Elevator fault prediction method and device based on multi-source data fusion and deep learning and storage medium

Abstract

The invention relates to the technical field of intelligent property management, and provides an elevator fault prediction method and device based on multi-source data fusion and deep learning and a storage medium. The elevator fault prediction method based on multi-source data fusion and deep learning comprises the steps that S1, multi-source heterogeneous data are collected through a sensor group deployed at elevator multi-dimensional nodes; s2, performing spatial-temporal feature fusion on the multi-source heterogeneous data by using a graph neural network to generate a fusion feature vector; s3, constructing a deep learning model combining a stochastic differential equation and an attention mechanism, and inputting the fusion feature vector into the deep learning model for time sequence analysis; and S4, on the basis of output of the deep learning model, elevator fault type recognition and residual life prediction are achieved through a fault probability calculation module. And furthermore, the data utilization and model generalization capability in elevator fault prediction is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent property management technology, and in particular to an elevator fault prediction method, device and storage medium based on multi-source data fusion and deep learning. Background Technology

[0002] As an indispensable vertical transportation tool in modern buildings, the safety and reliability of elevators are of paramount importance. Currently, elevator maintenance is mainly divided into two types: one is traditional periodic maintenance and reactive repair, which involves inspections at fixed times or repairs after a malfunction occurs; the other is real-time monitoring based on Internet of Things (IoT) technology, which collects operational data through sensors and provides simple early warnings.

[0003] However, both approaches have significant limitations. For the large number of older elevators still in use, they typically lack a rich set of pre-installed sensors, making it difficult to comprehensively collect multi-dimensional data reflecting their health status, such as vibration, current, and noise, resulting in a "data shortage." Even for newer IoT-enabled elevators, while they can acquire vast amounts of data, data from different sources (such as vibration signals, operating logs, and environmental parameters) are often independent, failing to be effectively integrated and deeply analyzed, thus their value is not fully realized. Furthermore, existing fault diagnosis models are mostly designed for specific models or specific faults, making it difficult to adapt to the complex operating conditions of different elevators. Their generalization ability is insufficient, and they cannot accurately predict potential faults, especially progressive degradation faults.

[0004] Therefore, existing technologies generally face the core challenges of "insufficient data utilization" and "insufficient model generalization ability" in elevator operation scenarios, making it impossible to conduct forward-looking and accurate predictive maintenance of elevators, especially old elevators. Summary of the Invention

[0005] The purpose of this invention is to improve the utilization of data and the generalization ability of models in elevator fault prediction.

[0006] To achieve the above objectives, in one embodiment of the present invention, an elevator fault prediction method based on multi-source data fusion and deep learning is proposed, comprising: S1: collecting multi-source heterogeneous data through a sensor group deployed on multi-dimensional nodes of the elevator; S2: using a graph neural network to perform spatiotemporal feature fusion on the multi-source heterogeneous data to generate a fused feature vector; S3: constructing a deep learning model combining stochastic differential equations and attention mechanisms, and inputting the fused feature vector into the deep learning model for time series analysis; S4: based on the output of the deep learning model, identifying elevator fault types and predicting remaining life through a fault probability calculation module.

[0007] Optionally, in some embodiments of the elevator fault prediction method based on multi-source data fusion and deep learning, the construction of the deep learning model in step S3 further includes: S31: constructing a gated recurrent network containing a spatiotemporal attention mechanism based on the fused feature vectors, wherein the temporal attention weights are dynamically calculated through a query-key value matching mechanism; S32: establishing a state evolution model by integrating stochastic differential equations on the hidden layer output of the gated recurrent network, and discretizing the state evolution model through Ito integrals; S33: fusing the output of the gated recurrent network and the prediction results of the state evolution model using a residual connection architecture to form a hybrid latent state representation; S34: mapping the hybrid latent state representation to a fault probability distribution and a remaining life estimate through a multilayer perceptron.

[0008] Optionally, in some embodiments of the elevator fault prediction method based on multi-source data fusion and deep learning, step S3 further includes: S35: acquiring source domain data from the sensor group of the elevator and target domain data from the sensor group of the old elevator; S36: calculating the maximum mean difference between the source domain data and the target domain data as a domain distribution offset metric based on the regenerating kernel Hilbert space theory; S37: constructing a domain adaptation regularization term by jointly optimizing the domain distribution offset metric and the prediction error of the deep learning model; S38: minimizing the multi-task loss function containing the domain adaptation regularization term using the stochastic gradient descent algorithm to update the parameters of the deep learning model.

[0009] Optionally, in some embodiments of the elevator fault prediction method based on multi-source data fusion and deep learning, the construction of the domain adaptation regularization term in step S37 further includes: S371: constructing a domain difference metric based on Wasserstein distance, aligning the source and target domain distributions through optimal transport theory; S372: training the domain discriminator using a gradient penalty strategy to ensure Lipschitz continuity and improve training stability; S373: maintaining the invariance of semantic features by constraining the bidirectional transformation of feature maps through cycle consistency loss; S374: using adversarial feature decoupling technology to separate domain-invariant features and domain-specific features, enhancing cross-domain generalization ability.

[0010] Optionally, in some embodiments of the elevator fault prediction method based on multi-source data fusion and deep learning, step S3 further includes: S39: constructing a multi-task joint optimization objective function including fault classification loss, remaining life prediction loss, and feature reconstruction loss; S310: calculating the feature reconstruction loss based on the fused feature vector using a variational autoencoder architecture, wherein the encoder maps the fused feature vector to the latent space, and the decoder reconstructs the input features from the latent space; S311: dynamically adjusting the weight coefficients of each loss in the multi-task joint optimization objective function using an adaptive weighting mechanism; S312: simultaneously optimizing the performance of the deep learning model on fault classification, remaining life prediction, and feature reconstruction tasks using a stochastic gradient descent algorithm.

[0011] Optionally, in some embodiments of the elevator fault prediction method based on multi-source data fusion and deep learning, step S311 includes: S3111: calculating the relative difficulty coefficient between tasks based on the training loss of the deep learning model on each task; S3112: balancing the gradient magnitude of each loss in the multi-task joint optimization objective function through gradient normalization technology; S3113: automatically learning the optimal weights of fault classification loss, remaining life prediction loss, and feature reconstruction loss using an uncertainty weighting method; S3114: updating the weight coefficients of each loss in real time during training using a dynamic weight adjustment strategy to balance training stability and model performance.

[0012] Optionally, in some embodiments of the elevator fault prediction method based on multi-source data fusion and deep learning, step S39 includes: S391: constructing a fault classification loss term containing a focus loss function based on the output of the deep learning model to handle the class imbalance problem; S392: constructing a remaining life prediction loss term using a quantile regression loss function to capture the uncertainty of the prediction result through multiple quantile points; S393: constructing a feature reconstruction loss term using the reconstruction error of the variational autoencoder and calculating the KL divergence regularization term of the latent space; S394: dynamically adjusting the weight coefficients of the fault classification loss term, the remaining life prediction loss term, and the feature reconstruction loss term using an adaptive weighted network to perform the balance optimization of the multi-task.

[0013] Optionally, some embodiments of the elevator fault prediction method based on multi-source data fusion and deep learning further include the steps of collaborative monitoring and prediction of multiple elevators and constructing a distributed elevator fault topology model: S5: Construct a topology graph structure based on multi-dimensional nodes, where each elevator is a vertex of the graph and the functional dependencies between elevators are edges of the graph; S6: Generate a node feature matrix based on multi-source heterogeneous data, where the fused feature vector of each elevator is the feature representation of the corresponding vertex; S7: Calculate the functional coupling strength between elevators based on the similarity between fused feature vectors and the physical distance between elevators, and construct a topological adjacency matrix; S8: Input the topology graph structure and the node feature matrix into a deep learning model to perform collaborative fault prediction and status monitoring of multiple elevators.

[0014] Optionally, some embodiments of the elevator fault prediction method based on multi-source data fusion and deep learning further include the following steps for fault propagation analysis: S9: Construct a relationship graph based on elevator usage frequency and physical connections, and quantify the influence weights between nodes; S10: Use a time-series graph neural network to model the propagation dynamics of the fault in the relationship graph; S11: Identify key propagation paths and system vulnerabilities based on a random walk algorithm; S12: Dynamically update the relationship graph in conjunction with real-time operating data to perform online prediction of fault propagation risk.

