Predictive maintenance method based on elevator operation and maintenance time sequence knowledge graph

Abstract

A predictive maintenance method based on an elevator operation and maintenance time sequence knowledge graph comprises the steps that firstly, the elevator operation and maintenance time sequence knowledge graph is constructed, predictive maintenance of electromechanical equipment parts is achieved based on elevator part maintenance period optimization, a theoretical distribution model based on Weibull distribution is constructed, and parameters of Weibull distribution are fitted through a least square method. Therefore, the model can accurately reflect the failure rule of the elevator parts; calculating an inference result of the elevator operation and maintenance time sequence knowledge graph through a graph convolution model and a Horkes process model, and carrying out recursive relation calculation of component fault rate functions in N maintenance cycles; an elevator component maintenance comprehensive cost simulation model is constructed, and the optimal maintenance cycle of elevator components is calculated through a continuous time Markov chain method; the fault prediction accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of predictive maintenance of electromechanical equipment, and more specifically to a method for optimizing the maintenance cycle of elevator components. Background Technology

[0002] Industry 4.0-driven intelligent monitoring technologies have generated massive amounts of equipment operation data, providing a data foundation for predictive maintenance of elevators. Current mainstream maintenance strategies are shifting from traditional periodic maintenance to data-driven predictive maintenance, using machine learning models to predict equipment status trends and provide early warnings of faults, thereby achieving precise allocation of maintenance resources. However, elevators, as typical electromechanical composite systems, exhibit differentiated degradation characteristics and dynamically correlated failure modes in their core components such as traction systems, door operators, and brakes, posing unique challenges to predictive maintenance.

[0003] Significant technical bottlenecks exist in the field of predictive maintenance of elevator components: First, there is insufficient modeling of the temporal correlation of multi-source heterogeneous data (vibration signals, fault logs, maintenance records) in elevators. Existing methods mostly use single sensor data analysis, which fails to effectively capture the coupling relationship between mechanical wear and electrical parameter drift. Second, there is a contradiction between the dynamic degradation characteristics of key components and fixed maintenance cycles. For example, the wear rate of brakes changes nonlinearly with the frequency of load, while current maintenance plans are still based on mean time between failures, resulting in maintenance redundancy or lag.

[0004] Knowledge graphs can be applied to elevator operation and maintenance to explore potential relationships between heterogeneous data. Conventional knowledge graphs cannot store the most important time information for predictive maintenance, which leads to the development of time-series knowledge graphs composed of quadruples. Predictive maintenance can be achieved by optimizing the reasoning results of time-series knowledge graphs through algorithms. At present, research on reasoning based on time-series knowledge graphs is very extensive. For example, Xu et al. combined contrastive learning to achieve prediction of time-series knowledge graphs. Bai et al. achieved completion of time-series knowledge graphs based on temporal logic rules. Although the above methods are used to complete time-series knowledge graphs for specific problems to achieve predictive maintenance, their research has made certain contributions to the field of predictive maintenance. However, the challenges of existing research include: (1) The correlation and temporal coupling characteristics of heterogeneous data have not been fully modeled. Existing research focuses on the static analysis of a single data source, which makes it difficult to capture the dynamic correlation between mechanical wear of components, electrical parameter drift and operation and maintenance events. (2) At present, elevator maintenance is mostly based on specifications and standardization requirements, and there is a lack of research on precise analysis of elevator components and optimization of maintenance cycles. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention proposes a predictive maintenance method based on elevator operation and maintenance time-series knowledge graphs. This method combines the prediction results of the time-series knowledge graph with aging and fault-increasing factors and introduces them into the failure rate model, updating the dynamic weights of the failure rate to reflect its evolution over time, thereby describing the dynamic changes in the performance recovery degree and aging trend of the equipment after maintenance. Furthermore, by comprehensively considering labor and time costs in the maintenance process, a comprehensive cost model for elevator components is established. Subsequently, a continuous-time Markov chain simulation method is used to iteratively solve the failure rate model and the comprehensive cost model, aiming at minimum cost and maximum utilization rate, to obtain the optimal maintenance cycle and maintenance frequency of the components during their service life. Finally, this method is applied to calculate the optimal maintenance cycle of elevator components.

[0006] This invention provides the following technical solution:

[0007] A predictive maintenance method based on elevator operation and maintenance time-series knowledge graph, the method comprising the following steps:

[0008] The first step is to construct a knowledge graph of elevator operation and maintenance time sequence, as follows:

[0009] Step (1.1) Data Sources: The data sources for constructing the knowledge graph in this invention include the elevator IoT platform and the elevator control cabinet, wherein the elevator IoT platform collects data through an open API interface;

[0010] Step (1.2) Data structure classification;

[0011] Step (1.3) Structured data processing;

[0012] For structured data, it is stored in a relational database because the entities and relations are clearly defined, and the knowledge graph triple conversion is performed directly in the subsequent step (1.5).

[0013] Step (1.4) Unstructured data relationship and entity extraction processing;

[0014] Step (1.5) Use graph database language to import data and draw an elevator operation and maintenance time sequence knowledge graph in the graph database;

[0015] The second step involves optimizing the maintenance cycle of elevator components to achieve predictive maintenance of electromechanical equipment components. The process is as follows:

[0016] Step (2.1) constructs a theoretical distribution model based on the Weibull distribution. The core function of the Weibull distribution is F(t). i ), ρ, λ(t) i );

[0017] Among them, the Weibull distribution is a statistical model widely used in failure analysis, prediction, and equipment aging reliability assessment, F(t) i ) represents the distribution function, ρ represents the linear correlation, and λ(t) represents the linear correlation function. i ) is the failure rate function, t i This represents the failure time, where i represents the i-th failure level.

[0018] Step (2.2) uses the least squares method to fit the parameters of the Weibull distribution to ensure that the model can accurately reflect the failure law of elevator components;

[0019] Least squares is a commonly used parameter estimation method, mainly used to minimize the sum of squared errors between model predictions and actual observations.

[0020] Step (2.3) calculates the recursive relationship of the component failure rate function within N maintenance cycles by using the inference results of the elevator operation and maintenance time sequence knowledge graph through graph convolution model and Hawkes process model;

[0021] Step (2.4) constructs a simulation model of the comprehensive maintenance cost of elevator components. This invention calculates the maintenance cost of elevator components from the following three aspects: the maintenance cost before failure C b Repair costs after a malfunction (C) a And downtime costs due to malfunctions C f ;

[0022] Step (2.5) uses the continuous-time Markov chain method to calculate the optimal maintenance cycle of elevator components;

[0023] Among them, the continuous-time Markov chain method is a commonly used algorithm in statistics. It controls the number of simulations through the state transition matrix and analyzes the impact of maintenance frequency on utilization and total cost.

[0024] Furthermore, the process of step (1.2) is as follows:

[0025] Step (1.2.1) The structured data comes partly from the structured operation data in the elevator control cabinet, including vibration signals, fault codes, operation status data, operation data, electrical parameters, temperature, and humidity; the other part comes from the time series data of car vibration data, elevator running speed, and elevator traction machine voltage, current, and car temperature provided by the elevator Internet of Things platform.

