Method and device for predicting service life of gearbox

Abstract

The invention discloses a gear box service life prediction method and device, and belongs to the technical field of mechanical performance prediction. By synchronously collecting vibration, oil liquid, temperature and other multi-dimensional data and performing fusion analysis, the limitation of single signal monitoring is broken through, and the operation state of the gearbox is comprehensively mastered; a deep learning and statistical analysis hybrid model is adopted, and a federated learning framework is combined, so that the accuracy of state evaluation and life prediction is remarkably improved while the data security is guaranteed; due to the integrated design and the dynamic optimization display function of the device, the practicability of technology landing is enhanced, and powerful support is provided for health management of the gearbox in an industrial scene.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical performance prediction, and in particular relates to a method and device for predicting the life of a gearbox. Background Art

[0002] As a core transmission component, wind turbine gearboxes operate in complex and harsh environments for extended periods of time. Their remaining lifespan is directly related to the unit's operational safety and maintenance costs. Accidental failure due to inaccurate lifespan predictions can result in significant economic losses.

[0003] Current gearbox life prediction technology has obvious shortcomings: First, most solutions rely on single-modal data (such as only vibration signals or oil data), but gearbox life degradation is the result of the combined action of multiple factors such as mechanical wear, oil degradation, and thermal damage. Single-modal data can only reflect local degradation characteristics - vibration signals can reflect mechanical impact but are difficult to reflect oil aging, and oil data can reflect the degree of wear but cannot quantify the impact of thermal fatigue, resulting in the inability to fully capture the life degradation law; second, the lack of a confidence-based multimodal fusion mechanism makes it difficult to effectively integrate features of different dimensions and achieve information complementarity, affecting the quality of basic data for life prediction; third, existing models mostly use a single algorithm, and do not introduce federated learning for distributed analysis, and do not dynamically determine weight aggregation parameters according to the proportion of data volume, resulting in a rough characterization of the health indicator degradation curve, low credibility of the probability interval of the remaining life prediction, and the error often exceeds 15%, which cannot meet the engineering requirements for life prediction accuracy.

[0004] Therefore, there is an urgent need for a gearbox life prediction method that can integrate multi-dimensional information and improve prediction accuracy. Summary of the Invention

[0005] The purpose of the present invention is to provide a gearbox life prediction method and device to solve the problems of low life prediction accuracy and poor interval reliability caused by the existing technology relying on single modal data, lack of fusion mechanism, and insufficient model design.

[0006] Existing technologies rely solely on single modal data for life prediction, which cannot fully reflect the degradation process of gearboxes with multi-factor coupling, resulting in large prediction errors and unreliable intervals, making it difficult to support accurate maintenance decisions.

[0007] To this end, the present invention achieves high-precision life prediction through multimodal data acquisition, adaptive fusion and hybrid model analysis: synchronously collect vibration signals, oil data, and temperature image data, and combine speed and load operating condition parameters as auxiliary features to capture degradation characteristics related to life from multiple dimensions; after preprocessing based on wavelet threshold denoising and sliding window normalization, each modal feature is extracted through CEEMDAN decomposition, ResNet-18, etc., and then a high-dimensional feature vector is obtained through confidence-based multimodal adaptive fusion to enhance information complementarity; an LSTM-Transformer hybrid network is constructed and physical constraints are embedded, and the federated learning framework is combined to determine the weight according to the data volume ratio for parameter aggregation, and finally the health indicator degradation curve and the 95% probability interval of the remaining life are output through edge computing, and the EWC algorithm is used to optimize the model stability to ensure that the prediction error does not exceed 5%, providing an accurate basis for gearbox life management.

[0008] To achieve the above object, the present invention adopts the following technical solution: a method for predicting the life of a gearbox, the method specifically comprising the following steps:

[0009] Multi-dimensional data acquisition: synchronously collects the vibration signal, oil data, and temperature image data of the gearbox, and records the speed and load operating parameters. Among them, the vibration signal is collected by a vibration sensor installed on the bearing seat near the meshing area (sampling frequency ≥ 25.6kHz), the oil data is collected by an oil sensor integrated into the oil return port of the oil circulation system (wear particle identification accuracy ≥ 5μm), and the temperature image data is collected by an infrared thermal imager deployed at the observation window facing the meshing surface (resolution ≥ 640×512 pixels); the speed and load parameters are collected by a photoelectric encoder and a torque sensor respectively, and the sampling frequency is synchronized with the vibration signal (frequency consistency ≤ ±10μs). This multi-dimensional monitoring method, combined with operating parameters, can accurately correlate state changes with operating conditions, overcome the limitations of single-dimensional monitoring, and thus more comprehensively and accurately grasp the gearbox status, timely capture various potential anomalies to avoid missing hidden faults, and provide comprehensive and accurate basic information for subsequent fault diagnosis and early warning.