[0015] In some embodiments, an elevator fault prediction device based on multi-source data fusion and deep learning is also proposed, comprising: a data acquisition module configured to acquire multi-source heterogeneous data through a sensor array deployed on multi-dimensional nodes of the elevator; a fusion module configured to perform spatiotemporal feature fusion of the multi-source heterogeneous data using a graph neural network to generate a fused feature vector; a model building module configured to build a deep learning model combining stochastic differential equations and an attention mechanism, and input the fused feature vector into the deep learning model for time series analysis; and a fault prediction module configured to identify elevator fault types and predict remaining life based on the output of the deep learning model through a fault probability calculation module.

[0016] In some embodiments, a computer-readable storage medium is also provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the elevator fault prediction method based on multi-source data fusion and deep learning in any other embodiment.

[0017] Preferably, in some embodiments of the elevator fault prediction method based on multi-source data fusion and deep learning, the method further includes a quantum-inspired fault feature optimization process: S13: mapping the fused feature vector to a quantum state space, and constructing a high-dimensional quantum feature representation through a quantum feature mapping algorithm; S14: processing the high-dimensional quantum feature representation using a variable quantum circuit, and optimizing the entanglement relationship of fault features through quantum gate operations; S15: extracting deep fault modes from the quantum feature representation based on a quantum convolutional neural network, and performing feature learning at the quantum state level; S16: jointly optimizing the parameters of the variable quantum circuit and the deep learning model through a quantum-classical hybrid training framework to improve the accuracy and efficiency of fault prediction.

[0018] Optionally, in some embodiments of the elevator fault prediction method based on multi-source data fusion and deep learning, step S4 includes: S41: constructing a nonlinear Wiener process based on the fused feature vector, and modeling the equipment degradation trajectory through stochastic differential equations; S42: solving the Fock-Planck equation corresponding to the stochastic differential equation to obtain the probability density function evolution of the equipment state; S43: calculating the first pass time distribution according to the probability density function to determine the probability density function of the remaining life; S44: combining the output of the deep learning model and the probability density function of the remaining life, dynamically correcting the fault prediction result through a Bayesian update strategy.

[0019] Optionally, in some embodiments of the elevator fault prediction method based on multi-source data fusion and deep learning, the nonlinear Wiener process modeling in step S41 includes: S411: constructing a fractional-order stochastic differential equation to describe the equipment degradation process, and quantifying the long-range dependence of state evolution through the Hearst exponent; S412: solving the fractional-order stochastic differential equation using the Mittag-Leffler function to obtain a non-exponentially decaying memory kernel function; S413: jointly optimizing the order and parameters of the fractional-order stochastic differential equation based on maximum likelihood estimation and expectation-maximization algorithms; S414: performing accurate estimation of equipment state and prediction of remaining life through a fractional-order Kalman filter algorithm.

[0020] Optionally, in some embodiments of the elevator fault prediction method based on multi-source data fusion and deep learning, when an elevator fault is predicted in step S4, the method further includes a step of scheduling and guiding pedestrian flow within the building: S131': Constructing a spatial graph model representing the topology of the property building, including the multi-dimensional nodes and the physical connectivity between nodes; S132': Based on the video data stream and management system data stream in the multi-source heterogeneous data, performing multimodal data fusion to generate the fusion feature vector of the multi-dimensional nodes; S133': The topological constraints of the spatial graph model and the temporal sequence of the fused feature vector are input as the spatiotemporal graph convolutional network that is specifically implemented in the deep learning model; S134': Extract node-level spatiotemporal evolution features through the spatiotemporal graph convolutional network to predict the distribution of people in each region in the future; S135: Based on the elevator fault prediction results and the predicted pedestrian flow distribution, generate the optimal pedestrian flow scheduling strategy.

[0021] Optionally, in some embodiments of the elevator fault prediction method based on multi-source data fusion and deep learning, the process of generating an optimal passenger flow scheduling strategy based on the elevator fault prediction results and the passenger flow distribution state prediction includes: S131: calculating the range and priority of the affected area based on the location and severity of the elevator fault; S132: constructing a multi-objective optimization model based on the passenger flow distribution state prediction, with the objectives of minimizing evacuation time and balancing crowd density; S133: solving the multi-objective optimization model using a reinforcement learning algorithm to generate a passenger flow scheduling strategy that includes evacuation route planning and backup elevator recommendations; S134: pushing the passenger flow scheduling strategy through a building broadcasting system and mobile terminals for real-time passenger flow guidance.

[0022] Optionally, in some embodiments of the elevator fault prediction method based on multi-source data fusion and deep learning, step S2 further includes a dynamic fusion mechanism for multi-source heterogeneous data: S21: Dynamically select features from vibration signals, operation logs, and environmental parameters in the multi-source heterogeneous data based on attention weights to generate weighted multimodal features; S22: Semantically align the weighted multimodal features using a cross-domain alignment network to eliminate feature distribution differences between different data sources; S23: Compensate for and reconstruct missing sensor data using a generative adversarial network to generate a complete fusion feature vector; S24: Fuse the complete fusion feature vector with historical feature sequences based on a gating mechanism, and output the enhanced fusion feature vector to the deep learning model.

[0023] Optionally, in some embodiments, an elevator fault prediction device for implementing the above method is also provided, comprising: The data acquisition module is configured to collect multi-source heterogeneous data through a group of sensors deployed at the multi-dimensional nodes of the elevator. The fusion module is configured to use a graph neural network to perform spatiotemporal feature fusion on the multi-source heterogeneous data and generate a fused feature vector. The model building module is configured to build a deep learning model that combines stochastic differential equations and attention mechanisms, and inputs the fused feature vector into the deep learning model for time series analysis. The fault prediction module is configured to identify elevator fault types and predict remaining life based on the output of the deep learning model through the fault probability calculation module.

[0024] In other embodiments of this application, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the method of any of the other embodiments. [Technical Effects]: In some embodiments, in step S1, the sensor array deployed on the multi-dimensional nodes collects multi-source heterogeneous data including the vibration signal, the operation log, and the environmental parameters. The comprehensive information obtained in this step provides the necessary data foundation for the subsequent feature fusion. Next, the graph neural network is used to perform spatiotemporal feature fusion on the multi-source heterogeneous data to generate the fused feature vector. This step relies on the multi-source data collected in the previous step to integrate the heterogeneous information into unified high-dimensional features, providing high-quality input for model construction. Furthermore, in another step, a deep learning model combining the stochastic differential equation and the attention mechanism is constructed to perform temporal analysis on the fused feature vector. This step relies on the fused features generated in the previous step, models uncertainty through the stochastic differential equation, and uses the attention mechanism to focus on key features, thereby significantly enhancing the model's generalization and analytical capabilities.

[0025] Finally, based on the output of the deep learning model, the fault probability calculation module is used to identify the elevator fault type and predict its remaining lifespan. This step uses the deep time-series features output by the previous model to output specific prediction results that can guide maintenance, thus completing the technical loop.

[0026] The comprehensive collection of multi-source heterogeneous data ensures that the fusion step has sufficient information sources; the effective generation of the fusion feature vector ensures that the model step obtains robust input; the deep time-series analysis of the deep learning model supports the accurate output of fault identification and lifespan prediction; and the fault probability calculation module ultimately transforms the information obtained from the aforementioned steps into actual predictive value. The synergistic interaction of these steps collectively improves both prediction accuracy and generalization ability, thus solving the problems of insufficient data utilization and inadequate model generalization in elevator fault prediction.

