A Fault Diagnosis Method for Electromechanical Servo Systems Based on Simulation-Driven and Subdomain Adaptation

Abstract

This invention discloses a fault diagnosis method for electromechanical servo systems based on simulation-driven and subdomain adaptation, comprising: constructing a simulation model of the electromechanical servo system; setting the parameters of the simulation model to obtain a high-accuracy simulation model; performing fault mechanism analysis, implementing fault injection, and obtaining a large number of simulation fault datasets; inputting the simulation fault datasets into the SDCTF method to achieve fault classification and obtain a pre-trained model; inputting experimental data, using the simulation data as the source domain and the experimental data as the target domain, minimizing the distribution differences between different subdomains of the source and target domains based on the local maximum mean difference distance metric method, fine-tuning the model parameters, and achieving fault diagnosis of the experimental data. This invention can simultaneously extract the local spatial features and global temporal features of multi-sensor signals from the electromechanical servo system and adaptively fuse the features of multi-sensor signals.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for electromechanical servo systems, and in particular to a fault diagnosis method for electromechanical servo systems based on simulation-driven and subdomain adaptation. Background Technology

[0002] Electromechanical servo systems are complex control systems integrating mechanical, electrical, and electronic technologies. They are important fly-by-wire actuators and have received widespread attention and application in the aerospace field. However, due to their long-term operation in complex environments with varying loads and speeds, they are prone to failure. Once a failure occurs, it can lead to serious consequences, even endangering the safety of the aircraft and its occupants. In short, the safety and reliability of electromechanical servo systems determine the safety and reliability of aircraft operation. Therefore, researching fault diagnosis methods and achieving automatic fault identification in electromechanical servo systems is of great significance for ensuring their safety and reliability.

[0003] Deep learning's powerful feature extraction and nonlinear representation capabilities have led to its widespread application in fault diagnosis. For example, Convolutional Neural Networks (CNNs) have strong spatial feature extraction capabilities when processing image data; Recurrent Neural Networks (RNNs), Long Memory Units (LSTMs), and Gated Recurrent Units (GRUs) have strong memory capabilities and the ability to capture contextual dependencies when processing sequence data. However, deep learning methods require a large amount of data to train the model; if the training samples are insufficient, the diagnostic rate of the model will be severely affected. In practice, however, fault experiments on electromechanical servo systems are costly, dangerous, and have limited fault data samples. Expanding the dataset by building simulation models to obtain simulation data is a common method to address the problem of insufficient samples. For example, simulation modeling of equipment such as engines and bearings solves the problem of missing samples in reality. Although expanding the dataset through simulation models solves the problem of insufficient samples, the distribution difference between simulation data and experimental data can still reduce the model's diagnostic rate. Domain adaptation, as a branch of transfer learning, can be used to address the problem of distribution differences. It reduces distribution differences by learning domain-invariant features. For example, distance metrics such as MMD and CORAL, and adversarial network methods can be used to reduce feature distribution differences. Although transfer learning has been widely used in fault diagnosis, its application in electromechanical servo systems still faces the following challenges:

[0004] (1) There are few reference materials on simulation models of electromechanical servo systems, making it difficult to establish a high-accuracy simulation model.

[0005] (2) Most fault diagnosis methods based on transfer learning use CNN as a feature extractor. However, CNN only focuses on the local spatial features of the signal and cannot capture the temporal dependence of the time series, thus losing important temporal information of the time series.

[0006] (3) Most of the fault diagnosis methods currently used for mechanical equipment are for single vibration signals, while electromechanical servo systems have many sensors, and such methods will lose a lot of useful information. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a fault diagnosis method for electromechanical servo systems based on simulation-driven and subdomain adaptation, which can simultaneously extract the local spatial features and global temporal features of multi-sensor signals of electromechanical servo systems, and adaptively fuse the features of multi-sensor signals. This solves the problem that CNN cannot capture the temporal information of signals and the problem that most current fault diagnosis algorithms are only used to process single vibration signals.

[0008] The present invention solves the technical problem by adopting the following technical solution:

[0009] A fault diagnosis method for electromechanical servo systems based on simulation-driven and subdomain adaptation includes the following steps:

[0010] Step S1: Construct a simulation model of the electromechanical servo system;

[0011] Step S2: Set the parameters of the electromechanical servo system simulation model to ensure that the simulation data of the electromechanical servo system is consistent with the experimental data of the experimental platform and has a small amplitude error, thereby obtaining a high-accuracy electromechanical servo system simulation model; on this basis, perform fault mechanism analysis, realize fault injection, and obtain a large number of simulation fault datasets.

