Elevator emergency rescue decision-making method and system based on reinforcement learning

Abstract

The present invention provides an elevator emergency rescue decision-making method and system based on reinforcement learning, which relates to the technical field of elevator emergency rescue, including real-time collection of elevator multimodal data through a distributed sensor network to generate an elevator state representation vector; inputting the obtained elevator state representation vector into a pre-trained graph convolutional neural network model, and simultaneously starting an edge computing unit for real-time simulation, simulating system responses under different rescue strategies, evaluating the feasibility of each rescue strategy, weightedly fusing the multi-dimensional risk score and the rescue strategy feasibility evaluation results, and generating a risk assessment report; sending a rescue instruction set to an elevator emergency rescue executive agency through a secure encrypted channel, and continuously monitoring the real-time feedback information of the elevator emergency rescue executive agency, and triggering a decision re-evaluation mechanism when an abnormal situation occurs, until a passenger safe evacuation signal is received and the elevator system restores a safe state signal, and the rescue mission is completed.

Description

Technical Field

[0001] The present invention relates to elevator emergency rescue technology, and in particular to an elevator emergency rescue decision-making method and system based on reinforcement learning. Background Art

[0002] As an indispensable vertical transportation tool in modern buildings, the safety and reliability of elevators are directly related to the safety of passengers' lives and property. Traditional elevator emergency rescue methods mainly rely on manual experience judgment and preset emergency plans, which have the following problems: single sensor data is difficult to fully reflect the complex state of the elevator system, resulting in deviations in fault diagnosis; it fails to fully consider multi-dimensional risk factors such as fault type, environmental factors and passenger conditions; preset emergency plans are difficult to adapt to complex and changeable fault scenarios, and the rescue efficiency is low; over-reliance on manual experience is easily affected by the subjective judgment and pressure of emergency personnel, increasing decision-making risks.

[0003] To solve the above problems, it is urgent to develop an intelligent elevator emergency rescue decision-making method based on reinforcement learning to achieve multimodal data fusion, intelligent fault diagnosis, multi-dimensional risk assessment and dynamic rescue strategy optimization, so as to improve the intelligence level and efficiency of elevator emergency rescue. Summary of the invention

[0004] The embodiments of the present invention provide an elevator emergency rescue decision-making method and system based on reinforcement learning, which can solve the problems in the prior art.

[0005] According to a first aspect of the embodiments of the present invention, A reinforcement learning-based elevator emergency rescue decision-making method is provided, comprising: Collect elevator multimodal data in real time through a distributed sensor network, including mechanical vibration signals, electrical system parameters, acoustic characteristics, visual images, and environmental indicators; preprocess the collected elevator multimodal data, perform feature fusion on the preprocessed elevator multimodal data through an attention mechanism, and generate an elevator state representation vector; The obtained elevator state representation vector is input into the pre-trained graph convolutional neural network model, which integrates the elevator system knowledge graph, outputs fault type and fault location information, combines fault severity, environmental conditions and passenger conditions for risk assessment, generates a multi-dimensional risk score, and starts the edge computing unit for real-time simulation to evaluate the feasibility of each rescue strategy. The multi-dimensional risk score and rescue strategy feasibility assessment results are weighted and fused to generate a risk assessment report. Based on the risk assessment report, the preset rescue strategy knowledge base is activated, which contains different rescue plan templates. The rescue plan templates are dynamically optimized using optimization algorithms to generate an initial rescue instruction sequence and simulate and predict the results of different decision paths. The optimal rescue decision path is selected, and a rescue instruction set is generated based on the optimal rescue decision path. The rescue instruction set is sent to the elevator emergency rescue executive agency, and the real-time feedback information of the elevator emergency rescue executive agency is continuously monitored. When an abnormality occurs, it is re-evaluated until the passengers are safely evacuated and the system returns to a safe state.

[0006] In an optional embodiment, The obtained elevator state representation vector is input into the pre-trained graph convolutional neural network model, which integrates the elevator system knowledge graph and outputs the fault type and fault location information including: Construct a multi-level elevator system knowledge graph that includes topological connections and causal dependencies between elevator components. The causal dependencies are represented by bidirectional directed edges, and the weights of the bidirectional directed edges represent the causal strength. The causal graph neural network is used to encode the constructed multi-level elevator system knowledge graph to obtain the causal representation of the nodes in the knowledge graph. Design a hybrid graph neural network model that includes a spatial-temporal graph convolution layer and a multi-scale attention mechanism, integrate the hybrid graph neural network model with the multi-level elevator system knowledge graph, input the elevator state representation vector into the hybrid graph neural network model, and process the input elevator state representation vector through multi-level information transmission and aggregation of the hybrid graph neural network model to generate multi-scale fault feature representation; Design a causal reasoning module based on the attention mechanism, use multi-scale fault feature representation to infer the root cause and propagation path of the fault, and obtain the causal explanation vector of the fault; The causal explanation vector of the fault is integrated with the multi-scale fault feature representation to generate a comprehensive fault feature. A multi-task learning framework with an adaptive weight allocation mechanism is designed to simultaneously classify the fault type and locate the fault location based on the comprehensive fault feature. At the same time, fault correlation constraints based on the multi-level elevator system knowledge graph structure are introduced, and preliminary fault type classification results and fault location location results are obtained using the multi-task learning framework. Multiple hybrid graph neural network models with different structures and parameters are constructed, and the same comprehensive fault features are used as input. The output results of multiple hybrid graph neural network models and multi-task learning frameworks are weightedly fused to generate the final fault type classification results and fault location positioning results.

[0007] In an optional embodiment, A multi-task learning framework with an adaptive weight allocation mechanism is designed to simultaneously classify fault types and locate fault locations based on comprehensive fault features. At the same time, fault correlation constraints based on the multi-level elevator system knowledge graph structure are introduced. The preliminary fault type classification results and fault location location results are obtained using the multi-task learning framework, including: Design a shared feature extraction network, which includes a residual connection and a hole convolution structure, input the comprehensive fault features into the shared feature extraction network, extract multi-scale features and generate a feature vector of fixed dimension, input the feature vector into the fault type classification task output layer and the fault location task output layer respectively, and output the initial fault type classification result and the initial fault location result respectively, wherein the fault type classification task output layer uses a fully connected layer, and the fault location task output layer uses a multi-layer perceptron structure and an attention mechanism; The fault type classification loss and the fault location loss are calculated based on the initial fault type classification results and the initial fault location location results, and a dynamic weight adjustment strategy based on task uncertainty is introduced to assign adaptive weights to each task. Based on the pre-built multi-level elevator system knowledge graph, a fault type correlation matrix is ​​generated, and the knowledge graph constraint loss is constructed using the fault type correlation matrix; Extracting intermediate features of the initial fault type classification result and the initial fault position location result, performing information interaction on the intermediate features to obtain interactive features, and updating the fault type classification result and the fault position location result based on the interactive features; Add random disturbances that obey Gaussian distribution to the comprehensive fault features to obtain the disturbed features, input the disturbed features into the shared feature extraction network, calculate the adversarial loss of the disturbed features, and combine the knowledge graph constraint loss, classification loss and positioning loss, as well as the adversarial loss for weighting to obtain the total loss; An alternating optimization strategy is adopted to optimize the model parameters and task weights, and a learning rate decay strategy and an early stopping mechanism are introduced during the optimization process. The iterations are repeated until the preset termination conditions are reached. The optimized model parameters and task weights are used to perform fault diagnosis on the new elevator operation data, and the final fault type classification results and fault location positioning results are output.

[0008] In an optional embodiment, The risk assessment is conducted based on the severity of the fault, environmental conditions, and passenger conditions to generate a multi-dimensional risk score. At the same time, the edge computing unit is started for real-time simulation to evaluate the feasibility of each rescue strategy. The multi-dimensional risk score and the rescue strategy feasibility assessment results are weighted and integrated to generate a risk assessment report including: Construct a risk assessment model based on multi-agent deep reinforcement learning. Each agent corresponds to a risk dimension, forming a multi-dimensional risk assessment system. The attention mechanism is used to process the information of fault type, fault location, fault severity, environmental conditions and passenger conditions to generate context feature vectors. Inputting the context feature vector into a pre-trained risk assessment model, and using a hierarchical soft actor-critic algorithm to generate a multi-dimensional dynamic risk score; Deploy an adaptive rescue strategy simulator based on a proximal strategy optimization algorithm on a distributed edge computing network, wherein the adaptive rescue strategy simulator uses a high-precision elevator physics engine that considers material fatigue and environmental factors, and the high-precision elevator physics engine performs simulation based on the dynamic mechanical characteristics of the elevator, the nonlinear electrical system response, the multi-layer elevator group control logic, and the group passenger behavior model; Using the adaptive rescue strategy simulator to simulate system responses under different rescue strategy combinations, and obtaining feasibility evaluation results and expected execution time of each rescue strategy combination; A weighted fusion network based on deep ensemble learning is constructed, and the multi-dimensional dynamic risk score and the feasibility assessment results of each rescue strategy combination are taken as input. The random dropout technology is applied in the weighted fusion network to obtain the uncertainty quantification result of the fusion process. The loss function is calculated according to the uncertainty quantification result, the weighted fusion network is trained, and the risk score report is output using the trained weighted fusion network.

