An elevator fault diagnosis method based on federated learning, an electronic device, a medium and a program product

Abstract

The application provides an elevator fault diagnosis method based on federated learning, an electronic device, a medium and a program product. The method comprises: obtaining a local vector of a first elevator through a local node; aligning the local vector with a reference vector based on a preset strategy to obtain a corrected vector; obtaining a fault type based on the corrected vector and a student model of the local node; obtaining an abnormal parameter related to the fault type; obtaining an abnormal result based on calculating a first similarity between each fault template and the abnormal parameter by pre-storing each fault template in the local node in the cloud; when the abnormal result indicates that there is a faulty component in the first elevator, determining the faulty component of the first elevator based on a preset knowledge graph according to the fault type and performing a warning. The application improves the accuracy and timeliness of elevator fault warning and reduces maintenance costs.

Description

Technical Field

[0001] This invention relates to the field of special equipment technology, and more specifically, to an elevator fault diagnosis method, electronic equipment, media, and program products based on federated learning. Background Technology

[0002] Traditional elevator fault diagnosis mainly relies on regular manual inspections, sensor threshold alarms, and post-incident maintenance, which suffers from problems such as slow response, high maintenance costs, and inability to predict hidden faults. With the development of IoT technology, intelligent diagnostic methods based on single elevator data have emerged, identifying faults by analyzing signals such as vibration, current, and noise. However, these methods have significant limitations: 1) Data silos: Fault data for a single elevator is limited, making it difficult to train robust and accurate diagnostic models, especially for rare faults; 2) Privacy and business confidentiality: Elevator operation data involves user privacy and the core knowledge of maintenance companies, and owners or maintenance providers are unwilling to directly share raw data; 3) Lack of knowledge reusability: Diagnostic experience for elevators of different models and in different environments is difficult to effectively transfer and share.

[0003] Based on this, federated learning has been gradually applied to the field of elevator fault diagnosis. Its core idea is to extract global fault knowledge through a cloud-based teacher model and distribute it to local nodes, combining it with local student models to achieve fault diagnosis, thus balancing data privacy protection and global knowledge reuse. However, the current application of federated learning in elevator fault diagnosis still has many key shortcomings: Due to the non-IID heterogeneity of data in each local node of the elevator cluster, the baseline vector distributed by the cloud and the local vector collected by the local nodes have not been standardized and aligned. Even if the two have the same dimension, the feature distribution will be disordered due to equipment heterogeneity, sensor bias, and time series offset, which will result in insufficient accuracy when the local nodes calculate the first similarity between the fault template and the abnormal parameters, and will easily lead to misjudgment of fault type. At the same time, the local student model does not fully rely on the baseline vector and global fault knowledge provided by the cloud-based teacher model, but only relies on its own local data for training, resulting in weak generalization ability and inability to effectively reuse the core value of the global fault template. In addition, under the constraints of privacy protection, the existing federated diagnosis schemes are difficult to achieve deep integration of global fault knowledge and local diagnosis, resulting in low accuracy of fault component location and poor generalization, which cannot meet the actual needs of large-scale and refined fault diagnosis of elevator clusters. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide an elevator fault diagnosis method, electronic device, medium and program product based on federated learning, which can improve the problem that the existing technology cannot meet the actual needs of large-scale and refined fault diagnosis of elevator clusters.

[0005] To achieve the above technical objectives, the technical solution adopted in this application is as follows:

[0006] Firstly, embodiments of this application provide an elevator fault diagnosis method based on federated learning.

[0007] Obtain the local vector of the first elevator through the local node;

[0008] Based on a preset strategy, the local vector is aligned with the reference vector to obtain a corrected vector, wherein the reference vector is obtained from the teacher model in the cloud, and the local vector and the reference vector have the same dimension.

[0009] Based on the correction vector, the fault type is obtained according to the student model of the local node;

[0010] Obtain the abnormal parameters related to the fault type;

[0011] Anomalies are obtained by calculating the first similarity between each fault template and anomaly parameters, based on each fault template pre-stored in the cloud on the local node.

[0012] When the abnormal result indicates that there is a faulty component in the first elevator, the faulty component of the first elevator is determined based on the fault type and a preset knowledge graph, and an early warning is issued.

[0013] According to the first aspect, aligning the local vector with the reference vector based on a preset strategy to obtain the corrected vector includes:

[0014] Calculate the second similarity between the local vector and the reference vector. When the second similarity is less than the first similarity threshold, extract the local key node time of the local vector and the reference key node time of the reference vector, and / or calculate the first reference magnitude of the local vector and the second reference magnitude of the reference vector.

[0015] Calculate the time difference between the local critical node time and the reference critical node time, and / or calculate the scaling factor of the first reference amplitude relative to the second reference amplitude;

[0016] The local vector is shifted as a whole based on the time difference, and / or each local feature in the local vector is scaled based on the scaling factor to obtain the corrected vector, so as to align the local vector with the reference vector.

[0017] According to the first aspect, the correction vector includes vibration characteristics, current characteristics, and door operator characteristics;

[0018] The fault type, obtained based on the correction vector and the student model of the local node, includes:

[0019] The correction vector is input into the student model, and the vibration result corresponding to the vibration feature, the current result corresponding to the current feature, and the gantry machine result corresponding to the gantry machine feature are obtained through the student model.

[0020] When one of the vibration result, current result, and gantry machine result is abnormal and the other two are normal, the third similarity and fourth similarity of the abnormal vibration result, current result, and gantry machine result are calculated with the other two normal vibration result, current result, and gantry machine result.

[0021] Based on the vibration results, current results, and gantry crane results, the fault type is obtained according to the third similarity and the fourth similarity.

[0022] According to the first aspect, the fault type is obtained based on the vibration results, current results, and gantry crane results, according to the third similarity and the fourth similarity, including:

[0023] When the third similarity and the fourth similarity are less than the second similarity threshold, the fault type is obtained based on one of the vibration result, current result, and gantry crane result that characterizes the abnormality. When the third similarity is greater than the second similarity threshold and the fourth similarity is less than the second similarity threshold, the fault type is obtained based on one of the vibration result, current result, and gantry crane result that characterize the abnormality, and one of the other two corresponding to the second similarity that characterize the normality. When the fourth similarity is greater than the second similarity threshold and the third similarity is less than the second similarity threshold, the fault type is obtained based on one of the vibration result, current result, and gantry crane result that characterize the abnormality, and one of the other two corresponding to the fourth similarity that characterize the normality.

