An elevator traction system early fault diagnosis method and system based on double-channel fusion "end-edge-cloud" collaboration

Abstract

The application discloses a kind of early fault diagnosis method and system of elevator traction system based on double-path fusion "end-edge-cloud" cooperation, belong to elevator intelligent operation and maintenance technical field.End side is through MEMS acceleration sensor, current sensor, temperature sensor and so on multi-source sensor synchronous acquisition operation signal of traction system;Edge computing node deploys lightweight time series Transform model, realizes real-time fault inference using linear attention mechanism and INT8 quantization technology;Cloud uses search enhancement generation technology to fuse edge side features and fault knowledge base, generates interpretable diagnostic report;Cloud is based on full data training optimization model and is updated to edge end and issues parameter.This application realizes the precise, real-time, interpretable diagnosis of early fault of elevator traction system, and the overall diagnostic accuracy reaches 92.0%, and is suitable for elevator predictive maintenance and safety monitoring field.

Description

Technical Field

[0001] This invention relates to the field of elevator equipment condition monitoring and intelligent fault diagnosis technology, and in particular to an early fault diagnosis method and system for elevator traction systems based on a dual-path fusion lightweight temporal Transformer model and adopting an "end-edge-cloud" collaborative architecture. Background Technology

[0002] The traction drive system is a core component of an elevator, mainly consisting of a traction machine, traction ropes, guide sheaves, and deflector sheaves. During long-term operation, early faults such as traction machine bearing wear, traction rope fatigue, and slippage between the traction sheave and the wire rope gradually accumulate. If these are not detected in time, they can easily develop into serious malfunctions or even safety accidents.

[0003] Currently, fault diagnosis of elevator traction systems mainly relies on three technical approaches: First, the traditional manual inspection and periodic maintenance model. Maintenance personnel are responsible for the periodic inspection of a large number of elevators, making it difficult to conduct a comprehensive investigation within a limited time. Industry data shows that about one-third of elevator faults cannot be detected in time during manual inspections, and potential safety hazards are difficult to detect in a timely manner. Second, diagnostic methods based on traditional signal processing, which extract features by performing spectral analysis on signals such as vibration acceleration, current, and temperature, but it is difficult to effectively capture non-stationary and nonlinear features as well as complex correlations between multiple source signals. Third, intelligent diagnostic methods based on traditional deep learning, such as convolutional neural networks and long short-term memory networks, although they have improved fault identification capabilities to some extent, still have the following shortcomings: (1) Early fault features are weak, and traditional deep learning models are not sensitive enough to weak signals, making it easy to miss early anomalies in the context of strong noise. (2) Elevator operation data has the characteristics of strong real-time, large total amount and high concurrency. A single elevator generates more than 100,000 data points such as vibration and temperature per day. Uploading all the data to the cloud for processing will face bandwidth bottleneck and communication delay, making it difficult to meet the real-time early warning requirements. (3) Most existing methods lack interpretability and only output fault classification results, making it difficult to provide specific diagnostic basis and maintenance suggestions to operation and maintenance personnel.

[0004] Therefore, how to accurately identify early fault characteristics of traction systems from massive, high-dimensional, and noisy operational data while ensuring real-time performance, and provide interpretable diagnostic results, is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an early fault diagnosis method and system for elevator traction systems based on dual-path fusion "end-edge-cloud" collaboration. It achieves millisecond-level real-time fault reasoning through a lightweight dual-path fusion model on the edge side, and outputs interpretable diagnostic results by combining cloud-based Retrieval-Augmented Generation (RAG) technology. At the same time, it establishes a cloud-edge collaborative model iteration mechanism to improve the accuracy of early weak fault identification and the practicality of diagnostic results for elevator traction systems while ensuring real-time performance.

[0006] To achieve the above-mentioned technical features, the objective of this invention is as follows: A first aspect provides an early fault diagnosis method for elevator traction systems based on dual-path fusion "end-edge-cloud" collaboration, comprising the following steps: S1. End-side multi-source signal acquisition: Deploy multiple types of sensors at key parts of the elevator traction system to synchronously acquire multi-source raw signals under the operating status of the traction system; S2. Edge-side data preprocessing and uploading: The collected raw signals are cleaned and standardized, and the preprocessed multi-dimensional time series data is transmitted to the edge computing node. S3. Edge-side dual-path fusion feature extraction and fault reasoning: A lightweight temporal Transformer diagnostic model is deployed on the edge computing node. The model uses a dual-path parallel structure to extract local fault features and global temporal degradation features respectively. After cross-path feature fusion, early fault reasoning is performed, and the fault type and corresponding confidence level are output. S4. Cloud-based Explainable Enhanced Diagnosis: For difficult fault samples with edge-side confidence levels below a preset threshold, their time-series data and intermediate features are uploaded to the cloud server. The cloud uses retrieval enhancement generation technology to integrate edge-side features with the traction system fault knowledge base to generate an explainable diagnostic report containing fault location, cause analysis, and maintenance suggestions. S5. Collaborative Iterative Model Update: The cloud-based diagnostic model is optimized based on full-scale running data and labeled fault samples. The updated lightweight model parameters are then distributed to edge computing nodes to achieve online upgrades of the edge diagnostic model.