[0009] In an optional embodiment, The formula for calculating the loss function based on the uncertainty quantification result is as follows: ; Among them, L represents the loss function, N represents the number of samples, and w i represents the weight of the i-th sample, w j represents the weight of the jth sample, x i represents the feature vector of the i-th sample, x j represents the feature vector of the jth sample, σ i represents the uncertainty measure of the i-th sample, and λ represents the regularization coefficient.

[0010] In an optional embodiment, Based on the risk assessment report, the preset rescue strategy knowledge base is activated, which contains different rescue plan templates. The rescue plan templates are dynamically optimized using the optimization algorithm to generate the initial rescue instruction sequence and simulate and predict the results of different decision paths. The optimal rescue decision path is selected, including: Perform semantic analysis based on the generated risk assessment report, match the preset rescue strategy knowledge base, and obtain the corresponding rescue plan template. The rescue strategy knowledge base is built based on the ontology and contains rescue plan templates for different fault scenarios as well as performance parameters and applicable conditions of rescue resources; The rescue plan template is dynamically optimized by using an optimization algorithm, three optimization objectives of rescue efficiency, resource consumption and secondary risk are constructed, the rescue resource deployment plan and rescue path planning are adaptively adjusted, and an initial rescue instruction sequence is generated, wherein the initial rescue instruction sequence includes the optimized rescue resource deployment plan and rescue path planning; Based on the initial rescue instruction sequence, a three-dimensional simulation model of the elevator system is constructed to simulate the rescue process under different decision paths, predict the rescue effect of each decision path, and statistically analyze the rescue indicators of each decision path to generate a decision path evaluation report, wherein the decision path evaluation report includes the rescue effect prediction results and rescue indicator statistical results of each decision path; The multi-attribute decision-making method is used to comprehensively evaluate the various rescue indicators in the decision path evaluation report, and the optimal rescue decision path is selected based on expert experience.

[0011] In an optional embodiment, The multi-attribute decision-making method is used to comprehensively evaluate the various rescue indicators in the decision path evaluation report, and the optimal rescue decision path is selected in combination with expert experience, including: Constructing an interval-type fuzzy decision matrix, wherein the interval-type fuzzy decision matrix includes the performance of multiple candidate rescue decision paths on multiple evaluation attributes, and the performance of each candidate rescue decision path on each evaluation attribute is represented by a membership interval and a non-membership interval; Determine an attribute weight vector, the attribute weight vector is used to represent the importance of each evaluation attribute, and use a pre-calculated mean operator to fuse the interval-type fuzzy decision matrix, and based on the fusion result, calculate the comprehensive utility interval of each alternative rescue decision path, the comprehensive utility interval consists of a membership interval and a non-membership interval; A ranking method based on possibility is used to compare the sizes of the comprehensive utility intervals of various alternative rescue decision paths, and the possibility between each pair of comprehensive utility intervals is calculated. The possibility is calculated by the upper and lower bounds of the membership interval of each comprehensive utility interval; The candidate rescue decision paths are sorted according to the calculation result of the possibility, and the candidate rescue decision path with the largest possibility is selected as the optimal rescue path.

[0012] According to a second aspect of the embodiments of the present invention, Provided is an elevator emergency rescue decision system based on reinforcement learning, comprising: The first unit is used to collect elevator multimodal data in real time through a distributed sensor network, including mechanical vibration signals, electrical system parameters, acoustic characteristics, visual images and environmental indicators; preprocess the collected elevator multimodal data, perform feature fusion on the preprocessed elevator multimodal data through an attention mechanism, and generate an elevator state representation vector; The second unit is used to input the obtained elevator state representation vector into a pre-trained graph convolutional neural network model, which integrates the elevator system knowledge graph, outputs fault type and fault location information, conducts risk assessment based on fault severity, environmental conditions, and passenger conditions, generates a multi-dimensional risk score, and simultaneously starts the edge computing unit for real-time simulation to evaluate the feasibility of each rescue strategy. The multi-dimensional risk score and the rescue strategy feasibility assessment results are weighted and fused to generate a risk assessment report; The third unit is used to activate the preset rescue strategy knowledge base based on the risk assessment report, which contains different rescue plan templates. It uses the optimization algorithm to dynamically optimize the rescue plan template, generate the initial rescue instruction sequence and simulate and predict the results of different decision paths, select the optimal rescue decision path, generate a rescue instruction set based on the optimal rescue decision path, and send it to the elevator emergency rescue execution agency. It continuously monitors the real-time feedback information of the elevator emergency rescue execution agency, and re-evaluates when an abnormality occurs until the passengers are safely evacuated and the system returns to a safe state.

[0013] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0014] According to a fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0015] In this embodiment, real-time collection and fusion of elevator multimodal data are achieved to improve the accuracy and comprehensiveness of fault diagnosis. The accuracy of fault identification and location is improved by combining graph convolutional neural networks and knowledge graphs. Multi-dimensional risk assessment is achieved, considering the severity of the fault, environmental conditions and passenger conditions, providing a more comprehensive risk assessment, using edge computing for real-time simulation, and optimizing the selection of rescue strategies. Dynamically optimize rescue plans to adapt to different fault scenarios and improve rescue efficiency. Real-time monitoring and dynamic adjustment of the rescue process are achieved to ensure the effectiveness and safety of rescue, improve the intelligence and automation level of elevator emergency rescue, and reduce human decision-making errors. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flow chart of an elevator emergency rescue decision-making method based on reinforcement learning according to an embodiment of the present invention; Figure 2 It is a structural schematic diagram of an elevator emergency rescue decision-making system based on reinforcement learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0019] Figure 1 FIG. 1 is a flow chart of an elevator emergency rescue decision-making method based on reinforcement learning according to an embodiment of the present invention. Figure 1 As shown, the method includes: S101. Collect elevator multimodal data in real time through a distributed sensor network, including mechanical vibration signals, electrical system parameters, acoustic characteristics, visual images and environmental indicators; preprocess the collected elevator multimodal data, perform feature fusion on the preprocessed elevator multimodal data through an attention mechanism, and generate an elevator state representation vector.

[0020] Exemplarily, first, a distributed sensor network is deployed in the elevator system. The network includes multiple types of sensors, which are used to collect data in different modes. Mechanical vibration sensors are installed in key parts such as the elevator car, counterweight and traction machine to collect vibration signals; electrical system sensors are connected to the elevator control cabinet and drive system to collect parameters such as voltage, current, and power; acoustic sensors are arranged inside and outside the car to collect sound characteristics during the operation of the elevator; visual sensors include high-definition cameras inside the car and in the shaft to collect real-time images; environmental sensors measure indicators such as temperature, humidity, and carbon dioxide concentration in the car.

[0021] These sensors are connected to edge computing nodes via wired or wireless means to form a complete data collection network. The edge computing nodes are responsible for preliminary data processing and temporary storage, while transmitting data to the central processing server through a secure encrypted channel.

[0022] Next, the collected multimodal data is preprocessed. For mechanical vibration signals, the wavelet denoising algorithm is used to remove high-frequency noise, and then the time domain and frequency domain features are extracted; for electrical system parameters, normalization is performed and short-term statistical features are calculated; acoustic features are extracted through Mel-frequency cepstral coefficients (MFCC); visual images use convolutional neural networks to extract deep features; and environmental indicators are standardized.

[0023] The preprocessed multimodal data has different dimensions and scales, and feature fusion is required. This method adopts a multimodal feature fusion algorithm based on the attention mechanism. First, an attention weight is assigned to the data of each modality, which reflects the importance of the modal data to the current elevator state judgment. The attention weight is calculated by a trainable neural network, which inputs the feature vector of each modality and outputs the corresponding weight coefficient.