[0010] Data preprocessing and feature fusion: Preprocess the above signals and data and extract features, then perform confidence-based multimodal adaptive fusion to obtain a high-dimensional feature vector (i.e., a 256-dimensional vector consisting of 128-dimensional time-frequency features extracted from vibration signals, 64-dimensional features extracted from oil data, and 64-dimensional features extracted from temperature image data). Specifically, this includes:

[0011] For preprocessing, adaptive noise elimination based on wavelet threshold and time-domain synchronous averaging are used to filter out interference on the vibration signal; sliding window normalization is performed on the oil data to unify the scale; finite element meshing is used to generate a thermal stress matrix for the temperature image data to quantify the temperature distribution.

[0012] Feature extraction: 128-dimensional time-frequency features of the vibration signal are extracted through CEEMDAN decomposition and 1D-CNN. ResNet-18 is used to classify the oil wear particle image morphology, and the dimension is reduced to 64 dimensions through PCA. Based on the graph neural network, the temperature field heat conduction relationship feature is constructed and regularized into a 64-dimensional vector.

[0013] Fusion processing, performs adaptive weighted fusion based on the confidence of each modal data to form a 256-dimensional comprehensive feature vector.

[0014] Targeted preprocessing effectively improves data quality, providing a solid and reliable foundation for subsequent analysis and avoiding interference of poor-quality data on the results. High-quality data, as a prerequisite for accurate analysis of the gearbox status, can ensure that subsequent feature extraction, model prediction and other links are based on reliable data, thereby improving the accuracy and credibility of the entire monitoring and early warning system. At the same time, the integration of multiple advanced algorithms to extract features can fully retain key information, carefully characterize the gearbox status, help to accurately identify the type and severity of faults, and provide a strong basis for formulating targeted maintenance plans.

[0015] Model deployment and collaborative analysis: deploy a hybrid model of deep learning and statistical analysis, combine it with the federated learning framework, and determine the weights according to the data volume of each gearbox to perform parameter aggregation analysis. Specifically including:

[0016] A hybrid model is constructed using a 2-layer LSTM (64 hidden units) to capture the timing degradation trend. The preset value of the number of LSTM hidden units is determined (such as "adaptive setting based on the amount of gearbox operating data"), and a Transformer encoder with an 8-head attention mechanism is used to extract long-range feature associations. The Paris crack propagation physical constraint is embedded in the loss function (its core formula is dN / da=C(ΔK)^m, where dN / da is the crack propagation rate, C and m are material constants, and ΔK is the stress intensity factor range). By introducing this physical law into the loss function, the fatigue crack propagation process of key components of the gearbox (such as gear tooth roots, bearing raceways, etc.) under cyclic loads can be quantified, so that the model prediction results not only conform to the data laws, but also fit the physical mechanism of mechanical damage evolution, thereby correcting the possible deviations of the purely data-driven model and improving the rationality and engineering applicability of the remaining life prediction.

[0017] Federated learning applications aggregate parameters every 24 hours and assign weights based on the proportion of data volume on each node, thus achieving privacy protection and collaborative optimization of distributed data.

[0018] The LSTM-Transformer hybrid network combined with physical constraints significantly improves the accuracy and rationality of remaining useful life (RUL) predictions, providing strong support for maintenance decisions and enabling enterprises to reasonably arrange maintenance time, avoid resource waste from excessive maintenance and sudden failures caused by untimely maintenance, thereby reducing maintenance costs and improving equipment utilization efficiency; at the same time, the federated learning framework, while strictly protecting data privacy, makes full use of distributed data optimization models, significantly improving the model's generalization ability, enabling it to maintain good performance in different industrial scenarios and equipment conditions, while taking into account data security and collaborative analysis effects.

[0019] Results presentation and dynamic optimization: Displays gearbox health indicator degradation curves and 95% probability intervals for remaining life. Edge computing nodes output RUL prediction results in real time. Elastic Weight Consolidation (EWC) algorithms are used for incremental learning to optimize model stability, ensuring the relative error between predicted and actual RUL does not exceed 5%. Display content is dynamically updated.

[0020] The combination of edge computing and incremental learning ensures the real-time and long-term accuracy of the results. The dynamic display function facilitates the formulation of scientific maintenance plans in advance to reduce sudden failures, enabling enterprises to promptly grasp the operating status of the gearbox and arrange maintenance work in advance, avoiding production interruptions caused by sudden failures, ensuring production continuity and stability, and improving production efficiency.

[0021] The present invention provides a gearbox life prediction device, which includes the following modules:

[0022] The data acquisition module integrates vibration sensors, oil sensors, infrared thermal imagers, photoelectric encoders and torque sensors to synchronously collect vibration signals, oil data, temperature image data, and speed and load operating parameters. Each sensor is deployed at the corresponding position of the gearbox according to actual conditions. The synchronous collection of multi-dimensional data provides the basis for comprehensive monitoring and ensures the integrity of state perception.

