Control cable fault diagnosis method based on three-phase current detection

Abstract

A control cable fault diagnosis method based on three-phase current detection comprises the steps that firstly, healthy samples are used for training AAE, then input and reconstructed input are projected into the trained AAE, paired activation values are obtained, novelty is recognized by calculating abnormal scores, a monitoring module checks the abnormal score of each given window, and a fault diagnosis module judges whether the abnormal scores are abnormal or not. And applying a moving average to these values to use it as a fault indicator, the AEE being a decoder with RaPP scoring technology. The beneficial effects are that the diagnosis model does not need additional sensors and signal generators, and has cost benefits. In addition, physical parameters and related information of the cable do not need to be filled in, the provided diagnosis model can effectively diagnose health and fault states, and the method is suitable for an automatic production line operating in various severe environments.

Description

Technical Field

[0001] The invention relates to the field of cable fault detection, in particular to a control cable fault diagnosis method based on three-phase current detection. Background Art

[0002] Control cables are widely used in industrial applications, including nuclear power plants, ship power systems, vehicles, and factory automation, as they play a vital role in controlling motors and sensor instruments. In order to maintain the stability of automation systems in various applications, it is essential to monitor the health of cables. Cable faults are divided into hard faults and soft faults. Hard faults include open circuits and short circuits, and soft faults are characterized by small impedance changes. Local deformation of cables due to harsh environments (soft faults) may be converted into hard faults due to local damage to components such as cable conductors, coatings, and shielding layers.

[0003] So far, the effectiveness of cable diagnostic methods based on reflection measurements has been demonstrated. In particular, time-frequency domain reflectometry (TFDR) has demonstrated robust fault diagnosis in a variety of situations due to the use of time-frequency domain information. Although these methods are effective in diagnosing hard faults due to the large amount of reflections, their diagnostic accuracy for small anomalies is not always effective due to physical difficulties such as the ambiguous response of the event signal. In addition, these reflection measurement-based methods require additional equipment such as signal generators and sensors for diagnosis. In addition, these methods require prior domain knowledge and manual feature extraction during testing and analysis. Summary of the invention

[0004] The purpose of the present invention is to solve the above problems and design a control cable fault diagnosis method based on three-phase current detection. The specific design scheme is:

[0005] A control cable fault diagnosis method based on three-phase current detection, first trains an AAE with healthy samples, then projects the input and its reconstructed input into the trained AAE to obtain paired activation values, and then identifies novelty by calculating the anomaly score. The monitoring module checks the anomaly score of each given window and applies a moving average to these values ​​to use them as fault indicators. The AEE is a decoder with RaPP scoring technology.

[0006] The AAE includes an encoder g and a decoder f. The encoder compresses the input and reduces the input dimension, while the decoder is its inverse mapping, that is, reconstructing the reduced-dimensional input into The selected features, such as Pa, Pb, Pc, I sum, are the input of the proposed network. The encoder encodes the input features into a simplified representation, namely the latent code. Therefore, the encoder acts as a latent code generator and tries to mislead the discriminator Dz into believing that the latent code comes from a normal distribution. The role of Dz is to distinguish whether a given sample is generated by the latent code of the AAE or a random vector sample from a normal distribution. The latent code has a distribution similar to the prior distribution during training.

[0007] Although AAE itself shows good results in novelty detection, the reconstruction error alone cannot fully utilize the information provided by the trained AAE, so using the information in its deep architecture can potentially improve the accuracy.

[0008] The RaPP process supports the decision process not only by comparing the input and the reconstructed input, but also by comparing the hidden activations in the latent space and the corresponding reconstruction. The hidden reconstruction can be indirectly calculated by feeding the reconstructed input to the trained AAE again, because the corresponding hidden reconstruction is equivalent to the hidden activation of the reconstructed input. The hidden reconstruction can be calculated by inputting each to g, where x^ is the reconstructed input of the trained AAE, and the input x is given by x^=A(x). Let x^=A(x) be the number of hidden layers in g and f. The part of g is calculated as follows:

[0009]

[0010] If x and x^ are substituted into A, their hidden representation pairs (hi, hi^) are obtained as follows:

[0011] h i (x) = g :i (x)

[0012]

[0013] Using these two hidden activations, we can get the novelty score. To calculate the score, RaPP uses the normalized path aggregation (S NAP )measure:

[0014] s NAP (x) = || (d(x) - μx) T ∨∑ -1 -1||

[0015] in x is a given training set, d is a matrix whose i-th row corresponds to d(xi) of xi∈x. In order to normalize the distance, we need to calculate d=UΣV T .