[0007] Preferably, in step S1, the multiple types of sensors include: a triaxial MEMS accelerometer installed on the bearing seats of the drive end and non-drive end of the traction machine; a Hall current sensor installed on the brake circuit and motor drive circuit of the traction machine; a patch-type temperature sensor installed on the surface of the traction machine housing; a speed sensor installed beside the traction sheave; and a non-contact wire rope vibration displacement sensor; wherein the sampling frequency of the vibration sensors is not less than 2kHz, covering all working conditions such as elevator no-load, rated load, start-stop speed change, etc.

[0008] Preferably, in step S3, the dual-path parallel structure specifically comprises: 1) Multi-receptive field convolutional feature pathway: Multi-scale parallel convolution kernels are used in combination with channel attention mechanism to capture instantaneous impact and minute local fault features such as pitting of traction machine bearings and local damage to wire ropes; 2) Lightweight temporal attention path: Based on the lightweight Transformer structure, model the long-cycle wear evolution law of the traction system and capture the global temporal degradation trend across start-stop and load.

[0009] Preferably, the dual-path feature fusion adopts a cross-branch multi-head attention mechanism to achieve bidirectional interaction between convolutional local detail features and Transformer global degradation features, thus solving the feature fragmentation problem caused by simple concatenation.

[0010] Preferably, the lightweight temporal Transformer model uses a linear attention mechanism to reduce the computational complexity from O(n²) to O(n), and combines INT8 quantization technology to compress the model volume, enabling millisecond-level inference on edge platforms.

[0011] Preferably, the lightweight temporal Transformer model is equipped with an industrial position encoding module, which uses a combination of trigonometric function position encoding and learnable operating condition encoding to enhance the model's sensitivity to elevator operating condition changes and event-driven signals.

[0012] Preferably, cross-layer contrast constraints are introduced during the model training phase to align the feature distribution of dual-path outputs, reduce the intra-class distance of similar fault features, expand the inter-class interval of different fault categories, and improve the robustness of early weak fault identification.

[0013] Preferably, in step S3, a graded early warning is performed based on the confidence level of the fault reasoning and the fault type: continuous real-time monitoring under normal conditions; when an early minor fault is detected, an early warning message is pushed and a planned maintenance reminder is triggered; when a severe fault is detected, an elevator safety shutdown command is triggered and fault information is reported simultaneously.

[0014] Preferably, the execution flow of the retrieval enhancement generation technology in step S4 is as follows: the physical features extracted from the edge side are mapped to natural language semantic descriptions, the semantic descriptions are used as retrieval statements to perform vector similarity retrieval in the fault knowledge base, the retrieved domain knowledge, edge reasoning results and semantic descriptions are used to jointly construct prompt words, and the prompt words are input into a large language model to generate a structured diagnostic report.

[0015] On the other hand, the present invention provides an early fault diagnosis system for elevator traction systems based on dual-path fusion "end-edge-cloud" collaboration, characterized in that it includes an end-side acquisition unit, an edge computing unit, a cloud-based diagnostic unit, and a model collaborative update unit: The end-side acquisition unit is deployed in key parts of the elevator traction system and is equipped with multiple types of sensors and acquisition controllers to synchronously acquire multi-source operating signals of the traction system and upload them after preprocessing the signals. The edge computing unit is communicatively connected to the end-side acquisition unit and is equipped with a lightweight time-series Transformer diagnostic model with dual-path fusion. It is used to receive time-series data uploaded from the end side and perform real-time fault reasoning, outputting fault type and confidence level. The cloud-based diagnostic unit is communicatively connected to the edge computing unit and is equipped with a traction system fault knowledge base and a retrieval enhancement generation engine. It is used to receive difficult fault samples uploaded from the edge side and generate interpretable diagnostic reports. The model collaborative update unit is deployed on the cloud and the edge respectively, and is used to realize the encrypted distribution and online update of the lightweight model parameters optimized in the cloud to the edge computing nodes.

[0016] Preferably, the end-side acquisition unit is further configured with a local data cache module and a wireless communication module. The local data cache module is used to temporarily store the acquired data when the network is interrupted, and to resume the transmission after the network is restored.

[0017] Preferably, the edge computing unit adopts an edge AI computing platform, which is deployed in the elevator machine room or control cabinet, supports INT8 quantization model inference, and the average latency of single-sample inference is no higher than 200ms.

[0018] Preferably, the cloud-based diagnostic unit is also equipped with a digital twin monitoring panel for visually displaying the operating status, health level, early warning records, and historical diagnostic reports of the connected elevator; the model collaborative update unit supports hot loading and updating of the model and version rollback.

[0019] The present invention has the following beneficial effects: 1. A three-level collaborative architecture for early fault diagnosis of elevator traction systems, namely "end-edge-cloud", was constructed. The end side completes multi-source signal acquisition and preliminary preprocessing, the edge side realizes real-time fault reasoning based on a dual-path fusion lightweight temporal Transformer model, and the cloud uses RAG technology to provide interpretable in-depth diagnostic analysis, which not only ensures the timeliness requirement of real-time early warning, but also realizes the refined diagnosis of complex faults.

[0020] 2. A lightweight temporal Transformer model with dual-path fusion is proposed. The convolutional path captures local weak fault impacts, while the Transformer path models the global degradation trend. Deep feature fusion is achieved through cross-branch attention. At the same time, a linear attention mechanism is adopted to reduce computational complexity, and INT8 quantization technology is combined to achieve model lightweighting. Millisecond-level inference is achieved on edge computing platforms, which is suitable for the low computing power and high real-time deployment requirements of elevator sites.

