Power supply system level fault diagnosis based on dynamic fuzzy network and multi-source information fusion

Abstract

The present invention discloses a power supply system-level fault diagnosis method based on dynamic fuzzy network and multi-source information fusion, the method comprising: collecting aircraft power supply system monitoring parameters; extracting multi-dimensional features of high-frequency signals; constructing a power supply system-level component association network; using Frechet algorithm to perform multi-dimensional information fusion; fusion of knowledge information, and simplification of graph model structure; constructing a dynamic fuzzy network library with historical parameters; and performing graph matching, search, and fault diagnosis on real-time monitoring parameter data. The present invention integrates multi-dimensional information such as frequency domain, time domain, and trend information in the power supply system monitoring parameters and multi-source information of knowledge information, deeply mines the association relationship between various components of the power supply system, relies on dynamic fuzzy network to store and fuse multi-source information of the power supply system, and realizes complex fault diagnosis of the power supply system based on graph theory methods. The method has the advantages of strong interpretability, high diagnostic efficiency, and strong scalability.

Description

Technical Field

[0001] The present invention relates to the technical field of aircraft power supply system fault diagnosis, and in particular to a power supply system-level fault diagnosis method based on dynamic fuzzy network and multi-source information fusion. Background Art

[0002] The aircraft power supply system is a complete system composed of the production, conversion and distribution of aircraft power. Its function is to provide all electrical equipment on the aircraft with power that meets the specified technical performance to ensure the normal operation of the electrical equipment. With the popularization of more-electric aircraft, the importance of aircraft power supply systems has gradually become prominent, and fault detection and diagnosis technologies have also been gradually applied to aircraft power supply systems. However, due to the scarcity of fault data in the power supply system of more-electric aircraft and the uncertainty of fault experience knowledge, as well as the complex structure of the power supply system, strong structural coupling, multi-parameters, and multi-components, fault detection and fault propagation methods in other fields are difficult to be directly applied to the aircraft power supply system. At present, there are few studies on fault diagnosis technology for aircraft power supply systems. Therefore, in view of the typical characteristics of the power supply system and the limitations of the existing fault diagnosis methods, a power supply system-level fault diagnosis method based on dynamic fuzzy networks and multi-source information fusion is proposed. The knowledge information, frequency domain and time domain information in the power supply system are integrated, and the dynamic fuzzy network is constructed to realize the fault diagnosis of the power supply system level in view of the high complexity of the numerous components in the power supply system. Summary of the invention

[0003] The purpose of the present invention is to provide a power supply system level fault diagnosis method based on dynamic fuzzy network and multi-source information fusion to realize aircraft power supply system level fault diagnosis.

[0004] The technical solution adopted by the present invention is a power supply system-level fault diagnosis method based on dynamic fuzzy network and multi-source information fusion, which specifically includes:

[0005] Use sensors to collect various operating monitoring parameters of each device in the aircraft power supply system;

[0006] The sorted parameter data of each sensor is processed with high frequency signals to decompose the multi-dimensional information such as time domain information, frequency domain information, trend information, etc. existing in the original signal;

[0007] Construct a component association network at the power supply system level, with each component as a node and the association relationship between components as an edge, and the initial state of each edge is set to 1;

[0008] Based on multi-dimensional information, the Frechet algorithm is used to fuse multi-dimensional information, calculate the correlation between components, and dynamically update the fuzzy graph network to construct a dynamic fuzzy network.

[0009] Combined with the power supply system structure information, knowledge information is integrated through graph theory to simplify the graph structure;

[0010] Based on fault data, dynamic fuzzy network construction at power supply system level under different fault modes is carried out;

[0011] By real-time monitoring of parameter data, a real-time dynamic fuzzy network at the power supply system level can be constructed. Through the graph model matching algorithm, the fault mode that best matches the measured power supply system dynamic fuzzy network can be found to achieve fault diagnosis at the power supply system level.

[0012] Preferably, the high-frequency signal processing method is a Hilbert-Huang transform algorithm and an effective value calculation method.

[0013] Preferably, the Hilbert-Huang transform algorithm is a high-frequency signal analysis method for processing non-stationary nonlinear signals, and mainly consists of two parts, the first part is empirical mode decomposition, and the second part is Hilbert transform.

[0014] (1) First, all the maximum points of the signal x(t) are linked to form the upper envelope x up (t); and all the minimum points x low (t) is linked to the lower envelope and the signal function obtained by the selected signal should be included in it

[0015] (2) Through the upper and lower envelopes, based on each time unit, the mean of each corresponding upper and lower envelope is calculated, and then the mean function is subtracted from the selected signal to obtain the difference function d(t). The specific formula is as follows:

[0016]

[0017] Repeat the above operation until the difference function d(t) satisfies two conditions: the number of extreme points of the data differs from the number of zero-crossing points by zero or by at most one; at any data point, the mean of the upper envelope composed of local maximum values ​​and the lower envelope composed of local minimum points is 0.

