A fault diagnosis method for electromechanical actuator based on virtual-real information fusion of digital twins

Abstract

The present invention provides a fault diagnosis method for an electromechanical actuator based on virtual-real information fusion of digital twins, comprising constructing an endogenous heat model of an electromechanical actuator, and further comprising the following steps: constructing a two-dimensional heat conduction model of the electromechanical actuator; reconstructing the digital twin temperature field of the electromechanical actuator system by using a physical information neural network; fusing the temperature field information reconstructed by the digital twin in a virtual space with the physical acquisition information to form a virtual-real data fusion data set of the electromechanical actuator; selecting feature extraction indicators and contribution indicators of principal component analysis for fault diagnosis; constructing a PCA-ELM fault diagnosis model by using MATLAB; using the virtual-real data fusion data set of the electromechanical actuator as training sample data to train the PCA-ELM fault diagnosis model; and diagnosing the measured signals of the actual electromechanical actuator and the information generated by the digital twin by using the PCA-ELM fault diagnosis model.

Description

Technical Field

[0001] The present invention relates to the technical field of electromechanical actuator fault diagnosis, and in particular to an electromechanical actuator fault diagnosis method based on digital twin virtual-real information fusion. Background Art

[0002] The high power density design of advanced electromechanical actuation systems is integrated with transmission structures in closed structures, which are usually used in limited spaces and harsh working environments. In such working environments, it is not possible to fully equip sensors for condition monitoring. Reconstruction of the temperature field, which is closely related to the health status of equipment operation, is an important part of condition monitoring. Usually, accurate temperature monitoring requires a large number of temperature sensors, which increases costs and reduces system availability. In most practical applications, there is not enough sensor information available, which increases the difficulty of reconstructing the entire temperature field and reduces the information and feature dimensions that can be used for fault diagnosis.

[0003] The application of electromechanical actuators has put forward higher requirements for condition monitoring and fault diagnosis technology. Digital twin (DT) is a broad concept that can reconstruct the state of physical entities in virtual space. There are different implementation paths for different devices. Digital twin has great benefits in monitoring the state of electromechanical actuators and fault diagnosis. Unlike traditional digital twins, traditional digital twins rely on deterministic physical simulation to approximate the monitored system. For complex systems, it is more recommended to create a deep digital twin (DDT), which uses deep learning to develop digital twins directly from operational data. Digital twins based on physical information neural networks belong to DDT. The reconstruction of mechanical fields, temperature fields, and electromagnetic fields has always been a research hotspot in the field of digital twins. In the actual experimental process, the measurable data is often very limited, and there is almost no direct correlation information between the measurable data and the field distribution. The traditional finite element simulation cannot integrate the actual measurement data. PINN can infer the field distribution of the entire calculation area from partial observation data, which is a classic inverse problem.

[0004] The task of fault diagnosis of complex electromechanical actuators is to identify whether the technical status of complex electromechanical actuator equipment is normal. If it is not normal, the faulty component needs to be determined. At present, the data-driven method is generally used for equipment fault diagnosis in the field of mechatronic actuator system research in China. For simple systems, it can correctly diagnose and locate common faults. However, since the newly developed electromechanical actuators are highly integrated complex systems, there are a large number of information nodes that cannot be tested in the actual installed actuators, such as the temperature field during operation. The data used for fault diagnosis can only be collected from a small number of collectable data nodes, and the fault data samples are even scarcer. Therefore, it is often difficult to conduct a comprehensive and thorough diagnosis of the system by detecting one or some local signals or using a single fault diagnosis method, and the diagnostic efficiency is extremely low. Summary of the invention

[0005] In order to solve the above technical problems, the present invention proposes a mechatronic actuator fault diagnosis method based on digital twin virtual and real information fusion, which can be applied to the maintenance and assurance of complex mechatronic equipment with complex structure, high energy density and strict volume requirements. Fault features are extracted on the basis of the fusion of virtual information reconstructed by digital twin and physical information actually collected, and mechatronic actuator health status samples are selected to learn the PCA-ELM fault diagnosis model of mechatronic actuator.

[0006] The object of the present invention is to provide a method for fault diagnosis of an electromechanical actuator based on virtual-real information fusion of digital twins, including constructing an endogenous heat model of the electromechanical actuator, and further comprising the following steps:

[0007] Step 1: Construct a two-dimensional heat conduction model of the electromechanical actuator;

[0008] Step 2: Use the physical information neural network to reconstruct the temperature field of the digital twin of the electromechanical actuation system;

[0009] Step 3: Fuse the temperature field information reconstructed by the digital twin in the virtual space with the physical acquisition information to form a virtual-real data fusion data set of the electromechanical actuator;

[0010] Step 4: Select the feature extraction index and the contribution index of principal component analysis used for fault diagnosis;

[0011] Step 5: Use MATLAB to build a PCA-ELM fault diagnosis model, set the model parameters including the number of neurons in the output layer, the number of neurons in the hidden layer, the connection weights and biases between the input layer and the hidden layer, and set the activation function of the model;

[0012] Step 6: Using the electromechanical actuator virtual and real data fusion data set as training sample data to train the PCA-ELM fault diagnosis model;

[0013] Step 7: Use the PCA-ELM fault diagnosis model to diagnose the measured signals of the actual electromechanical actuator and the digital twin generated information.

[0014] Preferably, in the thermal model of the electromechanical actuator, the energy loss of the motor is the part that generates the most heat.

