Elevator health assessment method based on time sequence knowledge graph and deep learning

Abstract

The application discloses an elevator health evaluation method based on a time sequence knowledge graph and deep learning, belongs to the field of elevator system health management, and comprises the following steps: firstly, constructing and dynamically updating an elevator fault time sequence knowledge graph; collecting structured operation data and unstructured maintenance text of the elevator, extracting entities and relations after preprocessing, and constructing a four-tuple time sequence knowledge graph integrating time information in a graph database; secondly, realizing precise evaluation of the health state of the elevator based on the time sequence knowledge graph; adopting a time sequence DistMult method to perform embedding representation on the graph, and generating a vector sequence integrating space-time information; extracting time sequence dependence features through a bidirectional gate recurrent unit, and combining a time step attention mechanism and a channel attention mechanism to dynamically focus on key degradation information and feature dimensions; finally, designing a residual multilayer perceptron classifier to classify the integrated features into health grades, introducing an incremental fine-tuning mechanism, and outputting probability evaluation results of each health grade of the elevator. The application improves accuracy.

Description

Technical Field

[0001] This invention belongs to the field of elevator system health management, specifically involving an elevator health assessment method based on temporal knowledge graphs and deep learning. Background Technology

[0002] Elevators have become a core component of modern urban vertical transportation, and their safe and stable operation directly impacts public safety and urban operational efficiency. This has led to a surge in elevator maintenance demands. However, traditional elevator maintenance primarily relies on periodic inspections and fault repairs, resulting in a significant waste of human and material resources. Deep learning models can uncover potential patterns in elevator operational data, enabling precise assessment of elevator health status. This shifts from reactive maintenance to proactive prevention, significantly improving maintenance efficiency and reducing safety risks. However, elevator malfunctions are characterized by a high probability of sudden occurrence and a short degradation cycle, making it difficult to accurately capture crucial pre-failure information, posing a unique challenge to health assessment.

[0003] Current elevator health assessment technology still faces several core bottlenecks: First, elevator operation data includes multiple heterogeneous types such as sensor time-series signals, operation and maintenance records, and fault reports. Existing technologies cannot effectively integrate this type of data and explore its potential correlations, resulting in a one-sided assessment basis. Second, conventional modeling methods lack accurate characterization of the time dimension, cannot capture the temporal dependencies of fault evolution, and are difficult to achieve early warning of early degradation.

[0004] The application of knowledge graphs in elevator health management can explore potential relationships between heterogeneous data. However, conventional knowledge graphs cannot store the crucial temporal information for health assessment. This leads to the development of temporal-series knowledge graphs composed of four-tuples. By optimizing these graphs through temporal embedding and incremental updates, the knowledge associations and temporal features of multi-source data are deeply integrated, providing "data + knowledge" dual-driven technical support for elevator health assessment. Sun et al. proposed a health assessment method based on knowledge graphs and link analysis. They extracted causal relationships from literature to construct a knowledge graph to determine key indicators and quantify their weights, then combined this with a stress-state-response model to construct a health index. Cheng et al. proposed a gas turbine health assessment method that integrates physical prior knowledge and spatiotemporal graph neural networks. These studies have achieved health assessments based on knowledge graphs for specific problems, making a significant contribution to the field of health management. However, the challenges of current research include: (1) The evolution of elevator health status is a continuous, nonlinear, and dynamic process affected by multiple coupled factors. Existing methods usually attach time as a static attribute to entities or relationships, or use simple temporal embedding methods, which makes it difficult to achieve deep fusion of multi-source heterogeneous data and accurate modeling of dynamic evolution. (2) New elevator fault data, component replacement, maintenance rule updates, etc. will bring about dynamic changes in data distribution and knowledge structure. Enabling the evaluation model to adapt to new data and integrate new knowledge at low cost without forgetting old knowledge is a key challenge for technology implementation. Summary of the Invention

[0005] To overcome the shortcomings of existing methods, this invention proposes an elevator health assessment method based on temporal knowledge graphs and deep learning. First, a temporal knowledge graph of elevator faults is constructed and dynamically updated. Structured operational data and unstructured maintenance text of the elevator are collected, preprocessed, and entities and relationships are extracted. A four-tuple temporal knowledge graph incorporating time information is then constructed in a graph database. Second, accurate assessment of elevator health status is achieved based on the temporal knowledge graph. The temporal DistMult method is used to embed the graph, generating a vector sequence incorporating spatiotemporal information. Temporal dependency features are extracted using a bidirectional gated recurrent unit, and time-step attention and channel attention mechanisms are combined to dynamically focus on key degradation information and feature dimensions. Finally, a residual multilayer perceptron (MLP) classifier is designed to classify the fused features into health levels, and an incremental fine-tuning mechanism is introduced to enable the model to continuously adapt to the distribution of new data, outputting probabilistic assessment results for each elevator health level.

[0006] The technical solution adopted in this invention is: An elevator health assessment method based on temporal knowledge graph and deep learning, the method comprising the following steps: The first step is to construct a knowledge graph of elevator fault timing. The process is as follows: Step (1.1) Acquisition of structured and unstructured data; Step (1.2) Structured data preprocessing; Fault data contains events that are not real faults, such as planned maintenance, elevator power outages, and downtime records caused by testing during maintenance. Such data needs to be filtered.

[0007] Step (1.3) Extraction of entities and relationships from unstructured data; Step (1.4) uses a graph database language to import data and constructs an elevator fault time sequence knowledge graph in the graph database; Step (1.5) Incremental data acquisition and preprocessing: Incremental data is the new data added within the time period from the last data acquisition to the current data acquisition. The incremental data is consistent with the initial data type and is divided into structured data and unstructured data. Step (1.6) Incremental storage of time-series knowledge graph: Import the new data into the graph database and merge duplicate nodes as in step (1.4).

[0008] The second step involves implementing elevator health assessment based on temporal knowledge graphs and deep learning. The process is as follows: Step (2.1) Embedding and incremental updating of the time-series knowledge graph; Step (2.2) Temporal feature extraction of embedded vectors; Step (2.3) Temporal attention-weighted fusion; Step (2.4) Channel attention mechanism: By dynamically allocating dimension weights, core features are strengthened, and the feature dimensions of the feature vector of elevator health assessment are selected by primary and secondary selection. Step (2.5) train the residual MLP classifier. Through the entire process of dimensionality mining, regularization to prevent overfitting, dimensionality reduction compression and residual optimization, the accurate mapping from multidimensional features to health level is achieved. Step (2.6) Incremental fine-tuning of the model, Step (2.7) Health assessment level classification: Step (2.8) Probability normalization; Step (2.9) Health level decision output; Step (2.10) uses a loss function to minimize the difference between the predicted probability and the true label, thereby optimizing the classification accuracy.

[0009] Furthermore, the process of step (1.1) is as follows: Step (1.1.1) Structured data acquisition: The structured data comes from the structured data in the elevator control cabinet. The elevator IoT platform obtains the data through API, including elevator employee number, internal number, elevator model, fault event name, fault code, 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.

