Cross-brand elevator fault prediction method and system based on transfer learning

Abstract

The invention discloses a cross-brand elevator fault prediction method and system based on transfer learning. The method comprises the steps that elevator data of multiple brands are collected, a multi-source elevator data set is constructed, and the multi-source elevator data set comprises sensor data and fault labels of the multiple brands of elevators; performing standardization processing on the multi-source elevator data set, and performing cross-brand feature alignment by adopting domain adaptation so as to minimize the data distribution difference between a source domain and a target domain; a deep transfer learning model is constructed, the model comprises a shared feature extraction layer and a domain adaptation module, the shared feature extraction layer extracts cross-brand common features through a pre-trained convolutional neural network, and the domain adaptation module aligns source domain and target domain feature spaces through an adversarial training strategy; and dynamically optimizing the model based on sample data, realizing online updating of model parameters in combination with an incremental learning algorithm, and outputting a fault type and a corresponding prediction probability. The cross-brand migration cost is low, the small sample scene adaptability is good, and the dynamic updating capacity is high.

Description

Technical Field

[0001] This invention relates to elevator fault prediction, and in particular to a cross-brand elevator fault prediction method and system based on transfer learning. Background Technology

[0002] With the increasing demand for intelligent elevator operation and maintenance, data-driven fault prediction technology has become a research focus, but it faces significant challenges in real-world cross-brand scenarios. Traditional methods mainly rely on rule-based thresholds or statistical models, which, while capable of basic anomaly detection, struggle to adapt to the dynamic changes in complex operating environments. While single-brand prediction models based on deep learning perform well on specific brand data, their cross-brand generalization ability declines sharply due to differences in sensor configuration, data dimensions, and fault modes across brands. Although existing research has attempted to reuse existing knowledge through traditional transfer learning, the heterogeneity of data distribution across different brands limits the transfer effect, and its heavy reliance on labeled data in the target domain makes it difficult to apply to small and medium-sized brands due to data scarcity. Furthermore, the dynamic nature of the elevator operating environment further exacerbates the performance degradation of static models. Existing technologies generally suffer from high cross-brand transfer costs, poor adaptability to small-sample scenarios, and insufficient dynamic update capabilities. Summary of the Invention

[0003] To address the aforementioned issues, this invention provides a cross-brand elevator fault prediction method and system based on transfer learning.

[0004] This invention provides the following technical solution: Firstly, a cross-brand elevator fault prediction method based on transfer learning is provided, including the following steps: S1. Collect elevator data from multiple brands to construct a multi-source elevator dataset, which includes sensor data and fault labels from multiple brands of elevators. S2. Standardize the multi-source elevator dataset and use domain adaptation to perform cross-brand feature alignment to minimize the data distribution difference between the source domain and the target domain, thereby obtaining a standard dataset. S3. Construct a deep transfer learning model, which includes a shared feature extraction layer and a domain adaptation module. The shared feature extraction layer extracts cross-brand common features of the standard dataset through a pre-trained convolutional neural network (CNN). The domain adaptation module aligns the source domain and target domain feature spaces of the standard dataset through an adversarial training strategy, thereby outputting the fault type and the corresponding predicted probability.

[0005] Furthermore, the sensor data includes vibration signals, current signals, and door status parameters; the fault label includes the fault type, occurrence time, and maintenance record.

[0006] Furthermore, the domain adaptation employs the maximum mean difference (MMD) metric or a domain adversarial neural network (DANN), and optimizes the adversarial training process between the domain classifier and the feature extractor through a gradient inversion layer (GRL). The domain adaptation is selected according to the data scenario: the maximum mean difference (MMD) is used for static feature distribution alignment; the domain adversarial neural network (DANN) combined with the gradient inversion layer (GRL) is used for dynamic adversarial training.

[0007] Furthermore, the training process of the deep transfer learning model adopts a hybrid loss function, including fault classification loss and domain adversarial loss. The domain adversarial loss is used to force the feature distribution of the source domain and the target domain to be consistent. The hyperparameter λ of the hybrid loss function is dynamically adjusted with each training round. The model is dynamically optimized based on sample data, and the model parameters are updated online by combining incremental learning algorithms.

[0008] Generally, the initial value of λ is 0.1, which is gradually increased linearly to 1.0. It can also be set as needed.

[0009] Furthermore, the dynamic optimization step includes: When the target brand data volume is less than n records, freeze the parameters of the shared feature extraction layer and only update the domain adaptation module. When the amount of data for the target brand increases to n or more, a few-shot learning strategy is used to fine-tune the entire model.

[0010] Generally, n is 500, but it can also be set as needed.