[0023] Data preprocessing and feature extraction module: Performs preprocessing such as adaptive noise elimination, sliding window normalization, and finite element meshing on the collected data. Multimodal features are extracted through technologies such as CEEMDAN+1D-CNN, ResNet-18+PCA, and graph neural networks. 256-dimensional vectors are generated through confidence-based adaptive fusion. Through precise data processing and feature fusion, high-quality input is provided for model analysis, ensuring the accuracy of subsequent predictions.

[0024] Model deployment and analysis module: Deploys the LSTM-Transformer hybrid model and federated learning framework, performs time series analysis and parameter aggregation of feature vectors, and outputs health status assessment results. Through advanced models and collaborative learning mechanisms, it achieves high-precision status assessment and life prediction, providing a scientific basis for maintenance decisions.

[0025] Result display module: Real-time display of health indicator degradation curve, remaining life 95% probability interval and RUL prediction value, through the EWC algorithm to achieve model incremental learning, dynamic update display content, through real-time and dynamic result presentation, so that staff can intuitively understand the gearbox status and respond to potential risks in a timely manner.

[0026] The aforementioned data preprocessing and feature extraction modules, model deployment and analysis modules, and result presentation modules are all integrated into the industrial computer, enabling data transmission and interaction via an internal industrial-grade data bus. This integrated design enhances practicality, making it easier to implement in various industrial scenarios and simplifying deployment. It also reduces connection complexity to minimize transmission interference and latency, improves device stability and responsiveness, and ensures efficient integration throughout the entire process.

[0027] Comparison of the advantages of the present invention with the prior art:

[0028] 1) Synchronously collect vibration, oil, and temperature image data and combine them with speed and load parameters to break through the limitations of a single mode and comprehensively capture the life attenuation characteristics of multi-factor coupling, solving the one-sidedness problem of single data and providing a complete information basis for prediction.

[0029] 2) Targeted preprocessing (wavelet denoising, finite element meshing, etc.) and feature extraction (CEEMDAN+1D-CNN, ResNet-18 classification, etc.), combined with confidence-based multimodal fusion, improves data quality and representation accuracy, and makes up for the shortcomings of the existing technology fusion mechanism.

[0030] 3) Build an LSTM-Transformer hybrid model and embed the Paris crack extension constraint. Combined with federated learning, weights are assigned according to data volume (aggregated once every 24 hours), taking into account the advantages of data-driven and mechanism analysis, improving the accuracy of degradation curve characterization and prediction while ensuring data privacy.

[0031] 4) Through real-time output of results through edge computing and incremental learning of the EWC algorithm, the remaining life prediction error is ensured to be ≤5%, and the reliability of the 95% probability interval is significantly improved, solving the problem of the existing technology error exceeding 15% and meeting the engineering accuracy requirements.

[0032] 5) The gearbox life prediction device integrates hardware such as vibration sensors (sampling ≥ 25.6kHz) and oil sensors (accuracy ≥ 5μm). The modules interact through the industrial-grade data bus inside the industrial computer, adapting to the on-site environment, facilitating deployment and real-time analysis, and having strong practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Schematic diagram of the gearbox life prediction system architecture provided by this embodiment;

[0034] Figure 2 This is a schematic diagram of the gearbox life prediction process provided by this embodiment;

[0035] Figure 3 Schematic diagram of the gearbox life prediction device provided in this embodiment. DETAILED DESCRIPTION

[0036] The following embodiments of the present invention are described in further detail with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0037] like Figure 1-Figure 3 As shown, a gearbox life prediction method utilizes a "multi-dimensional perception + intelligent feature analysis + dynamic early warning" technology approach. This method integrates information from multiple parameters, such as vibration, temperature, and oil, and combines the advantages of traditional methods with intelligent models to achieve full coverage from early fault identification to precise early warning. During implementation, the method focused on sensor selection and the compatibility of gearbox speed and power. It also focused on the targeted extraction of fault-sensitive features and ensured the early warning model was robust against operating condition fluctuations and noise, ultimately achieving the goal of "early detection, early diagnosis, and early maintenance."

[0038] Hardware deployment; vibration signal acquisition: A PCB352C33 triaxial accelerometer was selected and installed on the high-speed shaft bearing seat (near the gear meshing area). The parameters were set to a sampling frequency of 25.6 kHz, which meets the acquisition requirements of the gear meshing frequency (usually 5-10 kHz) and the bearing fault characteristic frequency (up to 15 kHz).

[0039] Oil data collection: The "SpectroScientificFluidScanQ1000" online oil monitor is installed on the lubricating oil return line; the parameters are set to an abrasive particle identification accuracy of 5μm, capable of detecting the concentrations of wear elements such as Fe (iron), Cu (copper), and Si (silicon), with a sampling period of 10 minutes.

[0040] Temperature image acquisition: A FLIRA655sc infrared thermal imager was selected and installed in the observation window, aligned with the gear meshing surface. The parameters were set to a resolution of 640 × 512 pixels, a frame rate of 60 Hz, and a temperature measurement range of -20°C to 150°C.

