Elevator fault prediction and intelligent maintenance decision method based on edge computing

Abstract

This invention belongs to the field of elevator intelligent operation and maintenance and edge computing technology, specifically disclosing a method for elevator fault prediction and intelligent maintenance decision-making based on edge computing. This method deploys multi-source sensors on key elevator components and constructs a dual-channel fusion analysis model through edge computing nodes: the first channel uses a temporal convolutional network to extract degradation trends from high-frequency data, and the second channel initializes a fuzzy cognitive graph based on expert knowledge; the temporal features are mapped to concept node activation values ​​and used to drive inference, generating health assessment results; the cloud platform uses historical fault data to optimize the fuzzy cognitive graph structure through an evolutionary algorithm and periodically sends back parameters to update the edge model; finally, based on the health trend prediction results, a three-level maintenance decision instruction is automatically generated. This invention achieves high-precision, interpretable early fault identification and closed-loop maintenance scheduling, improving the safety and efficiency of elevator operation and maintenance.

Description

Technical Field

[0001] This invention belongs to the field of elevator intelligent operation and maintenance and edge computing technology, specifically involving an elevator fault prediction and intelligent maintenance decision-making method based on edge computing. Background Technology

[0002] With the rapid development of smart cities and IoT technologies, elevators, as indispensable vertical transportation hubs in modern high-rise buildings, directly impact public safety and property security through their safe and reliable operation. To enhance the intelligence level of elevator operation and maintenance, edge computing technology is widely used in elevator monitoring systems, aiming to achieve real-time processing and status assessment of massive amounts of sensor data through edge computing capabilities.

[0003] Elevator fault prediction and health management technology based on condition monitoring has become a core research direction in this field. This type of technology mainly uses various sensors deployed on key elevator components to acquire real-time state parameters such as current, vibration, temperature, and operating displacement. Specific analytical models are then used to predict the deterioration trend of the elevator system, providing a basis for accurate maintenance decisions.

[0004] Existing elevator fault prediction methods still have problems when dealing with complex operating conditions: traditional models lack sufficient analytical depth for high-frequency sensor data, making it difficult to capture subtle fluctuations and dependencies in equipment operating parameters over long time scales. Mainstream deep learning models often exhibit black-box mechanisms, lacking the ability to reason about the causal logic and uncertainties of elevator faults, and their prediction logic cannot be deeply integrated with the prior knowledge of industry experts. Existing expert knowledge bases are mostly static and difficult to dynamically and adaptively evolve with the nonlinear changes during equipment aging, resulting in poor performance in discovering hidden fault propagation paths and responding to sudden abnormal states. Summary of the Invention

[0005] The purpose of this invention is to provide an elevator fault prediction and intelligent maintenance decision-making method based on edge computing, thereby solving the problems mentioned in the background art.

[0006] To achieve the above objectives, this invention provides an elevator fault prediction and intelligent maintenance decision-making method based on edge computing, comprising the following steps: Step 1: Deploy multi-source sensors on key elevator components to collect data in real time, including vibration signals, temperature data, current waveforms, and displacement information, and transmit the data synchronously to edge computing nodes; Step 2: Construct a dual-channel fusion analysis model on the edge computing node. The first channel uses a temporal convolutional network to extract multi-level features from high-frequency sensor data to identify the degradation trend of equipment status in a long time series. The second channel initializes a fuzzy cognitive graph model based on elevator domain expert knowledge. The fuzzy cognitive graph model contains multiple concept nodes that characterize the health status of the elevator subsystem and their initial causal weights. Step 3: Map the temporal features extracted from the first channel to the activation values ​​of the corresponding concept nodes in the fuzzy cognitive graph, and drive the fuzzy cognitive graph to perform forward reasoning to generate the comprehensive health assessment result of the elevator system at the current moment; Step 4: Upload the health assessment results and actual operating status labels generated at the edge to the cloud platform. Based on a historical fault case library, the cloud platform uses an evolutionary algorithm to iteratively optimize the connection relationships and causal weights of concept nodes in the fuzzy cognitive graph, forming a dynamically evolving cognitive graph structure. Step 5: Periodically send the cloud-optimized fuzzy cognitive map parameters back to the edge computing node to update the local model and continuously enhance the edge-side reasoning capability; Step 6: Based on the health trend prediction results output by the updated dual-channel model, and combined with the preset threshold, determine whether there is a potential risk of elevator failure, and automatically generate graded maintenance decision instructions according to the risk level, and push them to the maintenance dispatch system.

[0007] Preferably, the multi-source sensor includes a triaxial vibration accelerometer, a non-contact infrared temperature sensor, a Hall current sensor, and a high-precision encoder, which are respectively installed on the traction machine, guide rail, control cabinet, and car top. The edge computing node preprocesses the raw signals collected by the sensors, uses an adaptive noise cancellation operator to capture the ambient noise using a reference sensor installed on a non-stressed structural component, and subtracts the ambient noise from the mixed signal of the traction machine sensor to extract the component degradation characteristics. The edge computing node uses sliding window technology to cut the continuous data stream into sample frames of a preset length, performs a fast Fourier transform on each frame of data, and extracts its energy distribution on the frequency feature spectrum. The analog signals output by all sensors are converted into digital signals of a preset depth by an analog-to-digital converter module and aligned according to a preset timestamp to form a multidimensional time-series feature vector stream.

[0008] Preferably, the temporal convolutional network adopts a multi-level causal convolutional structure to ensure that the output at any time depends only on the current and previous input information; The temporal convolutional network includes dilated convolutional layers, whose dilation factor increases layer by layer with the number of convolutional layers according to a power law with base 2, so that the receptive field of the top layer of the network covers a preset number of time steps. Each convolutional operation is followed by a batch normalization layer, which accelerates the convergence of the network by standardizing the feature maps with zero mean and unit variance, and uses a modified linear unit as the activation function to introduce non-linear expressive power. The first channel of the temporal convolutional network also introduces a spatial transformation network layer to resample and geometrically transform the input temporal signal, eliminating the time axis stretching or compression effect caused by motor speed fluctuations. The final output layer of the first channel uses a fully connected structure to set the number of neurons to be consistent with the number of concept nodes in the subsequent fuzzy cognitive map.

