Escalator health assessment method and device based on deep learning and terminal equipment

Abstract

The invention is suitable for the technical field of escalator monitoring, and provides an escalator health assessment method and device based on deep learning and terminal equipment. The method comprises the steps that escalator operation data are collected through a sensor, and the escalator operation data comprise escalator vibration signals, escalator temperature, escalator operation noise and escalator operation current; local features used for representing the health state of the escalator are extracted from the escalator operation data through a convolutional neural network, the local features of various escalator operation data are fused through an attention mechanism, and global health state representation of the escalator is obtained; and an escalator domain knowledge graph is constructed, when the global health state characterization shows that the escalator state is abnormal, the reason of the escalator state abnormality is reasoned based on the relationship between entities in the escalator domain knowledge graph, and the escalator health condition is evaluated according to the reason of the escalator state abnormality. The escalator operation data can be comprehensively analyzed, the health state of the escalator is effectively evaluated, and the safety, reliability and operation efficiency of the escalator are improved.

Description

Technical Field

[0001] The present application relates to the field of escalator monitoring technology, and specifically to an escalator health assessment method, device and terminal equipment based on deep learning. Background Art

[0002] Escalators are a common vertical transportation equipment, widely used in public places such as shopping malls, subways, and airports. The safe operation of escalators is directly related to the safety of life and property of the public, so effective health monitoring and fault diagnosis of escalators are crucial. Traditional escalator maintenance usually adopts the method of regular inspection and manual inspection, which has problems such as long maintenance cycle, high labor cost, and untimely fault detection.

[0003] In recent years, with the development of the Internet of Things, big data and artificial intelligence technologies, predictive maintenance has begun to be applied in the field of escalators. By deploying various sensors on the escalator, collecting the escalator's operating data in real time, and analyzing the data using machine learning algorithms, it is possible to monitor the escalator's health status and provide early warning of failures. The inventors of this application found in their research that traditional escalator health monitoring methods are mainly based on a single type of data, such as vibration or temperature data, which is difficult to fully reflect the health status of the escalator; most methods use traditional machine learning models, such as support vector machines or decision trees, which have limited modeling capabilities for complex escalator systems. Summary of the invention

[0004] In view of this, the embodiments of the present application provide an escalator health assessment method, apparatus and terminal device based on deep learning, which can comprehensively analyze the escalator operation data and effectively evaluate the health status of the escalator.

[0005] In order to achieve the above objectives, this application adopts the following technical solutions:

[0006] In a first aspect, an embodiment of the present application provides an escalator health assessment method based on deep learning, comprising:

[0007] The escalator operation data is collected by a sensor, wherein the escalator operation data includes an escalator vibration signal, an escalator temperature, an escalator operation noise and an escalator operation current; wherein the collection frequency of the escalator operation data collected by the sensor and the transmission period of the escalator operation data are adjusted based on the escalator operation state and the information entropy of the escalator operation data;

[0008] Extracting local features for representing the health status of the escalator from the escalator operation data by using a convolutional neural network, and adaptively fusing the local features of various escalator operation data by using an attention mechanism to obtain a global health status representation of the escalator;

[0009] Construct an escalator domain knowledge graph, which contains entities and relationships between entities. Entities include escalator components, fault types, maintenance methods, and operating parameters. When constructing the escalator domain knowledge graph, a low-dimensional structural embedding representation of each entity and the relationship between entities is learned based on a graph convolutional network.

[0010] When the global health status representation indicates that the escalator state is abnormal, the cause of the abnormal escalator state is inferred based on the relationship between entities in the escalator domain knowledge graph, and the escalator health status is evaluated according to the cause of the abnormal escalator state.

[0011] Based on the first aspect, in some embodiments, when the global health status representation indicates that the escalator operation state is normal, the sensor collects escalator operation data at a first sampling frequency and transmits the escalator operation data at a first transmission period; when the global health status representation indicates that the escalator operation state is abnormal, the sensor collects escalator operation data at a second sampling frequency and transmits the escalator operation data at a second transmission period; wherein the second sampling frequency is greater than the first sampling frequency, and the second transmission period is less than the first transmission period;

[0012] And / or, when the model performance index decreases, the sensor increases the sampling frequency and reduces the transmission period; wherein the model performance index is the performance index of the model composed of the convolutional neural network and the attention mechanism, and the performance index of the model composed of the graph convolutional network and the escalator field knowledge graph;

[0013] And / or, when the information entropy of the escalator operation data increases, the sensor increases the sampling frequency and reduces the transmission period.

[0014] Based on the first aspect, in some embodiments, if the escalator speed at time t is Less than the escalator speed threshold v th , and the model performance index p at time t (t) Less than the model performance index threshold p th , then the sampling frequency at time t+1 is is the sampling frequency at time t, α is the adjustment factor, and 0<α<1;

[0015] If the escalator speed at time t is Greater than or equal to the escalator speed threshold v th , or the model performance index p at time t (t) Greater than or equal to the model performance index threshold p th , then the sampling frequency at time t+1 is β is the adjustment factor, and β>1.

[0016] Based on the first aspect, in some embodiments, the information entropy of the escalator operation data is X represents the data collected by the sensor, i represents the i-th value of the data, p(x i ) represents x i Probability of occurrence;

[0017] When the sensor transmits data, wavelet transform is used to compress the time series data of escalator operation, and principal component analysis is used to select the most important features for transmission.

[0018] Based on the first aspect, in some embodiments, the adaptively fusing local features of various escalator operation data through the attention mechanism to obtain the global health status representation of the escalator includes:

[0019] pass By integrating the local features of various escalator operation data, the global health status representation of the escalator is obtained; where v (a) is the global health status representation vector of the escalator, represents the feature vector of the data collected by the i-th sensor, is the attention weight of the i-th sensor, N is the number of sensors, and is the hidden representation of the data collected by the i-th sensor, u (a) is the attention vector, and The default parameters for linear transformation.

[0020] Based on the first aspect, in some embodiments, before adaptively fusing the local features of various escalator operation data through the attention mechanism, the method further includes:

[0021] Adaptively weighting the local features of each sensor according to their importance through the gating unit, and then performing the step of adaptively fusing the local features of various escalator operation data through the attention mechanism;

[0022] The process of adaptively weighting the local features of each sensor according to their importance by the gate control unit includes: Adaptively weight the local features of the data collected by each sensor. is the local feature of the data collected by the i-th sensor after adaptive weighting, is the local feature of the data collected by the ith sensor before adaptive weighting, is the gating signal corresponding to the i-th sensor in the gating unit, ⊙ represents element-by-element multiplication, and is the preset parameter of the gate unit, and σ is the sigmoid activation function.

