic diagram of signal transmission of a defective cable transmission model provided by an embodiment of the present invention;

[0040] 6 is a time domain waveform diagram of a reference signal corresponding to different center frequencies provided by an embodiment of the present invention;

[0041] 7 is a time domain waveform diagram of a reference signal corresponding to different pulse widths provided by an embodiment of the present invention;

[0042] FIG8 is a time domain waveform diagram of a reference signal corresponding to different frequency bandwidths provided by an embodiment of the present invention;

[0043] FIG9 is a schematic diagram of the structure of a convolutional neural network provided by an embodiment of the present invention;

[0044] FIG10 is a schematic structural diagram of a cable defect location system based on a convolutional neural network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0046] 1 is a flow chart of a cable defect location method based on a convolutional neural network according to an embodiment of the present invention, which includes steps S11 to S13, specifically:

[0047] Step S11: establishing a defective cable transmission model for a cable to be tested having n defects, and obtaining a total reflected signal of an initial incident signal after being reflected by the n defects according to the defective cable transmission model; wherein n is a positive integer.

[0048] Specifically, a defective cable transmission model is established for a cable to be tested with n defects, including sub-steps S111 to S113. Referring to FIG2 , it is a flow chart of establishing a defective cable transmission model according to an embodiment of the present invention, specifically:

[0049] Sub-step S111: Establishing a cable equivalent distributed parameter circuit model for a cable under test having n defects. This includes obtaining cable unit length parameters for the cable under test having n defects, and establishing the cable equivalent distributed parameter circuit model based on the cable unit length parameters; wherein the cable unit length parameters include resistance, inductance, conductance, and capacitance per unit length of the cable.

[0050] Figure 3 shows an equivalent distributed parameter circuit model for a cable according to an embodiment of the present invention. The figure shows a coaxial power cable, a typical transmission line. Transmission line theory can be used to describe the signal transmission characteristics of a cable. Under high-frequency conditions, the parameters of circuit components change. According to transmission line theory, an equivalent distributed parameter circuit model should be used to represent the cable circuit.

[0051] As a preferred embodiment, according to the equivalent distributed parameter circuit model shown in FIG3 , the high-frequency current flowing through the conductor under high-frequency conditions will produce a skin effect, and the cable unit length parameters are calculated based on the skin effect.

[0052] As a preferred embodiment, the resistance, inductance, conductance and capacitance per unit length of the cable can be expressed as:

[0053] Where, ω = 2πf is the signal angular frequency; r c is the cable core radius; r s is the inner radius of the shielding layer; ρ c is the cable core resistivity; ρ s is the resistivity of the shielding layer; μ0 is the magnetic permeability of vacuum; σ is the conductivity of the dielectric; ε r is the relative dielectric constant of the insulating material; ε0 is the dielectric constant of vacuum; R (Ω / m), L (H / m), G (S / m) and C (F / m) are the resistance, inductance, conductance and capacitance per unit length of the power cable, respectively.

[0054] Sub-step S112: Obtain the characteristic impedance of the cable according to the equivalent distributed parameter circuit model of the cable, including: according to Kirchhoff's law, obtain the voltage and current at the head end of the cable in the equivalent distributed parameter circuit model of the cable, and according to the voltage and current at the head end of the cable, obtain the reflection coefficient and input impedance of the cable section to be measured from the end of the cable.

[0055] As a preferred embodiment, according to the equivalent distributed parameter circuit model shown in Figure 3, according to Kirchhoff's law, the derivative of voltage with respect to the defect position and the derivative of current with respect to the defect position can be obtained.

[0056] As a preferred embodiment, the derivative of voltage with respect to the defect position and the derivative of current with respect to the defect position can be respectively expressed as:

[0057] where,

[0058] γ0 is the propagation constant of the power cable itself; Z0 is the characteristic impedance of the power cable itself.

[0059] As a preferred embodiment, the propagation constant and the characteristic impedance of the power cable itself can be respectively expressed as:

[0060] As a preferred embodiment, the propagation constant can also be expressed as:

[0061] where, α is the attenuation constant of the cable; β is the phase constant of the cable; v is the propagation speed of the electromagnetic wave signal in the cable.