[0024] According to the first aspect, after obtaining an anomaly result based on calculating a first similarity between each fault template and an anomaly parameter via a cloud-pre-stored fault template on the local node, the method further includes:

[0025] Obtain fault templates whose first similarity is greater than the fifth similarity threshold;

[0026] When the fault template with a similarity greater than the fifth similarity threshold indicates that the fault is caused by elevator aging, an abnormal result is obtained indicating that the fault of the first elevator is caused by aging. The step of determining the faulty component of the first elevator based on the fault type and the abnormal result and a preset knowledge graph is not executed.

[0027] According to the first aspect, determining the faulty component of the first elevator based on a preset knowledge graph according to the fault type includes:

[0028] The fault type is input into the knowledge graph, and the first elevator system corresponding to the fault type is obtained through matching in the knowledge graph;

[0029] Based on the system to which it belongs, several candidate components are identified;

[0030] The runtime, design life, number of historical failures, and time since the last maintenance are obtained for each candidate component in the local node.

[0031] Based on the runtime, design life, number of historical failures, and time since the last maintenance, a health score is calculated for each candidate component. The health score is positively correlated with the health of the corresponding candidate component. Candidate components with a health score less than a health threshold are considered faulty components.

[0032] According to the first aspect, after calculating the health score for each candidate component, the method further includes:

[0033] When the number of faulty components is greater than or equal to 2, the matching confidence of the faulty components and the failure probability of candidate components are obtained. The matching confidence is obtained based on the mapping weight between the faulty components and the fault type in the knowledge graph. The failure probability of candidate components is obtained from a list in the cloud, which includes the failure probability of candidate components of a second elevator that conforms to preset rules with the first elevator.

[0034] The confidence level of the corresponding candidate component is determined based on the matching confidence level, the failure probability of the candidate component, and the health score.

[0035] Secondly, embodiments of this application provide an electronic device, which includes a processor and a memory coupled to each other. The memory stores a computer program, and when the computer program is executed by the processor, the electronic device performs the method described in the first aspect.

[0036] Thirdly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when run on a computer, causes the computer to perform the method described in the first aspect.

[0037] Fourthly, embodiments of this application propose a program product, characterized in that it includes a computer program, which, when executed by a processor, implements the method described in the first aspect.

[0038] The invention employing the above technical solution has the following advantages:

[0039] In the technical solution provided in this application, the local node first obtains the local vector of the first elevator to provide basic data for fault diagnosis. Then, based on a preset strategy, the local vector is aligned with the reference vector of the same dimension obtained from the cloud teacher model to obtain a corrected vector, which effectively eliminates non-fault-causing interference such as equipment heterogeneity, sensor bias, and time series offset, and solves the problem of disordered feature distribution. Next, the corrected vector is input into the student model of the local node. Relying on the reasoning ability of the student model on standardized features, the fault type is obtained, which makes up for the shortcomings of the local model, which only relies on local data for training and has weak generalization ability. Then, the abnormal parameters related to the fault type are obtained. By calculating the first similarity between each fault template pre-stored in the cloud and the abnormal parameters of the local node, an abnormal result is obtained to characterize whether there is a faulty component in the first elevator, thus realizing the determination of the existence of the fault. When the abnormal result indicates that there is a faulty component, the fault type is matched with the system to which the fault belongs, the candidate component is screened, and the faulty component is located based on a preset knowledge graph, thus completing the fault warning. The technical solution provided in this application deeply integrates the global knowledge of the cloud-based teacher model with the real-time data and student model of the local node, and executes it independently on the local node throughout the process. This not only ensures the privacy and security of elevator operation data, but also effectively solves the problems of low diagnostic accuracy, poor generalization and insufficient real-time performance of existing solutions. It realizes the execution of elevator faults from feature collection, alignment, and type determination to component location and early warning, improves the accuracy and timeliness of elevator fault early warning, and reduces maintenance costs. Attached Figure Description

[0040] This application can be further illustrated by the non-limiting embodiments given in the accompanying drawings. It should be understood that the following drawings only illustrate some embodiments of this application and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained from these drawings without any inventive effort.

[0041] Figure 1 This application provides a federated learning-based elevator fault diagnosis method.

[0042] Figure 2 This is a sub-flowchart of S120 provided in an embodiment of this application.

[0043] Figure 3 This is a sub-flowchart of S130 provided in an embodiment of this application. Detailed Implementation

[0044] The present application will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that similar or identical parts are referred to by the same reference numerals in the drawings or description. Implementations not shown or described in the drawings are forms known to those skilled in the art. In the description of this application, terms such as "first" and "second" are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0045] Please refer to Figure 1 This application provides a federated learning-based elevator fault diagnosis method, which can be applied to electronic devices and whose steps can be executed or implemented by the electronic devices. The electronic devices can be, but are not limited to, personal computers, smartphones, and other electronic devices. The federated learning-based elevator fault diagnosis method may include the following steps:

[0046] S110, obtain the local vector of the first elevator through the local node;

[0047] S120, based on a preset strategy, align the local vector with the reference vector to obtain a corrected vector, wherein the reference vector is obtained from the teacher model in the cloud, and the local vector and the reference vector have the same dimension;

[0048] S130, Based on the correction vector, the fault type is obtained according to the student model of the local node;

[0049] S140, Obtain abnormal parameters related to the fault type;

[0050] S150, using each fault template pre-stored in the cloud on the local node, an abnormal result is obtained based on calculating the first similarity between each fault template and the abnormal parameters;

[0051] S160, when the abnormal result indicates that there is a faulty component in the first elevator, the faulty component of the first elevator is determined based on the fault type and a preset knowledge graph, and an early warning is issued.

[0052] In the above implementation, the local vector of the first elevator is first obtained through the local node, providing basic data support for fault diagnosis. Then, based on a preset strategy, the local vector is aligned with the baseline vector obtained from the cloud-based teacher model to obtain a corrected vector. This solves the problem of disordered feature distribution between the local vector and the baseline vector due to the non-IID heterogeneity of the elevator cluster's local node data, which results in the local vector and the baseline vector having the same dimension. This provides a standardized feature basis for subsequent fault determination. Next, based on the corrected vector, the fault type is obtained through the student model of the local node. Relying on the baseline vector and implicit global fault knowledge provided by the cloud-based teacher model, this compensates for the shortcomings of the local student model, which relies only on local data for training and has weak generalization ability, thus improving the accuracy of fault type determination. Then, abnormal parameters related to the fault type are obtained, and the first similarity between each fault template and the abnormal parameters is calculated using fault templates pre-stored in the cloud on the local node. Combined with the standardized correction vector, the problem of insufficient accuracy in similarity calculation in the existing scheme is solved, thus obtaining accurate abnormal results. Finally, when the abnormal result indicates the presence of a faulty component, the specific faulty component is determined according to the fault type and the preset knowledge graph, and an early warning is issued, forming a complete link of "benchmark alignment - fault type determination - similarity matching - component location - early warning". This effectively solves the defects of insufficient integration of global fault knowledge and local diagnosis and inaccurate fault component location in the existing federated diagnosis scheme, and realizes accurate and efficient diagnosis and early warning of elevator faults under privacy protection constraints.