[0015] In any of the above solutions, preferably, the energy loss of the motor includes copper loss, iron loss, mechanical loss and permanent magnet loss.

[0016] In any of the above schemes, preferably, the Joule heat caused by the heat generated by the current flowing through the winding generates the copper loss W cu , the formula is

[0017]

[0018] R s (T) is the resistance of the copper wire at a given temperature T, and the formula is

[0019] R s (T) = R s0 (1+α Cu (T winding -T winding_r ))

[0020] Among them, i a 、i b 、i c is the three-phase current, T is the duration of action, α cu is the temperature coefficient of resistance of the copper conductor, T winding is the winding temperature, R s0 is the nominal value of the single winding resistance, T winding_r is the reference temperature.

[0021] In any of the above schemes, preferably, the iron loss includes hysteresis loss P h , eddy current loss P e and additional loss, the formula is

[0022]

[0023] B m (T) = (1-0.0013*(T iron -T m ))B m

[0024] Among them, k h k is a constant proportional to the volume, quantity and unit of use of the core material. eis a constant value directly proportional to the volume of the core, resistivity, laminate thickness and the unit used, f is the frequency of the alternating flux, B m is the maximum magnetic flux density, x is a constant value with an interval of 1.5 to 2.5, T iron is the surface temperature of the ferromagnetic material, B m (T) is the maximum magnetic flux density under the influence of temperature, T m It is the reference temperature of ferromagnetic materials at a given magnetic flux density.

[0025] In any of the above schemes, preferably, during the rotation process, a large amount of heat is generated due to the friction between the rotor surface and the air, resulting in the mechanical energy loss G r , the calculation formula is

[0026]

[0027] Where k1 is the roughness coefficient of the rotor surface, C f is the air friction coefficient, ρ air is the air density, ω m is the angular velocity of the rotor, r rotor is the rotor radius, l rotor It refers to the axial length of the rotor.

[0028] In any of the above schemes, it is preferred that, due to the presence of tooth slots and air gap magnetic fields in the stator teeth of the motor that cannot be distributed in an ideal state, the harmonic components of the magnetic field generate eddy current losses in the permanent magnets, which are the permanent magnet losses P m , the formula is

[0029]

[0030] Among them, ρ m is the magnetic resistance, V m is the total volume of the magnet, is the magnetic flux density in the air gap, b m is the width of the magnet, ω is 1 / f, and f is the frequency.

[0031] In any of the above solutions, preferably, the parameters of the two-dimensional heat conduction model of the electromechanical actuator are set to a constant thermal conductivity, and the heat transfer equation is refined as follows:

[0032]

[0033] Where T is the temperature at a point in space (x, y), α is the thermal conductivity related to the thermal diffusivity of the material, and t is time.

[0034] In any of the above solutions, preferably, an internal heating part is added to the two-dimensional heat conduction model of the electromechanical actuator, and the formula is:

[0035]

[0036] Where g is the internal heat generation and k is the thermal conductivity of the heat source.

[0037] In any of the above schemes, it is preferred that the initial conditions and boundary conditions of the heat transfer equation are considered to be deterministic, and the formula is

[0038] T(x,y,t=0)=T0

[0039]

[0040] Where T0 is the initial temperature, T′ and q are boundary configuration parameters, T ∞ is the ambient temperature, h is the heat transfer coefficient, n is the normal vector of the space area, T(x, y, t =0) is the initial temperature at the x,y space coordinate point, T| x,y is the x,y space coordinate point, and k is the thermal conductivity.

[0041] In any of the above schemes, preferably, for the digital twin temperature field reconstruction of the electromechanical actuation system, the given observation data size is set to m, and the optimization task is

[0042]

[0043] The optimization task is transformed into an optimization problem,

[0044]

[0045] The activation function is selected as the tanh function.

[0046]

[0047] The temperature reconstruction network is expressed as

[0048] T θ :=L D σL D-1 σ…σL1

[0049]

[0050] The loss function is expressed as

[0051] L PINN_Ioss (θ) = w pde L pde +w bc L bc +w data L data

[0052]

[0053] Among them, T θ is the temperature reconstruction network, i is the number of network layers, is the i-th spatial coordinate x, is the i-th spatial coordinate y, is the temperature at the ith spatial coordinate xy, g i is the i-th internal heat source, L D is the representation of the Dth layer of the network, σ is the activation function, L1 is the first layer of the network, W1 is the weight of the first layer of the network, x is the input data node, b1 is the offset of the first layer, d1 is the number of dimensions of the first layer, d is the number of dimensions, D is the total number of network layers, w pde is the weight of partial differential loss, L pde is the partial differential loss, w bc is the boundary loss weight, L bc is the boundary loss function, w data is the data loss weight, L data is the data loss, N pde is the number of samples used to calculate the partial differential loss, N data is the number of samples used to calculate data loss, T data is the temperature data of the sample mark, N bc is the number of samples in the boundary case, W1 is the weight of the first layer network, W i is the weight of the i-th layer network, W D is the weight of the D-th layer network, b D is the offset of the D-th layer network, and N is the output dimension of the network.

[0054] In any of the above schemes, preferably, step 4 includes selecting 8 digital twin virtual information generation points for fault feature extraction, each dimension of virtual temperature information contains 5 feature quantities in the time domain, and also includes 10 groups of physical sensor signals, increasing the data dimension to 50 dimensions.