[0018] (3) Let d(t) be, c1(t) be the first IMF obtained by the selected signal function, and then calculate the corresponding residual, denoted as r1(t). The formula for calculating the residual is as follows:

[0019] r1(t)=x(t)-c1(t)

[0020] (4) Take r1(t) as the next selected starting signal, denoted as x(t), and recalculate according to the above algorithm. Through the above steps, the IMF components adaptively obtained for the signal can be calculated. When the residual is less than the threshold, the algorithm stops running. After empirical mode decomposition, x(t) can be converted into the following components:

[0021]

[0022] Where N is the number of IMFs, r N (t) is the final residual, c j (t) is the obtained IMF. The IMFs are mutually orthogonal and have the property of zero mean.

[0023] (5) After obtaining the signal IMF component, according to the provided signal component c j (t) is subjected to Hilbert transform, and the expression is as follows:

[0024]

[0025] Among them, y j (t) is the timing signal c j The Hilbert-Huang transform result of (t), P is the Cauchy principal value.

[0026] (6) c j (t) as the real part, and y j (t) as the imaginary part, the resulting expression is the corresponding expression, z(t) as follows:

[0027] z j (t) = c t (t)+iy j (t) = a(t)e iθ(t)

[0028] Among them, a(t) and θ t are the instantaneous amplitude and instantaneous phase of the analytical signal, which are c in the rectangular coordinate system j (t), y j (t) In polar coordinates, the two are expressed as follows:

[0029]

[0030] (7) By processing the instantaneous phase, we can obtain the value of its derivative, that is, the corresponding instantaneous frequency:

[0031]

[0032] (8) In order to save computational space, this method selects the calculated a1(t)ω1(t) as the instantaneous amplitude and instantaneous frequency characteristics calculated by Hilbert Huang.

[0033] Preferably, the effective value calculation method is:

[0034] Since there are both DC power signals and AC power signals in the power supply system, the instantaneous voltage at different times is different, and there is Um (peak value). In order to facilitate the measurement and calculation of alternating current, it is necessary to define the magnitude of alternating voltage and alternating current based on the effect of alternating current, that is, the effective value. Suppose a segment of electrical signal X is x1, x2,…, x N , then the effective value of this electrical signal x rms for:

[0035]

[0036] Preferably, the graph construction method is an undirected weighted graph, which means that the edges in the graph model are undirected and weighted. The so-called undirected means that any point pair (i, j) and (j, i) correspond to the same edge, and vertex i and vertex j are also called the two endpoints of the undirected edge (i, j). The adjacency matrix of the undirected weighted graph G is

[0037] A=(a ij ) n×n

[0038] Where i,j∈(0,N) represents the components of the power supply system, a ij Represents the association relationship between component i and component j, based on which the power supply system association graph model can be constructed. The initial state is set to a ij =1.

[0039] Preferably, the Frechet algorithm is a distance measure that calculates the distance between curves to evaluate the similarity between curves. For a given component P, Q, its parameters are set to P = {u1, u2, ..., u p} and Q = {v1,v2,...,v q}, we get a monotone pairing L between P and Q, which is a series of corresponding points Where a1=1, b1=1, a m =p,b m =q. For i=1,2,...,m, if and only if one of the following three conditions is met:

[0040]

[0041] Then we calculate

[0042]

[0043] in represent The Euclidean distance between the curves is:

[0044]

[0045] From this, we can calculate the association relationship between the components, that is, a PQ =d F (P,Q), based on which the dynamic fuzzy network A of the power supply system can be constructed. ij ) n×n , where i,j∈(0,N) represents the components of the power supply system, a ij Represents the association relationship between component i and component j, and the calculation method is Frechet algorithm.

[0046] Preferably, the graph simplification method is: based on the knowledge of the power supply system structure, construct a power supply system level component association matrix B = (b ij ) n×n , where i,j∈(0,N) represents the components of the power supply system, b ij =1 indicates that there is an association relationship between component i and component j, b ij =0 means there is no association between components. ij ) n×n With the correlation matrix B = (b ij ) n×n For comparison, if b ij =0, then it is considered that a ij =0, if b ij ≠0, then it is considered that a ij =a ij ,The graph reduction algorithm based on knowledge fusion can eliminate the edges that have no diagnostic significance in the dynamic fuzzy network based on the knowledge of the power supply system, simplify the network structure, and thus achieve the simplification of the graph model.