[0041] Working condition parameter collection: Speed ​​measurement uses a photoelectric encoder (resolution 1024 pulses / rev); load measurement uses a strain gauge torque sensor (range 0-5000Nm, accuracy ±0.1%); the synchronization mechanism uses the PTPv2 (IEEE1588-2008) precise time protocol to achieve sensor clock synchronization, and through a combination of hardware triggering and software calibration, ensure that the sampling time error of vibration signals, oil data, temperature images and working condition parameters is ≤±8μs, meeting the strict requirement of "frequency consistency ≤±10μs" in claim 2.

[0042] Data preprocessing process; vibration signal preprocessing: When performing adaptive noise elimination, the db6 wavelet is used for 5-layer decomposition, and the environmental noise and electrical interference are removed through the soft threshold noise reduction algorithm. The threshold calculation formula is: λ = 0.6745·median(|c_Dj|)·√(2log(n)), where cDj is the wavelet decomposition detail coefficient and n is the signal length; based on the gear rotation period (typical value: 50 periods), the vibration signal is synchronized and averaged in the time domain to highlight the gear meshing characteristics and suppress random noise; the time-frequency matrix is ​​generated through short-time Fourier transform (STFT), and the time-frequency resolution is set to 10ms×10Hz, taking into account the feature expression in both time and frequency domains.

[0043] Oil data preprocessing: The spectral data is standardized using a 7-day sliding window (a total of 1008 sampling points) to eliminate the influence of long-term sensor drift. The calculation formula is:

[0044] xnorm = (x-mean(W)) / std(W), where W is the data in the current sliding window. Grayscale normalization, size cropping (unified to 224 × 224 pixels), and Gaussian filtering (σ = 1.0) are performed on the wear particle image to enhance the wear particle edge features.

[0045] Temperature image preprocessing: The 640 × 512 pixel thermal image was divided into a finite element mesh with 5 mm × 5 mm cells, generating approximately 1300 mesh nodes. Based on the Fourier heat conduction equation, the temperature distribution T(x, y) was converted into a thermal stress matrix σ(x, y): σ(x, y) = 1-ναE·(T(x, y)-T0), where α is the thermal expansion coefficient, E is the elastic modulus, ν is the Poisson's ratio, and T0 is the reference temperature.

[0046] Feature extraction and multimodal fusion; vibration feature extraction: CEEMDAN decomposition is used to decompose the time-frequency matrix into 8 intrinsic mode function (IMF) components, the noise standard deviation is set to 0.2, the number of iterations is 100, and the first 3 main IMF components are retained; the time-frequency features are extracted through the 1D-CNN network, and the network structure is:

[0047] Conv1 D(64,k=5)→ReLU→MaxPooling(2)→Conv1D(64,k=3)→Flatten, and finally outputs a 128-dimensional time-frequency feature vector.

[0048] Oil feature extraction: A pre-trained ResNet-18 model was used to extract features from the wear particle images, intercepting the output of the penultimate fully connected layer (1024 dimensions). Principal component analysis was used to reduce the 1024-dimensional features to 64 dimensions, retaining 95% of the variance of the original data.

[0049] Temperature feature extraction: Construct a temperature field graph G = (V, E) for graph neural network modeling, where the node V is the grid unit and the weight of the edge E is the heat conduction coefficient. The heat conduction relationship features are extracted through the GCNConv layer; the feature vector output by the GNN is mapped to a 64-dimensional temperature field feature vector through a fully connected layer.

[0050] Multimodal fusion uses a confidence-based adaptive weighted fusion mechanism, calculated as: Ffusion = w1·Fvibration + w2·Foil + w3·Ftemperature. Weights w1 = 0.5 (vibration), w2 = 0.3 (oil), and w3 = 0.2 (temperature). The weights are dynamically adjusted based on the sensitivity of each modality at different fault stages. The final output is a 256-dimensional fusion feature vector (128 vibration, 64 oil, and 64 temperature).

[0051] LSTM-Transformer hybrid model training; model architecture design: The LSTM part adopts a two-layer stacked LSTM structure, with 64 hidden units in each layer and an input sequence length of 720 (corresponding to 5 days of data), which is used to capture short-term temporal dependencies such as vibration signals; the Transformer part adopts an 8-head attention mechanism with a hidden dimension of 128, and captures long-term degradation trends such as oil wear accumulation through multi-head self-attention.

[0052] Physical constraints: The Paris crack growth formula is embedded in the loss function: da / dN = C(ΔK)^m (where C = 1.2×10 -9, m = 3.2, a gear material constant), and ΔK is the stress intensity factor range. The physical constraint term is calculated as: L_phys = λ·|(da / dN)_pred - (a_pred - a_prev) / ΔN| (supplementing the absolute value sign and the cycle number difference ΔN to quantify the prediction deviation), where λ = 0.1 is the constraint weight, apred and aprev are the crack prediction values ​​at the current and previous moments, respectively.