[0053] The method in this embodiment is based on a federated learning model, which is established in three stages, specifically:

[0054] Phase 1: Construction of Cloud-Based Teacher Model and Global Knowledge Distillation

[0055] This phase relies on the cloud-based core nodes of the federated learning architecture to build the teacher model and distill and precipitate global fault knowledge, laying the knowledge foundation for the entire federated diagnostic system. The cloud, serving as the platform for global data aggregation and model training, first collects 128-dimensional feature vectors of elevators of different brands, service lives, and operating conditions from various local nodes, under the premise of privacy compliance. After deduplication and outlier cleaning, a standardized federated training dataset is constructed. Subsequently, a deep time-series Transformer teacher model with parameter sizes adapted to the high computing power of the cloud is built, and pre-trained using the standardized feature vectors to enable deep learning. The feature distribution patterns of the entire elevator and the fault-component mapping relationship are then established. Next, the teacher model uses a federated distillation algorithm to statistically aggregate the health features of the entire region to generate a global health baseline vector with the same dimension as the local vector. At the same time, it distills various fault feature templates and fault-component association knowledge, which are solidified into a fault template library pre-stored in the cloud. Finally, the baseline vector and fault template library are encrypted and distributed to all local nodes. This completes the construction of the cloud teacher model and the initial deployment of global knowledge, effectively solving the core problems of local node models lacking global fault cognition and having weak generalization ability, and providing a unified and standardized basis for subsequent feature alignment and local diagnosis.

[0056] Phase 2: Deployment and Local Adaptation of Local Nodes and Student Models

[0057] This phase completes the deployment and lightweight adaptation of the local execution entity of the federated learning model, meeting the real-time, privacy, and low-computing-power requirements of elevator edge diagnostics. Local nodes, serving as edge computing units for each elevator (such as ARM controllers and edge computing gateways within elevator control cabinets), are the direct execution scenario for fault diagnosis. Feature acquisition, alignment, inference, abnormal parameter calculation, and early warning modules need to be deployed in the local node of the first elevator to ensure real-time local vector acquisition, alignment with the baseline vector, and local computation. Simultaneously, addressing the issue of large parameter volumes in the teacher model, which cannot be directly deployed on low-computing-power edge hardware, cloud-based support for the teacher model is provided. The model undergoes lightweight processing, reducing the number of parameters by more than 85% through model pruning and quantization compression, generating a lightweight student model with initial weights that are then distributed to local nodes. Based on a limited amount of historical fault and health data collected by the local nodes, the student model is fine-tuned to adapt to the brand, service life, and operating conditions of the elevator company, addressing the Non-IID heterogeneity issue in federated learning. The local diagnostic link is then configured, ultimately achieving the adapted deployment of the local nodes and the lightweight student model. This allows the local nodes to perform accurate on-site fault diagnosis while protecting privacy, providing an execution platform for the implementation of federated learning.

[0058] Phase 3: Collaborative Iterative Optimization of Federated Learning Models

[0059] This phase involves collaborative iteration between cloud and local nodes to continuously optimize the diagnostic accuracy of the federated learning model, achieving deep integration of global knowledge and local data. After the local nodes of the first elevator complete student model inference and fault diagnosis, they extract only the local gradient information during the model inference process (without involving the original running data), encrypt it, and upload it to the cloud. The cloud receives the local gradients from all local nodes, uses the FedAvg federated averaging algorithm to aggregate the global gradients, updates the weights of the teacher model, and simultaneously re-distills the iterative global health baseline vector and the optimized fault template library based on the aggregated global data to improve the accuracy of global knowledge. Subsequently, the cloud transmits the updated teacher model and iterative baseline vector to the cloud. The quantity and optimization fault template library is redistributed to each local node. The student model on the local node performs a second local fine-tuning based on the new global knowledge, completing the iterative optimization of the model. The cloud continuously monitors the model iteration effect. When the accuracy improvement of the aggregated model is less than a preset threshold (e.g., 1%) or the local diagnostic accuracy reaches a preset standard (e.g., 98%), the federated iteration is terminated and the model enters the stable operation stage. This collaborative iteration process allows the cloud teacher model to continuously absorb the local data features of each local node, and the local student model to continuously adapt to the latest global fault knowledge. Ultimately, the accuracy of the federated learning model is continuously improved under privacy protection constraints, supporting the large-scale and refined fault diagnosis of elevator clusters.

[0060] Among them, the cloud represents the core scheduling and knowledge aggregation node of federated learning, which is responsible for privacy aggregation of elevator operation data across the entire domain, teacher model training, global fault knowledge distillation, baseline vector distribution and model iterative optimization.

[0061] The teacher model, deployed in the cloud, is the core model that carries the global elevator fault characteristics and patterns. It is responsible for extracting and distilling standardized feature knowledge of the entire healthy / fault state and generating a unified benchmark vector.

[0062] Local nodes are edge computing units corresponding to each elevator (such as ARM controllers and edge computing gateways in elevator control cabinets). They are the main entities responsible for fault diagnosis, including local vector acquisition, feature alignment, student model inference, fault parameter calculation, and local early warning.

[0063] The student model is a lightweight model deployed on local nodes. It is a diagnostic model adapted to low computing power at the edge. It is responsible for determining the fault type based on the corrected local vectors, while receiving knowledge distillation guidance from the cloud teacher model to make up for the lack of local data.

[0064] The steps of the elevator fault diagnosis method based on federated learning will be explained in detail below:

[0065] In S110, the local node, through its multimodal signal acquisition unit, collects three types of raw time-series data in real time during the operation of the first elevator: vibration, drive current, and door operator action. After preprocessing such as noise filtering, time slicing, and amplitude normalization at the local edge, the preprocessed signal is extracted in the time and frequency domains based on a preset feature extraction algorithm. The extracted features are then spliced ​​and standardized according to fixed-dimensional rules to generate a 128-dimensional standardized feature vector that is completely consistent with the baseline vector dimension issued by the cloud-based teacher model. This vector is the local vector of the first elevator. The entire process is completed independently on the local node without uploading the raw operating data.

