Elevator production process fault prediction method based on digital twinning and transfer learning

Abstract

An elevator production process fault prediction method based on digital twinning and transfer learning comprises the following steps that firstly, an elevator manufacturing workshop production process industrial Internet of Things system based on digital twinning is constructed; 2, constructing a model based on a deep residual error and a long and short-term memory network; step 3, building an elevator production process fault prediction model based on digital twinning and transfer learning: building an effective virtual model including a geometric model, a physical model, a behavior model and a rule model; the method comprises the steps of designing an adversarial migration network; designing a loss function; carrying out transfer learning network training on the marked virtual data Ds and the unmarked physical data Dt obtained by processing; in the training process, parameters in the network are updated by optimizing a loss function, and fault prediction in the production process of the elevator manufacturing workshop is achieved. The fault information of the elevator manufacturing workshop production process can be accurately predicted.

Description

Technical Field

[0001] This invention belongs to the field of elevator data processing technology, specifically relating to a method for predicting elevator production process faults based on digital twins and transfer learning. Background Technology

[0002] Guided by policy, intelligent manufacturing technologies have been widely applied. Elevator production, as a complex manufacturing system, encompasses multiple production stages and equipment. These devices not only play a crucial role in the overall quality of the elevator but are also closely related to its safety and stability. However, traditional methods for predicting elevator production failures often rely on the experience and intuition of equipment managers, judging equipment status through manual inspections and analysis of historical failure data. This approach is not only inefficient but also susceptible to human error, making it difficult to adapt to increasingly complex production environments and the diversity of equipment failures.

[0003] With the rapid development of Industrial Internet of Things (IIoT) technology, a large number of sensors and smart devices are deployed in elevator manufacturing workshops to collect various data in real time during the production process, including equipment operating status, temperature, humidity, pressure, and other parameters. While the acquisition of this data provides a rich source for fault prediction, the sheer volume and complexity of the data make extracting effective information and performing accurate fault prediction a pressing technical challenge.

[0004] Digital twin technology, by creating virtual models of physical equipment and production systems, enables real-time monitoring and simulation of the production process and equipment status. This technology provides accurate status assessments for equipment in elevator production, helping managers predict and make maintenance decisions before equipment failures occur. Through digital twin models, the operating environment of equipment can be simulated in real time, potential failure modes can be analyzed, and prediction algorithms can be continuously optimized through data feedback mechanisms to achieve predictive fault diagnosis. This allows elevator production workshops to improve equipment reliability and production efficiency without significantly increasing labor costs. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention proposes a fault prediction method for elevator manufacturing processes. First, an Industrial Internet of Things (IIoT) system is used to acquire sensor signals from equipment in the elevator manufacturing workshop. Then, digital twin technology is used to establish a correlation model between various stages of the workshop production process, and simulation data for common elevator manufacturing workshop processes is generated. Finally, a fault prediction model for the elevator manufacturing workshop production process is constructed based on Residual Networks (ResNet), Long Short-Term Memory (LSTM), and Transfer Learning (TL). Simulation data and updated data are used as the source and target domains, respectively, to pre-train the fault prediction model. Intelligent fault prediction for the elevator manufacturing workshop production process is achieved through model training and testing.

[0006] This invention provides the following technical solution:

[0007] A method for predicting elevator production process failures based on digital twins and transfer learning, the method comprising the following steps:

[0008] The first step is to build an industrial Internet of Things (IoT) system for the elevator manufacturing workshop based on digital twins. The process is as follows:

[0009] Step (1.1) The data collection layer collects physical signals and status information of equipment during the elevator manufacturing process;

[0010] Furthermore, the underlying equipment of the industrial Internet of Things system is deployed in the elevator manufacturing workshop, while collecting relevant data generated by mobile devices;

[0011] The underlying equipment of the industrial Internet of Things (IIoT) system includes production equipment, sensors, automated guided vehicles (AGVs), and robotic arms; mobile devices include mobile phones and computers.

[0012] Step (1.2) transmits the collected device physical signals and status information to the data analysis and storage layer through the industrial Internet of Things system;

[0013] Among them, the data analysis and storage layer is responsible for preprocessing, cleaning and format conversion of data, transforming it into a standard format that can be accepted and processed by machine learning models, and storing and managing the data;

[0014] Step (1.3) In the digital twin layer, a virtual-physical integrated digital twin model is constructed based on the actual production process of the elevator manufacturing workshop;

[0015] Among them, the digital twin model can accurately simulate the production process and equipment operation status of the elevator manufacturing workshop in virtual space, and perform real-time monitoring and dynamic simulation of the data collected by the data acquisition layer;

[0016] Step (1.4) transmits the real-time data collected by the field equipment to the digital twin layer, updates the state parameters of the virtual model, so that it can approximate the operating state of the real production environment and realize the dynamic synchronization between the virtual and real environments.