[0010] Furthermore, the process of step (1.2) is as follows: Step (1.2.1) Filter data with short intervals between fault reporting and fault resolution; Let the original event set be: Where N is the total number of events, For the i-th event, This is a condition separator used to describe the conditions for elements in a set. This is the elevator employee number. This refers to the time the fault was reported. The time for troubleshooting. The name of the fault event. This represents other data within structured data. The absolute value of the difference is used as the elevator maintenance time. Since events with excessively short maintenance times are false alarms caused by system malfunctions or temporary sensor interference, they do not reflect the true failures caused by wear and tear on the elevator's core components. Including these events in the dataset would interfere with the model's learning of real failure modes. Therefore, it is necessary to filter out events with excessively short maintenance times to obtain a complete event set. : ; in , For repair time, This is the minimum effective duration threshold.

[0011] Step (1.2.2) filters out test events caused by inspection actions to obtain real fault data; Step (1.2.2.1) defines the maintenance event set F: ; in, Indicates the name of the fault event. A set of event types for a maintenance / inspection state; Step (1.2.2.2) for each maintenance event Define the time window of its impact on the same elevator: ; in, It is the threshold for the duration of maintenance impact; The test events caused by the maintenance status in step (1.2.2.3) are represented as fault events occurring in the same elevator. The fault event started during a maintenance event. Impact Time Window Inside: ; Step (1.2.3) Initial Filtering Set Set of test events caused by maintenance Replenishment and maintenance events Taking the intersection of the complements of the two sets yields the valid fault event dataset: ; in, The intersection operator. It represents the complement of X, that is, the set of all elements that do not belong to X.

[0012] Furthermore, the process of step (1.3) is as follows: Step (1.3.1) uses the BERT pre-trained language model, Softmax classifier, and rule post-processing mechanism to extract entities from unstructured text data, as follows: Step (1.3.1.1) Text data embedding; For unstructured text data, the character sequence is first mapped to a word vector sequence. The BERT model is used to vectorize the text. By learning to generate context-dependent dynamic word vector representations, the semantic and syntactic information of words in specific contexts is deeply captured and converted into corresponding word vector sequences. BERT is a pre-trained language model based on the Transformer architecture. Through pre-training on a large-scale unlabeled corpus, it can adaptively adjust the vector representation according to the specific context of the words, thereby deeply capturing the semantic and syntactic information of words in specific contexts. Step (1.3.1.2) Project word vectors into the label space; Given a sequence of word vectors ,in For the word vector representation at position t, perform a linear transformation on the word vector: ; in, Here is the weight matrix, where C is the total number of label categories. For bias vectors, This is the score vector; Step (1.3.1.3) Softmax probability normalization transforms the original scores into a probability distribution: ; in, This indicates that at position t, given the hidden state... Under the given conditions, the probability of the model outputting label c, Let be the score for label c corresponding to position t, and exp be the natural exponentiation operation. The summation symbol; Step (1.3.1.4) Label prediction: For each word position t, select the label with the highest probability as the initial prediction. ; The arg max function represents finding the class c that maximizes the probability P. Step (1.3.1.5) calculates the optimal result by calculating the cross-entropy loss function of the model output. First, the model targets the input temporal features. For each tag Output raw score Then, by calculating the index of all tag scores... Combined with the score of the real label c The ratio of the index score of the true label to the sum of the index scores of all labels. This is transformed into the probability of the model predicting the true label c. Finally, take the negative logarithm of that probability. Furthermore, by transforming the quotient of the logarithms into the difference of the logarithms, a directly calculable loss value can be obtained: ; ; in, Indicates that in a given input sequence When, output the probability of label c. Represents a given input sequence and tags The scoring function is defined by log, where log represents the logarithmic operation and exp represents the natural exponential operation.

[0013] By continuously adjusting the parameters of each layer to minimize the loss value, the optimal recognition model is obtained. Step (1.3.2) involves relation extraction, using dependency syntax provided by the LTP platform; Among them, the LTP language platform is a Chinese natural language processing toolkit that provides integrated processing capabilities such as part-of-speech tagging, syntactic analysis, and relation extraction. It is a commonly used platform in natural language processing.

[0014] In step (1.4), the preprocessed data is converted to CSV format and deployed to the default resource directory of the graph database. Data from the CSV files is imported in batches. To express the temporal characteristics of the temporal knowledge graph, the traditional triplet... Add information about time t. Represented as a quadruple, where Indicates the head entity. The tail entity is represented by r, which indicates the relationship between the head entity and the tail entity. Since current graph databases only support triple storage, this invention embeds time information into the attributes of the entities to realize the construction of a time-series knowledge graph. A triple is the smallest semantic unit consisting of a head entity, a relation, and a tail entity. It is the basic data form for describing entity relations and attributes in a knowledge graph.

[0015] The process of step (1.5) is as follows: Step (1.5.1) Structured incremental data processing repeats step (1.2), obtaining events with excessively short maintenance times as described in step (1.2.1). ,in For incremental event sets; Step (1.5.2) follows step (1.2.2) to obtain the test events caused by maintenance: ; in, This is an incremental maintenance event set; The events obtained in step (1.5.3) and step (1.5.1) The new test event set in step (1.5.2) Supplementary collection and new maintenance events Taking the intersection of the complements of the two sets yields the effective incremental fault event dataset: ; Step (1.5.4) Unstructured incremental data uses step (1.3) to extract new entities and relations, and outputs an incremental triplet set.

[0016] The process of step (2.1) is as follows: Step (2.1.1) Embedding the elevator fault timing knowledge graph; Temporal knowledge graph embedding is achieved using the Temporal DistMult method, which is a general-purpose low-dimensional temporal knowledge graph embedding method. ; in, For embedded head and tail entities, For relational embedding, For the time embedding of timestamp t, diag( This indicates that the vector is converted into a diagonal matrix. This represents the transpose of vector x; Step (2.1.2) Incremental embedding update, as follows: Step (2.1.2.1) determines the incremental embedding range; Extracting the newly added quadruple association subgraph From the newly added set of quadruplets In the process, entities that meet the head criteria are selected. Or tail entity Compared with the existing knowledge graph Quadruples with direct relationships The final set is : ; in, For the set of incremental quadruplets, For the existing knowledge graph, This indicates that entity e has a direct relationship with the graph G, thus avoiding an excessively large range of associated subgraphs. For logical OR operator; Step (2.1.2.2) Incremental embedding computation; Using the temporal knowledge graph embedding method from step (2.1.1), only update the embedding vectors of entities, relations, and time in the associated subgraphs: ; in, For incremental embedding vectors of newly added or associated entities, For incremental embedding vectors of newly added or associated relationships, Incremental timestamps Temporal feature embedding vector; Step (2.1.3) Incremental embedding vector fusion; After the temporal knowledge graph is updated, the incremental embedding vectors are merged with the original embedding vectors to form a complete temporal knowledge vector sequence; ; Where x represents an entity or relation, and t is a timestamp. This is the incrementally updated embedding vector. The original embedding vector; Step (2.1.4) Temporal sequence construction: Based on the fused embedding vectors, a new temporal step embedding vector sequence is generated in time step order. Where T is the current total number of detection cycles, This is the embedding vector at time step i after fusion.