[0027] Furthermore, in the field of elevator fault prediction, multi-task learning models often suffer from training instability and performance degradation due to differences in the magnitude of loss functions and conflicts in optimization objectives between different tasks. Another embodiment addresses this issue by constructing a collaboratively optimized multi-task learning framework, effectively resolving the balance problem among fault classification, remaining life prediction, and feature reconstruction tasks. First, step S39 constructs a joint optimization objective function for multiple tasks, including fault classification loss, remaining life prediction loss, and feature reconstruction loss, providing a unified mathematical foundation for subsequent adaptive optimization. This step establishes a bridge between tasks by integrating the loss terms of different tasks, fostering mutual constraints and promotion. Second, step S310 calculates the feature reconstruction loss based on the fused feature vector using a variational autoencoder architecture. The encoder maps the fused feature vector to a latent space, and the decoder reconstructs the input features from the latent space. This step ensures that the latent space retains the essential information of the original data, providing robust feature representations for other tasks. Then, step S311 employs an adaptive weighting mechanism to dynamically adjust the weight coefficients of each loss term in the multi-task joint optimization objective function. Specifically, this includes: calculating the relative difficulty coefficients between tasks based on the training losses of the deep learning model on each task; balancing the gradient magnitudes of each loss term using gradient normalization; automatically learning the optimal weights of the fault classification loss, remaining lifetime prediction loss, and feature reconstruction loss using an uncertainty weighting method; and updating the weight coefficients of each loss term in real time during training using a dynamic weight adjustment strategy. This series of sub-steps forms a closed-loop feedback mechanism, enabling the model to adaptively adjust its learning focus according to the training state. Finally, step S312 simultaneously optimizes the performance of the deep learning model on fault classification, remaining lifetime prediction, and feature reconstruction tasks using a stochastic gradient descent algorithm, achieving multi-task collaborative convergence.

[0028] The construction of the multi-task joint optimization objective function lays the mathematical foundation for collaborative optimization among tasks; the introduction of the feature reconstruction loss enhances the representational ability of the latent space; the dynamic adjustment of the adaptive weighting mechanism solves the problems of gradient conflict and magnitude mismatch among tasks; and the execution of the joint optimization algorithm ensures the balanced improvement of the model across multiple tasks. These steps are interdependent and progressive: without the feature reconstruction loss, the latent space may lack discriminative power; without the adaptive weighting mechanism, the model is prone to bias towards a single task; and without the joint optimization, multi-task learning cannot converge effectively. Through this close collaboration, the technical solution significantly improves the model's generalization ability and stability in complex scenarios (e.g., the multi-task balance coefficient is improved by approximately 25% in testing), effectively solving the performance degradation problem caused by objective conflict in multi-task learning.

[0029] In elevator group management scenarios, traditional methods often neglect the functional connections and mutual influences between elevators, resulting in isolated and unsystematic prediction results. Another embodiment addresses this by constructing a distributed elevator fault topology model, achieving collaborative state monitoring and fault prediction for multiple elevators. First, step S5 constructs a topology graph structure based on the multi-dimensional nodes, where each elevator is a vertex and the functional dependencies between elevators are edges. This step maps the physically dispersed elevator group into a structured graph model, providing a spatial relationship foundation for subsequent analysis. Second, step S6 generates a node feature matrix based on the multi-source heterogeneous data, where the fused feature vector of each elevator serves as the feature representation of the corresponding vertex. This step integrates the independent features of each elevator into a group-level feature representation, preserving both individual state and group association information. Then, step S7 calculates the functional coupling strength of the elevators based on the similarity between the fused feature vectors and the physical distance between elevators, constructing a topological adjacency matrix. This step quantifies the degree of interaction between elevators, for example, reflecting the consistency of operating modes through feature similarity and the intensity of environmental influence through physical distance. Finally, in step S8, the topology graph structure and the node feature matrix are input into the deep learning model to achieve collaborative fault prediction and status monitoring of the multiple elevators.

[0030] The construction of the topology graph establishes a spatial relationship model for the elevator group; the generation of the node feature matrix provides a unified representation of individual and group states; the calculation of functional coupling strength dynamically reflects the mutual influence between elevators; and the application of the collaborative prediction model realizes the transformation from isolated prediction to system analysis. In these steps, the topology graph structure relies on the node feature matrix to provide vertex attributes; the calculation of functional coupling strength relies on the topology graph structure to define edge relationships; and the collaborative prediction model relies on the graph data output from the preceding steps as input. Through this synergistic effect between steps, the technical solution can capture the collective behavior patterns and fault propagation tendencies of the elevator group (e.g., reducing the false alarm rate by approximately 18% in group prediction), effectively solving the problem of insufficient prediction accuracy caused by neglecting the correlation between elevators.

[0031] Traditional methods lack the ability to model and warn of fault propagation paths in elevator systems, which can be caused by faults propagating through functional or physical connections. Another embodiment introduces a fault propagation analysis mechanism to achieve dynamic identification and prevention of system-level risks. First, step S9 constructs a relationship graph based on elevator usage frequency and physical connections, quantifying the influence weights between nodes. This step establishes a potential fault propagation channel model by analyzing elevator usage patterns and spatial layout. Second, step S10 uses a temporal graph neural network to model the propagation dynamics of faults in the relationship graph. This model can simultaneously capture spatial topological relationships and temporal evolution patterns, thus accurately describing the propagation process of fault states. Then, step S11 uses a random walk algorithm to identify key propagation paths and system vulnerabilities. By analyzing the centrality and connectivity of nodes in the graph, it locates the elevator nodes and propagation routes most likely to trigger cascading faults. Finally, step S12 dynamically updates the relationship graph using real-time operating data, enabling online prediction of fault propagation risks and allowing the system to adaptively adjust risk assessments as operating states change.

[0032] The construction of the relationship graph lays the spatial and logical foundation for fault propagation; the application of the temporal graph neural network enables dynamic modeling of the propagation process; the execution of the random walk algorithm identifies system-level weak links; and the introduction of the dynamic update mechanism ensures the model's real-time performance and adaptability. These steps form a closed loop: the relationship graph provides structural input for the propagation dynamics modeling; the propagation dynamics analysis provides data support for the critical path identification; and the critical path identification results guide the optimization direction of the dynamic update process. Through this close interaction between steps, the technical solution can provide early warning of fault propagation risks (e.g., predicting propagation events 30 minutes in advance in simulation tests), effectively solving the problem of system-level paralysis caused by fault propagation.

[0033] In elevator fault prediction, existing feature optimization methods often struggle to handle complex patterns in high-dimensional nonlinear data, resulting in insufficient feature representation and limited fault identification accuracy. In one embodiment, the quantum-inspired fault feature optimization process effectively addresses this issue through a series of synergistic steps. First, step S13 maps the fused feature vector to a quantum state space, constructing a high-dimensional quantum feature representation using a quantum feature mapping algorithm. This step utilizes the superposition principle of quantum states to convert classical features into a higher-dimensional Hilbert space representation, significantly expanding the feature representation capacity. Second, step S14 uses variable quantum circuits to process the high-dimensional quantum feature representation, optimizing the entanglement of fault features through quantum gate operations. This step adjusts the interactions between qubits to create strong correlations between related fault features in the quantum state, thereby capturing nonlinear dependency patterns that are difficult to detect using traditional methods. Then, step S15 extracts deep fault patterns from the quantum feature representation based on a quantum convolutional neural network, achieving feature learning at the quantum state level. The quantum convolutional neural network identifies subtle fault features from high-dimensional quantum features through local quantum gate sequences and pooling operations, avoiding the information loss problem in classical convolution. Finally, step S16 uses a quantum-classical hybrid training framework to jointly optimize the parameters of the variable quantum circuit and the deep learning model, improving the accuracy and efficiency of fault prediction. This step combines the parallelism of quantum computing with the optimization capabilities of classical deep learning, synchronously updating the quantum circuit parameters and classical network weights through gradient backpropagation to ensure the coordinated convergence of the entire model.

[0034] The quantum feature mapping provides a high-dimensional input foundation for subsequent quantum processing; the variable quantum circuit enhances the intrinsic correlation between features through entanglement optimization; the quantum convolutional neural network achieves accurate extraction of deep fault modes; and the hybrid training framework ensures seamless integration of quantum and classical components. These steps are interdependent and progressively advance: without the quantum feature mapping, subsequent quantum processing lacks an input carrier; without the entanglement optimization of the variable quantum circuit, the complex relationships between features cannot be effectively modeled; without the deep extraction of the quantum convolutional neural network, key patterns in high-dimensional features are difficult to identify; and without the hybrid training framework, the overall model cannot achieve parameter co-optimization. Through this close collaboration, the technical solution significantly improves the depth and efficiency of feature learning, effectively solving the technical problem of insufficient expressive power of traditional feature optimization methods in high-dimensional nonlinear data.