[0062] As a preferred embodiment, for a uniform transmission line, under high-frequency conditions, that is, R << jωL, G << jωC, then the characteristic impedance of the power cable itself is approximately a constant and can be expressed as:

[0063] where, c0 is the propagation speed of the electromagnetic wave in vacuum, μ r is the relative magnetic permeability of the insulating material, usually μ r = 1.

[0064] As a preferred embodiment, when the frequency of the electromagnetic wave signal in the cable is high enough, the propagation speed of the electromagnetic wave signal can be regarded as a constant value, and the value of the propagation speed of the electromagnetic wave signal is only related to the dielectric properties of the insulating material in the cable, that is, the cable insulation state can be reflected in the change of the propagation speed of the electromagnetic wave signal.

[0065] As a preferred embodiment, for a cable with a length of l, when x = l, the following can be obtained:

[0066] At the same time, the end load impedance can be expressed as:

[0067] Therefore, the voltage and current signal expressions at the distance x'=lx from the cable end can be expressed as:

[0068] As a preferred embodiment, according to the voltage and current signal expressions at a distance x'=lx from the cable end, when x'=0, the cable end reflection coefficient is recorded as Γ. The cable end reflection coefficient can be expressed as:

[0069] As a preferred embodiment, according to the expression of the cable end reflection coefficient, the end reflection coefficient is only related to the load impedance and the cable characteristic impedance. L = Z0, the end reflection coefficient Γ = 0; when Z L = 0, the end reflection coefficient Γ = -1; when Z L When =∞, the end reflection coefficient Γ=1.

[0070] Sub-step S113: establishing a defective cable transmission model according to the cable characteristic impedance.

[0071] It's worth noting that for a defect-free cable, when a pulse signal is injected into the cable headend via a signal generator, the signal is reflected at the cable end. Figure 4 illustrates the signal transmission process for a defect-free cable transmission model according to an embodiment of the present invention. When using the traveling wave method to locate defects, only a single catadioptric reflection is required. To simplify the calculation, only this single catadioptric reflection process is used for derivation.

[0072] It is worth noting that in sub-step S113, the cable characteristic impedance is obtained, and the cable end reflection coefficient is represented by the cable characteristic impedance. If the cable end reflection coefficient is recorded as ρ, it can be expressed as:

[0073] At this time, the reflection coefficient and input impedance of the defect-free cable head end can also be expressed as:

[0074] As a preferred embodiment, according to the defect-free cable transmission model of FIG4 , under the premise of considering only single catadioptric reflection, the signal is reflected only once at the end of the cable. The transfer function of the defect-free cable signal transmission model can be obtained, which can be expressed as:

[0075] It is worth noting that when a local defect occurs in a cable, an impedance discontinuity is formed, causing changes in the local electrical parameters of the cable. According to traveling wave theory, a series of refractions and reflections occur when a traveling wave propagates through a cable, and the reflection coefficient at the cable head end also changes accordingly.

[0076] Table 1 Refraction and reflection coefficient

[0077] See Figure 5, which is a schematic diagram of signal transmission in a defective cable transmission model according to an embodiment of the present invention. In the figure, when a local defect occurs in the cable, the characteristic impedance and propagation constant of that section will change. When the signal transmission encounters the defective section, due to the impedance mutation, a refraction-reflection phenomenon will occur. The reflection coefficients at the head end of each defect from defect 1 to defect n are ρ1 to ρ n , the reflection coefficient at the end of the cable is ρ, see Table 1, which is the refraction-reflection coefficient at the corresponding position.

[0078] According to the multi-defect cable model shown in Figure 5, when there are n defects, the total cable reflection signal can be divided into the sum of the reflection signal at each defect and the reflection signal at the cable end.

[0079] As a preferred embodiment, the reflection signal at defect n can be expressed as:

[0080] As a preferred embodiment, when there are n defects in the cable, the end reflection signal can be expressed as:

[0081] Among them, l (n+1)1 =l.

[0082] As a preferred embodiment, based on the reflection signal at point n and the reflection signal at the cable end, the transfer function G(ω) of the multi-defect cable model and the time domain positioning signal can be expressed as follows:

[0083] y(t)=IFFT(G(ω)·FFT(s(t))),

[0084] Where s(t) is the incident signal, y(t) is the time domain positioning signal, FFT(·) is the Fourier transform, and IFFT(·) is the inverse Fourier transform.

