An elevator predictive maintenance method based on artificial intelligence

Abstract

An AI-based predictive maintenance method for elevators, belonging to the field of predictive elevator maintenance, first collects structured and unstructured elevator data. After filtering out invalid data and eliminating maintenance interference, a valid fault dataset is obtained. Then, a health index is calculated from five dimensions: fault level, fault frequency, running time, number of runs, and maintenance time. Weights are determined using the analytic hierarchy process (AHP) and standardized to obtain a comprehensive health score. Next, a temporal knowledge graph is constructed, and entities and relationships are embedded temporally using a TransR-T model. Subsequently, local features of the health score and the embedded features of the temporal knowledge graph are fused and input into a BiLSTM layer to capture long-term temporal dependencies. An attention mechanism is used to enhance the identification of high-risk entities, and the output is a fault type probability distribution. A Bayesian network is used to correct the fault level probability, and routine maintenance is triggered based on the correction result, maintenance priority, and health score. This invention achieves accurate predictive elevator maintenance.

Description

Technical Field

[0001] This invention belongs to the field of predictive maintenance of elevator systems, and specifically relates to an artificial intelligence-based predictive maintenance method for elevators. Background Technology

[0002] With the continuous advancement of urbanization and the wave of urban renewal, the demand for upgrading and renovating old elevators is rapidly increasing. At the same time, industrial plants, smart parks, and other scenarios are placing higher demands on the continuous operational reliability of elevators. Relevant regulations explicitly strengthen the management of the entire elevator lifecycle, prompting the maintenance industry to shift from traditional passive repair to proactive prevention. Meanwhile, the rising costs of manual maintenance and the inefficiency of traditional maintenance models are becoming increasingly prominent issues. Against this backdrop, there is an urgent need to find efficient predictive maintenance solutions that meet the current needs of industrial and urban development.

[0003] The current field of predictive maintenance for elevators faces significant technical bottlenecks: First, existing technologies largely rely on fixed-cycle maintenance models or single-indicator fault warnings, exhibiting clear technical shortcomings and application limitations. Second, single-indicator warning methods focus only on localized data such as fault records or operating parameters, failing to comprehensively depict the elevator's health status and lacking the ability to capture the temporal correlation features of fault occurrences, making it difficult to accurately predict potential fault risks. Furthermore, the integration and application of knowledge graphs and deep learning models in existing technologies are not mature enough, lacking specific adaptation solutions for the temporal evolution patterns of faults in the elevator field, thus failing to meet the practical needs of high safety, high accuracy, and low cost in elevator operation and maintenance scenarios.

[0004] The combination of multi-dimensional health assessment and elevator operation and maintenance can quantify elevator status, construct a fault knowledge graph that integrates time-series information, introduce quadruples with added time information through traditional triples, and use deep learning models to accurately predict fault types and risk levels. Finally, based on dynamically calculated maintenance priorities, predictive maintenance of elevators can be achieved. At present, research on predictive maintenance of elevators has formed a multi-path exploration. For example, Wu et al. constructed an elevator health status prediction model by integrating multi-source sensor data and LSTM network, which effectively captured the equipment degradation trend; Zhang et al. introduced graph attention network (GAT) to model the elevator fault propagation relationship and realize collaborative early warning of related faults. Although the above methods have proposed effective data analysis and prediction models for specific problems of elevator systems and made positive contributions to promoting the practical application of predictive maintenance, the challenges of existing research include: (1) Most current elevator predictive maintenance research is based on sensor time-series data or static knowledge graphs, which fails to effectively unify the modeling of elevator structured operating parameters, unstructured maintenance texts and time-series dependencies between faults, resulting in insufficient ability to capture complex fault chains and hidden risks. (2) Currently, most prediction models lack an interpretable maintenance decision-making mechanism and cannot directly link the prediction results to the graded response strategy. In actual implementation, they still rely on human experience judgment. Summary of the Invention

[0005] To overcome the shortcomings of existing methods, this invention proposes an AI-based predictive maintenance method for elevators. This method first collects structured and unstructured elevator data, filtering out invalid data and eliminating maintenance interference to obtain a valid fault dataset. Then, it calculates health indicators from five dimensions: fault level, number of faults, running time, number of runs, and maintenance time. Weights are determined using the analytic hierarchy process (AHP) and standardized to obtain a comprehensive health score. Next, a temporal knowledge graph is constructed, and the TransR-T model is used to embed entities and relationships temporally. Subsequently, local features of the health score and the temporal knowledge graph embedding features are fused and input into a BiLSTM layer to capture long-term temporal dependencies. An attention mechanism is used to enhance the identification of high-risk entities. Finally, the probability distribution of fault types is output, and a Bayesian network is used to correct the fault level probability. Based on the correction results, maintenance priority, and health score, 7-day emergency, 15-day planned, or 30-day routine maintenance is triggered, achieving accurate predictive maintenance for elevators.

[0006] The technical solution adopted in this invention is: An AI-based predictive maintenance method for elevators, the method comprising the following steps: The first step is to calculate the elevator health status: First, collect structured and unstructured elevator data, filter invalid data and eliminate maintenance interference to obtain a valid fault dataset; then calculate the health status index from five dimensions: fault level, number of faults, running time, number of runs and maintenance time, determine the weights by combining the analytic hierarchy process and standardize to obtain a comprehensive health status score. The second step is the construction and embedding of the temporal knowledge graph: constructing a temporal knowledge graph and embedding entities and relationships in a temporal manner through the TransR-T model; The third step is predictive maintenance: the local features of the health score are integrated with the embedded features of the temporal knowledge graph, and the input is fed into the BiLSTM layer to capture long temporal dependencies. The attention mechanism is combined to enhance the identification of high-risk entities and output the probability distribution of fault types. The fault level probability is corrected through a Bayesian network. Based on the correction results, maintenance priority and health score, routine maintenance is triggered to achieve accurate predictive maintenance of elevators.