[0009] Preferably, the concept node set of the fuzzy cognitive graph model covers the traction system, gantry crane system, safety circuit, brake, and guide rail wear status; The initial causal weights are determined using the Delphi method. A predetermined number of senior engineers conduct a qualitative assessment of the causal influence strength between each node based on the fault tree analysis results, and the results are summarized through multiple iterations to form an initial causal correlation matrix. For older elevators, the fuzzy cognitive graph model is expanded to include mechanical wear concept nodes such as uneven wire rope tension, guide shoe clearance fluctuation, and worm gear transmission clearance. Its initial weights are converted into fuzzy causal strengths by analyzing historical maintenance records using natural language processing to identify the co-occurrence frequency of faults and symptoms.

[0010] Preferably, the reasoning process of the fuzzy cognitive graph adopts the maximum algebraic composition rule and membership function; The feature dimension output by the temporal convolutional network is processed by a non-linear activation function to compress the value range to a closed interval between zero and one, which serves as the initial activation level of the corresponding concept node. In each round of reasoning iteration, the state of each concept node at the current moment is determined by its state at the previous moment, the weighted cumulative value of the related nodes through weights, and the state decay coefficient. The state decay coefficient causes the current state value of the concept node to be multiplied by a preset decay factor in each iteration, so as to ensure the system's fault tolerance to instantaneous noise. The weighted cumulative value is transformed nonlinearly through a membership function until the state values ​​of all nodes tend to stabilize or reach the preset iteration limit. Finally, the output node is selected as the carrier of the health index. The mapping logic also introduces a confidence evaluation based on fuzzy set theory, which adds a confidence score to the state value of each concept node.

[0011] Preferably, the evolutionary algorithm adopts a differential evolution strategy, treating the weight parameters in the fuzzy cognitive map as individual genes; the differential evolution strategy generates a candidate weight set through mutation operation, the mutation vector is generated by scaling the difference between the current best individual and multiple randomly selected individuals, and the candidate genes are combined with existing genes through crossover operation. The fitness evaluation function is constructed based on the prediction accuracy and false alarm rate of the health index sequence within a preset time before the occurrence of historical failures, and the weight of maintenance costs is comprehensively considered. The cloud platform also establishes an elevator fault knowledge graph, uses graph neural networks to mine common fault patterns across devices and regions, and transforms the mining results into logical constraints to feed back into the crossover and mutation operators of the evolutionary algorithm. For newly installed elevators lacking historical data, the cloud platform selects digital twin models with similar environmental parameters from the fault knowledge graph and uses their causal weights as initial weights.

[0012] Preferably, the back-transmission process of the fuzzy cognitive graph parameters adopts an incremental update mechanism, transmitting only the changed weight matrix elements and newly added concept node parameters; The feedback cycle of the fuzzy cognitive map parameters is dynamically adjusted according to the frequency of elevator use. The update cycle is shortened in high-load scenarios where the average daily number of runs exceeds the preset threshold, and extended in low-frequency usage scenarios. After receiving updated parameters, the edge computing node first performs a consistency check in the shadow model. After the check passes, it updates the configuration parameters of the local dual-channel model through atomic replacement. The process of transmitting fuzzy cognitive graph parameters is encapsulated using an asymmetric encryption algorithm. Edge computing nodes decrypt and digest the transmitted packets through a built-in hardware security module. If data tampering or structural damage is detected, the node discards the current update and uses the previous version, while also reporting an abnormal status code to the cloud. The system also supports a model version rollback mechanism.

[0013] Preferably, the hierarchical maintenance decision-making instruction includes a three-level response mechanism; When the health index is higher than the first preset threshold, it is determined to be in a normal state, and the system only stores and archives the original data. When the health index is between the first preset threshold and the second preset threshold, it is determined to be a level 2 risk state, triggering an early warning notification, listing the hidden components and recommended inspection methods in detail and pushing them to the maintenance personnel's terminal. When the health index falls below the second preset threshold, it is determined to be a Level 1 emergency risk, an emergency work order is immediately generated, and maintenance personnel are assigned to handle the situation within the predetermined time limit. The first and second preset thresholds are dynamically adjusted according to the elevator usage scenario. In buildings with high-frequency use, the thresholds are increased to implement preventive maintenance. The generated maintenance decision instructions also include topology location information of the fault location and links to standard operating procedures; The system also introduces a closed-loop feedback mechanism to verify the effectiveness of maintenance. It uses computer vision technology to quantify the wear images of old parts uploaded after maintenance and compares them with the prediction results to trigger the model's self-correction.

[0014] Preferably, the edge computing node has a built-in lightweight model compression module that performs structured channel pruning on the temporal convolutional network. By evaluating the importance of the convolutional channels to the prediction contribution, non-core channels with redundancy exceeding a preset ratio are removed. The model compression module also converts the model parameters from floating-point numbers to fixed-point numbers for quantization to reduce the computational load of a single inference. During edge-side inference, a sparsity computation strategy is introduced for large-scale networks, where connection computation is only performed on nodes whose activation values ​​change beyond a preset minimum value. The edge computing nodes run on a real-time operating system and ensure that the prediction logic is not interfered with by background tasks by scheduling inference tasks with the highest priority. Edge computing nodes use fast wavelet transform technology to decompose the current waveform into multiple scales and analyze energy mutations at specific scales to identify early signs of thermal breakdown in the inverter power transistors.

[0015] Preferably, in a multi-ladder collaborative environment, a horizontal data sharing link is established between multiple edge computing nodes, and the fuzzy cognitive graph dynamically introduces external environment coupling nodes; If multiple elevators exhibit similar vibration characteristics at the same location, the cloud platform will reduce the causal weight of the deterioration of a single elevator component and increase the weight of environmental factors on the health index when optimizing. The edge computing node also has a self-diagnostic function, which monitors the health status of the sensors in real time. When a hardware failure occurs in the perception layer, the dual-channel model switches to the sensor failure compensation mode and uses the fuzzy cognitive graph to perform reverse reasoning through strongly correlated nodes to estimate the activation value of the missing node. The system also integrates a visual analysis module to monitor passenger behavior in the car in real time and introduce behavioral characteristics as dynamic input into a fuzzy cognitive map. When abnormal vibrations caused by abnormal passenger behavior are detected, the system automatically corrects the deviation of the health index to eliminate false alarms.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention constructs a dual-channel analysis architecture that integrates temporal convolutional networks and fuzzy cognitive graphs at the edge. This fully utilizes the strong feature extraction capabilities of deep learning models for high-frequency sensor data while retaining the interpretability of causal reasoning based on expert knowledge, thus overcoming the lack of logical transparency in traditional black-box models.