[0085] According to the defective cable transmission model, a total reflection signal of an initial incident signal after being reflected by the n defects is obtained, including: according to the defective cable transmission model, n first reflection signals of the initial incident signal reflected by the head ends of the n defects are obtained, and a second reflection signal of the n first reflection signals reflected by the end of the cable to be tested is obtained; the n first reflection signals are accumulated with the second reflection signal to obtain a total reflection signal, so that a transfer function of the total reflection signal obtains a second time domain signal, and the second time domain signal is used as the input of the first convolutional neural network.

[0086] Obtaining the incident signal includes: when the length of the cable to be tested is less than a first length threshold or greater than a second length threshold, adjusting the center frequency of an initial first reference signal to a preset center frequency, obtaining a second reference signal, and using the second reference signal as the incident signal.

[0087] It is worth noting that when the length of the tested cable is short, the center frequency f0 of the reference signal should be set higher to avoid waveform aliasing during the propagation of the signal, thereby ensuring that the defect can be completely distinguished at this time; at the same time, when the cable length is long, due to the large attenuation of the high-frequency component of the signal when propagating in the power cable, if a reference signal with a higher center frequency is selected, the energy reflected at the end will be extremely low. If the defect is far away from the test end at this time, it is very likely that it cannot be detected due to energy loss. Therefore, for long-distance cables, in order to reduce the impact of high-frequency signal loss, it is necessary to select a lower center frequency to ensure that the entire cable can be detected. According to the simulation, the present invention selects a suitable preset center frequency to adjust the initial reference signal.

[0088] As a preferred embodiment, based on the characteristics of signal distribution in the time-frequency domain, a Gaussian envelope linear frequency modulation signal is selected as the initial first reference signal.

[0089] As a preferred embodiment, the first reference signal can be expressed as:

[0090] Among them, α controls the duration of the signal, that is, the width of the Gaussian distribution, which is inversely proportional to the signal pulse width; α and β jointly affect the frequency bandwidth of the signal and are linearly related to the frequency bandwidth; ω0 is the center angular frequency of the signal, which can be expressed by the center frequency f0 as ω0=2πf0; t0 is the center time of the signal.

[0091] It is worth noting that the present invention uses the adjusted second reference signal obtained through TFDR as the time domain signal to serve as the input of the first convolutional neural network. Unlike TDR and FDR, the reference signal of TFDR has many advantages in signal processing. Since the reference signal has significant characteristics such as linear frequency increase, the reflected signal can be easily detected in the time and frequency domain. In addition, the reference signal has phase information, and the propagation characteristics of the cable can be obtained based on the phase information, and then the relationship between defects and categories can be established, which cannot be obtained from pulse or step signals. In addition, since the frequency bandwidth of the reference signal is relatively small, the specific frequency band suitable for use can be selected by freely changing the center frequency and frequency bandwidth according to information such as the actual length of the cable being tested.

[0092] As a preferred embodiment, the signal expression in the frequency domain obtained by performing Fourier transform processing on the incident signal s(t) is:

[0093] Among them, the time center of the reference signal on the time axis is t s And the duration of a single pulse width T s Can be expressed as: s =∫t|s(t)| 2 dt=t0,

[0094] Among them, the frequency center ω s and frequency bandwidth b s Can be expressed as: s =∫ω|S(ω)| 2 dt=ω0,

[0095] Among them, the parameters α and β are expressed as:

[0096] During actual testing, factors such as the length and material of the cable under test must be considered. The signal center frequency, pulse duration of a single signal, and signal frequency bandwidth must be adjusted to select an appropriate input signal. This ensures optimal resolution for defect identification and avoids waveform aliasing, ensuring that the test results fully reflect the location of local defects. The following discusses the effects of center frequency, pulse width, and frequency bandwidth on the reference signal's time-domain waveform. Recommendations for selecting reference signal parameters are provided for local defect location diagnosis in power cables of varying lengths and insulation conditions.