[0016] The control cable fault diagnosis method based on three-phase current detection obtained by the above technical solution of the present invention has the following beneficial effects:

[0017] The diagnostic model does not require additional sensors and signal generators, which is cost-effective. In addition, it does not need to fill in the physical parameters and related information of the cable. The proposed diagnostic model can effectively diagnose the health and fault status, and is suitable for automated production lines operating in various harsh environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is an equivalent circuit model diagram of the faulty cable of the present invention;

[0019] Figure 2 is a schematic diagram of the AAE architecture of the present invention;

[0020] Figure 3 This is the process of damaging the wire according to the present invention:

[0021] (a) A healthy wire inside a cable; (b) A wire without insulators; and (c) A faulty wire with damaged wires (part of the wire is cut off). DETAILED DESCRIPTION

[0022] The present invention will be described in detail below in conjunction with the accompanying drawings.

[0023] A control cable fault diagnosis method based on three-phase current detection, first trains an AAE with healthy samples, then projects the input and its reconstructed input into the trained AAE to obtain paired activation values, and then identifies novelty by calculating the anomaly score. The monitoring module checks the anomaly score of each given window and applies a moving average to these values ​​to use them as fault indicators. The AEE is a decoder with RaPP scoring technology.

[0024] The AAE includes an encoder g and a decoder f. The encoder compresses the input and reduces the input dimension, while the decoder is its inverse mapping, that is, reconstructing the reduced-dimensional input into The selected features, such as Pa, Pb, Pc, I sum, are the input of the proposed network. The encoder encodes the input features into a simplified representation, namely the latent code. Therefore, the encoder acts as a latent code generator and tries to mislead the discriminator Dz into believing that the latent code comes from a normal distribution. The role of Dz is to distinguish whether a given sample is generated by the latent code of the AAE or a random vector sample from a normal distribution. The latent code has a distribution similar to the prior distribution during training.

[0025] Although AAE itself shows good results in novelty detection, the reconstruction error alone cannot fully utilize the information provided by the trained AAE, so using the information in its deep architecture can potentially improve the accuracy.

[0026] The RaPP process supports the decision process not only by comparing the input and the reconstructed input, but also by comparing the hidden activations in the latent space and the corresponding reconstruction. The hidden reconstruction can be indirectly calculated by feeding the reconstructed input to the trained AAE again, because the corresponding hidden reconstruction is equivalent to the hidden activation of the reconstructed input. The hidden reconstruction can be calculated by inputting each to g, where x^ is the reconstructed input of the trained AAE, and the input x is given by x^=A(x). Let x^=A(x) be the number of hidden layers in g and f. The part of g is calculated as follows:

[0027]

[0028] If x and x^ are substituted into A, their hidden representation pairs (hi, hi^) are obtained as follows:

[0029] h i (x) = g :i (x)

[0030]

[0031] Using these two hidden activations, we can get the novelty score. To calculate the score, RaPP uses the normalized path aggregation (S NAP )measure:

[0032] s NAP (x) = || (d(x) - μx) T ∨∑ -1 -1||

[0033] in x is a given training set, d is a matrix whose i-th row corresponds to d(xi) of xi∈x. In order to normalize the distance, we need to calculate d=UΣV T .

[0034] Example 1 Control cable soft fault and characteristics

[0035] A soft fault is a fault with a small impedance change caused by local deformation of the cable, such as local deformation of the conductor or insulation layer in a harsh environment. The soft fault resistance Rf directly indicates the degree of damage to the conductor because it is proportional to the cross-sectional area reduction S of the faulty conductor. f Inversely proportional; the relationship can be expressed as:

[0036] S f =αS o ,0<α<1

[0037]

[0038] In the formula, S 0 is the cross-sectional area of ​​the normal cable, S f is the remaining cross-sectional area after damage. According to the damaged cross-sectional area of ​​the cable conductor, α indicates the degree of damage to the cable. 0 is the normal cable resistance, R f is the fault resistance of the faulty cable, where ρ is the resistivity of the conductor material and l is the length of the conductor. The value of a healthy cable is 1, and the value of a faulty cable is less than 1. For open circuits and short circuits, α is zero and greater than 1, respectively. The current through the fault location is affected by the damaged conductor and the fault resistance R f The influence of damaged insulators and leakage current flowing through the damaged insulation layer causes the power circuit to be unbalanced. The three-phase current in the control cable can be expressed as:

[0039] i a (t) = I ma cos(ωt+φ a )

[0040] i b (t) = I mb cos(ωt+φ b -120°)

[0041] i c (t) = I mc cos(ωt+φ c -120°)

[0042] Where I ma ,I mb ,I mc is the three-phase current amplitude, Φa, Φb, Φc are the corresponding phase angles. The ratio of each single-phase current to the total current directly reflects the degree of imbalance of the fault control network. j The definitions are as follows:

[0043]

[0044] P j =SC j / SSC

[0045] Where j represents any one of the phases a, b, and c. The square of the current ratio P j Comparison of each single-phase current (SC j ) to the total three-phase current (SSC).

[0046] Therefore, it reflects that the fault grid has a fault resistance R f . Single-phase cross-sectional area reduction Sf It will cause phase imbalance, which will have adverse effects on the equipment, such as power loss, additional heat generation, and reduced energy consumption. The asymmetry of phase impedance leads to unbalanced phase impedance, which in turn leads to unbalanced three-phase current.

[0047] Figure 1 It also shows that a soft fault is modeled as an additional resistor R f , which represents the reduction in the cross-sectional area of ​​the conductor. Zc represents the characteristic impedance of a healthy conductor. The insulation of the conductor is represented by R i and C i The dielectric response of an insulator can be described by its capacitive properties for storing charge and its conductive properties for conducting charge. The magnitude of the leakage current is smaller than the magnitude of the capacitive current in a normal conductor. However, when the conductor is damaged (R f Increase), when the insulator is damaged (R i reduction), i res Increase. Therefore, due to leakage, the vector sum of the three-phase current is not zero, and the leakage is caused by unhealthy wires. Therefore, the present invention takes the vector sum of the three-phase current of the wire input stage as a feature, which is expressed as:

[0048] i sum =i a +i b +i c

[0049] When a soft fault occurs in the conductor, i sum The value of is non-zero, but the normal wire i sum The value is zero.

[0050] In addition, the unbalance of the three-phase current can be described by the negative sequence current (NSC), which can be calculated by the following formula:

[0051]

[0052] where i NS is the NSC of the three-phase current. is the phase rotation operator. When unbalance exists, NSC is present in addition to the positive sequence current, so NSC is a suitable parameter for diagnosing control network imbalance; for example, if the load or transmission line is unbalanced. Due to the inherent asymmetry of the control network, the NSC value of normal motor and control cable components is small, but in the event of a soft fault, the NSC value fluctuates greatly. The more severe the fault, the greater the imbalance of the three-phase current.

[0053] Example 2 Experimental device

[0054] To simulate different degrees of wire damage, we artificially and gradually increase the resistance of the wire by reducing the cross-sectional area of ​​the faulty wire. Figure 3The process of causing damage to the wires is described. Figure 3 (a) depicts an intact and healthy conductor. To simulate damage, the insulator of the wire is removed, e.g. Figure 3 (b) In order to reduce the cross-sectional area of ​​the wire, a total of 30 wires were taken out. Four chains were cut in each experimental iteration, increasing S f , reduce α, such as Figure 3 As shown in (c), α decreases from 1.000 to 0.066. To simulate the cable fault condition, artificial soft faults of different degrees (0.066<α<1.000) are induced in the a-phase conductor. In order to control the damage of different degrees, the fault length is determined in the experiment.

[0055] Figure 3 Process of causing damage to wires: (a) Healthy wire inside a cable; (b) wire with insulator removed; (c) faulty wire with damaged wire (part of the wire is cut off).

[0056] According to the left cross-sectional area S of the target wire f , the cable working conditions are divided into 8 categories. In the experiment, a healthy target cable is prepared and gradually damaged to collect normal and fault data sets. 5000 selected windows of each fault severity level are selected from the measured three-phase current as training, validation and test data sets. The data sets collected for each severity are shown in Table 1:

[0057]

[0058]

[0059] Table 1 Total dataset for each damage severity level

[0060] In addition to the 8 classes corresponding to 1 normal condition and 7 fault conditions, an additional fault condition was tested. The damage to the wire conductor is mostly caused by the surrounding environment destroying the insulation layer outside the wire conductor. Therefore, the algorithm is also needed to diagnose the damage to the insulator. Figure 1 According to the equivalent model shown, the damage of the insulator will change R i Value and C i value, resulting in the current i res and i cap This method is based on the current of the target cable and can diagnose this type of soft fault.