[0017] 2. Through the cloud-driven dynamic evolution mechanism of fuzzy cognitive graphs, the system can continuously learn the implicit patterns in historical fault data, break through the knowledge boundaries of static expert systems, and improve the early identification capability of complex, rare and compound faults.

[0018] 3. This invention realizes a closed-loop system from state perception and trend prediction to maintenance decision-making, which greatly reduces the need for manual intervention, improves the scheduling efficiency and accuracy of maintenance resources, extends the service life of key components, and reduces the total life cycle maintenance cost while ensuring the safe operation of elevators. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall technical solution architecture according to the present invention; Figure 2 This is a schematic diagram illustrating the core principle framework of the dual-channel fusion analysis model based on temporal convolutional networks and fuzzy cognitive graphs according to the present invention. Figure 3 This is a logical flowchart of the dynamic evolution of cloud-driven fuzzy cognitive graph and edge-side model update according to the present invention. Figure 4 This is a schematic diagram of the data flow in temporal feature mapping and fuzzy cognitive graph reasoning according to the present invention; Figure 5 This is a flowchart illustrating the generation of graded maintenance decision instructions based on health indices according to the present invention. Detailed Implementation

[0020] Example 1: Reference Figures 1 to 5 In the elevator fault prediction and intelligent maintenance decision-making method based on edge computing, the overall system architecture is composed of a bottom perception layer, an intermediate edge computing layer, and a top cloud service layer.

[0021] In step 1, multi-source sensors are deployed on key elevator components to collect vibration signals, temperature data, current waveforms, and displacement information in real time, and the data is synchronously transmitted to edge computing nodes. During the sensor deployment phase, a triaxial vibration accelerometer is installed on the traction machine's base surface to monitor the mechanical vibrations generated during elevator startup, operation, and braking. The sensor can simultaneously acquire vibration data in three dimensions: horizontal, vertical, and lateral. The sampling frequency is set to 1000 Hz, meaning one set of vibration vectors is collected every 1 millisecond.

[0022] The non-contact infrared temperature sensor is fixed above the heat sink of the traction machine and the main board of the control cabinet. It obtains the surface temperature of key heat-generating components in real time by detecting infrared radiation energy, with a measurement accuracy better than 0.1 degrees Celsius.

[0023] The Hall current sensor is attached to the power cable at the output of the frequency converter to capture the phase current waveform of the traction motor under different loads in real time. Its bandwidth response is sufficient to cover high-frequency harmonic components.

[0024] The high-precision encoder is coupled to the speed limiter rope wheel at the end of the traction sheave shaft or the top of the car. It outputs displacement information through pulse counting to achieve accurate tracking of the elevator's running position, speed and acceleration.

[0025] All analog signals output from the sensors are converted into 16-bit digital signals by the analog-to-digital converter built into the edge computing node, and aligned according to a preset timestamp to form a multi-dimensional time-series feature vector stream. Data transmission uses wired shielded twisted-pair cabling or industrial-grade fieldbus protocols.

[0026] In the above method, step 2 involves constructing a dual-channel fusion analysis model at the edge computing node. The first channel uses a temporal convolutional network to extract multi-level features from high-frequency sensor data to identify subtle deterioration trends in equipment status over long time series. The second channel initializes a fuzzy cognitive graph model based on elevator domain expert knowledge. This fuzzy cognitive graph model contains multiple concept nodes representing the health status of the elevator subsystem and their initial causal weights. The first channel of the temporal convolutional network is implemented by constructing a multi-level causal convolutional structure.

[0027] Causal convolution ensures that the output at any given time depends only on the current and previous input information, strictly adhering to the causal logic of temporal evolution. The network depth is set to 8 layers or more, with the kernel size gradually increasing with the number of layers. For example, the first layer uses a kernel of size 3, and subsequent layers introduce dilation factors to achieve exponential expansion of the receptive field. After each convolutional operation, a batch normalization layer is configured to standardize the feature maps to zero mean and unit variance, accelerating the network's convergence and suppressing gradient vanishing. Modified linear units are used as activation functions, introducing non-linear expressive power into the model, enabling the network to capture subtle impact vibrations from sensor signals, such as pitting corrosion in traction machine bearings, or abnormal current fluctuations caused by aging contactors in control cabinets.

[0028] The final output layer of the first channel is a fully connected structure, with the number of neurons precisely set to match the number of concept nodes in the subsequent fuzzy cognitive graph, achieving feature dimension alignment from low-level physical signals to high-level logical concepts.

[0029] The second channel in step 2 introduces a fuzzy cognitive graph model to simulate the reasoning process of human experts regarding the causal chain of elevator malfunctions. The concept node set of the fuzzy cognitive graph covers the core physical units and functional logic of elevator operation, including but not limited to the thermal balance state of the traction machine, the flexibility of the brake opening and closing, the smoothness of the door operator system, the integrity of the safety circuit, the lubrication and wear of the guide rails, and the vibration level of the car.

[0030] The initial causal weights were determined using the Delphi method, involving more than 10 senior engineers with 15 years or more of industry experience. Each expert qualitatively assessed the strength of the causal influence between nodes based on the fault tree analysis results. For example, the influence of "continuous high temperature of the traction machine" on "bearing lubrication failure" was assigned a high initial positive value, while "sufficient guide rail lubrication" was assigned a negative correlation weight to "car horizontal vibration." The expert opinions were iteratively summarized, and the initial causal correlation matrix was finally formed by calculating the average value and dispersion. This initial causal correlation matrix describes the complex interaction logic within the system, providing a knowledge base for subsequent reasoning.

[0031] Step 3 involves mapping the temporal features extracted from the first channel to the activation values ​​of corresponding concept nodes in the fuzzy cognitive graph, and driving the fuzzy cognitive graph to perform forward inference to generate the comprehensive health assessment result of the elevator system at the current moment. During the mapping process, each feature dimension output by the temporal convolutional network is processed by a non-linear activation function, compressing its value range to a closed interval between 0 and 1. This value represents the initial activation level of the corresponding concept node; for example, when the current feature identifies abnormal harmonics, the activation value of the concept node in the door operator system will increase accordingly. During inference, the maximum algebraic synthesis rule is adopted, where the state of each concept node at the current moment is determined by its state at the previous moment and the cumulative value of all associated nodes influenced by their weights.