[0017] The process of step (2.2) is as follows: Step (2.2.1) BIGRU (Bidirectional Gated Recurrent Unit) calculation; Step (2.2.1.1) constructs the reset gate; through the activation function Control whether invalid information from the previous time step is forgotten, and avoid invalid information affecting the current feature extraction; ; in, Activation function; To reset the weight matrix of the gate, To reset the bias matrix of the gate, This represents the hidden state of the one-way GRU at time step i-1. Step (2.2.1.2) constructs an update gate to determine whether the state information from the previous time step is passed to the current time step; ; in, To reset the weight matrix of the gate, To reset the bias matrix of the gate; Step (2.2.1.3) uses the reset door. Filter the hidden state from the previous moment Only historical information relevant to the current state is retained, and then the filtered historical information is embedded into the current time step vector. Concatenate them into feature vectors, and finally pass through Activation functions generate candidate states, and nonlinear temporal features are extracted to obtain candidate hidden states. ; Where tanh is the hyperbolic tangent activation function. This indicates element-wise multiplication. Here is the weight matrix for the candidate states. Let be the bias matrix of the candidate states; Step (2.2.1.4) calculates the final hidden state by updating the gate weights. Merge with the previous state With candidate state Output the final time series characteristics of the current time step to avoid the random interference of data from a single time step; ; Step (2.2.2) involves forward GRU sequence computation using the input features at the i-th time step. Combined with the hidden state of the previous time step Calculate the forward hidden state at the current time step. Iterate through all state data of the elevator from the initial stage of operation to the present and extract the positive temporal dependencies; ; in, This represents the hidden state of the forward GRU at time step i-1; Step (2.2.3) involves feeding back the GRU sequence and using the input features from the i-th time step. Combined with the hidden state at a time step Calculate the backward hidden state at the current time step. Starting from the last time step T, we traverse backwards to the first time step to extract the reverse temporal dependency between the result and the cause. ; in, This represents the hidden state of the backward GRU at time step i+1; Step (2.2.4) Full sequence bidirectional feature fusion: concatenate the forward and backward feature vectors to retain the bidirectional features of the entire time step; ; in, This is a vector concatenation operator.

[0018] Step (2.3) Time-attention weighted fusion: the importance of knowledge varies at different time steps, through... Calculate the dynamic weight of each time step feature and assign high weights to critical time steps of elevator malfunctions; ; in, This is the weight matrix. This is the bias matrix; Fusion features .

[0019] The process of step (2.4) is as follows: Step (2.4.1) Global average pooling is performed on the features after attention fusion at each time step. Perform a global average to obtain all characteristic information of the elevator's health status; ; Where D is the feature dimension.

[0020] Step (2.4.2) Dimensionality reduction mapping: Select core feature dimensions to avoid redundant features interfering with the evaluation results; ; Where ReLU is the activation function. For the dimension reduction weight matrix, For dimensionality reduction bias; Step (2.4.3) Dimension weight calculation: Assign dynamic weights to each core feature dimension. ; in, For activation function, For the weight matrix of increased dimension, For dimensional bias; Step (2.4.4) Feature Dimension Weighting: Multiply the feature matrix and weight matrix element-wise to strengthen core dimension features and weaken secondary dimensions. ; in, This indicates element-wise multiplication.

[0021] The process of step (2.5) is as follows: Step (2.5.1) Fully connected dimension upscaling, using the core vector Using fully connected layers and ReLU activation functions as input, the features are nonlinearly transformed to uncover the deep nonlinear correlations in the elevator's health status. ; in, This is the weight matrix. This is the bias matrix; Step (2.5.2) regularization involves randomly dropping out neurons using dropout to prevent the model from overfitting to noise in the training data. ; Dropout is a regularization technique that prevents overfitting by randomly discarding neurons. Step (2.5.3) Fully connected dimensionality reduction compresses the feature dimensions while retaining the core information; ; in, This is the weight matrix. This is the bias matrix; Step (2.5.4) layer normalization: normalize the features after dimensionality reduction to avoid differences in the numerical range of different feature dimensions; ; in, The normalization operator is given by the formula: x is the input feature , This is the mean of all feature dimensions of the current sample. The variance of all feature dimensions of the current sample. and For learnable parameters, ; Step (2.5.5) involves, to prevent gradient vanishing in deep networks, the input features... Features after transformation Add, through Activate the output to avoid gradient vanishing in deep networks and enhance the transmission of core features; ; Step (2.5.6) will As input, it is re-inputted into step (2.5.1), and the dimensionality increase, normalization, and residual connection process is carried out according to steps (2.5.1-2.5.5) to gradually optimize the feature representation and obtain the final output of the residual connection. ; In step (2.6), to adapt to the new feature distribution brought about by the incremental data and avoid the high cost of full training, the health assessment model is lightly fine-tuned, and only the parameters of the residual MLP in step (2.5) are trained, selecting the new time step and the associated time step. The fusion features are used as fine-tuning samples, and the number of samples is the incremental data volume. times, Ensure data diversity The offset is a non-negative integer. In step (2.7), the health level score mapping will ultimately output... The mapping is to an n-dimensional score, where n is the total number of health levels. Each dimension corresponds to a health level, and the higher the score, the higher the risk level of the elevator and the lower its health level. ; in, This is the weight matrix. The bias matrix, Given an n-dimensional score vector, The larger the value, the higher the failure risk corresponding to the j-th health level.

[0022] In step (2.8), the n-dimensional score vector Z output in step (2.7) is transformed into a probability distribution, which intuitively outputs the probability of the elevator's health level: ; Where exp is an exponential function.

[0023] In step (2.9), based on the probability distribution output in step (2.8), the dimension index corresponding to the maximum probability is calculated using the arg max function. The dimension index j corresponds one-to-one with the elevator health level, thus determining the final elevator health level. ; In step (2.10), the loss function is as follows: ; Where N is the number of training samples, Let i be the true label of the j-th class of the i-th sample. Let i be the predicted probability of the j-th class for the i-th sample. Represents the L2 regularization coefficient. Let L2 norm be the square of all weight matrices.

[0024] The beneficial effects of this invention are mainly reflected in: 1. By using time-series knowledge graphs, multi-source heterogeneous information such as elevator structural parameters, operating data, historical maintenance records, and fault cases are uniformly transformed into structured embedded vectors, which solves the problems of multi-source data format conflicts and missing related information in traditional methods, enabling the evaluation model to have the dual advantages of data-driven and knowledge-driven approaches.

[0025] 2. By capturing bidirectional temporal correlations through a bidirectional gated loop unit, and superimposing time step attention and channel attention to dynamically focus core information, it can capture the causal relationship of faults from the past to the present and the reverse verification of risks from the future to the present. It can also dynamically amplify the feature weights of key time steps of faults, avoid effective information being diluted by noise, and improve the accuracy of identifying minor degradation states of elevators.