[0024] Then, each modal feature is multiplied by its corresponding attention weight to obtain a weighted feature. All weighted features are connected into a high-dimensional vector, which is then mapped to a fixed-dimensional space through a fully connected layer to finally generate an elevator state representation vector.

[0025] This state representation vector integrates information from various modes and can fully reflect the operating status of the elevator. It will serve as input for subsequent fault diagnosis and risk assessment, and provide a basis for intelligent rescue decisions. In this way, effective fusion of multi-source heterogeneous data is achieved, improving the accuracy and robustness of state perception.

[0026] S102. Input the obtained elevator state representation vector into a pre-trained graph convolutional neural network model, which integrates the elevator system knowledge graph, outputs fault type and fault location information, conducts risk assessment based on fault severity, environmental conditions, and passenger conditions, generates a multi-dimensional risk score, and simultaneously starts the edge computing unit for real-time simulation to evaluate the feasibility of each rescue strategy. The multi-dimensional risk score and rescue strategy feasibility assessment results are weighted and fused to generate a risk assessment report.

[0027] First, the obtained elevator state representation vector is input into the pre-trained graph convolutional neural network model. The core of this model is the integration of the elevator system knowledge graph, which contains information such as the relationship between the various components of the elevator, possible fault types, and fault propagation paths. The nodes in the knowledge graph represent the various components of the elevator and possible fault types, and the edges represent the connection relationship between the components or the causal relationship of the fault.

[0028] During the training process, the graph convolutional neural network model learned how to use the structural information in the knowledge graph to enhance the accuracy of fault diagnosis. When the elevator state representation vector is input, the model first maps this vector to the node of the knowledge graph, and then gradually aggregates the information of neighboring nodes through multi-layer graph convolution operations, and finally gives the prediction results of fault type and fault location at the output layer.

[0029] For example, suppose the input elevator state representation vector shows that the motor current is abnormally high, the temperature is high, and there is abnormal vibration. Based on these features and the relationship information in the knowledge graph, the graph convolutional neural network model will determine that it may be a motor bearing failure and locate the specific bearing position.

[0030] Next, based on the output fault type and fault location information, the system calls a pre-trained risk assessment model for risk assessment. This model is a multi-input and multi-output deep neural network that not only considers the fault type and location, but also combines factors such as the severity of the fault, the elevator's environmental conditions, and the passenger situation.

[0031] The severity of the fault is quantified by the fault type and the degree of deviation of the abnormal index; the environmental conditions include the type of building where the elevator is located (such as hospital, shopping mall, residential building), the current time, weather conditions, etc.; the passenger conditions include the current number of passengers, whether there are passengers with special needs (such as the elderly, children, and the disabled), etc. This information comes partly from the elevator state representation vector, and partly from the building management system and pre-set information.

[0032] The risk assessment model outputs a multi-dimensional risk score, including safety risk, rescue difficulty, time urgency, etc. For example, for the motor bearing fault diagnosed earlier, if it occurs in a hospital elevator and there is a patient in the car, the safety risk and time urgency scores will be higher.

[0033] For example, for a motor bearing failure, the simulation system will simulate the results of two strategies: "remotely reset the motor control system" and "emergency evacuation of passengers". The simulation results include strategy execution time, success probability, possible side effects, etc.

[0034] Finally, the system weightedly integrates the multi-dimensional risk scores and the rescue strategy feasibility assessment results to generate a risk assessment report. The fusion process uses an adjustable weight vector, and the initial weight is set based on historical rescue experience, which can be continuously optimized through reinforcement learning.

[0035] The risk assessment report includes the following: fault diagnosis results (type and location), risk scores in each dimension, recommended rescue strategies and feasibility analysis, estimated rescue time, resources to be mobilized, etc. For example, the report may recommend the immediate implementation of an emergency evacuation strategy and the arrangement of maintenance personnel to replace the bearing on site. The total rescue time is estimated to be 30 minutes, and two maintenance personnel and one medical staff need to be mobilized.

[0036] In an optional embodiment, The obtained elevator state representation vector is input into the pre-trained graph convolutional neural network model, which integrates the elevator system knowledge graph and outputs the fault type and fault location information including: Construct a multi-level elevator system knowledge graph that includes topological connections and causal dependencies between elevator components. The causal dependencies are represented by bidirectional directed edges, and the weights of the bidirectional directed edges represent the causal strength. The causal graph neural network is used to encode the constructed multi-level elevator system knowledge graph to obtain the causal representation of the nodes in the knowledge graph. Design a hybrid graph neural network model that includes a spatial-temporal graph convolution layer and a multi-scale attention mechanism, integrate the hybrid graph neural network model with the multi-level elevator system knowledge graph, input the elevator state representation vector into the hybrid graph neural network model, and process the input elevator state representation vector through multi-level information transmission and aggregation of the hybrid graph neural network model to generate multi-scale fault feature representation; Design a causal reasoning module based on the attention mechanism, use multi-scale fault feature representation to infer the root cause and propagation path of the fault, and obtain the causal explanation vector of the fault; The causal explanation vector of the fault is integrated with the multi-scale fault feature representation to generate a comprehensive fault feature. A multi-task learning framework with an adaptive weight allocation mechanism is designed to simultaneously classify the fault type and locate the fault location based on the comprehensive fault feature. At the same time, fault correlation constraints based on the multi-level elevator system knowledge graph structure are introduced, and preliminary fault type classification results and fault location location results are obtained using the multi-task learning framework. Multiple hybrid graph neural network models with different structures and parameters are constructed, and the same comprehensive fault features are used as input. The output results of multiple hybrid graph neural network models and multi-task learning frameworks are weightedly fused to generate the final fault type classification results and fault location positioning results.

[0037] For example, first, a multi-level elevator system knowledge graph is constructed. The knowledge graph contains the topological connection relationship and causal dependency relationship between the various elevator components. The topological connection relationship reflects the connection of elevator components in physical space, such as the connection between the motor and the traction wheel, the connection between the car and the guide rail, etc. The causal dependency relationship represents the fault propagation and influence relationship between the components, which is represented by bidirectional directed edges, and the weight of the edge represents the causal strength.

[0038] For example, there is a bidirectional causal relationship between motor bearings and motors: bearing failure can cause motor abnormalities, and motor overload may also accelerate bearing wear. This relationship is represented by a bidirectional directed edge in the knowledge graph, and the weights of the edges may be 0.8 and 0.6, indicating that bearing failure has a greater impact on the motor.

[0039] The construction of the knowledge graph is based on expert knowledge and historical fault data. By analyzing a large number of historical fault cases, extracting fault modes and propagation laws, the knowledge graph is continuously improved and updated. The knowledge graph adopts a multi-level structure, including component layer, subsystem layer and system layer, reflecting the hierarchical characteristics of the elevator system.

[0040] Next, a causal graph neural network is used to encode the constructed multi-level elevator system knowledge graph. The causal graph neural network considers the causal relationship between nodes and generates a causal representation of each node through multiple rounds of message passing. This process is similar to a random walk on the knowledge graph, but the direction and step length of the walk are affected by the causal relationship.

[0041] For example, for a motor node, its causal representation not only contains its own information, but also incorporates information about related components such as bearings and traction wheels, as well as information about possible faults caused by these components. This encoding method enables the representation of each node to contain rich system-level knowledge.

[0042] Then, a hybrid graph neural network model is designed and integrated with the multi-level elevator system knowledge graph. The model contains a space-time graph convolution layer and a multi-scale attention mechanism. The space-time graph convolution layer can simultaneously process the spatial structure and time series characteristics of the elevator system and capture the spatiotemporal evolution pattern of the fault. The multi-scale attention mechanism allows the model to focus on key information at different levels and scales.

[0043] The elevator state representation vector is input into the hybrid graph neural network model. The model first maps the input vector to the corresponding node of the knowledge graph, and then generates a multi-scale fault feature representation through multi-level information transmission and aggregation. For example, if the input vector shows that the motor current is abnormal, the model will pay special attention to the nodes of the motor and its related components, and generate fault feature representations of multiple scales including the motor system and transmission system through multiple rounds of information transmission.

[0044] Next, we design a causal reasoning module based on the attention mechanism. This module uses multi-scale fault feature representation and combines the causal relationship information in the knowledge graph to infer the root cause of the fault and the possible propagation path. This process is similar to reverse reasoning on a weighted causal graph to find the most likely fault source and propagation chain.