[0017] The second step is to construct a network model based on deep residuals and long short-term memory, as follows:

[0018] Step (2.1) constructs a fusion model (ResLSTM) based on deep residual networks and long short-term memory networks for feature extraction of fault data in the elevator manufacturing workshop. The core structure of the ResLSTM model includes multiple blocks, each block consisting of the following layers: Batch Normalization (BN) layer, ReLU activation layer, LSTM layer, Attention layer, Dense layer, Max Pooling layer, and Add layer;

[0019] Among them, batch normalization is a technique to improve the training speed and stability of neural networks; the activation layer maps the input of the neural network to the output; the LSTM layer is a special type of recurrent neural network that can process and predict sequential data; the attention mechanism layer enables the model to focus on important information by giving greater weight to certain parts of the input sequence; the dense layer is a fully connected layer that connects each neuron of the previous layer to each neuron of the current layer; the max pooling layer is used to extract the most important features from the input data; and the addition layer is used to combine multiple inputs, usually by adding the outputs of different previous layers.

[0020] In step (2.2), all blocks are connected sequentially to ensure that forward and backward signals can flow directly through any block, thereby improving the generalization ability and stability of the model.

[0021] In order to further optimize model performance, convolutional layers, as shortcuts for each block, are combined with the main chain (each layer in the block) to achieve skip connections and extract different features from multidimensional data.

[0022] Convolutional layers are a type of layer commonly used to process data with network structures.

[0023] Step (2.3) employs a pre-activation strategy to process the ReLU layer between the BN layer and the LSTM layer;

[0024] Step (2.4) After data cleaning and feature selection, the preprocessed time series data is input into the network for further processing;

[0025] The third step is to construct a fault prediction model for the elevator production process based on digital twins and transfer learning. The process is as follows:

[0026] Step (3.1) Establish an effective virtual model, including a geometric model, a physical model, a behavioral model, and a rule model.

[0027] Among them, the geometric model and physical model describe the heterogeneous elements in the workshop; the behavioral model determines the interaction mechanism by adding driving and perturbation factors to the geometric model and physical model, enabling the virtual model to perform complex behaviors; the rule model better describes the conventions and restrictions in the production system through a series of logical rules, conditional judgments and constraints.

[0028] Step (3.2) designs an adversarial migration network to reduce data discrepancies between virtual and physical spaces. This network consists of a classifier module and a domain discriminator module.

[0029] The classifier module is used to output the probability of fault prediction and is constructed using the classic Softmax method; the domain discriminator module adopts a binary classifier design to distinguish the learned features from physical space and virtual space.

[0030] Softmax is an activation function commonly used in multi-class classification problems.

[0031] Step (3.3) involves designing a loss function to further enhance the weights of the virtual samples;

[0032] Step (3.4) processes the labeled virtual data obtained from steps (3.2) and (3.3). and unlabeled physical data The transfer learning network is trained on it; during the training process, the parameters in the network are updated by optimizing the loss function, and finally the fault prediction of the elevator manufacturing workshop production process is realized.

[0033] Furthermore, the process of step (2.3) is as follows:

[0034] Step (2.3.1) optimizes the data distribution by normalizing the data in the activation layer, thereby further improving the optimizability and stability of the model;

[0035] Data normalization aims to improve the optimization performance and stability of a model by adjusting the scale or distribution of the input data to suit the training needs of the model.

[0036] In step (2.3.2), to avoid overfitting, the BN layer normalizes all signals, which enhances the robustness of the network.

[0037] Normalization is a technique used to prevent model overfitting.

[0038] The process of step (2.4) is as follows:

[0039] In step (2.4.1), the data is fed into the network through a batch normalization layer. The activation layer uses the nonlinear function ReLU to simplify the optimization process and achieve regularization.

[0040] Step (2.4.2) The LSTM layer extracts features from the input data, extracting temporal features from the time series data;

[0041] In the first block, convolutional layers reduce the dimensionality of the input data to accommodate subsequent operations. Dense layers extract the correlations between features through non-linear computation and provide mappings for the output layer. The attention mechanism layer refocuses the features, allocating different attention shares based on the correlations between signals.

[0042] In step (2.4.3), the residual features and quick features after max pooling are summed and input into the connection layer. By expanding the feature tensor, the connection layer effectively reduces computation and avoids overfitting, thus enhancing feature extraction capabilities.

[0043] Furthermore, the process of step (3.2) is as follows:

[0044] Step (3.2.1) reduces the difference between the virtual and physical spaces through adversarial training, ensuring that fault categories can be shared between the two spaces. The objective function is calculated as follows:

[0045]

[0046] in, and It is the objective function of the classifier; n v and n p These are the quantities of labeled virtual data and unlabeled physical data, respectively; L y G is represented as the cross-entropy loss of the classifier module. y and G d Represents a classifier; G f It is a generator; and These represent the feature vectors of the i-th sample of the virtual data and the j-th sample of the physical data, respectively, obtained from the feature extraction in the second step. It is the true label of the i-th sample;

[0047] Step (3.2.2) Design a weighted learning module based on the loss function.