[0097] In order to explore the influence of the center frequency on the reference signal waveform, the reference signal parameters are set: the center frequency is set to f0 = 50MHz, f0 = 100MHz, f0 = 150MHz; the signal pulse width Ts =5ns, frequency bandwidth b s =100MHz, center time t0 = 50ns. According to the expressions of α and β, α and β are only affected by T s and b s Therefore, the values ​​of α and β do not change in the above case, and we can get α=2×10 16 , due to 2b s T s =1, we can get β = 0. Referring to FIG6 , it is a time domain waveform diagram of reference signals corresponding to different center frequencies provided by an embodiment of the present invention.

[0098] As shown in Figure 6, changes in the center frequency have little effect on the signal pulse width; they only affect the waveform amplitude and spatial resolution. As the center frequency increases, the signal amplitude gradually increases, and the waveform spatial resolution improves. Therefore, when the test cable is short, a higher center frequency should be set to ensure higher spatial resolution, allowing the waveform to more fully reflect the defect location and condition and avoid waveform aliasing during signal propagation. Conversely, when the test cable is long, high-frequency signals experience significant attenuation during cable propagation. Selecting a higher center frequency will result in significant energy loss during signal propagation, making it impossible to effectively detect the defect location. Therefore, when testing long cables, a lower center frequency should be selected to minimize signal energy loss and ensure detection of the entire cable. Although high-frequency signals experience significant attenuation during cable propagation, a higher center frequency means higher spatial resolution of the signal waveform and greater sensitivity to localized, subtle defects in the cable. Therefore, in actual cable defect testing, the signal center frequency can be adjusted flexibly based on the cable's conditions.

[0099] In order to explore the influence of signal pulse width on the reference signal waveform, the reference signal parameters are set: the signal pulse width is set to T s =4ns, T s =6ns, T s =8ns; center frequency f0 = 100MHz, frequency bandwidth b s =100MHz, center time t0 = 50ns. Calculation shows that the values ​​of α and β in the above three cases are: α = 7.8 × 10 15 β=1.9×10 16 , α=1.4×10 16 β=9.2×10 15 , α=7.8×10 15 β=9.8×10 15 7 , which is a time domain waveform diagram of reference signals corresponding to different pulse widths provided by an embodiment of the present invention.

[0100] As shown in Figure 7, the signal pulse width is the parameter that controls the signal width. s As the signal amplitude increases, the signal width also increases, and the number of waveform oscillations increases. When the signal pulse width is small, the number of waveform oscillations also decreases. If the signal pulse width is too small, the waveform will be distorted. See Figure 7(b) and 7(c). When the parameters are selected so that 2b s T s When ≥1, the signal waveform is no longer symmetrical, that is, the signal exhibits frequency variation in the time domain, and the oscillation on the right side of the main peak is more intense than that on the left side.

[0101] In order to explore the influence of signal frequency bandwidth on the reference signal waveform, the reference signal parameters are set: the signal frequency bandwidth is set to b s =100MHz, b s =200MHz, b s =300MHz; center frequency f0 = 100MHz, frequency bandwidth T s =5ns, center time t0 = 50ns. Calculation shows that the values ​​of α and β in the above three cases are: α = 2×10 16 , β=0,α=2×10 16 β=3.5×10 16 , α=2×10 16 β=5.7×10 16 Referring to FIG8 , it is a time domain waveform diagram of reference signals corresponding to different frequency bandwidths provided by an embodiment of the present invention.

[0102] As shown in FIG8 , when the signal pulse width is fixed, the signal width remains basically unchanged. As the signal frequency bandwidth increases, the frequency variation of the signal waveform in the time domain becomes more obvious, that is, the signal distortion becomes more serious.

[0103] In summary, in order to better locate and diagnose local defects in power cables, the pulse width and frequency bandwidth of the reference signal should be kept as close as possible to 2b. s T s=1, so that the reference signal waveform is not distorted while maintaining good symmetry. For short cable samples, the signal pulse width should not be set too large, so that the signal waveform can fully reflect the defect location and status information. At the same time, a higher center frequency can be set to improve the signal time domain resolution. The signal frequency bandwidth can be set according to the signal pulse width and the principle of maintaining good waveform symmetry. For long cable samples, the signal pulse width can be set slightly larger. Since high-frequency signals are severely attenuated in cables, the center frequency setting should not be too large when testing long cables, and the frequency bandwidth can be adjusted according to the above parameters. In summary, in actual cable local defect location testing, the TFDR reference signal can be set according to the on-site cable length and insulation status, etc., and has high adaptability.