[0047] Preferably, the graph matching algorithm is a graph matching algorithm based on the Frobenius norm (F-norm). Let A = [a ij ] m×n is an adjacency matrix of an incident graph. is the F-norm of this matrix.

[0048] Let A=[a ij ] m×n is an adjacency matrix of an association graph, called is the first norm of this matrix

[0049] Based on the F-norm, we can measure the similarity of two matrices. Let the adjacency matrix of the two graphs be A = [aij ] m×n , B=[b ij ] m×n , the graph similarity calculation formula is:

[0050] C=AB

[0051]

[0052] Through the power supply system-level fault diagnosis method based on dynamic fuzzy networks and multi-source information fusion, we can combine multi-source information such as data information and knowledge information in the power supply system, construct a dynamic fuzzy network of each component of the power supply system, explore the deep correlation between the components of the power supply system, and realize accurate fault diagnosis at the aircraft power supply system level based on the dynamic fuzzy network to provide support for maintenance decisions.

[0053] The beneficial effects of the present invention are:

[0054] The present invention is based on the data and structure analysis of the power supply system to carry out multi-source information fusion. It can fuse the multi-dimensional information such as frequency domain, time domain and trend information in the power supply system monitoring parameters as well as the multi-source information of data information and knowledge information. With the association relationship of each component as the edge, a dynamic fuzzy network for power supply system fault diagnosis is established. The association relationship of each component of the power supply system is deeply mined and analyzed from the association relationship dimension, thereby relying on the dynamic fuzzy network to store and fuse the multi-source information of the power supply system, and realize the description of different working states of the power supply system. A power supply system-level fault diagnosis algorithm based on dynamic fuzzy network and multi-source information fusion proposes a power supply system-level fault diagnosis method that fuses data information and knowledge information, which can realize fast and efficient power supply system-level fault diagnosis and improve the accuracy and efficiency of the diagnosis model. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0056] Figure 1 It is a flow chart of power supply system-level fault diagnosis based on dynamic fuzzy network and multi-source information fusion in an embodiment of the present invention;

[0057] Figure 2 is the original data of various parameters of the generator equipment No. 1 in the embodiment of the present invention;

[0058] Figure 3 is the instantaneous frequency characteristic of the exciter output voltage signal after Hilbert-Huang transform;

[0059] Figure 4is the instantaneous amplitude characteristic of the exciter output voltage signal after Hilbert-Huang transform;

[0060] Figure 5 It is the feature extraction calculation result of the exciter output voltage signal. The first line is the instantaneous frequency, the second line is the instantaneous amplitude, and the third line is the effective value.

[0061] Figure 6 The adjacency matrix of the dynamic fuzzy network model of the power supply system obtained based on the Frechet algorithm is represented in the form of a grayscale heat map. The grayscale represents the association between different components. The figure shows the dynamic fuzzy network of the power supply system under three fault modes: normal, main generator armature winding phase short circuit, and exciter excitation winding single-phase open circuit;

[0062] Figure 7 It is the adjacency matrix heat map of the dynamic fuzzy network of the power supply system under three fault modes: normal, main generator armature winding phase short circuit, and exciter excitation winding single-phase open circuit after the graphical model is simplified. DETAILED DESCRIPTION

[0063] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention and are not limited to the present invention.

[0064] Figure 1 FIG. 1 is a flow chart of power supply system-level fault diagnosis based on dynamic fuzzy network and multi-source information fusion in an embodiment of the present invention. Figure 1 As shown, the method comprises the following steps:

[0065] Step S101: using sensors to collect various operation monitoring parameters of various devices in the aircraft power supply system;

[0066] Step S102: The sorted parameter data of each sensor is subjected to high-frequency signal processing to decompose multi-dimensional information such as time domain information, frequency domain information, trend information, etc. existing in the original signal;

[0067] The high-frequency signal processing method is a Hilbert-Huang transform algorithm and an effective value calculation method; the Hilbert-Huang transform algorithm is a high-frequency signal analysis method for processing non-stationary nonlinear signals, which mainly consists of two parts, the first part is empirical mode decomposition, and the second part is Hilbert transform.

[0068] (1) First, all the maximum points of the signal x(t) are linked to form the upper envelope x up (t); and all the minimum points x low (t) is linked to the lower envelope and the signal function obtained by the selected signal should be included in it

[0069] (2) Secondly, through the upper and lower envelopes, a range can be obtained, and then based on each time unit, the mean of each corresponding upper and lower envelope is obtained, and then the mean function is subtracted from the selected signal to obtain the difference function d(t). The specific formula is as follows:

[0070]

[0071] Repeat the above operation until the difference function d(t) satisfies two conditions: the number of extreme points of the data differs from the number of zero-crossing points by zero or by at most one; at any data point, the mean of the upper envelope composed of local maximum values ​​and the lower envelope composed of local minimum points is 0.