[0045] The output fault causal explanation vector contains information such as the inferred root cause, impact range, and severity. For example, for a motor overheating fault, the causal explanation vector may indicate that the root cause is bearing wear, the impact range includes the motor system and the transmission system, and the severity is medium.

[0046] The causal explanation vector of the fault is fused with the multi-scale fault feature representation to generate a comprehensive fault feature. This fusion process uses an attention mechanism to dynamically assign weights according to the importance of different features.

[0047] Based on the comprehensive fault characteristics, a multi-task learning framework with an adaptive weight allocation mechanism is designed to simultaneously classify fault types and locate fault locations. The adaptive weight allocation mechanism can dynamically adjust the weight of the loss function according to the difficulty and importance of different tasks. At the same time, fault correlation constraints based on the multi-level elevator system knowledge graph structure are introduced to improve the consistency of classification and location results.

[0048] In this embodiment, by integrating multi-level knowledge graphs and hybrid graph neural networks, domain knowledge and data-driven methods are fully utilized to significantly improve the accuracy of fault diagnosis. The causal reasoning module based on the attention mechanism provides the root cause and propagation path of the fault, enhancing the interpretability of the diagnosis results. The multi-model fusion strategy improves the robustness of the system and reduces the risk of misjudgment that may be caused by a single model. The adaptive weight allocation mechanism enables the system to be flexibly adjusted according to different fault scenarios, improving the adaptability of diagnosis. At the same time, fault type classification and location positioning are realized, providing comprehensive fault information for subsequent rescue decisions. Accurate diagnosis and positioning of elevator faults are achieved, laying the foundation for intelligent and automated elevator maintenance and emergency rescue.

[0049] In an optional embodiment, A multi-task learning framework with an adaptive weight allocation mechanism is designed to simultaneously classify fault types and locate fault locations based on comprehensive fault features. At the same time, fault correlation constraints based on the multi-level elevator system knowledge graph structure are introduced. The preliminary fault type classification results and fault location location results are obtained using the multi-task learning framework, including: Design a shared feature extraction network, which includes a residual connection and a hole convolution structure, input the comprehensive fault features into the shared feature extraction network, extract multi-scale features and generate a feature vector of fixed dimension, input the feature vector into the fault type classification task output layer and the fault location task output layer respectively, and output the initial fault type classification result and the initial fault location result respectively, wherein the fault type classification task output layer uses a fully connected layer, and the fault location task output layer uses a multi-layer perceptron structure and an attention mechanism; The fault type classification loss and the fault location loss are calculated based on the initial fault type classification results and the initial fault location location results, and a dynamic weight adjustment strategy based on task uncertainty is introduced to assign adaptive weights to each task. Based on the pre-built multi-level elevator system knowledge graph, a fault type correlation matrix is ​​generated, and the knowledge graph constraint loss is constructed using the fault type correlation matrix; Extracting intermediate features of the initial fault type classification result and the initial fault position location result, performing information interaction on the intermediate features to obtain interactive features, and updating the fault type classification result and the fault position location result based on the interactive features; Add random disturbances that obey Gaussian distribution to the comprehensive fault features to obtain the disturbed features, input the disturbed features into the shared feature extraction network, calculate the adversarial loss of the disturbed features, and combine the knowledge graph constraint loss, classification loss and positioning loss, as well as the adversarial loss for weighting to obtain the total loss; An alternating optimization strategy is adopted to optimize the model parameters and task weights, and a learning rate decay strategy and an early stopping mechanism are introduced during the optimization process. The iterations are repeated until the preset termination conditions are reached. The optimized model parameters and task weights are used to perform fault diagnosis on the new elevator operation data, and the final fault type classification results and fault location positioning results are output.

[0050] For example, first, a shared feature extraction network is designed. The network contains residual connections and dilated convolution structures, which can effectively extract multi-scale features. The residual connection alleviates the gradient vanishing problem of the deep network through short-circuit connections, while the dilated convolution expands the receptive field and captures a wider range of contextual information.

[0051] The comprehensive fault features are input into the shared feature extraction network. Next, the feature vectors are input into the fault type classification task output layer and the fault location location task output layer respectively. The fault type classification task output layer uses a fully connected layer. Assuming there are 10 types of faults, the number of neurons in the last layer is 10. The fault location location task output layer uses a multi-layer perceptron structure and an attention mechanism. Considering the complexity of the elevator system, it may contain multiple fault locations. This layer first passes through two fully connected layers, and then uses the attention mechanism to calculate the importance scores of different locations, and finally outputs the possible fault locations and their probabilities.

[0052] Then, the loss is calculated based on the initial fault type classification results and the initial fault location positioning results. The classification task uses the cross entropy loss, and the positioning task uses the square error loss. At the same time, a dynamic weight adjustment strategy based on task uncertainty is introduced. Specifically, the loss of each task is divided by the uncertainty estimate of the task, and the logarithm of the uncertainty is added. In this way, tasks with high uncertainty will receive smaller weights to avoid excessive impact on the overall optimization.

[0053] Next, based on the pre-built multi-level elevator system knowledge graph, a fault type correlation matrix is ​​generated. This matrix reflects the degree of association between different fault types. For example, "motor overheating" and "bearing wear" may have a high correlation (such as 0.8), while "motor overheating" and "door lock failure" may have a low correlation (such as 0.1). Using this correlation matrix, a knowledge graph constraint loss is constructed. This loss encourages the model to maintain consistency when predicting related faults and improves the reliability of diagnosis.

[0054] In order to further improve the generalization ability of the model, an adversarial learning strategy is introduced. A random perturbation that follows a Gaussian distribution is added to the comprehensive fault features to obtain the perturbed features. The perturbed features are input into the shared feature extraction network, and the difference between the output before and after the perturbation is calculated as the adversarial loss. This method enhances the robustness of the model to small perturbations and improves the performance in real noisy environments.

[0055] During the training process, the intermediate features of the initial fault type classification results and the initial fault location results are extracted, and information interaction is performed on these features. Specifically, the attention mechanism can be used to calculate the correlation between the features of the two tasks, and then feature fusion is performed based on the correlation. This interaction mechanism allows the two tasks to share useful information and improve overall performance.

[0056] Based on the interactive features, the fault type classification results and fault location positioning results are updated. This update process can be achieved through additional fully connected layers or convolutional layers to generate the final classification probability and positioning results.

[0057] Finally, the knowledge graph constraint loss, classification loss, localization loss, and adversarial loss are weighted and summed to get the total loss. The weights can be determined by the validation set performance or based on the importance of the task.

[0058] An alternating optimization strategy is used to optimize model parameters and task weights. Specifically, the task weights are first fixed and the model parameters are optimized; then the model parameters are fixed and the task weights are optimized. This process is repeated until the preset termination condition is reached (such as the performance of the validation set no longer improves or the maximum number of iterations is reached).

[0059] During the optimization process, a learning rate decay strategy and an early stopping mechanism are introduced. The learning rate starts from the initial value (such as 0.001) and decays once (such as multiplying by 0.9) every certain number of rounds (such as 10 rounds). The early stopping mechanism monitors the performance of the validation set. If the performance does not improve for multiple consecutive rounds (such as 5 rounds), the training is terminated early.

[0060] Finally, the optimized model parameters and task weights are used to perform fault diagnosis on the new elevator operation data, and the final fault type classification results and fault location positioning results are output.

[0061] In this embodiment, by sharing feature extraction and information interaction between tasks, the coordinated optimization of fault type classification and location positioning is achieved, and the overall diagnostic accuracy is improved. The dynamic weight adjustment strategy based on task uncertainty enables the model to adaptively allocate resources according to the difficulty of different tasks, thereby improving learning efficiency. The introduction of a multi-level elevator system knowledge graph enhances the domain knowledge of the model through correlation constraints, thereby improving the reliability and consistency of the diagnostic results. The use of an adversarial learning strategy improves the robustness of the model to input disturbances and enhances its adaptability in actual noisy environments. The alternating optimization strategy and learning rate adjustment mechanism improve the convergence speed and final performance of the model. The introduction of attention mechanisms and knowledge graph constraints improves the interpretability of the model decision-making process.