[0026] 3. By updating the incremental knowledge graph, only the embedding vectors and classification parameters associated with the incremental data are updated, avoiding full recalculation. At the same time, dynamic class weights are introduced into the loss function to solve the classification bias problem of small sample faults. Furthermore, low-dimensional knowledge embedding and lightweight residual blocks are used to effectively reduce the cost and running time of elevator health assessment. Attached Figure Description

[0027] Figure 1 This is a flowchart of the incremental update embedding process for time-series knowledge graphs. Figure 2 Flowchart from fusion features to elevator health level classification Detailed Implementation

[0028] The present invention will now be further described with reference to the accompanying drawings.

[0029] Reference Figure 1 and Figure 2 An elevator health assessment method based on temporal knowledge graphs and deep learning, the method comprising the following steps: The first step is to construct a knowledge graph of elevator fault timing. The process is as follows: Step (1.1) Collection of structured and unstructured data, the process is as follows: Step (1.1.1) Structured data acquisition; The structured data originates from the structured data in the elevator control cabinet and is obtained by the elevator IoT platform via API. This data includes elevator employee number, internal number, elevator model, fault event name, fault code, fault description, fault reporting time, fault resolution time, running time, and number of runs.

[0030] Step (1.1.2) Unstructured data includes fault causes, handling measures, fault phenomena, and maintenance records, which are text composed of natural language.

[0031] Step (1.2) Structured data preprocessing; Fault data contains events that are not actual faults, such as planned maintenance, elevator power outages, and downtime records caused by testing during maintenance. This type of data needs to be filtered out, and the process is as follows: Step (1.2.1) Filter data with short intervals between fault reporting and fault resolution; Let the original event set be: Where N is the total number of events, For the i-th event, This is a condition separator used to describe the conditions for elements in a set. This is the elevator employee number. This refers to the time the fault was reported. The time for troubleshooting. The name of the fault event. This represents other data within the structured data, such as fault codes, fault descriptions, and runtime. The absolute value of the difference is used as the elevator maintenance time. Since events with excessively short maintenance times are false alarms caused by system malfunctions or temporary sensor interference, they do not reflect the true failures caused by wear and tear on the elevator's core components. Furthermore, they can interfere with the model's learning of real failure modes in the dataset. Therefore, it is necessary to filter out events with excessively short maintenance times to obtain a complete event set. : ; in , For repair time, This is the minimum effective duration threshold.

[0032] Step (1.2.2) filters out test events caused by inspection actions to obtain real fault data; Step (1.2.2.1) defines the maintenance event set F: ; in, Indicates the name of the fault event. A set of event types for a maintenance / inspection state; Step (1.2.2.2) for each maintenance event Define the time window of its impact on the same elevator: ; in, It is the threshold for the duration of maintenance impact; The test events caused by the maintenance status in step (1.2.2.3) are represented as fault events occurring in the same elevator. The fault event started during a maintenance event. Impact Time Window Inside: ; Step (1.2.3) Initial Filtering Set Set of test events caused by maintenance Replenishment and maintenance events Taking the intersection of the complements of the two sets yields the valid fault event dataset: ; in, The intersection operator. It represents the complement of X, that is, the set of all elements that do not belong to X; Step (1.3) Extraction of unstructured data entities and relations, the process is as follows: Step (1.3.1) uses a BERT pre-trained language model, a Softmax classifier, and a rule post-processing mechanism to extract entities from unstructured text data; as follows: Step (1.3.1.1) Text data embedding; For unstructured text data, the character sequence needs to be mapped to the word vector sequence first. This invention uses the BERT model to realize text vectorization. By learning to generate context-dependent dynamic word vector representations, it deeply captures the semantic and syntactic information of words in specific contexts and converts them into corresponding word vector sequences. BERT is a pre-trained language model based on the Transformer architecture. Through pre-training on a large-scale unlabeled corpus, it can adaptively adjust the vector representation according to the specific context of the words, thereby deeply capturing the semantic and syntactic information of words in specific contexts. Step (1.3.1.2) Project word vectors into the label space; Given a sequence of word vectors ,in For the word vector representation at position t, perform a linear transformation on the word vector: ; in, Here is the weight matrix, where C is the total number of label categories. For bias vectors, This is the score vector; Step (1.3.1.3) Softmax probability normalization transforms the original scores into a probability distribution: ; in, This indicates that at position t, given the hidden state... Under the given conditions, the probability of the model outputting label c, Let be the score for label c corresponding to position t, and exp be the natural exponentiation operation. The summation symbol; Step (1.3.1.4) Label prediction: For each word position t, select the label with the highest probability as the initial prediction. ; The arg max function represents finding the class c that maximizes the probability P. Step (1.3.1.5) calculates the optimal result by calculating the cross-entropy loss function of the model output. First, the model targets the input temporal features. For each tag Output raw score Then, by calculating the index of all tag scores... Combined with the score of the real label c The ratio of the index score of the true label to the sum of the index scores of all labels. This is transformed into the probability of the model predicting the true label c. Finally, take the negative logarithm of that probability. Furthermore, by transforming the quotient of the logarithms into the difference of the logarithms, a directly calculable loss value can be obtained: ; ; in, Indicates that in a given input sequence When, output the probability of label c. Represents a given input sequence and tags The scoring function for time, where log is the logarithmic operation and exp is the natural exponential operation. By continuously adjusting the parameters of each layer to minimize the loss value, the optimal recognition model is obtained. Step (1.3.2) Relation extraction: This invention uses dependency syntax provided by the LTP platform; Among them, the LTP language platform is a Chinese natural language processing toolkit that provides integrated processing capabilities such as part-of-speech tagging, syntactic analysis, and relation extraction. It is a commonly used platform in natural language processing. Step (1.4) uses a graph database language to import data and constructs an elevator fault time sequence knowledge graph in the graph database; The preprocessed data is converted to CSV format and deployed to the default resource directory of the graph database. Data from the CSV files is imported in batches. To express the temporal characteristics of the temporal knowledge graph, the traditional triplet... Add information about time t. Represented as a quadruple, where Indicates the head entity. The tail entity is represented by r, which indicates the relationship between the head entity and the tail entity. Since current graph databases only support triple storage, this invention embeds time information into the attributes of the entities to realize the construction of a time-series knowledge graph. A triple is the smallest semantic unit consisting of a head entity, a relation, and a tail entity. It is the basic data form for describing entity relations and attributes in a knowledge graph. Step (1.5) Incremental data acquisition and preprocessing: Incremental data refers to newly added data within the time frame from the previous data acquisition to the current data acquisition. The incremental data has the same data type as the initial data and is divided into structured data and unstructured data; the process is as follows: Step (1.5.1) Structured incremental data processing repeats step (1.2), obtaining events with excessively short maintenance times as described in step (1.2.1). ,in For incremental event sets; Step (1.5.2) follows step (1.2.2) to obtain the test events caused by maintenance: ; in, This is an incremental maintenance event set; The events obtained in step (1.5.3) and step (1.5.1) The new test event set in step (1.5.2) Supplementary collection and new maintenance events Taking the intersection of the complements of the two sets yields the effective incremental fault event dataset: ; Step (1.5.4) Unstructured incremental data is processed by step (1.3) to extract new entities and relations, and outputs an incremental triplet set; Step (1.6) Incremental storage of time-series knowledge graph: Import the new data into the graph database and merge duplicate nodes as in step (1.4).