[0072] (3) At this time, d(t) is recorded as, c1(t) is considered to be the first IMF obtained by the selected signal function, and then the corresponding residual is calculated, recorded as r1(t). The formula for calculating the residual is as follows:

[0073] r1(t)=x(t)-c1(t)

[0074] (4) Then r1(t) is used as the next selected starting signal, recorded as x(t), and recalculated according to the above algorithm. Through the above steps, the IMF components adaptively obtained for the signal can be calculated. When the residual is less than the threshold, the algorithm stops running. After empirical mode decomposition, x(t) can be converted into the following components:

[0075]

[0076] Where N is the number of IMFs, r N (t) is the final residual, c j (t) is the obtained IMF. The IMFs are mutually orthogonal and have the property of zero mean.

[0077] (5) After obtaining the signal IMF component, according to the provided signal component c j (t) is subjected to Hilbert transform, and the expression is as follows:

[0078]

[0079] Among them, y j (t) is the timing signal c j The Hilbert-Huang transform result of (t), P is the Cauchy principal value.

[0080] (6) c j (t) as the real part, and y j (t) as the imaginary part, the resulting expression is the corresponding expression, z(t) as follows:

[0081] z j (t) = c t (t)+iy j (t) = a(t)e iθ(t)

[0082] Among them, a(t) and θ t are the instantaneous amplitude and instantaneous phase of the analytical signal, which are c in the rectangular coordinate system j (t), y j (t) In polar coordinates, the two are expressed as follows:

[0083]

[0084] (7) By processing the instantaneous phase, we can obtain the value of its derivative, that is, the corresponding instantaneous frequency:

[0085]

[0086] (8) In order to save computational space, this method selects the calculated a1(t)ω1(t) as the instantaneous amplitude and instantaneous frequency characteristics calculated by Hilbert Huang.

[0087] The effective value calculation method is as follows:

[0088] Since there are both DC power signals and AC power signals in the power supply system, the instantaneous voltage at different times is different, and there is Um (peak value). In order to facilitate the measurement and calculation of alternating current, it is necessary to define the magnitude of alternating voltage and alternating current based on the effect of alternating current, that is, the effective value. Suppose a segment of electrical signal X is x1, x2,…, x N , then the effective value of this electrical signal x rms for:

[0089]

[0090] Step S103: construct a power supply system-level component association network, with each component as a node, and the association relationship between components as an edge, and the initial state of each edge is set to 1;

[0091] The network is constructed as an undirected weighted graph. An undirected weighted graph means that the edges in the graph model are undirected and weighted. The so-called undirected means that any pair of points (i, j) and (j, i) correspond to the same edge. Vertex i and vertex j are also called the two endpoints of the undirected edge (i, j). The adjacency matrix of the undirected weighted graph G is

[0092] A=(a ij ) n×n

[0093] Where i,j∈(0,N) represents the components of the power supply system, a ij Represents the association relationship between component i and component j, based on which the power supply system association graph model can be constructed. The initial state is set to a ij =1

[0094] Step S104: Based on the multi-dimensional information, the Frechet algorithm is used to fuse the multi-dimensional information, calculate the correlation between the components, and dynamically update the fuzzy graph network to construct a dynamic fuzzy network;

[0095] The Frechet algorithm is a distance measure that calculates the distance between curves and evaluates the similarity between curves. For a given curve P = {u1,u2,...,u p} and Q = {v1,v2,...,v q}, we get a monotone pairing L between P and Q, which is a series of corresponding points

[0096] Where a1=1, b1=1, a m =p,b m =q. For i=1,2,...,m, if and only if one of the following three conditions is met:

[0097]

[0098] Then we calculate

[0099]

[0100] in represent The Euclidean distance between the curves is:

[0101]

[0102] From this, we can calculate the association relationship between the components, that is, a PQ =d F (P,Q)

[0103] Step S105: combining the power supply system structure information, integrating the knowledge information through graph theory, and simplifying the graph model structure;

[0104] The graph simplification method is: based on the knowledge of the power supply system structure, construct a power supply system level component association matrix B = (b ij ) n×n , where i,j∈(0,N) represents the components of the power supply system, b ij =1 indicates that there is an association relationship between component i and component j, b ij=0 means there is no association between components. ij ) n×n With the correlation matrix B = (b ij ) n×n For comparison, if b ij =0, then it is considered that a ij =0, if b ij ≠0, then it is considered that a ij =a ij ,The graph reduction algorithm based on knowledge fusion can eliminate the edges that have no diagnostic significance in the dynamic fuzzy network based on the knowledge of the power supply system, simplify the network structure, and thus achieve the simplification of the graph model.