[0012] Step S3: Input the simulated fault dataset into the SDCTF method to achieve fault classification and obtain a pre-trained model;

[0013] Step S4: Based on the obtained pre-trained model, input experimental data, take the simulation data as the source domain and the experimental data as the target domain, minimize the distribution difference between different subdomains of the source domain and the target domain based on the local maximum mean difference distance metric method, fine-tune the model parameters, and realize fault diagnosis of experimental data.

[0014] Furthermore, in step S1, the electromechanical servo system is divided into five modules through a modular modeling strategy: permanent magnet synchronous motor, planetary roller screw, controller, driver, and spring load. Mathematical models are performed on each of the five modules, and the design is completed in Simulink. Finally, all modules are connected according to logical relationships to obtain the simulation model of the electromechanical servo system.

[0015] Furthermore, the modeling of the permanent magnet synchronous motor includes three parts: the three-phase stator voltage equation, the electromagnetic torque equation, and the motor mechanical motion equation. The input is the three-phase voltage, and the output is the motor speed, electromagnetic torque, and three-phase current.

[0016] Furthermore, the specific method of step S2 includes:

[0017] Step S21: Based on the inherent parameters of the motor, lead screw, and load spring in the experimental platform, including motor winding resistance, motor winding inductance, motor winding mutual inductance, motor moment of inertia, motor damping coefficient, permanent magnet flux linkage, lead screw moment of inertia, lead screw damping coefficient, and elastic coefficient, initialize the inherent parameters of the model.

[0018] Step S22: Set the input commands of the experimental platform and the simulation model to be consistent. Based on the two standards of signal trend consistency and amplitude consistency, adjust the three closed-loop PI parameters from the inside out, first proportional and then integral, so that the simulation data and experimental data have the same trend and the amplitude error is small.

[0019] Step S23: Repeat step S22 until the trend is consistent and the amplitude error is small. At this point, a high-accuracy simulation model of the electromechanical servo system is obtained. Through fault mechanism analysis, faults are injected into the simulation model to obtain a simulation fault dataset.

[0020] Furthermore, in step S3, the method for inputting the simulated fault dataset into the SDCTF fault diagnosis method to achieve fault classification and obtain the pre-trained model is as follows:

[0021] Step S31: For a single sensor signal, the one-dimensional time series is converted into two-dimensional data through dimension transformation.

[0022] Step S32: The two-dimensional data from multiple sensors are combined into three-dimensional data, and the local spatial features of the multi-sensor signals are extracted using 2D-CNN.

[0023] Step S33: For two-dimensional data of single sensor signals, the data of each row is converted into a semantic vector through linear processing, and then the position information is mapped into the semantic vector through position encoding to retain the position information of the data.

[0024] Step S34: Input the semantic vector containing location information into the Transformer Encoder layer to extract global temporal features;

[0025] Step S35: The global temporal features of the multi-sensor signals extracted by the Transformer Encoder are adaptively fused using SE;

[0026] Step S36: The global temporal features and local spatial features are concatenated and input into a fully connected layer to achieve feature dimensionality reduction. Finally, the concatenation is input into a classifier to achieve fault classification, thus obtaining a pre-trained model.

[0027] Furthermore, in step S4, the squared distance between the mean empirical kernel values ​​of the source domain and the target domain is calculated. The formula for calculating the estimated value of LMMD is as follows:

[0028]

[0029] In the formula, C represents the total number of label categories. and The source domain and the target domain are respectively, n s and n t These represent the number of samples in the source domain and the target domain, respectively. and , respectively, are the weight coefficients of the sample with label k in the source domain and the target domain, H is the RKHS using kernel K, and Φ(*) is the mapping of RKHS;

[0030] The weighting coefficients are calculated as follows:

[0031]

[0032]

[0033] In the formula, D s For source domain data, D t For target domain data, Let be the probability that the i-th sample in the source domain belongs to label c. Let be the probability that the j-th sample in the source domain belongs to label c. Let be the predicted probability that the i-th sample in the target domain belongs to label c. Let be the predicted probability that the j-th sample in the target domain belongs to label c.