[0048] Among them, the weighted learning module reduces the impact of irrelevant virtual fault data by quantifying the contribution of different virtual fault categories, thereby avoiding the negative transfer problem;

[0049] The weighted learning module includes: an auxiliary domain discriminator, used to distinguish input data from physical space and virtual space; and an auxiliary classifier, used to assist learning through the label information of virtual data.

[0050] Step (3.2.3) The auxiliary classifier consists of a hidden layer and an output layer equipped with Leaky Softmax, whose output value is represented as...

[0051]

[0052] in, It is an auxiliary classifier The output for input z; k v Indicates the number of virtual fault categories; when moving from virtual space to... When providing input samples, The output value is large; otherwise, The output value is small. The output is represented as:

[0053]

[0054] in, This is the output of the discriminator, and its purpose is to distinguish the source of the input samples; G f (x i ) is the feature extractor G f For sample input x i Extracted feature representation; This is the output function of the auxiliary classifier for category k. Considering that physical fault data and virtual fault data of the same fault category share common fault knowledge, when the input sample completely belongs to the virtual fault category, The output is a large probability value, which means that it can be determined based on... The output identifies the shared fault categories between the physical and virtual spaces, therefore, and The loss function is calculated as follows:

[0055]

[0056] in, and These represent the loss functions of the auxiliary classifier; The true labels represent the sample sets i and k;

[0057] Step (3.2.4) designs an auxiliary domain discriminator. To increase the weight of the virtual samples, the loss function of the adversarial transfer network is expressed as:

[0058]

[0059] in, and These represent the loss functions of the adversarial transfer network; It is the weight function of the sample; this module assists learning by classifying the input samples, and its output value is larger when the input sample belongs to the virtual space.

[0060] Among them, the auxiliary domain discriminator is used to distinguish the source of the input data.

[0061] In step (3.3), the loss function formula is:

[0062]

[0063] in, and These represent the optimal parameters of the feature extractor, the classifier, and the domain discriminator, respectively, which can be iteratively updated using the Adams algorithm based on the adversarial domain adaptation method.

[0064] These are the parameters of the auxiliary classifier; argmin represents the optimization under the minimum constraint.

[0065] The process of step (3.4) is as follows:

[0066] Step (3.4.1) The training optimization objective of the transfer learning network model consists of two parts: the traditional loss function and the adaptive loss function. The loss function measures the difference between the predicted value and the true value, while the adaptive loss measures the difference in the domain distribution.

[0067]

[0068] Where l is the total loss of the transfer learning network; y s It is the source domain validation set, used to evaluate classification performance; It is the traditional classification loss of transfer learning networks on labeled data (source domain); λ is the adaptive loss of the transfer learning network; λ is the penalty factor.

[0069] Step (3.4.2) uses the cross-entropy function as the conventional classification loss function in the transfer learning process:

[0070]

[0071] Where M is the number of fault categories. * It is a sign function: it is 1 if the true class of sample i is equal to c, and 0 otherwise. * ic p represents the true label of the i-th sample for class c; * ic It is the predicted probability that observed sample i belongs to category c;

[0072] Step (3.4.3) uses Multi-kernel Maximum Mean Discrepancy (MK-MMD) to represent the similarity measure of sample features between the source and target domains. It is developed based on MMD. If the features are mapped to the Hilbert space through a kernel function, the distance between the two distributions can be measured by the average difference between the samples in the two domains after mapping. The mathematical expression of MMD is as follows:

[0073]

[0074] Where MMDX,Y represents the maximum average difference between the source domain X and the target domain Y; X is the set of samples in the source domain; and Y is the set of samples in the target domain. It is a kernel function mapping; It is the norm in Hilbert space;

[0075] The kernel function used in MK-MMD is a weighted kernel of several different kernel functions:

[0076]

[0077] Where κ is a multi-kernel function; k μ It is the μ-th basis kernel function; β μ is the weight of the μ-th kernel function; m is the number of basis kernel functions;

[0078] In this study, a portion of labeled source domain data and a portion of unlabeled target domain data are used as the training set. Features are extracted using a feature extraction model, and then converted into a probability distribution using a feature classifier. Next, the classification loss and distribution difference of the source domain features are calculated using the aforementioned set of loss functions. The goal is to find the model parameters that minimize the optimization function value, thereby achieving the target domain classification task.

[0079] Compared with existing methods, the beneficial effects of the present invention are as follows:

[0080] (1) A feature extraction method based on ResNet and LSTM is proposed, which comprehensively considers the change information of multi-sensor IIOT system data and accurately predicts the fault information of elevator manufacturing workshop production process.

[0081] (2) Applying ResNet to time series prediction solves the gradient vanishing problem. Replacing the Conv1D layer of ResNet with an LSTM layer improves the model's feature extraction of time series data.