[0032] In each round of inference iteration, the cumulative influence value of the nodes needs to undergo a nonlinear transformation through an sigmoid membership function. The smoothness of this sigmoid membership function ensures the stability of the state evolution and prevents numerical explosion. The inference process continues until the state values ​​of all nodes tend to stabilize or the preset iteration limit is reached. The output node is selected as the carrier of the health index, which represents the overall operational health of the elevator system. The health index is a continuous variable; a value close to 1 indicates that the system is in an excellent state, while a value close to 0 indicates that there are potential safety hazards in the system.

[0033] In the above method, step 4 involves uploading the health assessment results and actual operating status labels generated at the edge to the cloud platform. Based on a historical fault case library, the cloud platform uses an evolutionary algorithm to iteratively optimize the connection relationships and causal weights of concept nodes in the fuzzy cognitive graph, forming a dynamically evolving cognitive graph structure. The cloud platform constructs a massive elevator fault knowledge graph, extracting early warning features of faults by integrating historical operating data from elevators of different brands and models.

[0034] When data uploaded from the edge device shows a decline in the health index and a real fault repair record subsequently occurs, the cloud-based evolutionary algorithm initiates the optimization process. The evolutionary algorithm employs a differential evolution strategy, treating each weight parameter in the fuzzy cognitive graph as an individual gene. The algorithm generates a candidate weight set through mutation operations; the mutation vector is generated by scaling the difference between the current best individual and two randomly selected individuals. Subsequently, a crossover operation combines the candidate genes with existing genes to generate a new generation of weight configuration schemes. The fitness evaluation function is constructed based on the prediction accuracy of the health index sequence within the 24 hours prior to the historical fault occurrence. Only when the new weight configuration can identify fault precursors earlier and more accurately will the scheme be retained.

[0035] Through thousands of iterations, the fuzzy cognitive graph can automatically identify new fault propagation paths, such as discovering the implicit causal chain where ambient humidity affects the insulation performance of the control cabinet, ultimately leading to a door operator logic failure.

[0036] Step 5 involves periodically transmitting the cloud-optimized fuzzy cognitive graph parameters back to the edge computing nodes to update the local model, thereby continuously enhancing edge-side inference capabilities. The transmission of fuzzy cognitive graph parameters employs an incremental update mechanism, transmitting only changed weight matrix elements and newly added concept node parameters to conserve network bandwidth. The transmission cycle exhibits dynamic adaptive characteristics. In high-load elevator scenarios with over 500 daily operations, the update cycle is compressed to within 72 hours to cope with the rapid degradation of components; while in elevators used less frequently in residential communities, the update cycle can be extended to 168 hours.

[0037] Upon receiving an update command, the edge computing node first performs a consistency check in the shadow model to ensure that the new parameters will not trigger false alarms in the current operating environment. After the check passes, it updates the configuration parameters of the local dual-channel model using atomic replacement to ensure that the inference logic remains synchronized with the latest cloud-based collective intelligence.

[0038] Step 6 involves determining whether the elevator has potential fault risks based on the updated dual-channel model's health trend prediction results and preset thresholds. A graded maintenance decision instruction is automatically generated based on the risk level and pushed to the maintenance dispatch system. The system has two preset thresholds. The first preset threshold is typically set around 0.8. When the health index is higher than the first preset threshold, the elevator is considered healthy, and the system only stores and archives the raw data without triggering external intervention. When the health index drops to between the first and second preset thresholds (typically set to 0.5), it is classified as a level-two risk state. The system automatically triggers an early warning notification, detailing potentially hazardous components (such as excessive brake clearance) and recommended inspection methods. This information is pushed to the maintenance personnel's mobile terminals, requiring them to pay close attention during the next routine maintenance.

[0039] When the health index falls below the second preset threshold, it is judged as a level one emergency risk. The system immediately generates an emergency work order and forcibly assigns the nearest maintenance personnel to the site within 30 minutes through the maintenance dispatch platform. At the same time, a warning broadcast is sent to the elevator management, suggesting that speed limit operation or elevator shutdown for inspection be taken as appropriate.

[0040] To ensure real-time performance in resource-constrained edge computing environments, edge computing nodes incorporate a lightweight model compression module. This module performs structured channel pruning on temporal convolutional networks, eliminating non-core channels with redundancy exceeding 80% by evaluating the importance of each convolutional channel to the final prediction. Simultaneously, model parameters are quantized from 32-bit floating-point numbers to 8-bit fixed-point numbers. These optimizations reduce the computational load of a single inference iteration, ensuring that the entire link from sensor signal input to maintenance decision output is strictly controlled within 200 milliseconds.

[0041] The cloud platform has also established a cross-regional elevator fault knowledge graph, using graph neural networks to perform correlation analysis on the operation records of tens of thousands of elevators across the country. Graph neural networks can uncover common degradation patterns of specific brands in specific climatic environments (such as coastal areas with high salt spray). These findings are transformed into new logical constraints and fed back into the crossover and mutation operators of the evolutionary algorithm, enabling the fuzzy cognitive graph to have stronger transfer learning and generalization prediction capabilities when facing rare faults or newly emerging complex faults.

[0042] Example 2: Based on Example 1, this example further elaborates on the specific hardware implementation and parameter configuration of the elevator fault prediction method based on edge computing under complex working conditions.

[0043] For early identification of traction machine bearing failures, the edge computing nodes have undergone deep customization of the vibration signal preprocessing logic. Before entering the temporal convolutional network, the raw triaxial vibration signal is first bandpass filtered to remove high-frequency electrical noise (typically frequency components above 5000 Hz) unrelated to elevator leveling and switching operations, as well as low-frequency structural vibrations transmitted from the foundation. The edge nodes utilize sliding window technology to segment the continuous data stream into sample frames with a length of 2048 sampling points. Each frame undergoes a Fast Fourier Transform to extract its energy distribution along the frequency characteristic spectrum.

[0044] In the above method, the design of the dilated convolutional layers in the first channel temporal convolutional network of step 2 is particularly crucial. To capture the long-range dependencies of the elevator throughout its entire cycle from the bottom to the top (typically lasting 30 to 60 seconds), the network's dilation factor increases layer by layer in powers of 2: the first layer has a dilation factor of 1, the second layer 2, the third layer 4, and so on, until the eighth layer reaches a dilation factor of 128. This structure allows the receptive field of the top layer to cover thousands of time steps, enabling the identification of subtle car swaying trends caused by guide rail verticality deviations during long-distance elevator operation. The batch normalization operation after each convolution layer uses the cumulative mean and cumulative variance, which are then incorporated into the weights of the convolutional kernel during model deployment to further reduce the number of multiply-accumulate operations at the edges.

