Rotary bearing full life cycle health monitoring method based on multi-source data fusion

Abstract

This invention proposes a method for full lifecycle health monitoring of slewing bearings based on multi-source data fusion, relating to the fields of mechanical condition monitoring and intelligent manufacturing. The method includes: collecting multimodal raw signals from each slewing bearing during machine operation to obtain standardized feature data packets; writing the standardized feature data packets into a time-series database, and writing the historical operating data and raw equipment data corresponding to the slewing bearing into a relational database; constructing full lifecycle data for the slewing bearing based on the data stored in the time-series and relational databases; performing adaptive feature fusion based on the multimodal features in the full lifecycle data to obtain a fused health state feature vector; and inputting the fused health state feature vector into a hybrid prediction and diagnostic model to obtain the failure mode probability and remaining life prediction results for the slewing bearing. This application helps improve the accuracy and robustness of slewing bearing diagnosis.

Description

Technical Field

[0001] This invention relates to the field of mechanical condition monitoring and intelligent manufacturing, and in particular to a method for full life cycle health monitoring of slewing bearings based on multi-source data fusion. Background Technology

[0002] Slewing bearings are critical fundamental components in major technical equipment such as wind turbine generators, tunnel boring machines, heavy-duty port cranes, and radar antennas, enabling core rotational motion and bearing complex multidimensional loads. Because these components operate under extreme conditions such as low speed, heavy load, and high overturning moments for extended periods, they are prone to performance degradation or sudden failure. This can lead to unplanned downtime of the entire equipment, causing significant direct economic losses, and may also trigger serious cascading safety accidents. Therefore, a slewing bearing full lifecycle health monitoring system based on multi-source data fusion is employed to continuously and reliably monitor the health status of the slewing bearing, ensuring the safe, stable, and efficient operation of critical equipment.

[0003] Chinese patent CN117408112A discloses a fatigue life monitoring system for complex equipment components based on digital twins. This system establishes virtual models of the components in a virtual space and collects multi-source signal data from actual operating conditions. After A / D conversion and data preprocessing, it obtains the first data. The stress value and the first fatigue life value of the component are obtained through finite element analysis in a numerical calculation unit. This data is then fused with material properties, dimensions, and influencing factor data to train a machine learning model. The first data is then input into the machine learning model to calculate material coefficients, ultimately yielding the second fatigue life value. However, the finite element numerical calculation unit used in this scheme runs asynchronously on a dedicated offline server, primarily for generating training samples rather than participating in online inference. This makes it difficult to balance physical interpretability and real-time performance in predicting the remaining life under complex and variable operating conditions, resulting in limited accuracy in health assessment. Therefore, it is essential to provide a method and system for full life-cycle health monitoring of slewing bearings based on multi-source data fusion to improve the diagnostic precision and robustness of slewing bearings. Summary of the Invention

[0004] In view of this, the present invention proposes a method for full life cycle health monitoring of slewing bearings based on multi-source data fusion.

[0005] This invention provides a method for full life-cycle health monitoring of slewing bearings based on multi-source data fusion, the method comprising: Collect multimodal raw signals of each slewing bearing during machine operation, and preprocess the multimodal raw signals to obtain standardized feature data packets; The standardized feature data package is written into the time series database, and the historical operating data and original equipment data corresponding to the slewing bearing are written into the relational database. Based on the data stored in the time series database and the relational database, the full life cycle data corresponding to the slewing bearing is constructed. Adaptive feature fusion is performed based on the multimodal features in the full life cycle data to obtain a fused health status feature vector. The fused health status feature vector is then input into a hybrid prediction and diagnostic model to obtain the failure mode probability and remaining life prediction results corresponding to the slewing bearing. The hybrid prediction and diagnostic model includes a physical model and a data-driven model.

[0006] Based on the above technical solutions, preferably, the preprocessing of the multimodal raw signal to obtain standardized feature data packets specifically includes: The vibration signal in the original multimodal signal is bandpass filtered to obtain a standard vibration signal; The acoustic emission signal in the original multimodal signal is bandpass filtered for typical damaged frequency bands to obtain a standard acoustic emission signal; Within a fixed time window, the time-domain statistical characteristics and frequency-domain characteristics of each sensor are calculated. The time-domain statistical characteristics, the frequency-domain characteristics, the standard vibration signal, and the standard acoustic emission signal are packaged into a standardized feature data package. A timestamp and device identification code are added to each standardized feature data in the standardized feature data package.

