Multi-manufacturer elevator fault diagnosis model joint training method and system based on federated learning and privacy calculation

Abstract

The invention discloses a multi-manufacturer elevator fault diagnosis model joint training method and system based on federated learning and privacy calculation. Cross-manufacturer data collaborative training is achieved through a federated learning framework. Local nodes of manufacturers perform feature extraction on multi-modal data such as vibration signals and video streams, data privacy is protected by adopting differential privacy and homomorphic encryption technologies, and a central server dynamically adjusts aggregation weight according to data volume and distribution difference. According to the method, the problem of elevator industry data islands is solved for the first time, efficient aggregation of non-IID data is supported, the model accuracy is improved by 28.3%, meanwhile, the GDPR compliance requirement is met, and the method is suitable for an intelligent elevator Internet of Things scene.

Description

Technical Field

[0001] This invention lies at the intersection of distributed machine learning, data privacy protection, and elevator fault diagnosis technologies. Specifically, it relates to a method and system for collaborative training of multi-vendor elevator fault data based on a federated learning framework. By combining federated learning with privacy-preserving computing technology, this method enables the joint training and optimization of cross-vendor elevator fault diagnosis models while protecting the data privacy of each vendor. This addresses the lack of model generalization caused by data silos and fills a technological gap in the application of federated learning in the field of elevator electromechanical equipment health management. Background Art

[0002] Traditional ground transportation requires continuous maintenance and annual inspections through various policies and regulations. Similarly, vertical transportation facilities, which are crucial to the safety of passengers and property, also require inspection and maintenance. Vertical transportation facilities, such as elevators, are mechatronic devices, and their safety inspections are primarily manual, with professionals carrying testing equipment performing regular inspections, maintenance, and safety assessments. This approach is inefficient and heavily reliant on manual experience.

[0003] In addition, comprehensive monitoring of the elevator electromechanical equipment itself is carried out. Its status signal sampling points are widely distributed, the sampling frequency is high, and the operation data collection time is long. In addition, elevators are clustered distributed equipment. These characteristics make the overall operation data volume of elevator equipment present a massive data effect. Its data characteristics are reflected in the following aspects: 1) High dimensionality: The monitoring database is TB-level in scale. It is unrealistic to rely on traditional expert online analysis. It is necessary to study big data processing methods for intelligent analysis; 2) Multi-type: The monitoring data covers multiple signals such as vibration, ambient temperature and humidity, dust, video stream, deviation, pressure, current, voltage, etc., which are difficult to process under the same framework. It is easy to encounter problems such as weak fault signals being difficult to detect, input variables being unbalanced, and multi-variable fusion being difficult; 3) Strong correlation: The operation status data of elevators is mostly long-term time series. The current status data can only be used for static threshold alarm function: If the status monitoring and maintenance of clustered elevators is carried out, it is necessary to collect historical data from the monitoring database. While the massive amount of data generated by comprehensive monitoring of electromechanical equipment provides researchers with a great deal of value for online analysis and diagnosis of elevator health status, it also presents many challenges.

[0004] Existing elevator fault diagnosis relies on a large amount of historical operating data (such as vibration signals, temperature, current, and video streams). However, due to commercial confidentiality and data sovereignty regulations (such as GDPR), manufacturers cannot share this data, leading to the following problems:

[0005] Data silos: Insufficient data from a single vendor can lead to overfitting of models and an inability to cover complex failure scenarios.

[0006] Privacy leakage risk: Traditional centralized training requires uploading raw data to a central server, violating privacy compliance requirements;

[0007] Non-IID (non-independent and identically distributed) data challenges: Data distribution varies greatly among different manufacturers (e.g., elevator models and usage environments), making traditional federated learning aggregation strategies less effective.

[0008] Existing Federated Learning: Proposed by Google in 2016, it is used for collaborative training of mobile device data, but it does not involve multimodal data scenarios of industrial equipment.