[0026] Step (1.2.2) Unstructured data includes maintenance records, repair records, fault causes, and fault phenomena, which are text composed of natural language.

[0027] Furthermore, the process of step (1.4) is as follows:

[0028] Step (1.4.1) uses Word2vec (Word to Vector), Bi-LSTM (Bidirectional Long Short-Term Memory), and CRF (Conditional Random Field) techniques to extract entities and relationships from unstructured text data;

[0029] Step (1.4.2) Relation extraction: This invention uses dependency syntax provided by the LTP platform;

[0030] 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.

[0031] Preferably, the process of step (1.4.1) is as follows:

[0032] Step (1.4.1.1) Text data embedding:

[0033] Text data needs to be vectorized, mapping the corresponding character sequence to a word vector sequence. Through the word embedding method Word2vec, the word embedding lookup table is continuously trained, learned, and updated. Finally, the input text is matched through the lookup table and converted into the corresponding word vector sequence.

[0034] Word2vec is a commonly used algorithm for natural language processing. It generates word vectors by training a neural network model, mapping words to a continuous vector space and capturing the semantic and grammatical relationships between words.

[0035] Step (1.4.1.2) The text embedding vector is processed by a Bi-LSTM layer;

[0036] First, the initial parameters used in the Bi-LSTM layer are randomly initialized. The text embedding vector sequence output by the embedding layer is input into the forward LSTM layer and the backward LSTM layer model to learn the context information of the sequence vector. Finally, the output vectors of the two directions are concatenated as the output sequence of the Bi-LSTM layer, and the dropout technique is introduced to prevent overfitting in the experiment.

[0037] Dropout is a commonly used regularization technique that prevents overfitting by randomly discarding neurons. LSTM is a commonly used recurrent neural network. It controls the storage and retrieval of information by introducing memory units and gating mechanisms, thereby maintaining and updating information over long time spans. Bi-LSTM is an extended version of LSTM, which captures bidirectional contextual information of sequence data by using two LSTMs in the forward and backward directions of the time series.

[0038] Step (1.4.1.3) The output vector of the Bi-LSTM layer is processed by the CRF layer;

[0039] The CRF layer considers the relationship between adjacent labels to obtain the globally optimal sequence, and models the dependencies between entities based on the output of the Bi-LSTM layer, selecting the sequence with the highest probability from the label space;

[0040] CRF is a probabilistic graphical model used for labeling and segmenting sequence data. It can take into account contextual information and global features, thereby improving labeling accuracy.

[0041] Step (1.4.1.4) calculates the loss function output by the model to obtain the optimal result;

[0042] The output of the CRF layer is used as the final labeled sequence of the model, and the loss function is calculated by comparing it with the actual labeled sequence.

[0043]

[0044] Where P(Y|X) represents the probability of output sequence Y given input sequence X, s(X,Y′) represents the score function given input sequence X and label sequence Y′, log is the logarithm operation, and exp is the natural exponentiation operation.

[0045] By continuously adjusting the parameters of each layer to minimize the loss value, the optimal recognition model is obtained.

[0046] The process of step (1.5) is as follows:

[0047] Step (1.5.1) Structured data format conversion is as follows:

[0048] The structured data processed in step (1.3) is exported from the relational database as a CSV file. To express the temporal characteristics of the process state data, the traditional triplet format (e... h ,r,e t ) is expanded into a quadruple with temporal information (e h ,r,e t,t), where t represents the time information when the triple is established. However, currently graph databases only support triple storage. Therefore, this invention transforms the time sequence information and embeds it into the relationship between entities to realize the establishment of an elevator operation and maintenance time sequence knowledge graph based on graph database.

[0049] In data structures and databases, a triple refers to a set containing three elements, often used to represent entity relationships in relational databases.

[0050] Step (1.5.2) Unstructured data format conversion: Convert the entities and relationships extracted from the unstructured data in step (1.4) into CSV file format;

[0051] Step (1.5.3) Since the extracted unstructured data is all text, this invention converts the CSV file encoding format to UTF-8 encoding to establish a standardized data entry process;

[0052] Step (1.5.4) categorizes the preprocessed entity-relationship data (stored in CSV format) by type and deploys it in the default resource directory of the graph database;

[0053] Step (1.5.5) Load the CSV file in the resource directory, specify the first row as the header, and use each row of data as a line variable.

[0054] Step (1.5.6) merges nodes and relationships as follows:

[0055] Step (1.5.6.1) Create or merge a node named elevator:

[0056] t0: evelator:

[0057] Step (1.5.6.2) creates or merges a node named registerCode with the name attribute set to line.registerCode:

[0058] t1: registerCode:

[0059] Step (1.5.6.3) Create / merge a relational connection t1 and t0 named Elevator Code;

[0060] Step (1.5.6.4) creates or merges an information node, whose attributes are assigned values ​​by the corresponding fields in the line:

[0061] t2: information:

[0062] Step (1.5.6.5) Create / merge a relational connection t1 and t2 named Details;

[0063] Step (1.5.7) defines the field mapping, where the attribute values ​​of t2 come from the corresponding fields in the csv.

[0064] Step (1.5.8) constructs a data pipeline, traverses each row of data in the CSV file, and writes each row of data into the graph database through the graph traversal interface provided by the graph database to generate the corresponding entities and relations, thereby realizing the schema alignment and import of triple data.

[0065] Step (1.5.9) imports all the timestamp information of each row of data in the CSV file into the graph database. This timestamp information is used as time sequence information to establish a time relationship with the corresponding entities, thereby realizing the construction of the elevator operation and maintenance time sequence knowledge graph.

[0066] The process of step (2.2) is as follows:

[0067] Step (2.2.1) Slope estimate of the fitted line and the estimated intercept of the fitted line The calculation process is as follows:

[0068]

[0069] Where, x i The logarithmic time is the natural logarithm of the failure time, y i It is a double log-probability, i.e., the distribution function F(t) i The value after double natural logarithmic transformation, and x i With y i The average value.

[0070] Step (2.2.2) where the logarithmic time x i And the double logarithmic probability y i The calculation process is as follows:

[0071]

[0072] Step (2.2.3) F(t) i The calculation process is as follows:

[0073]

[0074] Where η is the scale parameter and β is the shape parameter.

[0075] The method for calculating the shape parameter β in the Weibull distribution in step (2.2.4) is as follows:

[0076]

[0077] The method for calculating the scale parameter η in the Weibull distribution in step (2.2.5) is as follows:

[0078]

[0079] The method for calculating the linear correlation ρ in step (2.2.6) is as follows:

[0080]

[0081] Here, ρ represents the linear correlation between the fitted data. The closer ρ is to 1, the higher the linear correlation between the data.