[0011] Secondly, a cross-brand elevator fault prediction system based on transfer learning is provided. This system, which implements the above method, includes: The multi-protocol data acquisition module is used to collect elevator data from multiple brands and construct a multi-source elevator dataset. The multi-source elevator dataset includes sensor data and fault tag data from elevators of multiple brands. It is used to access sensor data from elevators of different brands and complete protocol parsing and data normalization. The multi-protocol data acquisition module supports RS-485, CAN bus, and Modbus protocol parsing. The standardization processing module is used to standardize the multi-source elevator dataset and perform cross-brand feature alignment using domain adaptation to minimize the data distribution differences between the source domain and the target domain. The deep transfer learning model module is used to construct a deep transfer learning model. The model includes a shared feature extraction layer and a domain adaptation module. The shared feature extraction layer extracts cross-brand common features from a standard dataset through a pre-trained convolutional neural network (CNN). The domain adaptation module aligns the source domain and target domain feature spaces of the standard dataset through an adversarial training strategy, thereby outputting the fault type and the corresponding predicted probability.

[0012] Furthermore, the transfer learning engine module supports edge computing deployment and enables real-time inference on local elevator devices through lightweight model compression technology.

[0013] Furthermore, the fault early warning module is interconnected with the elevator operation and maintenance management platform, automatically generating work orders and assigning maintenance tasks based on the prediction results. The fault early warning module is interconnected with the elevator operation and maintenance management platform via a REST API, supporting automatic work order generation and MQTT protocol push.

[0014] Thirdly, an electronic device is provided, comprising: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the methods described above.

[0015] Fourthly, a computer-readable storage medium is provided that stores computer instructions thereon, which, when executed by a processor, implement the steps of the above-described method.

[0016] The beneficial effects of this invention are as follows: By employing a deep transfer learning model, shared features of elevators from different brands are extracted. A domain adaptation module is used to align the feature spaces of the source and target domains. Based on sample data, the model parameters are updated online using an incremental learning algorithm. This allows the output of fault types and their corresponding predicted probabilities. The model features low cross-brand transfer costs, good adaptability to small sample scenarios, and strong dynamic update capabilities. Attached Figure Description

[0017] Figure 1 The flowchart shows the cross-brand elevator fault prediction method based on transfer learning of this invention. Figure 2 This is a block diagram of the cross-brand elevator fault prediction system of the present invention. Detailed Implementation

[0018] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0019] This invention employs a deep transfer learning model to extract shared features from elevators of different brands, and uses a domain adaptation module to align the feature spaces of the source and target domains. Based on sample data, it combines incremental learning algorithms to achieve online updates of model parameters, thereby outputting fault types and predicted probabilities. It has the advantages of low cross-brand transfer costs, good adaptability to small sample scenarios, and strong dynamic update capabilities.

[0020] The embodiments of the present invention will be further described below with reference to several examples.

[0021] Example 1 like Figure 1 A cross-brand elevator fault prediction method based on transfer learning includes the following steps: Step 1: Collect elevator data from multiple brands to construct a multi-source elevator dataset, which includes sensor data and fault labels from elevators of multiple brands. Data source: Sensor data from multiple elevator brands were collected, including vibration sensors, current sensors, and door operator status sensors, to obtain vibration signals, current signals, and door status parameters. The data dimensions cover time-domain and frequency-domain characteristics such as peak value and mean value. Collect fault tags, including fault type, time of occurrence, and maintenance records.

[0022] Data preprocessing: Standardization: Z-score standardization is performed on the raw data of sensors from different brands (such as current in A / mA and vibration in g / m / s²) to eliminate dimensional differences; Denoising and Alignment: The Symlet wavelet basis threshold denoising algorithm is used to filter out high-frequency noise, and the multi-sensor data stream is aligned by timestamps; Feature enhancement: Perform continuous wavelet transform on the vibration signal to extract time-frequency domain energy features and construct a unified feature matrix across brands.

[0023] Step 2: Standardize the multi-source elevator dataset and use domain adaptation to align cross-brand features to minimize the data distribution differences between the source and target domains and obtain a standard dataset; Domain adaptation employs either maximum mean difference (MPD) or domain adversarial neural networks (NANNs) to optimize the adversarial training process between the domain classifier and the feature extractor through gradient inversion layers. The choice of domain adaptation depends on the data scenario: MPD is used for static feature distribution alignment; NAD is combined with gradient inversion layers for dynamic adversarial training. 1. Measurement of characteristic distribution differences: The difference in feature distributions between the source domain (Brand A) and the target domain (Brand B) is calculated, using the maximum mean difference (MMD) as the metric. The formula is as follows: ; Where ∅(⋅) is the kernel function (such as the Gaussian kernel). For source domain (brand A) data, Data for the target domain (Brand B).