[0082] (3) The effectiveness of the proposed method was verified on a real elevator manufacturing workshop production process dataset. Using RMSE, MAPE, R2, MAR, and FAR as evaluation metrics, the proposed model achieved excellent accuracy. Attached Figure Description

[0083] Figure 1 This is a schematic diagram of the ResLSTM model based on deep residual and long short-term memory networks. Detailed Implementation

[0084] The present invention will now be further described with reference to the accompanying drawings.

[0085] Reference Figure 1 A method for predicting elevator production process failures based on digital twins and transfer learning includes the following steps:

[0086] The first step is to build an industrial Internet of Things (IoT) system for the elevator manufacturing workshop based on digital twins. The process is as follows:

[0087] Step (1.1) The data collection layer collects physical signals and status information of equipment during the elevator manufacturing process;

[0088] Furthermore, the underlying equipment of the industrial Internet of Things system is deployed in the elevator manufacturing workshop, while collecting relevant data generated by mobile devices;

[0089] The underlying equipment of the industrial Internet of Things (IIoT) system includes production equipment, sensors, automated guided vehicles (AGVs), and robotic arms; mobile devices include mobile phones and computers.

[0090] Step (1.2) transmits the collected device physical signals and status information to the data analysis and storage layer through the industrial Internet of Things system;

[0091] Among them, the data analysis and storage layer is responsible for preprocessing, cleaning and format conversion of data, transforming it into a standard format that can be accepted and processed by machine learning models, and storing and managing the data;

[0092] Step (1.3) In the digital twin layer, a virtual-physical integrated digital twin model is constructed based on the actual production process of the elevator manufacturing workshop;

[0093] Among them, the digital twin model can accurately simulate the production process and equipment operation status of the elevator manufacturing workshop in virtual space, and perform real-time monitoring and dynamic simulation of the data collected by the data acquisition layer;

[0094] Step (1.4) transmits the real-time data collected by the field equipment to the digital twin layer, updates the state parameters of the virtual model, so that it can approximate the operating state of the real production environment and realize the dynamic synchronization between the virtual and real environments.

[0095] The second step is to construct a network model based on deep residuals and long short-term memory.

[0096] like Figure 1 As shown, the preprocessed data is fed as a time series to the batch normalization layer and convolutional layer. ReLU is a non-linear function in the activation layer used for normalization. Then, the LSTM layer performs feature extraction, and the CNN layer in Block 1 adapts the dimension of the input data to the output of Block 1. In addition, the dense layer extracts the correlation between features through non-linear computation and provides a mapping to the output space. Based on this, the features are refocused in the attention mechanism layer according to the correlation between the target variables to obtain different attention shares. Then, the new redistributed data is fed into the LSTM layer. After max pooling, the remaining features and convolutional layers are added and fed into the next block. In this model, the residual results are combined with the initial feature information of the input layer to obtain a deeper matrix. By widening the feature tensor, the cascaded layers can reduce the amount of computation, avoid overfitting, and improve the feature extraction capability.

[0097] The second step is as follows:

[0098] Step (2.1) Construct a fusion model (ResLSTM) based on deep residual network and long short-term memory network for feature extraction of fault data in elevator manufacturing workshop. The core structure of the ResLSTM model includes multiple blocks, each block consisting of the following layers: Batch Normalization (BN) layer, ReLU activation layer, LSTM layer, Attention mechanism layer, Dense layer, Max Pooling layer and Add layer;

[0099] Among them, batch normalization is a technique to improve the training speed and stability of neural networks; the activation layer maps the input of the neural network to the output; the LSTM layer is a special type of recurrent neural network that can process and predict sequential data; the attention mechanism layer enables the model to focus on important information by giving greater weight to certain parts of the input sequence; the dense layer is a fully connected layer that connects each neuron of the previous layer to each neuron of the current layer; the max pooling layer is used to extract the most important features from the input data; and the addition layer is used to combine multiple inputs, usually by adding the outputs of different previous layers.

[0100] In step (2.2), all blocks are connected sequentially to ensure that forward and backward signals can flow directly through any block, thereby improving the generalization ability and stability of the model.

[0101] In order to further optimize model performance, convolutional layers, as shortcuts for each block, are combined with the main chain (each layer in the block) to achieve skip connections and extract different features from multidimensional data.

[0102] Convolutional layers are a type of layer commonly used to process data with network structures.

[0103] Step (2.3) employs a pre-activation strategy to process the ReLU layer between the BN layer and the LSTM layer. The process is as follows:

[0104] Step (2.3.1) optimizes the data distribution by normalizing the data in the activation layer, thereby further improving the optimizability and stability of the model;

[0105] Data normalization aims to improve the optimization performance and stability of a model by adjusting the scale or distribution of the input data to suit the training requirements of the model.

[0106] In step (2.3.2), to avoid overfitting, the BN layer normalizes all signals, which enhances the robustness of the network.

[0107] Normalization is a technique used to prevent model overfitting.