[0045] For the second channel in step 2, the initial weight matrix between concept nodes in the fuzzy cognitive graph is stored in the form of an adjacency matrix. In the subgraph of the elevator door operator system, there are five core nodes: "door opening motor current," "door lock closing signal delay," "door operator guide rail frictional resistance," and "door operator controller communication frequency." The initial weights set by experts reflect direct physical connections. For example, an increase in the door operator guide rail frictional resistance directly leads to an increase in the door opening motor current; therefore, the weight between these two nodes is set to a positive correlation value of 0.75. The timely arrival of the door lock closing signal is strongly positively correlated with system health.

[0046] In the inference phase of step 3, a state decay coefficient is introduced to enhance the model's robustness to sudden anomalies. In each iteration of fuzzy inference, the current state value of a concept node is multiplied by a decay factor between 0.9 and 0.95, and then combined with the activation value from the external input and the weighted influence values ​​of other nodes. This mechanism is similar to the potential evolution of biological neurons, ensuring that the system has a certain tolerance to transient noise; only when abnormal features persist will the corresponding health index show a significant decline.

[0047] The evolutionary algorithm in step 4, when running in the cloud, fully considers the weight of maintenance costs in the construction of its fitness function. Specifically, the fitness value depends not only on the predicted lead time of failure but also on the false alarm rate. If a certain weight configuration leads to frequent false alarms, its fitness score will be penalized even if it can cover all real failures. The mutation factor in the differential evolution algorithm is set to 0.5, and the crossover probability is set to 0.9. This parameter combination balances the algorithm's global search capability and local refinement capability.

[0048] During the return transmission of fuzzy cognitive map parameters in step 5, to ensure the security and integrity of data transmission, the returned weight file is encapsulated using an asymmetric encryption algorithm. The edge computing node has a built-in hardware security module responsible for decrypting and digesting the returned packets. If data tampering or packet structure corruption is detected during transmission, the node will immediately discard the update, continue using the parameters from the previous version, and report an error status code to the cloud.

[0049] For the decision generation in step 6, the system introduces a dynamic decision threshold mechanism. For elevators in high-frequency public buildings such as hospitals and train stations, the first preset threshold is raised to 0.85 to implement a more stringent preventative maintenance strategy. For freight elevators or elevators in private villas with low usage frequency, the threshold is appropriately lowered to extend the service life of components and avoid over-maintenance. When the tiered maintenance instruction is sent to the maintenance scheduling system, it also includes topological location information of the fault location. For example, when an abnormal fluctuation in the safety circuit voltage is detected, the instruction will clearly identify the contact number most likely to fail, guiding maintenance personnel to perform targeted repairs.

[0050] Example 3: In this example, the evolution logic of the elevator fault prediction method based on edge computing in a multi-elevator collaborative environment and extreme working conditions is described in detail.

[0051] In the scenario of elevator group control in super high-rise buildings, a horizontal data sharing link is established between multiple edge computing nodes. When an elevator detects abnormal vibration in the middle section of the guide rail, the information is synchronously shared with other elevator edge nodes in the same shaft. The fuzzy cognitive map in step 2 dynamically introduces an "external environment coupling" node. If multiple elevators exhibit similar vibration characteristics at the same location, the cloud platform will automatically reduce the causal weight of individual elevator component deterioration and increase the weight of the impact of shaft structure deformation or drastic changes in ambient temperature on the health index when performing the optimization in step 4.

[0052] In the above method, the sampling logic in step 1 is optimized for extreme operating conditions. When the sensor detects that the elevator has entered an overspeed operation or abnormal oscillation state, the sampling frequency will automatically switch to a high-frequency mode, increasing from 1000 Hz to 5000 Hz, in order to capture a more refined transient failure process. The edge computing node uses a large-capacity non-volatile memory (NVM) as a circular buffer, which can store all high-frequency raw data from the most recent 30 minutes, providing a complete chain of evidence for subsequent in-depth accident backtracking in the cloud.

[0053] The dual-channel model in step 2 employs a multi-task learning architecture on the edge side. In addition to outputting feature vectors to the fuzzy cognitive map, the first channel of the temporal convolutional network simultaneously trains an auxiliary residual regression task to predict the changing trend of sensor signals over the next 3 seconds. By calculating the deviation between the predicted and actual observed values, the system can obtain an additional "novelty score," used to help determine whether a new type of fault not recorded in the historical database has occurred.

[0054] The fuzzy cognitive graph inference process in step 3 introduces a sparsity computation strategy for large-scale networks. In each iteration, only nodes whose activation values ​​change by more than a preset minimum value (e.g., 0.01) are connected, ignoring the propagation of minute fluctuations, thus reducing computational latency while ensuring inference accuracy. The mapping function employs adaptive gain control logic when processing TCN features, dynamically adjusting the mapping sensitivity based on sensor noise levels.

[0055] When step 4 is executed in the cloud, the evolutionary algorithm introduces a knowledge transfer mechanism. When a newly installed elevator lacks historical operating data, the cloud platform selects a "digital twin" model with similar environmental parameters from the fault knowledge graph and uses its mature causal weights as initial weights. In this way, the system can have a strong fault early warning capability on the first day of use.

[0056] In the feedback logic of step 5, a model version rollback mechanism has been added. If the edge side detects an increase in prediction error within 24 hours after the new parameters are deployed, the system will automatically trigger a rollback command to restore the most stable historical best version. At the same time, the edge side will mark the sample set that caused the increased error as "difficult samples" and prioritize uploading them to the cloud for algorithm engineers to perform offline analysis.

[0057] Specifically, regarding the maintenance decision in step 6, the system constructs a maintenance resource scheduling optimization model. When multiple elevators issue warnings simultaneously, the system not only sorts them according to the absolute value of their health index but also considers the skill matching of maintenance personnel, tool availability, and traffic conditions to make a multi-criteria decision. The generated decision instructions include links to standard operating procedures (SOPs) for maintenance operations. After arriving on-site, maintenance personnel can scan the identification code on the elevator controller to view the root cause analysis report of the fault diagnosed by the dual-channel model and the recommended maintenance steps on their mobile devices.

[0058] In practical engineering, to cope with the harsh electromagnetic environment inside the elevator control cabinet, the edge computing nodes employ a fully shielded metal casing and implement 4000-volt surge protection for all input / output interfaces. The power supply system adopts a wide voltage design, enabling stable output during power grid voltage fluctuations caused by elevator operation. On the software side, the edge nodes run on a real-time operating system (RTOS), and by scheduling inference tasks with the highest priority, it ensures that the prediction logic is not interfered with by background tasks such as data transmission and log storage.