[0034] The present invention discloses a fault diagnosis method for electromechanical servo systems based on simulation-driven and subdomain adaptation, which has the following beneficial effects:

[0035] This invention establishes a high-accuracy electromechanical servo system simulation model in MATLAB through a modular modeling strategy, and realizes fault injection through fault mechanism analysis, generating a large amount of simulation fault data, expanding the training dataset, and solving the problem of insufficient fault samples in electromechanical servo systems.

[0036] Compared with traditional CNN-based transfer learning fault diagnosis methods, the SDCTF fault diagnosis method proposed in this invention can simultaneously extract the local spatial features and global temporal features of multi-sensor signals from electromechanical servo systems, and adaptively fuse the features of multi-sensor signals. This solves the problem that CNN cannot capture signal temporal information and that most current fault diagnosis algorithms are only used to process single vibration signals. Attached Figure Description

[0037] Figure 1 A flowchart of a fault diagnosis method for electromechanical servo systems based on simulation-driven and subdomain adaptation;

[0038] Figure 2 This is a structural diagram of the overall simulation model of the electromechanical servo system in this invention;

[0039] Figure 3 This is a structural diagram of the simulation model of the permanent magnet synchronous motor in the electromechanical servo system of this invention;

[0040] Figure 4 This is a structural diagram of the simulation model of the planetary roller screw in the electromechanical servo system of this invention;

[0041] Figure 5 This is a schematic diagram of the relative motion between the lead screw and the nut;

[0042] Figure 6 This is a network structure diagram of the SDCTF method. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] To overcome the problems existing in the prior art, this invention proposes a fault diagnosis method for electromechanical servo systems based on simulation-driven and subdomain adaptation.

[0045] To address the issues of limited fault samples and insufficient simulation modeling research in electromechanical servo systems, this invention establishes a high-accuracy simulation model in Simulink and obtains a large amount of fault data through simulation to expand the training dataset.

[0046] To address the issue that CNN-based transfer learning-based fault diagnosis algorithms cannot capture the temporal information of time series signals and are mostly used to process single vibration signals, this invention integrates a two-dimensional convolutional network (2D-CNN), a Transformer, and a multi-channel attention mechanism (SE). The Transformer is used to solve the problem of CNNs failing to capture signal temporal information; and the SE adaptively fuses features from multiple sensor signals to address the issue that most current fault diagnosis algorithms only process single vibration signals.

[0047] refer to Figure 1 The present invention discloses a fault diagnosis method for electromechanical servo systems based on simulation-driven and subdomain adaptation, comprising the following steps:

[0048] Step S1: Construct a simulation model of the electromechanical servo system;

[0049] Step S2: Set the parameters of the electromechanical servo system simulation model to ensure that the simulation data of the electromechanical servo system is consistent with the experimental data of the experimental platform and has a small amplitude error, thereby obtaining a high-accuracy electromechanical servo system simulation model; on this basis, perform fault mechanism analysis, realize fault injection, and obtain a large number of simulation fault datasets.

[0050] Step S3: Input the simulated fault dataset into the SDCTF method to achieve fault classification, such as... Figure 6 As shown, the pre-trained model is obtained;

[0051] Step S4: Based on the obtained pre-trained model, input experimental data, take the simulation data as the source domain and the experimental data as the target domain, minimize the distribution difference between different subdomains of the source domain and the target domain based on the local maximum mean difference distance metric method, fine-tune the model parameters, and realize fault diagnosis of experimental data.

[0052] Furthermore, in step S1, the electromechanical servo system is divided into five modules—permanent magnet synchronous motor, planetary roller screw, controller, driver, and spring load—using a modular modeling strategy. Mathematical models are then created for each of the five modules, and the designs are completed in Simulink. Finally, all modules are connected according to their logical relationships to obtain the electromechanical servo system simulation model, such as... Figure 2 As shown.

[0053] Furthermore, such as Figure 3 As shown, the modeling of a permanent magnet synchronous motor includes three parts: three-phase stator voltage equation, electromagnetic torque equation, and motor mechanical motion equation. The input is three-phase voltage, and the output is motor speed, electromagnetic torque, and three-phase current.