[0105] Step S106: constructing a power supply system-level dynamic fuzzy network under different fault modes based on the fault data;

[0106] Step S107: A real-time dynamic fuzzy network at the power supply system level can be constructed by real-time monitoring parameter data. A graph matching search is performed in the dynamic fuzzy network library through a graph model matching algorithm. The fault mode with the highest matching degree is considered to be the fault mode corresponding to the current state, thereby realizing fault diagnosis at the power supply system level.

[0107] The graph matching algorithm is mainly based on the Frobenius norm (F-norm). Let A = [a ij ] m×n is an adjacency matrix of an incident graph. is the F-norm of this matrix.

[0108] Let A=[a ij ] m×n is an adjacency matrix of an association graph, called is the first norm of this matrix

[0109] Based on the F-norm, we can measure the similarity of two matrices. Let the adjacency matrix of the two graphs be A = [a ij ] m×n , B=[b ij ] m×n , the graph similarity calculation formula is:

[0110] C=AB

[0111]

[0112] Through the power supply system-level fault diagnosis method based on dynamic fuzzy networks and multi-source information fusion, we can combine multi-source information such as data information and knowledge information in the power supply system to construct a dynamic fuzzy network of each component of the power supply system, explore the deep correlation between the components of the power supply system, and build a dynamic fuzzy network library based on the dynamic fuzzy network in different states. Based on the graph search and matching algorithm, graph search and matching can be performed to achieve accurate fault diagnosis at the aircraft power supply system level and provide support for maintenance decisions.

[0113] The following is an explanation of this patent using a specific case:

[0114] Another embodiment of the present invention is to use a certain type of aircraft power supply system simulation model data. The data is obtained by Simulink simulation. In the model, the aircraft power supply system is powered by four three-phase AC generators, and the four motors are respectively connected to one AC load (replaced by a three-phase resistor), and the four ACs are not connected in parallel. Each channel is connected to a 270V transformer and a 28V transformer, respectively. At the same time, the DC power supplies of channels 1 and 2 are converged, and the DC power supplies of channels 3 and 4 are converged. The DC power supplies are respectively connected to two hydraulic pump loads (high-voltage loads, working in parallel) and two resistors (low-voltage loads, working in parallel). The sampling frequency of the simulation model is 20000, and the sampling time is 3s.

[0115] This study uses simulation data under different fault modes in the data set, including normal, main generator armature winding phase short circuit, main generator armature winding single-phase open circuit, exciter excitation winding single-phase open circuit, exciter armature winding single-phase open circuit, exciter armature winding phase short circuit, rotating rectifier single diode open circuit, 28V transformer rectifier impedance degradation, 270V transformer rectifier filter inductor open circuit and other 9 states. In this data set, in order to simulate the actual use environment, noise with a signal-to-noise ratio of 5 is added to generate 50 groups of samples. The present invention uses the first 30 samples for model training, which are considered to be historical monitoring data, and uses the last 20 samples for model verification and model accuracy evaluation, which are considered to be monitoring data.

[0116] Step 1: Use sensors to collect various operating monitoring parameters of various equipment in the aircraft power supply system;

[0117] The simulation data can be considered as the various operating monitoring parameters collected by the aircraft power supply system sensors. In this data set, there are 144 types of engine monitoring parameters, among which the original data of various parameters of the 1-channel generator are shown in the figure below. Figure 2 shown.

[0118] Step 2: The sorted sensor parameter data is processed with high frequency signals to decompose the multi-dimensional information such as time domain information, frequency domain information, trend information, etc. in the original signal;

[0119] The high frequency signal processing method is Hilbert-Huang transform algorithm and effective value calculation method;

[0120] The Hilbert-Huang transform algorithm is a high-frequency signal analysis method for processing non-stationary nonlinear signals, and mainly consists of two parts, the first part is empirical mode decomposition, and the second part is Hilbert transform.

[0121] (1) First, all the maximum points of the signal x(t) are linked to form the upper envelope x up (t); and all the minimum points x low (t) is linked to the lower envelope and the signal function obtained by the selected signal should be included in it

[0122] (2) Secondly, through the upper and lower envelopes, a range can be obtained, and then based on each time unit, the mean of each corresponding upper and lower envelope is obtained, and then the mean function is subtracted from the selected signal to obtain the difference function d(t). The specific formula is as follows:

[0123]

[0124] Repeat the above operation until the difference function d(t) satisfies two conditions: the number of extreme points of the data differs from the number of zero-crossing points by zero or by at most one; at any data point, the mean of the upper envelope composed of local maximum values ​​and the lower envelope composed of local minimum points is 0.