[0009] Elevator fault diagnosis technology: Existing solutions are mostly based on single-vendor data and use models such as CNN and LSTM (such as patent CN202010589806.4), but cross-vendor data fusion has not yet been solved; Privacy computing technology: Homomorphic encryption (Paillier), differential privacy (DP), etc. are used in medical and financial fields (such as patent CN202111060093.3), but have not been combined with the electromechanical data characteristics of elevators.

[0010] At least the above-mentioned existing technologies have not solved the following problems: 1. Federal feature alignment of elevator multimodal data (vibration spectrum, video stream); 2. Dynamic federal aggregation strategy under non-IID data; 3. Design of lightweight privacy computing solutions for elevator fault diagnosis. Summary of the Invention

[0011] Solve the problems of insufficient model training data, privacy leakage risks and low efficiency of non-IID data aggregation caused by data confidentiality among elevator manufacturers.

[0012] In order to solve the above technical problems, the technical methods provided by the present invention are as follows:

[0013] A method for jointly training a multi-vendor elevator fault diagnosis model based on federated learning is provided, comprising the following steps:

[0014] S1. Each manufacturer's local node extracts features from multimodal elevator data to generate a standardized training set. S2. The central server initializes the global diagnostic model and sends it to each local node. S3. The local node uses differential privacy technology to add noise to the training gradients, encrypts them, and uploads them to the central server.

[0015] S4. The central server calculates dynamic weights based on data volume and distribution differences, aggregates model parameters and updates the global model; S5. Iterates training and aggregation until the model converges.

[0016] In order to better achieve the purpose of the invention, the present invention also has the following better solutions:

[0017] In some embodiments, the multimodal data includes a vibration signal spectrum, temperature time series data, and video stream key frame features, and the feature extraction is specifically: performing fast Fourier transform on the vibration signal to generate a Mel spectrum; using YOLOv5 to detect abnormal behavior on the video stream and extracting HOG features of the key frames.

[0018] In some embodiments, the dynamic weight calculation satisfies the following formula: i =(n i ) / (∑n i )×1 / (1+KL(D i ‖D g local)) where ni is the amount of local data, KL(D i ‖D g local)) is the KL divergence between the local data distribution and the global distribution.

[0019] In some embodiments, the differential privacy technology uses adaptive noise injection, and the noise scale σ decays as the number of training rounds increases. The decay formula is: σ t =σ0×e -0.05t , where σ0 is the initial noise scale and t is the current round.

[0020] In some embodiments, the global diagnostic model has a two-branch structure, comprising:

[0021] Branch 1: ResNet network, the input is the vibration signal spectrum; Branch 2: BiLSTM network, the input is temperature and time series current data; Fusion layer: The output features of the two branches are spliced ​​and classified through the fully connected layer.

[0022] In some implementations, before model aggregation, the central server performs homomorphic decryption on the received encrypted parameters using a Paillier decryption algorithm.

[0023] The present invention also provides a federated learning system for a multi-vendor elevator fault diagnosis model joint training method based on federated learning, comprising:

[0024] Central server: includes model aggregation module, dynamic weight calculation module, and secure communication module;

[0025] Multiple local nodes: deployed within each manufacturer, including data preprocessing modules and privacy protection training modules;

[0026] Edge computing gateway: Deployed in the factory LAN, it uses local model reasoning and real-time fault warning.

[0027] In order to better achieve the purpose of the invention, the present invention also has the following better solutions:

[0028] In some embodiments, the local node supports multimodal data input and integrates a lightweight model pruning module with a pruning rate of 30%-50%.

[0029] In some embodiments, the secure communication module uses the TLS 1.3 protocol and generates a temporary key for each session.

[0030] In some embodiments, the edge computing gateway is connected to the central server via 5G-MEC to support low-latency model updates.