[0062] In an optional embodiment, The risk assessment is conducted based on the severity of the fault, environmental conditions, and passenger conditions to generate a multi-dimensional risk score. At the same time, the edge computing unit is started for real-time simulation to evaluate the feasibility of each rescue strategy. The multi-dimensional risk score and the rescue strategy feasibility assessment results are weighted and integrated to generate a risk assessment report including: Construct a risk assessment model based on multi-agent deep reinforcement learning. Each agent corresponds to a risk dimension, forming a multi-dimensional risk assessment system. The attention mechanism is used to process the information of fault type, fault location, fault severity, environmental conditions and passenger conditions to generate context feature vectors. Inputting the context feature vector into a pre-trained risk assessment model, and using a hierarchical soft actor-critic algorithm to generate a multi-dimensional dynamic risk score; Deploy an adaptive rescue strategy simulator based on a proximal strategy optimization algorithm on a distributed edge computing network, wherein the adaptive rescue strategy simulator uses a high-precision elevator physics engine that considers material fatigue and environmental factors, and the high-precision elevator physics engine performs simulation based on the dynamic mechanical characteristics of the elevator, the nonlinear electrical system response, the multi-layer elevator group control logic, and the group passenger behavior model; Using the adaptive rescue strategy simulator to simulate system responses under different rescue strategy combinations, and obtaining feasibility evaluation results and expected execution time of each rescue strategy combination; A weighted fusion network based on deep ensemble learning is constructed, and the multi-dimensional dynamic risk score and the feasibility assessment results of each rescue strategy combination are taken as input. The random dropout technology is applied in the weighted fusion network to obtain the uncertainty quantification result of the fusion process. The loss function is calculated according to the uncertainty quantification result, the weighted fusion network is trained, and the risk score report is output using the trained weighted fusion network.

[0063] For example, first, a risk assessment model based on multi-agent deep reinforcement learning is constructed. The model contains multiple agents, each of which is responsible for assessing a specific risk dimension, such as mechanical risk, electrical risk, passenger safety risk, etc. Each agent adopts a deep neural network structure, including multi-layer convolutional neural networks and fully connected layers.

[0064] In order to process the complex information of fault type, fault location, fault severity, environmental conditions and passenger conditions, an attention mechanism is introduced. Specifically, a multi-head self-attention layer is used to process the input information. Assuming there are 5 attention heads, each head independently calculates the attention weight, and then the results are concatenated and passed through a feedforward neural network to generate a context feature vector.

[0065] For example, the input information may include: fault type = "motor overheating", fault location = "12th floor motor compartment", fault severity = "moderate", environmental conditions = "high temperature and rainy", passenger conditions = "5 people, including 1 elderly person". The attention mechanism will automatically focus on the key parts of this information, such as the association between "motor overheating" and "high temperature".

[0066] Next, the context feature vector is fed into a pre-trained risk assessment model. The model uses a hierarchical soft actor-critic algorithm, where the actor network is responsible for generating the risk score and the critic network is responsible for evaluating the value of the current state. Specifically, the actor network contains three fully connected layers, and the last layer uses a softmax activation function to output a multi-dimensional risk score. The critic network also contains three fully connected layers and outputs a scalar value representing the estimated value of the current state.

[0067] By alternately optimizing the actor network and the critic network, the model is able to learn a more accurate risk assessment strategy. For example, for the above input, the model may output the following multi-dimensional dynamic risk scores: mechanical risk = 0.7, electrical risk = 0.8, passenger safety risk = 0.6, and overall risk = 0.75.

[0068] Then, an adaptive rescue strategy simulator based on the proximal policy optimization algorithm is deployed on the distributed edge computing network. The simulator uses a high-precision elevator physics engine that takes into account the influence of material fatigue and environmental factors. Specifically, the physics engine contains the following modules: Dynamic mechanical characteristics module: simulates the dynamic response of each mechanical component of the elevator, including the motion characteristics of the car, counterweight, wire rope, etc. Nonlinear electrical system response module: simulates the nonlinear response of electrical components such as motors and control systems. Multi-layer elevator group control logic module: simulates the scheduling logic of multiple elevators working together. Group passenger behavior model: simulates the behavior patterns of passengers in different situations. These modules work together to provide a highly realistic elevator system simulation.

[0069] The simulator evaluates the feasibility and expected execution time of each strategy combination. For example, for strategy 1, the simulation results may show: feasibility score = 0.8, expected execution time = 5 minutes.

[0070] Finally, a weighted fusion network based on deep ensemble learning is constructed. The network uses multiple deep neural networks with different structures (such as fully connected networks, convolutional neural networks, recurrent neural networks, etc.) as base learners. Each base learner independently processes the multi-dimensional dynamic risk score and the feasibility evaluation results of each rescue strategy combination.

[0071] Apply random dropout techniques in weighted fusion networks, that is, randomly shut down a portion of neurons during training. This technique not only prevents overfitting, but also provides an estimate of the uncertainty of the model's predictions. Specifically, multiple forward propagations (e.g., 100 times) are performed on the same input, each time using a different random dropout mask. The variance of these predictions can be used as a measure of uncertainty.

[0072] The loss function is calculated based on the uncertainty quantification results. For example, a negative log-likelihood loss can be used and an uncertainty-based regularization term can be added. This encourages the model to make more conservative predictions under uncertainty. The weighted fusion network is trained using the back-propagation algorithm. After training, the network is used to process the previously obtained risk scores and rescue strategy evaluation results to generate the final risk assessment report.

[0073] In an optional embodiment, The formula for calculating the loss function based on the uncertainty quantification result is as follows: ; Among them, L represents the loss function, N represents the number of samples, and w i represents the weight of the i-th sample, w j represents the weight of the jth sample, x i represents the feature vector of the i-th sample, x j represents the feature vector of the jth sample, σ i represents the uncertainty measure of the i-th sample, and λ represents the regularization coefficient.

[0074] In this embodiment, a multi-dimensional and dynamic evaluation of elevator failure risks is achieved through a multi-agent deep reinforcement learning model, which improves the accuracy and comprehensiveness of risk assessment. Based on real-time simulation of a high-precision physical engine, multiple rescue strategies can be generated and evaluated according to specific circumstances, which improves the pertinence and effectiveness of rescue decisions. Through deep integrated learning and random deactivation technology, the quantification of uncertainty in risk assessment and decision-making processes is achieved, and the reliability and robustness of the system are enhanced. Using a distributed edge computing network, real-time processing of complex simulation and decision-making processes is achieved, meeting the needs of rapid response in emergency situations. The multi-agent structure and modular design make the system easy to expand, and new risk dimensions or rescue strategies can be easily added. Through deep reinforcement learning methods, the system can learn from experience in each fault handling and continuously optimize the performance of risk assessment and rescue strategy generation.

[0075] S103. Based on the risk assessment report, activate the preset rescue strategy knowledge base, which contains different rescue plan templates, use the optimization algorithm to dynamically optimize the rescue plan template, generate the initial rescue instruction sequence and simulate and predict the results of different decision paths, select the optimal rescue decision path, generate a rescue instruction set based on the optimal rescue decision path, and send it to the elevator emergency rescue execution agency, continuously monitor the real-time feedback information of the elevator emergency rescue execution agency, and re-evaluate when an abnormality occurs until the passengers are safely evacuated and the system returns to a safe state.

[0076] Optionally, first, a preset rescue strategy knowledge base is activated based on the generated risk assessment report. The knowledge base is a structured database containing rescue plan templates for different fault scenarios. Each template includes the following key information: fault type, fault severity, environmental conditions, passenger conditions, rescue strategy steps, required resources, and expected execution time.

[0077] Next, the rescue plan template is dynamically optimized using an optimization algorithm to generate an initial rescue instruction sequence. Here, an optimization method based on a genetic algorithm is used. First, several templates that are most similar to the current fault scenario are selected from the knowledge base as the initial population. Then, new rescue strategy combinations are generated through crossover and mutation operations. The evaluation function considers multiple factors such as execution time, resource consumption, and passenger safety to select the optimal strategy combination.

[0078] Based on the initial rescue instruction sequence, the Monte Carlo tree search algorithm is used to simulate and predict the results of different decision paths. Specifically, each decision node is simulated thousands of times, and each simulation randomly selects subsequent actions until the rescue is completed, and the simulation results are recorded. By analyzing these simulation results, the optimal rescue decision path is selected. The evaluation criteria include average rescue completion time, passenger safety index, and resource utilization efficiency.

[0079] Generate a rescue instruction set based on the optimal rescue decision path. The instruction set includes three main parts: elevator system control instructions, emergency personnel dispatch instructions, and passenger guidance instructions. Next, the rescue instruction set is sent to the elevator emergency rescue execution agency through a secure encrypted channel. The AES-256 encryption algorithm is used to ensure the security of instruction transmission. At the same time, digital signature technology is used to verify the authenticity and integrity of the instructions.