[0023] Based on the first aspect, in some embodiments, the method further includes: fusing the attribute information of the entity with the low-dimensional structure embedding representation to obtain an enhanced entity; wherein, the attribute feature of the i-th entity is The low-dimensional structure embedding is represented as Then the enhanced entity e i for: and are the transformation matrices representing the attribute information and low-dimensional structure embedding respectively;

[0024] The reason for the abnormal escalator state based on the relationship between entities in the escalator domain knowledge graph includes: Reasoning about the abnormal state of the escalator, The reason for the abnormal escalator status is inferred. For the target entity The enhancement indicates that is the parameter matrix of the attention mechanism, | represents vector concatenation, For the target entity Related related entities The information obtained after weighted aggregation is For the target entity Related entities The attention weights between is the parameter matrix of the attention mechanism.

[0025] Based on the first aspect, in some embodiments, the method further includes:

[0026] For multiple escalator operating units, each escalator operating unit adopts a federated learning framework to locally train a model using its own escalator operation data; wherein the model is a model composed of the convolutional neural network and the attention mechanism, and a model composed of the graph convolutional network and the escalator domain knowledge graph;

[0027] After the training is completed, the escalator operating units exchange the trained model parameters through an encrypted communication protocol. In the process of training the model through the federated learning framework, differential privacy technology is introduced to add random noise to the gradient information shared between the escalator operating units.

[0028] In a second aspect, an embodiment of the present application provides an escalator health assessment device based on deep learning, comprising:

[0029] A data acquisition module, used for collecting escalator operation data through sensors, wherein the escalator operation data includes escalator vibration signal, escalator temperature, escalator operation noise and escalator operation current; wherein the collection frequency of the escalator operation data collected by the sensor and the transmission period of the escalator operation data are adjusted based on the escalator operation state and the information entropy of the escalator operation data;

[0030] A feature extraction and fusion module is used to extract local features for characterizing the health status of the escalator from the escalator operation data through a convolutional neural network, and adaptively fuse the local features of various escalator operation data through an attention mechanism to obtain a global health status characterization of the escalator;

[0031] A knowledge graph construction module is used to construct an escalator domain knowledge graph, which contains entities and relationships between entities. Entities include escalator components, fault types, maintenance methods, and operating parameters. When constructing the escalator domain knowledge graph, a low-dimensional structural embedding representation of each entity and the relationship between entities are learned based on a graph convolutional network.

[0032] The health assessment module is used to infer the cause of the abnormal escalator state based on the relationship between entities in the escalator domain knowledge graph when the global health status representation indicates that the escalator state is abnormal, and evaluate the escalator health status according to the cause of the abnormal escalator state.

[0033] In a third aspect, an embodiment of the present application provides a terminal device, including a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the computer program, the deep learning-based escalator health assessment method as described in any one of the first aspects is implemented.

[0034] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0035] In an embodiment of the present application, escalator operation data is collected by sensors, and the escalator operation data includes escalator vibration signals, escalator temperature, escalator operation noise, and escalator operation current; local features used to characterize the health status of the escalator are extracted from the escalator operation data through a convolutional neural network, and the local features of various escalator operation data are adaptively integrated through an attention mechanism to obtain a global health status representation of the escalator; an escalator domain knowledge graph is constructed, and when the global health status representation indicates that the escalator state is abnormal, the cause of the escalator state abnormality is inferred based on the relationship between entities in the escalator domain knowledge graph, and the escalator health status is evaluated according to the cause of the escalator state abnormality. The embodiment of the present application can comprehensively analyze the escalator operation data, effectively evaluate the escalator health status, promptly discover potential faults and give maintenance decision suggestions, thereby improving the safety, reliability and operation efficiency of the escalator.

[0036] Compared with traditional rule-based or single-model methods, the embodiments of the present application can mine the health degradation laws contained in the data from multiple dimensions, and have stronger predictive ability and robustness. At the same time, the embodiments of the present application adopt privacy protection measures such as data encryption, federated learning, and differential privacy during model training and deployment, which can maximize the data security of all parties while ensuring model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0038] Figure 1 It is a flowchart of an escalator health assessment method based on deep learning provided in an embodiment of the present application;

[0039] Figure 2 is a schematic diagram of a multi-view fusion model based on a convolutional neural network provided in an embodiment of the present application;

[0040] Figure 3 is a schematic diagram of a multi-view fusion model that integrates a gating unit and an attention mechanism provided in an embodiment of the present application;

[0041] Figure 4 It is a structural schematic diagram of an escalator health assessment device based on deep learning provided in an embodiment of the present application;

[0042] Figure 5 It is a schematic diagram of a terminal device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0043] The present application is described more clearly below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the effects of the present application, but are not intended to limit the present application in any form. It should be noted that, for those of ordinary skill in the art, several variations and improvements may be made without departing from the concept of the present application. These all fall within the scope of protection of the present application.

[0044] In order to make the purpose, technical solutions and advantages of the present application clearer, specific embodiments will be described below in conjunction with the accompanying drawings.

[0045] See also Figure 1 The escalator health assessment method based on deep learning provided in the embodiment of the present application may include the following steps:

[0046] Step 101, collecting escalator operation data through sensors, wherein the escalator operation data includes escalator vibration signal, escalator temperature, escalator operation noise and escalator operation current.

[0047] For example, various types of sensors such as vibration sensors, temperature sensors, noise sensors, current sensors, etc. can be installed on the key components of the escalator to collect escalator operation data in real time.

[0048] The frequency of the sensor collecting the escalator operation data and the transmission period of the escalator operation data are adjusted based on the escalator operation state and the information entropy of the escalator operation data. Specifically, an adaptive data sampling and transmission strategy can be adopted to dynamically adjust the sampling frequency and data transmission period of the sensor according to the escalator operation state (such as speed, passenger capacity, etc.) and the performance indicators of the deep learning model (such as loss function, accuracy, etc.).

[0049] This embodiment involves three deep learning models, namely a multi-view fusion model based on a convolutional neural network (CNN), a knowledge graph embedding model based on a graph convolutional network (GCN), and a distributed training model based on federated learning. These three deep learning models will be introduced one by one in the corresponding content later.

[0050] In one scenario, when the global health status characterization indicates that the escalator is in normal operation, the sensor collects escalator operation data at a first sampling frequency and transmits the escalator operation data at a first transmission period; when the global health status characterization indicates that the escalator is abnormal, the sensor collects escalator operation data at a second sampling frequency and transmits the escalator operation data at a second transmission period. The second sampling frequency is greater than the first sampling frequency, and the second transmission period is less than the first transmission period. That is, when the escalator is in normal operation, the sensor reduces the sampling frequency and transmission period to reduce the amount of data and transmission load.

[0051] For example, if the escalator speed v at time t is s (t) Less than the escalator speed threshold v th , and the model performance index p at time t (t) Less than the model performance index threshold p th , then the sampling frequency at time t+1 is is the sampling frequency at time t, α is the adjustment factor, and 0<α<1.