[0104] As a preferred embodiment, the signal pulse width and frequency bandwidth of the second reference signal should be kept as close as possible to 2b s T s =1, so that the center frequency of the first reference signal incident on the cables having a length greater than the second threshold value and the cables having a length less than the second threshold value is adjusted.

[0105] It is worth noting that the present invention adopts a multi-defect cable simulation model, uses the TFDR reference signal as the time domain signal, and uses the time domain signal as the input of the one-dimensional initial first convolutional neural network for training to obtain a trained second convolutional neural network.

[0106] In other words, based on the multi-defect cable simulation model, the TFDR method is used to obtain the time domain signals of different defect locations as the input of the one-dimensional convolutional neural network. The neural network is trained with these input data to learn the high-level features of the original time domain signals.

[0107] It is worth noting that convolutional neural network is a typical deep learning model, which is widely used in the field of image processing. The overall model structure of CNN is similar to that of MLP, and is also composed of an input layer, a hidden layer and an output layer. The difference is that the hidden layer of CNN is composed of a stack of convolutional layers and pooling layers, which are used to extract and simplify data features. The output layer is composed of a fully connected layer, which is mainly responsible for making final decisions based on the features extracted by the hidden layer. In the present invention, since the input data is a time domain signal collected by the TFDR method, and there is a problem of attenuation when the signal is transmitted in the cable, the amplitude difference between the end of the reflected signal and the amplitude of other positions is too large, and different operations need to be performed on different positions to extract corresponding features. However, different operations are performed on different positions to extract features, which will cause the obtained signal time-frequency energy distribution diagram to have cross-term effects. The present invention uses convolution operation to perform the same operation on data at any position. Therefore, a single connection layer is introduced after completing the multi-layer convolution operation to extract the features of the corresponding position, and finally achieve positioning, which can avoid the influence of cross-terms.

[0108] The second convolutional neural network is trained to obtain a trained first convolutional neural network, including: according to a multi-defect cable simulation model, first time domain signals at multiple defect positions are obtained respectively by time-frequency domain reflection method, and Gaussian distributions are added as training labels at the multiple defect positions respectively; the obtained multiple first time domain signals are respectively used as inputs of a one-dimensional initial second convolutional neural network, the corresponding second defect positions are output in sequence, and training is performed according to the second defect positions and the training labels to obtain a trained first convolutional neural network.

[0109] The present invention obtains training data through the traditional time-frequency domain reflection method, and supervises the initial convolutional neural network for each defect position. After learning effective features based on the traditional time-frequency domain reflection method, the cable to be tested is tested. Therefore, the present invention combines the time-frequency domain reflection method with the traditional method in the training stage to obtain a time domain signal with defect position characteristics, and uses the signal as the input of the convolutional neural network for training, which not only ensures the accuracy of defect location, but also improves the efficiency of defect location.

[0110] Figure 9 is a schematic diagram of the convolutional neural network architecture provided by an embodiment of the present invention. A convolutional neural network consists of two parts: forward propagation and backpropagation. Before backpropagation, a loss function must be defined. Based on the raw data, the present invention adds a Gaussian distribution to known defect locations as labels. The loss is calculated using the output of the forward propagation and this label. Common loss functions include mean squared error (MSE) and cross-entropy error. The present invention uses the MSE loss function.

[0111] It's worth noting that due to the significant difference between the amplitude of the positioning signal at the beginning and end and the amplitude at the defect, and the limited training data, if the mean squared error between the defect location and the network output is directly used as the optimization target for end-to-end training, the network will have difficulty learning. Therefore, based on this raw data, the present invention adds a Gaussian distribution as a label to the locations known to be defects, and calculates the loss using the output of the forward propagation and this label.

[0112] As a preferred embodiment, the mean square error loss function can be expressed as:

[0113] Among them: y k It represents the output of the neural network, i.e. the predicted value, t k Represents the supervised data, i.e., the true value, and k represents the dimension of the output data.