[0108] Step (2.4) After data cleaning and feature selection, the preprocessed time series data is input into the network for further processing. The process is as follows:

[0109] In step (2.4.1), the data will be fed into the network through a batch normalization layer. The activation layer uses the nonlinear function ReLU to simplify the optimization process and achieve regularization;

[0110] Step (2.4.2) The LSTM layer extracts features from the input data, extracting temporal features from the time series data;

[0111] In the first block, the convolutional layer reduces the dimensionality of the input data to accommodate subsequent operations, the dense layer extracts the correlation between features through nonlinear calculations and provides a mapping for the output layer, and the attention mechanism layer refocuses the features and allocates different attention shares according to the correlation between signals.

[0112] In step (2.4.3), the residual features and quick features after max pooling will be summed and input into the connection layer. The connection layer effectively reduces the amount of computation and avoids overfitting by expanding the feature tensor, while enhancing the feature extraction capability.

[0113] The third step is to construct a fault prediction model for the elevator production process based on digital twins and transfer learning. The process is as follows:

[0114] Step (3.1) Establish an effective virtual model, including a geometric model, a physical model, a behavioral model, and a rule model;

[0115] Among them, the geometric model and physical model describe the heterogeneous elements in the workshop; the behavioral model determines the interaction mechanism by adding driving and perturbation factors to the geometric model and physical model, enabling the virtual model to perform complex behaviors; the rule model better describes the conventions and restrictions in the production system through a series of logical rules, conditional judgments and constraints.

[0116] Step (3.2) Design an adversarial migration network to reduce data discrepancies between virtual and physical spaces. This network consists of a classifier module and a domain discriminator module.

[0117] The classifier module is used to output the probability of fault prediction and is constructed using the classic Softmax method; the domain discriminator module adopts a binary classifier design to distinguish the learned features from physical space and virtual space.

[0118] Among them, Softmax is an activation function commonly used in multi-class classification problems;

[0119] The process of step (3.2) is as follows:

[0120] Step (3.2.1) reduces the difference between the virtual and physical spaces through adversarial training, ensuring that fault categories can be shared between the two spaces. The objective function is calculated as follows:

[0121]

[0122] in, and It is the objective function of the classifier; n v and n pThese are the quantities of labeled virtual data and unlabeled physical data, respectively; L y G is represented as the cross-entropy loss of the classifier module. y and G d Represents a classifier; G f It is a generator; and These represent the feature vectors of the i-th and j-th samples obtained from the second step of feature extraction, respectively. It is the true label of the i-th sample;

[0123] Step (3.2.2) Design a weighted learning module based on the loss function;

[0124] Among them, the weighted learning module reduces the impact of irrelevant virtual fault data by quantifying the contribution of different virtual fault categories, thereby avoiding the negative transfer problem;

[0125] The weighted learning module includes: an auxiliary domain discriminator, used to distinguish input data from physical space and virtual space; and an auxiliary classifier, used to assist learning through the label information of virtual data.

[0126] Step (3.2.3) The auxiliary classifier consists of a hidden layer and an output layer equipped with Leaky Softmax, whose output value is represented as...

[0127]

[0128] in, It is an auxiliary classifier The output for input z; Indicates the number of virtual fault categories; when moving from virtual space to... When providing input samples, The output value is large; otherwise, The output value is small. The output is represented as:

[0129]

[0130] in, This is the output of the discriminator, and its purpose is to distinguish the source of the input samples; G f (x i ) is the feature extractor G f For sample input x i Extracted feature representation; This is the output function of the auxiliary classifier for category k. Considering that physical fault data and virtual fault data of the same fault category share common fault knowledge, when the input sample completely belongs to the virtual fault category, The output is a large probability value, which means that it can be determined based on... The output identifies the shared fault categories between the physical and virtual spaces, therefore, and The loss function is calculated as follows:

[0131]

[0132] in, and These represent the loss functions of the auxiliary classifier; The true labels represent the sample sets i and k;

[0133] Step (3.2.4) designs an auxiliary domain discriminator. To increase the weight of the virtual samples, the loss function of the adversarial transfer network is expressed as:

[0134]

[0135] in, and These represent the loss functions of the adversarial transfer network; It is the weight function of the sample; this module assists learning by classifying the input samples, and its output value is larger when the input sample belongs to the virtual space;

[0136] Among them, the auxiliary domain discriminator is used to distinguish the source of the input data.

[0137] Step (3.3) proposes the following loss function formula to further enhance the weights of the virtual samples:

[0138]

[0139] in, and These represent the optimal parameters of the feature extractor, the classifier, and the domain discriminator, respectively, which can be iteratively updated using the Adams algorithm based on the adversarial domain adaptation method.

[0140] These are the parameters of the auxiliary classifier; argmin represents the optimization under the minimum constraint.

[0141] Step (3.4) processes the labeled virtual data obtained from steps (3.2) and (3.3). and unlabeled physical data A transfer learning network was trained on it. During training, the parameters in the network were updated by optimizing the loss function, ultimately achieving fault prediction in the elevator manufacturing workshop; the process is as follows:

[0142] Step (3.4.1) The training optimization objective of the transfer learning network model consists of two parts: a traditional loss function and an adaptive loss function. The traditional loss function measures the difference between the predicted value and the true value, while the adaptive loss function measures the difference in the domain distribution.