[0007] Furthermore, in the first step, the elevator health calculation process is as follows: Step (1.1) Collection of structured and unstructured data. The structured data comes from the structured data in the elevator control cabinet and is obtained through the elevator IoT platform API. It includes elevator employee number, internal number, elevator model, fault event name, fault code, fault level, system status, fault description, fault reporting time, fault resolution time, running time and number of runs. The unstructured data includes fault cause, handling measures, fault phenomenon, and maintenance records, which are text composed of natural language. Step (1.2) Data preprocessing: Clean the original elevator data. First, remove invalid false alarm data with too short maintenance time, then filter the maintenance operation and its time window interference data, and finally obtain an effective dataset containing real faults. Step (1.3) Calculate the health indicators of five dimensions, including fault level, number of faults, running time, number of runs and maintenance time, and calculate the health score of the elevator from these five dimensions; Step (1.4) Comprehensive health calculation: The weights of fault level, number of faults, running time, number of runs and maintenance time are calculated according to the analytic hierarchy process. After calculating the health score, Z-score standardization based on global statistics is used to map it to the [0,1] interval to eliminate the scale differences between different elevator shafts and enhance the generalization ability of the model.

[0008] Furthermore, in the second step, the process of constructing and embedding the time-series knowledge graph is as follows: Step (2.1) Define the core elements of the time-series knowledge graph, including the valid fault dataset output from Step 1 and knowledge in the elevator maintenance domain; Step (2.2) Temporal knowledge graph construction: Based on the elements defined in step (2.1), the structured and unstructured data in step one are transformed into quadruplets containing head entity, relation, tail entity and time information to form a temporal knowledge graph; Step (2.3) Temporal knowledge graph embedding: The TransR-T embedding model is used to map entities and temporal relationships to a low-dimensional vector space, preserving temporal associations and dynamic evolution features.

[0009] Furthermore, in the third step, the predictive maintenance process is as follows: Step (3.1) Extract local anomaly features of health status, semantic-temporal features of temporal knowledge graph, and device similarity features. After all features are linearly transformed and output, they are spliced ​​and fused into the model input features. Step (3.2) will fuse multi-source features Input the BiLSTM layer to capture the temporal dependency features of elevator faults, then strengthen the feature representation of the high-incidence time steps of faults through the attention mechanism, and finally output the global core features of elevator faults that integrate multi-dimensional information. Step (3.3) introduces a learnable attention mechanism, which dynamically calculates the weights of each entity feature through a multilayer perceptron to enhance the model's ability to identify high-risk entities; Step (3.4) Output layer based on the feature matrix after attention fusion By using fully connected layers and activation functions, the probability distribution of fault types is output, and then the probability of fault level is corrected by a Bayesian network to achieve accurate maintenance decisions. Step (3.5) Maintenance phase output: output the global core features of elevator faults. Fault level probability corrected by Bayesian network The features are concatenated and then subjected to linear transformation and nonlinear activation. The probability distribution of the maintenance stage is output through Softmax classification, and the maximum probability is taken as the final maintenance stage decision result.

[0010] The beneficial effects of this invention are mainly reflected in: 1. This invention constructs an elevator health rating system based on five dimensions: fault level, number of faults, running time, number of runs, and maintenance time. It uses the analytic hierarchy process (AHP) to determine the weight of each dimension and employs Z-score standardization to eliminate scale differences between different elevators. This achieves the scientific integration and quantitative assessment of multi-dimensional health status data, providing reliable data support for maintenance decisions.

[0011] 2. This invention relies on a deep learning framework, integrates the local anomaly feature extraction of CNN with the long temporal dependency capture capability of BiLSTM, and combines the proposed TransR-T model to obtain semantic association features of temporal knowledge graph embedding, which effectively adapts to the characteristics of multi-factor induction and temporal evolution of elevator faults.

[0012] 3. This invention outputs the probability distribution of fault types through a deep learning model, and uses a Bayesian network to correct the probability of fault levels. Combined with health scores, it dynamically calculates maintenance priorities, and finally realizes a graded response strategy of 7-day emergency maintenance, 15-day planned maintenance and 30-day routine maintenance. Compared with the traditional fixed-cycle maintenance mode, it significantly improves the efficiency of emergency fault handling and the accuracy of maintenance resource scheduling. Attached Figure Description

[0013] Figure 1 This is a flowchart of an AI-based predictive maintenance method for elevators. Detailed Implementation

[0014] The present invention will now be further described with reference to the accompanying drawings.

[0015] Reference Figure 1 This paper presents an AI-based predictive maintenance method for elevators. First, it calculates a health index based on five dimensions: fault level, number of faults, running time, number of runs, and maintenance time. Weights are determined using the analytic hierarchy process (AHP), and after Z-score standardization and linear mapping, a comprehensive elevator health score is obtained. Next, a temporal knowledge graph is constructed based on a valid elevator fault dataset and maintenance domain knowledge. A time decay term is introduced, and the TransR-T model is used to embed entities and temporal relationships, resulting in a knowledge graph embedding representation. Then, local health features and knowledge graph embedding features are fused. After multi-source feature extraction and fusion, the data is input into a BiLSTM with an attention mechanism to capture long-term temporal dependencies in fault evolution and enhance the identification of high-risk entities. An initial fault probability distribution is output, and a Bayesian network is used to correct the fault level probability. Finally, based on the correction results, three levels of maintenance are output: 7-day emergency maintenance, 15-day planned maintenance, or 30-day routine maintenance, achieving accurate predictive maintenance for elevators.

[0016] The AI-based predictive maintenance method for elevators in this embodiment includes the following steps: The first step is to calculate the elevator's health status. The process is as follows: Step (1.1) Structured and unstructured data acquisition includes the following sub-steps; Step (1.1.1) Structured data acquisition; The structured data comes from the structured data in the elevator control cabinet and is obtained through the elevator IoT platform API. It includes elevator employee number, internal number, elevator model, fault event name, fault code, fault level, system status, fault description, fault reporting time, fault resolution time, running time and number of runs. Step (1.1.2) Unstructured data includes fault causes, handling measures, fault phenomena, and maintenance records, which are text composed of natural language.