[0125] (3) At this time, d(t) is recorded as, c1(t) is considered to be the first IMF obtained by the selected signal function, and then the corresponding residual is calculated, recorded as r1(t). The formula for calculating the residual is as follows:

[0126] r1(t)=x(t)-c1(t)

[0127] (4) Then r1(t) is used as the next selected starting signal, recorded as x(t), and recalculated according to the above algorithm. Through the above steps, the IMF components adaptively obtained for the signal can be calculated. When the residual is less than the threshold, the algorithm stops running. After empirical mode decomposition, x(t) can be converted into the following components:

[0128]

[0129] Where N is the number of IMFs, r N (t) is the final residual, c j (t) is the obtained IMF. The IMFs are mutually orthogonal and have the property of zero mean.

[0130] (5) After obtaining the signal IMF component, according to the provided signal component cj (t) is subjected to Hilbert transform, and the expression is as follows:

[0131]

[0132] Among them, y j (t) is the timing signal c j The Hilbert-Huang transform result of (t), P is the Cauchy principal value.

[0133] (6) c j (t) as the real part, and y j (t) as the imaginary part, the resulting expression is the corresponding expression, z(t) as follows:

[0134] z j (t) = c t (t)+iy j (t) = a(t)e iθ(t)

[0135] Among them, a(t) and θ t are the instantaneous amplitude and instantaneous phase of the analytical signal, which are c in the rectangular coordinate system j (t), y j (t) In polar coordinates, the two are expressed as follows:

[0136]

[0137] (7) By processing the instantaneous phase, we can obtain the value of its derivative, that is, the corresponding instantaneous frequency:

[0138]

[0139] (8) In order to save computational space, the calculated a1(t)ω1(t) is selected in this method as the instantaneous amplitude and instantaneous frequency characteristics obtained by Hilbert-Huang calculation.

[0140] Figure 3 , Figure 4 It is the instantaneous frequency and instantaneous amplitude results of the exciter output voltage signal after Hilbert-Huang transform, and the first dimension component is selected as the instantaneous amplitude and instantaneous frequency characteristics.

[0141] The effective value calculation method is as follows:

[0142] Since there are both DC power signals and AC power signals in the power supply system, the instantaneous voltage at different times is different, and there is Um (peak value). In order to facilitate the measurement and calculation of alternating current, it is necessary to define the magnitude of alternating voltage and alternating current based on the effect of alternating current, that is, the effective value. Suppose a segment of electrical signal X is x1, x2,…, xN , then the effective value of this electrical signal x rms for:

[0143]

[0144] Figure 5 This is the calculation result of feature extraction of the exciter output voltage signal. The first line is the instantaneous frequency, the second line is the instantaneous amplitude, and the third line is the effective value.

[0145] Step 3: Construct a component association network at the power supply system level, with each component as a node and the association relationship between components as an edge, and the initial state of each edge is set to 1;

[0146] The network is constructed as an undirected weighted graph. An undirected weighted graph means that the edges in the graph model are undirected and weighted. The so-called undirected means that any pair of points (i, j) and (j, i) correspond to the same edge. Vertex i and vertex j are also called the two endpoints of the undirected edge (i, j). The adjacency matrix of the undirected weighted graph G is

[0147] A=(a ij ) n×n

[0148] Where i,j∈(0,N) represents the components of the power supply system, a ij Represents the association relationship between component i and component j, based on which the power supply system association graph model can be constructed. The initial state is set to a ij =0

[0149] Step 4: Based on the multi-dimensional information, the Frechet algorithm is used to fuse the multi-dimensional information, calculate the correlation between the components, and dynamically update the fuzzy graph network to construct a dynamic fuzzy network;

[0150] The Frechet algorithm is a distance measure that calculates the distance between curves and evaluates the similarity between curves. For a given component P, Q, let its parameters be P = {u1,u2,…,u p} and Q = {v1,v2,…,v q}, we get a monotone pairing L between P and Q, which is a series of corresponding points

[0151] Where a1=1, b1=1, a m =p,b m =q. For i = 1, 2, ..., m, if and only if one of the following three conditions is met:

[0152]

[0153] Then we calculate

[0154]

[0155] in represent The Euclidean distance between the curves is:

[0156]

[0157] From this, we can calculate the association relationship between the components, that is, a PQ =d F (P,Q), based on which the dynamic fuzzy network A of the power supply system can be constructed. ij ) n×n , where i,j∈(0,N) represents the components of the power supply system, a ij Represents the association relationship between component i and component j, and the calculation method is Frechet algorithm.