[0143]

[0144] Where l is the total loss of the transfer learning network; y s It is the source domain validation set, used to evaluate classification performance; It is the traditional classification loss of transfer learning networks on labeled data (source domain); λ is the adaptive loss of the transfer learning network; λ is the penalty factor.

[0145] Step (3.4.2) uses the cross-entropy function as the conventional classification loss function in the transfer learning process:

[0146]

[0147] Where M is the number of fault categories. * It is a sign function: it is 1 if the true class of sample i is equal to c, and 0 otherwise. * ic p represents the true label of the i-th sample for class c; * ic It is the predicted probability that observed sample i belongs to category c;

[0148] Step (3.4.3) uses Multi-kernel Maximum Mean Discrepancy (MK-MMD) to represent the similarity measure of sample features between the source and target domains. It is developed based on MMD. If the features are mapped to the Hilbert space through a kernel function, the distance between the two distributions can be measured by the average difference between the samples in the two domains after mapping. The mathematical expression of MMD is as follows:

[0149]

[0150] Where MMDX,Y represents the maximum average difference between the source domain X and the target domain Y; X is the set of samples in the source domain; and Y is the set of samples in the target domain. It is a kernel function mapping; It is the norm in Hilbert space;

[0151] The kernel function used in MK-MMD is a weighted kernel of several different kernel functions:

[0152]

[0153] Where κ is a multi-kernel function; k μ It is the μ-th basis kernel function; β μ is the weight of the μ-th kernel function; m is the number of basis kernel functions;

[0154] In this method, a portion of labeled source domain data and a portion of unlabeled target domain data are used as the training set. Features are extracted using a feature extraction model, and the features are converted into a probability distribution using a feature classifier. Then, the classification loss and distribution difference of the source domain features are calculated using the aforementioned loss function set. The model parameters that minimize the optimization function value are then sought to achieve the classification task of the target domain.

[0155] One implementation scheme of this embodiment is as follows:

[0156] Step 1: Define the fault prediction model as follows:

[0157] Gated Recurrent Unit (GRU): An improved recurrent neural network (RNN) structure that controls the flow of information by introducing update and reset gates, thus solving the gradient vanishing problem of traditional RNNs.

[0158] Long Short-Term Memory (LSTM) Network: A special type of RNN that uses input gates, forget gates, and output gates to store and selectively forget information over a long period of time, effectively mitigating the gradient vanishing problem.

[0159] CNN-LSTM: Combines the advantages of Convolutional Neural Networks (CNN) and LSTM. CNN is used to extract local features from the input data, while LSTM handles sequence dependencies.

[0160] CNN-GRU: This is a hybrid model of CNN and GRU. CNN extracts spatial or local features, while GRU handles temporal dependencies.

[0161] Step Two: Experimental Dataset. The data used in this invention is derived from a real-world verification process conducted within a discrete manufacturing enterprise, using the elevator production process as an example. The elevator production process is a highly complex system involving multiple production stages and different types of processing equipment. Therefore, taking the door panel production line as an example, various intelligent and automated equipment, such as molding machines, cutting machines, punching machines, robotic arms, welding machines, and AGV carts, work closely together to complete the manufacturing task of elevator door panels through precise coordination and interaction. In addition, the workshop is equipped with other auxiliary equipment, such as transmission systems and material handling systems, which collectively participate in a series of processes including material processing, manufacturing, assembly, and storage.

[0162] Step 3: Define the evaluation indicators. This invention uses root mean square error (RMSE), evaluation absolute percentage error (MAPE), and coefficient of determination (R2) as evaluation indicators.

[0163] Where RMSE is the absolute value of the error; MAPE is the relative value of the error (in percentage form); R2 is the proportion of variance explained by the model, which is often used to compare the goodness of fit of different models.

[0164] Step four: Analyze and compare the results, as follows:

[0165] Referring to Table 1, as an improvement over LSTM, GRU combines the forget gate and input gate, simplifying the network structure and reducing information loss during training. Therefore, it exhibits better performance in terms of training rate and generalization ability. Simultaneously, the introduction of CNN layers allows for deeper data mining through local perception and weight sharing, enhancing the capabilities of CNN-GRU. Building upon this, the network structure in this invention further prevents gradient vanishing and optimizes nonlinear representation. Furthermore, the coefficient of determination in this invention is higher than other methods, thus achieving better model fitting results. Therefore, the method of this invention has a more significant ability in feature extraction from the elevator manufacturing process.

[0166] Table 1 is a comparison table with other existing models;

[0167]

[0168] Table 1

[0169] Another implementation scheme of this embodiment is as follows:

[0170] Step 1: Define the fault prediction model: This invention utilizes labeled virtual data and unlabeled physical data to train the developed partial-domain adversarial transfer network to complete the prediction task. To further demonstrate the superiority of the proposed method, this invention also compares it with three other methods.