[0007] Based on the above technical solutions, preferably, a static attribute table and a dynamic maintenance event table are established for each slewing bearing in the relational database. The static attribute table includes the design model, geometric dimensions, rated load, factory serial number, and initial installation clearance measurement value. The dynamic maintenance event table includes the type and dosage of grease used for each lubrication, bolt preload verification value, and component replacement history.

[0008] More preferably, the adaptive feature fusion based on the multimodal features in the full lifecycle data specifically includes: The multimodal features from different sensors in the full lifecycle data are converted into a multimodal feature matrix in the form of a time window. The multimodal feature matrix is ​​input into a front-end feature extraction network containing a one-dimensional convolutional layer to obtain the feature representation corresponding to the multimodal feature matrix; The feature representations are input into a multi-head self-attention layer based on a scaled dot product attention mechanism to obtain the attention weights of each feature representation in the current time window. Based on the attention weights, the feature representations of different sensors are weighted, summed, and deeply fused to obtain fused features. These fused features are then mapped through a multi-layer fully connected network to output a fused health status feature vector representing the current operating health status of the slewing bearing.

[0009] More preferably, after outputting the fused health status feature vector representing the current operating health status of the slewing bearing, the method further includes: The fused health status feature vector is input into the classifier so that the classifier outputs the probability of the fault mode corresponding to normal, raceway fatigue spalling, poor lubrication and fastener loosening, and generates the corresponding graded early warning signal when the probability of the fault mode is greater than the preset fault threshold.

[0010] More preferably, the training process of the hybrid prediction and diagnosis model includes: Multimodal data is collected and preprocessed, and the actual remaining lifetime labels corresponding to the multimodal data are obtained to construct a training sample set. The multimodal data includes stress data, load data, temperature data, and vibration data. Based on Lundberg-Palmgren fatigue life theory, the stress data, load data and material parameters in the training sample set are used as inputs to obtain the first remaining life prediction output of the physical model, and the multimodal data is input into the neural network to obtain the second remaining life prediction output of the data-driven model. The first remaining lifetime prediction output and the second remaining lifetime prediction output are weighted and fused with corresponding weighting coefficients to obtain the remaining lifetime prediction result output by the hybrid prediction and diagnostic model. The mean squared error between the predicted remaining lifetime and the corresponding actual remaining lifetime label is used as the loss function to calculate the loss value of the current batch of training samples. The weighting coefficients and the corresponding model parameters of the physical model and the data-driven model are updated jointly through backpropagation and gradient descent until the loss function converges or the number of training rounds reaches a preset value.

[0011] More preferably, the method further includes: Using preset weighted fusion rules, the failure probability, remaining life, and current operating load of the slewing bearing are comprehensively calculated to obtain a real-time comprehensive health score. Real-time fault warning information, remaining life prediction trends, and comprehensive health scores for each slewing bearing can be displayed via the web or mobile terminal.

[0012] A second aspect of this application provides a slewing bearing full life-cycle health monitoring system based on multi-source data fusion. The slewing bearing full life-cycle health monitoring system includes a data acquisition module, a data processing module, and a life prediction module. The data acquisition module is used to acquire the multimodal raw signals of each slewing bearing during machine operation, and to preprocess the multimodal raw signals to obtain standardized feature data packets. The data processing module is used to write the standardized feature data packet into the time series database, and write the historical operating data and original equipment data corresponding to the slewing bearing into the relational database. Based on the data stored in the time series database and the relational database, the module constructs the full life cycle data corresponding to the slewing bearing. The life prediction module is used to perform adaptive feature fusion based on the multimodal features in the full life cycle data to obtain a fused health status feature vector, and input the fused health status feature vector into the hybrid prediction and diagnosis model to obtain the failure mode probability and remaining life prediction results corresponding to the slewing bearing. The hybrid prediction and diagnosis model includes a physical model and a data-driven model.

[0013] A third aspect of this application provides an electronic device including a processor, a memory, a user interface, and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory.

[0014] A fourth aspect of this application provides a non-transitory computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of a method for full life-cycle health monitoring of slewing bearings based on multi-source data fusion.