[0017] Step (1.2) Data preprocessing: Clean the original elevator data. First, remove invalid false alarm data with too short maintenance time, then filter maintenance operations and interference data within their time windows, and finally obtain a valid dataset containing real faults. This includes the following sub-steps. Step (1.2.1) Data source definition: The original dataset generated by the elevator system within the time window T is as follows: ; Where N is the total number of fault events. This is the elevator employee number. This refers to the time the fault was reported. The time for troubleshooting. Fault level, For fault type codes, This represents the system state at the time the fault occurred. For other data in the dataset; Step (1.2.2) handles events with excessively short maintenance times. These events are false alarms caused by system misreporting or sensor interference, and are not actual elevator malfunctions. Therefore, events with excessively short maintenance times need to be filtered out from the dataset. Step (1.2.2.1) Calculate the repair time for each fault, in minutes. Calculate the repair time for each fault event: ; Step (1.2.2.2) filters out repair time that is less than the effective repair time threshold. The event obtains the dataset : ; in, The effective maintenance time threshold; Step (1.2.3) involves filtering maintenance status data to prevent fault data caused by maintenance activities from interfering with elevator health scores and predictive maintenance models. The process is as follows: Step (1.2.3.1) defines the maintenance impact window. : ; in, , The start and end times of the maintenance event. This serves as a buffer period, covering the preparation and completion phases before and after the maintenance. In step (1.2.3.2), all events where the system status is under maintenance or a fault occurs within the maintenance window are recorded as maintenance events. : ; Step (1.2.4) Initializing the dataset With dataset complement and The intersection of the complements of the two sets is used to obtain the valid fault dataset. ; in, The intersection operator. Let X represent the complement of X.

[0018] Step (1.3) calculates the health indicators of five dimensions, including fault level, number of faults, running time, number of runs, and maintenance time, and calculates the health score of the elevator from these five dimensions; including the following sub-steps; Step (1.3.1) Fault level dimension scoring: By quantifying the severity of the fault and the order of its occurrence, a weighted fusion is performed to transform subjective risk judgment into objective level classification; the process is as follows: Step (1.3.1.1) Basic Deduction Mapping: Deduction weights are set based on the severity of the fault; higher fault levels result in greater deductions. Learned from historical elevator fault maintenance data: ; in, Indicates the fault level , , , , , These are learnable parameters; Step (1.3.1.2) uses a time decay function, where recent faults have a greater impact and can mitigate the effects of older faults: ; in, The attenuation coefficient controls the attenuation rate. The number of days since the fault occurred ( ), The current time; Step (1.3.1.3) Total deduction for fault level: The weighted sum of deductions for all valid faults of each elevator is obtained from the base deduction and the time decay function. ; in, The total number of malfunctions for each elevator. Summation symbol; Step (1.3.1.4) Score the fault level dimension. With 100 points as the maximum score for the fault level dimension, subtract the weighted total deduction for all valid faults. To obtain the actual score for this dimension: ; in, This is a function that maximizes the score, ensuring that the score will not be lower than 0. Step (1.3.2) Scoring based on the number of failures, the process is as follows: Step (1.3.2.1) Count the number of valid faults: ; in, The cardinality of the set is the number of valid failure events. Step (1.3.2.2) uses logarithmic transformation to score the data, employing a logarithmic function to reflect the diminishing marginal returns effect. This means that as the number of failures increases, the impact of each new failure on the health status gradually decreases, consistent with the aging process of equipment. ; in, The adjustment coefficient controls the slope of the logarithmic curve; ln is the natural logarithm. Step (1.3.3) Scoring of the operating time dimension: Based on the life cycle pattern of elevator equipment, 5000 days is used as the dividing line for the operating time dimension score. When the operating time is less than 5000 days, the score drops linearly from 100 points to 80 points, and when it is greater than 5000 days, the score drops from 80 points to 50 points. ; in, The cumulative number of days the elevator has been in operation; Step (1.3.4) scores the number of runs. The more frequent the runs, the higher the equipment fatigue and the lower the score. The maximum percentage of score decay is limited, and the maximum magnitude of score decay is... This ensures that even if the number of runs far exceeds the threshold, the lowest score is still retained. The score is consistent with the actual fatigue characteristics of the equipment. ; in, This represents the average number of daily runs. This is a reasonable critical value for the number of daily elevator operations, used to determine whether the equipment's operating load exceeds the limit. The value is determined based on industry standards and the differences in elevator types. It is a minimum value function; Step (1.3.5) Score for maintenance time dimension: Based on elevator fault maintenance time, the score is dynamically adjusted by combining fault severity level weighting and maintenance time attenuation coefficient, and then output through normalization. The process is as follows: Step (1.3.5.1) first filters out elevators of the same model with fault levels of [missing information]. All historical valid repair times constitute a set Then, based on the actual operating status of a single elevator, the quantile thresholds corresponding to the lower limit of maintenance time, the optimal maintenance time, and the upper limit of maintenance time are dynamically calculated and adaptively adjusted. , , Then for the set Calculate the quantile values ​​corresponding to the above quantile thresholds to obtain the lower limit of repair time for this fault level. Optimal maintenance time Repair time limit This allows for the construction of a baseline parameter for maintenance time based on the fault level, adapted to the actual operating conditions of a single elevator. ; in, For elevators of the same model, the fault level is The set of all historical valid maintenance times, Fault level The corresponding minimum repair time, Fault level The corresponding optimal repair time, Fault level The corresponding maximum repair time, Represents a function that calculates the kth percentile of set X; Step (1.3.5.2) Single-fault repair time matching score: For the i-th valid fault, if the valid repair time of the fault is between the upper and lower limits of the baseline, the score is linearly adjusted according to the deviation ratio between the repair time and the optimal repair time, using the optimal repair time as the benchmark (corresponding to 100 points); if the repair time exceeds the upper limit, the score is adjusted from... Initially, points are deducted linearly based on the excess time; if the repair time is below the lower limit, starting from... Initially, points are added slightly based on the insufficient percentage, resulting in a final score for the repair time of the fault. : ; in, , , , , This is a learnable parameter, with a value range between 0 and 100; Step (1.3.5.3) Total score for maintenance time dimension: First, assign each fault to its level. Assign corresponding weights , The optimal weights are obtained through training with historical operation and maintenance data, satisfying the requirement that higher fault levels correspond to higher weights. The higher the value, the higher the score for each fault. Multiply the results by their corresponding weights and sum them up. Then sum the weights of all fault levels. Finally, divide the weighted sum of the former by the weighted sum of the latter to obtain the total score for the maintenance time dimension. : ; in, The level to which the i-th fault belongs The corresponding weights.

