Elevator fault diagnosis method based on digital twin

Abstract

A method for diagnosing elevator faults based on digital twins belongs to the field of elevator fault diagnosis. The method comprises the following steps: the first step is to collect elevator vibration signals; the second step is to construct a numerical model of the elevator in the vertical direction, first establish the dynamic equation of the system, and then calculate the kinetic energy of the system, the potential energy of the system and the dissipation coefficient of the system; the third step is to improve the PINNs model, introduce it as a physical information regulator in the loss function, and introduce a reactivation algorithm to optimize the loss function; step (3.2) designs a reactivation algorithm to obtain the optimal solution of the model; the fourth step is to input a small amount of elevator fault vibration signals collected in the first step into the PINNs model with motion equation constraints to obtain simulated elevator vibration data; the fifth step is to improve the GCN model for elevator fault diagnosis. The present invention can effectively obtain data that is not easy to obtain in reality, and effectively improve the accuracy of fault diagnosis.

Description

Technical Field

[0001] The present invention belongs to the technical field of elevator fault diagnosis, and in particular relates to an elevator fault diagnosis method based on digital twins. Background Art

[0002] Elevators are a convenient and indispensable means of transportation in high-rise buildings. Studies have shown that 54% of the world's population currently lives in urban areas. With the rapid development of urbanization, the number of elevators is also growing. Faced with the severe situation of elevator safety management, various places have built comprehensive platforms for elevator emergency response, deployed telephone hotlines or call systems, and configured professional response teams, providing important support for the rapid response and proper handling of elevator safety accidents. However, with the further increase in the number of elevators and the arrival of aging equipment, the phenomenon of abnormal vibration of elevators has become more and more common. The contradiction between the number of responders and the number of elevators, and the inability of accumulated professional experience to meet the needs of the surge in elevator abnormalities have become prominent. Therefore, it is urgent to carry out research and application of new intelligent processing technologies for elevators.

[0003] However, most of the current technical work has certain obstacles, mainly because real elevators are rarely connected to virtual models, which leads to the inability to effectively utilize the large amount of operating data generated by real elevators. At present, with the development of cyber-physical systems (CPS), it is crucial to pay attention to virtual space and achieve seamless integration of physical space and virtual space to improve elevator fault diagnosis. In this case, it is very important to introduce digital models of equipment and their data in virtual space to accurately depict the behavior of real entities.

[0004] With the rapid development of technologies such as the Internet, artificial intelligence, and digital twins, these technologies have gradually become valuable opportunities in elevator fault diagnosis. Digital twin technology can realize the digital expression of real scenes, and through the perception analysis and model construction of scene elements, it can realize the two-way real-time mapping between physical entities and virtual models. With the support of simulation verification methods, it helps to deduce the changing trends of real scenes, and also shows great application advantages in remote monitoring, status monitoring and control optimization of special equipment. In order to solve the problems of inaccurate characteristics and slow convergence speed of traditional elevator braking performance prediction models, some people have proposed elevator braking performance prediction and evaluation based on digital twins. By deploying sensors to monitor the multi-dimensional data of the elevator itself, and collecting the elevator in real time based on the intelligent gateway, a high-fidelity model of the elevator is built to realize the two-way mapping and dynamic interaction between the elevator physical entity and the twin. Some people have proposed a digital twin reference model for rotating machinery fault diagnosis, and discussed the construction requirements of the digital twin model. A model correction scheme based on parameter sensitivity analysis is proposed to enhance the adaptability of the digital twin model.

[0005] Deep learning is a novel intelligent method that can autonomously extract effective features from raw signals without relying on manual feature extraction. At present, deep learning technology has been widely used in speech classification, language translation, image classification and other fields, and has achieved great success. Deep learning has also made great progress in the field of elevator fault diagnosis. This technology can diagnose the operating health of mechanical equipment based on raw monitoring data. In response to the sample imbalance and hyperparameter tuning problems of the XGBoost model, a method for elevator fault diagnosis based on recursive feature elimination (RFE)-SMOTE-Tomek and IAO-XGBoost was proposed, which can better predict potential faults and take preventive maintenance measures, thereby extending the service life of the elevator and improving its reliability; a deep residual contraction network model based on the design idea of ​​the residual block in the deep learning model Res Net was proposed and developed to improve the learning ability of the high noise vibration signal characteristics of the elevator, and the data was used for fault diagnosis; based on the time series data of the elevator operation, deep learning was applied to the prediction of elevator faults, and neural networks were used to extract features from the time series vibration data of the elevator for corresponding research.

[0006] However, training deep learning diagnostic models requires a large amount of labeled fault data that contains all possible fault conditions. In actual operation, the collection and recording of elevator fault data is often limited by time, cost, and feasibility. Therefore, the acquisition of fault data is limited by many factors. Summary of the invention

[0007] In order to overcome the shortcomings of the existing technology, the present invention provides an elevator fault diagnosis method based on digital twins to improve the timeliness of real-time state detection and fault diagnosis results of digital twin models. By establishing an elevator digital twin model that is highly consistent with the physical entity, the virtual twin model is run in combination with the physical information neural network (Physics-Informed Neural Networks, PINNs) to obtain a large amount of simulation data that is highly consistent with reality. The e-RGCN (Epsilon Recurrent Network-Graph Convolutional Network, recursive-graph neural network) fault diagnosis algorithm is proposed, and the ε constraint is introduced to enhance the robustness and fault tolerance of the algorithm. The recursive network is organically combined with the graph convolutional network, and the vibration data is graphically processed through the ε constraint. Finally, an effective method is proposed to solve the problem of elevator fault diagnosis, which has good practical value.