[0052] If the escalator speed at time t is Greater than or equal to the escalator speed threshold v th , or the model performance index p at time t (t) Greater than or equal to the model performance index threshold p th, then the sampling frequency at time t+1 is β is the adjustment factor, and β>1.

[0053] In another scenario, when the model performance index decreases, the sensor increases the sampling frequency and reduces the transmission period. The model performance index is the performance index of the model composed of the above-mentioned convolutional neural network and the above-mentioned attention mechanism, and the performance index of the model composed of the above-mentioned graph convolutional network and the above-mentioned escalator field knowledge graph. The performance index can specifically be a loss function, accuracy, etc. When the model performance index decreases, increasing the sampling frequency and transmission period can obtain data with higher accuracy and real-time performance.

[0054] In another scenario, when the information entropy of the escalator operation data increases, the sensor increases the sampling frequency and reduces the transmission period. When adjusting the sensor sampling frequency and transmission period, not only the escalator operation status and model performance indicators can be considered, but also the information entropy of the data. When the information entropy of the data is high, it means that the data contains more useful information. At this time, the sampling frequency can be appropriately increased and the transmission period can be reduced; conversely, when the information entropy of the data is low, it means that the data redundancy is high. At this time, the sampling frequency can be reduced and the transmission period can be increased to reduce the amount of data.

[0055] Specifically, the information entropy of escalator operation data can be X represents the data collected by the sensor, i represents the i-th value of the data, p(x i ) represents x i Probability of occurrence.

[0056] In addition, when the sensor transmits data, wavelet transform can be used to compress the time series data of escalator operation, and principal component analysis can be used to select the most important features for transmission.

[0057] During data transmission, this embodiment introduces data compression and feature selection technology to reduce the amount of data transmitted. Specifically, methods such as wavelet transform can be used to compress the time series data of escalator operation to reduce the data length. Afterwards, methods such as principal component analysis (PCA) can be used to select the most important features from the compressed data for transmission to reduce the data dimension. Assume that the original data is X∈R n×m , where n is the number of sampling points and m is the feature dimension. Perform wavelet transform on X to get the compressed data Where n c <n. for X c Perform PCA to obtain the data after dimensionality reduction Where m r <m. The final transmitted data X r While retaining key information, the amount of data is greatly reduced, reducing transmission load and storage overhead.

[0058] In another scenario, when the global health status representation indicates that the escalator operation status is abnormal and the model performance indicators decrease, the sampling frequency of the sensor can be increased and the transmission cycle can be reduced, so that higher-precision and real-time data can be obtained to monitor the operation status of the escalator.

[0059] In some embodiments, a distributed big data platform (such as Hadoop, Spark, etc.) can also be used to store and manage the massive data collected by the sensor, supporting efficient data reading, writing and query operations. By designing a reasonable data storage format and indexing mechanism, it can meet the real-time writing of data and support fast data retrieval and analysis.

[0060] Step 102, extracting local features for characterizing the health status of the escalator from the escalator operation data through a convolutional neural network, and adaptively fusing the local features of various escalator operation data through an attention mechanism to obtain a global health status characterization of the escalator.

[0061] Specifically, after obtaining the raw data collected by the sensor, the raw data is cleaned, denoised, normalized and other preprocessing operations are performed to improve the data quality. Then, using multiple feature extraction methods such as time domain, frequency domain, time-frequency domain, etc., key features that can reflect the health status of the escalator are extracted from multiple sensor data such as vibration, temperature, noise, current, etc., such as the root mean square value, peak factor, spectral entropy, etc. of the vibration signal, as well as the trend degree and mutation point of the temperature signal.

[0062] The convolutional neural network and attention mechanism in this step are the first deep learning model mentioned above: a multi-view fusion model based on a convolutional neural network (CNN), which is used to fuse features extracted by different sensors and learn a unified representation of the escalator’s health status. Figure 2 As shown in the figure, the deep learning model designs an independent CNN sub-network for each sensor data to extract local features, and then adaptively fuses the features extracted by different sub-networks through the attention mechanism to obtain the global health status representation of the escalator.

[0063] In some embodiments, the step 102 of "adaptively fusing the local features of various escalator operation data through the attention mechanism to obtain the global health status representation of the escalator" may include: By integrating the local features of various escalator operation data, the global health status representation of the escalator is obtained; where v (a) is the global health status representation vector of the escalator, represents the feature vector of the data collected by the i-th sensor, is the attention weight of the i-th sensor, N is the number of sensors, and is the hidden representation of the data collected by the i-th sensor, u (a) is the attention vector, and The default parameters for linear transformation.

[0064] Through the attention mechanism, the system can adaptively adjust the importance of different sensor features. For example, in some cases, the vibration feature may receive a higher attention weight, indicating that the current health status of the escalator is mainly determined by the vibration data. This dynamic adjustment capability enables the system to more flexibly and accurately evaluate the overall health status of the escalator.

[0065] In one example, 40% of the attention weight can be allocated to vibration information, 20% of the attention weight can be allocated to temperature information, 10% of the attention weight can be allocated to noise information, and 30% of the attention weight can be allocated to current information. This allocation indicates that in the current situation, the system considers vibration and current information to be the most important, while noise information is relatively less critical. The system will combine all this information to form a global health status representation. This global health status representation can include multiple aspects of information, such as the overall state of the mechanical system, the smoothness of operation, energy efficiency, and potential failure risks.

[0066] In this way, the system can flexibly adjust the importance of various information according to different situations, so as to obtain a comprehensive and accurate assessment of the escalator health status. The advantage of this method is that it can automatically adapt to different situations and does not require fixed rules to be set manually, so it can better handle complex and changing situations.

[0067] As an implementable embodiment, before adaptively fusing the local features of various escalator operation data through the attention mechanism, the above-mentioned deep learning-based escalator health assessment method may also include: adaptively weighting the local features of each sensor according to importance through a gating unit, and then executing the step of adaptively fusing the local features of various escalator operation data through the attention mechanism.

[0068] Specifically, the process of adaptively weighting the local features of each sensor according to their importance by the gate control unit may include: Adaptively weight the local features of the data collected by each sensor. is the local feature of the data collected by the i-th sensor after adaptive weighting, is the local feature of the data collected by the ith sensor before adaptive weighting, is the gating signal corresponding to the i-th sensor in the gating unit, ⊙ represents element-by-element multiplication,

[0069] and is the preset parameter of the gate unit, and σ is the sigmoid activation function.