[0015] The slewing bearing full life cycle health monitoring method and system based on multi-source data fusion provided by this invention has the following advantages over existing technologies: (1) By collecting and preprocessing the multimodal raw signals and then performing adaptive feature fusion, compared with the method of relying solely on a single signal or manually selecting features, it can more comprehensively characterize the real operating state of the slewing bearing. Adaptive feature fusion can dynamically adjust the weight of each modal feature according to the changes in working conditions, reduce redundant information and noise interference, thereby improving the discrimination ability of the health status feature vector. Furthermore, by writing standardized feature data into the time series database and writing historical operating data and equipment raw data into the relational database, a unified full life cycle data system is constructed, making the operating status of the slewing bearing from commissioning to decommissioning recordable, queryable, and traceable. At the same time, the hybrid prediction and diagnosis model organically combines the physical model and the data-driven model. The physical model provides mechanistic constraints and boundary conditions, while the data-driven model captures complex nonlinear features and degradation laws. The two complement each other to overcome the limitations of a single model. By outputting the failure mode probability of the fused health status feature vector, it can not only identify whether there is a failure, but also distinguish the specific failure type, thereby improving the diagnostic precision and robustness.

[0016] (2) By constructing a multimodal feature matrix from the multimodal features of different sensors according to the time window, and then performing deep feature representation and adaptive weighted fusion through a front-end feature extraction network containing a one-dimensional convolutional layer and a multi-head self-attention layer based on the scaling dot product attention mechanism, it is possible to simultaneously mine the temporal features and cross-sensor correlations in the operation of the slewing bearing, automatically highlight the key signals that are more sensitive to the health status, suppress redundant and noise features, thereby obtaining a highly discriminative fused health status feature vector that represents the current operating health status. The fused feature vector is then input into a classifier to output the probabilities of various fault modes such as normal, raceway fatigue spalling, poor lubrication, and loose fasteners. When the probability of the fault mode exceeds a preset threshold, a graded warning signal is generated. This not only significantly improves the accuracy and robustness of fault identification and classification, but also realizes graded warnings for faults of different severity, which is conducive to taking targeted maintenance measures in advance. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating the method for full life-cycle health monitoring of slewing bearings based on multi-source data fusion provided by this invention; Figure 2 This is a schematic diagram of the sensor deployment corresponding to the slewing bearing provided by the present invention; Figure 3 This invention provides a hierarchical, full-lifecycle health monitoring architecture. Figure 4 The data fusion and processing flowchart of the intelligent analysis core provided by this invention; Figure 5 A simulation diagram of the remaining life prediction trend curve of the slewing bearing provided by the present invention; Figure 6 This is a schematic diagram of the slewing bearing full life cycle health monitoring system provided by the present invention; Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention.

[0019] Explanation of reference numerals in the attached diagram: 1. Slewing bearing full life cycle health monitoring system; 11. Data acquisition module; 12. Data processing module; 13. Life prediction module; 2. Slewing bearing; 21. Slewing bearing outer ring; 22. Rolling element; 23. Slewing bearing inner ring; 24. Vibration accelerometer; 25. Acoustic emission sensor; 26. Temperature sensor; 27. Dynamic strain gauge; 3. Electronic equipment; 31. Processor; 32. Communication bus; 33. User interface; 34. Network interface; 35. Memory. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] This invention discloses a method for full life-cycle health monitoring of slewing bearings based on multi-source data fusion, with reference to... Figure 1 The method includes steps S1 to S3.

[0022] Step S1: Collect multimodal raw signals of each slewing bearing during machine operation, and preprocess the multimodal raw signals to obtain standardized feature data packets.

[0023] In this step, a multi-source sensing and edge processing layer is established to determine the optimal placement of sensors within the slewing bearing. Based on the slewing bearing's failure physics model, this layer specifically deploys triaxial vibration accelerometers, acoustic emission sensors, temperature sensors, and strain gauges near the raceway in the bearing's load-bearing area. Simultaneously, an embedded edge computing unit is integrated, incorporating an adaptive filter and initial feature extraction algorithm to denoise and compress the original signal and extract primary time-domain and frequency-domain features, forming a standardized feature data package.