[0171] Non-Transfer Learning (Non-TL): To highlight the importance of knowledge transfer between virtual and physical spaces, the network structure of this method is consistent with the feature extractor and classifier of the method in this invention. It is trained only on virtual fault data, without using unlabeled physical data.

[0172] Maximum Mean Difference (MMD): Domain adaptation based on MMD is one of the most widely used transfer learning algorithms. The MMD loss is used to minimize the difference between virtual and physical data.

[0173] Adversarial learning method (ANN): It consists of a feature extractor, a classifier and a domain discriminator, and is trained by comparison using the labeled virtual fault data and unlabeled physical data in this invention.

[0174] Step two: Define the evaluation index. In this invention, accuracy is used as an important index for model evaluation.

[0175] Step 3: Analyze and compare the results, as follows:

[0176] Referring to Table 2, four experiments were conducted for each task of fault prediction in the elevator manufacturing workshop production process. Condition 1 (C1) represents the reinforcing rib forming area, and Condition 2 (C2) represents the robotic arm bending area. The test results of these four methods show that the average prediction accuracy rates obtained by the proposed method under different conditions are 94.8%, 72.8%, 90.1%, 78.7%, 94.8%, 73.2%, 86.2%, and 77.7%, respectively. The overall average prediction accuracy rate is 83.5%. Based on the above results, except for tasks C1-2 and C2-2, the results for other tasks are satisfactory, indicating that virtual fault knowledge can be effectively utilized through transfer strategies. It is important to emphasize that the method proposed in this invention is only performed with labeled virtual fault data and unlabeled physical data, meaning that the application scenario is highly consistent with the actual situation.

[0177] Table 2 shows the prediction results for different working conditions and tasks;

[0178]

[0179]

[0180] Table 2

[0181] The embodiments described in this specification are merely examples of implementations of the inventive concept and are for illustrative purposes only. The scope of protection of this invention should not be considered limited to the specific forms described in these embodiments; rather, it extends to equivalent technical means conceived by those skilled in the art based on the inventive concept.

Claims

1. A method for predicting elevator production process faults based on digital twins and transfer learning, characterized in that, The method includes the following steps: The first step is to build an industrial Internet of Things (IoT) system for the elevator manufacturing workshop based on digital twins. The process is as follows: Step (1.1) The data collection layer collects physical signals and status information of equipment during the elevator manufacturing process; Step (1.2) transmits the collected device physical signals and status information to the data analysis and storage layer through the industrial Internet of Things system; Step (1.3) In the digital twin layer, a virtual-physical integrated digital twin model is constructed based on the actual production process of the elevator manufacturing workshop; Step (1.4) transmits the real-time data collected by the field equipment to the digital twin layer, updates the state parameters of the virtual model, so that it can approximate the operating state of the real production environment and realize the dynamic synchronization between the virtual and real environments. The second step is to construct a network model based on deep residuals and long short-term memory, as follows: Step (2.1) Construct a fusion model ResLSTM based on deep residual network and long short-term memory network for feature extraction of fault data in elevator manufacturing workshop. The core structure of the ResLSTM model includes multiple blocks, each block consisting of the following layers: batch normalization layer, activation layer, LSTM layer, attention mechanism layer, dense layer, max pooling layer and addition layer. Step (2.2) Connect all blocks sequentially to ensure that forward and backward signals can flow directly through any block; Step (2.3) employs a pre-activation strategy to process the ReLU layer between the BN layer and the LSTM layer; Step (2.4) After data cleaning and feature selection, the preprocessed time series data is input into the network for further processing; The third step is to construct a fault prediction model for the elevator production process based on digital twins and transfer learning. The process is as follows: Step (3.1) Establish an effective virtual model, including a geometric model, a physical model, a behavioral model, and a rule model; Step (3.2) Design an adversarial migration network to reduce data differences between virtual and physical spaces; Step (3.3) involves designing a loss function to further enhance the weights of the virtual samples; Step (3.4) processes the labeled virtual data obtained from steps (3.2) and (3.3). and unlabeled physical data The transfer learning network is trained on it; during the training process, the parameters in the network are updated by optimizing the loss function, and finally the fault prediction of the elevator manufacturing workshop production process is realized.

2. The elevator production process fault prediction method based on digital twins and transfer learning as described in claim 1, characterized in that, The process of step (2.3) is as follows: Step (2.3.1) optimizes the data distribution by performing data normalization in the activation layer; Step (2.3.2) To avoid overfitting, the BN layer normalizes all signals.

3. The elevator production process fault prediction method based on digital twins and transfer learning as described in claim 1 or 2, characterized in that, The process of step (2.4) is as follows: In step (2.4.1), the data will enter the network through a batch normalization layer, and the activation layer will use the nonlinear function ReLU to simplify the optimization process and achieve regularization. Step (2.4.2) The LSTM layer extracts features from the input data, extracting temporal features from the time series data; In the first block, the convolutional layer reduces the dimensionality of the input data to accommodate subsequent operations; the dense layer extracts the correlation between features through non-linear computation and provides a mapping for the output layer; the attention mechanism layer refocuses the features and allocates different attention shares according to the correlation between signals. In step (2.4.3), the residual features and quick features after max pooling will be summed and input into the connection layer. The connection layer effectively reduces the amount of computation and avoids overfitting by expanding the feature tensor, while enhancing the feature extraction capability.