[0008] The technical solution adopted by the present invention to solve its technical problem is:

[0009] An elevator fault diagnosis method based on digital twins includes the following steps:

[0010] The first step is to collect elevator vibration signals;

[0011] The second step is to construct a numerical model of the elevator in the vertical direction. First, the dynamic equation of the system is established, and then the kinetic energy, potential energy and dissipation coefficient of the system are calculated.

[0012] The third step is to improve the PINNs model and introduce it into the loss function As a physical information regulator, the reactivation algorithm is introduced to optimize the loss function. The process is as follows:

[0013] Step (3.1) designs the loss function of the PINNs model into two parts, one of which is the loss of the neural network. , which is used to represent the error between the predicted value and the true value based on the network, and the other part is the loss of physical information , the loss of physical information includes the initial value loss , Boundary loss and the residual of the ordinary differential equation ;

[0014] Step (3.2) Design a reactivation algorithm to obtain the optimal solution of the model;

[0015] Step 4: Acquire simulated vibration data. Input a small amount of elevator fault vibration signals collected in the first step into the PINNs model with motion equation constraints to obtain simulated elevator vibration data.

[0016] Step 5: Improve Graph Convolutional Networks (GCN)

[0017] The model is used for elevator fault diagnosis. The process is as follows: the e-RGCN fault diagnosis algorithm is proposed by combining the recursive neural network (RNN) and the graph convolutional network GCN, a recursive radius is set, and a correlation graph is constructed through the recursive radius. Then, GCN can effectively identify the elevator fault status through the constructed correlation graph.

[0018] Furthermore, in the first step, the vibration signal of the elevator is collected at a frequency of 50 Hz by an acceleration sensor installed in the car top box on the top of the elevator car, and remote data acquisition is achieved through the TCP / IP communication protocol; the sensor, elevator and gateway are connected by serial port connection, and Modbus and RS485 protocols are used for data communication.

[0019] Going further, the process of the second step is as follows:

[0020] Step (2.1) establishes the dynamic equation of the system, the formula is as follows,

[0021] ;

[0022] Where T is the kinetic energy of the system; U is the potential energy of the system; D is the dissipation coefficient of the system; x is the generalized displacement (or angular displacement); is the generalized velocity (or angular velocity); Q is the generalized force (exciting force or exciting torque);

[0023] Step (2.2) Calculate the kinetic energy of the system , the potential energy of the system and system dissipation factor ,

[0024] ;

[0025] ;

[0026] ;

[0027] in, is the equivalent mass of the car and the loaded car; is the counterweight mass; , are the equivalent stiffness and damping of the wire rope between the car and the traction sheave respectively; , are the equivalent stiffness and damping of the wire rope between the counterweight and the traction sheave respectively; , and They are the car vibration linear displacement, counterweight vibration linear displacement, and traction sheave vibration angular displacement; , are the moment of inertia of the traction sheave and the radius of the rope groove of the traction sheave respectively;

[0028] Step (2.3) Arrange the formulas in step (2.2) and step (2.1) as follows:

[0029] Step (2.3.1) for generalized coordinates ,have,

[0030] ;

[0031] Step (2.3.2) Since the elevator car is only affected by the vertical wire rope tension and gravity, substitute the result of step (2.3.1) into the Lagrange equation of step (2.1) to obtain,

[0032] ;

[0033] Step (2.3.3) for generalized coordinates ,have,

[0034] ;

[0035] Step (2.3.4) The elevator counterweight is only affected by the vertical wire rope tension and gravity. Substituting the result of step (2.3.3) into the Lagrange equation of step (2.1), we get:

[0036] ;

[0037] Step (2.3.5) for generalized coordinates ,have,

[0038] ;

[0039] Step (2.3.6) for generalized coordinates , the system inertia force in the elevator start-brake stage is used as the excitation source, and the exciting force is expressed as,

[0040] ;

[0041] Step (2.4), Step (2.3.2), Step (2.3.2) and Step (2.3.2) constitute the differential equation of motion, which can be expressed in matrix form as follows,

[0042] ;

[0043] in,

[0044] ;

[0045] ;

[0046] ;

[0047] ;

[0048] ;

[0049] Among them, M is the system mass matrix, C is the system damping matrix, K is the system stiffness matrix, and F is the exciting force matrix; , , They represent the acceleration, velocity and displacement of the system respectively.