[0070] For example, Figure 3 As shown in the figure, a corresponding CNN sub-network is set for each sensor to extract the local features of the data collected by each sensor. After each CNN sub-network extracts the local features, the local features of the data collected by different sensors are adaptively weighted through the gating unit, and then cross-view fusion is performed through the attention mechanism. The gating unit automatically adjusts the weight according to the importance of the local features of the data collected by each sensor to highlight the role of key sensors. The attention mechanism calculates the fusion weight according to the correlation between sensors and utilizes the complementary information of different sensors.

[0071] When cross-view attention fusion is performed, the correlation between the viewpoints (i.e., sensors) is calculated first, and then the fusion weight is calculated based on the correlation. The calculation formula for the correlation between the i-th sensor and the j-th sensor is:

[0072]

[0073] in, is the parameter matrix for correlation calculation, is the correlation between the i-th sensor and the j-th sensor.

[0074] The calculation formula of fusion weight is:

[0075]

[0076]

[0077] in, is the fusion weight of the i-th sensor, v (c) Represents the final health status.

[0078] Through the synergy of the gating unit and the attention mechanism, the importance of different sensor data can be adaptively adjusted, and the complementary information from different perspectives can be fully utilized to obtain a more comprehensive and accurate escalator health status assessment result.

[0079] Step 103, constructing an escalator domain knowledge graph, which includes entities and relationships between entities, and the entities include escalator components, fault types, maintenance methods, and operating parameters.

[0080] The knowledge graph in the escalator field is a structured knowledge base that contains various entities related to escalators and the relationships between them. The specific content of the knowledge graph in the escalator field can include entity types, relationship types, attribute information, rules and constraints, and historical data.

[0081] 1. Entity types may include escalator components, fault types, maintenance methods, and operating parameters.

[0082] Escalator components include: drive system (including motor, reducer, drive chain, etc.), step system (including steps, step chain, etc.), handrail system (including handrail belt, handrail drive device, etc.), control system (including control cabinet, sensor, safety switch, etc.), truss structure (including truss, cover plate, etc.), braking system (including working brake, additional brake, etc.).

[0083] Fault types include: mechanical failure (such as stuck steps, broken handrails, etc.), electrical failure (such as motor overheating, control system failure, etc.), safety device failure (such as safety switch failure, etc.), abnormal noise, abnormal vibration, and abnormal speed.

[0084] Maintenance methods include: daily inspection, regular maintenance, fault repair, parts replacement, system upgrade, etc.

[0085] Operating parameters include: speed, load, operating time, energy consumption, etc.

[0086] 2. Relationship types can include component relationships, fault-symptom relationships, fault-cause relationships, fault-solution relationships, component-parameter relationships, and maintenance-cycle relationships.

[0087] Composition relationship: for example, "drive system-composition-motor", the drive system is composed of the motor.

[0088] Fault-symptom relationship: For example, "step stuck - causes - abnormal vibration", step stuck causes abnormal vibration.

[0089] Fault-cause relationship: for example, "insufficient lubrication-causes-abnormal noise", insufficient lubrication causes abnormal noise.

[0090] Fault-solution relationship: For example, "Broken handrail - Need to replace handrail", the solution to a broken handrail is to replace the handrail.

[0091] Component-parameter relationship: for example, "motor-impact-energy consumption", where energy consumption is a parameter of the motor.

[0092] Maintenance-cycle relationship: For example, "Lubrication-Recommended-Monthly", the lubrication maintenance cycle is monthly maintenance.

[0093] 3. Attribute information may include component attributes, fault attributes, maintenance attributes, etc.

[0094] Component attributes: e.g. manufacturer, model, age, etc. Failure attributes: e.g. severity, frequency of occurrence, etc. Maintenance attributes: e.g. tools required, estimated time, skill requirements, etc.

[0095] 4. Rules and constraints can include safety regulations, maintenance rules, etc.

[0096] Safety regulations: For example, "The load must not exceed the rated value." Maintenance rules: For example, "A safety test must be performed after replacing a part."

[0097] 5. Historical data may include: past fault records, maintenance history, parts replacement records, etc.

[0098] By building such a comprehensive knowledge graph in the escalator field, the system can achieve the following:

[0099] 1. Diagnose faults more accurately: For example, quickly locate problems by analyzing symptoms and possible causes.

[0100] 2. Provide targeted maintenance suggestions: for example, provide appropriate maintenance methods based on the type and severity of the fault.

[0101] 3. Predict potential problems: For example, predict possible failures based on component life and historical data.

[0102] 4. Optimize maintenance plan: For example, formulate a reasonable maintenance plan based on the status and maintenance cycle of each component.

[0103] 5. Support decision making: for example, provide comprehensive information support to managers to assist in decision making.

[0104] The above-mentioned knowledge graph in the escalator field not only contains static structured information, but can also be continuously updated and expanded through machine learning algorithms, incorporating dynamic information such as new failure modes and maintenance experience, thereby continuously improving the intelligence level and adaptability of the system.

[0105] When constructing the escalator domain knowledge graph, the low-dimensional structure embedding representation of each entity and the relationship between entities are learned based on the graph convolutional network. In this step, the escalator domain knowledge graph and the graph convolutional network constitute the second deep learning model mentioned above: the knowledge graph embedding model based on the graph convolutional network (GCN).

[0106] The deep learning model first constructs a knowledge graph in the escalator field, which includes entities and their relationships such as escalator component structure, fault type, and maintenance method. Then, GCN is used to learn a low-dimensional embedding representation of each entity while considering the semantic associations between entities. When an abnormal state of the escalator is identified, the fault propagation relationship in the knowledge graph is used to infer the possible cause to assist in fault diagnosis and prediction. The forward propagation process of GCN is:

[0107]

[0108] Among them, H (l) is the node representation matrix of the lth layer, To add the adjacency matrix of self-loops, for The degree matrix, W (l) is the weight matrix of the lth layer, and σ is the activation function.

[0109] Low-dimensional structural embedding maps entities from the original high-dimensional space (which may contain a large number of attributes and relationships) to a low-dimensional vector space. For example, the original "motor" entity may have many attributes (model, power, usage time, etc.). Through low-dimensional structural embedding, the motor can be represented as a vector such as [0.5, 0.3, -0.2, 0.8], which can capture the key features of the motor and its relationship with other entities.

[0110] The relationship between entities can be specifically a semantic association between entities, which refers to various relationships between entities. In escalator data, it can include: "component" relationship, such as the motor is part of the drive system; "cause" relationship, such as bearing wear may cause abnormal vibration; "influence" relationship, such as load increase will affect energy consumption; "need" relationship, such as abnormal noise may require lubrication maintenance.