[0061] Example 2 Results Analysis

[0062] We tested 11 scenarios listed in Table 2 to demonstrate the proposed method under different operating and fault conditions. Scenarios 1-7 were prepared for each damage level, i.e., 4 cut degrees-28 cut degrees, and experiments were conducted separately. The 4 cut degree level is a slightly damaged dataset, which means that only 13.3% of the conductors are damaged in this level. The most severely damaged level is the 28 cut degree level, which means that 93.3% of the conductors are damaged in this level; however, the damaged wires are still working and do not show any fault symptoms. Scenario 8 tests relatively mild damage levels, i.e., 4 cut degrees-12 cut degrees, where the cross-sectional area of ​​the conductor is more than half, while scenario 9 tests severe damage levels, i.e., 16 cut degrees-28 cut degrees, where the cross-sectional area of ​​the conductor is less than half. Scenario 10 is to check whether the proposed method diagnoses the damage of the insulator when the conductor under the insulation layer is intact. According to the equivalent model, the algorithm proposed in the present invention can be diagnosed theoretically. The last scenario is to test the entire dataset at once to verify the overall diagnostic ability of the proposed algorithm.

[0063]

[0064]

[0065] Table 2 Experimental scenarios

[0066] Table 3 gives the AUROCs of the standard AAE and the RaPP improved AAE. The best score in each case is underlined. The improved AAE provides good results for most scenarios, except for scenario 8. Scenario 5 has the largest improvement, with the difference between the standard AUROC and the RaPP AAE exceeding 20%. For mild damage levels (scenario 8), the standard AAE detects damage better than the AAE using RaPP. However, the AUROCs of both methods are greater than 0.900, which is high enough for the diagnosis of soft faults, because soft faults tend to gradually transform into hard faults in chronological order from mild to severe. Therefore, ongoing soft fault development can be detected during the development process. Moreover, the AUROC under scenario 10 is calculated to be 0.980, which is sufficient to prove that the method of the present invention is suitable for insulator damage diagnosis. As shown in Figure 1 The equivalent circuit shown works properly; therefore, the proposed method can be used to diagnose the damage of the insulation layer. Scenario 11 is tested for various damage levels found during the development of soft faults. The experimental results show that the proposed diagnostic model is applicable to practical problems.

[0067]

[0068]

[0069] Table 3 AUROC of novel datasets

[0070] The main differences between the existing methods and the proposed method are: (1) The proposed method does not require manual feature extraction and prior domain knowledge, such as the physical parameters of the cable. (2) No additional equipment is required (cost-effective); (3) Very short cables can be diagnosed. (4) Even if the operating conditions change continuously, the operating cable can be diagnosed online. The proposed method is applicable to the field of factory automation where various operating conditions and different types of cables exist.

[0071] 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 fault diagnosis method based on three-phase current detection, first trains AAE with healthy samples, then projects the input and its reconstructed input into the trained AAE to obtain paired activation values, and then identifies novelty by calculating the anomaly score. The monitoring module checks the anomaly score of each given window and applies a moving average to these values ​​to use them as fault indicators, characterized in that, The AEE is a decoder with RaPP scoring technology.

2. The control cable fault diagnosis method based on three-phase current detection according to claim 1 is characterized in that: The AAE consists of an encoder g and a decoder f, the encoder compresses the input and reduces the input dimension, while the decoder is its inverse mapping.

3. According to the control cable fault diagnosis method based on three-phase current detection according to claim 1, it is characterized in that: The input x is given by x^=A(x), set to the number of hidden layers in g and f, and the part of g is calculated as follows: If x and x^ are substituted into A, their hidden representation pairs (hi, hi^) are obtained as follows: h i (x)=g :i (x) Using these two hidden activations, we can get the novelty score. To calculate the score, RaPP uses the normalized path aggregation (S NAP )measure: s NAP (x)=||(d(x)-μx) T ∨∑ -1 -1|| in x is a given training set, d is a matrix whose i-th row corresponds to d(xi) of xi∈x. In order to normalize the distance, we need to calculate d=UΣV T .

Images