[0019] Step (1.4) Comprehensive health score calculation includes the following sub-steps: Step (1.4.1) calculates the weights of each dimension using the Analytic Hierarchy Process (AHP). The AHP is a multi-criteria decision-making method that decomposes complex decision problems into multiple levels and determines the weights of each factor through pairwise comparisons. In this invention, this method is used to determine the weights of five health dimensions: fault level, number of faults, running time, number of runs, and maintenance time. The process is as follows: Step (1.4.1.1) determines the weights and constructs the judgment matrix using the 1-9 scaling method: ; in, This indicates the importance of indicator i relative to indicator j, satisfying the following condition: The 1-9 scale uses integers from 1 to 9 and their reciprocals to represent the relative importance of two factors, where 1 indicates equal importance and 9 indicates that the former is extremely more important than the latter. Step (1.4.1.2) calculates the largest eigenvalue of the matrix using the eigenvalue method. The weight vector is obtained by normalizing the feature vector. ; ; ; in, For feature vectors, This is the normalized weight vector; Step (1.4.1.3) Consistency check: This quantifies the rationality of human subjective judgment and ensures that the judgment matrix does not contain logical contradictions. The process is as follows: Step (1.4.1.3.1) Calculate the consistency index CI: ; Where n=5 is the matrix order; Step (1.4.1.3.2) Obtain the RI value according to the RI reference table of the general random consistency index of the analytic hierarchy process. This table is obtained by summarizing the average consistency index of judgment matrices of different orders. Based on the order n of the judgment matrix in step 1.4.1.1, the corresponding RI value is obtained by matching from the table. Step (1.4.1.3.3) Calculate the consistency ratio CR: ; if Consistency is acceptable, if The judgment matrix needs to be adjusted; After calculating the health score in step (1.4.2), Z-score standardization based on global statistics is used to map it to the [0,1] interval, eliminating scale differences between different elevator shafts and enhancing the model's generalization ability. The process is as follows: Step (1.4.2.1) Calculate the overall score based on the weights obtained from the analytic hierarchy process and the scores of the five indicators: ; in, Fault level, The number of failures. Score based on runtime. Score based on the number of runs. Score based on repair time. The weights are obtained in step (1.4.1.2).

[0020] Step (1.4.2.2) calculates the global mean and standard deviation of the health scores for all elevators within the historical time window: ; ; in, Let j be the number of time steps for the j-th elevator. This represents the total number of elevators. Step (1.4.2.3) standardizes the health score of each elevator at each time step using Z-score: ; in, This indicates the elevator at time step t. value; Step (1.4.2.4) linearly maps the standardized scores to the [0,1] interval: ; in, and These are the minimum and maximum values ​​of all standardized scores.

[0021] The second step, the construction and embedding of the time-series knowledge graph, is as follows: Step (2.1) Define the core elements of the time-series knowledge graph, including the valid fault dataset output from Step 1. And knowledge in the field of elevator maintenance.

[0022] Step (2.2) Temporal knowledge graph construction: Based on the elements defined in step (2.1), the structured and unstructured data from step one are transformed into quadruplets containing head entities, relations, tail entities, and time information to form a temporal knowledge graph. This includes the following sub-steps: Step (2.2.1) Data alignment and cleaning, the process is as follows: Step (2.2.1.1) unifies the time granularity by normalizing all data to a daily granularity to ensure consistency of time information; Step (2.2.1.2) extracts entity, relation, and temporal attributes from the unstructured data in step (1.1.2) using natural language processing (NLP) techniques; Step (2.2.2) Temporal knowledge graph storage: A graph database is used to store quadruples, replacing the traditional triples. Expanded into quadruples with temporal information s,o represent the head entity and the tail entity, and r represents the relationship between the entities. Since current graph databases only support triple storage, this invention converts the time-series information and embeds it into the relationships between entities to establish a time-series knowledge graph based on a graph database. Step (2.3) Temporal knowledge graph embedding: The TransR-T embedding model is used to map entities and temporal relationships to a low-dimensional vector space, preserving temporal associations and dynamic evolution features; The basic TransR model is a classic knowledge graph embedding method. This invention proposes a temporally extended TransR-T embedding model, which, by introducing a time decay mapping matrix and a loss function with a temporal penalty term, enables the model to dynamically capture the evolutionary pattern of elevator faults that are closer to newer and farther from older. The model includes the following sub-steps: Step (2.3.1) Initial embedding of the temporal knowledge graph: A normal distribution random initialization strategy is adopted to randomly initialize all entities s, t and relations r in the temporal knowledge graph to a low-dimensional space, thereby obtaining the initial embedding vectors of the head entity, tail entity and relation, as well as the basic mapping matrix corresponding to the relation. : ; in, Represents the space of real numbers. The vector space representing the entity embedding vectors, The vector space representing the embedding vectors of relations; Step (2.3.2) involves the temporal mapping of entities to the relation space, introducing a time decay term. The fundamental mapping matrix obtained from step (2.3.1) Combined with the time decay term, the time factor optimized mapping matrix is ​​obtained. By using matrix multiplication, the initial entity embedding vector is mapped to the relation space, resulting in the mapped relation space embedding vectors of the head and tail entities. and The mapping matrix of recent relationships has higher weights, allowing entity embedding to dynamically adjust over time. This adapts to faults or events that occurred more recently, increasing their impact on the current elevator health status and fault prediction, while older faults have a smaller impact. ; ; ; in, Let r be the mapping matrix of the time series relation r at time τ. The time decay coefficient, For relationship timestamps, For the current time, , Let be the embedding vectors of the relation space of the head entity and the tail entity after mapping; Step (2.3.3) optimizes the time-series loss function. This time-series loss function iterates through all real-existing positive elevator operation and maintenance sample quadruples and constructed spurious negative sample quadruples, using L2 distance to measure the rationality of the quadruples. It requires that the irrationality of the negative samples is at least higher than the rationality of the positive samples by a marginal parameter. At the same time, a time-series penalty term is introduced. Constraining positive and negative sample pairs with excessively large time differences, and through By ensuring that the loss is non-negative, the semantic structure and temporal consistency of the knowledge graph are optimized simultaneously during training, enabling the model to accurately learn the semantic associations of elevator maintenance events. ; in, For the set of positive sample quadruplets, This is a set of negative sample quadruplets, obtained by randomly replacing the head entity, tail entity, or timestamp in the quadruplets. It is a function with maximum value. As a marginal parameter, the distance between positive and negative samples is increased. L2 distance function, which measures vector similarity. This is a time-series penalty term that penalizes positive and negative samples with excessively large time differences, thereby strengthening time-series consistency. Step (2.3.4) Vector standardization: The converged entity and relation embedding vectors are mapped to the [0,1] interval through Min-Max standardization to eliminate dimensional differences. ; in, The original embedding vector, This is the standardized vector; Step (2.3.5) integrates all quadruples from the 1st to the Tth time step by performing a union, and then concatenates the normalized head entity embedding, normalized relation embedding, and normalized tail entity embedding of the qth quadruple in each time step element by element, ultimately forming a complete temporal knowledge graph embedding representation. : ; in, This represents the number of quadruplets in the t-th time step. This represents the head entity of the q-th quadruple at time step t. The relationship between the t-th time step and the q-th quadruple. This represents the tail entity of the q-th quadruple at time step t. The union symbol represents taking the union of all sets from 1 to T, integrating all elements by time step.