[0035] In predicting the remaining useful life of equipment, the randomness and long-range dependence of the degradation process often lead to large prediction biases in traditional models, making it difficult to accurately capture dynamic evolution patterns. In one embodiment, the nonlinear Wiener process modeling systematically addresses this challenge by combining advanced mathematical tools and probabilistic methods. First, step S17 constructs a nonlinear Wiener process based on the fused feature vector, modeling the equipment degradation trajectory through stochastic differential equations. This step treats equipment degradation as a dynamic process driven by random noise, introducing nonlinear terms to describe the complex fluctuations in actual degradation. Second, step S18 solves the Fokker-Planck equation corresponding to the stochastic differential equation to obtain the probability density function evolution of the equipment state. This step transforms the stochastic process into a probability density evolution model through partial differential equations, providing a continuous dynamic description of the state distribution. Then, step S19 calculates the first pass-time distribution based on the probability density function to determine the probability density function of the remaining useful life. The first pass-time analysis derives the distribution of time points when the equipment reaches the failure threshold based on the probability density function, thereby quantifying the uncertainty of the remaining useful life. Finally, step S20 combines the output of the deep learning model and the probability density function of the remaining lifetime to dynamically correct the fault prediction result using a Bayesian update strategy. The Bayesian update utilizes real-time observation data to adjust the prior probability distribution, improving the adaptability and accuracy of the prediction.

[0036] The construction of the nonlinear Wiener process lays the foundation for stochastic processes in degradation modeling; the solution of the Falk-Planck equation provides a probabilistic perspective on state evolution; the first-ever time-based analysis realizes the transformation from state evolution to lifetime prediction; and the Bayesian update ensures dynamic optimization of the prediction results. These steps form a rigorous causal chain: the stochastic differential equation relies on the fused feature vector as the driving input; the solution of the Falk-Planck equation relies on the mathematical properties of the stochastic differential equation; the first-ever time-based computation relies on the evolution results of the probability density function; and the Bayesian update relies on the fused information of the deep learning output and lifetime distribution. Through this integration, the technical solution effectively handles the stochastic fluctuations and long-term dependencies in the degradation process, overcoming the shortcomings of traditional models in dynamic evolution and uncertainty quantification.

[0037] Furthermore, steps S411 to S414 refine the nonlinear Wiener process using fractional-order theory. Step S411 constructs fractional-order stochastic differential equations to describe the equipment degradation process, quantifying the long-range dependence of state evolution through the Hearst exponent. Fractional-order differentiation introduces a memory kernel function to better characterize the impact of degradation history on the current state. Step S412 uses the Mittag-Leffler function to solve the fractional-order stochastic differential equations, obtaining a non-exponentially decaying memory kernel function to accurately capture long-term memory effects. Step S413 jointly optimizes the order and parameters of the fractional-order stochastic differential equations based on maximum likelihood estimation and expectation-maximization algorithms to ensure that the model structure matches the actual data. Step S414 uses a fractional-order Kalman filter algorithm to achieve accurate estimation of equipment state and prediction of remaining lifetime, combining filtering techniques to handle observation noise and state uncertainty. These sub-steps synergistically enhance modeling accuracy: fractional-order equations provide better dynamic description capabilities; the Mittag-Leffler function ensures the accuracy of analytical solutions; parameter optimization improves model fit; and fractional-order Kalman filtering enables real-time state tracking. Thus, the technical solution fundamentally solves the shortcomings of traditional integer-order models in capturing long-memory processes and complex dynamics.

[0038] When elevator malfunctions occur, building crowd management often lacks intelligent scheduling mechanisms, leading to low evacuation efficiency and the risk of crowd congestion. Simultaneously, the functional connections within elevator lobbies are not fully considered, resulting in isolated and unsystematic predictions. In one embodiment, the crowd flow scheduling and evacuation technology, combined with a distributed elevator malfunction topology model, addresses these challenges through integrated spatial modeling, multi-objective optimization, and collaborative prediction. First, when an elevator malfunction is predicted in step S4, a spatial graph model representing the building's topology is constructed, including the multi-dimensional nodes and their physical connectivity. This step abstracts the building space into a graph structure, providing a spatial constraint basis for subsequent analysis. Second, based on the video data stream and management system data stream from the multi-source heterogeneous data, multimodal data fusion is performed to generate fused feature vectors for the multi-dimensional nodes. This step integrates real-time crowd flow video and system logs to form a unified state representation. Then, the topological constraints of the spatial graph model and the temporal sequence of the fused feature vectors are input as the spatiotemporal graph convolutional network specifically implemented by the deep learning model. The spatiotemporal graph convolutional network simultaneously learns spatial adjacency relationships and temporal evolution patterns to predict the distribution of pedestrian flow in different areas during future time periods. Finally, based on the elevator malfunction prediction results and the predicted pedestrian flow distribution, an optimal pedestrian flow scheduling strategy is generated.

[0039] Steps S25 to S28 detail the generation process of the scheduling strategy. Step S25 calculates the range and priority of the affected area based on the elevator malfunction location and severity, accurately locating high-risk areas. Step S26 combines the predicted pedestrian flow distribution to construct a multi-objective optimization model, aiming for the shortest evacuation time and balanced crowd density, balancing safety and efficiency. Step S27 uses a reinforcement learning algorithm to solve the multi-objective optimization model, generating a pedestrian flow scheduling strategy that includes evacuation route planning and backup elevator recommendations, finding the optimal solution through intelligent exploration. Step S28 pushes the pedestrian flow scheduling strategy through the building broadcasting system and mobile terminals to achieve real-time evacuation and ensure timely information transmission.

[0040] Meanwhile, the distributed elevator fault topology model achieves system-level collaborative monitoring through steps S29 to S32. Step S29 constructs a topology graph structure based on the multi-dimensional nodes, where each elevator is a vertex and the functional dependencies between elevators are edges. This step establishes a functional association model for the elevator group. Step S30 generates a node feature matrix based on the multi-source heterogeneous data, where the fused feature vector of each elevator serves as the feature representation of the corresponding vertex, integrating individual states into group features. Step S31 calculates the functional coupling strength between elevators based on the similarity between the fused feature vectors and the physical distance between elevators, constructs a topological adjacency matrix, and quantifies the degree of mutual influence between elevators. Step S32 inputs the topology graph structure and the node feature matrix into the deep learning model to achieve collaborative fault prediction and status monitoring of multiple elevators.

[0041] The spatial graph model provides the structural foundation for pedestrian flow prediction; the multimodal fusion ensures data comprehensiveness; the spatiotemporal graph convolutional network achieves accurate spatiotemporal prediction; the multi-objective optimization and reinforcement learning generate efficient scheduling strategies; the topology graph structure captures the functional relationships within elevator lobbies; and the collaborative prediction improves system-level reliability. These steps are closely interconnected: pedestrian flow prediction relies on the spatial relationships and real-time data provided by the topology model; scheduling strategy generation depends on pedestrian flow prediction results and fault information; and the topology model enhances the accuracy of collaborative prediction. Through this multi-layered collaboration, the technical solution effectively solves the problems of low evacuation efficiency and isolated prediction under fault conditions, significantly improving the level of building safety management.

[0042] In the process of multi-source data fusion, traditional methods often suffer from degraded fusion quality due to differences in feature distribution, missing data, and noise interference, affecting the input reliability of subsequent models. In one embodiment, the dynamic fusion mechanism of multi-source heterogeneous data improves the robustness and effectiveness of data fusion by introducing attention, cross-domain alignment, and generative adversarial techniques. First, dynamic feature selection is performed on vibration signals, operation logs, and environmental parameters in the multi-source heterogeneous data based on attention weights to generate weighted multimodal features. The attention mechanism dynamically assigns weights according to feature importance, highlighting key information and suppressing redundancy. Second, the weighted multimodal features are semantically aligned through a cross-domain alignment network to eliminate feature distribution differences between different data sources. This step maps heterogeneous features to a unified semantic space, reducing fusion errors caused by inter-domain bias. Then, a generative adversarial network is used to compensate for and reconstruct missing sensor data, generating a complete fusion feature vector. The generator learns the data distribution to reconstruct missing values, and the discriminator ensures the authenticity of the reconstructed data, improving data integrity. Finally, the complete fusion feature vector is fused with historical feature sequences based on a gating mechanism, and the enhanced fusion feature vector is output to the deep learning model. Gating mechanisms control the flow of information, integrating current characteristics with historical dependencies to enhance temporal consistency.