[0114] There are some internal parameters in the network model, such as weight W and bias b, which are used to calculate the degree of deviation between the predicted value and the true value output by the training set. These parameters constitute the loss function L(x). The optimization algorithm mainly achieves the purpose of minimizing / maximizing the loss function by modifying the internal parameters W and b of the model. The optimization of the internal parameters of the model plays a very important role in effectively training the model and producing accurate results. Therefore, it is necessary to adopt various optimization algorithms to update and calculate the network parameters that affect model training and model output, so that they approach or reach the optimal value. The optimization algorithm selected by the present invention is the Adam gradient descent method, and the formula for parameter update is:

[0115] Among them, β1 and β2 are attenuation factors, V t is the momentum term. In the calculation process, V t and s t The initial value is 0, which causes the moment estimate to be biased towards 0, especially in the initial time step or when the attenuation factors β1 and β2 are small. t and s t All have been corrected.

[0116] Through the forward propagation and back propagation process of the one-dimensional convolutional neural network, the high-level features of the input original time domain signal are finally obtained. These high-level features contain the information of the input data, are highly distinguishable, and are conducive to cable defect location.

[0117] Step S12: Using the total reflection signal as input to a trained first convolutional neural network, wherein the first convolutional neural network extracts time domain features corresponding to the n defects.

[0118] Step S13: establishing a probability distribution model for the time domain features, and obtaining first defect positions corresponding to n defects one by one according to the probability distribution model.

[0119] Specifically, the first convolutional neural network extracts the first time domain feature corresponding to each reflected signal, outputs several consecutive first probabilities corresponding to the cable segment to be tested, and takes the cable sub-segment to be tested corresponding to the largest second probability among the several first probabilities as the first defect location, thereby obtaining the first defect locations corresponding to n defects one by one.

[0120] It's worth noting that cable defect location is considered a probabilistic problem. The probability of a defect occurring at a specific location in the cable is calculated, and the location with the highest probability is the desired defect location. After inputting the test data into a trained one-dimensional convolutional neural network, the network learns high-level features, establishes a mapping relationship between cable defects and their locations, and achieves defect location.

[0121] It is worth noting that the probability of a defect is the local maximum value, that is, when there are multiple defects, each defect position will have a corresponding local maximum probability. Each time, the local maximum probability is selected as the located defect position. Therefore, for a one-dimensional time domain signal, the corresponding local defect positions of multiple defects can be output simultaneously.

[0122] The present invention uses a convolutional neural network to process the input total reflection signal and obtain the defect position according to the corresponding probability distribution model. In fact, the probability of a defect occurring at each position can be obtained, and the traditional binary classification problem of whether a defect exists at each position is converted into a probability problem of a defect occurring at each position. The position with the highest probability of a defect is located as the defect position, thereby improving the accuracy of cable defect positioning; in addition, the position where a defect is about to occur can also be obtained, except that the sub-segment of the cable to be tested with the highest probability in the cable segment to be tested is used as the defect position.

[0123] 10 , which is a schematic structural diagram of a cable defect location system based on a convolutional neural network according to an embodiment of the present invention, comprising: a reflection module 101 , a time domain feature acquisition module 102 , and a location module 103 .

[0124] It is worth noting that the reflection module 101 is used to obtain the total reflection signal of the cable to be tested with n defects, and transmit the obtained total reflection signal to the time domain feature acquisition module 102; after receiving the total reflection signal, the time domain feature acquisition module 102 obtains the time domain characteristics of the total reflection signal, and transmits the total reflection signal to the positioning module 103; after receiving the time domain characteristics, the positioning module 103 locates the cable to be tested according to the probability distribution model to obtain the defect position.

[0125] The reflection module 101 is used to establish a defective cable transmission model for a cable to be tested having n defects, and obtain a total reflected signal of an initial incident signal after being reflected by the n defects according to the defective cable transmission model; wherein n is a positive integer.

[0126] Specifically, a defective cable transmission model is established for a cable to be tested with n defects, including: establishing a cable equivalent distributed parameter circuit model for the cable to be tested with n defects, obtaining the cable characteristic impedance based on the cable equivalent distributed parameter circuit model, and establishing the defective cable transmission model based on the cable characteristic impedance.