[0080] The system continuously monitors the real-time feedback information of the elevator emergency rescue executive body. The feedback information includes: the current state of the elevator (speed, position, door status, etc.), environmental parameters (temperature, humidity, air quality, etc.), passenger status (number of people, vital signs, etc.), rescue progress (completed rescue steps, difficulties encountered, etc.), and this information is collected and processed in real time through IoT devices and edge computing nodes, and updated once a second.

[0081] When an abnormal situation occurs, such as a delay in the arrival of rescuers or a passenger's urgent medical needs, the system will trigger a decision re-evaluation mechanism. Specifically, the system will re-execute the process of risk assessment, rescue strategy optimization, and decision path simulation to generate a new set of rescue instructions.

[0082] The system continues to execute the above process until it receives two key signals: the passenger safe evacuation signal and the elevator system restored to a safe state signal. These two signals are sent by on-site rescue personnel and the elevator control system respectively, and require multiple confirmations to ensure accuracy. When both signals are confirmed, the system determines that the rescue mission is completed and enters the summary and evaluation stage. At this stage, the system will generate a detailed rescue report, including timeline, decision-making process, resource consumption and other information, to provide a basis for future rescue optimization.

[0083] In an optional embodiment, The preset rescue strategy knowledge base is activated in the risk assessment report, which contains different rescue plan templates. The rescue plan templates are dynamically optimized using the optimization algorithm to generate the initial rescue instruction sequence and simulate and predict the results of different decision paths. The optimal rescue decision path is selected, including: Perform semantic analysis based on the generated risk assessment report, match the preset rescue strategy knowledge base, and obtain the corresponding rescue plan template. The rescue strategy knowledge base is built based on the ontology and contains rescue plan templates for different fault scenarios as well as performance parameters and applicable conditions of rescue resources; The rescue plan template is dynamically optimized by using an optimization algorithm, three optimization objectives of rescue efficiency, resource consumption and secondary risk are constructed, the rescue resource deployment plan and rescue path planning are adaptively adjusted, and an initial rescue instruction sequence is generated, wherein the initial rescue instruction sequence includes the optimized rescue resource deployment plan and rescue path planning; Based on the initial rescue instruction sequence, a three-dimensional simulation model of the elevator system is constructed to simulate the rescue process under different decision paths, predict the rescue effect of each decision path, and statistically analyze the rescue indicators of each decision path to generate a decision path evaluation report, wherein the decision path evaluation report includes the rescue effect prediction results and rescue indicator statistical results of each decision path; The multi-attribute decision-making method is used to comprehensively evaluate the various rescue indicators in the decision path evaluation report, and the optimal rescue decision path is selected based on expert experience.

[0084] Exemplarily, first, semantic analysis is performed based on the generated risk assessment report to match the preset rescue strategy knowledge base. The risk assessment report contains key information such as the elevator fault type, fault severity, environmental conditions, and passenger conditions. Natural language processing technology is used to perform semantic analysis on the report and extract key features. The rescue strategy knowledge base is built based on ontology and described in OWL (Web Ontology Language). The core concepts of the knowledge base include: fault type, environmental conditions, passenger type, rescue resources, and rescue strategies. These concepts are interconnected through semantic relationships to form a complete knowledge network. Through the semantic matching algorithm, the extracted features are matched with the concepts in the knowledge base to obtain the corresponding rescue plan template. In this example, the "Main motor overheating moderate fault high temperature environment rescue plan" template may be matched.

[0085] Next, the rescue plan template is dynamically optimized using an optimization algorithm. Three optimization objectives are constructed: rescue efficiency, resource consumption, and secondary risks. Rescue efficiency is measured by the estimated rescue completion time, resource consumption includes the use of manpower, equipment, and materials, and secondary risks consider the additional dangers that may be caused during the rescue process.

[0086] An improved particle swarm optimization algorithm is used for multi-objective optimization. Each particle represents a possible rescue plan, including rescue resource deployment plan and rescue path planning. The position and velocity of the particle are updated iteratively and gradually converge to the optimal solution. During the optimization process, the performance parameters and applicable conditions of the rescue resources are considered, and the resource configuration is adaptively adjusted.

[0087] Based on the initial rescue command sequence, a 3D simulation model of the elevator system is constructed. The simulation environment is developed using the Unity3D engine, including accurate models of key components such as the elevator car, hoistway, and machine room. The elevator movement and heat transfer process are simulated through the physical engine to achieve highly realistic rescue scene reproduction.

[0088] In the simulation environment, the rescue process under different decision paths is simulated. For example, the following decision paths may be simulated: Path A: execute according to the initial rescue instruction sequence, Path B: evacuate passengers first and then repair equipment, Path C: add a rescuer to speed up the deployment of cooling equipment.

[0089] Each decision path is simulated multiple times, taking into account various random factors (such as fluctuations in the arrival time of rescuers, probability of equipment failure, etc.). The key indicators of each simulation are recorded, and a decision path evaluation report is generated through statistical analysis. The report contains the mean, standard deviation and confidence interval of each indicator of each decision path.

[0090] Finally, the multi-attribute decision-making method is used to comprehensively evaluate each decision path. The decision hierarchy structure is constructed using the analytic hierarchy process (AHP) and the weights of the evaluation indicators are set. Experts in the field of elevator rescue are invited to score each path, and the comprehensive score of each decision path is calculated by combining AHP.

[0091] Based on the comprehensive score and expert discussion, path C is selected as the optimal rescue decision path. This path performs best in terms of rescue efficiency and passenger safety. Although the resource consumption is slightly higher, it is still the best choice after comprehensive consideration.

[0092] In this embodiment, through semantic analysis and knowledge base matching, the system can quickly understand complex fault scenarios and extract relevant rescue knowledge, greatly improving the accuracy and efficiency of decision-making. By using the improved particle swarm algorithm for multi-objective optimization, the rescue resource configuration and path planning can be adaptively adjusted according to the actual situation to achieve optimal resource utilization. By constructing a high-fidelity three-dimensional simulation model, the rescue effect of different decision paths can be accurately predicted, providing a reliable basis for decision-making. Considering multiple dimensions such as rescue efficiency, resource consumption, and passenger safety, and making a comprehensive evaluation through a multi-attribute decision-making method, it is ensured that the selected rescue plan is the global optimal. The rescue strategy knowledge base constructed based on ontology not only supports current decisions, but also continuously absorbs new rescue experience to achieve continuous accumulation and optimization of knowledge. Through clear consideration and simulation prediction of secondary risks, the additional dangers that may occur during the rescue process are effectively reduced, and the overall safety of rescue is improved. Generate a detailed decision path evaluation report to make the decision-making process more transparent and explainable, which is conducive to the understanding and acceptance of the final rescue decision by the relevant parties.

[0093] In an optional embodiment, The multi-attribute decision-making method is used to comprehensively evaluate the various rescue indicators in the decision path evaluation report, and the optimal rescue decision path is selected in combination with expert experience, including: Constructing an interval-type fuzzy decision matrix, wherein the interval-type fuzzy decision matrix includes the performance of multiple candidate rescue decision paths on multiple evaluation attributes, and the performance of each candidate rescue decision path on each evaluation attribute is represented by a membership interval and a non-membership interval; Determine an attribute weight vector, the attribute weight vector is used to represent the importance of each evaluation attribute, and use a pre-calculated mean operator to fuse the interval-type fuzzy decision matrix, and based on the fusion result, calculate the comprehensive utility interval of each alternative rescue decision path, the comprehensive utility interval consists of a membership interval and a non-membership interval; A ranking method based on possibility is used to compare the sizes of the comprehensive utility intervals of various alternative rescue decision paths, and the possibility between each pair of comprehensive utility intervals is calculated. The possibility is calculated by the upper and lower bounds of the membership interval of each comprehensive utility interval; The candidate rescue decision paths are sorted according to the calculation result of the possibility, and the candidate rescue decision path with the largest possibility is selected as the optimal rescue path.

[0094] Exemplarily, first, an interval-type fuzzy decision matrix is ​​constructed. The matrix contains the performance of multiple alternative rescue decision paths on multiple evaluation attributes. The performance of each alternative rescue decision path on each evaluation attribute is represented by a membership interval and a non-membership interval. The membership interval represents the degree range of the path belonging to the attribute, and the non-membership interval represents the degree range of not belonging to the attribute.

[0095] For example, suppose there are three alternative rescue decision paths (A, B, C) and four evaluation attributes (rescue timeliness, resource consumption, safety, and operability).