[0055] In any of the above solutions, preferably, the step 4 further comprises reducing the dimension of the data by using a principal component analysis method, in which the principal component contribution rate is designed to be no less than 95%, that is,

[0056]

[0057] Among them, k is the number of eigenvalues ​​that can be indexed, j is the eigenvalue index, and λ j is the eigenvalue of the j-dimensional feature, and p is the selected number of features.

[0058] In any of the above schemes, it is preferred that the network parameters of the PCA-ELM model set the number of input layer neurons to 7, the number of output layer neurons to 6, the number of hidden layer neurons to 36, the connection weight ω and the bias b between the input layer and the hidden layer are randomly generated before the training starts, and the activation function selects the Sigmoidal function: g(x)=1 / (1+exp(-x)).

[0059] In any of the above schemes, it is preferred that the connection weights of the input layer and the hidden layer are randomly assigned to be

[0060]

[0061] Among them, ω ln is the connection weight between the lth node in the input layer and the nth node in the hidden layer, l is the lth node in the input layer, and n is the nth node in the hidden layer.

[0062] In any of the above schemes, preferably, the output weights of the hidden layer and the output layer are

[0063]

[0064] Among them, β lm is the connection output weight between the lth node in the hidden layer and the mth node in the output layer, and m is the mth node in the output layer.

[0065] In any of the above schemes, preferably, the deviation value of the randomly assigned hidden layer neuron is

[0066]

[0067] Calculate the output of the network corresponding to the training sample data as

[0068] T=[t1 t2…t Q ] m×Q

[0069]

[0070] Among them, b l is the offset vector of layer l, t Q is the Q-th dimension output temperature, m is the number of output layer nodes, b i is the offset of the i-th layer, ω i =[ω i1 ω i2 …ω in ] and g(x) is the activation function.

[0071] In any of the above schemes, it is preferred that the output weight β of the hidden layer and the output layer is calculated as

[0072] β=H+ T′

[0073]

[0074] Among them, H + is the Moore-Penrose generalized inverse of the output matrix from the hidden layer to the output layer, and T′ is the temperature output matrix.

[0075] In any of the above solutions, preferably, step 6 includes testing the accuracy ACC of the PCA-ELM fault diagnosis model on the test set after the training reaches convergence, and when ACC>P min Save the model.

[0076] The present invention proposes a fault diagnosis method for electromechanical actuators based on the fusion of virtual and real information of digital twins, which uses virtual and real information for fault diagnosis at the same time. It overcomes the problem that there are few measured signal nodes, scarce fault data samples, and it is difficult to conduct a comprehensive and thorough diagnosis of the system, and expands the data dimension and feature dimension. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 It is a flow chart of a preferred embodiment of the electromechanical actuator fault diagnosis method based on digital twin virtual-real information fusion according to the present invention.

[0078] Figure 2 It is a flowchart of another preferred embodiment of the electromechanical actuator fault diagnosis method based on digital twin virtual-real information fusion according to the present invention.

[0079] Figure 3 The present invention is a flowchart of an embodiment of a method for reconstructing a virtual temperature field of a digital twin of an electromechanical actuator according to a method for diagnosing an electromechanical actuator fault based on digital twin virtual-real information fusion according to the present invention.

[0080] Figure 4 It is a schematic diagram of an embodiment of the framework of the extreme learning machine information fusion fault diagnosis method based on principal component analysis of the electromechanical actuator fault diagnosis method based on digital twin virtual-real information fusion according to the present invention.

[0081] Figure 5 It is a schematic diagram of an embodiment of the extreme learning machine structure of the electromechanical actuator fault diagnosis method based on digital twin virtual-real information fusion according to the present invention. DETAILED DESCRIPTION

[0082] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0083] Embodiment 1

[0084] like Figure 1As shown, step 100 is executed to construct an internal heat model of the electromechanical actuator. In the electromechanical actuator thermal model, the energy loss of the motor is the part that generates the most heat. The energy loss of the motor includes copper loss, iron loss, mechanical loss and permanent magnet loss.

[0085] The heat generated by the current flowing through the winding causes Joule heat to generate the copper loss W cu , the formula is

[0086]

[0087] R s (T) is the resistance of the copper wire at a given temperature T, and the formula is

[0088] R s (T) = R s0 (1+α cu (T winding -T winding_r ))

[0089] Among them, i a 、i b 、i c is the three-phase current, T is the duration of action, α cu is the temperature coefficient of resistance of the copper conductor, T winding is the winding temperature, R s0 is the nominal value of the single winding resistance, T winding_r is the reference temperature.

[0090] The iron loss includes hysteresis loss P h , eddy current loss P e and additional loss, the formula is

[0091]

[0092]

[0093] B m (T) = (1-0.0013*(T iron -T m ))B m

[0094] Among them, k h k is a constant proportional to the volume, quantity and unit of use of the core material. e is a constant value directly proportional to the core volume, resistivity, laminate thickness and unit used, f is the frequency of the alternating flux, B m is the maximum magnetic flux density, x is a constant value with an interval of 1.5 to 2.5, T iron is the surface temperature of the ferromagnetic material, B m(T) is the maximum magnetic flux density under the influence of temperature, T m It is the reference temperature of ferromagnetic materials at a given magnetic flux density.