[0127] For a cable to be tested with n defects, a cable equivalent distributed parameter circuit model is established, including: obtaining cable unit length parameters for the cable to be tested with n defects, and establishing a cable equivalent distributed parameter circuit model based on the cable unit length parameters; wherein the cable unit length parameters include: resistance, inductance, conductance and capacitance per unit length of the cable.

[0128] Obtaining the cable characteristic impedance according to the cable equivalent distributed parameter circuit model includes: obtaining the cable head end voltage and current in the cable equivalent distributed parameter circuit model according to Kirchhoff's law, and obtaining the reflection coefficient and input impedance of the cable segment to be measured from the cable end according to the cable head end voltage and current.

[0129] According to the defective cable transmission model, a total reflection signal of an initial incident signal after being reflected by the n defects is obtained, including: according to the defective cable transmission model, n first reflection signals of the initial incident signal reflected by the head ends of the n defects are obtained, and a second reflection signal of the n first reflection signals reflected by the end of the cable to be tested is obtained; the n first reflection signals are accumulated with the second reflection signal to obtain a total reflection signal, so that a transfer function of the total reflection signal obtains a second time domain signal, and the second time domain signal is used as the input of the first convolutional neural network.

[0130] Obtaining the incident signal includes: when the length of the cable to be tested is less than a first length threshold or greater than a second length threshold, adjusting the center frequency of an initial first reference signal to a preset center frequency, obtaining a second reference signal, and using the second reference signal as the incident signal.

[0131] Obtaining a trained first convolutional neural network includes: according to a multi-defect cable simulation model, obtaining first time domain signals at multiple defect positions respectively by a time-frequency domain reflection method, and adding Gaussian distributions as training labels at the multiple defect positions respectively; using the obtained multiple first time domain signals as inputs of a one-dimensional initial second convolutional neural network, outputting the corresponding second defect positions in sequence, and training according to the second defect positions and the training labels to obtain a trained first convolutional neural network.

[0132] The time domain feature acquisition module 102 is configured to use the total reflection signal as input to a trained first convolutional neural network, and the first convolutional neural network extracts the time domain features corresponding to the n defects.

[0133] The positioning module 103 is configured to establish a probability distribution model for the time domain features, and obtain first defect positions corresponding to n defects according to the probability distribution model.

[0134] Specifically, the first convolutional neural network extracts the first time domain feature corresponding to each reflected signal, outputs several consecutive first probabilities corresponding to the cable segment to be tested, and takes the cable sub-segment to be tested corresponding to the largest second probability among the several first probabilities as the first defect location, thereby obtaining the first defect locations corresponding to n defects one by one.

[0135] The present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the convolutional neural network-based cable defect location method as described in the first aspect is implemented.

[0136] The present invention obtains the total reflection signal of the cable under test after reflection from multiple defects. Since the obtained total reflection signal contains the reflection signals after reflection from all defects, the convolutional neural network performs a unified convolution operation on the reflection signals of multiple defect positions, and can automatically obtain all corresponding time domain features for each reflection signal at one time, without relying on human experience, thereby improving the accuracy of cable defect positioning and improving the efficiency of defect positioning. Moreover, the convolutional neural network obtains the defect position according to the corresponding probability distribution model of the input total reflection signal, and converts the traditional two-classification problem of whether there is a defect at each position into a probability problem, thereby further improving the accuracy of cable defect positioning.

[0137] Those skilled in the art will appreciate that the embodiments of the present application may also provide computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0138] The present application is described with reference to the flow chart and / or block diagram of the method, device (system), and computer program product according to the embodiment of the present application. It should be understood that each flow process and / or box in the flow chart and / or block diagram and the combination of the flow process and / or box in the flow chart and / or block diagram can be realized by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processing machine or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device for realizing the function specified in one flow chart flow or multiple flows and / or one box or multiple boxes of the block diagram.

[0139] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a product including an instruction device that implements the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0140] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0141] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A cable defect location method based on convolutional neural network, characterized in that: include: A defective cable transmission model is established for a cable to be tested having n defects, and a total reflection signal of an initial incident signal after being reflected by the n defects is obtained according to the defective cable transmission model; wherein n is a positive integer; Using the total reflection signal as an input of a trained first convolutional neural network, wherein the first convolutional neural network extracts time domain features corresponding to the n defects; A probability distribution model is established for the time domain feature, and first defect positions corresponding to n defects are obtained according to the probability distribution model.