[0066] After obtaining the local vector from the local node, as follows Figure 2 As shown, S120 specifically includes the following steps:

[0067] S121, calculate the second similarity between the local vector and the reference vector. When the second similarity is less than the first similarity threshold, extract the local key node time of the local vector and the reference key node time of the reference vector, and / or calculate the first reference amplitude of the local vector and the second reference amplitude of the reference vector.

[0068] S122, calculate the time difference between the local key node time and the reference key node time, and / or, calculate the scaling factor of the first reference amplitude relative to the second reference amplitude;

[0069] S123, based on the time difference, the local vector is translated as a whole, and / or, based on the scaling factor, each local feature in the local vector is scaled to obtain the corrected vector, so as to align the local vector with the reference vector.

[0070] The second similarity between the local vector and the reference vector is calculated to quantitatively evaluate the degree of feature distribution offset of the first elevator's local vector caused by non-fault-causing factors such as equipment heterogeneity, sensor drift, and timing phase misalignment. This provides an accurate quantitative basis for subsequent feature alignment correction. When the second similarity is less than the first similarity threshold, it can be determined that the local vector has significant invalid feature interference. It is necessary to extract the key node time and reference amplitude of the local and reference vectors to carry out timing and amplitude correction, thereby removing non-fault interference components and retaining real fault features, ensuring the effectiveness of subsequent corrected vectors and the accuracy of model diagnosis. The reference vector is a 128-dimensional standardized feature vector generated by the cloud teacher model based on the federated learning architecture, which aggregates multimodal temporal features under the healthy steady-state operation of elevators across the entire domain and performs standardized distillation. It uniquely represents the common health feature distribution law of elevators without faults or anomalies, and its dimensional structure is completely consistent with the local vector. It serves as the global unified standard anchor point for local feature alignment in the federated learning system.

[0071] In this embodiment, the first similarity threshold can be 85%. When the second similarity is greater than or equal to 85%, it proves that the local vector and the reference vector have been aligned. When it is less than 85%, alignment is required. Alignment mainly includes the following two parts:

[0072] The specific implementation process of timing alignment (time difference shifting of critical nodes):

[0073] When timing correction is required based on the second similarity, the local node accurately extracts the local key node time of the elevator operation corresponding to the local vector (such as the timing points of core actions like traction machine starting and braking, and car door opening and closing) and the global standard reference key node time corresponding to the reference vector. The time difference between the two is calculated by subtraction. This time difference represents the leading or lagging offset of the local vector's timing relative to the reference vector. Then, based on this time difference, the overall time domain translation operation is performed on the full-dimensional timing features corresponding to the local vector to accurately align the local key node with the reference key node. Only the invalid feature offset caused by timing phase misalignment is eliminated, and the fault-induced feature distortion in the local vector is completely preserved, thus completing the alignment correction of the timing dimension.

[0074] The specific implementation process of amplitude alignment (scaling factor feature scaling)

[0075] When amplitude correction is required based on the second similarity, the local node calculates the first baseline amplitude of the overall features of the local vector (such as feature mean, peak value, and other global amplitude indicators) and the second baseline amplitude of the baseline vector (healthy state standard global amplitude indicator). The standardized amplitude scaling factor is calculated by the ratio of the second baseline amplitude to the first baseline amplitude. Based on this scaling factor, each local feature value in the 128 dimensions of the local vector is scaled proportionally. Only the overall amplitude offset caused by sensor drift and heterogeneous aging of equipment is corrected, without changing the relative distribution law of features and the distortion morphology of fault features. Finally, the amplitude-standardized correction vector is obtained, completing the alignment correction of the amplitude dimensions.

[0076] In this embodiment, both the correction vector and the local vector include vibration features, current features, and door operator features. The correction vector and the local vector are feature vectors of the same dimension and structure. Both are composed of three parts: vibration features, current features, and door operator features. The vibration features are used to characterize the operating state of mechanical structures such as elevator traction machines and guide rails. The current features are used to reflect the electrical operating characteristics of traction motors and electrical circuits. The door operator features are used to reflect the action rules of the opening and closing process of car doors and landing doors. During the alignment process of time-series translation and amplitude scaling, the local node uniformly performs overall time-series translation and proportional amplitude scaling on the vibration features, current features, and door operator features in the local vector. This does not change the internal combination relationship and fault distortion information of the three types of features. Finally, a correction vector containing vibration features, current features, and door operator features is obtained. This allows the correction vector to completely retain the effective feature distribution related to faults while eliminating non-fault interference such as sensor drift and time-series misalignment. This provides structurally consistent and standardized feature inputs for the subsequent fault type determination by the student model.

[0077] In S130, such as Figure 3 The main steps include the following:

[0078] S131: Input the correction vector into the student model, and obtain the vibration result corresponding to the vibration feature, the current result corresponding to the current feature, and the gantry machine result corresponding to the gantry machine feature through the student model;

[0079] S132: When one of the vibration result, current result and door machine result is abnormal and the other two are normal, calculate the third similarity and fourth similarity between the abnormal vibration result, current result and door machine result and the other two normal vibration result, current result and door machine result.

[0080] S133: Based on the vibration results, current results, and gantry crane results, the fault type is obtained according to the third similarity and the fourth similarity.

[0081] Specifically, S133 is:

[0082] When the third similarity and the fourth similarity are less than the second similarity threshold, the fault type is obtained based on one of the vibration result, current result, and gantry crane result that characterizes the abnormality. When the third similarity is greater than the second similarity threshold and the fourth similarity is less than the second similarity threshold, the fault type is obtained based on one of the vibration result, current result, and gantry crane result that characterize the abnormality, and one of the other two corresponding to the second similarity that characterize the normality. When the fourth similarity is greater than the second similarity threshold and the third similarity is less than the second similarity threshold, the fault type is obtained based on one of the vibration result, current result, and gantry crane result that characterize the abnormality, and one of the other two corresponding to the fourth similarity that characterize the normality.

[0083] In this embodiment, the core judgment indicator in this solution is the second similarity threshold. This threshold is a pre-set inter-modal feature association judgment standard for local nodes, used to measure the degree of correlation between the distribution of abnormal modal features and normal modal features. The preset value is 0.7 (which can be dynamically fine-tuned according to elevator brand and operating conditions). The threshold is essentially a quantitative boundary for whether there is a synergistic effect between abnormal and normal modes. When the similarity is >0.7, it indicates that there is a significant correlation between the two types of modal features, which may form a joint fault. When the similarity is <0.7, it indicates that there is no obvious correlation between the two types of modal features, and it is only a single modal independent anomaly.