[0096] First, an interval-type fuzzy decision matrix is ​​constructed. This matrix contains the performance of multiple alternative rescue decision paths on multiple evaluation attributes. The performance of each alternative rescue decision path on each evaluation attribute is represented by a membership interval and a non-membership interval. The membership interval represents the degree range to which the path belongs to the attribute, and the non-membership interval represents the degree range to which it does not belong to the attribute. This representation method can effectively capture the uncertainty and ambiguity of decision information.

[0097] For example, suppose there are three alternative rescue decision paths (A, B, C) and four evaluation attributes (rescue timeliness, resource consumption, safety, and operability). The constructed interval fuzzy decision matrix may be as follows: Path A performs as follows in terms of rescue timeliness ([0.7, 0.8], [0.1, 0.2]), in terms of resource consumption ([0.6, 0.7], [0.2, 0.3]), in terms of safety ([0.8, 0.9], [0.0, 0.1]), and in terms of operability ([0.7, 0.8], [0.1, 0.2]). The performance of paths B and C can be expressed similarly.

[0098] Next, determine the attribute weight vector. The attribute weight vector is used to indicate the importance of each evaluation attribute. It can be determined by expert scoring or analytic hierarchy process. For example, the attribute weight vector obtained may be (0.3, 0.2, 0.3, 0.2), which corresponds to the weights of rescue timeliness, resource consumption, safety, and operability, respectively.

[0099] Then, the interval-type fuzzy decision matrix is ​​fused using the pre-calculated mean operator. The mean operator takes into account the attribute weights and performs a weighted average of the performance of each path on each attribute. During the fusion process, the membership interval and non-membership interval are calculated separately. The purpose of this step is to integrate multi-dimensional decision information into a one-dimensional comprehensive evaluation.

[0100] Based on the fusion results, the comprehensive utility interval of each alternative rescue decision path is calculated. The comprehensive utility interval is also composed of membership interval and non-membership interval, reflecting the overall performance of the path.

[0101] Next, the possibility-based ranking method is used to compare the size of the comprehensive utility intervals of each alternative rescue decision path. The possibility between each pair of comprehensive utility intervals is calculated, and the possibility is calculated by the upper and lower bounds of the membership interval of each comprehensive utility interval. The possibility indicates the degree to which one interval is greater than another interval, and its value is between 0 and 1. The larger the value, the higher the possibility that one interval is greater than another interval.

[0102] Finally, the alternative rescue decision paths are sorted according to the calculation results of the possibility. For each path, calculate the average possibility of its possibility greater than all other paths. The path with the largest average value is the optimal rescue path. For example, the average possibility of path A is 0.55, the average possibility of path B is 0.30, and the average possibility of path C is 0.65. According to the calculation results, the average possibility of path C is the largest, so path C is selected as the optimal rescue path.

[0103] In this embodiment, by using interval-type fuzzy sets to represent decision information, the uncertainty and ambiguity in the rescue decision-making process can be better captured and processed, and the reliability of the decision can be improved. The multi-attribute decision-making method is adopted to comprehensively consider multiple aspects such as rescue timeliness, resource consumption, safety and operability, so as to ensure that the selected rescue path has good performance in all aspects. By introducing the attribute weight vector, the importance of each evaluation attribute can be flexibly adjusted according to the characteristics of different rescue scenarios, so that the decision is more in line with actual needs. The sorting method based on possibility provides a scientific method for comparing interval-type fuzzy numbers, which can accurately sort the alternative rescue paths and avoid the problem of unclear sorting that may occur in traditional methods. The pre-calculated mean operator is used for information fusion, which simplifies the calculation process and improves the decision-making efficiency, which is particularly suitable for emergency rescue situations that require rapid response. It is realized to select the optimal elevator rescue path while considering multiple decision attributes and uncertainty information, and provides an innovative solution for improving the decision-making quality of elevator emergency rescue.

[0104] Figure 2 FIG. 1 is a schematic diagram of the structure of an elevator emergency rescue decision-making system based on reinforcement learning according to an embodiment of the present invention. Figure 2 As shown, the system comprises: The first unit is used to collect elevator multimodal data in real time through a distributed sensor network, including mechanical vibration signals, electrical system parameters, acoustic characteristics, visual images and environmental indicators; preprocess the collected elevator multimodal data, perform feature fusion on the preprocessed elevator multimodal data through an attention mechanism, and generate an elevator state representation vector; The second unit is used to input the obtained elevator state representation vector into a pre-trained graph convolutional neural network model, which integrates the elevator system knowledge graph, outputs fault type and fault location information, conducts risk assessment based on fault severity, environmental conditions, and passenger conditions, generates a multi-dimensional risk score, and simultaneously starts the edge computing unit for real-time simulation to evaluate the feasibility of each rescue strategy. The multi-dimensional risk score and the rescue strategy feasibility assessment results are weighted and fused to generate a risk assessment report; The third unit is used to activate the preset rescue strategy knowledge base based on the risk assessment report, which contains different rescue plan templates. It uses the optimization algorithm to dynamically optimize the rescue plan template, generate the initial rescue instruction sequence and simulate and predict the results of different decision paths, select the optimal rescue decision path, generate a rescue instruction set based on the optimal rescue decision path, and send it to the elevator emergency rescue execution agency. It continuously monitors the real-time feedback information of the elevator emergency rescue execution agency, and re-evaluates when an abnormality occurs until the passengers are safely evacuated and the system returns to a safe state.

[0105] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0106] According to a fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0107] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An elevator emergency rescue decision-making method based on reinforcement learning, characterized in that: include: Collect elevator multimodal data in real time through a distributed sensor network, including mechanical vibration signals, electrical system parameters, acoustic characteristics, visual images, and environmental indicators; Preprocess the collected elevator multimodal data, perform feature fusion on the preprocessed elevator multimodal data through the attention mechanism, and generate an elevator state representation vector; The obtained elevator state representation vector is input into the pre-trained graph convolutional neural network model, which integrates the elevator system knowledge graph, outputs fault type and fault location information, combines fault severity, environmental conditions and passenger conditions for risk assessment, generates a multi-dimensional risk score, and starts the edge computing unit for real-time simulation to evaluate the feasibility of each rescue strategy. The multi-dimensional risk score and rescue strategy feasibility assessment results are weighted and fused to generate a risk assessment report. Based on the risk assessment report, the preset rescue strategy knowledge base is activated, which contains different rescue plan templates. The rescue plan templates are dynamically optimized using optimization algorithms to generate an initial rescue instruction sequence and simulate and predict the results of different decision paths. The optimal rescue decision path is selected, and a rescue instruction set is generated based on the optimal rescue decision path. The rescue instruction set is sent to the elevator emergency rescue executive agency, and the real-time feedback information of the elevator emergency rescue executive agency is continuously monitored. When an abnormality occurs, it is re-evaluated until the passengers are safely evacuated and the system returns to a safe state.

2. The method according to claim 1, characterized in that The obtained elevator state representation vector is input into the pre-trained graph convolutional neural network model, which integrates the elevator system knowledge graph and outputs the fault type and fault location information including: Construct a multi-level elevator system knowledge graph that includes topological connections and causal dependencies between elevator components. The causal dependencies are represented by bidirectional directed edges, and the weights of the bidirectional directed edges represent the causal strength. The causal graph neural network is used to encode the constructed multi-level elevator system knowledge graph to obtain the causal representation of the nodes in the knowledge graph. Design a hybrid graph neural network model that includes a spatial-temporal graph convolution layer and a multi-scale attention mechanism, integrate the hybrid graph neural network model with the multi-level elevator system knowledge graph, input the elevator state representation vector into the hybrid graph neural network model, and process the input elevator state representation vector through multi-level information transmission and aggregation of the hybrid graph neural network model to generate multi-scale fault feature representation; Design a causal reasoning module based on the attention mechanism, use multi-scale fault feature representation to infer the root cause and propagation path of the fault, and obtain the causal explanation vector of the fault; The causal explanation vector of the fault is integrated with the multi-scale fault feature representation to generate a comprehensive fault feature. A multi-task learning framework with an adaptive weight allocation mechanism is designed to simultaneously classify the fault type and locate the fault location based on the comprehensive fault feature. At the same time, fault correlation constraints based on the multi-level elevator system knowledge graph structure are introduced, and preliminary fault type classification results and fault location location results are obtained using the multi-task learning framework. Multiple hybrid graph neural network models with different structures and parameters are constructed, and the same comprehensive fault features are used as input. The output results of multiple hybrid graph neural network models and multi-task learning frameworks are weightedly fused to generate the final fault type classification results and fault location positioning results.