[0021] 3. The cloud-based system employs retrieval-enhanced generation technology to deeply integrate physical features extracted from the edge with the knowledge base of the elevator traction system domain, generating structured and traceable diagnostic reports. This significantly reduces the illusion rate of content generated from large models, providing maintenance personnel with clear causes of failures and maintenance suggestions, and improving the engineering practicality of diagnostic results.

[0022] 4. Establish a collaborative iterative mechanism for cloud-based training optimization and edge incremental updates. The cloud continuously optimizes model performance based on full data and periodically sends the lightweight model parameters to the edge to achieve continuous evolution of diagnostic model performance and adapt to different elevator models and operating conditions.

[0023] 5. It can accurately identify various early faults such as traction machine bearing wear, traction rope fatigue, traction sheave slippage, and brake abnormalities, promoting the upgrade of elevator operation and maintenance from passive response maintenance to early warning predictive maintenance, effectively reducing the risk of elevator safety accidents and the maintenance cost throughout the entire life cycle. Attached Figure Description

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] Figure 1 This is a schematic diagram of the overall architecture of the "end-edge-cloud" collaborative early fault diagnosis system for elevator traction systems of the present invention; Figure 2 This is a schematic diagram of the overall process of the early fault diagnosis method for elevator traction system of the present invention; Figure 3 This is a schematic diagram of the network structure of the dual-path fusion lightweight temporal Transformer model of the present invention; Figure 4 This is a schematic diagram showing the deployment location of the end-side multi-source sensor of the present invention on the elevator traction system; Figure 5 This is a schematic diagram of the edge-side real-time fault reasoning and hierarchical early warning process of the present invention; Figure 6 This is a schematic diagram of the process for generating interpretable diagnostic reports through cloud-based retrieval enhancement according to the present invention; Figure 7 This is a schematic diagram of the collaborative update process of the "end-edge-cloud" model of the present invention; Figure 8 This is a schematic diagram comparing the diagnostic effects of early-stage bearing failures in traction machines according to an embodiment of the present invention. Detailed Implementation

[0026] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0027] Example 1: Complete Implementation Process of Early Fault Diagnosis Method for Elevator Traction System This embodiment describes in detail the complete execution process of the diagnostic method of the present invention. The overall process is as follows: Figure 2 As shown, the specific steps are as follows: Step S1: End-side multi-source signal acquisition Multiple types of sensors are deployed at key monitoring points in the elevator traction system, with specific locations as follows: Figure 4 As shown: (1) Install a triaxial MEMS accelerometer on the drive end bearing housing and the non-drive end bearing housing of the traction machine respectively. The measurement range is ±8g, the noise density is 25μg / √Hz, the sampling frequency is set to 2kHz, and the vibration acceleration signals in three directions are collected. (2) A Hall current sensor is connected in series between the traction machine brake circuit and the motor drive circuit, with a measurement range of ±50A and a sampling frequency of 1kHz, to collect the brake engagement / release current and the three-phase current of the motor. (3) Attach a PT100 thermocouple temperature sensor to the surface of the traction machine housing, with a sampling frequency of 1Hz, and collect temperature signals of key parts of the traction machine; (4) Install a photoelectric encoder on the side of the traction sheave, with a sampling frequency of 100Hz, to collect the traction sheave speed signal; (5) Install a non-contact laser displacement sensor next to the traction steel wire rope, with a sampling frequency of 2kHz, to collect the vibration displacement signal of the steel wire rope.

[0028] All sensors achieve synchronous acquisition of multi-source signals through a synchronous triggering mechanism, covering all operating conditions of the elevator, including no-load, half-load, full-load, start-stop and speed change, and capturing weak fault characteristics such as early wear, pitting, and foreign object intrusion.

[0029] Step S2: Edge data preprocessing and edge upload The end-side acquisition controller performs the following preprocessing operations on the raw signal in sequence: 1) Data cleaning: Remove outliers that exceed the sensor's range, use linear interpolation to fill in short-term data gaps, and use median filtering to eliminate impulse noise; 2) Timing slicing: The multi-channel signal is sliced ​​according to a window length of 1024 points and a non-overlapping step size to generate a single frame timing sample; 3) Normalization and filtering: Z-score normalization is used for each frame sample to eliminate load amplitude differences, and 50Hz power frequency environmental noise is removed by band-stop filtering to obtain normalized multidimensional time series samples; The preprocessed time series samples are uploaded to the edge computing node in real time via a 4G / 5G wireless communication module or an industrial Ethernet.

[0030] Step S3: Edge-side dual-path fusion feature extraction and fault reasoning After receiving multi-channel time-series data, the edge computing node constructs a multi-dimensional time-series matrix with dimensions [L,C], where L is the time-series length (value 1024) and C is the total number of sensor channels. The time-series matrix is ​​then input into a pre-trained and quantized dual-path fusion lightweight time-series Transformer model to perform forward inference.

[0031] The model first extracts local impact features through a multi-receptive field convolutional path and extracts global degradation features through a lightweight temporal attention path. After the two features are fused through cross-branch multi-head attention interaction, the output layer outputs the probability distribution of K types of faults, where K is the total number of fault types, including traction machine bearing wear, traction rope fatigue, slippage between traction sheave and wire rope, brake malfunction, foreign object intrusion, etc.