[0033] The second step involves implementing elevator health assessment based on temporal knowledge graphs and deep learning. The process is as follows: Step (2.1) Temporal Knowledge Graph Embedding and Incremental Update: After the initial knowledge graph embedding, to avoid the high computational cost of full embedding, this invention only embeds and updates newly added knowledge and related knowledge, such as... Figure 1 As shown, the process is as follows: Step (2.1.1) Embedding the elevator fault timing knowledge graph; This invention uses the temporal DistMult method to embed temporal knowledge graphs. Temporal DistMult is a general-purpose low-dimensional temporal knowledge graph embedding method. ; in, For embedded head and tail entities, For relational embedding, For the time embedding of timestamp t, diag( This indicates that the vector is converted into a diagonal matrix. This represents the transpose of vector x; Step (2.1.2) Incremental embedding update, as follows: Step (2.1.2.1) determines the incremental embedding range; Extracting the newly added quadruple association subgraph From the newly added set of quadruplets In the process, entities that meet the head criteria are selected. Or tail entity Compared with the existing knowledge graph Quadruples with direct relationships The final set is : ; in, For the set of incremental quadruplets, For the existing knowledge graph, This indicates that entity e has a direct relationship with the graph G, thus avoiding an excessively large range of associated subgraphs. For logical OR operator; Step (2.1.2.2) Incremental embedding computation; Using the temporal knowledge graph embedding method from step (2.1.1), only update the embedding vectors of entities, relations, and time in the associated subgraphs: ; in, For incremental embedding vectors of newly added or associated entities, For incremental embedding vectors of newly added or associated relationships, Incremental timestamps Temporal feature embedding vector; Step (2.1.3) Incremental embedding vector fusion; After the temporal knowledge graph is updated, the incremental embedding vectors are merged with the original embedding vectors to form a complete temporal knowledge vector sequence; ; Where x represents an entity or relation, and t is a timestamp. This is the incrementally updated embedding vector. The original embedding vector; Step (2.1.4) Temporal sequence construction: Based on the fused embedding vectors, a new temporal step embedding vector sequence is generated in time step order. Where T is the current total number of detection cycles, This is the embedding vector at time step i after fusion; Step (2.2) Temporal feature extraction of the embedded vector, such as Figure 2 The BIGRU bidirectional feature extraction part is shown below, and the process is as follows: Step (2.2.1) BIGRU (Bidirectional Gated Recurrent Unit) calculation; Step (2.2.1.1) constructs the reset gate; through the activation function Control whether invalid information from the previous time step is forgotten, and avoid invalid information affecting the current feature extraction; ; in, Activation function; To reset the weight matrix of the gate, To reset the bias matrix of the gate, This represents the hidden state of the one-way GRU at time step i-1. Step (2.2.1.2) constructs an update gate to determine whether the state information from the previous time step is passed to the current time step; ; in, To reset the weight matrix of the gate, To reset the bias matrix of the gate; Step (2.2.1.3) uses the reset door. Filter the hidden state from the previous moment Only historical information relevant to the current state is retained, and then the filtered historical information is embedded into the current time step vector. Concatenate them into feature vectors, and finally pass through Activation functions generate candidate states, and nonlinear temporal features are extracted to obtain candidate hidden states. ; Where tanh is the hyperbolic tangent activation function. This indicates element-wise multiplication. Here is the weight matrix for the candidate states. Let be the bias matrix of the candidate states; Step (2.2.1.4) calculates the final hidden state by updating the gate weights. Merge with the previous state With candidate state Output the final time series characteristics of the current time step to avoid the random interference of data from a single time step; ; Step (2.2.2) involves forward GRU sequence computation using the input features at the i-th time step. Combined with the hidden state of the previous time step Calculate the forward hidden state at the current time step. Iterate through all state data of the elevator from the initial stage of operation to the present and extract the positive temporal dependencies; ; in, This represents the hidden state of the forward GRU at time step i-1; Step (2.2.3) involves feeding back the GRU sequence and using the input features from the i-th time step. Combined with the hidden state at a time step Calculate the backward hidden state at the current time step. Starting from the last time step T, we traverse backwards to the first time step to extract the reverse temporal dependency between the result and the cause. ; in, This represents the hidden state of the backward GRU at time step i+1.

[0034] Step (2.2.4) Full sequence bidirectional feature fusion: concatenate the forward and backward feature vectors to retain the bidirectional features of the entire time step; ; in, This is a vector concatenation operator; Step (2.3) Time-attention weighted fusion: the importance of knowledge varies at different time steps, through... Calculate the dynamic weight of each time step feature, and assign high weights to critical time steps of elevator malfunctions, such as... Figure 2 The time attention module is shown below; ; in, This is the weight matrix. This is the bias matrix; Fusion features ; Step (2.4) uses a channel attention mechanism to strengthen core features by dynamically allocating dimensional weights and prioritizing the feature dimensions of the elevator health assessment feature vector. Figure 2 The channel attention enhancement module is shown below; Step (2.4.1) Global average pooling is performed on the features after attention fusion at each time step. Perform a global average to obtain all characteristic information of the elevator's health status; ; Where D is the feature dimension; Step (2.4.2) Dimensionality reduction mapping: Select core feature dimensions to avoid redundant features interfering with the evaluation results; ; Where ReLU is the activation function. For the dimension reduction weight matrix, For dimensionality reduction bias. Step (2.4.3) Dimension weight calculation: Assign dynamic weights to each core feature dimension. ; in, For activation function, For the weight matrix of increased dimension, For dimensional bias; Step (2.4.4) Feature Dimension Weighting: Multiply the feature matrix and weight matrix element-wise to strengthen core dimension features and weaken secondary dimensions. ; in, This indicates element-wise multiplication; Step (2.5) trains a residual MLP classifier. Through a complete process of dimensionality mining, regularization to prevent overfitting, dimensionality reduction compression, and residual optimization, it achieves accurate mapping from multidimensional features to health levels, such as... Figure 2 The residual MLP classification module is shown below; Step (2.5.1) Fully connected dimension upscaling, using the core vector Using the input as input, the features are nonlinearly transformed through a fully connected layer and the ReLU activation function to uncover the deep nonlinear correlation of elevator health status; ; in, This is the weight matrix. This is the bias matrix; Step (2.5.2) regularization involves randomly dropping out neurons using dropout to prevent the model from overfitting to noise in the training data. ; Dropout is a regularization technique that prevents overfitting by randomly discarding neurons. Step (2.5.3) Fully connected dimensionality reduction compresses the feature dimensions while retaining the core information; ; in, This is the weight matrix. This is the bias matrix.