[0095] During the rotation process, a large amount of heat is generated due to the friction between the rotor surface and the air, resulting in the mechanical energy loss G r , the calculation formula is

[0096]

[0097] Where k1 is the roughness coefficient of the rotor surface, C f is the air friction coefficient, ρ air is the air density, ω m is the angular velocity of the rotor, r rotor is the rotor radius, l rotor It refers to the axial length of the rotor.

[0098] Since there are tooth slots and air gap magnetic fields in the stator teeth of the motor that cannot be distributed in an ideal state, the harmonic components of the magnetic field generate eddy current losses in the permanent magnets, which are the permanent magnet losses P m , the formula is

[0099]

[0100] Among them, ρ m is the magnetic resistance, V m is the total volume of the magnet, is the magnetic flux density in the air gap, b m is the width of the magnet, ω is 1 / f, and f is the frequency.

[0101] Execute step 110 to construct a two-dimensional heat conduction model of the electromechanical actuator, set the parameters of the two-dimensional heat conduction model of the electromechanical actuator to a constant thermal conductivity, and refine the heat transfer equation to:

[0102]

[0103] Where T is the temperature at a point in space (x, y), α is the thermal conductivity related to the thermal diffusivity of the material, and t is time.

[0104] Adding the internal heat generation part to the two-dimensional heat conduction model of the electromechanical actuator, the formula is:

[0105]

[0106] Where g is the internal heat generation and k is the thermal conductivity of the heat source.

[0107] The initial and boundary conditions of the heat transfer equation are considered to be deterministic, and the formula is

[0108] T(x,y,t=0)=T0

[0109]

[0110] Where T0 is the initial temperature, T′ and q are boundary configuration parameters, T ∞ is the ambient temperature, h is the heat transfer coefficient, n is the normal vector of the space area, T(x, y, t =0) is the initial temperature at the x,y space coordinate point, T| x,y is the x,y space coordinate point, and k is the thermal conductivity.

[0111] Execute step 120 to reconstruct the digital twin temperature field of the electromechanical actuation system using the physical information neural network. For the digital twin temperature field reconstruction of the electromechanical actuation system, the given observation data size is set to m, and the optimization task is

[0112]

[0113] Transform the optimization task into an optimization problem.

[0114]

[0115]

[0116] The activation function is selected as the tanh function.

[0117]

[0118] The temperature reconstruction network is expressed as

[0119] T θ :=L D σL D-1 σ…σL1

[0120]

[0121] L PINN_loss (θ) = w pde L pde +w bc L bc +w data L data

[0122]

[0123] Among them, T θ is the temperature reconstruction network, i is the number of network layers, is the i-th spatial coordinate x, is the i-th spatial coordinate y, is the temperature at the ith spatial coordinate xy, g i is the i-th internal heat source, L D is the representation of the Dth layer of the network, σ is the activation function, L1 is the first layer of the network, W1 is the weight of the first layer of the network, x is the input data node, b1 is the offset of the first layer, d1 is the dimension of the first layer, d is the number of dimensions, D is the total number of network layers, w pde is the weight of partial differential loss, L pde is the partial differential loss, w bc is the boundary loss weight, L bc is the boundary loss function, w data is the data loss weight, L data is the data loss, N pde is the number of samples used to calculate the partial differential loss, N data is the number of samples used to calculate data loss, T data is the temperature data of the sample mark, N bc is the number of samples in the boundary case, W1 is the weight of the first layer network, W i is the weight of the i-th layer network, W D is the weight of the D-th layer network, b D is the offset of the D-th layer network, and N is the output dimension of the network.

[0124] Execute step 130 to fuse the temperature field information reconstructed by the digital twin in the virtual space with the physical acquisition information to form a virtual-real data fusion data set of the electromechanical actuator;

[0125] Execute step 140 to select feature extraction indicators and principal component analysis contribution indicators for fault diagnosis. Select 8 digital twin virtual information generation points for fault feature extraction. Each dimension of virtual temperature information contains 5 feature quantities in the time domain and 10 groups of physical sensor signals. The data dimension is increased to 50 dimensions. The principal component analysis method is used to reduce the data dimension. In the principal component analysis method, the principal component contribution rate is designed to be no less than 95%, that is,

[0126]

[0127] Among them, k is the number of eigenvalues ​​that can be indexed, j is the eigenvalue index, and λ j is the eigenvalue of the j-dimensional feature, and p is the selected number of features.

[0128] Execute step 150, use MATLAB to build a PCA-ELM fault diagnosis model, set model parameters including the number of neurons in the output layer, the number of neurons in the hidden layer, the connection weights and biases between the input layer and the hidden layer, and set the activation function of the model. The network parameters of the PCA-ELM model set the number of neurons in the input layer to 7, the number of neurons in the output layer to 6, the number of neurons in the hidden layer to 36, the connection weights ω and the bias b between the input layer and the hidden layer are randomly generated before the training starts, and the activation function selects the Sigmoidal function: g(x)=1 / (1+exp(-x)).

[0129] The connection weights between the input layer and the hidden layer are randomly assigned to be

[0130]

[0131] Among them, ω ln is the connection weight between the lth node in the input layer and the nth node in the hidden layer, l is the lth node in the input layer, and n is the nth node in the hidden layer.

[0132] The output weights of the hidden layer and the output layer are

[0133]

[0134] Among them, β lm is the connection output weight between the lth node in the hidden layer and the mth node in the output layer, and m is the mth node in the output layer.