[0053] Training configuration: The dataset consists of two years of operating data from five gearboxes (87,600 samples, sampled every 10 minutes). These five gearboxes are all 1.5MW doubly-fed wind turbine gearboxes, operating in a wind farm with a temperate continental climate, at an altitude of approximately 800 meters, with an average annual temperature of 10-15°C, wind speeds ranging from 3-25 m / s, and loads ranging from 30% to 100% of rated power. The model operates for approximately 20 hours per day, during which it experiences various operating conditions, including normal operation, gusty winds, and low-temperature startups. The optimizer uses Adam, an initial learning rate of 1e-4, and cosine annealing learning rate scheduling. The loss function is: L = LMSE + 0.1·Lphys.

[0054] System verification data: The data collection conditions for "15 identical 1.5MW doubly-fed wind turbine gearboxes at a wind farm" are as follows: The wind farm is located in a coastal area with a maritime climate, approximately 50 meters above sea level, with an average annual temperature of 15-20°C, high humidity, wind speeds ranging from 4-30 m / s, and loads ranging from 20% to 100% of the rated power. During operation, the gearboxes are exposed to environmental factors such as salt spray corrosion and high humidity, and have experienced operation in extreme weather conditions such as typhoons.

[0055] The training strategy is: batch size 32, training for 100 epochs, and early stopping strategy (training is stopped if the validation set loss does not decrease for 5 consecutive epochs).

[0056] Federated learning and incremental learning deployment; Federated learning architecture: The central node is deployed on an AWS EC2p3.2xlarge server (GPU: NVIDIA V100, 32GB video memory); the edge nodes are 10 NVIDIA Jetson AGX Xavier devices (64-core Cortex-A72 CPU, 512-core Volta GPU) deployed locally in the wind farm.

[0057] Federated Training Process: For data partitioning, each edge node manages the data of 1 gearbox, with approximately 8,760 samples per device (data volume accounts for 10%); the edge node trains the model for 10 Epochs based on local data, using the Adam optimizer (learning rate 1e-5); every 24 hours, each edge node uploads the model parameters to the central node, and the central node performs weighted averaging according to the data volume ratio (10%): W_global = ∑(N_i / N_total)·W_i (i = 1 to 10), and the central node distributes the global model to each edge node, and the edge node replaces the local parameters with the global parameters.

[0058] Incremental Learning Mechanism: When the new data accumulates to 100 groups and the prediction error > 5%, the Elastic Weight Consolidation (EWC) algorithm is triggered: Calculate the Fisher information matrix:

[0059] Add a regularization term to the loss function:

[0060] L_EWC = L(θ)+(λ / 2)∑F_i(θ_i - θ_i^*) 2 , where θ_i^* is the old model parameter, λ = 1000 is the regularization strength, and the relative error is ensured to be ≤ 5% by fine-tuning the model parameters.

[0061] Health Index and Remaining Useful Life (RUL) Prediction; Prediction Output Design: The health index (HI) is calculated and defined as HI ∈ [0, 1], and the model output is mapped to the [0, 1] interval through the Sigmoid function; the stages are divided as follows: HI > 0.9 is the normal period (green warning), 0.6 < HI ≤ 0.9 is the slow degradation period (yellow warning), and HI ≤ 0.6 is the accelerated degradation period (red warning).

[0062] The remaining useful life (RUL) prediction adopts the Monte Carlo Dropout method (p = 0.2), and the prediction distribution is generated by repeating the inference 100 times; calculate the 95% confidence interval: [RUL2.5%, RUL97.5%].

[0063] Dynamic Update Mechanism: The data update triggers an incremental update of the model once every 100 newly added data groups (about 1 week); the warning trigger condition is that when HI ≤ 0.6 or RUL ≤ 30 days, the system automatically pushes warning information to the operation and maintenance terminal, including the warning level (yellow / red), predicted RUL and its confidence interval, and visualization of key fault features (vibration spectrum, oil particle image, etc.).

[0064] System Verification and Performance Evaluation; Test Environment: The verification object is the gearboxes of 15 1.5MW doubly-fed wind turbines of the same model in a wind farm; the test period is from January 2023 to June 2024 (a total of 18 months).

[0065] Evaluation indicators: In terms of prediction accuracy, the average relative error of RUL is 3.8% (the error of the traditional single-modal method is 15.2%); the fault warning accuracy is 98.7% (15,245 days in advance warning); the correlation coefficient between the health index (HI) curve and the actual degradation trend is 0.96; in terms of generalization ability, the cross-device prediction variance is reduced from 5.7 days (non-federated learning) to 2.1 days (after federated learning optimization); in terms of real-time performance, the single sample processing time at the edge node is <200ms, meeting the online monitoring requirements; in terms of economic benefits, compared with the period from January 2021 to June 2022 before the wind farm adopted the present invention, after the implementation of the present invention, the number of unplanned shutdowns was reduced by 85% and the maintenance cost was reduced by 42%.