[0023] The third step, predictive maintenance, is as follows: Step (3.1) extracts local anomaly features of health status, semantic-temporal features of temporal knowledge graph, and device similarity features. After all features are linearly transformed and output, they are spliced ​​and fused into the model input features, including the following sub-steps; Step (3.1.1) scores the overall health of the elevator. One-dimensional convolution and pooling operations are performed to extract local temporal anomaly features from the health status sequence. The process is as follows: Step (3.1.1.1) 1D convolution operation to capture anomalies in the elevator health sequence: ; in, This is a one-dimensional convolution operation function. Activation function ; Step (3.1.1.2) involves pooling to reduce the feature dimension, retain key anomaly features, and output a local anomaly feature vector. : ; in, It is the max pooling function; Step (3.1.2) Semantic-temporal feature extraction of the temporal knowledge graph, embedding the elevator temporal knowledge graph representation. Perform a linear transformation to extract the core features of the faulty entity and the temporal causal relationship: ; in, It is a linear transformation function; Step (3.1.3) Elevator equipment similarity feature extraction: Based on elevators of the same model and installation batch, common features are extracted using historical health data of similar elevators through average pooling to solve the problem of insufficient fault data for single elevators. ; in, The quantity of elevators of the same model and batch as the target elevator. Calculate the current overall health score for the j-th similar elevator; Step (3.1.4) Multi-source feature fusion, which integrates local anomaly features of fixed-dimensional health status. Semantic features of temporal knowledge graph Equipment similarity characteristics The data is then concatenated and dimensionality reduced to a fixed dimension from the model input through a linear transformation. And the feature distribution differences are eliminated by layer normalization: ; ; in, This is a feature vector concatenation operation.

[0024] Step (3.2) will fuse multi-source features Input the BiLSTM layer to capture the temporal dependency features of elevator faults, then strengthen the feature representation of the high-incidence time steps of faults through the attention mechanism, and finally output the global core features of elevator faults that integrate multi-dimensional information. Bidirectional Long Short-Term Memory (BiLSTM) is an extended model of Long Short-Term Memory (LSTM). Its core is to enhance the context modeling capability by processing sequences in both directions in parallel. It is widely used in feature extraction tasks of sequence data. Step (3.2) includes the following sub-steps: Step (3.2.1) involves updating the hidden state of the forward LSTM to capture the long-term temporal dependency features of elevator faults from the forward time series: ; ; in, It is a forward output gate. This represents the forward-looking cell state, storing long-term temporal information. Forward input gate, Forward forgetting gate, Forward weight matrix, Forward bias term, This is element-wise multiplication; Step (3.2.2) involves updating the hidden state of the LSTM to capture the long-term temporal dependency features of elevator faults from the reverse time series: ; ; in, It is a back-output gate. This is a backward cell state. It is a back-input gate. For backward forgetting gate, This is the backward weight matrix. This is a backward bias term; Step (3.2.3) involves bidirectional feature concatenation to obtain the final output of the BiLSTM layer at time step t: ; Step (3.3) introduces a learnable attention mechanism, which dynamically calculates the weights of each entity feature through a multilayer perceptron to enhance the model's ability to identify high-risk entities. This includes the following sub-steps: Step (3.3.1) outputs the features of BiLSTM Vector concatenation is performed with each entity embedding vector to avoid information loss during feature fusion. Then, a linear transformation is used to map the concatenated high-dimensional features to a low-dimensional space. Finally, the tanh activation function is used to perform a non-linear transformation on the fused features to obtain the deep fused entity features. ; in, This is the weight matrix. For bias terms, This is a vector concatenation operation. This represents the hyperbolic tangent activation function. Let be the normalized embedding vector of the i-th entity; Step (3.3.2) calculates the attention score for each entity using a fully connected layer: ; in, This is a learnable attention weight vector; Step (3.3.3) uses the Softmax function to normalize and obtain the attention weights: ; Where M is the total number of entities, that is, the number of all entities in the time-series knowledge graph; Step (3.3.4) performs a weighted summation of the entity embeddings based on attention weights and fuses them with BiLSTM features to obtain an attention-weighted feature vector. : ; ; in, and These are learnable parameters; Step (3.3.5) involves processing all time steps. of Concatenate them sequentially along the time dimension to obtain the complete feature matrix. ; Step (3.4) Output layer based on the feature matrix after attention fusion By using fully connected layers and activation functions, the probability distribution of fault types is output, and then the probability of fault level is corrected by a Bayesian network to achieve accurate maintenance decisions, including sub-steps. Step (3.4.1) Fault level probability prediction: The output is mapped to a probability distribution using the Softmax activation function to obtain the fault type probability distribution matrix. ; in, This is the weight matrix. For bias terms; Step (3.4.2) involves constructing a Bayesian network and using it to correct the fault level probability output by the deep learning model. The process is as follows: Step (3.4.2.1) constructs a fault level As the parent node, with M key operational features A Naive Bayesian network with child nodes. The set of values ​​is , Step (3.4.2.2) involves parameter learning, calculating prior probabilities and conditional probabilities from historical operation and maintenance data, as follows: Step (3.4.2.2.1) obtains the prior probability from the historical fault level frequency statistics. : ; in, Fault level in historical data Number of times it appears The Laplace smoothing coefficient is... Total number of fault levels; Step (3.4.2.2.2) involves running features divided into continuous features and discrete features, which need to be modeled separately when calculating conditional probabilities; Step (3.4.2.2.2.1) for continuous features Based on the Gaussian normal distribution, the fault level is calculated. Likelihood probability of the following feature value x Normalization coefficient Used to ensure that the integral sum of the probability density function is 1, the exponent term represents the sum of x and the mean. The greater the deviation, the lower the likelihood probability. ; in, and The fault levels are respectively Lower features The historical mean and standard deviation.