[0032] When the output probability of a certain type of fault exceeds the warning threshold of 0.75, it is determined to be an early-stage fault, and an edge-level warning signal is immediately generated. Tiered handling is then implemented according to the severity of the fault, as follows: Figure 5 As shown.

[0033] Step S4: Cloud-based interpretable enhanced diagnostics For difficult fault samples with an inference confidence level below 0.75 on the edge side, or abnormal samples that do not match known fault types, the edge nodes will upload the corresponding multi-source time series data and the model intermediate layer feature vectors to the cloud server.

[0034] The cloud server employs retrieval-enhanced generation technology, integrating edge-side physical features with a pre-built traction system fault knowledge base to generate a structured diagnostic report containing fault type, fault location, severity classification, fault cause analysis, and recommended maintenance measures. The specific process is as follows: Figure 6 As shown, the diagnostic report is simultaneously pushed to the maintenance personnel's mobile device and the cloud platform management backend.

[0035] Step S5: Collaborative Iterative Update of the Model The cloud server aggregates all operational data uploaded by edge nodes and manually labeled fault samples weekly, performing incremental training or full retraining on the basic diagnostic model. After training, the model undergoes INT8 quantization compression to generate a lightweight model file adapted to the edge platform. The new model parameters are then distributed to each edge computing node via an encrypted channel. Once the edge nodes verify the data, a hot-load update is performed, completing the model upgrade without interrupting the monitoring service. The model version number is recorded, supporting historical version rollback. The specific process is as follows: Figure 7 As shown.

[0036] Example 2: Specific Implementation of a Dual-Path Fusion Lightweight Temporal Transformer Model This embodiment details the network structure and core algorithm of a dual-path fusion lightweight temporal Transformer model. All complete calculation formulas are explained in this embodiment, and the overall model structure is as follows: Figure 3 As shown.

[0037] 1. Input Embedding Layer The multidimensional time-series input matrix of dimension [L,C] Through learnable linear transformation matrices Mapping to the embedding dimension D (D=128 in this embodiment), we obtain a vector embedding sequence of dimension [L,D]. The calculation method is as follows: ; In the formula: The input matrix is ​​a multidimensional time series matrix with dimensions [L,C]. This is the length of a single frame of temporal samples; in this embodiment, the value is 1024. The total number of input sensor channels; Let be a learnable linear transformation weight matrix with dimensions [C,D]. For the feature embedding dimension, this embodiment takes a value of 128; It is the mapped vector embedding sequence with dimensions [L,D].

[0038] 2. Industrial Location Coding Layer This invention designs a position encoding system adapted for industrial scenarios, which consists of two parts: a trigonometric function position encoding and a learnable working condition encoding. (1) Trigonometric function position encoding: used to encode temporal position information, for temporal position index pos Hidden Dimension Index The encoding formula is: Even-numbered dimensions: ; Odd-numbered dimensions: ; In the formula: PE This is a trigonometric function position encoding vector with dimensions [L,D]. pos This is a time-series location index, with values ​​ranging from 0 to L-1; i This is the index variable for the hidden dimension, with a value range from 0 to D / 2-1; D represents the total number of feature embedding dimensions.

[0039] (2) Learnable operating condition coding: For three typical load conditions of elevators—no load, half load, and full load—three learnable embedding vectors are set up, and the corresponding coding vector is matched according to the current elevator operating condition. .

[0040] The final position-enhanced input sequence is: ; In the formula: The final input sequence after location augmentation has dimensions [L,D]. The learnable embedding vector corresponding to the load condition has dimensions [1,D].

[0041] 3. Dual-path feature extraction module (1) Multi-receptive field convolutional feature pathways Four sets of one-dimensional parallel convolutional kernels of different sizes (3, 7, 15, and 31 ohms) are used to perform multi-scale feature extraction on the input sequence, capturing local impact features in different frequency bands. The convolutional output is then fed into a channel attention module to dynamically weight the features of each channel, enhancing the response of fault-related channels. This pathway outputs a local feature matrix. The dimensions are [L,D].

[0042] (2) Lightweight temporal attention pathway It adopts a lightweight Transformer encoder structure with a core linear attention layer to replace the quadratic complexity of standard self-attention computation.

[0043] For the input sequence The query matrix is ​​obtained through three linear transformation matrices respectively. Key matrix Value matrix : ; In the formula: The input feature sequence for the attention layer has dimensions [L,D]. There are three sets of learnable linear transformation weight matrices, each with dimensions [D, d_k], where The feature dimension of a single attention head; The query matrix has dimensions [L, d_k]. The key matrix has dimensions [L, d_k]. It is a value matrix with dimensions [L, d_k].

[0044] The standard formula for calculating self-attention is: ; In the formula: This is the standard self-attention output. This is a normalized activation function used to convert attention weights into a probability distribution; is the feature dimension of the key matrix K, used to scale the dot product result to avoid excessively large values; The time complexity of this calculation is... ,in The length of the input sequence.

[0045] This invention employs a linear attention mechanism and introduces a kernel feature mapping function. The calculation order is adjusted as follows: ; The kernel feature mapping function is in the form of the ELU activation function plus 1, that is... In the formula: This is the output of linear attention; For kernel feature mapping function; This is the activation function for the Exponential Linear Unit. The time complexity of this calculation is reduced to ,in By specifying the input sequence length, the computational overhead of long sequence inference is significantly reduced.