[0066] As demonstrated in the preceding examples, the present invention utilizes multimodal fusion, physical constraints, and federated learning to construct a comprehensive, closed-loop management system, from data acquisition to intelligent early warning. In practical applications, this system significantly improves the accuracy of gearbox life prediction and its cross-device adaptability, fully encompassing the technical features enumerated in the patent claims and providing reliable technical support for the full lifecycle management of wind turbine gearboxes.

[0067] like Figure 3 The figure shows a gearbox life prediction method and device. This method uses wind farm gearboxes as monitoring targets. Through the complete process of "hardware configuration → data acquisition → processing and analysis → early warning output," it accurately monitors gearbox status and predicts remaining useful life (RUL). The device parameters, algorithm parameters, and experimental data involved are for illustrative purposes only and can be adjusted according to actual scenarios.

[0068] In order to achieve the synchronous collection, real-time processing and intelligent analysis of multimodal data, it is necessary to build an "edge-cloud" collaborative architecture. The specific hardware configuration is as follows: The following, combined with the layered architecture diagram, details the system architecture and core functions of each layer of the present invention, further clarifies the full process logic of "data collection → edge processing → cloud analysis → feedback optimization", and integrates the federated learning architecture to improve the feedback optimization mechanism.

[0069] Overview of the layered architecture. The system adopts a layered architecture, which is divided into data collection layer, edge computing layer, cloud analysis layer, and feedback optimization layer from top to bottom. Each layer achieves collaborative linkage through data interfaces and communication protocols to ensure the efficient flow and intelligent analysis of multimodal data.

[0070] Data acquisition layer: Synchronous perception of multi-source signals. The data acquisition layer is the "perception terminal" of the system. It realizes real-time collection of key status parameters of the gearbox by deploying multiple types of sensors. The core components are as follows: 3D schematic diagram of the gearbox: marking the sensor installation position, including the high-speed shaft bearing seat, lubricating oil pipeline, and gear meshing surface, and intuitively displaying the distribution of monitoring points.

[0071] Sensor deployment and signal visualization; High-speed shaft bearing housing: Install a PCB352C33 triaxial accelerometer (sampling frequency 25.6kHz) to collect vibration signals and simultaneously display the time domain waveform (reflecting the change in signal amplitude over time) and FFT spectrum (highlighting key frequency bands such as gear meshing frequency and bearing fault characteristic frequency).

[0072] Lubricating oil pipeline: Integrates the SpectroScientific FluidScan Q1000 online oil monitor (wear particle identification accuracy 5μm) and simultaneously collects oil data: the upper part is a spectral analysis histogram (displaying the concentration of elements such as Fe, Cu, and Si, in ppm); the lower part is a microscopic image of ferrography wear particles (marking the morphological characteristics of cutting and fatigue wear particles).

[0073] Gear meshing surface: A FLIRA655sc infrared thermal imager (resolution 640×512 pixels) was used to collect temperature image data. The surface temperature distribution of the gearbox was displayed in the form of a thermal image, with emphasis on marking high-temperature areas (e.g., areas above 35°C) such as the tooth surface contact area.

[0074] Working condition parameter collection: The working condition parameters are recorded synchronously through the photoelectric encoder (speed) and the torque sensor (load). The sampling frequency is consistent with the vibration signal, and the time error is ≤±10μs.

[0075] Synchronization mechanism: Time synchronization of all sensors is achieved through the PTPv2 protocol, ensuring that the sampling time error of vibration signals, oil data, temperature data and operating parameters is ≤±10μs, providing the time consistency basis for subsequent multimodal fusion, corresponding to "frequency consistency ≤±10μs".

[0076] Edge computing layer: real-time preprocessing and feature fusion. The edge computing layer is deployed on the NVIDIA Jetson AGX Xavier embedded device and is responsible for real-time data preprocessing and feature extraction, reducing cloud transmission pressure. The core modules are as follows:

[0077] Preprocessing module; vibration signal processing chain: original vibration signal → adaptive noise elimination based on wavelet threshold (filter order 128) → time domain synchronous averaging (number of cycles 50) → time-frequency analysis (generating time-frequency matrix through STFT) to filter out environmental interference and highlight the effective signal within the gear meshing cycle.

[0078] Oil data processing chain: Spectral concentration data → 7-day sliding window normalization (to eliminate detection bias); wear particle image → grayscale normalization → size cropping (224×224 pixels) → morphological classification preprocessing, and sliding window normalization operation is performed on the oil data.

[0079] Temperature data processing chain: original thermal image → finite element meshing (unit size 5mm) → thermal stress matrix mapping, converting the temperature distribution into quantifiable matrix data, which complies with the preprocessing method of "generating a thermal stress matrix by using finite element meshing for temperature image data" in claim 3.