[0084] In S131, 128 maintenance positive vectors containing vibration features (first 48 dimensions), current features (middle 48 dimensions), and gantry crane features (last 32 dimensions) are input into the student model of the local node. The model automatically splits the correction vector according to the feature modality dimension, performs independent reasoning and discrimination on the three types of feature sub-vectors, and outputs vibration results, current results, and gantry crane results that respectively represent the operating state of the corresponding modality. Each result is clearly marked as "normal" or "abnormal" for the corresponding modality, and also includes the quantitative value of the modality feature (for subsequent similarity calculation).

[0085] Upon entering S132, the three types of results are logically filtered, retaining only the scenario where "only one of the three types of results is judged as abnormal, and the remaining two are normal" (excluding the case where two or more modalities are abnormal at the same time). The inter-modal similarity calculation process is then initiated: using the feature results (quantified values) of the abnormal modality as the reference object, the improved cosine similarity algorithm is used to perform distribution matching calculation with the feature results (quantified values) of the two normal modalities respectively, to obtain two sets of similarity values, namely the third similarity (similarity between the abnormal modality and the first normal modality) and the fourth similarity (similarity between the abnormal modality and the second normal modality). The two sets of values ​​are used to accurately measure the degree of feature correlation between the abnormal modality and each normal modality, providing a core basis for subsequent differentiation between single faults and joint faults.

[0086] In S133, based on the relationship between the third similarity, the fourth similarity, and the second similarity threshold (0.7), the fault type is refined in three scenarios, clearly distinguishing between single-mode faults and joint faults: When both the third and fourth similarities are less than the second similarity threshold (0.7), it indicates that the abnormal mode has no obvious correlation with the feature distributions of the two normal modes. The anomaly is a single-mode fault independently generated by the corresponding mode, and the fault type is directly determined based on the result corresponding to the abnormal mode (only the fault template of the abnormal mode is associated); when the third similarity is greater than the second similarity threshold (0.7) and the fourth similarity is less than the second similarity threshold (0.7), it indicates that the fault type is a single-mode fault generated independently by the corresponding mode. If the abnormal mode is significantly associated only with the normal mode corresponding to the third similarity, the fault is a joint fault caused by the combined effect of the abnormal mode and the normal mode. The fault type needs to be determined by combining the results of the abnormal mode and the normal mode (fault template of the combined effect of the two modes); when the fourth similarity is greater than the second similarity threshold (0.7) and the third similarity is less than the second similarity threshold (0.7), it indicates that the abnormal mode is significantly associated only with the normal mode corresponding to the fourth similarity. The fault is also a joint fault. The fault type needs to be determined by combining the results of the abnormal mode and the normal mode (fault template of the combined effect of the two modes).

[0087] Assuming the second similarity threshold in this example is preset to 0.7, and the correction vector is 128-dimensional (48-dimensional vibration, 48-dimensional current, and 32-dimensional gantry crane), the specific execution process is as follows: Execute S131: Input the 128-dimensional positive maintenance vector into the student model. After the model is split, the output results are: vibration result (abnormal, indicating abnormal vibration of the traction machine), current result (normal, indicating normal current of the traction motor), and gantry crane result (normal, indicating normal gantry crane operation), which meets the triggering condition of "only one type of abnormality, two types of normality"; Execute S132: Using the abnormal vibration result as a reference, calculate the similarity with the normal current result and gantry crane result respectively, obtaining the third similarity (similarity between vibration and current) = 0.75, and the fourth similarity... = 0.62; Execute S133: Combine the second similarity threshold (0.7) to determine the scenario - the third similarity (0.75) > 0.7, the fourth similarity (0.62) < 0.7, which belongs to the scenario of "only the third similarity is greater than the threshold". Therefore, the fault is a joint fault of vibration and current. The fault type is determined by combining the abnormal vibration result (abnormal vibration of traction machine) and the normal current result (but it is related to vibration). Finally, it is determined to be "a joint abnormal vibration-current fault caused by the wear of traction machine bearings" (because although the wear of traction machine bearings mainly manifests as abnormal vibration, it will indirectly affect the stability of motor current, causing the correlation between the two features to exceed the threshold, forming a joint fault). If the similarity values ​​in the example are adjusted to 0.65 for the third similarity and 0.68 for the fourth similarity, both less than 0.7, then it is determined to be a single-mode fault, and based solely on the abnormal vibration result, it is determined to be a "single fault of traction machine bearing wear". If the third similarity is 0.63 and the fourth similarity is 0.72, then it is determined to be a combined fault of vibration and gantry crane, and based on the abnormal vibration and normal gantry crane result, it is determined to be a "combined fault of traction machine vibration abnormality indirectly affecting the stability of gantry crane operation".

[0088] In S140, after the fault type is determined by the local node based on the multimodal feature similarity, the feature dimension interval corresponding to the fault type in the 128 maintenance positive vector is first located according to the feature mode to which the fault type belongs and whether it is a joint fault type. If it is a single-mode fault, the key feature values ​​in the corresponding vibration, current or gantry machine sub-vector are extracted. If it is a joint fault, the core feature data of the two related modes are extracted simultaneously. Then, the correction parameters such as the time-series translation time difference and amplitude scaling coefficient obtained in the previous feature alignment process, as well as the quantitative scores of each mode state output by the student model, are used together as basic anomaly features. The basic anomaly features are then matched with the corresponding fault templates pre-stored in the local node and distilled by the cloud teacher model. Redundant feature parameters that are not related to the current fault type are removed, and the time domain indicators, frequency domain indicators, amplitude indicators and time series offset indicators that are strongly correlated with the fault type are retained. Finally, after standardization and integration, the anomaly parameters that are highly matched with the current fault type are obtained and used to calculate the first similarity with the fault template to further confirm the faulty component.

[0089] In S150, the fault template is a standardized feature template set pre-stored on each local node after the cloud-based teacher model aggregates the standardized feature patterns of various single-modal faults, joint faults, and aging-related faults of elevators across the entire domain, refined and solidified by the federated distillation algorithm. Each fault template corresponds to a specific fault type (including single-modal faults, joint-modal faults, and aging-related faults), and its dimensional structure is completely consistent with the local anomaly parameters and the corrected feature dimensions (adapting to the modal division of 128-dimensional features). It contains the standard range of core feature parameters such as time domain, frequency domain, and amplitude corresponding to this type of fault. Essentially, it is the "feature fingerprint" of various faults (including aging-related faults). It can be directly used for local anomaly matching and fault determination without real-time interaction with the cloud. Specifically, it includes single-modal fault templates, joint-modal fault templates, and aging-related fault templates (specifically corresponding to the feature anomalies caused by long-term use and natural aging of elevators, covering the standard range of vibration, current, and door operator-related feature parameters caused by aging). It is the core reference for local nodes to quickly match and determine faults.