2. The cable defect location method based on convolutional neural network according to claim 1, characterized in that: The first convolutional neural network is obtained, comprising: According to the multi-defect cable simulation model, first time domain signals at the multi-defect positions are obtained respectively by time-frequency domain reflection method, and Gaussian distribution is added as training labels at the multi-defect positions respectively; The obtained multiple first time domain signals are respectively used as inputs of the one-dimensional initial second convolutional neural network, and the corresponding second defect positions are output in sequence. Training is performed according to the second defect positions and the training labels to obtain a trained first convolutional neural network.

3. The cable defect location method based on convolutional neural network according to claim 1, characterized in that: The method of establishing a probability distribution model for the time domain feature and obtaining first defect positions corresponding to n defects according to the probability distribution model comprises: The first convolutional neural network extracts the first time domain feature corresponding to each reflected signal, outputs a number of continuous first probabilities corresponding to the cable segment to be tested, and selects the sub-segment of the cable to be tested corresponding to the largest second probability from the several first probabilities as the first defect position, and obtains the first defect positions corresponding to n defects one by one.

4. The cable defect location method based on convolutional neural network according to claim 1, characterized in that: The step of obtaining a total reflected signal of an initial incident signal after the initial incident signal is reflected by the n defects according to the defective cable transmission model includes: According to the defective cable transmission model, n first reflection signals of the initial incident signal reflected by the n defect head ends and second reflection signals of the n first reflection signals reflected by the end of the cable to be tested are obtained; The n first reflection signals are accumulated with the second reflection signal to obtain a total reflection signal, so that a transfer function of the total reflection signal obtains a second time domain signal, and the second time domain signal is used as an input of the first convolutional neural network.

5. The cable defect location method based on convolutional neural network according to claim 4, characterized in that: Acquiring the incident signal includes: When the length of the cable to be tested is less than a first length threshold or greater than a second length threshold, the center frequency of the initial first reference signal is adjusted to a preset center frequency to obtain a second reference signal, and the second reference signal is used as the incident signal.

6. The cable defect location method based on convolutional neural network according to claim 1, characterized in that: The method of establishing a defective cable transmission model for a cable to be tested having n defects comprises: For a cable to be tested with n defects, a cable equivalent distributed parameter circuit model is established, and according to the cable equivalent distributed parameter circuit model, the The cable characteristic impedance is determined, and a defective cable transmission model is established based on the cable characteristic impedance.

7. The cable defect location method based on convolutional neural network according to claim 6, characterized in that: The method of establishing a cable equivalent distributed parameter circuit model for a cable to be tested having n defects comprises: For a cable to be tested with n defects, the cable unit length parameters are obtained, and a cable equivalent distributed parameter circuit model is established according to the cable unit length parameters; wherein the cable unit length parameters include: resistance, inductance, conductance and capacitance of the cable unit length.

8. The cable defect location method based on convolutional neural network according to claim 6, characterized in that: The obtaining the cable characteristic impedance according to the cable equivalent distributed parameter circuit model comprises: According to Kirchhoff's law, the voltage and current at the head end of the cable in the cable equivalent distributed parameter circuit model are obtained, and based on the voltage and current at the head end of the cable, the reflection coefficient and input impedance of the cable segment to be tested from the end of the cable are obtained.

9. A cable defect location system based on convolutional neural network, characterized in that: include: Reflection module, time domain feature acquisition module and positioning module; among them, The reflection module is used to establish a defective cable transmission model for a cable to be tested having n defects, and obtain a total reflection signal of an initial incident signal after being reflected by the n defects according to the defective cable transmission model; wherein n is a positive integer; The time domain feature module is used to use the total reflection signal as an input of a trained first convolutional neural network, and the first convolutional neural network extracts the time domain features corresponding to the n defects; The positioning module is used to establish a probability distribution model for the time domain feature, and obtain the first defect positions corresponding to n defects according to the probability distribution model.

10. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the cable defect location method based on a convolutional neural network as described in any one of claims 1 to 8 is implemented.

Images