[0135] The bias value of the randomly assigned hidden layer neurons is

[0136] b=[b1 b2…b l ] T

[0137] Calculate the output of the network corresponding to the training sample data as

[0138] T=[t1 t2…t Q ] m×Q

[0139]

[0140] Among them, b l is the offset vector of layer l, t Q is the Q-th dimension output temperature, m is the number of output layer nodes, b i is the offset of the i-th layer, ω i =[ω i1 ω i2 …ω in ] and g(x) is the activation function.

[0141] Calculate the output weight β of the hidden layer and the output layer as

[0142] β=H + T′

[0143]

[0144] Among them, H + is the Moore-Penrose generalized inverse of the output matrix from the hidden layer to the output layer, and T′ is the temperature output matrix.

[0145] Execute step 160, use the electromechanical actuator virtual and real data fusion data set as training sample data, train the PCA-ELM fault diagnosis model, and test the accuracy ACC of the PCA-ELM fault diagnosis model on the test set after the training reaches convergence. min Save the model.

[0146] Execute step 170 to diagnose the measured signal of the actual electromechanical actuator and the digital twin generated information using the PCA-ELM fault diagnosis model.

[0147] Embodiment 2

[0148] The present invention provides a method for fault diagnosis of an electromechanical actuator based on digital twin virtual-real information fusion. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0149] A fault diagnosis method for electromechanical actuators based on digital twin virtual-real information fusion is implemented in the following steps:

[0150] Step 1: Construct a heat model for the electromechanical actuator;

[0151] The part that generates the most heat in the thermal model of the electromechanical actuator is the energy loss of the motor. The losses of the motor mainly include three categories, namely copper loss, iron loss and mechanical loss. Due to the heat generated by multiple structures of the motor, in order to simplify the model, the power loss of the motor components mainly considers copper loss, iron loss, mechanical loss and permanent magnet loss.

[0152] Step 2: Build a 2D heat transfer model of the electromechanical actuator

[0153] In heat transfer analysis, the heat conduction equation is solved in the canonical space under the assumption that the basic heat conduction law is followed. Since the focus of the internal temperature of the electromechanical actuator is the entire temperature field rather than a single point, a two-dimensional heat conduction model needs to be constructed.

[0154] Step 3: Construct the temperature field reconstruction of the digital twin of the electromechanical system;

[0155] Reconstruct the temperature field of the digital twin of the electromechanical actuation system using physical information neural network;

[0156] Step 4: Forming a virtual and real data fusion data set of the electromechanical actuator;

[0157] The temperature field information reconstructed by the digital twin in the virtual space and the physical acquisition information are fused to form a virtual-real data fusion data set of the electromechanical actuator.

[0158] Step 5: Set feature extraction index and principal component contribution rate;

[0159] Select feature extraction indicators for fault diagnosis and contribution indicators for principal component analysis.

[0160] Step 6: Construct PCA-ELM fault diagnosis model;

[0161] Use MATLAB to build the PCA-ELM model, set the model parameters including the number of neurons in the output layer, the number of neurons in the hidden layer, the connection weights and biases between the input layer and the hidden layer, and set the activation function of the model;

[0162] Step 7: Train the PCA-ELM fault diagnosis model;

[0163] Using the data set obtained in step 4 as training sample data, the extreme learning machine fault diagnosis model is learned;

[0164] Step 8: Diagnose the measured signals of the actual electromechanical actuator and the information generated by the digital twin;

[0165] Use the ELM fault diagnosis model trained in step 7 to diagnose the measured signals of the actual electromechanical actuator and the digital twin generated information.

[0166] The present invention discloses a method for fault diagnosis of electromechanical actuators based on virtual-real information fusion of digital twins in the technical field of fault diagnosis of electromechanical actuators with virtual-real data fusion, which can be applied to the maintenance and guarantee of complex mechatronic equipment with complex structures, high energy density and strict volume requirements. Fault features are extracted on the basis of the virtual information reconstructed by the fusion of digital twins and the actual collected physical information, and the health status samples of the electromechanical actuators are selected to learn the PCA-ELM fault diagnosis model of the electromechanical actuators. In this diagnostic method, virtual and real information are used for fault diagnosis at the same time, which overcomes the problem that there are few measured signal nodes, few fault data samples, and it is difficult to conduct a comprehensive and thorough diagnosis of the system, and the data dimension and feature dimension are expanded.

[0167] Embodiment 3

[0168] The basic process of the electromechanical actuator fault diagnosis method based on digital twin virtual-real information fusion of the present invention is as follows: Figure 2 As shown, the technical approach of the method is implemented by the following steps:

[0169] Step 1: Model the internal heat generation in electromechanical actuators

[0170] The part that generates the most heat in the thermal model of the electromechanical actuator is the energy loss of the motor. The losses of the motor mainly include three categories, namely copper loss, iron loss and mechanical loss. Due to the heat generated by multiple structures of the motor, in order to simplify the model, the power loss of the motor components mainly considers copper loss, iron loss, mechanical loss and permanent magnet loss. The power loss analysis of the motor components is shown in Table 1.

[0171] Table 1 Proportion of power loss in each part of the motor

[0172]

[0173] In the motor, the main heat sources are the copper loss in the winding and the stator loss, which cannot be ignored. At the same time, due to the heat resistance of the permanent magnet, the heating of the permanent magnet is also a key consideration in this paper. The following assumptions are made:

[0174] 1) The loss distribution in the winding and core is uniform.

[0175] 2) The axial thermal conductivity is significantly lower than the radial thermal conductivity.