[0054] The modeling method for the three-phase stator voltage equations is as follows:

[0055] The three-phase voltage equations and flux linkage equations of a permanent magnet synchronous motor are as follows:

[0056]

[0057] Ψ 3s =L 3s i 3s +ψ f F 3s (θ e (2)

[0058] In the formula, Ψ 3s For the total flux linkage of the three-phase winding, u 3s R 3s i 3s These represent the phase voltage, resistance, and current of the three-phase winding, respectively, L 3s For the inductance of the three-phase winding, ψ f For permanent magnet flux linkage, F 3s (θ e θ represents the flux linkage generated by the three-phase winding. e Let be the rotor position angle, and satisfy:

[0059]

[0060]

[0061] In the formula, i a i b and i c Let A be the current, B be the current, and C be the current, R be the winding resistance, and ψ be the current. a ψ b and ψ c For phase A flux linkage, phase B flux linkage, and phase C flux linkage, u a u b and u c Let L be the voltage of phase A, phase B, and phase C. aa L bb and L cc For the stator winding self-inductance, M ab M ac M ba M bc M ca and M cb Mutual inductance between the three-phase stator windings;

[0062] The permanent magnet synchronous motor used in the experiment is a surface-mounted type, therefore the self-inductance and mutual inductance of the windings are independent of the rotor position, and thus, we have...

[0063]

[0064] In the formula, L mFor the stator winding self-inductance, M m The mutual inductance of the three-phase stator windings.

[0065] The three-phase windings of the permanent magnet synchronous motor used in the experimental setup are connected in a star configuration, therefore...

[0066]

[0067] Substituting into the flux linkage equation, we get:

[0068]

[0069] Substituting into the voltage equation, we get:

[0070]

[0071] In the formula, E a E b and E c Let be the components of the back electromotive force generated by the permanent magnet in the stator winding, respectively, and satisfy the following:

[0072]

[0073] In the formula, w e Let be the electric angular velocity of the motor.

[0074] The method for modeling the electromagnetic torque equation is as follows:

[0075] In permanent magnet synchronous motors, the angles between the stator and rotor flux linkages are not 90°, resulting in strong coupling and making independent control of the magnetic field and electromagnetic torque impossible. Therefore, the natural coordinate system abc is transformed into the synchronous coordinate system dq through Clark and Park coordinate transformations. The Clark and Park transformation formulas are as follows:

[0076]

[0077]

[0078] In the formula, f a f a and f a Let f be the A-axis, B-axis, and C-axis variables in the natural coordinate system. α f β f0 represents the α-axis, β-axis, and 0-axis variables in the stationary coordinate system. d and f q These are the d-axis and q-axis variables in the synchronized coordinate system.

[0079] In the synchronous rotating coordinate system, the three-phase voltage equations are equivalent to a series connection of a resistor, inductor, and voltage source in both the d-axis and q-axis directions, achieving complete decoupling. Therefore, the system's electric power in the synchronous rotating coordinate system dq is:

[0080]

[0081] In the formula, i d i q For the d-axis current and q-axis current in the synchronous coordinate system, u d u q These are the d-axis voltage and q-axis voltage in the synchronous coordinate system.

[0082] Because the Park transformation process uses a constant amplitude transformation, the electrical power becomes 2 / 3 of the actual system power. Furthermore, the first three terms in the above equation represent energy conversion within the electrical system and do not translate to the actual mechanical system. Therefore, the actual torque power output by the motor is:

[0083]

[0084] Among them, P m This refers to the torque power of the motor system.

[0085] The final electromagnetic torque equation is written as follows:

[0086]

[0087] In the formula, ω m T is the rotor speed of the motor. e ψ is the electromagnetic torque of the motor, p is the number of pole pairs of the motor, and ψ is the electromagnetic torque of the motor. f For permanent magnet flux linkage, i q This is the q-axis current;

[0088] The mechanical motion equations of the electric motor are as follows:

[0089]

[0090] In the formula, J is the moment of inertia of the motor, and T L Let B be the load torque, and B be the motor damping coefficient; and satisfy:

[0091]

[0092] Furthermore, such as Figure 4 As shown, the method for modeling planetary roller screws is as follows:

[0093] The kinematic equations between the lead screw and the nut are established, with the motor speed signal as input and the axial displacement of the nut as output. The specific process is as follows:

[0094] In a roller screw, the axial output displacement of the screw is generated solely by the rotation of the screw, and is linearly related to the rotation angle of the screw, but independent of the roller motion.