[0158] Figure 6 The adjacency matrix of the dynamic fuzzy network model of the power supply system obtained based on the Frechet algorithm is represented in the form of a grayscale heat map. The grayscale represents the association between different components. The figure shows the dynamic fuzzy network of the power supply system under three fault modes: normal, main generator armature winding phase short circuit, and exciter excitation winding single-phase open circuit.

[0159] Step 5: Combine the power supply system structure information, integrate knowledge information through graph theory, and simplify the graph model structure;

[0160] The graph simplification method is: based on the knowledge of the power supply system structure, construct a power supply system level component association matrix B = (b ij ) n×n , where i,j∈(0,N) represents the components of the power supply system, b ij =1 indicates that there is an association relationship between component i and component j, b ij =0 means there is no association between components. ij ) n×n With the correlation matrix B = (b ij ) n×n For comparison, if b ij =0, then it is considered that a ij =0, if b ij ≠0, then it is considered that a ij =a ij ,The graph reduction algorithm based on knowledge fusion can eliminate the edges that have no diagnostic significance in the dynamic fuzzy network based on the knowledge of the power supply system, simplify the network structure, and thus achieve the simplification of the graph model.

[0161] Step 6: Construct a power supply system-level dynamic fuzzy network under different fault modes based on fault data;

[0162] Based on the monitoring parameter data under 9 states, including normal, main generator armature winding interphase short circuit, main generator armature winding single-phase open circuit, exciter excitation winding single-phase open circuit, exciter armature winding single-phase open circuit, exciter armature winding interphase short circuit, rotating rectifier single diode open circuit, 28V transformer rectifier impedance degradation, 270V transformer rectifier filter inductor open circuit, etc., a dynamic fuzzy network of the power supply system is constructed, thereby constructing a dynamic fuzzy network library. Figure 7 The dynamic fuzzy network of the power supply system under three simplified fault modes: normal, main generator armature winding phase short circuit, and exciter excitation winding single-phase open circuit. It can be seen that the relationships between most of the edges in the matrix are simplified, which greatly saves computing resources.

[0163] Step 7: A real-time dynamic fuzzy network at the power supply system level can be constructed by real-time monitoring of parameter data. A graph matching search is performed in the dynamic fuzzy network library through a graph model matching algorithm. The fault mode with the highest matching degree is considered to be the fault mode corresponding to the current state, thereby realizing fault diagnosis at the power supply system level.

[0164] The graph matching algorithm is mainly based on the Frobenius norm (F-norm). Let A = [a ij ] m×n is an adjacency matrix of an incident graph. is the F-norm of this matrix.

[0165] Let A=[a ij ] m×n is an adjacency matrix of an association graph, called is the first norm of this matrix

[0166] Based on the F-norm, we can measure the similarity of two matrices. Let the adjacency matrix of the two graphs be A = [a ij ] m×n , B=[b ij ] m×n , the graph similarity calculation formula is:

[0167] C=AB

[0168]

[0169] In this case, the first 30 samples are used to construct the network model, which are considered as historical monitoring data, and the last 20 samples are used to verify the model and evaluate the model accuracy, which are considered as monitoring data. It has been verified that the algorithm diagnosis accuracy is 98.3%, which has a very high diagnostic accuracy. At the same time, when a fault mode that does not exist in the dynamic fuzzy network occurs, the dynamic fuzzy network established under the new fault mode can be stored in the model library to achieve model expansion.

[0170] In summary, the present invention proposes a power supply system-level fault diagnosis method based on dynamic fuzzy network and multi-source information fusion. Based on the power supply system data and structure analysis, multi-source information fusion is performed, and multi-dimensional information such as frequency domain, time domain and trend information in the power supply system monitoring parameters and multi-source information of data information and knowledge information can be integrated. With the association relationship of each component as the edge, a dynamic fuzzy network for power supply system fault diagnosis is established, and the association relationship of each component of the power supply system is deeply mined and analyzed from the association relationship dimension. The multi-source information of the power supply system is stored and integrated by relying on the dynamic fuzzy network to realize the description of different working states of the power supply system. Based on real-time monitoring data, through feature extraction and network model construction, a dynamic fuzzy network library under the working state of the power supply system can be constructed. Based on the graph matching algorithm, a graph matching search can be performed in the dynamic fuzzy network library. The fault mode with the highest matching degree is considered to be the fault mode corresponding to the current state, and the fault diagnosis of the power supply system level is realized. A power supply system-level fault diagnosis algorithm based on dynamic fuzzy network and multi-source information fusion proposes a power supply system-level fault diagnosis method that integrates data information and knowledge information, which can realize fast and efficient power supply system-level fault diagnosis and improve the accuracy and efficiency of the diagnosis model.