[0176] 3) It can usually be assumed that the heat flow around is symmetrical.

[0177] 4) Heat flows evenly from the inner boundary to the outer boundary.

[0178] (1) Copper loss

[0179] Copper losses are caused by Joule heating due to the heat generated by the current flowing through the windings.

[0180]

[0181] Among them, i a 、i b 、i c is the three-phase current [A], R s (T) is the resistance value of the stator winding [], and T is the duration of action [s]. The current flowing through the resistor generates heat, causing the temperature to rise, and the resistance value of the resistor also changes with the increase in temperature. Therefore, at a given temperature T, the resistance of the copper wire is:

[0182] R s (T) = R s0 (1+α cu (T winding -T winding_r )) (2)

[0183] In the formula, α cuis the temperature coefficient of resistance of the copper conductor, which is 0.004141. winding is the winding temperature [℃], R s0 is the nominal value of the single winding resistance [℃]; T winding_r is the reference temperature [℃], set to 20℃.

[0184] (2) Iron loss

[0185] Iron losses can be divided into hysteresis losses, eddy current losses and additional losses. Iron losses are determined by weight and depend on frequency and maximum flux density. Magnet losses are usually small, but this is critical because magnets usually do not have good temperature resistance. Iron losses can be formulated as:

[0186]

[0187] B m (T) = (1-0.0013*(TT m ))B m

[0188] Among them, P h is the hysteresis loss, P e is the eddy current loss, k h is a constant proportional to the volume, quantity and unit of use of the core material, k e It is a constant value directly proportional to the volume of the core, resistivity, laminate thickness and the unit used, B m is the maximum magnetic flux density [Wb / m2], f is the frequency of the alternating magnetic flux [Hz], and x is a constant value with an interval of 1.5 to 2.5. T is the surface temperature of the ferromagnetic material [℃], T m is the reference temperature of a ferromagnetic material at a given magnetic flux density [°C],

[0189] (3) Mechanical energy loss

[0190] During the rotation process, a lot of heat is generated due to the friction between the rotor surface and the air. The calculation formula is:

[0191]

[0192] Where k1 is the roughness coefficient of the rotor surface, C f is the air friction coefficient, ρ air is the air density [kg / m3], ω m is the angular velocity of the rotor [rad / s], r rotor is the rotor radius [m]; l rotor It refers to the axial length of the rotor [m].

[0193] (4) Permanent magnet loss

[0194] In actual operation, due to the existence of tooth slots and air gap magnetic fields in the motor stator teeth that cannot be distributed in an ideal state, the harmonic components of the magnetic field generate eddy current losses in the permanent magnets. The calculation formula is:

[0195]

[0196] Where f is the frequency [Hz], ρ m is the magnetic resistance [H / m], V m is the total volume of the magnet [m 2 ], B m is the magnetic flux density in the air gap [Wb / m 3 ], B m is the width of the magnet [m].

[0197] Step 2: Build a 2D heat transfer model of the electromechanical actuator

[0198] In heat transfer analysis, the heat conduction equation is solved in the canonical space under the assumption that the basic heat conduction law is followed. Since the focus of the internal temperature of the electromechanical actuator is the entire temperature field rather than a single point, a two-dimensional heat conduction model needs to be constructed.

[0199] Considering rectangular coordinates with two dimensions and setting the parameters to constant thermal conductivity, the heat transfer equation can be refined as:

[0200]

[0201] Where T is the temperature at a point (x, y) in space, α is the thermal conductivity related to the thermal diffusivity of the material, ξ is the density of the material, and C is the specific heat capacity.

[0202] Especially when considering internal heating, an internal heating section should be added as follows:

[0203]

[0204] Where g is the internal heat generation. In addition, the initial and boundary conditions of the heat transfer equation are considered deterministic in this paper.

[0205] T(x,y,t=0)=T0

[0206]

[0207] Where T0 is the initial temperature, T′ and q are boundary configuration parameters, T ∞ is the ambient temperature, h is the heat transfer coefficient, and n is the normal vector of the space area.

[0208] Step 3: Construct the digital twin temperature field reconstruction of the electromechanical system

[0209] Using physical information neural network to reconstruct the temperature field of the digital twin of the electromechanical actuation system (PINN-HT). Figure 3 shown.

[0210] For the PINN-HT method, given the observed data size is set to m, the optimization task can be written as:

[0211]

[0212] In summary, the task can be transformed into an optimization problem as shown in formula (10).

[0213]

[0214] The activation function is selected as the tanh function:

[0215]

[0216] The temperature reconstruction network is expressed as formula (12):

[0217] T θ :=L D σL D-1 σ…σL1

[0218]

[0219] The loss function of the PINN-HT method is expressed as:

[0220] L PINN_loss (θ) = w pde L pde +w bc L bc +w data L data (13)

[0221]

[0222] Step 4: Forming a virtual-real data fusion dataset of the electromechanical actuator

[0223] The temperature field information reconstructed by the digital twin in the virtual space and the physical acquisition information are combined to form a virtual-real data fusion data set of the electromechanical actuator.

[0224] Step 5: Set feature extraction index and principal component contribution rate

[0225] Select feature extraction indicators for fault diagnosis and contribution indicators for principal component analysis.

[0226] Temperature information cannot be directly entered into the fault diagnosis module. It is necessary to extract the features that change significantly when the system is abnormal through feature extraction to improve the efficiency of fault identification.