[0031] The present invention also provides an electronic device comprising a memory, a processor, a computer program stored on the memory, and a computer-readable storage medium storing the computer program. When the processor executes the program, the above-mentioned multi-vendor elevator fault diagnosis model joint training method based on federated learning is implemented.

[0032] The specific method and process include: local preprocessing of multimodal data: vibration signal: time domain to frequency domain conversion (FFT), extraction of Mel-frequency spectrum; video stream: detection of abnormal elevator behavior based on YOLOv5, extraction of keyframe features; construction of local training set: annotation of fault types (such as traction motor wear, door system failure). Federated model initialization: the central server sends the global model (a two-branch CNN-LSTM network), branch 1 processes the spectrum, and branch 2 processes the time series data; local nodes load the model and initialize local parameters. Privacy-preserving local training: dynamic mask encryption is used: random noise is added to the gradient to meet (ε, δ)-differential privacy; local training loss function: cross entropy loss + feature distribution consistency constraint (to alleviate the Non-IID problem).

[0033] Weighted dynamic federated aggregation:

[0034] Calculate the data volume weight of each node:

[0035] w i =(n i ) / (∑n i )×1 / (1+KL(D i ‖D g lobal), KL divergence measures the difference in data distribution; aggregation formula:

[0036] Model iteration and evaluation: The central server verifies the global model accuracy and triggers the early stopping mechanism (termination occurs if the loss decreases by <1% for three consecutive rounds).

[0037] System Architecture

[0038] Central server module:

[0039] Model management: initialization, version control, aggregation algorithm library; secure communication: TLS1.3 protocol to prevent man-in-the-middle attacks; dynamic weight calculation: weight allocation based on data volume and distribution differences.

[0040] Local node module: data preprocessing: multimodal data standardization, feature extraction;

[0041] Local model training: lightweight model pruning (reducing communication overhead); privacy protection: differential privacy noise injection, homomorphic encryption transmission.

[0042] Thanks to the adoption of the above technical solutions, the first elevator federated training framework was created: for the first time, federated learning and privacy computing were combined to solve the data silo problem in the elevator industry; multimodal federated feature alignment: a dual-branch model was designed to be compatible with the joint training of vibration spectrum and time series data; dynamic weighted aggregation strategy: the KL divergence was introduced to adjust the weights to improve the model convergence speed under non-IID data; lightweight privacy protection: dynamic mask encryption reduces the accuracy loss by 30% compared to traditional DP. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 :Federated learning system architecture and interaction flow chart;

[0044] Figure 2 : Multimodal data processing flow chart;

[0045] Figure 3 : Accuracy comparison curve of federated learning and traditional methods in the embodiment. DETAILED DESCRIPTION

[0046] Below in conjunction with embodiment, the present invention is described in further detail:

[0047] refer to Figures 1 to 3 As shown, the present invention provides a multi-vendor elevator fault diagnosis model joint training method based on federated learning, including the following steps:

[0048] S1. Each manufacturer's local node extracts features from multimodal elevator data to generate a standardized training set. S2. The central server initializes the global diagnostic model and sends it to each local node. S3. The local node uses differential privacy technology to add noise to the training gradients, encrypts them, and uploads them to the central server.

[0049] S4. The central server calculates dynamic weights based on data volume and distribution differences, aggregates model parameters, and updates the global model. S5. Iterates training and aggregation until the model converges. Multimodal data includes vibration signal spectrograms, temperature time series data, and key frame features of video streams. The feature extraction is specifically as follows: Fast Fourier transform of vibration signals to generate Mel spectrum; YOLOv5 is used to detect abnormal behavior in video streams and extract HOG features of key frames. The dynamic weight calculation satisfies the following formula: w i =(n i ) / (∑n i )×1 / (1+KL(D i ‖D g local)) where ni is the amount of local data, KL(D i ‖D g lobal)) is the KL divergence between the local data distribution and the global distribution. Differential privacy technology uses adaptive noise injection, and the noise scale σ decays with the increase of training rounds. The decay formula is: σ t =σ0×e -0.05t , where σ0 is the initial noise scale and t is the current round.