[0111] Through low-dimensional structural embedding, a concise vector can be used to represent the complex escalator state, which contains information such as component status, operating parameters, potential faults, etc. Graph convolutional networks can learn implicit relationships between entities, such as abnormal combinations of certain parameters may indicate specific faults; for example, increased vibration and increased energy consumption may jointly point to bearing problems. By learning semantic associations between entities, knowledge of known faults can be transferred to new situations; if it learns that "bearing wear causes vibration", it may be able to infer that wear of other similar components may also cause vibration. Low-dimensional structural embedding allows the system to consider sensor data, historical fault records, maintenance information, etc. at the same time. This fusion can provide more comprehensive and accurate abnormal state identification. By analyzing the relationship between entities, the system can provide more meaningful explanations; for example, not only indicating the existence of an anomaly, but also explaining the possible causes and effects of this anomaly.

[0112] By setting up a knowledge graph in the escalator field, the system can more intelligently understand the overall status of the escalator, not only identifying current anomalies, but also predicting potential problems and providing targeted maintenance suggestions. It combines the advantages of data-driven and knowledge-driven, making the identification of abnormal conditions more accurate, comprehensive and explainable.

[0113] In some embodiments, the above-mentioned escalator health assessment method based on deep learning may further include: fusing the attribute information of the entity with the low-dimensional structure embedding representation to obtain an enhanced entity; wherein, the attribute feature of the i-th entity is The low-dimensional structure embedding is represented as Then the enhanced entity e i for: and They are the transformation matrices for embedding representation of attribute information and low-dimensional structure, respectively.

[0114] Correspondingly, the above-mentioned inference of the reasons for the abnormal escalator status based on the relationship between entities in the escalator domain knowledge graph may include: Reasoning about the abnormal state of the escalator, The reason for the abnormal escalator status is inferred. For the target entity The enhancement indicates that is the parameter matrix of the attention mechanism, | represents vector concatenation, For the target entity Related related entities The information obtained after weighted aggregation is For the target entity Related entities The attention weights between is the parameter matrix of the attention mechanism.

[0115] Step 104, when the global health status representation indicates that the escalator state is abnormal, the cause of the escalator state abnormality is inferred based on the relationship between entities in the escalator domain knowledge graph, and the escalator health status is evaluated according to the cause of the escalator state abnormality.

[0116] For example, after identifying that the escalator state is abnormal, the escalator domain knowledge graph is used to assist in fault diagnosis and prediction and infer the cause of the escalator state abnormality, which can be divided into the following steps:

[0117] 1. Abnormal state mapping:

[0118] First, the identified abnormal state is mapped to the relevant entity in the escalator domain knowledge graph. For example, if abnormal vibration is detected, the entity "abnormal vibration" is located in the escalator domain knowledge graph.

[0119] 2. Related entity retrieval:

[0120] Retrieve entities directly related to abnormal status, including possible causes, affected components, related parameters, etc. For example, for "abnormal vibration", possible related entities include: possible causes include bearing wear, step looseness, motor imbalance, etc.; affected components include drive system, step system, etc.; related parameters include vibration frequency, vibration amplitude, etc.

[0121] 3. Path analysis:

[0122] Analyze the relationships between entities in the escalator domain knowledge graph to find paths from abnormal states to possible causes. These paths can include: direct relationships, such as "abnormal vibration" is directly connected to "bearing wear"; indirect relationships, such as "abnormal vibration" is indirectly connected to "motor imbalance" through "drive system".

[0123] 4. Probabilistic Reasoning:

[0124] Based on historical data and expert knowledge, a probability is assigned to each possible fault path. For example, the probability that "abnormal vibration" is caused by "bearing wear" is 60%; the probability that "abnormal vibration" is caused by "loose steps" is 30%; the probability that "abnormal vibration" is caused by "motor imbalance" is 10%.

[0125] 5. Evidence Integration:

[0126] The system integrates other sensor data and operating parameters as additional evidence. For example, if a temperature increase is detected at the same time, it can increase the probability of "bearing wear"; if the noise level also increases, it can further support the hypothesis of "step loosening".

[0127] 6. Multi-source information fusion:

[0128] Combining the inference results of the escalator domain knowledge graph and the output of the deep learning model, the system performs multi-source information fusion. This can be achieved by: using Bayesian networks to integrate probabilistic information from different sources; using fuzzy logic methods to handle uncertain information; and using methods such as decision trees or random forests to integrate multiple factors.

[0129] 7. Generation of diagnostic results:

[0130] Based on the above analysis, the system generates a ranked fault diagnosis list, which may include: possible fault causes and their probabilities, severity assessment of each fault, and recommended further inspection or test items.

[0131] 8. Predictive Analytics:

[0132] Using the causal relationships and historical data in the knowledge graph, the system can also perform predictive analysis: evaluate the chain reactions that may be caused by the current failure, predict the impact that other components may be affected, and estimate the possible development trend of the failure if no maintenance is performed.

[0133] 9. Maintenance suggestion generation:

[0134] Finally, based on the diagnosis and prediction results and the maintenance information in the knowledge graph, the system generates maintenance recommendations, including: providing specific repair steps for the most likely causes of failure, recommending preventive maintenance measures to avoid potential problems from worsening, and providing a prioritized maintenance task list.

[0135] 10. Continuous learning and updating:

[0136] The system records the results of each diagnosis and actual maintenance status, which are used to update the probabilities and relationships in the escalator knowledge graph, optimize the fault diagnosis and prediction models, and accumulate new failure modes and solutions.

[0137] Through this process, the escalator domain knowledge graph not only provides structured domain knowledge, but can also be combined with data-driven anomaly detection methods to achieve a more accurate, explainable and predictive fault diagnosis system. This approach is particularly suitable for dealing with complex systems such as escalators because it can capture the complex interactions between various components and potential fault propagation paths.

[0138] In some embodiments, the above-mentioned deep learning-based escalator health assessment method may also include: for multiple escalator operating units, each escalator operating unit adopts a federated learning framework to locally train a model using its own escalator operation data; wherein the model is a model composed of the convolutional neural network and the attention mechanism, and a model composed of the graph convolutional network and the escalator domain knowledge graph; after the training is completed, the escalator operating units exchange the trained model parameters through an encrypted communication protocol; wherein, in the process of training the model through the federated learning framework, differential privacy technology is introduced to add random noise to the gradient information shared between the escalator operating units.

[0139] Among them, the gradient is the partial derivative of the loss function with respect to the model parameters. During the model training process, the gradient information indicates how to adjust the parameters to reduce the loss and is used to update the model parameters. Each escalator operating unit does not directly share the original data or complete model parameters, but shares the gradient information or parameter updates. This shared information is used to aggregate and update the global model on the central server.

[0140] This embodiment involves the third deep learning model mentioned above: a distributed training model based on federated learning. This deep learning model uses a federated learning framework between multiple escalator operating units to achieve collaborative training of the model. Each operating unit uses its own data to train CNN and GCN models locally, and then exchanges model parameters through encrypted communication protocols (such as homomorphic encryption) to optimize the global model without sharing the original data. In order to protect data privacy, differential privacy technology is introduced in the federated learning process to add random noise to the shared gradient information.