[0046] A feedforward network layer, consisting of two fully connected layers, is connected after the linear attention layer. The GELU activation function is used in between, and the dimensionality is first expanded to 4D and then compressed back to D. This pathway outputs a global temporal feature matrix. The dimensions are [L,D].

[0047] 4. Cross-branch attention fusion module A cross-branch multi-head attention mechanism is used to achieve deep interaction between two sets of features: convolutional features... As a query, Transformer features As keys and values, calculate the cross-path attention output; then use the Transformer features... As a query, convolutional features The back attention output is calculated using the two attention outputs as keys and values; after element-wise addition of the two attention outputs, the result is normalized to obtain the fused feature matrix. The dimensions are [L,D].

[0048] This fusion method achieves bidirectional complementarity between local details and global trends, solves the feature fragmentation problem caused by simple splicing and fusion, and improves the feature recognition of early weak faults.

[0049] 5. Cross-layer contrast constraints during training phase During model training, a cross-layer contrastive loss function is introduced to align the feature distributions of the convolutional and Transformer pathways. For the same batch of samples, global average pooling features from both outputs are extracted, and the contrastive loss is calculated. ; In the formula: The cross-layer contrastive loss function value; The global average pooling feature vector is the output of the same sample through the convolutional path; This is the global average pooling feature vector output from the Transformer path for the same sample; Let be the feature vector of the j-th other sample in the batch; This is the function for calculating cosine similarity. The temperature coefficient of the comparison loss is used to adjust the interval size of the characteristic distribution.

[0050] This loss function makes the two features of the same sample closer in the semantic space, while widening the feature distance between samples of different categories, thus improving the model's robustness in identifying early and minor faults.

[0051] The model's total loss function is a weighted sum of the cross-entropy classification loss and the contrastive loss: ; In the formula: This represents the total loss function value of the model. The value of the cross-entropy classification loss function; To compare the balancing weight of the loss, this embodiment uses 0.3.

[0052] 6. INT8 Quantization Module After model training is complete, symmetric quantization is used to quantize the model weights and activation values ​​from FP32 floating-point format to INT8 integer format. The quantization formula is as follows: ; In the formula: This is a floating-point value with the original FP32 precision; This is the quantized integer value with INT8 precision; This is the quantization scaling factor; To quantize the zero-point offset, the value is set to 0 under symmetric quantization; round is the rounding function.

[0053] After quantization, the model size is about 1 / 4 of the original FP32 model, and the inference speed is increased by 2 to 3 times, adapting to the computing power and storage constraints of edge computing platforms.

[0054] 7. Output Layer Fusion feature matrix Global average pooling yields a global feature vector of dimension D. By connecting a fully connected layer and using the Softmax activation function, the probability distribution of K types of faults is output: ; In the formula: This is a one-dimensional feature vector after global average pooling, with dimension [D]. The weight matrix of the output layer has dimensions [D,K]. Let K be the bias vector of the output layer, with dimension [K]. The normalized activation function; Let K be the probability distribution vector corresponding to the K types of faults, with dimension [K]. This represents the total number of fault categories.

[0055] Example 3: Implementation of Cloud-Based Search Enhancement for Generating Interpretable Diagnostic Reports This embodiment details the complete process of generating interpretable diagnostic reports in the cloud based on search-enhanced generation technology. The process is as follows: Figure 6 As shown, the specific steps are as follows: Step 1: Data and Feature Reception The cloud server receives troublesome fault samples uploaded from the edge side, including multi-source time-series raw data, feature vectors output by the intermediate layer of the lightweight Transformer model, and preliminary inference results and confidence levels from the edge side.

[0056] Step 2: Feature semantic mapping A pre-trained feature-text mapping model is used to convert the physical feature vectors extracted from the edge side into natural language semantic descriptions. For example, the high-frequency energy enhancement of the vibration signal and the appearance of specific characteristic frequencies in the envelope spectrum are mapped to a standardized semantic description of "the high-frequency component of the vibration of the drive end bearing is enhanced, and the outer ring fault characteristic frequency appears in the envelope spectrum".

[0057] Step 3: Knowledge Base Vector Retrieval The pre-built traction system fault knowledge base includes a fault mechanism library, a typical case library, a maintenance standard library, and a regulatory standard library. All knowledge entries have been annotated by domain experts and stored in vector form.

[0058] The semantic description generated in step 2 is used as a query statement. After being converted into a query vector, cosine similarity retrieval is performed in the knowledge base to return the top 5 knowledge items with the highest similarity and their relevance scores, covering the physical mechanism of the fault, historical fault cases of elevators of the same model, standard maintenance procedures, and other content.

[0059] Step 4: Enhance prompt word construction The preliminary inference results from the edge side, semantic physical feature descriptions, and retrieved knowledge entries are structurally combined to construct enhanced prompt words that conform to the input format of a large language model. An example of a prompt word template is shown below: json { "task":"Generate elevator traction system fault diagnosis report", "edge_inference_result":{ "fault_type":"string", "confidence":"float", "top_features":["feature1","feature2"] }, "physical_features":"string", "retrieved_knowledge":[ {"entry":"knowledge_text1","relevance":0.95}, {"entry":"knowledge_text2","relevance":0.89} ], "output_format":"markdownwithsections:fault type, fault location, severity level, possible causes of the fault, recommended maintenance measures"} ``` The prompt explicitly requires all conclusions to cite sources of knowledge, and clearly states any uncertainties when information is insufficient.