[0050] The global diagnostic model has a dual-branch structure, including: Branch 1: ResNet network, whose input is the vibration signal spectrum; Branch 2: BiLSTM network, whose input is temperature and time-series current data; Fusion layer: The output features of the two branches are concatenated and classified through the fully connected layer. Before model aggregation, the central server performs homomorphic decryption on the received encrypted parameters using the Paillier decryption algorithm.

[0051] This federated learning system, based on a multi-vendor elevator fault diagnosis model joint training method, consists of a central server with a model aggregation module, a dynamic weight calculation module, and a secure communication module; multiple local nodes deployed within each vendor, including a data preprocessing module and a privacy-preserving training module; and an edge computing gateway deployed within the factory's local area network, which uses local models for inference and real-time fault warning. The local nodes support multimodal data input and integrate a lightweight model pruning module with a pruning rate of 30%-50%. The secure communication module uses the TLS 1.3 protocol, generating a temporary key for each session. The edge computing gateway connects to the central server via 5G-MEC, supporting low-latency model updates.

[0052] In addition, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory, as well as a computer-readable storage medium storing the computer program. When the processor executes the program, the above-mentioned multi-vendor elevator fault diagnosis model joint training method based on federated learning is implemented.

[0053] The specific method and process include: local preprocessing of multimodal data: vibration signal: time domain to frequency domain conversion (FFT), extraction of Mel-frequency spectrum; video stream: detection of abnormal elevator behavior based on YOLOv5, extraction of keyframe features; construction of local training set: annotation of fault types (such as traction motor wear, door system failure). Federated model initialization: the central server sends the global model (a two-branch CNN-LSTM network), branch 1 processes the spectrum, and branch 2 processes the time series data; local nodes load the model and initialize local parameters. Privacy-preserving local training: dynamic mask encryption is used: random noise is added to the gradient to meet (ε, δ)-differential privacy; local training loss function: cross entropy loss + feature distribution consistency constraint (to alleviate the Non-IID problem).

[0054] Weighted dynamic federated aggregation:

[0055] Calculate the data weight of each node: w i =(n i ) / (∑n i )×1 / (1+KL(D i ‖D g lobal), KL divergence measures the difference in data distribution; aggregation formula:

[0056] Model iteration and evaluation: The central server verifies the global model accuracy and triggers the early stopping mechanism (termination occurs if the loss decreases by <1% for three consecutive rounds).

[0057] The system architecture includes a central server module: model management (initialization, version control, and aggregation algorithm library); secure communication (TLS 1.3 protocol to prevent man-in-the-middle attacks); dynamic weight calculation (weight allocation based on data volume and distribution differences); local node module: data preprocessing (multimodal data standardization and feature extraction); local model training (lightweight model pruning to reduce communication overhead); and privacy protection (differential privacy noise injection and homomorphic encryption transmission).

[0058] Example 1: Standard federated training scenario

[0059] Data preparation includes:

[0060] Participating manufacturers: 3 (A / B / C), providing 10,000 / 8,000 / 12,000 data items respectively; Fault types: 6 categories (traction motor failure, door system abnormality, etc.).

[0061] Model configuration: Global model: ResNet-34 (spectral branch) + BiLSTM (temporal branch); Optimizer: AdamW, learning rate 0.001, weight decay 1e-4.

[0062] Privacy parameters: Differential privacy: ε = 2.0, δ = 1e-5, noise scale σ = 0.8;

[0063] Homomorphic encryption: Paillier algorithm, key length 2048 bits.

[0064] Training results:

[0065] Federated model accuracy: 94.7% (28.3% higher than the highest single-vendor model); communication overhead: 62% reduction in data transmission per round (model pruning + gradient compression).