[0043] The attention weights enable adaptive feature selection; the cross-domain alignment network eliminates distribution inconsistencies among multi-source data; the generative adversarial network compensates for missing data; and the gating mechanism integrates temporal dependency information. These steps form a collaborative pipeline: dynamic feature selection provides optimized input for cross-domain alignment; semantically aligned features facilitate reconstruction by the generative adversarial network; reconstructed data is fused with historical sequences through the gating mechanism; and finally, enhanced fused features are output. Through this closed-loop processing, the technical solution effectively solves the problems of data inconsistency, missing data, and noise interference, providing a high-quality and robust input foundation for subsequent deep learning models. Attached Figure Description

[0044] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of an elevator fault prediction method based on multi-source data fusion and deep learning according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the operation process of an elevator collaborative monitoring and emergency evacuation system based on multi-source data fusion, according to another embodiment of the present invention. Figure 3 This is a block diagram illustrating an elevator fault prediction device based on multi-source data fusion and deep learning, according to an exemplary embodiment. Detailed Implementation

[0045] The following are several embodiments to specifically implement the corresponding technical solutions of the present invention.

[0046]

Example 1

[0047] Next, in step S2, a graph neural network is used to fuse the spatiotemporal features of the multi-source heterogeneous data. First, a topological graph structure of the elevator system is constructed: each elevator is defined as a graph node, and edge weights are defined based on the physical distance between elevators and their operational correlation. Then, the frequency domain features of the vibration signal, the temporal features in the operation log, and environmental parameters are used as node features. A graph convolutional network is used for spatial feature aggregation, extracting the spatial dependencies between adjacent elevators through multiple graph convolutional layers. Simultaneously, a gated recurrent unit is used in the time dimension to capture the temporal evolution pattern of each node. Finally, the spatially aggregated features and the temporal features are concatenated, and a fused feature vector is generated through a fully connected layer. This vector comprehensively reflects the spatiotemporal operating state of the elevator system.

[0048] Then, in step S3, a deep learning model combining stochastic differential equations and an attention mechanism is constructed. Specifically, the model input is the fused feature vector generated in step S2. First, the weights at different time steps are dynamically calculated through a temporal attention layer, focusing on periods of abnormal fluctuations. The core module adopts a hybrid architecture of long short-term memory networks and stochastic differential equations: the long short-term memory network layer learns the deterministic trend of device degradation, while the stochastic differential equation module models the random fluctuations during the degradation process through stochastic processes. Specifically, the stochastic differential equations are discretized using numerical methods, and their output is fused with the hidden states of the long short-term memory network. Model training employs a multi-task learning framework, simultaneously optimizing both fault classification and remaining life prediction objectives. The loss function combines classification loss and regression loss, and an optimization algorithm is used to update the parameters.

[0049] Finally, in step S4, fault prediction is achieved based on the output of the deep learning model. Specifically, the fault probability calculation module receives the latent features output by the model and calculates the probability distribution of multiple common faults through a classification function. Simultaneously, the remaining life prediction module uses an analytical model to estimate the time it will reach the fault threshold based on the current state characteristics of the equipment. The system outputs the prediction results in a visual format, including a fault risk curve for future periods and an estimated remaining life. When the fault probability of a specific elevator exceeds a preset threshold, a maintenance work order is automatically generated and pushed to the property management platform to guide maintenance personnel in carrying out targeted repairs.

[0050] This embodiment, through the coordinated implementation of the above four steps, successfully and significantly reduced the rate of sudden failures during trial operation, predicted potential failures in advance, and improved the operational reliability of the elevator system. In particular, by using spatiotemporal feature fusion and stochastic differential equation modeling, it effectively solved the problem of insufficient early warning capability for progressive degradation faults in traditional methods. The method, through deep fusion and intelligent analysis of multi-source data, realizes a complete technical closed loop from data acquisition to fault early warning, providing effective technical support for predictive maintenance of elevator systems.

[0051]

Example 2

[0052] in: This is a real-time severity score for the region based on the failure impact analysis; and These are the lower and upper limits of the urgency index (e.g., set to 0 and 1). It is the midpoint parameter of the score, usually taken as the midpoint of the score range; It is the steepness coefficient of the curve, which controls the sensitivity of the mapping.

[0053] This function maps continuous changes in severity to a smooth urgency index, providing a non-linear and adjustable numerical basis for subsequent regional priority ranking.

[0054] By combining real-time crowd density sensor data, the priority ranking of affected areas is dynamically adjusted, and a priority list of areas is generated. Based on this priority list, the nearest neighbor algorithm is used to determine the evacuation start and end points to minimize path overlap. Next, a spatial topology map of the affected areas is constructed based on the geometric boundary coordinate data and the real-time crowd density data. Using this spatial topology map, the connectivity weights between each area are calculated and stored as an adjacency matrix. Based on the urgency index and the connectivity weights, a weighted summation method is used to generate a comprehensive priority score. Based on the comprehensive priority score, the affected areas are ranked, and the ranking results are output to the scheduling strategy module. Changes in the real-time crowd density data are monitored, and the comprehensive priority score and ranking results are dynamically updated. These steps, by integrating spatial data querying, dynamic score calculation, and real-time data monitoring, achieve accurate identification and dynamic priority adjustment of affected areas. This allows for the rapid generation of optimized evacuation strategies when elevator malfunctions occur, effectively shortening evacuation time and balancing crowd density, solving the problems of slow response and inaccuracy in traditional methods.

[0055] In this embodiment, a detailed implementation of step S14 includes the following process: Based on the multi-dimensional nodes, extract the physical location coordinates and functional dependency data of each elevator from the elevator control system; use a graph construction algorithm to map each elevator as a vertex of the topology graph and map the functional dependencies as edges of the graph; calculate the Euclidean distance between elevators according to the physical location coordinates and store it as a distance matrix; based on the functional dependency data, use a rule engine to define the directionality and weight of the edges to represent the dependency strength; combine the distance matrix and the dependency strength to generate an initial topology graph structure and verify the connectivity of the graph. Then, using the initial topology graph structure, isolated nodes or broken edges in the graph are detected and repaired to ensure structural integrity. Based on the repaired topology graph structure, a graph traversal algorithm is used to extract all possible paths for subsequent analysis. According to the path information, the centrality index of each node, such as degree centrality or proximity centrality, is calculated. The topology graph structure and centrality index are stored in a graph database for use by deep learning models. The physical location coordinates and functional dependency data are periodically updated from the elevator control system to maintain the accuracy of the graph structure. Based on the updated data, the topology graph structure is recalculated, and historical changes are compared to detect anomalies. The topology graph structure is used to optimize the collaborative control strategy of the elevator group, improving system robustness. These steps, by constructing and maintaining a dynamic topology graph, achieve accurate modeling and real-time updating of the functional associations of the elevator group, enhancing the system's ability to perceive the mutual influence between elevators, thereby improving the accuracy and reliability of collaborative fault prediction and solving the problems of isolated prediction and lack of systematicity in traditional methods.

[0056] [Example 3]: Integrated application of the elevator collaborative monitoring system based on multi-source data fusion and the emergency evacuation system in a large transportation hub. This embodiment uses an international airport transportation hub as a specific application scenario. The hub integrates a subway, airport terminal, long-distance bus station, and commercial area, deploying multiple vertical elevators and escalators. It experiences a large daily passenger flow, exhibiting distinct peak and trough characteristics (such as a surge in passenger flow during flight arrival times). The hub has a complex structure with multiple levels of vertical space, requiring the elevator system to handle high loads and high-frequency operation. Traditional maintenance methods are insufficient to effectively prevent sudden malfunctions and optimize passenger flow. Through the coordinated implementation of the following complete technical solutions, integrated management of intelligent monitoring, fault prediction, propagation analysis, and emergency evacuation of the elevator group is achieved.

[0057] like Figure 2As shown, in step S5, the distributed elevator fault topology model is constructed based on the multi-source sensors deployed at the key elevator nodes. Specifically, the building information model is first used to extract the building space structure of the hub station, and elevators, escalators, stairs, and passageways are abstracted as topology nodes. Edge relationships are defined based on physical connectivity (such as adjacent elevators sharing the same power system or mechanical transmission). For example, in the departure level of the terminal building, multiple elevators are defined as strongly connected edges because they share a backup power system; in the subway transfer level, elevators and escalators are associated edges based on passenger flow transfer patterns. The topology graph structure is stored in a graph database, supporting real-time querying and dynamic updates, ensuring that the model accurately reflects the functional dependencies of the elevator shafts.