[0090] The anomaly result is obtained by the local node by calculating the first similarity between each fault template and its corresponding anomaly parameter (here, "first similarity" refers to the matching degree between each fault template and the current elevator anomaly parameter), combined with a preset first similarity threshold (a value of 0.8 is recommended and can be dynamically fine-tuned). The core content includes three parts: first, a ranking table of the first similarity between all fault templates and anomaly parameters; second, a clear determination of whether the elevator has a faulty component (if the first similarity of at least one fault template is greater than or equal to the first similarity threshold, then a faulty component is determined to exist; if the first similarity of all templates is less than the threshold, then no faulty component is determined to exist); and third, the fault association information corresponding to the fault template with the highest similarity, which provides a direct reference for subsequent location of specific faulty components based on the knowledge graph. The entire calculation is completed independently on the local node without cloud intervention.

[0091] This embodiment achieves accurate identification of faulty components by performing local similarity matching between pre-stored global fault templates (including aging-related templates) and the current elevator's abnormal parameters. The specific steps are as follows: The local node retrieves all pre-stored fault templates (including single-modal fault templates, joint-modal fault templates, and aging-related fault templates) to ensure complete matching between template dimensions and abnormal parameter dimensions (adapting to 128-dimensional feature modal division); it extracts previously determined abnormal parameters corresponding to the fault type (including modal features of single / joint faults, correction parameters, etc.), and uses an improved cosine similarity algorithm to calculate the first similarity between each fault template and the abnormal parameter, quantifying the degree of feature matching; and then performs local similarity matching on all first-phase parameters. The similarity is compared with a preset first similarity threshold (a value of 0.8 is recommended and can be dynamically fine-tuned) to filter out fault templates with a similarity ≥ the threshold (i.e., templates that highly match the current elevator's abnormal parameters); all matching template information and similarity data are integrated to generate anomaly results, clearly determining whether the current elevator has a faulty component, and outputting the fault association information corresponding to the template with the highest similarity, providing core support for subsequent location of specific faulty components based on knowledge graphs; the entire process is executed independently on local nodes without cloud interaction, ensuring both data privacy and security, and ensuring the real-time nature of fault determination. At the same time, relying on aging-related fault templates, it accurately matches abnormal features caused by aging, avoiding confusion between normal aging and fault aging.

[0092] The specific method for calculating the first similarity is as follows: The local node first divides the abnormal parameters to be matched and all fault templates (including single-mode fault templates, joint-mode fault templates, and aging-related fault templates) into vibration feature sub-vectors (first 48 dimensions), current feature sub-vectors (middle 48 dimensions), and gantry crane feature sub-vectors (last 32 dimensions) according to a fixed physical partitioning rule of 128-dimensional features, ensuring that the abnormal parameters completely correspond to the feature dimension ranges of each fault template; then, it presets differentiated weighting coefficients for the three types of feature sub-vectors (vibration feature weighting coefficient 0.45, current feature weighting coefficient 0.35, gantry crane feature weighting coefficient 0.2, plus... The weighting coefficients are dynamically adjusted based on the sensitivity of each mode to elevator faults. Each sub-vector of the abnormal parameter and the corresponding sub-vector of the fault template are pre-processed with weights to amplify the feature contribution ratio of core fault-sensitive modes such as vibration and current, and weaken the redundant interference of auxiliary modes such as door operator. Finally, an improved cosine similarity algorithm is used to calculate the matching similarity between the weighted abnormal parameter sub-vector and the fault template sub-vector for each mode. Then, the similarity of the three modes is integrated by weighted summation to obtain the first similarity between the abnormal parameter and a single fault template. The value of this similarity ranges from [0,1]. The closer the value is to 1, the more consistent the feature distributions of the two are. The closer the value is to 0, the less significant the feature correlation is.

[0093] In this embodiment, a fault template with a first similarity greater than a fifth similarity threshold is obtained. When the fault template with a first similarity greater than the fifth similarity threshold indicates that the fault is caused by elevator aging, an abnormal result is obtained indicating that the fault of the first elevator is caused by aging, and S160 is not executed.

[0094] When the abnormal result indicates that there is a faulty component in the first elevator, execute S160, specifically:

[0095] The fault type is input into the knowledge graph, and the first elevator system corresponding to the fault type is obtained through matching in the knowledge graph. Based on the system, several candidate components are determined. The runtime, design life, number of historical faults, and time since last maintenance are obtained for each candidate component in the local node. Based on the runtime, design life, number of historical faults, and time since last maintenance, a health score is calculated for each candidate component. The health score is positively correlated with the health of the corresponding candidate component. Candidate components with a health score less than a health threshold are the faulty components.

[0096] This embodiment describes the core process of accurately locating the specific faulty component of the first elevator based on a knowledge graph and local component data after determining the presence of a faulty component in the abnormal results. The entire process is executed independently on the local node without cloud interaction. The specific steps are as follows:

[0097] Knowledge graph matching of the system: Local nodes input the previously determined fault types (including single-modal faults, joint faults, and aging-related faults) into the pre-stored elevator fault domain knowledge graph. This knowledge graph is generated by the cloud-based teacher model based on the distillation of elevator fault data across the entire domain. It has a built-in precise mapping relationship between "fault type and elevator system" (e.g., abnormal vibration faults correspond to the traction system, abnormal door machine faults correspond to the door machine system, and aging faults correspond to the aging-prone system). Through the graph's association matching algorithm, the first elevator system corresponding to the fault type is quickly located, clarifying the approximate scope of the fault and avoiding blind investigation.

[0098] Identifying candidate components: Based on the matched system and the "system-component" relationships in the knowledge graph, all components within that system that could potentially cause this type of fault are selected as candidate components. For example, if the system is a traction system and the fault type is abnormal vibration (including vibration caused by aging), candidate components could include the traction machine front bearing, traction machine rear bearing, traction wire rope, guide rails, etc.; if the system is a gantry crane system and the fault type is abnormal gantry crane operation, candidate components could include the gantry crane motor, gantry crane controller, gantry crane guide rails, etc., ensuring that candidate components comprehensively cover all possible sources of fault and do not overlook any potential faulty components.

[0099] Local retrieval of core parameters for candidate components: The local node retrieves four types of core operational data for each candidate component through its built-in data acquisition and storage module: runtime (cumulative runtime of the component since it was put into use), design life (the standard service life preset at the time of manufacture), historical failure count (the cumulative number of failures that have occurred in the past, including aging-related failures), and time since last maintenance (the time interval from the last maintenance to the present). If a certain type of parameter is missing, it is supplemented by "filling with the average of the same type of component in the same system" (e.g., if the historical failure count of a candidate component is missing, the average failure count of the same type of component of the same brand and service life is retrieved to fill it), ensuring the completeness of the parameters and providing a reliable basis for calculating the health score.