[0024] Furthermore, such as Figure 2 As shown, taking a single-row rolling slewing bearing 2 as an example, it includes a slewing bearing outer ring 21, rolling elements 22, a slewing bearing inner ring 23, and a slewing bearing cage. Multiple vibration accelerometers 24 are used on the slewing bearing 2 for vibration monitoring. The vibration accelerometers 24 are installed at multiple end faces along the circumferential direction of the outer ring 21. The sensitive axis of the vibration accelerometer 24 is perpendicular to the end face and points towards the center of rotation to capture radial vibration from all directions. The mounting surface needs to be flattened and fixed using a combination of high-reliability adhesive and mechanical fastening. Acoustic emission sensors 25 with resonant frequencies within a specific range are selected for acoustic emission monitoring. They are precisely installed on the end face of the outer ring 21 corresponding to the main load-bearing area. A special acoustic coupling agent is filled between the acoustic emission sensor 25 and the end face of the outer ring 21 to ensure low-loss transmission of high-frequency stress wave signals. The probe of a temperature sensor 26 is embedded in a pre-machined temperature measuring hole in the outer ring 21 of the slewing bearing. The probe is positioned close to the sealing structure to monitor frictional temperature rise. Dynamic strain gauge 27 is disposed between the outer ring 21, rolling element 22 and inner ring 23 of slewing bearing. A non-contact current sensor is installed in the power circuit of the drive motor. By monitoring the motor current in real time and combining it with the transmission model, the real-time load spectrum acting on the slewing bearing 2 is indirectly calculated.

[0025] All sensor signals are connected to an edge intelligent acquisition unit installed near the device. This intelligent acquisition unit integrates a multi-channel synchronous analog-to-digital converter and an embedded processor. Its workflow is as follows: Raw analog signals from all channels are synchronously acquired at a sampling rate of at least 20kHz. Subsequently, the built-in preprocessing firmware performs bandpass filtering on the vibration signal to eliminate low-frequency oscillations and high-frequency noise, and bandpass filtering on the acoustic emission signal to extract typical damage frequency bands, and calculates their root mean square (RMS) values. The system then packages the preprocessed time-domain statistical characteristics (such as RMS, peak value, and kurtosis) and frequency-domain characteristics (such as spectral centroid and sideband energy) of each sensor into a standard-format feature dataset within a fixed time window.

[0026] The failure process of slewing bearings is mainly dominated by factors such as physical wear, fatigue accumulation, and material damage. If the failure of a slewing bearing can be reflected through the analysis of vibration and strain signals, it can be modeled as a multi-physics coupled nonlinear system. For the establishment of the failure model, a dual-coupling model based on stress and vibration is used to describe the fatigue process of the slewing bearing.

[0027] The failure model of a slewing bearing is expressed as follows: in, R Indicates the remaining life of the slewing bearing. Indicates the first t Stress at any moment Indicates the first t Instantaneous load on the slewing bearing at any given moment. and These constants, determined by the slewing bearing material and operating conditions, characterize the cumulative nature of fatigue damage. The failure model of the slewing bearing describes the change in its lifespan over time under load and stress. By monitoring changes in stress and load in real time, the remaining service life of the slewing bearing is estimated.

[0028] The goal of optimal sensor placement is to minimize the cost of the monitoring system while ensuring sufficient monitoring accuracy. Assuming that sensors are rationally arranged based on the physical model of the slewing bearing to obtain the most effective monitoring data, the location and number of sensor points can be optimized through strategies that minimize information redundancy and maximize information capture.

[0029] The mathematical objective function for optimizing sensor placement is expressed as: in, Represents the sensor's position vector. Indicates the first i The sensor at the first t The signal of the moment, Indicates the start time of monitoring. Indicates the end time of monitoring. m Indicates the number of sensors. n This indicates the number of monitoring points for the slewing bearing. This represents a penalty factor used to control the density of the sensor arrangement. P j The first term represents the position vector P of the sensor. j The objective function consists of several components. The first term is the effective information content of the sensor signal, and the second term is the cost of sensor deployment. The introduction of this concept means that the optimization problem not only considers maximizing the amount of information, but also balances the number of sensors with the cost of deployment. By optimizing this objective function, the optimal sensor deployment locations and number can be determined.