[0095]

[0096] In the formula, l s v is the lead screw pitch. s R is the axial displacement velocity of the lead screw. s Let β be the radius of the lead screw engagement point. s The lead screw helix angle;

[0097] Therefore, the roller screw is simplified during the modeling process, and only the relative motion between the screw and the nut is considered.

[0098] In a roller screw, the helix angle is β. s The helix angle of the roller and nut is β. n Because of β s and β n Approaching 0, i.e., cosβ s ≈cosβ n Therefore, in the direction perpendicular to the raceway, the normal force of the raceway between the screw and the roller, and between the roller and the nut, is approximately F. N cosβ s F N The load force is spring load; the clearance between the lead screw and roller, and between the roller and nut is simplified to the clearance between the lead screw and nut; the friction inside the roller lead screw is simplified to the friction generated by the relative motion between the lead screw and nut. That is, the lead screw bears the load through the direction perpendicular to the raceway, and the friction is calculated through the direction parallel to the raceway. At the same time, the rotational inertia generated by the rotation of the roller is not considered.

[0099] The motor input drives the lead screw to rotate, and the circumferential linear velocity V of the lead screw is... ts for:

[0100] V ts =ω m ×R s (17)

[0101] The model of the lead screw and nut and the schematic diagram of their relative motion are as follows: Figure 5 As shown. For every revolution of the leadscrew, the nut moves axially a distance equal to one pitch, that is, the nut moves along the direction of the leadscrew's helix angle; the speed V of the leadscrew relative to the nut, parallel to the leadscrew raceway. ps and the speed V perpendicular to the raceway vs for

[0102] V ps =V ts ×cosβ s (18)

[0103] V vs =V ts ×sinβ s (19)

[0104] The speed V of the lead screw perpendicular to the raceway vs After passing through the gap between the lead screw and the nut, the speed of the driving nut along the raceway perpendicular to the raceway is V. vr Thus, the axial velocity V of the nut can be obtained. ar and the speed V along the raceway pr for:

[0105] V ar =V vr ×(1 / cosβ s (20)

[0106] V pr =V ar ×sinβ s (twenty one)

[0107] The nut moves along the axial direction at a speed of V ar The axial displacement of the nut is obtained by integrating over time;

[0108] The speed V of the lead screw parallel to the lead screw raceway ps The speed V of the nut along the raceway pr The speeds are opposite, and the absolute values ​​of the two are added together to obtain the relative velocity, i.e.

[0109] V prs =|V ps |+|V pr | (22)

[0110] The relative motion between the lead screw and the nut generates friction, which is simulated as a function of relative velocity and assumed to be the sum of Stribeck, Coulomb, and viscous components. Stribeck friction is a negative slope characteristic occurring at low speeds. Coulomb friction produces a constant force at any speed. Viscous friction is a force that opposes motion with a force proportional to the relative velocity. Near zero velocity, the sum of Coulomb and Stribeck friction is usually called sliding friction, which can be considered static friction. Friction is approximated by the following equation:

[0111]

[0112] F brk =μ brk ·N·sign(v)

[0113] F C =μ C ·N·sign(v)

[0114]

[0115] v Coul =v brk / 10 (23)

[0116] In the formula, F is the frictional force. C For Coulomb friction, F brk For sliding friction, v brk Let N be the sliding friction velocity, N be the normal force, and μ be the sliding friction velocity. brk μ is the coefficient of sliding friction. C v is the Coulomb friction coefficient. St For the Stribeck velocity threshold, v Coul is the Coulomb velocity threshold, v is the relative velocity, and f is the viscous friction coefficient.

[0117] Furthermore, the method for controller modeling is as follows:

[0118] The control algorithm employs a three-loop PI control algorithm encompassing displacement, speed, and current, with the current loop using an i... d The control strategy with a value of 0, since the damping winding of the d-axis generates magnetic flux, together with the magnetic flux of the permanent magnet, constitutes the magnetic field of the motor; i d As the overall magnetic flux changes, a coupling relationship will occur; when i d When the magnetic flux is 0, the magnetic flux is entirely provided by the permanent magnet, thus achieving decoupling of the motor.