[0035] Step (2.5.4) layer normalization: normalize the features after dimensionality reduction to avoid differences in the numerical range of different feature dimensions; ; in, The normalization operator is given by the formula: x is the input feature , This is the mean of all feature dimensions of the current sample. The variance of all feature dimensions of the current sample. and For learnable parameters, ; Step (2.5.5) involves, to prevent gradient vanishing in deep networks, the input features... Features after transformation Add, through Activate the output to avoid gradient vanishing in deep networks and enhance the transmission of core features; ; Step (2.5.6) will As input, it is re-inputted into step (2.5.1), and the dimensionality increase, normalization, and residual connection process is carried out according to steps (2.5.1-2.5.5) to gradually optimize the feature representation and obtain the final output of the residual connection. ; Step (2.6) Incremental model fine-tuning: To adapt to the new feature distribution brought by the incremental data and avoid the high cost of full training, this invention performs lightweight fine-tuning on the health assessment model, training only the parameters of the residual MLP in step (2.5), and selecting the newly added time step and the associated time step. The fusion features are used as fine-tuning samples, and the number of samples is the incremental data volume. times, Ensure data diversity The offset is a non-negative integer. Step (2.7) Health assessment level classification: The health level score mapping will ultimately output... The mapping is to an n-dimensional score, where n is the total number of health levels. Each dimension corresponds to a health level, and the higher the score, the higher the risk level of the elevator and the lower its health level. ; in, This is the weight matrix. The bias matrix, Given an n-dimensional score vector, The larger the value, the higher the failure risk corresponding to the j-th health level. Step (2.8) Probability normalization transforms the n-dimensional score vector Z output in step (2.7) into a probability distribution, visually outputting the probability of the elevator's health level: ; Where exp is an exponential function; Step (2.9) Health level decision output; Based on the probability distribution output in step (2.8), the dimension index corresponding to the maximum probability is calculated using the arg max function. The dimension index j corresponds one-to-one with the elevator health level, thus determining the final elevator health level. ; Step (2.10) uses a loss function to minimize the difference between the predicted probability and the true label, optimizes the classification accuracy, and prevents overfitting; ; Where N is the number of training samples, Let i be the true label of the j-th class of the i-th sample. Let i be the predicted probability of the j-th class for the i-th sample. Represents the L2 regularization coefficient. Let L2 norm be the square of all weight matrices.

[0036] The implementation process of the elevator health assessment method based on temporal knowledge graph and deep learning in this embodiment is as follows: Step 1: Define the following 8 health assessment models: ANN: By simulating the connection mode of biological neurons, it uses a multilayer perceptron structure to perform nonlinear fitting of data, which can achieve mapping learning of complex features, but it is difficult to capture temporal dependencies when processing sequence data.

[0037] CNNs 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, but they are weak in capturing long-term temporal dependencies.

[0038] LSTM: By introducing a gating mechanism, it can effectively capture long-term dependencies in sequence data, solving the gradient vanishing and gradient explosion problems that occur in traditional RNNs during long sequence training.

[0039] GRU: Through a simplified gating mechanism, it can efficiently capture dependencies in sequence data while alleviating the gradient vanishing problem of traditional RNNs.

[0040] BIGRU: Based on GRU, it introduces a bidirectional structure, which can simultaneously capture the forward and reverse temporal dependencies of sequence data, make more comprehensive use of contextual information, and improve the learning ability of bidirectional temporal features.

[0041] SB-BiGRU: Based on BiGRU, it incorporates spatial and temporal attention mechanisms, which can adaptively focus on key time periods or core features in sequence data, further enhancing the ability to capture important information, while maintaining the learning advantage of BiGRU for bidirectional temporal dependencies.

[0042] MMNN is a model framework for transformer health assessment. It first divides the health status of transformers into 5 levels, then integrates data features through a multimodal interactive network structure, and finally uses transformer experimental data to verify its effectiveness in the health assessment task.

[0043] TKG-BiGRUDA: The health assessment method of this invention. Based on an incremental update model constructed using temporal knowledge graphs, dual attention mechanisms, and residual MLP, it enables cross-source correlation mining of elevator operation data and maintenance knowledge, ultimately achieving multi-level health status classification.

[0044] 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.

[0045] 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.

[0046] The dataset was divided into training set, test set and incremental dataset in a ratio of 7:2:1.

[0047] Step 3: Define evaluation indicators To verify the accuracy of the algorithm's health assessment, this invention uses the following evaluation metrics to evaluate the algorithm: Classification accuracy (ACC) refers to the proportion of samples in which the model's predicted elevator health level matches the actual label out of the total number of samples. It is used to quantify the overall classification accuracy of the model for elevator health levels, and the formula is as follows: ; in, This represents the number of samples whose predicted health level matches the actual label. This represents the total number of elevator operation data samples included in the evaluation.

[0048] The macro average (F1-mac) is used to balance the classification bias of different health levels, ensuring the identification accuracy of a few high-risk levels and avoiding missed detection of high-risk states due to sample bias. The formula is expressed as follows: ; ; ; Where C represents the total number of elevator health levels, and in this embodiment, C equals 5. This represents the number of samples that are actually of class i and are predicted to be of class i. The number of samples that are not actually of class i but are predicted to be of class i. This represents the number of samples that are actually of class i but are predicted to be of class non-class i.

[0049] The differential average (F1-min) is used to evaluate the overall recognition ability of all samples in a multi-class classification task. The formula is as follows: ; ; ; Step 4: Analyze the comparison results Referring to Table 1, which shows the comparison results of eight models, the present invention demonstrates significant advantages in core indicators based on the comparison data of elevator health level classification. The ACC of the present invention far surpasses that of basic models such as ANN and CNN, as well as classic time-series models such as LSTM and GRU, and even outperforms advanced models like BiGRU, SB-BiGRU, and the multimodal model MMNN, demonstrating better overall classification ability. This indicates that the present invention has a stronger ability to capture features from elevator operation data. Furthermore, the F1-mac is 0.14-0.28 higher than other methods, indicating that the present invention achieves more balanced accuracy in identifying high-risk elevator levels, avoiding the problem of most models missing minority classes. Although the F1-mic of the present invention is slightly lower than that of SB-BiGRU, this difference reflects the accuracy of the present invention in identifying high-risk levels, as F1-mic is easily inflated by multiple low-risk samples. Overall, the present invention outperforms existing methods in both overall accuracy and balanced level identification ability in elevator health level classification tasks.

[0050] Table 1 shows the comparison results of the eight models;

[0051] This embodiment demonstrates the effectiveness of the improved elevator health scoring method of the present invention through ablation experiments. The steps are as follows: Step 1: Define 4 health assessment models: TKG: The health assessment method of this invention has removed the temporal knowledge graph part.

[0052] NoA: The health assessment method of this invention removes the dual attention mechanism part.