[0030] The placement of sensors is achieved by minimizing this objective function, ultimately resulting in a sensor layout capable of accurately monitoring the condition of the slewing bearing. The sensor layout must consider not only the distribution characteristics of physical quantities such as vibration, temperature, and strain, but also potential interference factors encountered during the placement process, such as interference between sensors and environmental noise.

[0031] This step also includes steps S11 to S13.

[0032] Step S11: Bandpass filtering is performed on the vibration signal in the original multimodal signal to obtain a standard vibration signal.

[0033] Step S12: Perform bandpass filtering on the acoustic emission signal in the original multimodal signal for typical damaged frequency bands to obtain a standard acoustic emission signal.

[0034] Step S13: Calculate the time-domain statistical characteristics and frequency-domain characteristics of each sensor within a fixed time window, package the time-domain statistical characteristics, frequency-domain characteristics, standard vibration signal and standard acoustic emission signal into a standardized feature data package, and add a timestamp and device identification code to each standardized feature data in the standardized feature data package.

[0035] Step S2: Write the standardized feature data package into the time series database, and write the historical operating data and original equipment data corresponding to the slewing bearing into the relational database. Based on the data stored in the time series database and the relational database, construct the full life cycle data corresponding to the slewing bearing.

[0036] In this step, a static attribute table and a dynamic maintenance event table are created for each slewing bearing in a relational database. These tables are then linked. The static attribute table includes the design model, geometric dimensions, rated load, serial number, and initial installation clearance measurement. The dynamic maintenance event table includes the type and dosage of grease used for each lubrication, bolt preload verification values, and component replacement history. The feature dataset, containing timestamps and equipment identification codes, is then transmitted to the data center in real time. Upon receiving the data, the data center automatically parses the data packets, writes the time-series features into the corresponding sequence in the time-series database, and associates key events with the relational database.

[0037] Furthermore, data synchronization between the time-series database and the relational database is achieved through a periodic or real-time data synchronization mechanism. Time-series data is typically stored in the time-series database to support efficient time-series queries and historical data analysis; while static or semi-static data such as slewing bearing design parameters, assembly records, historical loads, and maintenance work orders are stored in the relational database. To ensure data consistency and integrity, the system uses a custom data synchronization module to periodically or under event-driven triggers synchronize key data from the time-series database to the relational database, and simultaneously synchronize updated data from the relational database to the time-series database. Key parameters during the synchronization process include synchronization frequency, data filtering and transformation rules, and data consistency verification. This synchronization mechanism ensures that the data in both databases remains up-to-date.

[0038] Step S3: Perform adaptive feature fusion based on the multimodal features in the full life cycle data to obtain a fused health status feature vector. Input the fused health status feature vector into the hybrid prediction and diagnosis model to obtain the failure mode probability and remaining life prediction results of the slewing bearing. The hybrid prediction and diagnosis model includes a physical model and a data-driven model.

[0039] This step also includes steps S31 to S35.

[0040] Step S31: Convert the multimodal features from different sensors in the full lifecycle data into a multimodal feature matrix in the form of a time window.

[0041] Step S32: Input the multimodal feature matrix into the front-end feature extraction network containing a one-dimensional convolutional layer to obtain the feature representation corresponding to the multimodal feature matrix.

[0042] Step S33: Input the feature representation into a multi-head self-attention layer based on the scaling dot product attention mechanism to obtain the attention weight of each feature representation in the current time window.

[0043] Step S34: Perform weighted summation and deep fusion operations on the feature representations of different sensors according to the attention weights to obtain fused features and map the fused features through a multi-layer fully connected network to output a fused health status feature vector representing the current operating health status of the slewing bearing.

[0044] In this step, a neural network model based on the scaling dot product attention mechanism is used to dynamically assign weights and deeply fuse multimodal features from different sensors to generate feature vectors that best reflect the current state.

[0045] Specifically, the neural network model based on the scaling dot product attention mechanism adopts an adaptive feature fusion module, which uses the scaling dot product attention mechanism to dynamically allocate the weights of each sensor feature to generate the feature vector that best reflects the current state of the slewing bearing. The network structure of the neural network model consists of an input layer, an attention mechanism layer, a feature fusion layer, and an output layer.