[0119] Furthermore, the method for driver modeling is as follows:

[0120] The driver consists of an SVPWM algorithm and a three-phase inverter. The control current signal is processed by the SVPWM algorithm to obtain a PWM wave. The duty cycle of the PWM is used to control the switching of the three-phase IGBT transistors, thereby outputting a control voltage signal to drive the motor to rotate.

[0121] The implementation of the SVPWM algorithm mainly includes sector determination of the reference voltage vector, calculation of the duration of non-zero and zero vectors in each sector, and determination of the switching point of each sector vector. Finally, a triangular carrier signal of a certain frequency is compared with the switching point of each sector vector to generate the PWM pulse signal required by the converter. In Simulink, the sector determination of the reference voltage vector, the calculation of the duration of non-zero and zero vectors in each sector, and the calculation of the switching point of each sector vector are implemented by function programming.

[0122] Furthermore, the method for modeling spring loads is as follows:

[0123] According to Hooke's law of elasticity, when a spring undergoes elastic deformation, the spring force is directly proportional to the spring's elongation, that is:

[0124] F = -kx (24)

[0125] In the formula, x is the deformation and k is the elastic coefficient.

[0126] Furthermore, the specific method of step S2 includes:

[0127] Step S21: Based on the inherent parameters of the motor, lead screw, and load spring in the experimental platform, including motor winding resistance, motor winding inductance, motor winding mutual inductance, motor moment of inertia, motor damping coefficient, permanent magnet flux linkage, lead screw moment of inertia, lead screw damping coefficient, and elastic coefficient, initialize the inherent parameters of the model.

[0128] Step S22: Set the input commands of the experimental platform and the simulation model to be consistent. Based on the two standards of signal trend consistency and amplitude consistency, adjust the three closed-loop PI parameters from the inside out, first proportional and then integral, so that the simulation data and experimental data have the same trend and the amplitude error is small.

[0129] Step S23: Repeat step S22 until the trend is consistent and the amplitude error is small. At this point, a high-accuracy simulation model of the electromechanical servo system is obtained. Through fault mechanism analysis, faults are injected into the simulation model to obtain a simulation fault dataset.

[0130] Furthermore, in step S3, the method for inputting the simulated fault dataset into the SDCTF fault diagnosis method to achieve fault classification and obtain the pre-trained model is as follows:

[0131] Step S31, the input sensor signal is a one-dimensional time series X∈R L×1×1 L is the signal length. It is transformed into two-dimensional data X∈R through dimensionality transformation. W×H×1 W and H represent the number of rows and columns of the two-dimensional data, respectively.

[0132] Step S32: Combine the two-dimensional data from multiple sensors into three-dimensional data X∈R. W×H×C C represents the number of sensors. Multi-sensor signals X∈R are extracted using 2D-CNN. W×H×C The local spatial features. The 2D-CNN network consists of three convolutional layers, three ReLU activation layers, two Dropout layers, and one Flatten layer. Multi-sensor signal X∈R W×H×C The spatial features X are output by 2D-CNN. s ∈ S×1 , where S is the length of the local spatial feature.

[0133] Step S33, for a single sensor signal X∈RW×H×1 For each row of data X∈R 1×H×1 After linear processing, it is transformed into a semantic vector Z∈R. 1×N×1 Where N is the length of the semantic vector, the final semantic matrix Z∈R is obtained. W×N×1 Then, through positional encoding, the positional information is mapped to the semantic vector. Positional encoding tells the transformer model the position of the data in the sequence, preserving the data's positional information. The positional encoding formula is as follows:

[0134]

[0135]

[0136] In the formula, PE represents the position code, pos represents the position of the current data in the sequence, and d model The dimension of the current data.

[0137] Step S34, the semantic matrix Z∈R containing location information W×N×1 The input is fed into the Transformer Encoder layer to extract global temporal features X. T ∈R W×M×1 , where M is the length of the latent features. The multi-head attention mechanism is the core of the TransformerEncoder, and its calculation process is as follows:

[0138]

[0139]

[0140] X MHSA =LayerNorm(A n (X T )+X T )

[0141] In the formula, H i For the output of the i-th head attention, Q i K i and V i For the query matrix, key matrix, and value matrix in the i-th head attention, W 0 For a linear projection with multiple connections, H n X is the output of the nth head attention. T For global time series features, A n (X T This represents the output of multi-head attention. For the normalization layer, X MHSA This is the output of the multi-head self-attention mechanism after residual connection and normalization.