[0050] Furthermore, the process of step (3.1) is as follows:

[0051] Step (3.1.1) The loss of the neural network is as follows,

[0052] ;

[0053] in, is the number of sampled data, The operator representing the mean square error (MSE), represents the predicted value of the neural network, Indicates the actual measured value;

[0054] Step (3.1.2) A core idea of ​​PINNs is to make the neural network satisfy the given physical laws. The physical laws are expressed in the form of ODE. The ODE residual loss is used to measure whether the solution of the network satisfies these equations. The residual of the ordinary differential equation is as follows,

[0055] ;

[0056] in, is the number of sampled data, The operator representing the mean square error (MSE), represents the actual measured value, for any time;

[0057] Step (3.1.3) The boundary conditions define the behavior of the solution at the boundary of the problem region, and the initial conditions define the state of the system at time t = 0. The boundary conditions and initial conditions of PINNs are obtained through actual measurement and estimation as follows:

[0058] ;

[0059] Step (3.1.4) The boundary loss function compares the boundary conditions predicted by the model with the actual physical boundary conditions, so that the solution of the model in the boundary area conforms to the physical laws. The boundary loss is as follows:

[0060] ;

[0061] in, is the number of sampled data, The operator representing the mean square error MSE, represents the predicted value of the neural network, Indicates the actual measured value;

[0062] Step (3.1.5) The initial value loss function measures the difference between the model prediction value and the actual initial conditions to ensure that the network output meets the physical constraints at the initial moment. The initial value loss is as follows:

[0063] ;

[0064] in, is the number of sampled data, The operator representing the mean square error (MSE), represents the predicted value of the neural network, Indicates the actual measured value;

[0065] Step (3.1.6) Add the losses mentioned in steps (3.1.1) to (3.1.6) to get the final loss function of the PINNs model. The formula is as follows:

[0066] ;

[0067] in, , It is a physical information regulator. By introducing λ, PINNs can accurately estimate the output of the equation. By introducing physical information into the loss function, λ plays an important role in determining the contribution of physics to the final loss function. As λ increases, the output of PINNs will gradually rely on the constraints of the physical equation, while reducing λ will make the actual measured data more influential on the output; if the value of λ is set to 0, the entire loss function does not contain the physical information term and becomes a traditional ANN, then the network only relies on the data during training.

[0068] In the step (3.2), using PINNs to solve the differential equation is an optimization problem of minimizing the loss function. For the vibration dynamics equation of the elevator, when using an optimization algorithm with line search and convergence criteria, a local optimal solution can be predicted. The local optimal solution can be regarded as the exact solution obtained. However, sometimes the local optimal solution obtained may have a large deviation. In order to avoid the situation where the solution has a large deviation, a reactivation technology is proposed to reactivate the optimization algorithm so that it can get the optimal solution again; the process is as follows:

[0069] Step (3.2.1) PINNs uses the Adma optimizer with strong Wolfe. In the kth iteration, it will find a step size αk that satisfies the following conditions,

[0070] Function value decreasing condition,

[0071] ;

[0072] Strong curvature condition,

[0073] ;

[0074] in, is the search direction, and is a constant of the strong Wolfe condition, 0< < <1, is the current point, is the gradient of the current point. If the optimization falls into a local optimal solution at the beginning of the kth iteration, there is no αk that satisfies the equation. If no αk is found to satisfy the equation, the optimization is stopped.

[0075] Step (3.2.2) In order to restart the optimization process, introduce The constant makes the amplification factor slightly greater than zero, giving the following formula,

[0076] ;

[0077] The above formula is used to replace the function reduction condition. The strong curvature condition is only used in the first iteration of reactivation. The introduction of the above formula allows Greater than , making it easier to find αk that satisfies the equation; once found, the reactivation optimization algorithm will continue to use the standard strong Wolfe condition, that is, step (3.2.1).

[0078] The process of the fifth step is as follows:

[0079] Step (5.1) constructs the association graph. e-RNN is one of the methods for complex network analysis of time series. Recursive network is an intuitive method that uses the recursiveness of mechanical systems to convert time series into graphs. Unlike traditional time series analysis, which focuses more on the sequence of samples, recursive network analysis focuses more on the spatial topological structure of samples and the relationship between samples.

[0080] Step (5.2) Construction of the e-RGCN fault diagnosis network. The proposed e-RGCN classification network consists of two GCN layers, one fully connected layer and an activation function. The GCN layer extracts graph features from the association graph, and the fully connected layer is used as a classifier for elevator fault diagnosis.

[0081] Step (5.3) inputs the graph into the network, extracts features through two GCN layers, and obtains deep features; then classifies the deep features through a softmax classification layer, and outputs the fault type classification result;

[0082] Among them, the Softmax function is a commonly used activation function, which is often used in multi-category classification problems. The Softmax function can convert a real number vector into a vector representing a probability distribution, so that each element in the vector is between 0 and 1, and the sum of all elements is 1.

[0083] The processing process of step (5.1) is as follows:

[0084] Step (5.1.1) constructs a graph that requires a relationship between the central vertex and other vectors. Two vertices can only be There is an edge between two vertices only when is the Euclidean norm between two vectors, is the recursive radius, which is the global threshold of RNN and can adjust the criteria of recursive pairs; two vertices are recursive pairs only when the conditions are met; in addition, the topological structure of all samples is obtained by performing pairwise operations on all vectors, and the structure of the network is encoded into a recursive matrix expressed as follows,

[0085] ;

[0086] in, is the Heaviside function, expressed as follows,

[0087] ;

[0088] in, is the Euclidean norm of a vector, if and There is an edge between The value of is 1, otherwise The value of is 0;

[0089] Step (5.1.2) The adjacency matrix of the graph can be constructed by recursive matrix as follows,

[0090] ;

[0091] in is the Kronecker function, expressed as,

[0092] ;

[0093] The graph of the vibration signal is obtained through the obtained adjacency matrix.