[0082] The process of step (2.3) is as follows:

[0083] Step (2.3.1) uses a graph convolution model to realize the dynamic embedding of the temporal knowledge graph of electromechanical equipment;

[0084] Use a dynamic graph convolution model to obtain the entity embedding vector H of the temporal knowledge graph at time t. e ;

[0085] The entity at time t is embedded based on a dynamic graph convolution model;

[0086]

[0087] in, Let t be the hidden state of entity e. It is the set of entities e connected through relation r at time t, i.e. The normalization constant is This represents the hidden state of entity e at time t-1. Here, H0 represents the manually defined diagonal relation weight matrix for relation r at time t-1, where H0 represents the hidden state embedded in the entity's initial state, and ReLU is used as the activation function. is a trainable parameter matrix.

[0088] Step (2.3.2) implements entity prediction using the Hawkes process model;

[0089] Step (2.3.3) Dynamically adjust the anti-aging factor α i And increase factor b i This enables the model to reflect the fault behavior of components under actual operating conditions.

[0090]

[0091] in, K is the initial age-reduction factor, k1 is the adjustment coefficient, and the effect of the balancing inference results on the age-reduction factor is... k1 is the initial increment factor, and k2 is the adjustment coefficient used to control the weight of the reasoning results;

[0092] Step (2.3.4) The initial failure rate function λ1 (within the first maintenance cycle) is expressed as:

[0093]

[0094] Step (2.3.5) After the first maintenance time T (within the second maintenance cycle), following the maintenance operation, the component's age regresses, and the failure rate becomes:

[0095] λ2=b1λ1(t2+T-α1T),t2∈[T,2T]

[0096] In step (2.3.6), 2T after the second maintenance time (within the third maintenance cycle), the failure rate function further changes as follows:

[0097] λ3=b2λ2(t3+T-α2T),t3∈[2T,3T]

[0098] In step (2.3.7), the failure rate function during the j-th maintenance cycle is:

[0099] λ j =b j-1 λ j-1 (t j +T-α j-1 T), t j ∈[(j-1)T,jT]

[0100] Step (2.3.8) states that within a finite usage time, the failure rate function for the Nth cycle is:

[0101]

[0102] Preferably, the process of step (2.3.2) is as follows:

[0103] Step (2.3.2.1) constructs intensity functions to achieve entity prediction and temporal prediction. The temporal conditional intensity function constructed for all entities is as follows:

[0104]

[0105] Among them, W λ It is a projection matrix. At query time t q The hidden state, This represents the embedding of the query relation, and <, ·> represents the matrix inner product function. It is a softplus function with parameter β, the purpose of which is to ensure the positive value of the intensity function.

[0106] Step (2.3.2.2) predicts the faulty entity e(t) that occurs at query time by maximizing the conditional strength function. q The probability τ of ) i ;

[0107]

[0108] Here, softmax is a common activation function for multi-class classification problems, used to calculate probabilities, and arg max represents finding the activation function that makes the probability of a class class equal to or greater than the probability of a class class class equal to or greater than the probability of a class class class class equal to or greater than the probability of a class class class class class equal to the probability of a class ... The predicted largest failure entity;

[0109] Step (2.3.2.3) uses the conditional time density function to predict the failure time offset Δt, which indicates the trend of earlier or later occurrence of the faulty entity. i ;

[0110]

[0111] Where τ is a time parameter used for integration.

[0112] The process of step (2.4) is as follows:

[0113] Step (2.4.1) Maintenance cost C b The calculation is as follows:

[0114] Step (2.4.1.1) In this invention, the maintenance cost before the failure is divided into labor cost C. p and material costs C s Two parts. Among them, labor costs are linearly related to pre-failure maintenance time:

[0115] C p (i)=Pt b (i)

[0116] Among them, t b This indicates the duration of a single maintenance operation, and P represents the unit labor cost coefficient.

[0117] The maintenance duration in step (2.4.1.2) will dynamically change with the maintenance frequency, which can be expressed as:

[0118]

[0119] In this function, e represents the technician's experience level, and r reflects the intensity of the experience curve effect.

[0120] Step (2.4.1.3) cumulative pre-failure maintenance cost C b Integrated into:

[0121]

[0122] Step (2.4.2) Maintenance Cost C a The calculation is as follows:

[0123] The number of faults occurring within the maintenance cycle [(j-1)T,jT] in step (2.4.2.1) can be modeled as follows:

[0124]

[0125] The cumulative number of failures of the part in step (2.4.2.2) during the entire maintenance cycle is:

[0126]

[0127] Step (2.4.2.3) combines the average cost per repair (c) o The total maintenance cost is:

[0128] C a =c o F total

[0129] Step (2.4.3) Downtime cost C f The calculation is as follows;

[0130] The downtime in step (2.4.3.1) includes the duration t of a single maintenance operation. b and single fault handling time t a The downtime loss per unit time is c. f The total cost of downtime losses consists of the downtime cost per unit time, c. f Total maintenance time T b The total time T spent on repairs a A decision is expressed as:

[0131] C f =c f (T b +T a )

[0132] The total time T spent on maintenance in step (2.4.3.2) b The total time T spent on repairs a It is expressed as follows:

[0133]

[0134] Step (2.4.4) Total cost C required for maintenance of each component total Represented as:

[0135] C total =C b +C a +C f .

[0136] The process of step (2.5) is as follows:

[0137] Step (2.5.1) defines the state space and transition rate matrix as follows;

[0138] The state space S for step (2.5.1.1) is defined as follows:

[0139] S={(j,k)∣j∈[1,2,…,N],k∈[0,1]}

[0140] Where j∈[1,2,…,N] represents the j-th maintenance cycle, and k∈[0,1] represents the component status (0 = normal, 1 = failure);

[0141] The transfer and repair rate matrix Q in step (2.5.1.2) is defined as follows:

[0142] Transition rate from normal state to failure state:

[0143] Q fault =λ j

[0144] Repair rate from failure state to normal state:

[0145]

[0146] Step (2.5.2) Steady-state probability calculation: For each state (j,k), the steady-state probability equation is satisfied:

[0147] ∑π((j′,k′))q {(j′,k′)→(j,k)} =π((j,k))∑q {(j,k)→(j′,k′)}

[0148] Where (j,k) represents the current state, (j′,k′) represents the changed state, and π(·) is the steady-state probability.

[0149] Step (2.5.3) Calculate the number of failures N within the maintenance cycle. fault Total maintenance time T maintenance The calculation method is as follows:

[0150]

[0151] Step (2.5.4) combines time and cost to iteratively optimize the number of maintenance times N. When the minimum cost assessment index E(N) and the maximum utilization rate A(N) are reached, the iteration stops to obtain the optimal maintenance cycle T.

[0152] Preferably, the process of step (2.5.4) is as follows:

[0153] The total cost assessment indicator E(N) for step (2.5.4.1) is:

[0154]

[0155] Among them, T total Total time used for the component;

[0156] The utilization rate and effective service life of components in step (2.5.4.2) can be expressed as:

[0157] T available =T total -T b -T a

[0158]

[0159] The utilization rate A(N) is represented by the ratio of the effective usage time of a component to the total usage time.

[0160] In step (2.5.4.3), the number of repairs N is the decision variable, and the relationship between the maintenance cycle T and the number of repairs N is expressed as follows:

[0161] T = [int(T total -T ρ ) / N)]

[0162] Here, int is the integer function, ensuring that the period is an integer.