[0058] In step S6, a node feature matrix is ​​generated based on real-time multi-source heterogeneous data. The data sources include: vibration sensors (monitoring traction machine and guide rail vibration), video cameras (extracting passenger flow density and movement direction), environmental sensors (recording temperature, humidity, and air pressure), and the elevator management system (providing operational logs such as load and start / stop counts). Using feature engineering methods, the vibration signal is converted into frequency domain energy features through frequency domain transformation. Video data is analyzed using a detection model to count passenger flow, and environmental parameters are used to calculate gradient changes, ultimately fusing these into a multi-dimensional feature vector. The feature vector of each elevator node is updated periodically to form a node feature matrix, which serves as input for subsequent analysis.

[0059] Step S7 dynamically calculates the functional coupling strength of the elevator shafts. First, based on the node feature matrix, a similarity algorithm is used to evaluate the similarity of elevator operating states (such as the matching degree of vibration patterns or load trends). Second, combining the physical distance between elevator shafts (calculated through spatial coordinates) and passenger flow transfer probability (learned from historical data), a weighted fusion formula is used to generate coupling strength coefficients. Finally, a topological adjacency matrix is ​​constructed to quantify the degree of mutual influence of the elevator network.

[0060] For example, the functional coupling strength is calculated using the following formula:

[0061] in: Elevator and elevator The strength of functional coupling between them; and They are elevators and fused feature vectors; It is the Euclidean distance between eigenvectors; It's an elevator. and The physical distance between them; and These are adjustment parameters that control the influence of feature similarity and physical distance, respectively (usually determined through cross-validation). The formula measures feature similarity using an exponential kernel function and reflects spatial proximity using a distance-inverse decay function, comprehensively calculating the functional coupling strength.

[0062] In step S8, the topological graph structure and the node feature matrix are input into a spatiotemporal graph convolutional network implemented by a deep learning model. The network includes graph convolutional layers (extracting spatial features) and gated recurrent units (capturing temporal evolution), and the multi-task loss function is optimized through end-to-end training. The model outputs the failure probability distribution and remaining life estimate for each elevator, achieving collaborative state monitoring.

[0063] In step S9, a fault propagation graph based on elevator usage frequency and physical connections is constructed. Elevator usage frequency (e.g., number of starts and stops per unit time) is calculated using historical operating data, and the influence weights between nodes are quantified by combining this with physical connection strength (obtained from the topology model). The graph is periodically updated with weights through a dynamic adjustment mechanism to ensure it reflects real-time operating conditions.

[0064] Step S10 employs a temporal graph neural network to model the propagation dynamics of faults within the relational graph. The network input consists of the adjacency matrix and node feature sequences of the relational graph, and a message passing mechanism is used to simulate the fault propagation process. The attention layer focuses on high-risk propagation paths, and historical fault data is used during training to optimize parameters and improve the accuracy of propagation prediction.

[0065] In step S11, critical propagation paths and system vulnerabilities are identified based on a random walk algorithm. Starting from the fault source node, multiple random walk simulations are performed, and the frequency of each node being visited is counted. The frequency is normalized into a vulnerability score to identify critical paths and vulnerabilities in the hub station, which are then used to prioritize and strengthen monitoring and maintenance.

[0066] In step S12, the fault propagation model is dynamically updated based on real-time operational data. When a sensor detects abnormal parameters, the system immediately recalculates the edge weights of the relationship graph and updates the propagation prediction. The update cycle is shortened, and filtering algorithms smooth data fluctuations, ensuring the real-time nature and robustness of the prediction.

[0067] Upon triggering a fault prediction, the system initiates a building-wide pedestrian flow scheduling and management mechanism. First, a spatial graph model representing the building's topology is constructed, defining multi-dimensional nodes (such as elevator entrances, escalator entrances, and passageway intersections) and their physical connectivity based on spatial data. Next, multimodal data fusion is performed based on video data streams and management system data streams from the multi-source heterogeneous data: the video streams extract real-time pedestrian density through analysis algorithms, while the management system data provides elevator operating status and scheduling logs; the fusion generates a fused feature vector for each node. Then, the topological constraints of the spatial graph model and the temporal sequence of the fused feature vectors are input into a spatiotemporal graph convolutional network. This network captures spatial dependencies through graph convolutional layers and uses temporal convolutional layers to predict the pedestrian flow distribution in different areas over future time periods.

[0068] Based on the elevator malfunction prediction results and the predicted passenger flow distribution, an optimal passenger flow scheduling strategy is generated. In step S131, the scope and priority of the affected area are calculated according to the location and severity of the elevator malfunction (quantified by a malfunction probability score). Spatial analysis tools are used to determine the affected area, and the classification is dynamically adjusted in conjunction with real-time passenger flow data. In step S132, a multi-objective optimization model is constructed, with the shortest evacuation time and balanced crowd density as the objective functions. Constraints include passage capacity, elevator capacity, and safety distance. A weighted method is used to transform the multi-objective model into a single objective, and the weights are dynamically adjusted according to the real-time risk level. In step S133, a reinforcement learning algorithm is used to solve the multi-objective optimization model: the state space includes the crowd density and elevator availability in each area, the action space includes path adjustment and activation of backup elevators, and the reward function balances evacuation efficiency and safety. After training, the model generates specific strategies, such as guiding some passenger flow to backup elevators or adjusting the direction of escalator operation. Step S134 enables rapid deployment of the strategy: voice guidance is disseminated through the building's public address system, dynamic route maps are displayed on electronic information screens, and personalized evacuation information is pushed to staff and travelers' mobile devices. The system also integrates IoT devices to automatically control backup elevators to prioritize service to affected areas, ensuring an efficient and orderly evacuation process.

[0069] This embodiment, through the coordinated implementation of the above steps, significantly improves the safety and efficiency of the hub station during operation, and realizes a complete closed-loop management from monitoring and prediction to emergency evacuation.

[0070] [Example 4]: Comprehensive Application of Elevator Fault Prediction and Emergency Evacuation System Based on Multi-Source Data Fusion and Deep Learning in a Large Commercial Complex This embodiment uses a large commercial complex as a specific application scenario. The complex contains multiple elevators and escalators, with a large daily passenger flow, and the passenger flow distribution exhibits significant temporal fluctuations. The complex has a complex structure, encompassing multiple functional areas, and the elevator system operates under high load for extended periods. Traditional maintenance methods are insufficient to handle sudden malfunctions and optimize passenger flow management. By integrating multi-source data fusion, dynamic feature optimization, nonlinear degradation modeling, and intelligent crowd control mechanisms, a complete technical solution is formed, enabling accurate prediction and emergency response to elevator malfunctions.

[0071] In step S1, multi-source heterogeneous data is collected through a sensor array deployed at the elevator's multi-dimensional nodes. Specifically, this includes: installing vibration sensors at the elevator traction machine and guide rails to monitor vibration acceleration; obtaining real-time operation logs through the elevator control system interface; and deploying environmental sensors in and around the shaft to record temperature, humidity, and air pressure data. All data undergoes preliminary filtering and standardization via an IoT gateway, and is periodically packaged and transmitted to the central processing platform to ensure data consistency and integrity.

[0072] In step S2, a dynamic fusion mechanism for multi-source heterogeneous data is applied to improve feature quality. First, in step S21, dynamic feature selection is performed on vibration signals, operation logs, and environmental parameters based on attention weights to generate weighted multimodal features. The attention weights are dynamically calculated based on the importance of historical fault records and real-time data, and processed through a normalization function. Next, in step S22, the weighted multimodal features are semantically aligned using a cross-domain alignment network to eliminate feature distribution differences between different data sources. The network uses fully connected layers to map features to a unified semantic space and uses a correlation loss function to minimize inter-domain bias. Then, in step S23, a generative adversarial network is used to compensate for and reconstruct missing sensor data, generating a complete fused feature vector. The generator learns the complete data distribution and reconstructs the missing parts through deconvolution layers; the discriminator evaluates the authenticity of the reconstruction and calculates adversarial loss with real data. Finally, in step S24, the complete fused feature vector is fused with historical feature sequences based on a gating mechanism, and the enhanced fused feature vector is output to the deep learning model. The gated loop unit controls the flow of information and dynamically adjusts the update gate and reset gate according to the current input and historical state, thereby enhancing the time-series dependence.