[0080] Feature extraction module; uses parallel channels to process multimodal data to improve feature extraction efficiency: the vibration channel inputs the time-frequency matrix, and is decomposed by CEEMDAN (noise standard deviation 0.2, 100 iterations) to obtain 8 IMF components, which are then

[0081] 1D-CNN (Conv1 D(64,k=5)→ReLU→MaxPooling→Flatten) extracts 128-dimensional time-frequency feature vectors.

[0082] Oil channel: Input the wear particle image, classify the morphology through the ResNet-18 model, extract the 1024-dimensional deep features, and reduce the dimension to 64-dimensional feature vector through PCA (retaining 95% variance), which corresponds to the step of "ResNet-18 classification of oil wear particle morphology, and PCA reduction to 64 dimensions" in claim 4.

[0083] In the temperature channel, the thermal stress matrix is ​​input, and the relationship characteristics of the temperature field and the thermal conduction of the structural parts are constructed based on the graph neural network (GNN). After dimension regularization, a 64-dimensional feature vector is obtained.

[0084] Cross-modal attention fusion performs adaptive weighted fusion based on the confidence of each modal data, with vibration features as query, oil features as key, and temperature features as value. The weight is calculated through the attention mechanism: Attention(Q,K,V)=Softmax(QK T / √d_k)V, where dk=64 (key dimension), and finally outputs a 256-dimensional fusion feature vector (128 vibration + 64 oil + 64 temperature), realizing the complementary enhancement of multimodal information and the technical feature of "confidence-based multimodal adaptive fusion to obtain a 256-dimensional vector".

[0085] Cloud analytics layer: Deep modeling and lifespan prediction. The cloud analytics layer is deployed on AWS EC2p3.2xlarge servers and uses deep learning models to accurately predict health indicators and remaining lifespan (RUL). Key components include the following: LSTM-Transformer hybrid model:

[0086] The left half is the expanded diagram of the LSTM unit (3 time steps), which captures short-term temporal dependencies such as vibration signals through a 2-layer LSTM (64 hidden units).

[0087] The right part is a Transformer encoder (8-head attention mechanism), which extracts the equal-length cycle features of oil particle accumulation, enhancing the model's ability to capture long-term degradation trends.

[0088] After the output features of the two are concatenated, they are input into a fully connected layer to obtain a preliminary prediction result, which is consistent with the model structure in claim 5 where "the LSTM has 2 layers and 64 hidden units, and the Transformer encoder uses an 8-head attention mechanism".

[0089] Physical constraint module, embedding the Paris crack growth formula: dN / da = C(ΔK)^m in the model loss function, where C = 1.2e-9 and m = 3.2 (gear material parameters). Using the physical law of gear crack growth as a constraint to correct the model prediction deviation and make the output result more in line with engineering practice.

[0090] Life prediction output, calculating the 95% probability interval of the remaining useful life (RUL) through the Monte Carlo Dropout method (repeating the inference 100 times), and synchronously outputting the degradation curve of the health index (HI): The value range of HI is 0-1, where HI > 0.9 is the normal period, 0.6 < HI ≤ 0.9 is the slow degradation period, and HI ≤ 0.6 is the accelerated degradation period. The prediction results are displayed in the form of a visualization chart (the horizontal axis is the running time, and the vertical axis is the HI value and the RUL prediction value), supporting historical data backtracking, meeting the requirement of "displaying the degradation curve of the gearbox health index and the 95% probability interval of the remaining useful life (RUL)".

[0091] Feedback optimization layer: Incremental learning and federated collaboration. The feedback optimization layer continuously optimizes the model performance through incremental learning and federated learning mechanisms to adapt to the heterogeneous characteristics of different gearboxes. The core modules are as follows:

[0092] Incremental learning mechanism: Flowchart, new collected data input → calculating the error (relative error) with the model prediction result → if the error > 5% or the new data volume > 100 groups → triggering the Elastic Weight Consolidation (EWC) algorithm → updating the model parameters while retaining historical knowledge → outputting the optimized model.

[0093] This mechanism ensures that the model can still maintain stable prediction accuracy when the gearbox working conditions change, avoiding "catastrophic forgetting".

[0094] Federated learning architecture: Architecture diagram, with the cloud server as the central node, connecting 10 wind farm edge nodes (each node corresponds to 1 fan gearbox), forming a "central-edge" collaborative network.

[0095] Operation process: Each edge node trains a model based on local data → uploads model parameters (not original data) to the cloud every 24 hours → the cloud allocates aggregation weights according to the proportion of node data volume (for example, node A accounts for 30% of the data volume and has a weight of 0.3) → aggregates parameters through the formula: W_global = ∑(N_i / N_total)·W_i (i=1to10) → sends global model parameters to each edge node → each node updates its local model, which is consistent with the federated learning mechanism of "aggregating parameters once every 24 hours and allocating weights according to the amount of node data".