[0163] The beneficial effects of this invention are mainly reflected in:

[0164] 1. Construct a quadruple time-series knowledge graph to integrate structured and unstructured data, solve the problem of insufficient time-series correlation modeling in traditional methods, and improve the accuracy of fault prediction.

[0165] 2. By combining the Weibull distribution, aging factor and failure factor to update the failure rate model, and solving iteratively through a continuous-time Markov chain, the maintenance cycle and frequency can be dynamically optimized, reducing maintenance redundancy or lag.

[0166] 3. Establish a comprehensive model that includes maintenance, repair, and downtime costs, quantify cost and utilization indicators, and achieve maintenance decisions with the goal of minimizing costs, combining economic efficiency and engineering practicality. Attached Figure Description

[0167] Figure 1 This is a flowchart of a predictive maintenance method based on elevator operation and maintenance time sequence knowledge graph. Detailed Implementation

[0168] The present invention will now be further described with reference to the accompanying drawings.

[0169] Reference Figure 1 A predictive maintenance method based on elevator operation and maintenance time-series knowledge graph, the method comprising the following steps:

[0170] The first step is to construct a knowledge graph of elevator operation and maintenance time sequence, as follows:

[0171] Step (1.1) Data Sources: The data sources for constructing the knowledge graph in this invention include the elevator IoT platform and the elevator control cabinet, wherein the elevator IoT platform collects data through an open API interface;

[0172] Step (1.2) Data structure classification, the process is as follows:

[0173] Step (1.2.1) The structured data comes partly from the structured operation data in the elevator control cabinet, including vibration signals, fault codes, operation status data, operation data, electrical parameters, temperature, and humidity; the other part comes from the time series data of car vibration data, elevator running speed, and elevator traction machine voltage, current, and car temperature provided by the elevator Internet of Things platform.

[0174] Step (1.2.2) Unstructured data includes maintenance records, repair records, fault causes, and fault phenomena, which are text composed of natural language.

[0175] Step (1.3) Structured data processing;

[0176] For structured data, it is stored in a relational database. Because the entities and relations are clearly defined, the knowledge graph triple conversion is performed directly in the subsequent step (1.5).

[0177] Step (1.4) Unstructured data relationship and entity extraction processing, the process is as follows:

[0178] Step (1.4.1) uses Word2vec (Word to Vector), Bi-LSTM (Bidirectional Long Short-Term Memory), and CRF (Conditional Random Field) techniques to extract entities and relationships from unstructured text data; the process is as follows:

[0179] Step (1.4.1.1) Text data embedding:

[0180] Text data needs to be vectorized to map the corresponding character sequence to a word vector sequence. This invention uses the Word2vec word embedding method to continuously train, learn and update the word embedding lookup table, and finally matches the input text through the lookup table to convert it into the corresponding word vector sequence.

[0181] Word2vec is a commonly used algorithm for natural language processing. It generates word vectors by training a neural network model, mapping words to a continuous vector space and capturing the semantic and grammatical relationships between words.

[0182] Step (1.4.1.2) The text embedding vector is processed by a Bi-LSTM layer;

[0183] First, the initial parameters used in the Bi-LSTM layer are randomly initialized. The text embedding vector sequence output by the embedding layer is input into the forward LSTM layer and the backward LSTM layer model to learn the context information of the sequence vector. Finally, the output vectors of the two directions are concatenated as the output sequence of the Bi-LSTM layer, and the dropout technique is introduced to prevent overfitting in the experiment.

[0184] Dropout is a commonly used regularization technique that prevents overfitting by randomly discarding neurons. LSTM is a commonly used recurrent neural network. It controls the storage and retrieval of information by introducing memory units and gating mechanisms, thereby maintaining and updating information over long time spans. Bi-LSTM is an extended version of LSTM, which captures bidirectional contextual information of sequence data by using two LSTMs in the forward and backward directions of the time series.

[0185] Step (1.4.1.3) The output vector of the Bi-LSTM layer is processed by the CRF layer;

[0186] The CRF layer considers the relationship between adjacent labels to obtain the globally optimal sequence, and models the dependencies between entities based on the output of the Bi-LSTM layer, selecting the sequence with the highest probability from the label space;

[0187] CRF is a probabilistic graphical model used for labeling and segmenting sequence data. It can take into account contextual information and global features, thereby improving labeling accuracy.

[0188] Step (1.4.1.4) calculates the loss function output by the model to obtain the optimal result;

[0189] The output of the CRF layer is used as the final labeled sequence of the model, and the loss function is calculated by comparing it with the actual labeled sequence.

[0190]

[0191] Where P(Y|X) represents the probability of output sequence Y given input sequence X, s(X,Y′) represents the score function given input sequence X and label sequence Y′, log is the logarithm operation, and exp is the natural exponentiation operation.

[0192] By continuously adjusting the parameters of each layer to minimize the loss value, the optimal recognition model is obtained.

[0193] Step (1.4.2) Relation extraction: This invention uses dependency syntax provided by the LTP platform;

[0194] 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.

[0195] Step (1.5) involves importing data using a graph database language and drawing an elevator operation and maintenance time-series knowledge graph in the graph database; the process is as follows:

[0196] Step (1.5.1): Structured data format conversion;

[0197] The structured data processed in step (1.3) is exported from the relational database as a CSV file. To express the temporal characteristics of the process state data, the traditional triplet format (e... h ,r,e t ) is expanded into a quadruple with temporal information (e h ,r,e t The triple is defined as follows: (t) where t represents the time information when the triple is formed. However, current graph databases only support triple storage. Therefore, this invention transforms and embeds the time-series information into the relationships between entities to establish a time-series knowledge graph for elevator operation and maintenance based on a graph database.

[0198] In data structures and databases, a triple refers to a set containing three elements, often used to represent entity relationships in relational databases.

[0199] Step (1.5.2) Unstructured data format conversion;

[0200] Convert the entities and relationships extracted from the unstructured data in step (1.4) into CSV file format;

[0201] Step (1.5.3) Since the extracted unstructured data is all text, this invention converts the CSV file encoding format to UTF-8 encoding to establish a standardized data entry process;

[0202] Step (1.5.4) categorizes the preprocessed entity-relationship data (stored in CSV format) by type and deploys it in the default resource directory of the graph database;

[0203] Step (1.5.5) Load the CSV file in the resource directory, specify the first row as the header, and use each row of data as a line variable.

[0204] Step (1.5.6) merges nodes and relationships as follows:

[0205] Step (1.5.6.1) Create or merge a node named elevator:

[0206] t0: evelator:

[0207] Step (1.5.6.2) creates or merges a node named registerCode with the name attribute set to line.registerCode:

[0208] t1: registerCode:

[0209] Step (1.5.6.3) Create / merge a relational connection t1 and t0 named Elevator Code;

[0210] Step (1.5.6.4) creates or merges an information node, whose attributes are assigned values ​​by the corresponding fields in the line:

[0211] t2: information:

[0212] Step (1.5.6.5) Create / merge a relational connection t1 and t2 named Details;

[0213] Step (1.5.7) defines the field mapping, where the attribute values ​​of t2 come from the corresponding fields in the csv.