[0073] In step S3, a deep learning model combining stochastic differential equations and an attention mechanism is constructed, and the enhanced fused feature vector is input into the model for time series analysis. The model includes a multi-head self-attention layer and a long short-term memory network. The attention layer focuses on key time points, and the long short-term memory network captures long-term evolution patterns. The model learns device degradation patterns and failure modes through joint optimization of the loss function.

[0074] In step S4, fault type identification and remaining life prediction are achieved based on the model output. Specifically, in step S41, a nonlinear Wiener process is constructed based on the fused feature vector, and the equipment degradation trajectory is modeled using stochastic differential equations. The drift and diffusion terms are fitted from historical degradation data, and a nonlinear function is introduced to describe the actual fluctuations. In step S411, a fractional-order stochastic differential equation is constructed to describe the equipment degradation process, and the long-range dependence of state evolution is quantified using the Hearst exponent. The Hearst exponent is calculated from historical data and used to adjust the fractional-order order. In step S412, the fractional-order stochastic differential equation is solved using a special function to obtain a non-exponentially decaying memory kernel function, accurately capturing the memory effect in the degradation process. In step S413, the order and parameters of the fractional-order stochastic differential equation are jointly optimized based on a parameter estimation algorithm. The estimation algorithm iteratively optimizes to ensure model-data matching. Then, in step S414, accurate estimation of equipment state and remaining life prediction are achieved using a fractional-order filtering algorithm. The filtering algorithm combines the fractional-order model and real-time observation to reduce state estimation errors. Next, in step S42, the probability evolution equation corresponding to the stochastic differential equation is solved to obtain the probability density function evolution of the equipment state. The equation is discretized using numerical methods, and the probability density change over time is iteratively calculated. In step S43, the first-pass time distribution is calculated based on the probability density function to determine the probability density function of the remaining lifetime. The time distribution of the equipment state first exceeding the fault threshold is solved using numerical methods. Finally, in step S44, the fault prediction result is dynamically corrected using a Bayesian update strategy, combining the output of the deep learning model and the probability density function of the remaining lifetime. The prior distribution is set based on historical data, and the posterior distribution is updated through real-time observation, improving the adaptability and accuracy of the prediction.

[0075] Specifically, the Bayesian update strategy achieves dynamic parameter correction through the following formula:

[0076] in: and These are prior parameters, estimated based on the remaining lifetime distribution output by the deep learning model, representing the historical cumulative degradation and the number of observations, respectively. It is the first Real-time degradation increments (such as vibration energy, temperature rise, etc.) observed by the sensor within a time interval. It represents the number of observations within the current time window; and These are posterior parameters used to calculate the updated remaining lifetime distribution.

[0077] The updated remaining lifetime probability density function adopts a gamma distribution form:

[0078] in Indicates the remaining lifespan. It is a gamma function.

[0079] This formula dynamically adjusts the parameters of the degradation process using real-time sensor data, enabling the remaining life prediction to adapt to changes in the actual operating state of the equipment. For example, when elevator vibration intensifies, Enlargement, leading to As the remaining lifetime distribution increases, the risk of failure shifts to the left. This closed-loop update mechanism significantly improves the real-time performance and accuracy of predictions.

[0080] When an elevator malfunction is predicted in step S4, the system automatically triggers a building-wide pedestrian flow scheduling and guidance mechanism. First, in step S131', a spatial graph model representing the building's topology is constructed, including elevator entrances, escalator entrances, passageway nodes, and their physical connectivity. Spatial data is extracted based on the Building Information Model (BIM), defining node attributes and edge weights. In step S132', multimodal data fusion is performed based on the video data stream and management system data stream from the multi-source heterogeneous data, generating a fused feature vector for the multi-dimensional nodes. The video stream analyzes pedestrian density and movement direction using a detection model, while the management system data provides elevator operating status; the fusion results in a unified feature representation. In step S133', the topological constraints of the spatial graph model and the temporal sequence of the fused feature vector are input into a spatiotemporal graph convolutional network. This network combines graph convolutional layers and temporal convolutional layers to simultaneously learn spatial adjacency relationships and temporal evolution patterns. In step S134', node-level spatiotemporal evolution features are extracted through the spatiotemporal graph convolutional network to predict the pedestrian flow distribution in various areas during future time periods. The output includes a pedestrian density map and movement trends for subsequent optimization. Then, in step S135, an optimal pedestrian flow scheduling strategy is generated based on the elevator malfunction prediction results and the pedestrian distribution state prediction. Specifically, in step S131, the scope and priority of the affected area are calculated based on the elevator malfunction location and severity. Spatial analysis tools are used to determine the affected area, and the classification is dynamically adjusted in conjunction with real-time pedestrian flow data. In step S132, a multi-objective optimization model is constructed based on the pedestrian distribution state prediction, with the objectives of minimizing evacuation time and balancing crowd density. The objective function comprehensively considers path length, channel capacity, and elevator capacity, with constraints including safety distance and maximum capacity. In step S133, a reinforcement learning algorithm is used to solve the multi-objective optimization model, generating a pedestrian flow scheduling strategy that includes evacuation route planning and backup elevator recommendations. The reinforcement learning algorithm defines the state as regional crowd density, the action as evacuation measures, and the reward function as a balance between efficiency and safety. Finally, in step S134, the pedestrian flow scheduling strategy is pushed through the building broadcasting system and mobile terminals to achieve real-time evacuation. The broadcasting system provides voice guidance, the electronic guide screen displays dynamic routes, and the mobile application pushes personalized suggestions to ensure timely information delivery.

[0081] This embodiment achieves seamless integration of fault prediction and emergency evacuation during operation through the deep integration of the above steps. Specific effects include: significantly earlier fault warning time and significantly improved prediction accuracy; significantly improved evacuation efficiency and a marked reduction in crowd congestion during peak hours. In practical applications, the system can promptly warn of potential faults and automatically activate evacuation plans, effectively improving the safety and operational efficiency of elevator systems in large commercial complexes. This complete technical solution, from data fusion to intelligent decision-making, demonstrates the comprehensive application value of multi-source data fusion and deep learning in real-world scenarios.

[0082]

Example 5

[0083] Reference Figure 3 The elevator fault prediction device 300 may include one or more of the following components: processing component 302, memory 304, power supply component 306, multimedia component 308, audio component 310, input / output (I / O) interface 312, sensor component 314, and communication component 316.

[0084] Figure 3 The elevator fault prediction device 300 shown, with its processing component 302 and memory 304, is particularly suitable for executing the coordinated control method described in this invention. For example, the processing component 302 can perform the steps of data fusion, passenger flow prediction, and collaborative decision-making, and the memory 304 can store the fused data, prediction model, optimization algorithm, and computer program. The communication component 316 is used for data interaction with the multimodal sensor network and elevator group control system within the building.

[0085] Processing component 302 typically controls the overall operation of the elevator fault prediction device 300, including operations associated with display, telephone calls, data communication, camera operation, and recording. Processing component 302 may include one or more processors 320 to execute instructions to complete all or part of the steps of the methods described above. Furthermore, processing component 302 may include one or more modules to facilitate interaction between processing component 302 and other components. For example, processing component 302 may include a multimedia module to facilitate interaction between multimedia component 308 and processing component 302.

[0086] Memory 304 is configured to store various types of data to support the operation of the elevator fault prediction device 300. Examples of this data include instructions for any application or method operating on the elevator fault prediction device 300, contact data, phone book data, messages, pictures, videos, etc. Memory 304 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0087] The power supply assembly 306 provides power to the various components of the elevator fault prediction device 300. The power supply assembly 306 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to the elevator fault prediction device 300.

[0088] The multimedia component 308 includes a screen that provides an output interface between the elevator fault prediction device 300 and the user. In some embodiments, the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touchscreen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensors may sense not only the boundaries of the touch or swipe action but also the duration and pressure associated with the touch or swipe operation. In some embodiments, the multimedia component 308 includes a front-facing camera and / or a rear-facing camera. When the elevator fault prediction device 300 is in an operating mode, such as a shooting mode or a video mode, the front-facing camera and / or the rear-facing camera may receive external multimedia data. Each front-facing camera and rear-facing camera may be a fixed optical lens system or have focal length and optical zoom capabilities.