[0094] The processing process of step (5.2) is as follows:

[0095] Step (5.2.1) uses the cross-entropy loss function as the loss function of the network output. The cross-entropy loss function is a loss function in machine learning and deep learning. Especially in classification problems, it is used to measure the difference between the model's predicted value and the true label. It is particularly suitable for classification tasks (such as binary classification and multi-classification tasks). The loss function of the e-RGCN model is Indicates the degree of match between the predicted label and the actual label, as follows,

[0096] ;

[0097] Where N is the number of samples, C is the number of categories, is the probability of the jth category in the true label of the i-th sample, is the jth category in the model's predicted probability value for the i-th sample;

[0098] Step (5.2.2) During the e-RGCN training process, multi-layer GCN mixes the features of the node itself and its neighbors. Even if an edge is deleted, information about the node can still be obtained from its neighbors. Drop Edge randomly deletes some edges in the original graph, which increases the generalization of the network.

[0099] ;

[0100] Among them, A is the graph structure input of GCN, This is the graph structure after randomly selecting some edges.

[0101] The beneficial effects of the present invention are mainly manifested in:

[0102] (1) Integrate elevator fault diagnosis and digital twin technology, build an elevator numerical model and combine it with a deep learning algorithm to establish an elevator digital twin system. This method can effectively obtain data that is difficult to obtain in reality, facilitate various subsequent studies, and provide effective materials for similar studies;

[0103] (2) The traditional PINNs model is improved. The present invention introduces λ as a physical information regulator in the entire loss function to adjust the proportion of physical information in the entire loss function to obtain a more accurate prediction result;

[0104] (3) An e-RGCN fault diagnosis algorithm is proposed, and the ε constraint is introduced to enhance the robustness and fault tolerance of the algorithm. The recursive network is organically combined with the graph convolutional network. The vibration data is graphically processed through the ε constraint, which can effectively improve the fault diagnosis accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0105] Figure 1 It is a schematic diagram of the 3-DOF vertical vibration dynamics model of a 1:1 traction elevator.

[0106] Figure 2 It is a specific technical roadmap for constructing the elevator fault diagnosis model.

[0107] Figure 3 This is a comparison chart of the accuracy of seven models for seven experiments on four labels. DETAILED DESCRIPTION

[0108] The present invention will be further described below in conjunction with the accompanying drawings.

[0109] Reference Figure 1~Figure 3, an elevator fault diagnosis method based on digital twins, comprising the following steps:

[0110] The first step is to collect elevator vibration signals;

[0111] In the first step, the vibration signal of the elevator is collected at a frequency of 50 Hz by an acceleration sensor installed in the car top box on the top of the elevator car, and remote data acquisition is achieved through the TCP / IP communication protocol; the sensor, elevator and gateway are connected using a serial port connection, and Modbus and RS485 protocols are used for data communication.

[0112] The second step is to construct a numerical model of the elevator in the vertical direction. First, the dynamic equation of the system is established, and then the kinetic energy, potential energy and dissipation coefficient of the system are calculated.

[0113] The process of the second step is as follows:

[0114] Step (2.1) is based on Figure 1 The three-degree-of-freedom vertical vibration dynamics model of the elevator established in the paper assumes that all displacements are positive upward and angular displacements are positive counterclockwise. The dynamic equation of the system is established according to the Lagrange equation. The formula is as follows:

[0115] ;

[0116] Where T is the kinetic energy of the system; U is the potential energy of the system; D is the dissipation coefficient of the system; x is the generalized displacement (or angular displacement); is the generalized velocity (or angular velocity); Q is the generalized force (exciting force or exciting torque);

[0117] Step (2.2) Calculate the kinetic energy of the system , the potential energy of the system and system dissipation factor ,

[0118] ;

[0119] ;

[0120] ;

[0121] in, is the equivalent mass of the car and the loaded car; is the counterweight mass; , are the equivalent stiffness and damping of the wire rope between the car and the traction sheave respectively; , are the equivalent stiffness and damping of the wire rope between the counterweight and the traction sheave respectively; , and They are the car vibration linear displacement, counterweight vibration linear displacement, and traction sheave vibration angular displacement; , are the moment of inertia of the traction sheave and the radius of the rope groove of the traction sheave respectively;

[0122] Step (2.3) Arrange the formulas in step (2.2) and step (2.1) as follows:

[0123] Step (2.3.1) for generalized coordinates ,have,

[0124] ;

[0125] Step (2.3.2) Since the elevator car is only affected by the vertical wire rope tension and gravity, substitute the result of step (2.3.1) into the Lagrange equation of step (2.1) to obtain,

[0126] ;

[0127] Step (2.3.3) for generalized coordinates ,have,

[0128] ;

[0129] Step (2.3.4) The elevator counterweight is only affected by the vertical wire rope tension and gravity. Substituting the result of step (2.3.3) into the Lagrange equation of step (2.1), we get:

[0130] ;

[0131] Step (2.3.5) for generalized coordinates ,have,

[0132] ;

[0133] Step (2.3.6) for generalized coordinates , the system inertia force in the elevator start-brake stage is used as the excitation source, and the exciting force is expressed as,

[0134] ;

[0135] Step (2.4), Step (2.3.2), Step (2.3.2) and Step (2.3.2) constitute the differential equation of motion, which can be expressed in matrix form as follows,

[0136] ;

[0137] in,

[0138] ;

[0139] ;

[0140] ;

[0141] ;

[0142] ;

[0143] Among them, M is the system mass matrix, C is the system damping matrix, K is the system stiffness matrix, and F is the exciting force matrix; , , They represent the acceleration, velocity and displacement of the system respectively.