[0059] This invention also relates to an interpretability analysis method for fuzzy cognitive graph weights. While generating a health index, the system automatically identifies the core influencing factors causing the index to decline by calculating contribution operators. For example, if the health index drops from 0.9 to 0.6, the system can indicate that 70% of the impact comes from "traction machine current waveform distortion," 20% from "increased ambient temperature," and 10% from "displacement tracking error." This transparent output allows maintenance personnel to understand the logic behind the model, enhancing the trustworthiness of artificial intelligence systems in the field of industrial safety.

[0060] Example 4: This example focuses on describing the implementation details of the elevator fault prediction method based on edge computing in the renovation of old elevators and under non-standard communication protocols.

[0061] For older elevators lacking digital interfaces, the data acquisition in step 1 primarily relies on externally installed sensors. Due to the poor smoothness of operation in older elevators, the triaxial vibration accelerometer is subject to significant background mechanical noise interference during data acquisition. Therefore, an adaptive noise cancellation operator is introduced into the edge computing node. This operator utilizes reference sensors mounted on non-load-bearing structural components to capture ambient noise, and by calculating a correlation function, subtracts the ambient noise from the mixed signal from the traction machine sensors to extract the true characteristics of component degradation.

[0062] In the above method, the first channel temporal convolutional network in step 2 introduces a spatial transformation network layer to address the variable operating speeds of older elevators. This spatial transformation network layer can resample and geometrically transform the input temporal signal, eliminating the time axis stretching or compression effects caused by motor speed fluctuations, thus ensuring that subsequent convolutional feature extraction remains invariant to frequency shifts. Residual connections in the network ensure that even with low signal quality, the original features from shallow layers can still be passed to higher-level decision logic.

[0063] The second channel fuzzy cognitive map in step 2 expands the nodes to address the unique mechanical wear logic of older elevators. New conceptual nodes include "uneven wire rope tension," "guide shoe clearance fluctuation," and "worm gear transmission clearance." Since the physical parameters of these nodes are difficult to measure directly, the initial weights are obtained through natural language processing (NLP) analysis of historical maintenance records. The cloud platform scans past maintenance work orders to identify the co-occurrence frequency of faults and symptoms, converting this into fuzzy causal strength as cold-start parameters for the model.

[0064] The reasoning process in step 3 addresses the nonlinear degradation process of aging elevators by employing a nonlinear membership function family. The steepness of the membership functions is adjusted differently for elevators of varying service life. For elevators that have been in service for over 15 years, the functions are more sensitive to small-scale parameter changes, enabling them to detect impending fatigue failures earlier.

[0065] Step 4, which involves evolution in the cloud, introduces a co-evolution strategy. The elevator group is divided into different subpopulations, each optimized for specific fault types. After a period of independent evolution, each subpopulation shares its optimal causal link through a "gene exchange" operator, ultimately synthesizing a comprehensive cognitive graph model with all-round diagnostic capabilities. This method improves the efficiency of evolutionary algorithms in large-scale search spaces.

[0066] In step 5, the transmission of fuzzy cognitive map parameters utilizes breakpoint resume and low baud rate modulation techniques in the weak signal environment of older buildings. Edge nodes communicate with the cloud via leaky cables or low-power wide-area networks (LPWANs) installed in elevator shafts. In poor signal conditions, the system prioritizes transmitting critical threshold update parameters, while the complete weight matrix update is queued for execution when network conditions improve.

[0067] Regarding the decision-making logic in step 6, this embodiment introduces a closed-loop feedback mechanism for maintenance effectiveness verification. After the maintenance instruction is executed, maintenance personnel need to upload photos or measurement data of the replaced old parts via an app. The cloud platform uses computer vision technology to quantify the wear and tear of the old parts and compares it with the health status predicted by the model. If the prediction result does not match the actual wear and tear, the sample will be assigned a high weight, forcibly triggering the next round of evolutionary algorithm optimization, thus achieving a closed loop of model self-correction.

[0068] When processing current waveforms, edge computing nodes employ Fast Wavelet Transform (DWT) technology to decompose the current signal into detailed components at different scales. By analyzing energy abrupt changes at specific scales, the system can identify early signs of thermal breakdown in the inverter power transistors, which are easily overlooked in traditional time-domain analysis. Simultaneously, the edge nodes also integrate power monitoring functionality, indirectly reflecting the overall operating efficiency of the traction system from an energy efficiency perspective by calculating changes in energy consumption per kilometer of operation.

[0069] Example 5: This example further explores the application and security enhancement of the elevator fault prediction method based on edge computing in the intelligent building integrated management system.

[0070] In highly integrated smart buildings, elevator edge computing nodes interact with building automation systems via the CoAP (Co-Restricted Application Protocol) or MQTT (MQTT) protocol. The displacement information collected in step 1 is not only used for fault prediction but also synchronously output to the building dispatch algorithm to achieve elevator pre-deployment based on passenger flow prediction.

[0071] In the above method, the dual-channel model in step 2 has a self-diagnostic function on the edge side. The system monitors the health status of the sensors in real time, for example, by detecting the DC offset of the Hall current sensor to determine whether temperature drift has occurred. Once a hardware failure occurs in the perception layer, the dual-channel model will automatically switch to the "sensor failure compensation" mode. In the "sensor failure compensation" mode, the activation values ​​of missing physical nodes in the fuzzy cognitive map will be estimated by other strongly correlated nodes through reverse reasoning, ensuring that the prediction logic can still maintain more than 80% reliability even when some perception is missing.

[0072] The mapping logic in step 3 introduces a confidence evaluation based on fuzzy set theory. Each mapped concept node state value is accompanied by a confidence score between 0 and 1. This confidence score is determined by the signal-to-noise ratio of the input features and the model's historical performance. When the confidence score is lower than a preset threshold (e.g., 0.6), the inference engine triggers an "uncertainty handling" mechanism, expressing it as a range value in the final health index output, for example, a health score between 0.55 and 0.65. This representation provides richer risk information for maintenance decisions.

[0073] Step 4, when processing large-scale multidimensional data in the cloud, employs a parallel computing architecture based on containerization technology. The evolutionary computation tasks for thousands of elevators are distributed across a dynamically scalable cloud host cluster. The population size in the evolutionary algorithm is automatically adjusted based on the current computing resource load. To prevent the model evolution from getting trapped in local optima, the cloud platform periodically introduces "explorer" genes with random noise into the population to maintain the diversity of causal weight distribution.