3. The method according to claim 2, characterized in that A multi-task learning framework with an adaptive weight allocation mechanism is designed to simultaneously classify fault types and locate fault locations based on comprehensive fault features. At the same time, fault correlation constraints based on the multi-level elevator system knowledge graph structure are introduced. The preliminary fault type classification results and fault location location results are obtained using the multi-task learning framework, including: Design a shared feature extraction network, which includes a residual connection and a hole convolution structure, input the comprehensive fault features into the shared feature extraction network, extract multi-scale features and generate a feature vector of fixed dimension, input the feature vector into the fault type classification task output layer and the fault location task output layer respectively, and output the initial fault type classification result and the initial fault location result respectively, wherein the fault type classification task output layer uses a fully connected layer, and the fault location task output layer uses a multi-layer perceptron structure and an attention mechanism; The fault type classification loss and the fault location loss are calculated based on the initial fault type classification results and the initial fault location location results, and a dynamic weight adjustment strategy based on task uncertainty is introduced to assign adaptive weights to each task. Based on the pre-built multi-level elevator system knowledge graph, a fault type correlation matrix is ​​generated, and the knowledge graph constraint loss is constructed using the fault type correlation matrix; Extracting intermediate features of the initial fault type classification result and the initial fault position location result, performing information interaction on the intermediate features to obtain interactive features, and updating the fault type classification result and the fault position location result based on the interactive features; Add random disturbances that obey Gaussian distribution to the comprehensive fault features to obtain the disturbed features, input the disturbed features into the shared feature extraction network, calculate the adversarial loss of the disturbed features, and combine the knowledge graph constraint loss, classification loss and positioning loss, as well as the adversarial loss for weighting to obtain the total loss; An alternating optimization strategy is adopted to optimize the model parameters and task weights, and a learning rate decay strategy and an early stopping mechanism are introduced during the optimization process. The iterations are repeated until the preset termination conditions are reached. The optimized model parameters and task weights are used to perform fault diagnosis on the new elevator operation data, and the final fault type classification results and fault location positioning results are output.

4. The method according to claim 1, characterized in that: The risk assessment is conducted based on the severity of the fault, environmental conditions, and passenger conditions to generate a multi-dimensional risk score. At the same time, the edge computing unit is started for real-time simulation to evaluate the feasibility of each rescue strategy. The multi-dimensional risk score and the rescue strategy feasibility assessment results are weighted and integrated to generate a risk assessment report including: Construct a risk assessment model based on multi-agent deep reinforcement learning. Each agent corresponds to a risk dimension, forming a multi-dimensional risk assessment system. The attention mechanism is used to process the information of fault type, fault location, fault severity, environmental conditions and passenger conditions to generate context feature vectors. Inputting the context feature vector into a pre-trained risk assessment model, and using a hierarchical soft actor-critic algorithm to generate a multi-dimensional dynamic risk score; Deploy an adaptive rescue strategy simulator based on a proximal strategy optimization algorithm on a distributed edge computing network, wherein the adaptive rescue strategy simulator uses a high-precision elevator physics engine that considers material fatigue and environmental factors, and the high-precision elevator physics engine performs simulation based on the dynamic mechanical characteristics of the elevator, the nonlinear electrical system response, the multi-layer elevator group control logic, and the group passenger behavior model; Using the adaptive rescue strategy simulator to simulate system responses under different rescue strategy combinations, and obtaining feasibility evaluation results and expected execution time of each rescue strategy combination; A weighted fusion network based on deep ensemble learning is constructed, and the multi-dimensional dynamic risk score and the feasibility assessment results of each rescue strategy combination are taken as input. The random dropout technology is applied in the weighted fusion network to obtain the uncertainty quantification result of the fusion process. The loss function is calculated according to the uncertainty quantification result, the weighted fusion network is trained, and the risk score report is output using the trained weighted fusion network.

5. The method according to claim 4, characterized in that The formula for calculating the loss function based on the uncertainty quantification result is as follows: ； Among them, L represents the loss function, N represents the number of samples, and w i represents the weight of the i-th sample, w j represents the weight of the jth sample, x i represents the feature vector of the i-th sample, x j represents the feature vector of the jth sample, σ i represents the uncertainty measure of the i-th sample, and λ represents the regularization coefficient.

6. The method according to claim 1, characterized in that Based on the risk assessment report, the preset rescue strategy knowledge base is activated, which contains different rescue plan templates. The rescue plan templates are dynamically optimized using the optimization algorithm to generate the initial rescue instruction sequence and simulate and predict the results of different decision paths. The optimal rescue decision path is selected, including: Perform semantic analysis based on the generated risk assessment report, match the preset rescue strategy knowledge base, and obtain the corresponding rescue plan template. The rescue strategy knowledge base is built based on the ontology and contains rescue plan templates for different fault scenarios as well as performance parameters and applicable conditions of rescue resources; The rescue plan template is dynamically optimized by using an optimization algorithm, three optimization objectives of rescue efficiency, resource consumption and secondary risk are constructed, the rescue resource deployment plan and rescue path planning are adaptively adjusted, and an initial rescue instruction sequence is generated, wherein the initial rescue instruction sequence includes the optimized rescue resource deployment plan and rescue path planning; Based on the initial rescue instruction sequence, a three-dimensional simulation model of the elevator system is constructed to simulate the rescue process under different decision paths, predict the rescue effect of each decision path, and statistically analyze the rescue indicators of each decision path to generate a decision path evaluation report, wherein the decision path evaluation report includes the rescue effect prediction results and rescue indicator statistical results of each decision path; The multi-attribute decision-making method is used to comprehensively evaluate the various rescue indicators in the decision path evaluation report, and the optimal rescue decision path is selected based on expert experience.

7. The method according to claim 6, characterized in that The multi-attribute decision-making method is used to comprehensively evaluate the various rescue indicators in the decision path evaluation report, and the optimal rescue decision path is selected in combination with expert experience, including: Constructing an interval-type fuzzy decision matrix, wherein the interval-type fuzzy decision matrix includes the performance of multiple candidate rescue decision paths on multiple evaluation attributes, and the performance of each candidate rescue decision path on each evaluation attribute is represented by a membership interval and a non-membership interval; Determine an attribute weight vector, the attribute weight vector is used to represent the importance of each evaluation attribute, and use a pre-calculated mean operator to fuse the interval-type fuzzy decision matrix, and based on the fusion result, calculate the comprehensive utility interval of each alternative rescue decision path, the comprehensive utility interval consists of a membership interval and a non-membership interval; A ranking method based on possibility is used to compare the sizes of the comprehensive utility intervals of various alternative rescue decision paths, and the possibility between each pair of comprehensive utility intervals is calculated. The possibility is calculated by the upper and lower bounds of the membership interval of each comprehensive utility interval; The candidate rescue decision paths are sorted according to the calculation result of the possibility, and the candidate rescue decision path with the largest possibility is selected as the optimal rescue path.

8. An elevator emergency rescue decision system based on reinforcement learning, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to collect elevator multimodal data in real time through a distributed sensor network, including mechanical vibration signals, electrical system parameters, acoustic characteristics, visual images and environmental indicators; preprocess the collected elevator multimodal data, perform feature fusion on the preprocessed elevator multimodal data through an attention mechanism, and generate an elevator state representation vector; The second unit is used to input the obtained elevator state representation vector into a pre-trained graph convolutional neural network model, which integrates the elevator system knowledge graph, outputs fault type and fault location information, conducts risk assessment based on fault severity, environmental conditions, and passenger conditions, generates a multi-dimensional risk score, and simultaneously starts the edge computing unit for real-time simulation to evaluate the feasibility of each rescue strategy. The multi-dimensional risk score and the rescue strategy feasibility assessment results are weighted and fused to generate a risk assessment report; The third unit is used to activate the preset rescue strategy knowledge base based on the risk assessment report, which contains different rescue plan templates. It uses the optimization algorithm to dynamically optimize the rescue plan template, generate the initial rescue instruction sequence and simulate and predict the results of different decision paths, select the optimal rescue decision path, generate a rescue instruction set based on the optimal rescue decision path, and send it to the elevator emergency rescue execution agency. It continuously monitors the real-time feedback information of the elevator emergency rescue execution agency, and re-evaluates when an abnormality occurs until the passengers are safely evacuated and the system returns to a safe state.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images