[0144] The third step is to improve the PINNs model and introduce it into the loss function As a physical information regulator, a reactivation algorithm is introduced to optimize the loss function.

[0145] Among them, Physics-informed Neural Networks (PINNs) are neural networks that can be used to solve supervised learning tasks and were proposed by a team from Brown University in 2019. Compared with traditional data-based artificial neural networks (ANNs), PINNs not only have the ability to learn the distribution laws of data samples, but can also guide the model through physical laws expressed in mathematical equations, thereby better utilizing physics knowledge to enhance its prediction and classification capabilities;

[0146] The process is as follows:

[0147] Step (3.1) designs the loss function of the PINNs model into two parts, one of which is the loss of the neural network. , which is used to represent the error between the predicted value and the true value based on the network, and the other part is the loss of physical information , the loss of physical information includes the initial value loss , Boundary loss and the residual of the ordinary differential equation ;

[0148] The process of step (3.1) is as follows:

[0149] Step (3.1.1) The loss of the neural network is as follows,

[0150] ;

[0151] in, is the number of sampled data, The operator representing the mean square error (MSE), represents the predicted value of the neural network, Indicates the actual measured value;

[0152] Step (3.1.2) A core idea of ​​PINNs is to make the neural network satisfy the given physical laws. The physical laws are expressed in the form of ODE. The ODE residual loss is used to measure whether the solution of the network satisfies these equations. The residual of the ordinary differential equation is as follows,

[0153] ;

[0154] in, is the number of sampled data, The operator representing the mean square error (MSE), represents the actual measured value, for any time;

[0155] Step (3.1.3) The boundary conditions define the behavior of the solution at the boundary of the problem region, and the initial conditions define the state of the system at time t = 0. The boundary conditions and initial conditions of PINNs are obtained through actual measurement and estimation as follows:

[0156] ;

[0157] Step (3.1.4) The boundary loss function compares the boundary conditions predicted by the model with the actual physical boundary conditions, so that the solution of the model in the boundary area conforms to the physical laws. The boundary loss is as follows:

[0158] ;

[0159] in, is the number of sampled data, The operator representing the mean square error MSE, represents the predicted value of the neural network, Indicates the actual measured value;

[0160] Step (3.1.5) The initial value loss function measures the difference between the model prediction value and the actual initial conditions to ensure that the network output meets the physical constraints at the initial moment. The initial value loss is as follows:

[0161] ;

[0162] in, is the number of sampled data, The operator representing the mean square error (MSE), represents the predicted value of the neural network, Indicates the actual measured value;

[0163] Step (3.1.6) Add the losses mentioned in steps (3.1.1) to (3.1.6) to get the final loss function of the PINNs model. The formula is as follows:

[0164] ;

[0165] in, It is a physical information regulator. By introducing λ, PINNs can accurately estimate the output of the equation. By introducing physical information into the loss function, λ plays an important role in determining the contribution of physics to the final loss function. As λ increases, the output of PINNs will gradually rely on the constraints of the physical equation, while reducing λ will make the actual measured data more influential on the output; if the value of λ is set to 0, the entire loss function does not contain the physical information term and becomes a traditional ANN, then the network only relies on the data during training.

[0166] Step (3.2) Design a reactivation algorithm to obtain the optimal solution of the model;

[0167] In the step (3.2), using PINNs to solve the differential equation is an optimization problem of minimizing the loss function. For the vibration dynamics equation of the elevator, when using an optimization algorithm with line search and convergence criteria, a local optimal solution can be predicted. The local optimal solution can be regarded as the exact solution obtained. However, sometimes the local optimal solution obtained may have a large deviation. In order to avoid the situation where the solution has a large deviation, a reactivation technology is proposed to reactivate the optimization algorithm so that it can get the optimal solution again; the process is as follows:

[0168] Step (3.2.1) is as follows Figure 2 As shown in the PINNs section, PINNs uses the Adma optimizer with strong Wolfe. In the kth iteration, a step size αk is found that satisfies the following conditions:

[0169] Function value decreasing condition,

[0170] ;

[0171] Strong curvature condition,

[0172] ;

[0173] in, is the search direction, and is a constant of the strong Wolfe condition, 0< < <1, is the current point, is the gradient of the current point. If the optimization falls into a local optimal solution at the beginning of the kth iteration, there is no αk that satisfies the equation. If no αk is found to satisfy the equation, the optimization is stopped.