[0074] The data transmission mechanism in step 5 employs dual-mirror backup technology. The storage area of ​​the edge computing node is divided into a working area and a backup area. Newly transmitted parameters are first written to the backup area and verified. Only after a system restart or a formal switchover command is received are they loaded from the backup area into the running memory. This mechanism ensures that even if an unexpected power outage occurs during the data transmission process, the elevator monitoring system can immediately return to its previous stable state.

[0075] The decision instructions generated in step 6 include an analysis of the elevator's life-cycle value (LCC). The decision system not only recommends when maintenance should be performed, but also calculates the cost-benefit ratio between "continued maintenance" and "complete elevator replacement" based on the elevator's age and failure frequency. When the health index frequently triggers a Level 1 warning and maintenance costs exceed a preset ratio, the system automatically generates an assessment report recommending major renovations and pushes it to the property management's decision-making level.

[0076] In terms of security, the system implements strict access control. Two-way certificate authentication is used between edge computing nodes and sensors, and between edge nodes and the cloud. All control commands and sensitive parameters are processed in a hardware-level encryption engine. Simultaneously, the system establishes an anomaly detection mechanism based on behavior auditing. If an edge node generates an abnormally large amount of uplink data within a short period or performs unauthorized parameter modifications, the cloud platform will immediately block the node's communication and issue a security alert to the system administrator to prevent network attacks targeting the elevator network system.

[0077] This invention also supports remote virtual maintenance. After the early warning is generated in step 6, cloud-based technical experts can remotely diagnose the problem by retrieving all high-frequency raw signals from the past hour through edge nodes. Using triaxial vibration signal trajectory reconstruction technology, experts can observe the car's motion trajectory in a virtual environment, more intuitively locating guide rail installation deviations. This cloud-edge collaborative expert system reduces the cost of resolving complex problems.

[0078] Finally, the system also integrates voice interaction and visual analysis modules. Inside the elevator car, the visual analysis module monitors passenger behavior in real time, such as jumping or blocking the elevator door for an extended period. These behavioral characteristics are dynamically input into the fuzzy cognitive map. When abnormal vibrations caused by passenger misoperation are detected, the system can automatically provide a voice alert and automatically correct any abnormal deviations in the health index calculation, avoiding false alarms caused by passenger behavior.

[0079] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

[0080] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0081] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An elevator fault prediction and intelligent maintenance decision-making method based on edge computing, characterized in that, Includes the following steps: Step 1: Deploy multi-source sensors on key elevator components to collect data in real time, including vibration signals, temperature data, current waveforms, and displacement information, and transmit the data synchronously to edge computing nodes; Step 2: Construct a dual-channel fusion analysis model on the edge computing node. The first channel uses a temporal convolutional network to extract multi-level features from high-frequency sensor data to identify the degradation trend of equipment status in a long time series. The second channel initializes a fuzzy cognitive graph model based on elevator domain expert knowledge. The fuzzy cognitive graph model contains multiple concept nodes that characterize the health status of the elevator subsystem and their initial causal weights. Step 3: Map the temporal features extracted from the first channel to the activation values ​​of the corresponding concept nodes in the fuzzy cognitive graph, and drive the fuzzy cognitive graph to perform forward reasoning to generate the comprehensive health assessment result of the elevator system at the current moment; Step 4: Upload the health assessment results and actual operating status labels generated at the edge to the cloud platform. Based on a historical fault case library, the cloud platform uses an evolutionary algorithm to iteratively optimize the connection relationships and causal weights of concept nodes in the fuzzy cognitive graph, forming a dynamically evolving cognitive graph structure. Step 5: Periodically send the cloud-optimized fuzzy cognitive map parameters back to the edge computing node to update the local model and continuously enhance the edge-side reasoning capability; Step 6: Based on the health trend prediction results output by the updated dual-channel model, and combined with the preset threshold, determine whether there is a potential risk of elevator failure, and automatically generate graded maintenance decision instructions according to the risk level, and push them to the maintenance dispatch system.

2. The elevator fault prediction and intelligent maintenance decision-making method based on edge computing according to claim 1, characterized in that, The multi-source sensors include a triaxial vibration accelerometer, a non-contact infrared temperature sensor, a Hall current sensor, and a high-precision encoder, which are respectively installed on the traction machine, guide rail, control cabinet, and top of the car. The edge computing node preprocesses the raw signals collected by the sensors, uses an adaptive noise cancellation operator to capture the ambient noise using a reference sensor installed on a non-stressed structural component, and subtracts the ambient noise from the mixed signal of the traction machine sensor to extract the component degradation characteristics. The edge computing node uses sliding window technology to cut the continuous data stream into sample frames of a preset length, performs a fast Fourier transform on each frame of data, and extracts its energy distribution on the frequency feature spectrum. The analog signals output by all sensors are converted into digital signals of a preset depth by an analog-to-digital converter module and aligned according to a preset timestamp to form a multidimensional time-series feature vector stream.

3. The elevator fault prediction and intelligent maintenance decision-making method based on edge computing according to claim 2, characterized in that, The temporal convolutional network adopts a multi-level causal convolutional structure to ensure that the output at any given time depends only on the current and previous input information. The temporal convolutional network includes dilated convolutional layers, whose dilation factor increases layer by layer with the number of convolutional layers according to a power law with base 2, so that the receptive field of the top layer of the network covers a preset number of time steps. Each convolutional operation is followed by a batch normalization layer, which accelerates the convergence of the network by standardizing the feature maps with zero mean and unit variance, and uses a modified linear unit as the activation function to introduce non-linear expressive power. The first channel of the temporal convolutional network also introduces a spatial transformation network layer to resample and geometrically transform the input temporal signal, eliminating the time axis stretching or compression effect caused by motor speed fluctuations. The final output layer of the first channel uses a fully connected structure to set the number of neurons to be consistent with the number of concept nodes in the subsequent fuzzy cognitive map.

4. The elevator fault prediction and intelligent maintenance decision-making method based on edge computing according to claim 3, characterized in that, The concept node set of the fuzzy cognitive graph model covers the traction system, gantry crane system, safety circuit, brake and guide rail wear status; The initial causal weights are determined using the Delphi method. A predetermined number of senior engineers conduct a qualitative assessment of the causal influence strength between each node based on the fault tree analysis results, and the results are summarized through multiple iterations to form an initial causal correlation matrix. For older elevators, the fuzzy cognitive graph model is expanded to include mechanical wear concept nodes such as uneven wire rope tension, guide shoe clearance fluctuation, and worm gear transmission clearance. Its initial weights are converted into fuzzy causal strengths by analyzing historical maintenance records using natural language processing to identify the co-occurrence frequency of faults and symptoms.