[0060] Step 5: Generate Large Model Report The constructed enhanced prompts are input into a large language model fine-tuned for the elevator domain. The model combines retrieval knowledge with physical features to generate a structured diagnostic report. All conclusions are linked to knowledge base sources, avoiding unfounded and misleading outputs. Real-world testing data shows that adopting this RAG architecture reduces the factual error rate of the diagnostic report by 71.4%, significantly improving its practicality.

[0061] Step 6: Report Output and Push The generated diagnostic reports are stored in a cloud database in structured JSON format and sent to the mobile applications of the corresponding maintenance personnel via push notification service. They are also displayed visually on the cloud platform's web management backend.

[0062] Example 4: Implementation of Collaborative Update of the "End-Edge-Cloud" Model Example 4: Implementation of Collaborative Update of the "End-Edge-Cloud" Model This embodiment details the model iteration and update mechanism for cloud-edge collaboration, as follows: Figure 7 As shown, the specific steps are as follows: Step 1: Data Aggregation and Difficult Case Screening The cloud server periodically aggregates operational data uploaded by all connected edge nodes, focusing on collecting low-confidence samples from the edge side, manually verified and labeled fault samples, and newly added fault type samples to construct an incremental training dataset. Simultaneously, the data is cleaned and deduplicated to ensure training data quality.

[0063] Step 2: Cloud-based model training and optimization The cloud-based system performs full retraining or incremental fine-tuning on the basic dual-path Transformer model based on the full dataset. The training uses a joint loss function of cross-entropy loss and contrastive loss, employs the Adam optimizer, sets the initial learning rate to 0.001, and sets the batch size to 64. The number of training epochs is dynamically adjusted according to the dataset size.

[0064] Data augmentation techniques are used during training to improve the model's generalization ability, including adding random noise, amplitude scaling, and time shifting to the time series signal.

[0065] Step 3: Lightweight Quantization of the Model The high-precision model obtained through cloud training is FP32 precision and requires quantization compression to adapt to the computing power of edge platforms. A quantization-aware training method is adopted to simulate quantization errors during training, ultimately generating a lightweight model file with INT8 precision. This significantly compresses model size and inference latency while ensuring that the precision loss is less than 1%.

[0066] Step 4: Model Encryption and Distribution The cloud distributes the quantized lightweight model file to each edge computing node via a TLS encrypted transmission channel, with the model version number, checksum, and update instructions included during the distribution process. A model version repository is established in the cloud to manage all historical model versions, supporting version tracking and rollback.

[0067] Step 5: Edge-side hot-reload update After receiving the model file, the edge computing node first verifies the file's integrity using a checksum. Once confirmed, it performs a hot model reload: without interrupting the current monitoring service, it loads the new model weights into memory, completing a seamless switch of the inference engine. After the update is complete, it sends the update results and the current model version back to the cloud.

[0068] If diagnostic errors occur after the update, you can roll back to the previous stable version with one click to ensure the continuity of on-site monitoring services.

[0069] Example 5: Complete Implementation of an Early Fault Diagnosis System for Elevator Traction Systems This embodiment is a complete implementation plan of the system, which elaborates on the hardware composition, deployment method and working logic of each unit of the system. The overall system architecture is as follows: Figure 1 As shown.

[0070] This system consists of four parts: edge acquisition unit, edge computing unit, cloud diagnostic unit, and model collaborative update unit. It adopts a hierarchical collaborative architecture of "edge acquisition - edge inference - cloud enhancement - collaborative evolution".

[0071] 1. End-side data acquisition unit The end-side acquisition unit is deployed in the elevator machine room traction system and consists of a sensor group, acquisition controller, local cache module and communication module.

[0072] Sensor group: includes 2 triaxial MEMS accelerometers, 2 Hall current sensors, 1 PT100 temperature sensor, 1 photoelectric encoder, and 1 laser displacement sensor, which are deployed at the corresponding monitoring points of the traction system to realize the synchronous acquisition of multi-source signals such as vibration, current, temperature, speed, and displacement.

[0073] Acquisition controller: It adopts the high-performance STM32H7 series microcontroller, which is responsible for multi-sensor synchronous triggering, A / D conversion, data preprocessing and transmission control; it supports high-frequency sampling above 2kHz and has real-time data processing capabilities.

[0074] Local caching module: Equipped with a 4GB high-speed storage card, it is used to temporarily store collected data when the network is interrupted, and automatically resumes the transmission after the network is restored to avoid data loss.

[0075] Communication module: integrates a 4G / 5G wireless communication module and an industrial Ethernet interface, supporting stable data transmission with edge computing units.

[0076] 2. Edge computing unit The edge computing unit is deployed in the elevator machine room or control cabinet. It adopts the NVIDIA Jetson OrinNX edge AI computing platform, pre-installed with the Ubuntu operating system and PyTorch deep learning inference framework. A single edge unit can connect to the end-side acquisition devices of 5 to 10 elevators.