[0066] Example 2: Non-IID data scenario

[0067] Data distribution: Manufacturer A focuses on traction motor failures, while Manufacturer B focuses on door system anomalies;

[0068] Dynamic weight adjustment: The KL divergence threshold is set to 0.3, triggering weight redistribution;

[0069] Results: The recall rate of the model on manufacturer B’s data increased by 19.6%.

[0070] The technical effects obtained according to these embodiments are as follows:

[0071] Data efficiency: Cross-vendor joint training increases model coverage of fault types by 45%;

[0072] Privacy and security: Meet GDPR compliance requirements, encryption parameters cannot be reversed to the original data;

[0073] Economic benefits: Manufacturer collaboration reduces individual data collection costs by approximately 37%;

[0074] Technical compatibility: supports 5G edge computing node deployment and is compatible with the smart elevator IoT architecture.

[0075] The above are only some embodiments of the present invention. For those skilled in the art, several modifications and improvements can be made without departing from the inventive concept of the present invention, which all fall within the scope of protection of the present invention.

Claims

1. A multi-vendor elevator fault diagnosis model joint training method based on federated learning, characterized in that: The following steps are involved: S1. Each manufacturer's local node extracts features from multimodal elevator data and generates a standardized training set. S2. The central server initializes the global diagnostic model and sends it to each local node; S3, the local node uses differential privacy technology to add noise to the training gradient, encrypts it and uploads it to the central server; S4. The central server calculates dynamic weights based on data volume and distribution differences, aggregates model parameters, and updates the global model. S5. Iterate the training and aggregation until the model converges.

2. The method according to claim 1, characterized in that The multimodal data includes vibration signal spectrogram, temperature time series data and video stream key frame features, and the feature extraction is specifically as follows: Perform fast Fourier transform on the vibration signal to generate Mel spectrum; YOLOv5 is used to detect abnormal behaviors in video streams and extract HOG features of key frames.

3. The method according to claim 1, characterized in that The dynamic weight calculation satisfies the following formula: i =(n i ) / (∑n i )×1 / (1+KL(D i ‖D g local)) where ni is the amount of local data, KL(D i ‖D g local)) is the KL divergence between the local data distribution and the global distribution.

4. The method according to claim 1, wherein The differential privacy technology uses adaptive noise injection, and the noise scale σ decays with the increase of training rounds. The decay formula is: σ t =σ0×e -0.05t , where σ0 is the initial noise scale and t is the current round.

5. The method according to claim 1, wherein The global diagnostic model has a dual-branch structure, including: Branch 1: ResNet network, input is vibration signal spectrum; Branch 2: BiLSTM network, with temperature and time series current data as input; Fusion layer: The output features of the two branches are concatenated and classified through the fully connected layer.

6. The method according to claim 1, wherein Before model aggregation, the central server performs homomorphic decryption on the received encrypted parameters using the Paillier decryption algorithm.

7. A federated learning system implementing the method according to any one of claims 1 to 6, characterized in that: include: Central server: includes model aggregation module, dynamic weight calculation module, and secure communication module; Multiple local nodes: deployed within each manufacturer, including data preprocessing modules and privacy protection training modules; Edge computing gateway: deployed in the factory local area network for local model reasoning and real-time fault warning.

8. The federated learning system according to claim 7, characterized in that: The local node supports multimodal data input and integrates a lightweight model pruning module with a pruning rate of 30%-50%.

9. The federated learning system according to claim 7, characterized in that: The secure communication module adopts the TLS 1.3 protocol and generates a temporary key for each session.

10. The federated learning system according to claim 7, characterized in that: The edge computing gateway is connected to the central server via 5G-MEC, supporting low-latency model updates.

11. An electronic device comprising a memory, a processor, and a computer program stored in the memory, wherein: When the processor executes the program, the method according to any one of claims 1 to 6 is implemented.

12. A computer-readable storage medium storing a computer program, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images