[0174] Step (3.2.2) In order to restart the optimization process, introduce The constant makes the amplification factor slightly greater than zero, giving the following formula,

[0175] ;

[0176] The above formula is used to replace the function reduction condition. The strong curvature condition is only used in the first iteration of reactivation. The introduction of the above formula allows Greater than , making it easier to find αk that satisfies the equation; once found, the reactivation optimization algorithm will continue to use the standard strong Wolfe condition, that is, step (3.2.1).

[0177] Step 4: Acquire simulated vibration data. Input a small amount of elevator fault vibration signals collected in the first step into the PINNs model with motion equation constraints to obtain simulated elevator vibration data.

[0178] like Figure 2 As shown, the acquired simulated vibration data is divided into normal data and fault data and labeled as the input data of the e-RGCN network;

[0179] The fifth step is to improve the GCN model for elevator fault diagnosis. The process is: combine the recursive neural network RNN ​​and the graph convolutional network GCN to propose the e-RGCN fault diagnosis algorithm, set a recursive radius, construct a correlation graph through the recursive radius, and then GCN can effectively identify the elevator fault status through the constructed correlation graph.

[0180] The process of the fifth step is as follows:

[0181] Step (5.1) constructs the correlation graph. e-RNN is one of the methods for complex network analysis of time series. Recursive network is an intuitive method that uses the recursiveness of mechanical systems to convert time series into graphs. Unlike traditional time series analysis that focuses more on the sequence of samples, recursive network analysis focuses more on the spatial topological structure of samples and the relationship between samples.

[0182] The processing process of step (5.1) is as follows:

[0183] Step (5.1.1) constructs a graph that requires a relationship between the central vertex and other vectors. Two vertices can only be There is an edge between two vertices only when is the Euclidean norm between two vectors, is the recursive radius, which is the global threshold of RNN and can adjust the criteria of recursive pairs; two vertices are recursive pairs only when the conditions are met; in addition, the topological structure of all samples is obtained by performing pairwise operations on all vectors, and the structure of the network is encoded into a recursive matrix expressed as follows,

[0184] ;

[0185] in, is the Heaviside function, expressed as follows,

[0186] ;

[0187] in, is the Euclidean norm of a vector, if and There is an edge between The value of is 1, otherwise The value of is 0;

[0188] Step (5.1.2) The adjacency matrix of the graph can be constructed by recursive matrix as follows,

[0189] ;

[0190] in is the Kronecker function, expressed as,

[0191] ;

[0192] The graph of the vibration signal is obtained through the obtained adjacency matrix.

[0193] Step (5.2) Construction of the e-RGCN fault diagnosis network. The proposed e-RGCN classification network consists of two GCN layers, one fully connected layer and an activation function. The GCN layer extracts graph features from the association graph, and the fully connected layer is used as a classifier for elevator fault diagnosis.

[0194] The processing process of step (5.2) is as follows:

[0195] Step (5.2.1) uses the cross-entropy loss function as the loss function of the network output. The cross-entropy loss function is a loss function in machine learning and deep learning. Especially in classification problems, it is used to measure the difference between the model's predicted value and the true label. It is particularly suitable for classification tasks (such as binary classification and multi-classification tasks). The loss function of the e-RGCN model is Indicates the degree of match between the predicted label and the actual label, as follows,

[0196] ;

[0197] Where N is the number of samples, C is the number of categories, is the probability of the jth category in the true label of the i-th sample, is the jth category in the model's predicted probability value for the i-th sample;

[0198] Step (5.2.2) During the e-RGCN training process, multi-layer GCN mixes the features of the node itself and its neighbors. Even if an edge is deleted, information about the node can still be obtained from its neighbors. Drop Edge randomly deletes some edges in the original graph, which increases the generalization of the network.

[0199] ;

[0200] Among them, A is the graph structure input of GCN, This is the graph structure after randomly selecting some edges.

[0201] Step (5.3) inputs the graph into the network, extracts features through two GCN layers, and obtains deep features; then classifies the deep features through a softmax classification layer, and outputs the fault type classification result;

[0202] Among them, the Softmax function is a commonly used activation function, which is often used in multi-category classification problems. The Softmax function can convert a real number vector into a vector representing a probability distribution, so that each element in the vector is between 0 and 1, and the sum of all elements is 1.

[0203] The implementation process of this embodiment is as follows:

[0204] Step 1: Define 6 fault diagnosis models.

[0205] LeNet-5: Through convolution operations, local image features are extracted and the number of parameters of the fully connected layer is reduced. The parameters of the same convolution kernel are shared across the entire input image, which greatly reduces the model parameters. The pooling layer is used to reduce the size of the feature map, retain important features, and reduce computational complexity.

[0206] ResNet: By introducing residual connections, the gradient vanishing problem in deep network training is solved, making it possible to stably train extremely deep network structures and improve performance.

[0207] DGCN: Enhances the flexibility and expressiveness of graph convolutional networks by dynamically adjusting graph structures and node features, thereby improving the modeling effect of complex graph data.

[0208] GAT: Dynamically assigns weights to different neighbors in the graph through the self-attention mechanism, thereby effectively capturing important information in the graph structure and the complex relationships between nodes.

[0209] SGCN: Improves the computational efficiency and storage performance of graph convolutional networks on large-scale graph data by introducing a sparse graph convolution mechanism while maintaining high expressive power.

[0210] GraphSAGE: It performs efficient feature learning in graphs by sampling and aggregating node neighbors. It is suitable for processing large-scale graph data and dynamic graph structures.