[0100] Calculating the health score of candidate components: A multi-dimensional weighted scoring algorithm is used, combining four core parameters to calculate the health score of each candidate component. The score ranges from 0 to 100, and the score is positively correlated with the component's health (the higher the score, the healthier the component; the lower the score, the higher the probability of aging or failure). The specific calculation logic is as follows: First, the four parameters are standardized to the [0,1] interval (eliminating dimensional differences), and a differential weighting coefficient is preset (running time weight 0.3, design life weight 0.25, historical failure count weight 0.25, and time since last maintenance weight 0.2; the weights can be dynamically adjusted according to the elevator's operating conditions); then, the score is calculated using the formula "Health score = (1 - standardized running time value) × 0.3 + standardized design life value × 0.25 + (1 - standardized historical failure count value) × 0.25 + (1 - standardized time since last maintenance value) × 0.2"; finally, a health threshold is preset (60 points is recommended, and can be dynamically fine-tuned). Candidate components with a score less than 60 are judged as faulty or severely aged components.

[0101] Final confirmation of faulty components: All candidate components with health scores < health threshold are screened out. If only one candidate component meets the criteria, it is directly determined to be the faulty component of the first elevator. If two or more candidate components meet the criteria, the previously preset confidence calculation process is initiated (the location confidence of each candidate component is calculated by multi-dimensional weighted fusion) to screen out the final faulty component and ensure that the location result is unique and accurate.

[0102] This solution uses a knowledge graph to accurately associate fault types with their respective systems and candidate components, effectively narrowing the scope of fault investigation. It also quantifies health scores using multiple parameters, including runtime, design life, historical fault count, and time since last maintenance. This provides a clear and objective reflection of component health status, distinguishing between normal aging and faulty aging, significantly improving the accuracy of fault location and avoiding misjudgments and omissions. Furthermore, the entire process is executed independently on local nodes without uploading raw data, strictly adhering to federated learning privacy requirements and eliminating cloud interaction latency. This makes it suitable for low-computing-power real-time inference scenarios at the elevator edge, ensuring efficient fault location. In addition, this solution can accurately pinpoint faulty components, reducing the manpower and time costs associated with blind maintenance investigations. It also supports preventative maintenance based on quantified health scores. Combined with a federated, iteratively updated knowledge graph, it adapts to various elevator brands and age ranges, including single faults, combined faults, and aging faults. Alignment with front-end features and the fault type determination process form a complete closed loop, significantly improving the overall intelligence level and generalization ability of elevator fault diagnosis.

[0103] The confidence level calculation process is as follows:

[0104] The matching confidence score and candidate component failure probability of the faulty component are obtained. The matching confidence score is obtained based on the mapping weight between the faulty component and the fault type in the knowledge graph. The candidate component failure probability is obtained from a list in the cloud, which includes the failure probabilities of candidate components of a second elevator that conform to preset rules with the first elevator. The confidence score of the corresponding candidate component is determined based on the matching confidence score, the candidate component failure probability, and the health score.

[0105] This embodiment focuses on the core step of quantitatively evaluating the reliability of faulty component determination after identifying faulty components with health scores below a health threshold. This step involves fusing knowledge graph mapping associations, global fault probabilities, and local component health status. The entire process is completed independently on the local node, and the specific execution process is as follows:

[0106] Obtain the matching confidence of faulty components: The local node first retrieves the pre-stored elevator fault domain knowledge graph and extracts the mapping weights between the previously determined faulty components and the corresponding fault types from the graph. These mapping weights are generated by the cloud teacher model based on the distillation of the full-domain elevator fault data and have a value range of [0,1]. The closer the value is to 1, the closer the feature association between the faulty component and the fault type is and the more stable the mapping relationship is. The mapping weight is directly used as the matching confidence of the faulty component to quantify the degree of matching between the fault type and the faulty component.

[0107] Obtaining candidate component failure probabilities: The local node retrieves the failure probabilities of candidate components corresponding to the second elevator that meets the preset rules from the locally stored elevator component failure probability list, which is aggregated and distributed by the cloud federation. The preset rules include that the second elevator and the first elevator are of the same brand, the same service life, the same operating conditions, and belong to the same system. This list is generated by the cloud through federated learning, which aggregates the historical failure statistics of all elevators that meet the rules. Only the failure probability statistics of each candidate component in the same type of elevators in the whole domain are retained (value range [0,1]). Without real-time interaction with the cloud, the local node directly obtains the failure probability value of the corresponding candidate component as a global failure probability reference indicator.

[0108] The weighted fusion calculation of the final confidence level is as follows: First, the obtained matching confidence level, candidate component failure probability, and previously calculated health score are standardized, and the health score (0-100 points) is linearly normalized to the [0,1] interval to unify the dimensions of the three indicators. Then, a differential weighting coefficient is preset, where the matching confidence level weight α=0.4 (reflecting the close correlation between the failure type and the component, with the highest weight), the candidate component failure probability weight β=0.35 (reflecting the failure probability of elevators of the same type in the whole domain), and the health score weight γ=0.25 (reflecting the health status of the local component itself). Finally, the final confidence level is calculated by weighted summation formula, that is, final confidence level = (α× matching confidence level) + (β× candidate component failure probability) + (γ× standardized health score value). The calculation result is limited to the range of [0,1]. If it needs to be converted to a percentage form, it is multiplied by 100 and one decimal place is retained. The higher the confidence level value, the more reliable and credible the judgment result of the failure component is.

[0109] The knowledge graph in this embodiment is a domain-specific knowledge network for elevator faults, built by the cloud based on a federated learning architecture, aggregating elevator fault data from across the entire domain. This network is constructed through knowledge distillation and structured modeling, and then encrypted and distributed to local nodes for storage. The knowledge graph uses fault type, elevator system, core components, and component aging attributes as entity nodes, and causal, attributional, and induced associations between nodes as connecting edges. Each edge is assigned a quantified mapping weight to characterize the degree of association between the fault type and the corresponding component. It not only incorporates a hierarchical attribution mapping relationship of "fault type—system—candidate component," but also integrates the corresponding association rules for single faults, joint faults, aging faults, and various components. When locating faulty components at local nodes, system matching and candidate component selection can be directly performed based on the input fault type, and the corresponding mapping weights can be extracted to calculate the matching confidence. This achieves local reuse of global fault knowledge without relying on real-time cloud interaction, while strictly adhering to federated learning data privacy protection requirements, providing standardized and structured knowledge support for accurate fault component determination.