[0024] 2. Domain-based adversarial training: Constructing Domain Adversarial Neural Networks (DANNs), including shared feature extractors Fault classifier Sum Domain Classifier ; Adversarial training mechanism: Feature extractor Objective: To generate domain-invariant features, so that Unable to distinguish between the source domain and the target domain; Domain classifier Objective: To maximize domain classification accuracy; During backpropagation, the gradient inversion layer (GRL) is used to... Inverting the gradient enables adversarial optimization.

[0025] 3. Hybrid loss function design: The total loss function is a weighted sum of the classification loss and the domain adversarial loss: ; Classification loss Cross-entropy loss optimizes fault prediction accuracy; Domain confrontation loss Binary cross-entropy loss forces the feature distributions of the source and target domains to be consistent. The hyperparameter λ is dynamically adjusted with each training round, with an initial value of 0.1 and gradually increasing to 1.0.

[0026] Step 3: Construct a deep transfer learning model, which includes a shared feature extraction layer and a domain adaptation module. The shared feature extraction layer extracts cross-brand common features from the standard dataset through a pre-trained convolutional neural network. The domain adaptation module aligns the source and target domain feature spaces of the standard dataset through an adversarial training strategy, thereby outputting the fault type and the corresponding predicted probability. The training process of the deep transfer learning model adopts a hybrid loss function, including fault classification loss and domain adversarial loss. The domain adversarial loss is used to force the feature distribution of the source domain to be consistent with that of the target domain. The hyperparameter λ of the hybrid loss function is dynamically adjusted with each training round. The model is dynamically optimized based on sample data, and the online updating of model parameters is achieved by combining incremental learning algorithms. This application constructs a multi-source elevator dataset and eliminates data distribution differences between brands through standardization and feature alignment techniques. It designs a fault prediction model based on deep transfer learning, utilizing a pre-trained convolutional neural network (CNN) to extract common features and combining a domain adversarial network (DANN) to align the feature spaces of the source and target domains. A dynamic few-shot optimization strategy is employed to achieve online model updates. In particular, a hybrid loss function (classification loss + domain adversarial loss) is introduced to enhance cross-brand generalization ability and support incremental learning to adapt to scenarios with incremental data. This invention significantly reduces data dependence in the target domain, is suitable for smart building elevator operation and maintenance scenarios, and has the advantages of high efficiency, low cost, and easy scalability.

[0027] like Figure 2 A cross-brand elevator fault prediction system based on transfer learning includes: The multi-protocol data acquisition module is used to collect elevator data from multiple brands and construct a multi-source elevator dataset, which includes sensor data and fault labels from elevators of multiple brands. The standardization processing module is used to standardize multi-source elevator datasets and perform cross-brand feature alignment using domain adaptation to minimize the data distribution differences between the source and target domains. The deep transfer learning model module is used to build a deep transfer learning model. The model includes a shared feature extraction layer and a domain adaptation module. The shared feature extraction layer extracts cross-brand common features from the standard dataset through a pre-trained convolutional neural network. The domain adaptation module aligns the source domain and target domain feature spaces of the standard dataset through an adversarial training strategy, thereby outputting the fault type and the corresponding predicted probability.

[0028] The dynamic optimization steps include: when the target brand data volume is less than 500, freezing the parameters of the shared feature extraction layer and updating only the domain adaptation module; when the target brand data volume increases to 500 or more, using a few-shot learning strategy to fine-tune the entire model.

[0029] 1. Small sample fine-tuning strategy: Initial stage (target data volume < 500 records): Freeze shared feature extraction layer The parameters are updated only in the domain adaptation module. and classifier ; Focal Loss is used to mitigate the class imbalance problem.

[0030] Incremental phase (target data volume ≥ 500 records): Unfreeze the shared layers and fine-tune the entire model with a lower learning rate (e.g., 1e-5); The Elastic Weighted Merging (EWC) algorithm preserves important parameters from the source domain and prevents catastrophic forgetting.

[0031] The transfer learning engine module supports edge computing deployment and enables real-time inference on local elevator devices through lightweight model compression technology.

[0032] 2. Edge-cloud collaborative deployment: At the edge: Deploy lightweight models to process sensor data in real time and generate fault probabilities; Cloud: Receives data from multiple edge nodes, periodically aggregates and updates the global model, and distributes it to the edge for synchronization.

[0033] The cloud aggregates edge data and updates the model every 24 hours, then distributes it to the edge via HTTPS.

[0034] Multi-protocol data acquisition module: Supports RS-485, CAN bus, and Modbus protocol parsing, converting data from different brands into a unified JSON format; Apache Kafka is used to implement high-concurrency data stream buffering to ensure real-time performance.