[0053] FT: The health assessment method of this invention has removed the incremental update part.

[0054] TKG-BiGRUDA: The health assessment method of the present invention.

[0055] Step 2: 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, test set and incremental dataset in a ratio of 7:2:1.

[0056] Step 3: Define evaluation indicators This invention employs the following classification evaluation metrics to quantitatively analyze the classification performance of the algorithm used. These include recall (R), precision (P), and F1-Score.

[0057] ; ; ; Here, TP represents the number of samples correctly identified by the model based on the true label, FN represents the number of samples where the high-risk level was incorrectly classified as another level by the model, and FP represents the number of samples where the low-risk level was incorrectly classified as a high-risk level by the model. Recall indicates the proportion of high-risk levels that the model can identify, precision measures the accuracy with which the model correctly identifies high-risk levels, and F1-Score strikes a balance between precision and recall, providing a comprehensive evaluation of the model's classification performance.

[0058] Step 4: Analyze and compare the results Referring to Table 2, it shows the recall (R), precision (P), and F1-Score of four models (TKG, NoA, FT, and TKG-BiGRUDA) under five label classifications. The experimental results in Table 2 show that the metrics of the method described in this invention are higher than those of the TKG group. This indicates that the temporal knowledge graph can deeply mine cross-source temporal correlations in the data, reduce data noise interference, and improve the accuracy of health level classification and the precision of health assessment. The NoA group's ability to identify health levels for all five labels is far less than that of the method described in this invention, indicating that the attention mechanism is key to focusing on core information and that the dual-channel attention mechanism is indispensable. FT's three core metrics show the smallest difference from the method described in this invention, indicating that full retraining is theoretically close to incremental updates in accuracy, but lacks the ability to dynamically adapt to new data. Therefore, the temporal knowledge graph, dual attention mechanism, and incremental update mechanism of this invention all make substantial contributions to performance improvement, and the combined effect of these modules makes the method described in this invention far surpass other modules in R, P, and F1-Score, demonstrating its extremely strong ability to control false negatives and false positives for high-risk levels. In summary, the experimental results verify the effectiveness of the improved method of this invention.

[0059] Table 2 shows the experimental results;

[0060] The embodiments described in this specification are merely examples of implementations of the inventive concept and are for illustrative purposes only. The scope of protection of this invention should not be considered limited to the specific forms described in these embodiments; rather, it extends to equivalent technical means conceived by those skilled in the art based on the inventive concept.

Claims

1. A method for elevator health assessment based on temporal knowledge graphs and deep learning, characterized in that, The method includes the following steps: The first step is to construct a knowledge graph of elevator fault timing. The process is as follows: Step (1.1) Acquisition of structured and unstructured data; Step (1.2) Structured data preprocessing: Fault data contains events that are not real faults, such as planned maintenance, elevator power outages, and downtime records caused by testing during maintenance. Filter out this type of data. Step (1.3) Extraction of entities and relationships from unstructured data; Step (1.4) uses a graph database language to import data and constructs an elevator fault time sequence knowledge graph in the graph database; Step (1.5) Incremental data acquisition and preprocessing: Incremental data is the new data added within the time period from the last data acquisition to the current data acquisition. The incremental data is consistent with the initial data type and is divided into structured data and unstructured data. Step (1.6) Incremental storage of time-series knowledge graph: Import the new data into the graph database and merge duplicate nodes as in step (1.4); The second step involves implementing elevator health assessment based on temporal knowledge graphs and deep learning. The process is as follows: Step (2.1) Embedding and incremental updating of the time-series knowledge graph; Step (2.2) Temporal feature extraction of the embedding vector is as follows: Step (2.2.1) Calculation of the bidirectional gated loop unit; Step (2.2.2) Forward GRU sequence calculation: Traverse all state data of the elevator from the initial stage of operation to the current state and extract the forward temporal dependencies; Step (2.2.3) involves backward computation of the GRU sequence, traversing backward from the last time step T to the first time step, and extracting the reverse temporal dependencies of the result to the cause. Step (2.2.4) Full sequence bidirectional feature fusion: concatenate the forward and backward feature vectors to retain the bidirectional features of the entire time step; Step (2.3) Temporal attention-weighted fusion; Step (2.4) Channel attention mechanism: By dynamically allocating dimension weights, core features are strengthened, and the feature dimensions of the feature vector of elevator health assessment are selected by primary and secondary selection. Step (2.5) train the residual MLP classifier. Through the entire process of dimensionality mining, regularization to prevent overfitting, dimensionality reduction compression and residual optimization, the accurate mapping from multidimensional features to health level is achieved. Step (2.6) Incremental fine-tuning of the model, Step (2.7) Health assessment level classification: Step (2.8) Probability normalization; Step (2.9) Health level decision output; Step (2.10) uses a loss function to minimize the difference between the predicted probability and the true label, thereby optimizing the classification accuracy.

2. The elevator health assessment method based on temporal knowledge graph and deep learning as described in claim 1, characterized in that, The process of step (1.1) is as follows: Step (1.1.1) Structured data acquisition: The structured data comes from the structured data in the elevator control cabinet. The elevator IoT platform obtains the data through API, including elevator employee number, internal number, elevator model, fault event name, fault code, 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.

3. The elevator health assessment method based on temporal knowledge graph and deep learning as described in claim 1, characterized in that, The process of step (1.2) is as follows: Step (1.2.1) Filter data with short intervals between fault reporting and fault resolution: Assume the original event set Filter out events with excessively short repair times to obtain an event set. ; Step (1.2.2) filters out test events caused by inspection actions to obtain actual fault data, as follows: Step (1.2.2.1) defines the maintenance event set F; Step (1.2.2.2) for each maintenance event Define the time window of its impact on the same elevator. ; The test events caused by the maintenance status in step (1.2.2.3) are represented as fault events occurring in the same elevator. This fault event The start time is during the maintenance event. Impact Time Window Inside; Step (1.2.3) Initial Filtering Set Set of test events caused by maintenance Replenishment and maintenance events The intersection of the complements of the two sets yields the valid fault event dataset.

4. The elevator health assessment method based on temporal knowledge graph and deep learning as described in claim 1, characterized in that, The process of step (1.3) is as follows: Step (1.3.1) uses the BERT pre-trained language model, Softmax classifier, and rule post-processing mechanism to extract entities from unstructured text data, as follows: Step (1.3.1.1) Text data embedding; For unstructured text data, the character sequence is first mapped to a word vector sequence. The BERT model is used to vectorize the text. By learning to generate context-dependent dynamic word vector representations, the semantic and syntactic information of words in specific contexts is deeply captured and converted into corresponding word vector sequences. Step (1.3.1.2) Project word vectors into the label space; Given a sequence of word vectors ,in Let t be the word vector representation of position t, and then perform a linear transformation on the word vector; Step (1.3.1.3) Softmax probability normalization transforms the original scores into a probability distribution; Step (1.3.1.4) Label prediction: For each word position t, select the label with the highest probability as the initial prediction; Step (1.3.1.5) calculates the optimal result by calculating the cross-entropy loss function of the model output. First, the model targets the input temporal features. For each tag Output raw score Then, by calculating the index of all tag scores... Combined with the score of the real label c The ratio of the index score of the true label to the sum of the index scores of all labels. This is transformed into the probability of the model predicting the true label c. Finally, take the negative logarithm of that probability. Furthermore, by transforming the quotient of the logarithm into the difference of the logarithms, a directly calculable loss value is obtained; by continuously adjusting the parameters of each layer to minimize the loss value, the optimal recognition model is obtained. Step (1.3.2) Relation extraction uses dependency syntax provided by the LTP platform.