[0214] Step (1.5.8) constructs a data pipeline, traverses each row of data in the CSV file, and writes each row of data into the graph database through the graph traversal interface provided by the graph database to generate the corresponding entities and relations, thereby realizing the schema alignment and import of triple data.

[0215] Step (1.5.9) imports all the timestamp information of each row of data in the CSV file into the graph database. This timestamp information is used as time sequence information to establish a time relationship with the corresponding entities, thereby realizing the construction of the elevator operation and maintenance time sequence knowledge graph.

[0216] The second step involves optimizing the maintenance cycle of elevator components to achieve predictive maintenance of electromechanical equipment components. The process is as follows:

[0217] Step (2.1) constructs a theoretical distribution model based on the Weibull distribution. The core function of the Weibull distribution is F(t). i ), ρ, λ(t) i ).

[0218] The Weibull distribution is a statistical model widely used in failure analysis, prediction, and equipment aging reliability assessment. F(t)i ) represents the distribution function, ρ represents the linear correlation, and λ(t) represents the linear correlation function. i ) is the failure rate function, t i This represents the failure time, where i represents the i-th failure level.

[0219] Step (2.2) uses the least squares method to fit the parameters of the Weibull distribution to ensure that the model can accurately reflect the failure law of elevator components.

[0220] Least squares is a commonly used parameter estimation method, mainly used to minimize the sum of squared errors between model predictions and actual observations.

[0221] The process of step (2.2) is as follows:

[0222] Step (2.2.1) Slope estimate of the fitted line and the estimated intercept of the fitted line The calculation process is as follows:

[0223]

[0224]

[0225] Where, x i The logarithmic time is the natural logarithm of the failure time, y i It is a double log-probability, i.e., the distribution function F(t) i The value after double natural logarithmic transformation, and x i With y i The average value.

[0226] Step (2.2.2) where the logarithmic time x i And the double logarithmic probability y i The calculation process is as follows:

[0227]

[0228] Step (2.2.3) F(t) i The calculation process is as follows:

[0229]

[0230] Where η is the scale parameter and β is the shape parameter.

[0231] The method for calculating the shape parameter β in the Weibull distribution in step (2.2.4) is as follows:

[0232]

[0233] The method for calculating the scale parameter η in the Weibull distribution in step (2.2.5) is as follows:

[0234]

[0235] The method for calculating the linear correlation ρ in step (2.2.6) is as follows:

[0236]

[0237] Here, ρ represents the linear correlation between the fitted data. The closer ρ is to 1, the higher the linear correlation between the data.

[0238] Step (2.3) involves calculating the elevator operation and maintenance time-series knowledge graph inference results using a graph convolution model and a Hawkes process model, and then performing a recursive calculation of the component failure rate function over N maintenance cycles. The process is as follows:

[0239] Step (2.3.1) uses a graph convolution model to realize the dynamic embedding of the temporal knowledge graph of electromechanical equipment;

[0240] Use a dynamic graph convolution model to obtain the entity embedding vector H of the temporal knowledge graph at time t. e ;

[0241] The entity at time t is embedded based on a dynamic graph convolution model;

[0242]

[0243] in, Let t be the hidden state of entity e. It is the set of entities e connected through relation r at time t, i.e. The normalization constant is This represents the hidden state of entity e at time t-1. Here, H0 represents the manually defined diagonal relation weight matrix for relation r at time t-1, where H0 represents the hidden state embedded in the entity's initial state, and ReLU is used as the activation function. is a trainable parameter matrix.

[0244] Step (2.3.2) implements entity prediction using the Hawkes process model. The process is as follows:

[0245] Step (2.3.2.1) constructs intensity functions to achieve entity prediction and temporal prediction. The temporal conditional intensity function constructed for all entities is as follows:

[0246]

[0247] Among them, W λ It is a projection matrix. At query time t qThe hidden state, This represents the embedding of the query relation, and <, ·> represents the matrix inner product function. It is a softplus function with parameter β, the purpose of which is to ensure the positive value of the intensity function.

[0248] Step (2.3.2.2) predicts the faulty entity e(t) that occurs at query time by maximizing the conditional strength function. q The probability τ of ) i ;

[0249]

[0250] Here, softmax is a common activation function for multi-class classification problems, used to calculate probabilities, and arg max represents finding the activation function that makes the probability of a class class equal to or greater than the probability of a class class class equal to or greater than the probability of a class class class class equal to or greater than the probability of a class class class class class equal to the probability of a class ... The predicted largest failure entity;

[0251] Step (2.3.2.3) uses the conditional time density function to predict the failure time offset Δt, which indicates the trend of earlier or later occurrence of the faulty entity. i ;

[0252]

[0253] Where τ is a time parameter used for integration;

[0254] Step (2.3.3) Dynamically adjust the anti-aging factor α i And increase factor b i This enables the model to reflect the fault behavior of components under actual operating conditions.

[0255]

[0256] in, K is the initial age-reduction factor, k1 is the adjustment coefficient, and the effect of the balancing inference results on the age-reduction factor is... k1 is the initial increment factor, and k2 is the adjustment coefficient used to control the weight of the inference results.

[0257] Step (2.3.4) The initial failure rate function λ1 (within the first maintenance cycle) is expressed as:

[0258]

[0259] Step (2.3.5) After the first maintenance time T (within the second maintenance cycle), following the maintenance operation, the component's age regresses, and the failure rate becomes:

[0260] λ2=b1λ1(t2+T-α1T),t2∈[T,2T]

[0261] In step (2.3.6), 2T after the second maintenance time (within the third maintenance cycle), the failure rate function further changes as follows:

[0262] λ3=b2λ2(t3+T-α2T),t3∈[2T,3T]

[0263] In step (2.3.7), the failure rate function during the j-th maintenance cycle is:

[0264] λ j =b j-1 λ j-1 (t j +T-α j-1 T), t j ∈[(j-1)T,jT]

[0265] Step (2.3.8) states that within a finite usage time, the failure rate function for the Nth cycle is:

[0266]

[0267] Step (2.4) constructs a simulation model of the comprehensive maintenance cost of elevator components. This invention calculates the maintenance cost of elevator components from the following three aspects: the maintenance cost before failure C b Repair costs after a malfunction (C) a And downtime costs due to malfunctions C f The process is as follows:

[0268] Step (2.4.1) Maintenance cost C b The calculation is as follows:

[0269] Step (2.4.1.1) In this invention, the maintenance cost before the failure is divided into labor cost C. p and material costs C s Two parts. Among them, labor costs are linearly related to pre-failure maintenance time:

[0270] C p (i)=Pt b (i)

[0271] Among them, t b This indicates the duration of a single maintenance operation, and P represents the unit labor cost coefficient.

[0272] The maintenance duration in step (2.4.1.2) will dynamically change with the maintenance frequency, which can be expressed as:

[0273]

[0274] In this function, e represents the technician's experience level, and r reflects the intensity of the experience curve effect.