[0142] The output of the multi-head attention mechanism is passed through two fully connected feedforward layers to obtain:

[0143] FF(X MHSA )=δ(0,W1X MHSA +b1)W2+b2

[0144] X T =LayerNorm(FF(X) MHSA )+X MHSA )

[0145] In the formula, W1 and W2 are the weights of the feedforward network, b1 and b2 are the biases of the feedforward network, δ is the ReLU activation function, and FF(X) = 1 / 2. MHSA ) represents the output of the feedforward network.

[0146] Step S35: Adaptively fuse the global temporal features X of the multi-sensor signals extracted by the TransformerEncoder using SE. MT ∈R W×M×S SE first performs compression on the input multi-channel global features. It compresses single-channel features through global average pooling, and then compresses the information of all data points in the feature space into a single real number to represent the overall information of the entire channel through averaging. The calculation process of the compression operation is as follows:

[0147]

[0148] In the formula, Z c This is the output result after compression; X MT,c This represents the global feature of the c-th sensor channel.

[0149] Then, two fully connected layers are used to automatically analyze the correlation between different channels and obtain the weight value of each channel. The specific calculation process is as follows:

[0150] O c =σ(K1δ(K0(Z) c )))

[0151]

[0152] In the formula, σ and δ are the Sigmoid activation function and the ReLU activation function, respectively; K0 and K1 are the fully connected weight coefficients, and O c Y represents the channel weighting coefficient. MT,c This represents the global feature of the c-th sensor channel after weighting.

[0153] The weighted multi-channel global time series feature Y MT ∈R W×M×S The final global temporal feature Y is obtained through the flattening layer. MT∈R T×1 , where T is the length of the global temporal feature.

[0154] Step S36, the global temporal feature Y MT ∈R T×1 With local spatial features X s ∈ S×1 The model is then spliced ​​together. Since the dimensionality may be high, feature dimensionality reduction is achieved by inputting into a fully connected layer. Finally, fault classification is performed in a classifier to obtain a pre-trained model.

[0155] Furthermore, in step S4, the squared distance between the mean empirical kernel values ​​of the source domain and the target domain is calculated. The formula for calculating the estimated value of LMMD is as follows:

[0156]

[0157] In the formula, C represents the total number of label categories. and The source domain and the target domain are respectively, n s and n t These represent the number of samples in the source domain and the target domain, respectively. and , respectively, are the weight coefficients of the sample with label k in the source domain and the target domain, H is the RKHS using kernel K, and Φ(*) is the mapping of RKHS;

[0158] The weighting coefficients are calculated as follows:

[0159]

[0160]

[0161] In the formula, D s For source domain data, D t For target domain data, Let be the probability that the i-th sample in the source domain belongs to label c. Let be the probability that the j-th sample in the source domain belongs to label c. Let be the predicted probability that the i-th sample in the target domain belongs to label c. Let be the predicted probability that the j-th sample in the target domain belongs to label c.

[0162] The present invention provides a simulation-driven and subdomain-adaptive fault diagnosis method for electromechanical servo systems. First, the mechanism of the electromechanical servo system is analyzed, and a simulation model is established in Simulink. Next, a high-accuracy simulation model is obtained by initializing the model's inherent parameters and adjusting control parameters. Fault mechanism analysis is then performed to inject fault data, generating a large amount of simulated fault data. The simulation dataset is then input into the fault diagnosis algorithm for model pre-training. Finally, experimental data is input, and model transfer is achieved based on domain adaptation. This invention, through a modular modeling strategy, establishes a high-accuracy simulation model of the electromechanical servo system in MATLAB and achieves fault injection through fault mechanism analysis, generating a large amount of simulated fault data, expanding the training dataset, and solving the problem of insufficient fault samples in electromechanical servo systems. Compared with traditional CNN-based transfer learning fault diagnosis methods, the SDCTF fault diagnosis method proposed in this invention can simultaneously extract local spatial features and global temporal features of multi-sensor signals from the electromechanical servo system and adaptively fuse multi-sensor signal features, solving the problem that CNN cannot capture signal temporal information and that most current fault diagnosis algorithms only process single vibration signals.