[0141] In some embodiments, model interpretation and decision support are also involved, using technologies such as SHAP (SHapley Additive exPlanations) to analyze the prediction results of the deep learning model, calculate the contribution of each input feature to the prediction results, and provide an interpretable report. At the same time, combined with domain expert knowledge and historical maintenance data, the module gives a comprehensive score of the escalator's health status, as well as maintenance priorities and maintenance decision recommendations for potential faults.

[0142] In some embodiments, the above-mentioned deep learning-based escalator health assessment method also includes: displaying the real-time operating status, health metrics, prediction results and other information of the escalator through web and mobile applications, so that managers can understand the operation status of the escalator; when an abnormality is detected in the escalator status or an impending failure is predicted, an alarm notification is issued to relevant personnel to facilitate rapid response and disposal.

[0143] In some embodiments, the above-mentioned deep learning-based escalator health assessment method further includes: using active learning strategies, selecting sample points that have the greatest impact on the judgment results according to the uncertainty of model prediction, actively requesting expert annotations or conducting experimental verification, so as to expand the training data in a targeted manner. In addition, the deep learning model can be incrementally trained and fine-tuned regularly using newly collected data, so that the deep learning model can adapt to changes in escalator working conditions.

[0144] In some embodiments, before extracting local features for characterizing the health status of the escalator from the escalator operation data through a convolutional neural network, the above-mentioned escalator health assessment method based on deep learning may also include: performing spectral analysis on the escalator operation data through fast Fourier transform, and calculating the spectral energy and spectral entropy of the escalator operation data; comparing the spectral energy with a preset spectral energy range, and comparing the spectral entropy with a preset spectral entropy range; if the spectral energy exceeds the preset spectral energy range, and / or the spectral entropy exceeds the preset spectral entropy range, inputting the escalator operation data into a convolutional neural network, and extracting local features for characterizing the health status of the escalator from the escalator operation data through the convolutional neural network.

[0145] The spectral energy and spectral entropy of escalator operation data are calculated by fast Fourier transform, and the escalator operation data that may have abnormalities are pre-determined according to the spectral energy and spectral entropy. The escalator operation data that may have abnormalities are then input into the convolutional neural network to extract local features for characterizing the health status of the escalator. This can reduce the data processing amount of the convolutional neural network and improve the processing efficiency.

[0146] Specifically, you can Calculate the spectrum energy E of the escalator operation data by Calculate the spectrum entropy H of the escalator operation data, x(t) is the escalator operation data.

[0147] The above escalator health assessment method based on deep learning has the following beneficial effects:

[0148] 1. Comprehensively utilize multi-source heterogeneous data to comprehensively evaluate the health status of escalators. The present invention uses multimodal sensors to collect escalator operation data, covering multiple dimensions such as vibration, temperature, noise, and current. Through a deep learning model, feature extraction and fusion of different modal data are achieved, and a more comprehensive and accurate escalator health assessment model is constructed, which effectively improves the accuracy of fault detection and early warning.

[0149] 2. Adaptively adjust data sampling and transmission strategies to improve data processing efficiency. The present invention adaptively adjusts the sensor sampling frequency and data transmission cycle according to the escalator operation status and model performance, reduces data redundancy and transmission overhead while ensuring data quality, and improves the real-time performance and efficiency of the system.

[0150] 3. Introduce domain knowledge to enhance the interpretability of fault diagnosis and prediction. The present invention constructs a knowledge graph in the escalator domain, embeds the knowledge graph into a deep learning model, and uses a graph convolutional network to learn the representation of entities and relationships, so that the model can use domain knowledge for reasoning and decision-making, thereby improving the interpretability and credibility of fault diagnosis and prediction results.

[0151] 4. Adopting a federated learning framework to achieve secure sharing and collaborative modeling of multi-party data. The present invention adopts a federated learning framework between multiple escalator operating units. Each unit uses its own data to train the model locally, only sharing model parameters instead of original data. This protects data privacy while achieving collaborative training and optimization of the model, and improves the generalization performance of the model.

[0152] 5. Introduce active learning and incremental learning to achieve continuous optimization and adaptive updating of the model. The present invention adopts an active learning strategy to select the most valuable samples for annotation and learning, which improves the annotation efficiency and model performance. At the same time, the incremental learning method is used to regularly fine-tune and update the model with newly collected data, so that the model can adapt to changes in escalator working conditions and maintain a high prediction accuracy.

[0153] 6. Provide visual and interactive decision support to facilitate monitoring and maintenance by managers. The present invention has developed web and mobile applications to display the real-time operating status, health indicators, warning information, etc. of the escalator in an intuitive and interactive way, and provide maintenance decision suggestions based on the diagnosis results, providing managers with an efficient and convenient decision support tool.

[0154] The above-mentioned escalator health assessment method based on deep learning, through key technologies such as adaptive data sampling, multi-view feature fusion, knowledge graph embedding, federated learning, and active learning, can comprehensively analyze the escalator operation data, effectively evaluate the escalator health status, timely discover potential faults and give maintenance decision suggestions, thereby improving the safety, reliability and operation efficiency of the escalator. Compared with traditional rule-based or single model-based methods, the embodiments of the present application can mine the health degradation laws contained in the data from multiple dimensions, and have stronger predictive ability and robustness. At the same time, the embodiments of the present application adopt privacy protection measures such as data encryption, federated learning, and differential privacy during the model training and deployment process, which can maximize the protection of data security of all parties while ensuring model performance.

[0155] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0156] Corresponding to the escalator health assessment method based on deep learning described in the above embodiment, Figure 4 A structural block diagram of an escalator health assessment device based on deep learning provided in an embodiment of the present application is shown. For ease of explanation, only the parts related to the embodiment of the present application are shown.

[0157] See also Figure 4 , an embodiment of the present application provides an escalator health assessment device based on deep learning, including a data acquisition module 201, a feature extraction and fusion module 202, a knowledge graph construction module 203 and a health assessment module 204.

[0158] The data acquisition module 201 is used to collect escalator operation data through sensors, and the escalator operation data includes escalator vibration signal, escalator temperature, escalator operation noise and escalator operation current. The acquisition frequency of the escalator operation data collected by the sensor and the transmission period of the escalator operation data are adjusted based on the escalator operation state and the information entropy of the escalator operation data.

[0159] The feature extraction and fusion module 202 is used to extract local features for characterizing the health status of the escalator from the escalator operation data through a convolutional neural network, and adaptively fuse the local features of various escalator operation data through an attention mechanism to obtain a global health status characterization of the escalator.