5. The elevator fault prediction and intelligent maintenance decision-making method based on edge computing according to claim 4, characterized in that, The reasoning process of the fuzzy cognitive graph adopts the maximum algebra composition rule and membership function; The feature dimension output by the temporal convolutional network is processed by a non-linear activation function to compress the value range to a closed interval between zero and one, which serves as the initial activation level of the corresponding concept node. In each round of reasoning iteration, the state of each concept node at the current moment is determined by its state at the previous moment, the weighted cumulative value of the related nodes through weights, and the state decay coefficient. The state decay coefficient causes the current state value of the concept node to be multiplied by a preset decay factor in each iteration, so as to ensure the system's fault tolerance to instantaneous noise. The weighted cumulative value is transformed nonlinearly through a membership function until the state values ​​of all nodes tend to stabilize or reach the preset iteration limit. Finally, the output node is selected as the carrier of the health index. The mapping logic also introduces a confidence evaluation based on fuzzy set theory, which adds a confidence score to the state value of each concept node.

6. The elevator fault prediction and intelligent maintenance decision-making method based on edge computing according to claim 5, characterized in that, The evolutionary algorithm employs a differential evolution strategy, treating the weight parameters in the fuzzy cognitive graph as individual genes. The differential evolution strategy generates a candidate weight set through mutation operations. The mutation vector is generated by scaling the difference between the current best individual and multiple randomly selected individuals, and the candidate genes are combined with existing genes through crossover operations. The fitness evaluation function is constructed based on the prediction accuracy and false alarm rate of the health index sequence within a preset time before the occurrence of historical failures, and the weight of maintenance costs is comprehensively considered. The cloud platform also establishes an elevator fault knowledge graph, uses graph neural networks to mine common fault patterns across devices and regions, and transforms the mining results into logical constraints to feed back into the crossover and mutation operators of the evolutionary algorithm. For newly installed elevators lacking historical data, the cloud platform selects digital twin models with similar environmental parameters from the fault knowledge graph and uses their causal weights as initial weights.

7. The elevator fault prediction and intelligent maintenance decision-making method based on edge computing according to claim 6, characterized in that, The back-transmission process of the fuzzy cognitive graph parameters adopts an incremental update mechanism, transmitting only the changed weight matrix elements and newly added concept node parameters. The feedback cycle of the fuzzy cognitive map parameters is dynamically adjusted according to the frequency of elevator use. The update cycle is shortened in high-load scenarios where the average daily number of runs exceeds the preset threshold, and extended in low-frequency usage scenarios. After receiving updated parameters, the edge computing node first performs a consistency check in the shadow model. After the check passes, it updates the configuration parameters of the local dual-channel model through atomic replacement. The process of transmitting fuzzy cognitive graph parameters is encapsulated using an asymmetric encryption algorithm. Edge computing nodes decrypt and digest the transmitted packets through a built-in hardware security module. If data tampering or structural damage is detected, the node discards the current update and uses the previous version, while also reporting an abnormal status code to the cloud. The system also supports a model version rollback mechanism.

8. The elevator fault prediction and intelligent maintenance decision-making method based on edge computing according to claim 7, characterized in that, The hierarchical maintenance decision-making instructions include a three-level response mechanism; When the health index is higher than the first preset threshold, it is determined to be in a normal state, and the system only stores and archives the original data. When the health index is between the first preset threshold and the second preset threshold, it is determined to be a level 2 risk state, triggering an early warning notification, listing the hidden components and recommended inspection methods in detail and pushing them to the maintenance personnel's terminal. When the health index falls below the second preset threshold, it is determined to be a Level 1 emergency risk, an emergency work order is immediately generated, and maintenance personnel are assigned to handle the situation within the predetermined time limit. The first and second preset thresholds are dynamically adjusted according to the elevator usage scenario. In buildings with high-frequency use, the thresholds are increased to implement preventive maintenance. The generated maintenance decision instructions also include topology location information of the fault location and links to standard operating procedures; The system also introduces a closed-loop feedback mechanism to verify the effectiveness of maintenance. It uses computer vision technology to quantify the wear images of old parts uploaded after maintenance and compares them with the prediction results to trigger the model's self-correction.

9. The elevator fault prediction and intelligent maintenance decision-making method based on edge computing according to claim 8, characterized in that, The edge computing node has a built-in lightweight model compression module that performs structured channel pruning on the temporal convolutional network. By evaluating the importance of the convolutional channel to the prediction contribution, non-core channels with redundancy exceeding a preset ratio are removed. The model compression module also converts the model parameters from floating-point numbers to fixed-point numbers for quantization to reduce the computational load of a single inference. During edge-side inference, a sparsity computation strategy is introduced for large-scale networks, where connection computation is only performed on nodes whose activation values ​​change beyond a preset minimum value. The edge computing nodes run on a real-time operating system and ensure that the prediction logic is not interfered with by background tasks by scheduling inference tasks with the highest priority. Edge computing nodes use fast wavelet transform technology to decompose the current waveform into multiple scales and analyze energy mutations at specific scales to identify early signs of thermal breakdown in the inverter power transistors.

10. The elevator fault prediction and intelligent maintenance decision-making method based on edge computing according to claim 9, characterized in that, In a multi-tier collaborative environment, a horizontal data sharing link is established between multiple edge computing nodes, and the fuzzy cognitive graph dynamically introduces external environment coupling nodes. If multiple elevators exhibit similar vibration characteristics at the same location, the cloud platform will reduce the causal weight of the deterioration of a single elevator component and increase the weight of environmental factors on the health index when optimizing. The edge computing node also has a self-diagnostic function, which monitors the health status of the sensors in real time. When a hardware failure occurs in the perception layer, the dual-channel model switches to the sensor failure compensation mode and uses the fuzzy cognitive graph to perform reverse reasoning through strongly correlated nodes to estimate the activation value of the missing node. The system also integrates a visual analysis module to monitor passenger behavior in the car in real time and introduce behavioral characteristics as dynamic input into a fuzzy cognitive map. When abnormal vibrations caused by abnormal passenger behavior are detected, the system automatically corrects the deviation of the health index to eliminate false alarms.

Images