5. The elevator health assessment method based on temporal knowledge graph and deep learning as described in claim 3, characterized in that, The process of step (1.5) is as follows: Step (1.5.1) Structured incremental data processing repeats step (1.2), obtaining events with excessively short maintenance times as described in step (1.2.1). ; Step (1.5.2) obtains the test events caused by maintenance as in step (1.2.2); The events obtained in step (1.5.3) and step (1.5.1) The new test event set in step (1.5.2) Supplementary collection and new maintenance events The intersection of the complements of the two sets is used to obtain the effective incremental fault event dataset. Step (1.5.4) Unstructured incremental data uses step (1.3) to extract new entities and relations, and outputs an incremental triplet set.

6. The elevator health assessment method based on temporal knowledge graph and deep learning as described in claim 1, characterized in that, The process of step (2.1) is as follows: Step (2.1.1) Elevator fault temporal knowledge graph embedding: Temporal DistMult method is used to embed the temporal knowledge graph. Step (2.1.2) Incremental embedding update, as follows: Step (2.1.2.1) determines the incremental embedding range; Extracting the newly added quadruple association subgraph From the newly added set of quadruplets In the process, entities that meet the head criteria are selected. Or tail entity Compared with the existing knowledge graph Quadruples with direct relationships The final set is ; Step (2.1.2.2) Incremental embedding computation; Using the temporal knowledge graph embedding method in step (2.1.1), only the embedding vectors of entities, relations, and time in the associated subgraphs are updated; Step (2.1.3) Incremental embedding vector fusion; After the temporal knowledge graph is updated, the incremental embedding vectors are merged with the original embedding vectors to form a complete temporal knowledge vector sequence; Step (2.1.4) Temporal sequence construction: Based on the fused embedding vectors, a new temporal step embedding vector sequence is generated in time step order. Where T is the current total number of detection cycles, This is the embedding vector at time step i after fusion.

7. The elevator health assessment method based on temporal knowledge graph and deep learning as described in claim 1, characterized in that, The process of step (2.2) is as follows: Step (2.2.1) Calculation of the bidirectional gated loop unit; Step (2.2.1.1) constructs the reset gate; through the activation function Control whether invalid information from the previous time step is forgotten, and avoid invalid information affecting the current feature extraction; Step (2.2.1.2) constructs an update gate to determine whether the state information from the previous time step is passed to the current time step; Step (2.2.1.3) uses the reset door. Filter the hidden state from the previous moment Only historical information relevant to the current state is retained, and then the filtered historical information is embedded into the current time step vector. Concatenate them into feature vectors, and finally pass through Activation functions generate candidate states, and nonlinear temporal features are extracted to obtain candidate hidden states; Step (2.2.1.4) calculates the final hidden state by updating the gate weights. Merge with the previous state With candidate state Output the final time series characteristics of the current time step to avoid the random interference of data from a single time step; Step (2.2.2) involves forward GRU sequence computation using the input features at the i-th time step. Combined with the hidden state of the previous time step Calculate the forward hidden state at the current time step. Iterate through all state data of the elevator from the initial stage of operation to the present and extract the positive temporal dependencies; Step (2.2.3) involves feeding back the GRU sequence and using the input features from the i-th time step. Combined with the hidden state at a time step Calculate the backward hidden state at the current time step. Starting from the last time step T, we traverse backwards to the first time step to extract the reverse temporal dependency between the result and the cause. Step (2.2.4) Full sequence bidirectional feature fusion: concatenate the forward and backward feature vectors to retain the bidirectional features of the entire time step.

8. The elevator health assessment method based on temporal knowledge graph and deep learning as described in claim 1, characterized in that, The process of step (2.4) is as follows: Step (2.4.1) Global average pooling is performed on the features after attention fusion at each time step. Perform a global average to obtain all characteristic information of the elevator's health status; Step (2.4.2) Dimensionality reduction mapping: Select core feature dimensions to avoid redundant features interfering with the evaluation results; Step (2.4.3) Dimension weight calculation: Assign dynamic weights to each core feature dimension; Step (2.4.4) Feature Dimension Weighting: Multiply the feature matrix and weight matrix element-wise to strengthen the core dimension features and weaken the secondary dimensions, thus obtaining the core vector. .

9. The elevator health assessment method based on temporal knowledge graph and deep learning as described in claim 1, characterized in that, The process of step (2.5) is as follows: Step (2.5.1) Fully connected dimension upscaling, using the core vector Using fully connected layers and ReLU activation functions as input, the features are nonlinearly transformed to uncover the deep nonlinear correlations in the elevator's health status. Step (2.5.2) regularization involves randomly dropping out neurons using dropout to prevent the model from overfitting to noise in the training data. Step (2.5.3) Fully connected dimensionality reduction compresses the feature dimensions while retaining the core information; Step (2.5.4) layer normalization: normalize the features after dimensionality reduction to avoid differences in the numerical range of different feature dimensions; Step (2.5.5) involves, to prevent gradient vanishing in deep networks, the input features... Features after transformation Add, through Activate the output to avoid gradient vanishing in deep networks and enhance the transmission of core features; Step (2.5.6) will As input, it is re-inputted into step (2.5.1), and the dimensionality increase, normalization, and residual connection process is carried out according to steps (2.5.1-2.5.5) to gradually optimize the feature representation and obtain the final output of the residual connection. .

10. The elevator health assessment method based on temporal knowledge graph and deep learning as described in claim 1, characterized in that, In step (2.6), to adapt to the new feature distribution brought about by the incremental data and avoid the high cost of full training, the health assessment model is lightly fine-tuned, and only the parameters of the residual MLP in step (2.5) are trained, selecting the new time step and the associated time step. The fusion features are used as fine-tuning samples, and the number of samples is the incremental data volume. times, Ensure data diversity The offset is a non-negative integer. In step (2.7), the health level score mapping will ultimately output... The mapping is to an n-dimensional score, where n is the total number of health levels. Each dimension corresponds to a health level, and the higher the score, the higher the risk level of the elevator and the lower its health level. In step (2.8), the n-dimensional score vector Z output in step (2.7) is transformed into a probability distribution, which intuitively outputs the probability of the elevator's health level. In step (2.9), based on the probability distribution output in step (2.8), the dimension index corresponding to the maximum probability is solved by the arg max function. The dimension index j corresponds one-to-one with the elevator health level, thus determining the final elevator health level.

Images