[0160] The knowledge graph construction module 203 is used to construct an escalator domain knowledge graph, which contains entities and relationships between entities. The entities include escalator components, fault types, maintenance methods, and operating parameters. When constructing the escalator domain knowledge graph, the low-dimensional structure embedding representation of each entity and the relationship between entities are learned based on the graph convolutional network.

[0161] The health assessment module 204 is used to infer the cause of the abnormal escalator state based on the relationship between entities in the escalator domain knowledge graph when the global health status representation indicates that the escalator state is abnormal, and evaluate the escalator health status according to the cause of the abnormal escalator state.

[0162] Optionally, when the global health status representation indicates that the escalator operation state is normal, the sensor collects escalator operation data at a first sampling frequency and transmits the escalator operation data at a first transmission period; when the global health status representation indicates that the escalator operation state is abnormal, the sensor collects escalator operation data at a second sampling frequency and transmits the escalator operation data at a second transmission period; wherein the second sampling frequency is greater than the first sampling frequency, and the second transmission period is less than the first transmission period;

[0163] And / or, when the model performance index decreases, the sensor increases the sampling frequency and reduces the transmission period; wherein the model performance index is the performance index of the model composed of the convolutional neural network and the attention mechanism, and the performance index of the model composed of the graph convolutional network and the escalator field knowledge graph;

[0164] And / or, when the information entropy of the escalator operation data increases, the sensor increases the sampling frequency and reduces the transmission period.

[0165] Optionally, if the escalator speed at time t is Less than the escalator speed threshold v th , and the model performance index p at time t (t) Less than the model performance index threshold p th, then the sampling frequency at time t+1 is is the sampling frequency at time t, α is the adjustment factor, and 0<α<1;

[0166] If the escalator speed at time t is Greater than or equal to the escalator speed threshold v th , or the model performance index p at time t (t) Greater than or equal to the model performance index threshold p th , then the sampling frequency at time t+1 is β is the adjustment factor, and β>1.

[0167] Optionally, the information entropy of the escalator operation data is X represents the data collected by the sensor, i represents the i-th value of the data, p(x i ) represents x i Probability of occurrence;

[0168] When the sensor transmits data, wavelet transform is used to compress the time series data of escalator operation, and principal component analysis is used to select the most important features for transmission.

[0169] Optionally, the feature extraction and fusion module 202 is specifically used to: By integrating the local features of various escalator operation data, the global health status representation of the escalator is obtained; where v (a) is the global health status representation vector of the escalator, represents the feature vector of the data collected by the i-th sensor, is the attention weight of the i-th sensor, N is the number of sensors, and is the hidden representation of the data collected by the i-th sensor, u (a) is the attention vector, and The default parameters for linear transformation.

[0170] Optionally, the above-mentioned deep learning-based escalator health assessment device may also include a gated weighting module, which is used to adaptively weight the local features of each sensor according to importance through a gating unit, and then the feature extraction and fusion module 202 executes a process of adaptively fusing the local features of various escalator operation data through an attention mechanism.

[0171] The process of adaptively weighting the local features of each sensor according to their importance by the gate control unit includes: Adaptively weight the local features of the data collected by each sensor. is the local feature of the data collected by the i-th sensor after adaptive weighting, is the local feature of the data collected by the ith sensor before adaptive weighting, is the gating signal corresponding to the i-th sensor in the gating unit, ⊙ represents element-by-element multiplication, and is the preset parameter of the gate unit, and σ is the sigmoid activation function.

[0172] Optionally, the above-mentioned escalator health assessment device based on deep learning may also include an information fusion module, which is used to fuse the attribute information of the entity with the low-dimensional structure embedding representation to obtain an enhanced entity. The low-dimensional structure embedding is represented as The enhanced entity for: and They are the transformation matrices for embedding representation of attribute information and low-dimensional structure, respectively.

[0173] Correspondingly, the health assessment module 204 is specifically used to: Reasoning about the abnormal state of the escalator, The reason for the abnormal escalator status is inferred. For the target entity The enhancement indicates that is the parameter matrix of the attention mechanism, | represents vector concatenation, For the target entity Related related entities The information obtained after weighted aggregation is For the target entity Related entities The attention weights between is the parameter matrix of the attention mechanism.

[0174] Optionally, the above-mentioned deep learning-based escalator health assessment device may also include a federated learning module, which is specifically used for: for multiple escalator operating units, each escalator operating unit adopts a federated learning framework to locally train a model using its own escalator operation data; wherein the model is a model composed of the convolutional neural network and the attention mechanism, and a model composed of the graph convolutional network and the escalator domain knowledge graph; after the training is completed, the escalator operating units exchange the trained model parameters through an encrypted communication protocol; wherein, in the process of training the model through the federated learning framework, differential privacy technology is introduced to add random noise to the gradient information shared between the escalator operating units.

[0175] Figure 5 Schematic diagram of a terminal device provided by an embodiment of the present invention. Figure 5 As shown, the terminal device 300 of this embodiment includes a processor 310 and a memory 320, wherein the memory 320 stores a computer program that can be run on the processor 310, such as an escalator health assessment program based on deep learning. When the processor 310 executes the computer program, the steps in the above-mentioned escalator health assessment method embodiment based on deep learning are implemented, such as Figure 1 Alternatively, when the processor 310 executes the computer program, the functions of each module / unit in the above-mentioned device embodiments are realized, for example, Figure 4 The functions of modules 201 to 204 are shown.

[0176] Exemplarily, the computer program may be divided into one or more modules / units, which are stored in the memory 320 and executed by the processor 310 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of completing specific functions, which are used to describe the execution process of the computer program in the terminal device 300. For example, the computer program may be divided into a data acquisition module, a feature extraction and fusion module, a knowledge graph construction module, and a health assessment module.

[0177] The terminal device 300 may be a computing device such as a desktop computer, a notebook, a PDA, or a cloud server. The terminal device may include, but is not limited to, a processor 310 and a memory 320. Those skilled in the art will appreciate that Figure 5 It is only an example of the terminal device 300 and does not constitute a limitation of the terminal device 300. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the terminal device may also include input and output devices, network access devices, buses, etc.

[0178] The processor 320 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0179] The memory 320 may be an internal storage unit of the terminal device 300, such as a hard disk or memory of the terminal device 300. The memory 320 may also be an external storage device of the terminal device 300, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the terminal device 300. Further, the memory 320 may also include both an internal storage unit of the terminal device 300 and an external storage device. The memory 320 is used to store the computer program and other programs and data required by the terminal device. The memory 320 may also be used to temporarily store data that has been output or is to be output.