[0163] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for fault diagnosis of an electromechanical servo system based on simulation driving and sub-domain adaptation, characterized in that, The method comprises the following steps: Step S1, constructing a mechatronic servo system simulation model; Step S2, setting the parameters of the mechatronic servo system simulation model to ensure that the simulation data of the mechatronic servo system and the experimental data of the experimental bench are consistent in trend and have small amplitude errors, thereby obtaining a high-accuracy mechatronic servo system simulation model; on this basis, fault mechanism analysis is performed, fault injection is realized, and a large number of simulation fault data sets are obtained; Step S3, inputting the simulation fault data set into the SDCTF method to realize fault classification and obtain a pre-trained model; Step S4, inputting experimental data on the basis of the obtained pre-trained model, taking simulation data as a source domain and experimental data as a target domain, minimizing the distribution difference between different sub-domains of the source domain and the target domain based on a local maximum mean difference distance measurement method, fine-tuning model parameters, and realizing fault diagnosis on experimental data; The specific method of step S2 comprises: Step S21, initializing the inherent parameters of the model according to the inherent parameters of the motor, lead screw and load spring in the experimental bench, including motor winding resistance, motor winding inductance, motor winding mutual inductance, motor rotational inertia, motor damping coefficient, permanent magnet flux linkage, lead screw rotational inertia, lead screw damping coefficient and spring coefficient; Step S22, setting the input instructions of the experimental bench and the simulation model to be consistent, based on two standards of signal trend consistency and amplitude consistency, adjusting the three-closed-loop PI parameters in the order of inside-out and proportion-integration first, so that the simulation data and the experimental data have the same trend and small amplitude errors; Step S23, repeating step S22 until the trend is consistent and the amplitude error is small, at which time a high-accuracy simulation model of the mechatronic servo system is obtained, and faults are injected into the simulation model through fault mechanism analysis to obtain a simulation fault data set.

2. The method according to claim 1, wherein, In step S1, the mechatronic servo system is divided into five modules of permanent magnet synchronous motor, planetary roller screw, controller, driver and spring load through a modular modeling strategy, mathematical modeling is performed on the five modules, design is completed in Simulink, and finally all modules are connected according to the logical relationship to obtain a mechatronic servo system simulation model.

3. The method according to claim 2, wherein, Permanent magnet synchronous motor modeling includes three parts of three-phase stator voltage equation, electromagnetic torque equation and motor mechanical motion equation, the input is three-phase voltage, and the output is motor speed, electromagnetic torque and three-phase current.

4. The method according to claim 3, wherein, In step S3, the method for inputting the simulation fault data set into the SDCTF fault diagnosis method to realize fault classification and obtain a pre-trained model is as follows: Step S31, for single sensor signals, one-dimensional time series is converted into two-dimensional data through dimension conversion; Step S32, two-dimensional data of multiple sensors are combined into three-dimensional data, and local spatial features of multi-sensor signals are extracted through 2D-CNN; Step S33, for two-dimensional data of single sensor signals, the data of each row is converted into a semantic vector through linear processing, and then the position information is mapped into the semantic vector through position coding to retain the position information of the data; Step S34, the semantic vector containing position information is input into a Transformer Encoder layer to extract global time sequence features; Step S35, the global timing features of the multisensor signals extracted by the SE adaptive fusion Transformer Encoder are extracted; Step S36, the global timing features are spliced with the local spatial features, and input into a full connection layer to realize feature dimension reduction, and finally input into a classifier to realize fault classification, to obtain a pre-training model.

5. The method according to claim 4, wherein, In step S4, the squared distance between the source domain and the target domain empirical kernel means is calculated As the estimate of the LMMD, the following formula is used: where C is the total number of label classes, and are the source and target domains, respectively, and are the number of samples in the source and target domains, respectively, and are the weight coefficients of samples with label k in the source and target domains, respectively, H is the RKHS using kernel K, is the mapping of RKHS; The weight coefficient is calculated as follows: wherein, is the source domain data, is the target domain data, is the probability that the i-th sample in the source domain belongs to label c, is the probability that the j-th sample in the source domain belongs to label c, is the predicted probability that the i-th sample in the target domain belongs to label c, is the predicted probability that the j-th sample in the target domain belongs to label c.

Images