[0077] The edge computing unit has a built-in dual-path fusion INT8 quantized lightweight Transformer model. Its core functions include: receiving and parsing time-series data uploaded by multiple elevators and caching it by device and channel; performing fault inference in real time with an average inference latency of 114ms per sample, meeting the requirements for millisecond-level early warning; performing graded early warning according to fault level, pushing early warning information for minor faults and triggering a safety shutdown by linking the elevator control system for severe faults; and uploading difficult fault samples to the cloud and receiving and executing model updates.

[0078] 3. Cloud-based diagnostic unit The cloud-based diagnostic unit is deployed on a cloud server cluster and includes a data storage module, a fault knowledge base, a search-enhanced generation engine, a visualization monitoring module, and a user management module.

[0079] Data storage module: Uses a time-series database to store all elevator operation data and diagnostic records, supporting efficient retrieval of massive amounts of data; Fault Knowledge Base: Structured storage of knowledge in the fields of traction system fault mechanisms, typical cases, maintenance specifications, regulations and standards, supporting vector retrieval; Search Enhancement Generation Engine: Integrates vector retrieval components with a domain-fine-tuned large language model to achieve interpretable deep diagnosis of complex faults; Visualized monitoring module: Provides a digital twin monitoring panel that displays the real-time operating status, health index, early warning records and historical reports of all connected elevators, and supports multi-dimensional statistical analysis.

[0080] 4. Model Collaborative Update Unit The model collaborative update unit is divided into a cloud management end and an edge execution end: the cloud management end is responsible for model training, quantization compression, version management and distribution scheduling, and supports targeted push of adapted models according to elevator model and working condition; the edge execution end is responsible for receiving model files, verification, hot reloading updates and version feedback, and supports one-click rollback.

[0081] When the system is working, the edge acquisition unit collects the traction system's operating signals in real time, preprocesses them, and then uploads them to the edge computing unit. The edge computing unit performs real-time fault reasoning through a lightweight model, continuously monitors the normal state, handles faults according to their severity, and uploads difficult faults to the cloud. The cloud generates interpretable diagnostic reports for difficult faults and periodically optimizes the model based on the full amount of data, then sends the updated data to the edge, forming a complete closed loop of "monitoring-diagnosis-optimization".

[0082] Example 6: Experimental Verification and Typical Fault Diagnosis Use Cases To verify the technical effect of the present invention, a three-month field test was conducted on a passenger elevator with a rated load of 1350kg and a rated speed of 2.5m / s in a hospital's medical technology building. A complete set of data acquisition and diagnostic equipment was deployed to collect time-series datasets including normal operating conditions and various early simulated faults, and the results were compared with existing mainstream methods.

[0083] 1. Comparison of experimental results The traditional CNN-LSTM-based method, the VMD-GRU-Transformer combined method, and the standard Transformer method were selected for comparison. The test results are shown in Table 1 below. Table 1. Experimental results of different methods

[0084] The experimental results above demonstrate that the method proposed in this invention significantly outperforms existing methods in terms of diagnostic accuracy, real-time performance, and model lightweighting. Particularly in edge real-time inference scenarios, this invention achieves low-latency inference of 214 milliseconds, meeting the requirements for real-time fault detection in elevator traction systems; the overall diagnostic accuracy reaches 92.0%, with the accuracy for complex fault diagnosis improved to 82.5%; the application of RAG technology significantly reduces the illusion rate of intelligent diagnostic reports, effectively ensuring the reliability and practicality of the generated reports.

[0085] 2. Typical Fault Diagnosis Use Cases Use case: Early wear of the outer ring of the traction machine drive end bearing During testing, the lightweight Transformer model on the edge side continuously monitored the vibration signal of the drive-end bearing housing in the X direction, which showed energy enhancement in the high-frequency range (2kHz~5kHz), and specific fault characteristic frequency components appeared in the envelope spectrum of the vibration signal. The model output the probability of bearing wear failure as 0.89, triggering an edge-level early warning.

[0086] After receiving the sample and its edge features, the cloud-based RAG engine retrieves bearing fault-related entries from the fault knowledge base and ultimately generates the following diagnostic report: Fault type: Early wear of the outer ring of the rolling bearing at the drive end of the traction machine. Fault location: Traction machine drive end bearing housing Severity level: Level 2 (early warning level, it is recommended to arrange specific testing within two weeks) Possible causes of failure: poor lubrication due to aging of bearing grease, or excessive bearing preload leading to prolonged overload. Diagnostic criteria: Increased energy in the vibration signal within the 2kHz~5kHz frequency band; presence of bearing outer ring fault characteristic frequencies in the envelope spectrum, with peak values ​​3.2 times the normal baseline. Recommended maintenance measures: (1) Arrange maintenance personnel to replenish or replace bearing grease; (2) Check bearing installation clearance and preload; (3) Increase the frequency of monitoring the elevator's operating status.

[0087] The diagnostic report was verified on-site after unpacking and confirmed to be consistent with the actual situation, proving that the diagnostic method of the present invention has high accuracy and practicality.