4. The elevator production process fault prediction method based on digital twins and transfer learning as described in claim 1 or 2, characterized in that, The process of step (3.2) is as follows: Step (3.2.1) reduces the difference between the virtual and physical spaces through adversarial training, ensuring that fault categories can be shared between the two spaces. The objective function is calculated as follows: in, and It is the objective function of the classifier; n v and n p These are the quantities of labeled virtual data and unlabeled physical data, respectively; L y G is represented as the cross-entropy loss of the classifier module. y and G d Represents a classifier; G f It is a generator; and These represent the feature vectors of the i-th sample of the virtual data and the j-th sample of the physical data, respectively, obtained from the feature extraction in the second step. It is the true label of the i-th sample; Step (3.2.2) Design a weighted learning module based on the loss function; Among them, the weighted learning module reduces the impact of irrelevant virtual fault data by quantifying the contribution of different virtual fault categories, thereby avoiding the negative transfer problem; The weighted learning module includes: an auxiliary domain discriminator, used to distinguish input data from physical space and virtual space; and an auxiliary classifier, used to assist learning through the label information of virtual data. Step (3.2.3) The auxiliary classifier consists of a hidden layer and an output layer equipped with Leaky Softmax, whose output value is represented as... in, It is an auxiliary classifier The output for input z; k v Indicates the number of virtual fault categories; when moving from virtual space to... When providing input samples, The output value is large; otherwise, The output value is small. The output is represented as: in, It is the output of the discriminator, and its purpose is to distinguish the source of the input samples. G f (x i ) is the feature extractor G f For sample input x i Extracted feature representation; It is the output function of the auxiliary classifier for category k; and The loss function is calculated as follows: in, and These represent the loss functions of the auxiliary classifier; The true labels represent the sample sets i and k; Step (3.2.4) designs an auxiliary domain discriminator. To increase the weight of the virtual samples, the loss function of the adversarial transfer network is expressed as: in, and These represent the loss functions of the adversarial transfer network; It is the weight function of the sample; this module assists learning by classifying the input samples, and its output value is larger when the input sample belongs to the virtual space.

5. The elevator production process fault prediction method based on digital twins and transfer learning as described in claim 4, characterized in that, In step (3.3), the loss function formula is: in, and These represent the optimal parameters of the feature extractor, the classifier, and the domain discriminator, respectively, which can be iteratively updated using the Adams algorithm based on the adversarial domain adaptation method. These are the parameters of the auxiliary classifier; argmin represents the optimization under the minimum constraint.

6. The elevator production process fault prediction method based on digital twins and transfer learning as described in claim 5, characterized in that, The process of step (3.4) is as follows: Step (3.4.1) The training optimization objective of the transfer learning network model consists of two parts: the traditional loss function and the adaptive loss function. The loss function measures the difference between the predicted value and the true value, while the adaptive loss measures the difference in the domain distribution. in, It is the total cost of the transfer learning network; y s It is the source domain validation set, used to evaluate classification performance; It is the traditional classification loss of transfer learning networks on labeled data, i.e., the source domain; λ is the adaptive loss of the transfer learning network; λ is the penalty factor. Step (3.4.2) uses the cross-entropy function as the conventional classification loss function in the transfer learning process: Where M is the number of fault categories, y * It is a sign function: 1 if the true class of sample i is equal to c, otherwise 0, y * ic p represents the true label of the i-th sample for class c; * ic It is the predicted probability that observed sample i belongs to category c; Step (3.4.3) uses Multi-Kernel Maximum Mean Difference (MK-MMD) to represent the similarity measure of sample features between the source and target domains. It is developed based on MMD. If the features are mapped to the Hilbert space through a kernel function, the distance between the two distributions is measured by the average difference between the samples in the two domains after mapping. The mathematical expression of MMD is as follows: Where MMDX,Y represents the maximum average difference between the source domain X and the target domain Y; X is the set of samples in the source domain; and Y is the set of samples in the target domain. It is a kernel function mapping; It is the norm in Hilbert space; The kernel function used in MK-MMD is a weighted kernel of multiple different kernel functions: Where κ is a multi-kernel function; k μ It is the μ-th basis kernel function; β μ It is the weight of the μ-th kernel function; m is the number of basis kernel functions; In this process, a portion of labeled source domain data and a portion of unlabeled target domain data are used as the training set. Features are extracted using a feature extraction model, and the features are converted into a probability distribution using a feature classifier. Then, the classification loss and distribution difference of the source domain features are calculated using the aforementioned loss function set. The model parameters that minimize the optimization function value are then sought to achieve the classification task of the target domain.

Images