[0171] Although the present invention is described in detail above, the present invention is not limited thereto, and those skilled in the art can make various modifications based on the principles of the present invention. Therefore, all modifications made based on the principles of the present invention should be understood to fall within the protection scope of the present invention.

Claims

1. A power supply system-level fault diagnosis method based on dynamic fuzzy network and multi-source information fusion, characterized by: The sensors collect various operating monitoring parameters of each device in the aircraft power supply system; the sensor parameter data is processed by high-frequency signal to decompose the time domain information, frequency domain information, trend information and other multi-dimensional information existing in the original signal; Construct the power supply system-level component association network. The graph construction method is an undirected and weighted graph. An undirected and weighted graph means that the edges in the graph model are undirected and weighted. The so-called undirected means that any point pair (i, j) and (j, i) correspond to the same edge. Vertex i and vertex j are also called the two endpoints of the undirected edge (i, j). The adjacency matrix A of the undirected and weighted graph G is (a ij ) n×n , where i,j∈(0,N) represents the components of the power supply system, a ij Represents the association relationship between component i and component j, and constructs the power supply system association graph model based on it. The initial state is set to a ij =1; Based on multidimensional information, the Frechet algorithm is used to fuse multidimensional information, calculate the association between components, and dynamically update the fuzzy graph network to build a dynamic fuzzy network. The Frechet algorithm is a distance measure that calculates the distance between curves to evaluate the similarity between curves. For given components P and Q, let its parameters be P = {u1, u2, ..., u p } and Q = {v1,v2,...,v q }, we get a monotone pairing L between P and Q, which is a series of corresponding points Where a1=1, b1=1, a m =p,b m =q, for i=1,2,...,m, if and only if one of the following three conditions is met: Then we calculate: in represent The Euclidean distance between the curves is: The relationship between the components is calculated, that is, a PQ =d F (P,Q), based on which the dynamic fuzzy network A of the power supply system is constructed = (a ij ) n×n , where i,j∈(0,N) represents the components of the power supply system, a ij Represents the association relationship between component i and component j, and the calculation method is Frechet algorithm; Combined with the power supply system structure information, the knowledge information is integrated through graph theory to simplify the structure. The graph simplification method is as follows: Based on the power supply system structure knowledge, the power supply system level component association matrix B = (b ij ) n×n , where i,j∈(0,N) represents the components of the power supply system, b ij =1 indicates that there is an association relationship between component i and component j, b ij =0 means there is no association between components. The constructed dynamic fuzzy network A = (a ij ) n×n With the correlation matrix B = (b ij ) n×n For comparison, if b ij =0, then it is considered that a ij =0, if b ij ≠0, then it is considered that a ij =a ij ,The graph reduction algorithm based on knowledge fusion can remove the edges that have no diagnostic significance in the dynamic fuzzy network based on the knowledge of the power supply system, simplify the network structure, and thus realize the simplification of the graph model; Based on fault data, dynamic fuzzy network construction at power supply system level under different fault modes is carried out; By real-time monitoring parameter data, a real-time dynamic fuzzy network at the power supply system level is constructed. Through the graph model matching algorithm, the most matching fault mode of the measured power supply system dynamic fuzzy network is found to realize fault diagnosis at the power supply system level. The graph matching algorithm is as follows: The graph matching algorithm is based on the Frobenius norm. Let A = [a ij ] m×n is an adjacency matrix of an incident graph. is the Frobenius norm of this matrix, let A = [a ij ] m×n is an adjacency matrix of an association graph, called is the first norm of this matrix. Based on the Frobenius norm, the similarity of two matrices can be measured. Suppose the adjacency matrix of the two graphs is A = [a ij ] m×n , B=[b ij ] m×n , the graph similarity calculation formula is: C=AB, 2. According to the method described in claim 1, the high-frequency signal processing method is the Hilbert-Huang transform algorithm and the effective value calculation method: the Hilbert-Huang transform algorithm is a high-frequency signal analysis method for processing non-stationary nonlinear signals, which mainly consists of two parts, the first part is the empirical mode decomposition, and the second part is the Hilbert transform.

3. According to the method of claim 2, the effective value calculation method is: Since there are both DC power signals and AC power signals in the power supply system, the instantaneous voltage at different times is different, and there is a Um peak value. In order to facilitate the measurement and calculation of alternating current, it is necessary to define the alternating voltage and alternating current based on the effect of alternating current, that is, the effective value. Suppose a section of electrical signal X is x1, x2,, x N , then the effective value of this electrical signal x rms for:

Images