[0110] This application provides an electronic device that may include a processing module and a memory. The memory stores a computer program, which, when executed by the processor, enables the electronic device to perform the corresponding steps in the aforementioned federated learning-based elevator fault diagnosis method.

[0111] In this embodiment, the processor can be an integrated circuit chip with signal processing capabilities. For example, the processor can be a central processing unit (CPU), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application.

[0112] The memory can be, but is not limited to, random access memory, read-only memory, programmable read-only memory, erasable programmable read-only memory, electrically erasable programmable read-only memory, etc.

[0113] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the electronic device described above can be referred to the corresponding steps in the aforementioned method, and will not be elaborated further here.

[0114] This application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program that, when run on a computer, causes the computer to execute the federated learning-based elevator fault diagnosis method as described in the above embodiments.

[0115] Computer-readable storage media may be magnetic disks, optical disks, read-only memory, random access memory, flash memory, USB flash drives, hard disks, or solid-state drives, etc., and may also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implement the methods shown in the above embodiments.

[0116] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described federated learning-based elevator fault diagnosis method. The computer program product may exist in a computer-readable storage medium in forms including, but not limited to, source files, executable files, and installation package files.

[0117] Based on the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by hardware or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application can be embodied in the form of a software product. This software product can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as a personal computer, electronic device, or network device, etc.) to execute the methods described in the various implementation scenarios of this application.

[0118] In the embodiments provided in this application, it should be understood that the disclosed methods can also be implemented in other ways. The method embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, program segment, or part of code, which includes one or more executable instructions for implementing a specified logical function. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions. Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0119] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for elevator fault diagnosis based on federated learning, characterized in that, Obtain the local vector of the first elevator through the local node; Based on a preset strategy, the local vector is aligned with the reference vector to obtain a corrected vector, wherein the reference vector is obtained from the teacher model in the cloud, and the local vector and the reference vector have the same dimension. Based on the correction vector, the fault type is obtained according to the student model of the local node; Obtain the abnormal parameters related to the fault type; Anomalies are obtained by calculating the first similarity between each fault template and anomaly parameters, based on each fault template pre-stored in the cloud on the local node. When the abnormal result indicates that there is a faulty component in the first elevator, the faulty component of the first elevator is determined based on the fault type and a preset knowledge graph, and an early warning is issued. The step of aligning the local vector with the reference vector based on a preset strategy to obtain the corrected vector includes: Calculate the second similarity between the local vector and the reference vector. When the second similarity is less than the first similarity threshold, extract the local key node time of the local vector and the reference key node time of the reference vector, and / or calculate the first reference magnitude of the local vector and the second reference magnitude of the reference vector. Calculate the time difference between the local critical node time and the reference critical node time, and / or calculate the scaling factor of the first reference amplitude relative to the second reference amplitude; The local vector is shifted as a whole based on the time difference, and / or each local feature in the local vector is scaled based on the scaling factor to obtain the corrected vector, so as to align the local vector with the reference vector.

2. The method according to claim 1, characterized in that, The correction vector includes vibration characteristics, current characteristics, and gantry crane characteristics; The fault type, obtained based on the correction vector and the student model of the local node, includes: The correction vector is input into the student model, and the vibration result corresponding to the vibration feature, the current result corresponding to the current feature, and the gantry machine result corresponding to the gantry machine feature are obtained through the student model. When one of the vibration result, current result, and gantry machine result is abnormal and the other two are normal, the third similarity and fourth similarity of the abnormal vibration result, current result, and gantry machine result are calculated with the other two normal vibration result, current result, and gantry machine result. Based on the vibration results, current results, and gantry crane results, the fault type is obtained according to the third similarity and the fourth similarity.

3. The method according to claim 2, characterized in that, Based on the vibration results, current results, and gantry crane results, and according to the third and fourth similarity scores, the fault type is obtained, including: When both the third and fourth similarities are less than the second similarity threshold, the fault type is obtained based on one of the abnormal vibration results, current results, and gantry crane results. When the third similarity is greater than the second similarity threshold and the fourth similarity is less than the second similarity threshold, the fault type is obtained based on one of the abnormal vibration results, current results, and gantry crane results, and one of the other two normal results corresponding to the second similarity. When the fourth similarity is greater than the second similarity threshold and the third similarity is less than the second similarity threshold, the fault type is obtained based on one of the abnormal vibration results, current results, and gantry crane results, and one of the other two normal results corresponding to the fourth similarity.

4. The method according to claim 1, characterized in that, After obtaining an anomaly result based on calculating a first similarity between each fault template pre-stored in the cloud and an anomaly parameter on the local node, the method further includes: Obtain fault templates whose first similarity is greater than the fifth similarity threshold; When the fault template with a similarity greater than the fifth similarity threshold indicates that the fault is caused by elevator aging, an abnormal result is obtained indicating that the fault of the first elevator is caused by aging. The step of determining the faulty component of the first elevator based on the fault type and the abnormal result and a preset knowledge graph is not executed.

5. The method according to claim 1, characterized in that, The step of determining the faulty component of the first elevator based on the fault type and a preset knowledge graph includes: The fault type is input into the knowledge graph, and the first elevator system corresponding to the fault type is obtained through matching in the knowledge graph; Based on the system to which it belongs, several candidate components are identified; The runtime, design life, number of historical failures, and time since the last maintenance are obtained for each candidate component in the local node. Based on the runtime, design life, number of historical failures, and time since the last maintenance, a health score is calculated for each candidate component. The health score is positively correlated with the health of the corresponding candidate component. Candidate components with a health score less than a health threshold are considered faulty components.

6. The method according to claim 5, characterized in that, After calculating the health score for each candidate component, the method further includes: When the number of faulty components is greater than or equal to 2, the matching confidence of the faulty components and the failure probability of candidate components are obtained. The matching confidence is obtained based on the mapping weight between the faulty components and the fault type in the knowledge graph. The failure probability of candidate components is obtained from a list in the cloud, which includes the failure probability of candidate components of a second elevator that conforms to preset rules with the first elevator. The confidence level of the corresponding candidate component is determined based on the matching confidence level, the failure probability of the candidate component, and the health score.

7. An electronic device, characterized in that, The electronic device includes a processor and a memory coupled together, the memory storing a computer program that, when executed by the processor, causes the electronic device to perform the method as described in any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when run on a computer, causes the computer to perform the method as described in any one of claims 1 to 6.

9. A program product, characterized in that, It includes a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 6.

Images