[0275] Step (2.4.1.3) cumulative pre-failure maintenance cost C b Integrated into:

[0276]

[0277] Step (2.4.2) Maintenance Cost C a The calculation is as follows:

[0278] The number of faults occurring within the maintenance cycle [(j-1)T,jT] in step (2.4.2.1) can be modeled as follows:

[0279]

[0280] The cumulative number of failures of the part in step (2.4.2.2) during the entire maintenance cycle is:

[0281]

[0282] Step (2.4.2.3) combines the average cost per repair (c) o The total maintenance cost is:

[0283] C a =c o F total

[0284] Step (2.4.3) Downtime cost C f The calculation is as follows:

[0285] The downtime in step (2.4.3.1) includes the duration t of a single maintenance operation. b and single fault handling time t a The downtime loss per unit time is c. f The total cost of downtime losses consists of the downtime cost per unit time, c. f Total maintenance time T b The total time T spent on repairs a A decision is expressed as:

[0286] C f =c f (T b +T a )

[0287] The total time T spent on maintenance in step (2.4.3.2) b The total time T spent on repairs a It is expressed as follows:

[0288]

[0289] Step (2.4.4) Total cost C required for maintenance of each component totalIt can be represented as:

[0290] C total =C b +C a +C f

[0291] Step (2.5) uses the continuous-time Markov chain method to calculate the optimal maintenance cycle of elevator components;

[0292] Among them, the continuous-time Markov chain method is a commonly used algorithm in statistics. It controls the number of simulations through the state transition matrix and analyzes the impact of maintenance frequency on utilization and total cost.

[0293] The process of step (2.5) is as follows:

[0294] Step (2.5.1) defines the state space and transition rate matrix, the process is as follows:

[0295] The state space S for step (2.5.1.1) is defined as follows:

[0296] S={(j,k)∣j∈[1,2,…,N],k∈[0,1]}

[0297] Where j∈[1,2,…,N] represents the j-th maintenance cycle, and k∈[0,1] represents the component status (0 = normal, 1 = failure).

[0298] The transfer and repair rate matrix Q in step (2.5.1.2) is defined as follows:

[0299] Transition rate from normal state to failure state:

[0300] Q fault =λ j

[0301] Repair rate from failure state to normal state:

[0302]

[0303] Step (2.5.2) Steady-state probability calculation: For each state (j,k), the steady-state probability equation is satisfied:

[0304] ∑π((j′,k′))q {(j′,k′)→(j,k)} =π((j,k))∑q {(j,k)→(j′,k′)}

[0305] Where (j,k) represents the current state, (j′,k′) represents the changed state, and π(·) is the steady-state probability.

[0306] Step (2.5.3) Calculate the number of failures N within the maintenance cycle. faultTotal maintenance time T maintenance The calculation method is as follows:

[0307]

[0308] Step (2.5.4) iteratively optimizes the number of maintenance operations N by combining time and cost. The iteration stops when the minimum cost assessment index E(N) and the maximum utilization rate A(N) are reached, yielding the optimal maintenance cycle T. The process is as follows:

[0309] The total cost assessment indicator E(N) for step (2.5.4.1) is:

[0310]

[0311] Among them, T total Total time used for the component;

[0312] The utilization rate and effective service life of components in step (2.5.4.2) can be expressed as:

[0313] T available =T total -T b -T a

[0314]

[0315] The utilization rate A(N) is represented by the ratio of the effective usage time of a component to the total usage time.

[0316] In step (2.5.4.3), the number of repairs N is the decision variable, and the relationship between the maintenance cycle T and the number of repairs N is expressed as follows:

[0317] T = [int(T total -T ρ ) / N)]

[0318] Here, `int` is the integer function, ensuring that the period is an integer.

[0319] The processing procedure for application examples of this invention is as follows:

[0320] Step 1: Introduction to the experimental dataset. The experimental dataset is provided by the elevator operation and maintenance time-series knowledge graph constructed by the method proposed in this invention in combination with actual elevator data, containing 15,356 entities and semantic relationships.

[0321] Step two: Define evaluation metrics. To verify the predictive performance of the algorithm, this invention uses the following evaluation metrics to assess the algorithm:

[0322] Cost assessment metric: This metric measures the ratio between total cost and total usage time, corresponding to step (2.5.4.1).

[0323]

[0324] Utilization rate: This indicator measures the ratio between the effective usage time of a component and the total usage time, corresponding to step (2.5.4.2).

[0325]

[0326] Step 3: Analysis and comparison of results. This invention analyzed the fault record data of the four components with the highest elevator failure rates in the graph. Specific component failure statistics are shown in Table 1.

[0327] Table 1 is a statistical table of the number of failures and failure rate;

[0328]

[0329] Statistical analysis of component performance revealed a significant concentration of failures in the car door system, with a failure rate of 22.6%. In contrast, the drive motor demonstrated superior operational stability, maintaining a failure rate below industry benchmarks. While individual failure probabilities varied among other functional modules, they generally fell within a similar range of 10%.

[0330] Table 2 shows the parameter estimation results for commonly used elevator components based on the failure rate model, and Table 3 shows the cost parameter statistics. Experimental results show that the failure distribution of each component follows a Weibull distribution during its service life. When using linear regression to fit the distribution of discrete failure data, the optimal solutions for the shape parameter β and scale parameter η are obtained through least squares estimators. The fitting results show that the failure data are distributed near the fitted line, and the fitting accuracy meets the requirements. It is particularly noteworthy that β = 1.85 for the car door indicates a significant wear-type failure process, while β = 0.97 for the motor is close to a random failure mode.

[0331] Table 2 shows the estimation results of the modeling parameters;

[0332]

[0333] Table 3 shows the statistical results of cost parameters;

[0334]

[0335] The optimization results of the maintenance cycle were obtained through iterative analysis using the continuous-time Markov chain method, as shown in Table 4. The experimental results demonstrate that the 37-day regular maintenance strategy effectively controls the failure frequency of elevator car doors in the current dataset, achieving optimal maintenance costs while ensuring equipment availability. This optimization method not only improves resource utilization efficiency but also reduces the systemic operational risks of elevators by optimizing the maintenance cycle, demonstrating the engineering value of optimized maintenance decisions.

[0336] Table 4 shows the results of the maintenance cycle optimization.