[0227] In order to maximize the efficiency of the algorithm, the principle is to simplify the data dimension as much as possible. Some dimensionless characteristic indicators are selected: the mean sub-C of the time domain, the variance C var , variance C s , kurtosis K and skewness S.

[0228] Specifically, 8 digital twin virtual information generation points are selected for fault feature extraction. Each dimension of virtual temperature information contains 5 feature quantities in the time domain. In addition, it also includes 10 groups of physical sensor signals such as voltage, current, and magnetic flux. The data dimension is therefore increased to 50 dimensions, which includes temperature information that is not sensitive to specific fault information and other redundant features. Its existence increases the pressure on the fault diagnosis system and affects the judgment of faults. Therefore, the principal component analysis (PCA) method is used to reduce the dimension of the data.

[0229] In the principal component analysis method, the vector variance represents the contribution rate of the corresponding principal component to the overall data. In order to retain as many data features of the original data as possible, the principal component contribution rate is designed to be no less than 95%, that is:

[0230]

[0231] The dimension of the data is reduced to the greatest extent, and most of the information of the data is retained. As the input of fault diagnosis, it will greatly reduce the pressure of the diagnosis module and reduce the complexity of the algorithm.

[0232] Step 6: Construct PCA-ELM fault diagnosis model

[0233] Diagnostic procedures such as Figure 4 As shown in the figure, the structure of the more specific diagnosis model network is as follows Figure 5 As shown in the figure, the number of input layer neurons is set to 7, the number of output layer neurons is set to 6, and the number of hidden layer neurons is set to 36. The connection weights ω and bias b between the input layer and the hidden layer are randomly generated before the training begins, and the activation function is the Sigmoidal function: g(x) = 1 / (1+exp(-x)).

[0234] Specifically, the connection weights between the input layer and the hidden layer are randomly assigned:

[0235]

[0236] In the formula, ω ln is the connection weight between the lth node in the input layer and the nth node in the hidden layer.

[0237] Assume the output weights of the hidden layer and the output layer are:

[0238]

[0239] In the formula, β lm is the connection output weight between the lth node in the hidden layer and the mth node in the output layer.

[0240] Randomly assign bias values ​​to hidden layer neurons:

[0241] b=[b1 b2 … b l ] T (17)

[0242] The output of the network corresponding to the training sample data is calculated as:

[0243] T=[t1 t2 …t Q ] m×Q (18)

[0244]

[0245] In the formula, ω i =[ω i1 ω i2 …ω in ] and g(x) is the activation function.

[0246] It can be written as: Hβ = T', where H is the hidden layer output matrix:

[0247]

[0248] The output weight β of the hidden layer and the output layer is calculated as:

[0249] β=H + T′ (20)

[0250] Among them, H + is the Moore-Penrose generalized inverse of the output matrix from the hidden layer to the output layer.

[0251] When the training reaches convergence, the accuracy ACC of the PCA-ELM fault diagnosis model on the test set is tested. When ACC>P min Save the model.

[0252] Step 7: Train the PCA-ELM fault diagnosis model

[0253] Using the data set obtained in step 4 as training sample data, the extreme learning machine fault diagnosis model of the electromechanical actuation system is learned;

[0254] Specifically, 100 sets of data in each of the normal state, five typical faults and multiple faults of the solid-state power distribution unit in the data set, a total of 600 sets of data, are divided into training set, validation set and test set according to 6:2:2. The data after dimensionality reduction by principal component analysis is used as input, and a total of five faults plus normal states are output. The test set and validation set are used for training and verification of the PCA-ELM model.

[0255] Step 8: Diagnose the measured signals of the actual electromechanical actuator and the information generated by the digital twin

[0256] Use the ELM fault diagnosis model trained in step 7 to diagnose the measured signals of the actual electromechanical actuator and the digital twin generated information.

[0257] In order to better understand the present invention, the above is described in detail in conjunction with the specific embodiments of the present invention, but it is not intended to limit the present invention. Any simple modifications made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention. Each embodiment in this specification focuses on the differences from other embodiments, and the same or similar parts between the embodiments can be referenced to each other. For the system embodiment, since it basically corresponds to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

Claims

1. A method for fault diagnosis of an electromechanical actuator based on digital twin virtual-real information fusion, comprising constructing an endogenous heat model of the electromechanical actuator, characterized in that: The following steps are also included: Step 1: Construct a two-dimensional heat conduction model of the electromechanical actuator; Step 2: Use the physical information neural network to reconstruct the temperature field of the digital twin of the electromechanical actuation system; Step 3: Fuse the temperature field information reconstructed by the digital twin in the virtual space with the physical acquisition information to form a virtual-real data fusion data set of the electromechanical actuator; Step 4: Select the feature extraction index and the contribution index of principal component analysis used for fault diagnosis; Step 5: Use MATLAB to build a PCA-ELM fault diagnosis model, set the model parameters including the number of neurons in the output layer, the number of neurons in the hidden layer, the connection weights and biases between the input layer and the hidden layer, and set the activation function of the model; Step 6: Using the electromechanical actuator virtual and real data fusion data set as training sample data to train the PCA-ELM fault diagnosis model; Step 7: Use the PCA-ELM fault diagnosis model to diagnose the measured signals of the actual electromechanical actuator and the digital twin generated information.