[0025] Step (3.4.2.2.2.2) for discrete features The fault level is calculated using a multinomial distribution. Likelihood probability of the following feature value x Smoothing coefficients are introduced into the numerator and denominator respectively. and This avoids the zero-probability problem of class frequencies being zero, ensuring the predictive stability of the model. ; in, Fault level Lower features The number of times the value k is taken. Fault level Total number of times Features The number of possible values, For smoothing coefficients; Step (3.4.2.3) posterior probability inference, for the current elevator feature vector The posterior probability distribution of the fault level L is calculated using Bayes' theorem: ; ; ; in, For a given fault level The likelihood of the observed eigenvector x is... This is a normalization constant, calculated using the law of total probability. Step (3.5) Maintenance phase output: output the global core features of elevator faults. Fault level probability corrected by Bayesian network The features are concatenated into a fusion feature set. After linear transformation and nonlinear activation, the probability distributions for the three maintenance stages are output through Softmax classification. The maximum probability is taken as the final maintenance stage decision result. ; ; ; ; in, , This is the weight matrix. , This is a bias term.

[0026] The implementation process of the AI-based predictive maintenance method for elevators in this embodiment is as follows: Step 1: Define the following 8 predictive maintenance models: CNNs (Convolutional Networks) can automatically extract local spatial features of data through structures such as convolutional layers and pooling layers, and reduce model complexity by using weight sharing. This model performs well in extracting local anomaly features of health sequences, but its ability to capture long-term temporal dependencies is weak.

[0027] LSTM: By introducing a gating mechanism consisting of an input gate, a forget gate, and an output gate, it can effectively capture long-term dependencies in sequential data, solving the gradient vanishing and gradient exploding problems that occur in traditional RNNs during long sequence training. However, it can only model from the forward direction and cannot fully utilize future information.

[0028] GRU: Through a simplified gating mechanism, it can efficiently capture dependencies in sequential data while alleviating the gradient vanishing problem of traditional RNNs. Compared to LSTM, GRU has fewer parameters and faster training speed, but its expressive power is slightly inferior to LSTM when modeling complex temporal patterns.

[0029] BiLSTM: Based on LSTM, it introduces a bidirectional structure, enabling it to simultaneously capture both forward and reverse temporal dependencies in sequence data, making more comprehensive use of contextual information. This model can utilize both past and future health sequence information, resulting in a stronger ability to model the temporal evolution of elevator faults.

[0030] BiGRU introduces a bidirectional structure based on GRU, processing time series in both forward and backward directions to efficiently capture forward and backward temporal dependencies. This model combines bidirectional modeling capabilities with the advantages of low parameter count, but its ability to express complex temporal patterns is slightly inferior to BiLSTM.

[0031] Transformer: A sequence modeling architecture based on self-attention mechanism, it captures the dependencies between arbitrary positions in a sequence through multi-head attention layers, enabling parallel processing of sequence data. In predictive maintenance of elevators, this method can globally model the health time series and has a strong ability to capture the evolution patterns of long-span faults, but the model has a large number of parameters and requires a large amount of training data.

[0032] GCN: By performing convolution operations on graph-structured data and utilizing adjacency matrices to propagate node features, it can learn spatial dependencies between nodes. It can model the correlation of fault types based on elevator fault propagation graphs, but it cannot effectively handle temporal dynamic information and unstructured text.

[0033] Proposed Model: The model of this invention integrates multi-source features, uses BiLSTM and attention mechanism to capture temporal dependencies, and combines TransR-T knowledge graph and Bayesian correction to achieve accurate fault prediction and emergency, planned and routine graded maintenance, effectively solving the problems of insufficient multi-source data fusion and poor decision interpretability.

[0034] Step 2: Introduction to the Experimental Dataset The experimental dataset is provided by an elevator time-series knowledge graph constructed using the method proposed in this invention and actual elevator data. The data comes from real-time elevator fault operation data obtained remotely using the TCP / IP communication protocol and contains 39,188 entities and semantic relationships.

[0035] The dataset includes real-time operational fault data for 14,259 elevators, with 89,188 fault data entries collected. The collected data includes each elevator's elevator number, internal number, elevator model, fault event name, fault code, fault description, fault reporting time, fault resolution time, running time, and number of runs.

[0036] The dataset was divided into training and test sets in a 7:3 ratio.

[0037] Step 3: Define evaluation indicators: To verify the predictive maintenance accuracy of the algorithm, this invention uses the following evaluation metrics to assess the algorithm: Accuracy (Acc) represents the proportion of correctly predicted samples out of all predicted samples. This metric directly reflects the overall predictive performance of the model, and the formula is as follows:

[0038] Among them, TP is the number of positive samples that are predicted to be positive, TN is the number of negative samples that are predicted to be negative, FN is the number of positive samples that are predicted to be negative, and FP is the number of negative samples that are predicted to be positive.

[0039] Precision (P) measures how many of the fault warnings issued by the model are actually actual faults. The formula is as follows:

[0040] Recall (R) represents the proportion of samples that were actually positive but were correctly predicted as positive by the model. This metric measures how many true faults the model can detect, and the formula is as follows:

[0041] The F1 score strikes a balance between precision and recall, providing a comprehensive evaluation of the model's classification performance. The formula is as follows:

[0042] Step 4: Analyze and compare the results: Refer to Table 1, which shows the comparison results of eight models in the elevator fault prediction task.

[0043] Table 1 shows the performance comparison results of eight models in the elevator fault prediction maintenance task;

[0044] From the perspective of overall classification performance, the method of this invention demonstrates significant advantages in core metrics: the ACC of this invention reaches 93.2%, far exceeding basic time-series models such as CNN and LSTM, and also surpassing advanced models such as BiLSTM, BiGRU, and Transformer, reflecting stronger overall classification capabilities. This indicates that the method of this invention has a more comprehensive ability to capture complex features in elevator operation data. In terms of precision and recall, the method of this invention reaches 88.7% and 89.8% respectively, both higher than all comparative models, indicating that this method achieves a better balance between warning accuracy and fault detection rate. The F1 score, as the harmonic mean of precision and recall, reaches 89.2%, which is 7.6 percentage points higher than the second-best BiLSTM, fully verifying the superiority of the method of this invention in comprehensive performance. Overall, this invention outperforms existing comparative methods in elevator predictive maintenance tasks.

[0045] This embodiment demonstrates the effectiveness of the improved elevator predictive maintenance method of the present invention through ablation experiments. The steps are as follows: Step 1: Define the following four predictive maintenance models: w / o TKG: The predictive maintenance method of this invention removes the time-series knowledge graph part.

[0046] w / o Attention: The predictive maintenance method of this invention removes the attention mechanism.