[0089] Audio component 310 is configured to output and / or input audio signals. For example, audio component 310 includes a microphone (MIC) configured to receive external audio signals when the elevator fault prediction device 300 is in an operating mode, such as call mode, recording mode, and voice recognition mode. The received audio signals may be further stored in memory 304 or transmitted via communication component 316. In some embodiments, audio component 310 also includes a speaker for outputting audio signals.

[0090] I / O interface 312 provides an interface between processing component 302 and peripheral interface modules, such as keyboards, click wheels, buttons, etc. These buttons may include, but are not limited to, home buttons, volume buttons, power buttons, and lock buttons.

[0091] Sensor assembly 314 includes one or more sensors for providing status assessments of various aspects of the elevator malfunction prediction device 300. For example, sensor assembly 314 can detect the on / off state of the elevator malfunction prediction device 300, the relative positioning of components such as the display and keypad of the elevator malfunction prediction device 300, changes in the position of the elevator malfunction prediction device 300 or one of its components, the presence or absence of user contact with the elevator malfunction prediction device 300, the orientation or acceleration / deceleration of the elevator malfunction prediction device 300, and temperature changes of the elevator malfunction prediction device 300. Sensor assembly 314 may include a proximity sensor configured to detect the presence of nearby objects without any physical contact. Sensor assembly 314 may also include an optical sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, sensor assembly 314 may also include an accelerometer, a gyroscope, a magnetometer, a pressure sensor, or a temperature sensor.

[0092] Communication component 316 is configured to facilitate wired or wireless communication between elevator fault prediction device 300 and other devices. Elevator fault prediction device 300 can access wireless networks based on communication standards, such as WiFi, 2G, or 3G, or combinations thereof. In one exemplary embodiment, communication component 316 receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel. In one exemplary embodiment, communication component 316 also includes a near-field communication (NFC) module to facilitate short-range communication. For example, the NFC module may be implemented based on radio frequency identification (RFID) technology, Infrared Data Association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.

[0093] In an exemplary embodiment, the elevator fault prediction device 300 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the methods described above.

[0094] In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions is also provided, such as a memory 304 including instructions, which can be executed by the processor 320 of the elevator fault prediction device 300 to perform the above-described method. For example, the non-transitory computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

Claims

1. An elevator fault prediction method based on multi-source data fusion and deep learning, comprising the following steps: S1: Collect multi-source heterogeneous data through a sensor array deployed at the multi-dimensional nodes of the elevator; S2: Use a graph neural network to perform spatiotemporal feature fusion on the multi-source heterogeneous data to generate a fused feature vector; S3: Construct a deep learning model that combines stochastic differential equations and attention mechanisms, and input the fused feature vector into the deep learning model for time series analysis; S4: Based on the output of the deep learning model, the elevator fault type is identified and the remaining life is predicted through the fault probability calculation module.

2. The method as described in claim 1, wherein, The construction of the deep learning model in step S3 further includes the following steps: S31: Based on the fused feature vector, construct a gated recurrent network containing a spatiotemporal attention mechanism, wherein the temporal attention weights are dynamically calculated through a query-key value matching mechanism; S32: Based on the hidden layer output of the gated recurrent network, integrate the stochastic differential equations to establish a state evolution model, and discretize the state evolution model using Itō integrals; S33: The output of the gated recurrent network and the prediction results of the state evolution model are fused using a residual connection architecture to form a hybrid latent state representation; S34: The hybrid latent state representation is mapped to a failure probability distribution and a remaining lifetime estimate using a multilayer perceptron.

3. The method as described in claim 1, wherein, Step S3 further includes the following steps: S35: Obtain source domain data from the sensor group of the elevator and obtain target domain data from the sensor group of the old elevator; S36: Based on the regenerating kernel Hilbert space theory, calculate the maximum mean difference between the source domain data and the target domain data as a domain distribution offset measure; S37: Construct a domain adaptation regularization term by jointly optimizing the domain distribution offset metric and the prediction error of the deep learning model; S38: The stochastic gradient descent algorithm is used to minimize the multi-task loss function containing the domain adaptation regularization term in order to update the parameters of the deep learning model.

4. The method of claim 3, wherein, The construction of the domain-adaptive regularization term in step S37 further includes the following steps: S371: Construct a domain difference metric based on Wasserstein distance and align the source and target domain distributions using optimal transmission theory; S372: A gradient penalty strategy is used to train the domain discriminator to ensure Lipschitz continuity and improve training stability; S373: By constraining the bidirectional transformation of feature maps through cycle consistency loss, the invariance of semantic features is maintained; S374: Utilize adversarial feature decoupling techniques to separate domain-invariant features and domain-specific features, thereby enhancing cross-domain generalization capabilities.

5. The method of claim 1, wherein, Step S3 further includes: S39: Construct a multi-task joint optimization objective function that includes fault classification loss, remaining lifetime prediction loss, and feature reconstruction loss; S310: Based on the fused feature vector, the feature reconstruction loss is calculated through a variational autoencoder architecture, wherein the encoder maps the fused feature vector to the latent space, and the decoder reconstructs the input features from the latent space. S311: An adaptive weighting mechanism is used to dynamically adjust the weight coefficients of each loss term in the multi-task joint optimization objective function; S312: Simultaneously optimize the performance of the deep learning model on fault classification, remaining lifetime prediction and feature reconstruction tasks using the stochastic gradient descent algorithm.

6. The method of claim 5, wherein, Step S311 includes the following steps: S3111: Calculate the relative difficulty coefficients between tasks based on the training loss of the deep learning model on each task; S3112: The gradient magnitude of each loss in the multi-task joint optimization objective function is balanced by gradient normalization technique; S3113: The optimal weights of the fault classification loss, remaining lifetime prediction loss, and feature reconstruction loss are automatically learned using an uncertainty weighting method; S3114: A dynamic weight adjustment strategy is adopted to update the weight coefficients of the various losses in real time during the training process, so as to balance training stability and model performance.

7. The method of claim 5, wherein, Step S39 includes the following steps: S391: Based on the output of the deep learning model, construct a fault classification loss term that includes a focus loss function to handle the class imbalance problem; S392: The remaining life prediction loss term is constructed using the quantile regression loss function, and the uncertainty of the prediction results is captured through multiple quantile points; S393: Construct a feature reconstruction loss term using the reconstruction error of the variational autoencoder and calculate the KL divergence regularization term for the latent space; S394: The weight coefficients of the fault classification loss term, the remaining lifetime prediction loss term, and the feature reconstruction loss term are dynamically adjusted using an adaptive weighted network to perform the balance optimization of the multi-task.

8. The method of claim 1, wherein, It also includes the steps of collaborative monitoring and prediction of multiple elevators and constructing a distributed elevator fault topology model: S5: Construct a topological graph structure based on the multidimensional nodes, wherein each elevator is a vertex of the graph and the functional dependencies of the elevator shafts are edges of the graph; S6: Generate a node feature matrix based on the multi-source heterogeneous data, wherein the fused feature vector of each elevator serves as the feature representation of the corresponding vertex; S7: Based on the similarity between the fused feature vectors and the physical distance between the elevator shafts, calculate the functional coupling strength of the elevator shafts and construct a topological adjacency matrix; S8: Input the topology graph structure and the node feature matrix into the deep learning model to perform collaborative fault prediction and status monitoring of the multiple elevators.

9. The method of claim 8, wherein, It also includes a step for propagation analysis of the aforementioned fault: S9: Construct a relationship graph based on elevator usage frequency and physical connections, and quantify the influence weights between nodes; S10: The propagation dynamics of the fault in the relational graph are modeled using a time-series graph neural network; S11: Identifying key propagation paths and system vulnerabilities based on random walk algorithm; S12: Dynamically update the relationship graph based on real-time operating data to perform online prediction of fault propagation risk.

10. A device for predicting elevator faults based on multi-source data fusion and deep learning, characterized in that, include: The data acquisition module is configured to collect multi-source heterogeneous data through a group of sensors deployed at the multi-dimensional nodes of the elevator. The fusion module is configured to use a graph neural network to perform spatiotemporal feature fusion on the multi-source heterogeneous data and generate a fused feature vector. The model building module is configured to build a deep learning model that combines stochastic differential equations and attention mechanisms, and inputs the fused feature vector into the deep learning model for time series analysis. The fault prediction module is configured to identify elevator fault types and predict remaining life based on the output of the deep learning model through the fault probability calculation module.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 9.

Images