2. The electromechanical actuator fault diagnosis method based on digital twin virtual-real information fusion according to claim 1 is characterized in that: In the thermal model of the electromechanical actuator, the energy loss of the motor is the part that generates the most heat, and the energy loss of the motor includes copper loss, iron loss, mechanical loss and permanent magnet loss.

3. The electromechanical actuator fault diagnosis method based on digital twin virtual-real information fusion according to claim 2 is characterized in that: The heat generated by the current flowing through the winding causes Joule heat to generate the copper loss W cu , the formula is R s (T) is the resistance of the copper wire at a given temperature T, and the formula is Among them, i a 、i b 、i c is the three-phase current, T is the duration of action, α cu is the temperature coefficient of resistance of the copper conductor, T winding is the winding temperature, R s0 is the nominal value of the single winding resistance, T winding_r is the reference temperature.

4. The electromechanical actuator fault diagnosis method based on digital twin virtual-real information fusion according to claim 3 is characterized in that: The parameters of the two-dimensional heat conduction model of the electromechanical actuator are set to a constant thermal conductivity, and the heat transfer equation is refined to: Where T is the temperature at a point in space (x, y), α is the thermal conductivity related to the thermal diffusivity of the material, and t is time.

5. The electromechanical actuator fault diagnosis method based on digital twin virtual-real information fusion according to claim 4 is characterized in that: Adding the internal heat generation part to the two-dimensional heat conduction model of the electromechanical actuator, the formula is: Where g is the internal heat generation and k is the thermal conductivity of the heat source.

6. The electromechanical actuator fault diagnosis method based on digital twin virtual-real information fusion according to claim 5 is characterized in that: The initial and boundary conditions of the heat transfer equation are considered to be deterministic, and the formula is Where T0 is the initial temperature, and q are boundary configuration parameters, T ∞ is the ambient temperature, h is the heat transfer coefficient, n is the normal vector of the space area, T(x,y,t=0) is the initial temperature at the x,y space coordinate point, T| x,y is the temperature at the x,y space coordinate point, and k is the thermal conductivity.

7. The electromechanical actuator fault diagnosis method based on digital twin virtual-real information fusion according to claim 6 is characterized in that: For the digital twin temperature field reconstruction of the electromechanical actuation system, the given observation data size is set to m, and the optimization task is Transform the optimization task into an optimization problem. The activation function is selected as the tanh function. The temperature reconstruction network is expressed as The loss function is expressed as Among them, T θ is the temperature reconstruction network, i is the number of network layers, is the i-th spatial coordinate x, is the i-th spatial coordinate y, is the temperature at the ith spatial coordinate xy, g i is the i-th internal heat source, L D is the representation of the Dth layer of the network, σ is the activation function, L1 is the first layer of the network, W1 is the weight of the first layer of the network, x is the input data node, b1 is the offset of the first layer, d1 is the number of dimensions of the first layer, d is the number of dimensions, D is the total number of network layers, w pde is the weight of partial differential loss, L pde is the partial differential loss, w bc is the boundary loss weight, L bc is the boundary loss function, w data is the data loss weight, L data is data loss, |N pde | is the number of samples used to calculate the partial differential loss, |N data | is the number of samples used to calculate data loss, T data is the temperature data of the sample mark, |N bc | is the number of samples in the boundary case, W1 is the weight of the first layer network, W i is the weight of the i-th layer network, W D is the weight of the D-th layer network, b D is the offset of the D-th layer network, and N is the output dimension of the network.

8. The electromechanical actuator fault diagnosis method based on digital twin virtual-real information fusion according to claim 7 is characterized in that: The step 4 includes selecting 8 digital twin virtual information generation points for fault feature extraction, each dimension of virtual temperature information includes 5 feature quantities in the time domain, and also includes 10 groups of physical sensor signals, increasing the data dimension to 50 dimensions; The principal component analysis method is used to reduce the dimension of the data. In the principal component analysis method, the contribution rate of the principal component is designed to be no less than 95%, that is, Among them, k is the number of eigenvalues ​​that can be indexed, j is the eigenvalue index, and λ j is the eigenvalue of the j-dimensional feature, and p is the selected number of features.

9. The electromechanical actuator fault diagnosis method based on digital twin virtual-real information fusion according to claim 8, characterized in that: The network parameters of the PCA-ELM model set the number of input layer neurons to 7, the number of output layer neurons to 6, the number of hidden layer neurons to 36, the connection weight ω and the bias b between the input layer and the hidden layer are randomly generated before the training starts, and the activation function selects the Sigmoidal function: g(x)=1 / (1+exp(-x)).

10. The electromechanical actuator fault diagnosis method based on digital twin virtual-real information fusion according to claim 9 is characterized in that: The connection weights between the input layer and the hidden layer are randomly assigned to Among them, ω ln is the connection weight between the lth node in the input layer and the nth node in the hidden layer, l is the lth node in the input layer, and n is the nth node in the hidden layer; The output weights of the hidden layer and the output layer are Among them, β lm is the connection output weight between the lth node in the hidden layer and the mth node in the output layer, and m is the mth node in the output layer; The bias value of the randomly assigned hidden layer neurons is b =[b1 b2 … b l ] T Calculate the output of the network corresponding to the training sample data as T=[t1 t2 …t Q ] m×Q Among them, b l is the offset vector of layer l, t Q is the Q-th dimension output temperature, m is the number of output layer nodes, b i is the offset of the i-th layer, ω i =[ω i1 ω i2 …ω in ] and g(x) are activation functions.

Images