[0047] w / o Similarity: The predictive maintenance method of the present invention removes the equipment similarity feature portion.

[0048] Proposed Model: The predictive maintenance method of this invention.

[0049] Step 2: Introduction to the Experimental Dataset The experimental data used in this embodiment is the same as that in specific embodiment 2, and the dataset is divided into training set and test set in a 7:3 ratio.

[0050] Step 3: Define evaluation indicators: The classification performance of the algorithm was quantitatively analyzed using the classification evaluation metrics of precision (ACC), recall (R), precision (P), and F1-Score.

[0051] Step 4: Analyze and compare the results: Table 2 shows the comparison results of the ablation experiments, which demonstrates the impact of removing each module on the model performance.

[0052] Table 2 shows the comparison results of the ablation experimental performance of the method of the present invention;

[0053] The data shows that the method of this invention achieves optimal performance across all metrics, fully validating the necessity of the synergistic effect of each module. After removing the temporal knowledge graph, the ACC drops to 80.2%, and the F1 score drops to 78.9%, representing reductions of 13.0 and 10.3 percentage points respectively compared to the method of this invention, the most significant performance decrease among all ablation models. This indicates that the temporal knowledge graph embedding module is the core contribution of the method of this invention, effectively improving the model's ability to capture complex faults by fusing structured operating parameters with semantic associations and temporal evolution information in unstructured maintenance texts. After removing the attention mechanism, the ACC decreases by 10.4 percentage points and the F1 score decreases by 7.6 percentage points compared to the method of this invention. This indicates that the attention mechanism is crucial for the differentiated weighting of high-risk entities; its removal makes the model prone to missing important signals in long time series, leading to increased false negatives and false positives. After removing the device similarity feature, the ACC and F1 scores decrease by 9.7 and 6.8 percentage points respectively compared to the method of this invention. The effectiveness of utilizing historical health data from similar elevators in mitigating the data sparsity problem of individual elevators was verified. The performance of the three ablation models was lower than that of the method proposed in this invention, validating the necessity of the synergistic effect of the three modules: temporal knowledge graph embedding, attention mechanism, and device similarity features. In summary, the experimental results validate the effectiveness of the improved method proposed in this invention.

[0054] The embodiments described in this specification are merely examples of implementations of the inventive concept and are for illustrative purposes only. The scope of protection of this invention should not be considered limited to the specific forms described in these embodiments; rather, it extends to equivalent technical means conceived by those skilled in the art based on the inventive concept.

Claims

1. An AI-based predictive maintenance method for elevators, characterized in that, The method Includes the following steps: The first step is to calculate the elevator health status: First, collect structured and unstructured elevator data, filter invalid data and eliminate maintenance interference to obtain a valid fault dataset; then calculate the health status index from five dimensions: fault level, number of faults, running time, number of runs and maintenance time, determine the weights using the analytic hierarchy process and obtain a comprehensive health status score through standardization. The second step is the construction and embedding of the temporal knowledge graph: constructing a temporal knowledge graph and embedding entities and relationships in a temporal manner through the TransR-T model; The third step is predictive maintenance: the local features of the health score are integrated with the embedded features of the temporal knowledge graph, and the input is fed into the BiLSTM layer to capture long temporal dependencies. The attention mechanism is combined to enhance the identification of high-risk entities and output the probability distribution of fault types. The fault level probability is corrected through a Bayesian network. Based on the correction results, maintenance priority and health score, routine maintenance is triggered to achieve accurate predictive maintenance of elevators.

2. The predictive maintenance method for elevators based on artificial intelligence as described in claim 1, characterized in that, In the first step, the elevator health calculation process is as follows: Step (1.1) Collection of structured and unstructured data. The structured data comes from the structured data in the elevator control cabinet and is obtained through the elevator IoT platform API. It includes elevator employee number, internal number, elevator model, fault event name, fault code, fault level, system status, fault description, fault reporting time, fault resolution time, running time and number of runs. The unstructured data includes fault cause, handling measures, fault phenomenon, and maintenance records, which are text composed of natural language. Step (1.2) Data preprocessing: Clean the original elevator data. First, remove invalid false alarm data with too short maintenance time, then filter the maintenance operation and its time window interference data, and finally obtain an effective dataset containing real faults. Step (1.3) Calculate the health indexes of five dimensions, including fault level, number of faults, running time, number of runs and maintenance time, and calculate the health score of the elevator from these five dimensions; Step (1.4) Comprehensive health calculation: The weights of fault level, number of faults, running time, number of runs and maintenance time are calculated according to the analytic hierarchy process. After calculating the health score, Z-score standardization based on global statistics is used to map it to the [0,1] interval to eliminate the scale differences between different elevator shafts and enhance the generalization ability of the model.

3. The predictive maintenance method for elevators based on artificial intelligence as described in claim 2, characterized in that, Step (1.2) includes the following sub-steps: Step (1.2.1) Data source definition: Define the original dataset generated by the elevator system within the time window T. ; Step (1.2.2) Handling events with excessively short repair times: Calculate the repair time for each fault event, and obtain the dataset by filtering events whose repair times are less than the effective repair time threshold. Filter out the obtained dataset This means filtering out events with excessively short repair times; Step (1.2.3) filters the maintenance status data. Events where the system status is under maintenance or a fault occurs within the maintenance window are recorded as maintenance datasets. ; Step (1.2.4) Initializing the dataset With dataset complement and The intersection of the complements of the two sets yields the valid fault dataset.

4. The predictive maintenance method for elevators based on artificial intelligence as described in claim 2, characterized in that, Step (1.3) includes the following sub-steps: Step (1.3.1) Fault level dimension score: by quantifying the severity of the fault and the order of the fault occurrence, a weighted fusion is performed to transform subjective risk judgment into objective level classification; Step (1.3.2) Score the number of failures, count the number of valid failures, use a logarithmic function to reflect the diminishing marginal effect, and logarithmically transform the score; Step (1.3.3) Running time dimension score: Based on the life cycle pattern of elevator equipment, a dividing line for the running time dimension score is set, and a score is given according to the running time. Step (1.3.4) score based on the number of runs: the more frequently the equipment runs, the higher the equipment fatigue level and the lower the score. Step (1.3.5) Maintenance time dimension score: Based on the elevator fault maintenance time, combined with the fault severity level weighting and maintenance time decay coefficient dynamic adjustment, the maintenance time score is output through normalization.