[0035] Transfer learning engine: Built-in MMD calculation and DANN training interfaces, supporting custom kernel functions; The lightweight model compression technique employs channel pruning and 8-bit quantization to compress the model volume to 30% of its original size.

[0036] Provides a PyTorch / TensorFlow model conversion tool, compatible with ONNX format for cross-platform deployment.

[0037] The fault early warning module is interconnected with the elevator operation and maintenance management platform, automatically generating work orders and assigning maintenance tasks based on the prediction results. The fault early warning module connects to the elevator operation and maintenance management platform via a REST API, supporting automatic work order generation and MQTT protocol push.

[0038] Fault warning module: The fault probability threshold is automatically adjusted based on the distribution of historical data; the dynamic threshold is automatically calculated based on the 95th percentile of the historical fault probability, and an early warning is triggered if the threshold is exceeded.

[0039] It interfaces with the operation and maintenance system API to push fault locations, suggested repair parts, and priorities.

[0040] An electronic device, comprising: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors perform the methods described above.

[0041] In the embodiments provided in this application, it should be understood that the disclosed methods and systems can also be implemented in other ways. The method and system embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0042] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0043] On the other hand, a computer-readable storage medium stores computer instructions thereon, which, when executed by a processor, implement the steps of the methods described above. When the computer program is executed by a processor, it implements the methods described in any of the first aspects above. If the functionality is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0044] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A cross-brand elevator fault prediction method based on transfer learning, characterized in that, Includes the following steps: S1. Collect elevator data from multiple brands to construct a multi-source elevator dataset, which includes sensor data and fault labels from multiple brands of elevators. S2. Standardize the multi-source elevator dataset and use domain adaptation to perform cross-brand feature alignment to minimize the data distribution difference between the source domain and the target domain, thereby obtaining a standard dataset. S3. Construct a deep transfer learning model, which includes a shared feature extraction layer and a domain adaptation module. The shared feature extraction layer extracts cross-brand common features of the standard dataset through a pre-trained convolutional neural network. The domain adaptation module aligns the source domain and target domain feature spaces of the standard dataset through an adversarial training strategy, thereby outputting the fault type and the corresponding prediction probability.

2. The method according to claim 1, characterized in that, The sensor data includes vibration signals, current signals, and door status parameters; the fault label includes the fault type, occurrence time, and maintenance record.

3. The method according to claim 1, characterized in that, The domain adaptation employs either maximum mean difference metric or domain adversarial neural network, and optimizes the adversarial training process between the domain classifier and the feature extractor through gradient inversion layer; the domain adaptation is selected based on the data scenario: maximum mean difference is used for static feature distribution alignment. Domain adversarial neural networks combined with gradient inversion layers are used for dynamic adversarial training.

4. The method according to claim 1, characterized in that, The training process of the deep transfer learning model adopts a hybrid loss function, including fault classification loss and domain adversarial loss. The domain adversarial loss is used to force the feature distribution of the source domain to be consistent with that of the target domain. The hyperparameter λ of the hybrid loss function is dynamically adjusted with each training round; the model is dynamically optimized based on sample data, and the model parameters are updated online by combining an incremental learning algorithm.

5. The method according to claim 4, characterized in that, The dynamic optimization steps include: when the amount of target brand data is less than n, freezing the parameters of the shared feature extraction layer and updating only the domain adaptation module; when the amount of target brand data increases to n or more, using a few-shot learning strategy to fine-tune the entire model.

6. A cross-brand elevator fault prediction system based on transfer learning, characterized in that, include: The multi-protocol data acquisition module is used to collect elevator data from multiple brands and construct a multi-source elevator dataset, which includes sensor data and fault labels from elevators of multiple brands. The standardization processing module is used to standardize the multi-source elevator dataset and perform cross-brand feature alignment using domain adaptation to minimize the data distribution differences between the source domain and the target domain. The deep transfer learning model module is used to construct a deep transfer learning model. The model includes a shared feature extraction layer and a domain adaptation module. The shared feature extraction layer extracts cross-brand common features of the standard dataset through a pre-trained convolutional neural network. The domain adaptation module aligns the source domain and target domain feature spaces of the standard dataset through an adversarial training strategy, thereby outputting the fault type and the corresponding predicted probability.

7. The system according to claim 6, characterized in that, The transfer learning engine module supports edge computing deployment and enables real-time inference on local elevator devices through lightweight model compression technology.

8. The system according to claim 6, characterized in that, The fault early warning module is interconnected with the elevator operation and maintenance management platform. It automatically generates work orders and assigns maintenance tasks based on the prediction results. The fault early warning module is interconnected with the elevator operation and maintenance management platform through REST API, which supports automatic work order generation and MQTT protocol push.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-5.

10. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, this instruction implements the steps of the method as described in any one of claims 1-5.

Images