[0180] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A deep learning-based escalator health assessment method, characterized in that: include: The escalator operation data is collected by a sensor, wherein the escalator operation data includes an escalator vibration signal, an escalator temperature, an escalator operation noise and an escalator operation current; wherein the collection frequency of the escalator operation data collected by the sensor and the transmission period of the escalator operation data are adjusted based on the escalator operation state and the information entropy of the escalator operation data; Extracting local features for representing the health status of the escalator from the escalator operation data by using a convolutional neural network, and adaptively fusing the local features of various escalator operation data by using an attention mechanism to obtain a global health status representation of the escalator; Construct an escalator domain knowledge graph, which contains entities and relationships between entities. Entities include escalator components, fault types, maintenance methods, and operating parameters. When constructing the escalator domain knowledge graph, a low-dimensional structural embedding representation of each entity and the relationship between entities is learned based on a graph convolutional network. When the global health status representation indicates that the escalator state is abnormal, the cause of the abnormal escalator state is inferred based on the relationship between entities in the escalator domain knowledge graph, and the escalator health status is evaluated according to the cause of the abnormal escalator state.

2. The escalator health assessment method based on deep learning according to claim 1 is characterized in that: When the global health status representation indicates that the escalator operation state is normal, the sensor collects escalator operation data at a first sampling frequency and transmits the escalator operation data at a first transmission period; when the global health status representation indicates that the escalator operation state is abnormal, the sensor collects escalator operation data at a second sampling frequency and transmits the escalator operation data at a second transmission period; wherein the second sampling frequency is greater than the first sampling frequency, and the second transmission period is less than the first transmission period; And / or, when the model performance index decreases, the sensor increases the sampling frequency and reduces the transmission period; wherein the model performance index is the performance index of the model composed of the convolutional neural network and the attention mechanism, and the performance index of the model composed of the graph convolutional network and the escalator field knowledge graph; And / or, when the information entropy of the escalator operation data increases, the sensor increases the sampling frequency and reduces the transmission period.

3. The escalator health assessment method based on deep learning according to claim 2 is characterized in that: If the escalator speed at time t is Less than the escalator speed threshold v th , and the model performance index p at time t (t) Less than the model performance index threshold p th , then the sampling frequency at time t+1 is is the sampling frequency at time t, α is the adjustment factor, and 0<α<1; If the escalator speed at time t is Greater than or equal to the escalator speed threshold v th , or the model performance index p at time t (t) Greater than or equal to the model performance index threshold p th , then the sampling frequency at time t+1 is β is the adjustment factor, and β>

1.

4. The escalator health assessment method based on deep learning according to claim 2 is characterized in that: The information entropy of escalator operation data is: X represents the data collected by the sensor, i represents the i-th value of the data, p(x i ) represents x i Probability of occurrence; When the sensor transmits data, wavelet transform is used to compress the time series data of escalator operation, and principal component analysis is used to select the most important features for transmission.

5. The escalator health assessment method based on deep learning according to claim 1 is characterized in that: The local features of various escalator operation data are adaptively integrated through the attention mechanism to obtain the global health status representation of the escalator, including: pass By integrating the local features of various escalator operation data, the global health status representation of the escalator is obtained; where v (a) is the global health status representation vector of the escalator, represents the feature vector of the data collected by the i-th sensor, is the attention weight of the i-th sensor, N is the number of sensors, and is the hidden representation of the data collected by the i-th sensor, u (a) is the attention vector, and The default parameters for linear transformation.

6. The escalator health assessment method based on deep learning according to claim 5 is characterized in that: Before adaptively fusing the local features of various escalator operation data through the attention mechanism, the method further includes: Adaptively weighting the local features of each sensor according to their importance through the gating unit, and then performing the step of adaptively fusing the local features of various escalator operation data through the attention mechanism; The process of adaptively weighting the local features of each sensor according to their importance by the gate control unit includes: Adaptively weight the local features of the data collected by each sensor. is the local feature of the data collected by the i-th sensor after adaptive weighting, is the local feature of the data collected by the ith sensor before adaptive weighting, is the gating signal corresponding to the i-th sensor in the gating unit, ⊙ represents element-by-element multiplication, and is the preset parameter of the gate unit, and σ is the sigmoid activation function.

7. The escalator health assessment method based on deep learning according to claim 1 is characterized in that: The method further includes: fusing the attribute information of the entity with the low-dimensional structure embedding representation to obtain an enhanced entity; wherein, the attribute feature of the i-th entity is The low-dimensional structure embedding is represented as The enhanced entity for: and are the transformation matrices representing the attribute information and low-dimensional structure embedding respectively; The reason for the abnormal escalator state based on the relationship between entities in the escalator domain knowledge graph includes: Reasoning about the abnormal state of the escalator, The reason for the abnormal escalator status is inferred. For the target entity The enhancement indicates that is the parameter matrix of the attention mechanism, | represents vector concatenation, For the target entity Related related entities The information obtained after weighted aggregation is For the target entity Related entities The attention weights between is the parameter matrix of the attention mechanism.

8. The escalator health assessment method based on deep learning according to claim 1, characterized in that: The method further comprises: For multiple escalator operating units, each escalator operating unit adopts a federated learning framework to locally train a model using its own escalator operation data; wherein the model is a model composed of the convolutional neural network and the attention mechanism, and a model composed of the graph convolutional network and the escalator domain knowledge graph; After the training is completed, the escalator operating units exchange the trained model parameters through an encrypted communication protocol. In the process of training the model through the federated learning framework, differential privacy technology is introduced to add random noise to the gradient information shared between the escalator operating units.

9. An escalator health assessment device based on deep learning, characterized in that: include: A data acquisition module, used for collecting escalator operation data through sensors, wherein the escalator operation data includes escalator vibration signal, escalator temperature, escalator operation noise and escalator operation current; wherein the collection frequency of the escalator operation data collected by the sensor and the transmission period of the escalator operation data are adjusted based on the escalator operation state and the information entropy of the escalator operation data; A feature extraction and fusion module is used to extract local features for characterizing the health status of the escalator from the escalator operation data through a convolutional neural network, and adaptively fuse the local features of various escalator operation data through an attention mechanism to obtain a global health status characterization of the escalator; A knowledge graph construction module is used to construct an escalator domain knowledge graph, which contains entities and relationships between entities. Entities include escalator components, fault types, maintenance methods, and operating parameters. When constructing the escalator domain knowledge graph, a low-dimensional structural embedding representation of each entity and the relationship between entities are learned based on a graph convolutional network. The health assessment module is used to infer the cause of the abnormal escalator state based on the relationship between entities in the escalator domain knowledge graph when the global health status representation indicates that the escalator state is abnormal, and evaluate the escalator health status according to the cause of the abnormal escalator state.

10. A terminal device, comprising a memory and a processor, wherein the memory stores a computer program that can be run on the processor, characterized in that: When the processor executes the computer program, the steps of the escalator health assessment method based on deep learning as described in any one of claims 1 to 8 are implemented.

Images