[0337]

[0338] 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 predictive maintenance method based on elevator operation and maintenance time-series knowledge graph, characterized in that, The method includes the following steps: The first step is to construct a knowledge graph of elevator operation and maintenance time sequence, as follows: The data sources for constructing the knowledge graph in step (1.1) include the elevator IoT platform and the elevator control cabinet, with the elevator IoT platform collecting data through an open API interface; Step (1.2) Data structure classification; Step (1.3) Structured data processing; For structured data, it is stored in a relational database because the entities and relations are clearly defined, and the knowledge graph triple conversion is performed directly in the subsequent step (1.5). Step (1.4) Unstructured data relationship and entity extraction processing; Step (1.5) Use graph database language to import data and draw an elevator operation and maintenance time sequence knowledge graph in the graph database; The second step involves optimizing the maintenance cycle of elevator components to achieve predictive maintenance of electromechanical equipment components. The process is as follows: Step (2.1) constructs a theoretical distribution model based on the Weibull distribution. The core function of the Weibull distribution is F(t). i ), ρ, λ(t) i ); Wherein, F(t) i ) represents the distribution function, ρ represents the linear correlation, and λ(t) represents the linear correlation function. i ) is the failure rate function, t i This represents the failure time, where i represents the i-th failure level. Step (2.2) uses the least squares method to fit the parameters of the Weibull distribution to ensure that the model can accurately reflect the failure law of elevator components; Step (2.3) calculates the recursive relationship of the component failure rate function within N maintenance cycles by using the inference results of the elevator operation and maintenance time sequence knowledge graph through graph convolution model and Hawkes process model; Step (2.4) constructs a comprehensive cost simulation model for elevator component maintenance, calculating the maintenance cost of elevator components from the following three aspects: maintenance cost before failure C b Repair costs after a malfunction (C) a And downtime costs due to malfunctions C f ; Step (2.5) uses the continuous-time Markov chain method to calculate the optimal maintenance cycle of elevator components.

2. The predictive maintenance method based on elevator operation and maintenance time-series knowledge graph as described in claim 1, characterized in that, The process of step (1.2) is as follows: Step (1.2.1) The structured data comes partly from the structured operation data in the elevator control cabinet, including vibration signals, fault codes, operation status data, operation data, electrical parameters, temperature, and humidity; the other part comes from the time series data of car vibration data, elevator running speed, and elevator traction machine voltage, current, and car temperature provided by the elevator Internet of Things platform. Step (1.2.2) Unstructured data includes maintenance records, repair records, fault causes, and fault phenomena, which are text composed of natural language.

3. The predictive maintenance method based on elevator operation and maintenance time-series knowledge graph as described in claim 1 or 2, characterized in that, The process of step (1.4) is as follows: Step (1.4.1) uses Word2vec, Bi-LSTM, and CRF techniques to extract entities and relations from unstructured text data; Step (1.4.2) involves relation extraction, using dependency syntax provided by the LTP platform.

4. The predictive maintenance method based on elevator operation and maintenance time-series knowledge graph as described in claim 3, characterized in that, The process of step (1.4.1) is as follows: Step (1.4.1.1) Text data embedding: Text data needs to be vectorized, mapping the corresponding character sequence to a word vector sequence. Through the word embedding method Word2vec, the word embedding lookup table is continuously trained, learned, and updated. Finally, the input text is matched through the lookup table and converted into the corresponding word vector sequence. Word2vec is a commonly used algorithm for natural language processing. It generates word vectors by training a neural network model, mapping words to a continuous vector space and capturing the semantic and grammatical relationships between words. Step (1.4.1.2) The text embedding vector is processed by a Bi-LSTM layer; First, the initial parameters used in the Bi-LSTM layer are randomly initialized. The text embedding vector sequence output by the embedding layer is input into the forward LSTM layer and the backward LSTM layer model to learn the context information of the sequence vector. Finally, the output vectors of the two directions are concatenated as the output sequence of the Bi-LSTM layer, and the dropout technique is introduced to prevent overfitting in the experiment. Step (1.4.1.3) The output vector of the Bi-LSTM layer is processed by the CRF layer; The CRF layer considers the relationship between adjacent labels to obtain the globally optimal sequence, and models the dependencies between entities based on the output of the Bi-LSTM layer, selecting the sequence with the highest probability from the label space; Step (1.4.1.4) calculates the loss function output by the model to obtain the optimal result; The output of the CRF layer is used as the final labeled sequence of the model, and the loss function is calculated by comparing it with the actual labeled sequence. Where P(Y|X) represents the probability of outputting sequence Y given input sequence X, s(X,Y′) represents the score function given input sequence X and label sequence Y′, log is the logarithm operation, and exp is the natural exponentiation operation; By continuously adjusting the parameters of each layer to minimize the loss value, the optimal recognition model is obtained.

5. The predictive maintenance method based on elevator operation and maintenance time-series knowledge graph as described in claim 1 or 2, characterized in that, The process of step (1.5) is as follows: Step (1.5.1) Structured data format conversion is as follows: The structured data processed in step (1.3) is exported from the relational database as a CSV file. To express the temporal characteristics of the process state data, the traditional triplet format (e... h ,r,e t ) is expanded into a quadruple (e) with temporal information. h ,r,e t ,t), where t represents the time information when the triple is established. However, currently graph databases only support triple storage. Therefore, this invention transforms the time sequence information and embeds it into the relationship between entities to realize the establishment of an elevator operation and maintenance time sequence knowledge graph based on graph database. Step (1.5.2) Unstructured data format conversion: Convert the entities and relationships extracted from the unstructured data in step (1.4) into CSV file format; Step (1.5.3) Since the extracted unstructured data is all text, this invention converts the CSV file encoding format to UTF-8 encoding to establish a standardized data entry process; Step (1.5.4) categorizes the preprocessed entity-relationship data (stored in CSV format) by type and deploys it in the default resource directory of the graph database; Step (1.5.5) Load the CSV file in the resource directory, specify the first row as the header, and use each row of data as a line variable; Step (1.5.6) merges nodes and relationships as follows: Step (1.5.6.1) Create or merge a node named elevator: t0: evelator: Step (1.5.6.2) creates or merges a node named registerCode with the name attribute set to line.registerCode: t1: registerCode: Step (1.5.6.3) Create / merge a relational connection t1 and t0 named Elevator Code; Step (1.5.6.4) creates or merges an information node, whose attributes are assigned values ​​by the corresponding fields in the line: t2: information: Step (1.5.6.5) Create / merge a relational connection t1 and t2 named Details; Step (1.5.7) defines the field mapping, where the attribute values ​​of t2 come from the corresponding fields in the CSV file; Step (1.5.8) constructs a data pipeline, traverses each row of data in the CSV file, and writes each row of data into the graph database through the graph traversal interface provided by the graph database to generate the corresponding entities and relations, thereby realizing the schema alignment and import of triple data; Step (1.5.9) imports all the timestamp information of each row of data in the CSV file into the graph database. This timestamp information is used as time sequence information to establish a time relationship with the corresponding entities, thereby realizing the construction of the elevator operation and maintenance time sequence knowledge graph.

6. The predictive maintenance method based on elevator operation and maintenance time-series knowledge graph as described in claim 1 or 2, characterized in that, The process of step (2.2) is as follows: Step (2.2.1) Estimated slope of the fitted line and the estimated intercept of the fitted line The calculation process is as follows: Where, x i The logarithmic time is the natural logarithm of the failure time, y i It is a double log-probability, i.e., the distribution function F(t) i The value after double natural logarithmic transformation, and x i With y i The average value; Step (2.2.2) where the logarithmic time x i And the double logarithmic probability y i The calculation process is as follows: x i =ln t i Step (2.2.3) F(t) i The calculation process is as follows: Where η is the scale parameter and β is the shape parameter; The method for calculating the shape parameter β in the Weibull distribution in step (2.2.4) is as follows: The method for calculating the scale parameter η in the Weibull distribution in step (2.2.5) is as follows: The method for calculating the linear correlation ρ in step (2.2.6) is as follows: Here, ρ represents the linear correlation between the fitted data. The closer ρ is to 1, the higher the linear correlation between the data.