[0211] Step 2: Analyze and compare the results.

[0212] Reference Figure 3 It can be seen from the results that the two methods without graph structure: LeNet-5 and ResNet did not perform very well in the experiment, because these two models did not take into account the structural characteristics of the graph. They completely relied on data-driven methods for training. Due to the limitation of training data, the model could not fully exert its potential and was prone to overfitting, which may lead to poor performance of the model in identifying fault information, thus affecting its overall performance; other methods and e-RGCN are all based on graph structure. GCN captures the dependencies between time series from the perspective of graph topology, which enables them to learn and extract feature information of nodes and capture important information of edges between nodes. Therefore, the latter model is more accurate than the above two models. The reason why the e-RGCN model of the present invention is superior to several other GCN models is that the construction method of graph data directly determines the quality of graph data, so it is crucial to ensure high-quality fault sample graphs. In the present invention, e-RGCN is used to construct high-quality fault sample graphs, so that structured feature blocks can better identify faults. It can be shown from the figure that the method proposed by the present invention has good stability.

[0213] The contents described in the embodiments of this specification are merely enumerations of implementation forms of the inventive concept and are for illustrative purposes only. The protection scope of the present invention should not be considered to be limited to the specific forms described in this embodiment, and the protection scope of the present invention also extends to equivalent technical means that can be thought of by ordinary technicians in this field based on the inventive concept.

Claims

1. An elevator fault diagnosis method based on digital twins, characterized in that: The method comprises the following steps: The first step is to collect elevator vibration signals; The second step is to construct a numerical model of the elevator in the vertical direction. First, the dynamic equation of the system is established, and then the kinetic energy, potential energy and dissipation coefficient of the system are calculated. The third step is to improve the PINNs model and introduce it into the loss function As a physical information regulator, the reactivation algorithm is introduced to optimize the loss function. The process is as follows: Step (3.1) designs the loss function of the PINNs model into two parts, one of which is the loss of the neural network. , which is used to represent the error between the predicted value and the true value based on the network, and the other part is the loss of physical information , the loss of physical information includes the initial value loss , Boundary loss and the residual of the ordinary differential equation ; Step (3.2) Design a reactivation algorithm to obtain the optimal solution of the model; Step 4: Acquire simulated vibration data. Input a small amount of elevator fault vibration signals collected in the first step into the PINNs model with motion equation constraints to obtain simulated elevator vibration data. The fifth step is to improve the GCN model for elevator fault diagnosis. The process is: combine the recursive neural network RNN ​​and the graph convolutional network GCN to propose the e-RGCN fault diagnosis algorithm, set a recursive radius, construct a correlation graph through the recursive radius, and then GCN can effectively identify the elevator fault status through the constructed correlation graph.

2. The elevator fault diagnosis method based on digital twin according to claim 1, characterized in that: In the first step, the vibration signal of the elevator is collected at a frequency of 50 Hz by an acceleration sensor installed in the car top box on the top of the elevator car, and remote data acquisition is achieved through the TCP / IP communication protocol; the sensor, elevator and gateway are connected using a serial port connection, and Modbus and RS485 protocols are used for data communication.

3. The elevator fault diagnosis method based on digital twin according to claim 1 or 2, characterized in that: The process of the second step is as follows: Step (2.1) establishes the dynamic equation of the system, the formula is as follows, ； Where T is the kinetic energy of the system; U is the potential energy of the system; D is the dissipation coefficient of the system; x is the generalized displacement; is the generalized velocity; Q is the generalized force; Step (2.2) Calculate the kinetic energy of the system , the potential energy of the system and system dissipation factor , ； ； ； in, is the equivalent mass of the car and the loaded car; is the counterweight mass; , are the equivalent stiffness and damping of the wire rope between the car and the traction sheave respectively; , are the equivalent stiffness and damping of the wire rope between the counterweight and the traction sheave respectively; , and They are the car vibration linear displacement, counterweight vibration linear displacement, and traction sheave vibration angular displacement; , are the moment of inertia of the traction sheave and the radius of the rope groove of the traction sheave respectively; Step (2.3) Arrange the formulas in step (2.2) and step (2.1) as follows: Step (2.3.1) for generalized coordinates ,have, ； Step (2.3.2) Since the elevator car is only affected by the vertical wire rope tension and gravity, substitute the result of step (2.3.1) into the Lagrange equation of step (2.1) to obtain, ； Step (2.3.3) for generalized coordinates ,have, ； Step (2.3.4) The elevator counterweight is only affected by the vertical wire rope tension and gravity. Substituting the result of step (2.3.3) into the Lagrange equation of step (2.1), we get: ； Step (2.3.5) for generalized coordinates ,have, ； Step (2.3.6) for generalized coordinates , the system inertia force in the elevator start-brake stage is used as the excitation source, and the exciting force is expressed as, ； Step (2.4), Step (2.3.2), Step (2.3.2) and Step (2.3.2) constitute the differential equation of motion, which can be expressed in matrix form as follows, ； in, ； ； ； ； ； Among them, M is the system mass matrix, C is the system damping matrix, K is the system stiffness matrix, and F is the exciting force matrix; , , They represent the acceleration, velocity and displacement of the system respectively.