[0088] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for early fault diagnosis of elevator traction systems based on dual-path fusion "end-edge-cloud" collaboration, characterized in that, Includes the following steps: S1. End-side multi-source signal acquisition: Deploy multiple types of sensors at key parts of the elevator traction system to synchronously acquire multi-source raw signals under the operating status of the traction system; S2. Edge-side data preprocessing and uploading: The collected raw signals are cleaned and standardized, and the preprocessed multi-dimensional time series data is transmitted to the edge computing node. S3. Edge-side dual-path fusion feature extraction and fault reasoning: A lightweight temporal Transformer diagnostic model is deployed on the edge computing node. The model uses a dual-path parallel structure to extract local fault features and global temporal degradation features respectively. After cross-path feature fusion, early fault reasoning is performed, and the fault type and corresponding confidence level are output. S4. Cloud-based Explainable Enhanced Diagnosis: For difficult fault samples with edge-side confidence levels below a preset threshold, their time-series data and intermediate features are uploaded to the cloud server. The cloud uses retrieval enhancement generation technology to integrate edge-side features with the traction system fault knowledge base to generate an explainable diagnostic report containing fault location, cause analysis, and maintenance suggestions. S5. Collaborative Iterative Model Update: The cloud-based diagnostic model is optimized based on full-scale running data and labeled fault samples. The updated lightweight model parameters are then distributed to edge computing nodes to achieve online upgrades of the edge diagnostic model.

2. The method for early fault diagnosis of elevator traction system according to claim 1, characterized in that, In step S1, the multiple types of sensors include a triaxial acceleration sensor installed in the bearing housing of the traction machine, a current sensor installed in the electrical circuit of the traction machine, a temperature sensor installed in the housing of the traction machine, a speed sensor installed at the traction wheel, and a vibration displacement sensor installed at the traction wire rope; the sampling frequency of all vibration sensors is not less than 2kHz.

3. The method for early fault diagnosis of elevator traction system according to claim 1, characterized in that, The dual-path parallel structure described in step S3 includes a multi-receptive-field convolutional feature path and a lightweight temporal attention path. The multi-receptive-field convolutional feature path extracts local instantaneous impact and micro-damage features through multi-scale parallel convolution and channel attention mechanisms. The lightweight temporal attention path extracts global temporal degradation features based on a lightweight Transformer structure. Cross-path feature fusion employs a cross-branch multi-head attention mechanism to achieve bidirectional interaction and deep fusion of features from two paths.

4. The method for early fault diagnosis of elevator traction system according to claim 1, characterized in that, The lightweight temporal Transformer diagnostic model described in step S3 uses a linear attention mechanism to replace the standard self-attention calculation to reduce computational complexity, and compresses the model size and inference latency through INT8 quantization technology; the model is equipped with an industrial position coding module, which uses a combination of trigonometric function position coding and learnable operating condition coding to adapt to elevator variable operating condition operation signals.

5. The method for early fault diagnosis of elevator traction system according to claim 1, characterized in that, In step S3, a graded early warning is executed based on the confidence level of the fault reasoning and the fault type: continuous real-time monitoring is performed under normal conditions; When an early, minor fault is detected, an early warning message is sent and a planned maintenance reminder is triggered; When a severe fault is detected, an elevator safety shutdown command is triggered and fault information is reported simultaneously.

6. The method for early fault diagnosis of an elevator traction system according to claim 1, characterized in that, The execution flow of the retrieval enhancement generation technology in step S4 is as follows: the physical features extracted from the edge side are mapped to natural language semantic descriptions, the semantic descriptions are used as retrieval statements to perform vector similarity retrieval in the fault knowledge base, the retrieved domain knowledge, edge reasoning results and semantic descriptions are used to jointly construct prompt words, and the prompt words are input into a large language model to generate a structured diagnostic report.

7. An early fault diagnosis system for elevator traction systems based on dual-path fusion "end-edge-cloud" collaboration, the system being based on the method described in any one of claims 1-6, characterized in that... It includes an edge acquisition unit, an edge computing unit, a cloud diagnostic unit, and a model collaborative update unit: The end-side acquisition unit is deployed in key parts of the elevator traction system and is equipped with multiple types of sensors and acquisition controllers to synchronously acquire multi-source operating signals of the traction system and upload them after preprocessing the signals. The edge computing unit is communicatively connected to the end-side acquisition unit and is equipped with a lightweight time-series Transformer diagnostic model with dual-path fusion. It is used to receive time-series data uploaded from the end side and perform real-time fault reasoning, outputting fault type and confidence level. The cloud-based diagnostic unit is communicatively connected to the edge computing unit and is equipped with a traction system fault knowledge base and a retrieval enhancement generation engine. It is used to receive difficult fault samples uploaded from the edge side and generate interpretable diagnostic reports. The model collaborative update unit is deployed on the cloud and the edge respectively, and is used to realize the encrypted distribution and online update of the lightweight model parameters optimized in the cloud to the edge computing nodes.

8. The elevator traction system early fault diagnosis system according to claim 7, characterized in that, The end-side acquisition unit is also equipped with a local data cache module and a wireless communication module. The local data cache module is used to temporarily store the acquired data when the network is interrupted, and to resume the transmission after the network is restored.

9. The early fault diagnosis system for elevator traction systems according to claim 7, characterized in that, The edge computing unit adopts an edge AI computing platform and is deployed in the elevator machine room or control cabinet. It supports INT8 quantization model inference, and the average latency of single-sample inference is no higher than 200ms.

10. The elevator traction system early fault diagnosis system according to claim 7, characterized in that, The cloud-based diagnostic unit is also equipped with a digital twin monitoring panel, which is used to visually display the operating status, health level, early warning records and historical diagnostic reports of the connected elevator; the model collaborative update unit supports hot loading and updating of the model and version rollback.

Images