7. The predictive maintenance method based on elevator operation and maintenance time-series knowledge graph as described in claim 1 or 2, characterized in that, The process of step (2.3) is as follows: Step (2.3.1) uses a graph convolution model to realize the dynamic embedding of the temporal knowledge graph of electromechanical equipment; Use a dynamic graph convolution model to obtain the entity embedding vector H of the temporal knowledge graph at time t. e ; The entity at time t is embedded based on a dynamic graph convolution model; in, Let t be the hidden state of entity e. It is the set of entities e connected through relation r at time t, i.e. The normalization constant is This represents the hidden state of entity e at time t-1. Here, H0 represents the manually defined diagonal relation weight matrix for relation r at time t-1, where H0 represents the hidden state embedded in the entity's initial state, and ReLU is used as the activation function. is a trainable parameter matrix; Step (2.3.2) implements entity prediction using the Hawkes process model; Step (2.3.3) Dynamically adjust the anti-aging factor α i And increase factor b i This enables the model to reflect the fault behavior of components under actual operating conditions. in, Let k1 be the initial age-reduction factor, k1 be the adjustment coefficient, and the effect of the balancing inference results on the age-reduction factor be... k1 is the initial increment factor, and k2 is the adjustment coefficient used to control the weight of the reasoning results; Step (2.3.4) The initial failure rate function λ1 is expressed as: Step (2.3.5) After the first maintenance time T (within the second maintenance cycle), following the maintenance operation, the component's age regresses, and the failure rate becomes: λ2=b1λ1(t2+T-α1T),t2∈[T,2T] In step (2.3.6), 2T after the second maintenance time (within the third maintenance cycle), the failure rate function further changes as follows: λ3=b2λ2(t3+T-α2T),t3∈[2T,3T] In step (2.3.7), the failure rate function during the j-th maintenance cycle is: λ j =b j-1 λ j-1 (t j +T-α j-1 T),t j ∈[(j-1)T,jT] Step (2.3.8) states that within a finite usage time, the failure rate function for the Nth cycle is:

8. The predictive maintenance method based on elevator operation and maintenance time-series knowledge graph as described in claim 7, characterized in that, The process of step (2.3.2) is as follows: Step (2.3.2.1) constructs intensity functions to achieve entity prediction and temporal prediction. The temporal conditional intensity function constructed for all entities is as follows: Among them, W λ It is a projection matrix. At query time t q The hidden state, This represents the embedding of the query relation, and <, ·> represents the matrix inner product function. It is a softplus function with parameter β, the purpose of which is to ensure the positive value of the intensity function; Step (2.3.2.2) predicts the faulty entity e(t) that occurs at query time by maximizing the conditional strength function. q The probability τ of ) i ; Here, softmax is a common activation function for multi-class classification problems, used to calculate probabilities, and arg max represents finding the activation function that makes the probability of a class class equal to or greater than the probability of a class class class equal to or greater than the probability of a class class class class equal to or greater than the probability of a class class class class class equal to the probability of a class ... The predicted largest failure entity; Step (2.3.2.3) uses the conditional time density function to predict the failure time offset Δt, which indicates the trend of earlier or later occurrence of the faulty entity. i ; Where τ is a time parameter used for integration.

9. The predictive maintenance method based on elevator operation and maintenance time-series knowledge graph as described in claim 1 or 2, characterized in that, The process of step (2.4) is as follows: Step (2.4.1) Maintenance cost C b The calculation is as follows: Step (2.4.1.1) In this invention, the maintenance cost before the failure is divided into labor cost C. p and material costs C s The two parts are related in that labor costs are linearly related to pre-failure maintenance time: C p (i)=Pt b (i) Among them, t b This indicates the duration of a single maintenance operation, and P represents the unit labor cost coefficient. The maintenance duration in step (2.4.1.2) will dynamically change with the maintenance frequency, which can be expressed as: In this function, e represents the technician's experience level, and r reflects the intensity of the experience curve effect; Step (2.4.1.3) cumulative pre-failure maintenance cost C b Integrated into: Step (2.4.2) Maintenance cost C a The calculation is as follows: The number of faults occurring within the maintenance cycle [(j-1)T,jT] in step (2.4.2.1) can be modeled as follows: The cumulative number of failures of the part in step (2.4.2.2) during the entire maintenance cycle is: Step (2.4.2.3) combines the average cost per repair (c) o The total maintenance cost is: C a =c o F total Step (2.4.3) Downtime cost C f The calculation is as follows; The downtime in step (2.4.3.1) includes the duration t of a single maintenance operation. b and single fault handling time t a The downtime loss per unit time is c. f The total cost of downtime losses consists of the downtime cost per unit time, c. f Total maintenance time T b The total time T spent on repairs a A decision is expressed as: C f =c f (T b +T a ) The total time T spent on maintenance in step (2.4.3.2) b The total time T spent on repairs a It is expressed as follows: Step (2.4.4) Total cost C required for maintenance of each component total Represented as: C total =C b +C a +C f 。 10. The predictive maintenance method based on elevator operation and maintenance time-series knowledge graph as described in claim 1 or 2, characterized in that, The process of step (2.5) is as follows: Step (2.5.1) defines the state space and transition rate matrix as follows; The state space S for step (2.5.1.1) is defined as follows: S={(j,k)∣j∈[1,2,…,N],k∈[0,1]} Where j∈[1,2,…,N] represents the j-th maintenance cycle, and k∈[0,1] represents the component status (0 = normal, 1 = failure); The transfer and repair rate matrix Q in step (2.5.1.2) is defined as follows: Transition rate from normal state to failure state: Q fault =λ j Repair rate from failure state to normal state: Step (2.5.2) Steady-state probability calculation: For each state (j,k), the steady-state probability equation is satisfied: ∑π((j′,k′))q {(j′,k′)→(j,k)} =π((j,k))∑q {(j,k)→(j′,k′)} Where (j,k) represents the current state, (j′,k′) represents the changed state, and π(·) is the steady-state probability. Step (2.5.3) Calculate the number of failures N within the maintenance cycle. fault Total maintenance time T maintenance The calculation method is as follows: Step (2.5.4) combines time and cost to iteratively optimize the number of maintenance operations N. The iteration stops when the minimum cost assessment index E(N) and the maximum utilization rate A(N) are reached, yielding the optimal maintenance cycle T. The process is as follows: The total cost assessment indicator E(N) for step (2.5.4.1) is: Among them, T total Total time used for the component; The utilization rate and effective service life of components in step (2.5.4.2) can be expressed as: T available =T total -T b -T a The utilization rate A(N) is represented by the ratio of the effective usage time of a component to the total usage time. In step (2.5.4.3), the number of repairs N is the decision variable, and the relationship between the maintenance cycle T and the number of repairs N is expressed as follows: T=[int(T total -T ρ ) / N)] Here, int is the integer function, ensuring that the period is an integer.

Images