4. The elevator fault diagnosis method based on digital twin according to claim 1 or 2, characterized in that: The process of step (3.1) is as follows: Step (3.1.1) The loss of the neural network is as follows, ； in, is the number of sampled data, The operator representing the mean square error MSE, represents the predicted value of the neural network, Indicates the actual measured value; Step (3.1.2) A core idea of ​​PINNs is to make the neural network satisfy the given physical laws. The physical laws are expressed in the form of ODE. The ODE residual loss is used to measure whether the solution of the network satisfies these equations. The residual of the ordinary differential equation is as follows, ； in, is the number of sampled data, The operator representing the mean square error (MSE), represents the actual measured value, for any time; Step (3.1.3) The boundary conditions define the behavior of the solution at the boundary of the problem region, and the initial conditions define the state of the system at time t = 0. The boundary conditions and initial conditions of PINNs are obtained through actual measurement and estimation as follows: ； Step (3.1.4) The boundary loss function compares the boundary conditions predicted by the model with the actual physical boundary conditions, so that the solution of the model in the boundary area conforms to the physical laws. The boundary loss is as follows: ； in, is the number of sampled data, The operator representing the mean square error MSE, represents the predicted value of the neural network, Indicates the actual measured value; Step (3.1.5) The initial value loss function measures the difference between the model prediction value and the actual initial conditions to ensure that the network output meets the physical constraints at the initial moment. The initial value loss is as follows: ； in, is the number of sampled data, The operator representing the mean square error (MSE), represents the predicted value of the neural network, Indicates the actual measured value; Step (3.1.6) Add the losses mentioned in steps (3.1.1) to (3.1.6) to get the final loss function of the PINNs model. The formula is as follows: ； in, , is a physical information regulator, and by introducing λ, PINNs can accurately estimate the output of the equation.

5. The elevator fault diagnosis method based on digital twin according to claim 4, characterized in that: The process of step (3.2) is as follows: Step (3.2.1) PINNs uses the Adma optimizer with strong Wolfe. In the kth iteration, it will find a step size αk that satisfies the following conditions, Function value decreasing condition, ； Strong curvature condition, ； in, is the search direction, and is a constant of the strong Wolfe condition, 0< < <1, is the current point, is the gradient of the current point. If the optimization falls into a local optimal solution at the beginning of the kth iteration, there is no αk that satisfies the equation. If no αk is found to satisfy the equation, the optimization is stopped. Step (3.2.2) In order to restart the optimization process, introduce The constant makes the amplification factor slightly greater than zero, giving the following formula, ； The above formula is used to replace the function reduction condition. The strong curvature condition is only used in the first iteration of reactivation. The introduction of the above formula allows Greater than , making it easier to find αk that satisfies the equation; once found, the reactivation optimization algorithm will continue to use the standard strong Wolfe condition, that is, step (3.2.1).

6. The elevator fault diagnosis method based on digital twin according to claim 1 or 2, characterized in that: The process of the fifth step is as follows: Step (5.1) construct an association graph; Step (5.2) Construction of the e-RGCN fault diagnosis network. The proposed e-RGCN classification network consists of two GCN layers, one fully connected layer and an activation function. The GCN layer extracts graph features from the association graph, and the fully connected layer is used as a classifier for elevator fault diagnosis. Step (5.3) inputs the graph into the network, extracts features through two GCN layers, and obtains deep features; then classifies the deep features through a softmax classification layer, and outputs the fault type classification result.

7. The elevator fault diagnosis method based on digital twin according to claim 6, characterized in that: The processing process of step (5.1) is as follows: Step (5.1.1) constructs a graph that requires a relationship between the central vertex and other vectors. Two vertices can only be There is an edge between two vertices only when is the Euclidean norm between two vectors, is the recursive radius, which is the global threshold of RNN and can adjust the criteria of recursive pairs; two vertices are recursive pairs only when the conditions are met; in addition, the topological structure of all samples is obtained by performing pairwise operations on all vectors, and the structure of the network is encoded into a recursive matrix expressed as follows, ； in, is the Heaviside function, expressed as follows, ； in, is the Euclidean norm of a vector, if and There is an edge between The value of is 1, otherwise The value of is 0; Step (5.1.2) The adjacency matrix of the graph can be constructed by recursive matrix as follows, ； in is the Kronecker function, expressed as, ； The graph of the vibration signal is obtained through the obtained adjacency matrix.

8. The elevator fault diagnosis method based on digital twin according to claim 6, characterized in that: The processing process of step (5.2) is as follows: Step (5.2.1) uses the cross entropy loss function as the loss function of the network output, and the loss function of the e-RGCN model is Indicates the degree of match between the predicted label and the actual label, as follows, ； Where N is the number of samples, C is the number of categories, is the probability of the jth category in the true label of the i-th sample, is the jth category in the model's predicted probability value for the i-th sample; Step (5.2.2) During the e-RGCN training process, multi-layer GCN mixes the features of the node itself and its neighbors. Even if an edge is deleted, information about the node can still be obtained from its neighbors. Drop Edge randomly deletes some edges in the original graph, which increases the generalization of the network. ； Among them, A is the graph structure input of GCN, This is the graph structure after randomly selecting some edges.

Images