[0096] Early warning feedback closed loop: When the health indicator HI ≤ 0.6 or the remaining life (RUL) enters the early warning threshold (such as ≤ 30 days), the system automatically pushes the early warning information to the operation and maintenance terminal, and at the same time incorporates the actual fault data fed back by the operation and maintenance into incremental learning, forming a closed-loop mechanism of "prediction-early warning-operation and maintenance-optimization". The RUL prediction results are output in real time through the edge computing nodes and the display content is dynamically updated. It has the characteristics of "real-time output of RUL prediction results through edge computing nodes" and "dynamic update of display content".

[0097] Through the above-mentioned layered architecture, the system realizes the real-time collection of multimodal data of gearboxes, rapid processing at the edge, in-depth analysis and dynamic optimization in the cloud. Combined with federated learning and physical constraints, it significantly improves the accuracy of health monitoring and cross-device adaptability, providing reliable technical support for the full life cycle management of gearboxes.

Claims

1. A gearbox life prediction method, characterized in that: The method comprises: Synchronously collecting vibration signals, oil data, and temperature image data of the gearbox; The above signals and data are preprocessed and features are extracted, and high-dimensional feature vectors are obtained through multi-modal adaptive fusion based on confidence; A hybrid model of deep learning and statistical analysis was deployed, combined with a federated learning framework. Parameter aggregation analysis was performed based on the weights determined by the data volume of each gearbox to display the gearbox health indicator degradation curve and the 95% probability interval for the remaining life.

2. The method according to claim 1, characterized in that The method further comprises: Record the gearbox speed and load condition parameters and use them as auxiliary features in fusion analysis; The speed and load operating parameters are collected through a photoelectric encoder and a torque sensor. The sampling frequency is synchronized with the vibration signal, and the frequency consistency is ≤±10μs.

3. The method according to claim 1, characterized in that The pretreatment includes: Performing adaptive noise elimination based on wavelet threshold and time-domain synchronous averaging on the vibration signal to filter out interference signals; performing sliding window normalization on the oil data to unify the data scale; Finite element meshing is used to generate a thermal stress matrix for the temperature image data to quantify temperature distribution characteristics.

4. The method according to claim 1, wherein The extracted features include: Extract 128-dimensional time-frequency features of vibration signals through CEEMDAN decomposition and 1D-CNN; ResNet-18 is used to classify the oil wear particle image morphology, and the dimension is reduced to 64 dimensions through PCA; The relationship characteristics between the temperature field and the heat conduction of the structural parts are constructed based on the graph neural network, and a 64-dimensional feature vector is obtained after dimension regularization.

5. The method according to claim 1, wherein The deployment of the deep learning and statistical analysis hybrid model includes: Construct an LSTM-Transformer hybrid network, where the LSTM has at least one layer and the number of hidden units is a preset value, which is adaptively adjusted according to the gearbox model or operating conditions. The Transformer encoder uses an 8-head attention mechanism. Paris crack growth physics constraints are embedded in the loss function to quantify the fatigue damage evolution of key gearbox components.

6. The method according to claim 1, characterized in that The combined federated learning framework includes: Parameters are aggregated every 24 hours, and weights are allocated based on the amount of node data.

7. The method according to claim 1, characterized in that The display of the gearbox health indicator degradation curve and the 95% probability interval of the remaining life includes: Output the remaining life (RUL) prediction results in real time through edge computing nodes; The elastic weight curing (EWC) algorithm is used for incremental learning to optimize model stability, ensure that the relative error between the predicted remaining life and the actual remaining life does not exceed 5%, and dynamically update the display content.

8. A gearbox life prediction device, characterized in that: The device comprises: Data acquisition module, used to synchronously collect gearbox vibration signals, oil data, temperature image data, and record speed and load condition parameters; The data preprocessing and feature extraction module is used to preprocess and extract features from the collected data, and then obtain a high-dimensional feature vector through multi-modal adaptive fusion based on confidence; The model deployment and analysis module is used to deploy a hybrid model of deep learning and statistical analysis. In combination with the federated learning framework, it determines the weights based on the data volume of each gearbox and performs parameter aggregation analysis. The result display module is used to display the gearbox health indicator degradation curve and the 95% probability interval of the remaining life. It outputs the remaining life (RUL) prediction results in real time through the edge computing node, uses the elastic weight solidification algorithm for incremental learning to optimize the model stability, and dynamically updates the display content.

9. The device according to claim 8, characterized in that The data acquisition module includes: The vibration sensor is installed in the gearbox bearing seat and close to the meshing area, with a sampling frequency of ≥25.6kHz; the oil sensor is integrated into the oil return port of the gearbox oil circulation system, with a wear particle identification accuracy of ≥5μm; The infrared thermal imager is deployed in the gearbox observation window and aligned with the meshing surface, with a resolution of ≥640×512 pixels.

10. The device according to claim 8, characterized in that The data preprocessing and feature extraction module, model deployment and analysis module, and result display module are all integrated in the industrial computer, and data transmission and interaction are achieved through an internal industrial-grade data bus.

Images