5. The predictive maintenance method for elevators based on artificial intelligence as described in claim 4, characterized in that, The process of step (1.3.1) is as follows: Step (1.3.1.1) Basic Deduction Mapping: Deduction weights are set according to the severity of the fault; the higher the fault level, the more points are deducted. It was learned from historical elevator fault maintenance data; Step (1.3.1.2) sets the time decay function. Recently occurring faults have a greater impact, which can reduce the impact of older faults. Step (1.3.1.3) Total deduction for fault level: The weighted sum of deductions for all valid faults of each elevator is obtained from the base deduction and the time decay function; Step (1.3.1.4) Score the fault level dimension. With 100 points as the maximum score for the fault level dimension, subtract the weighted total deduction for all valid faults. This yields the actual score for that dimension.

6. The predictive maintenance method for elevators based on artificial intelligence as described in claim 4, characterized in that, The process of step (1.3.5) is as follows: Step (1.3.5.1) first filters out elevators of the same model with fault levels of [missing information]. All historical valid repair times constitute a set Then, based on the actual operating status of a single elevator, the quantile thresholds corresponding to the lower limit of maintenance time, the optimal maintenance time, and the upper limit of maintenance time are dynamically calculated and adaptively adjusted. , , Then for the set Calculate the quantile values ​​corresponding to the above quantile thresholds to obtain the lower limit of repair time for this fault level. Optimal maintenance time Maximum repair time In order to construct a maintenance time baseline parameter that adapts to the actual operating status of a single elevator; Step (1.3.5.2) Single fault repair time matching score: For the i-th valid fault, if the effective repair time of the fault is between the upper and lower limits of the baseline, the score is linearly adjusted according to the deviation ratio between the repair time and the optimal repair time, based on the optimal repair time. If the repair time exceeds the limit, from Initially, points are deducted linearly based on the excess time; if the repair time is below the lower limit, starting from... Initially, points are added slightly based on the insufficient percentage, resulting in a final score for the repair time of the fault. ; Step (1.3.5.3) Total score for maintenance time dimension: First, assign each fault to its level. Assign corresponding weights , The optimal weights are obtained through training with historical operation and maintenance data, satisfying the requirement that higher fault levels correspond to higher weights. The higher the value, the higher the score for each fault. Multiply the results by their corresponding weights and sum them up. Then sum the weights of all fault levels. Finally, divide the weighted sum of the former by the weighted sum of the latter to obtain the total score for the maintenance time dimension. .

7. The predictive maintenance method for elevators based on artificial intelligence as described in claim 2, characterized in that, Step (1.4) includes the following sub-steps: Step (1.4.1) calculates the weights of each dimension using the analytic hierarchy process, as follows: Step (1.4.1.1) determines the weights and constructs the judgment matrix using the 1-9 scaling method; Step (1.4.1.2) calculates the largest eigenvalue of the matrix using the eigenvalue method. The weight vector is obtained by normalizing the feature vector. ; Step (1.4.1.3) Consistency check: Calculate the consistency index CI. Obtain the RI value according to the general random consistency index RI lookup table of the analytic hierarchy process. This table is obtained by summarizing the average consistency indices of judgment matrices of different orders. Based on the order n of the judgment matrix, the corresponding RI value is obtained by matching from the table. Then calculate the consistency ratio CR. Consistency is acceptable, if The judgment matrix needs to be adjusted; Step (1.4.2) Calculate the comprehensive score based on the weights obtained by the analytic hierarchy process and the scores of the five indicators. Calculate the global mean and standard deviation of the health scores of all elevators within the historical time window. Perform Z-score standardization on the health scores of each elevator at each time step and linearly map the standardized scores to the [0,1] interval.

8. A predictive maintenance method for elevators based on artificial intelligence as described in any one of claims 1 to 7, characterized in that, In the second step, the construction and embedding process of the time-series knowledge graph is as follows: Step (2.1) Define the core elements of the time-series knowledge graph, including the valid fault dataset output from Step 1 and knowledge in the elevator maintenance domain; Step (2.2) Temporal knowledge graph construction: Based on the elements defined in step (2.1), the structured and unstructured data in step one are transformed into quadruplets containing head entity, relation, tail entity and time information to form a temporal knowledge graph; Step (2.3) Temporal knowledge graph embedding: The TransR-T embedding model is used to map entities and temporal relationships to a low-dimensional vector space, preserving temporal associations and dynamic evolution features.

9. The predictive maintenance method for elevators based on artificial intelligence as described in claim 8, characterized in that, In the third step, the predictive maintenance process is as follows: Step (3.1) Extract local anomaly features of health status, semantic-temporal features of temporal knowledge graph, and device similarity features. After all features are linearly transformed and output, they are spliced ​​and fused into the model input features. Step (3.2) will fuse multi-source features Input the BiLSTM layer to capture the temporal dependency features of elevator faults, then strengthen the feature representation of the high-incidence time steps of faults through the attention mechanism, and finally output the global core features of elevator faults that integrate multi-dimensional information. Step (3.3) introduces a learnable attention mechanism, which dynamically calculates the weights of each entity feature through a multilayer perceptron to enhance the model's ability to identify high-risk entities; Step (3.4) Output layer based on the feature matrix after attention fusion By using fully connected layers and activation functions, the probability distribution of fault types is output, and then the probability of fault level is corrected by a Bayesian network to achieve accurate maintenance decisions. Step (3.5) Maintenance phase output: output the global core features of elevator faults. Fault level probability corrected by Bayesian network The features are concatenated and then subjected to linear transformation and nonlinear activation. The probability distribution of the maintenance stage is output through Softmax classification, and the maximum probability is taken as the final maintenance stage decision result.

10. The predictive maintenance method for elevators based on artificial intelligence as described in claim 9, characterized in that, Step (3.4) includes the following sub-steps: Step (3.4.1) Fault level probability prediction: The output is mapped to a probability distribution through the Softmax activation function to obtain the fault type probability distribution matrix; Step (3.4.2) involves constructing a Bayesian network and using it to correct the fault level probability output by the deep learning model; the process is as follows: Step (3.4.2.1) constructs a fault level As the parent node, with M key operational features Naive Bayesian network with child nodes The set of values ​​is ; Step (3.4.2.2) parameter learning: calculate prior probability and conditional probability from historical operation and maintenance data; Step (3.4.2.3) posterior probability inference, for the current elevator feature vector The posterior probability distribution of the fault level L is calculated using Bayes' theorem.

Images