[0046] At the input layer, multimodal features from different sensors are represented as a set of matrices. ,in N This indicates the number of time steps or data points. D This represents the dimension of each sensor feature. In the attention mechanism layer, the importance of each feature in the final representation is determined by calculating the correlation between features. Let the input feature matrix be X. Then, a scaled dot product attention mechanism is defined to calculate the similarity between the query, key, and value, as shown below:

[0047] in, Represents the query matrix. Represents the key matrix. Represents a value matrix, d k K represents the dimension of the key. T The key matrix Q is represented by its transpose, softmax() represents the normalization exponential function, and Attention() represents the similarity calculation of the query matrix Q, the key matrix K, and the value matrix V. Used for scaling operations. By calculating attention weights between input features, the weights of each sensor's features are automatically adjusted based on their correlation at the current moment, thereby better fusing data from different sensors.

[0048] In the feature fusion layer, the feature matrix is ​​processed by the attention mechanism. The data is fed into a deep neural network for further fusion and mapping to generate the final feature representation. This process can be implemented using multiple fully connected layers. The final output layer generates the health assessment or fault diagnosis results for the slewing bearing.

[0049] Step S35: Input the fused health status feature vector into the classifier so that the classifier outputs the probability of the fault mode corresponding to normal, raceway fatigue spalling, poor lubrication and fastener loosening, and generates the corresponding graded early warning signal when the probability of the fault mode is greater than the preset fault threshold.

[0050] In this embodiment, by constructing a multimodal feature matrix from the multimodal features of different sensors according to time windows, and then sequentially passing it through a front-end feature extraction network containing a one-dimensional convolutional layer and a multi-head self-attention layer based on a scaling dot product attention mechanism for deep feature representation and adaptive weighted fusion, it is possible to simultaneously mine the temporal features and cross-sensor correlations during the operation of the slewing bearing, automatically highlight key signals that are more sensitive to the health status, and suppress redundant and noise features, thereby obtaining a highly discriminative fused health status feature vector that characterizes the current operating health status. This fused feature vector is then input into a classifier to output the probabilities of various fault modes, such as normal operation, raceway fatigue spalling, poor lubrication, and loose fasteners. When the probability of a fault mode exceeds a preset threshold, a graded warning signal is generated. This not only significantly improves the accuracy and robustness of fault identification and classification, but also realizes graded warnings for faults of different severity, which is conducive to taking targeted maintenance measures in advance.

[0051] In one example, the training process for the hybrid prediction and diagnostic model includes: Multimodal data is collected and preprocessed, and the actual remaining life labels corresponding to the multimodal data are obtained to construct a training sample set. The multimodal data includes stress data, load data, temperature data, and vibration data. Based on Lundberg-Palmgren fatigue life theory, stress data, load data and material parameters in the training sample set are used as inputs to obtain the first remaining life prediction output of the physical model, and multimodal data are input into the neural network to obtain the second remaining life prediction output of the data-driven model. The first remaining lifetime prediction output and the second remaining lifetime prediction output are weighted and fused with corresponding weighting coefficients to obtain the remaining lifetime prediction result of the hybrid prediction and diagnostic model output. The mean squared error between the remaining lifetime prediction result and the corresponding actual remaining lifetime label is used as the loss function to calculate the loss value of the current batch of training sample set. The weighting coefficients and the corresponding model parameters of the physical model and the data-driven model are updated jointly through backpropagation and gradient descent until the loss function converges or the number of training rounds reaches the preset value.

[0052] Furthermore, by constructing a hybrid prediction and diagnostic model that couples the Lundberg-Palmgren lifetime theory equation with a long short-term memory network, the physical trend constraint of slewing bearing life decay provided by the Lundberg-Palmgren lifetime theory equation is combined with the nonlinear deviation of the long short-term memory network under actual working conditions to jointly achieve high-precision fault classification and prediction of remaining service life probability.

[0053] The Lundberg-Palmgren life theory is used as the core algorithm in the hybrid prediction and diagnostic module, combining the physical model with the data-driven component to accurately predict the remaining life of slewing bearings. The Lundberg-Palmgren theory is primarily based on a cumulative damage model of fatigue life, suitable for describing the life decay process of mechanical components under cyclic loads. Specifically, the Lundberg-Palmgren life theory characterizes the fatigue damage accumulation process using the following formula:

[0054] in, Indicates the first t The cumulative amount of fatigue damage over time. For the first Instantaneous stress at a given moment Indicates the fatigue strength of the material. Indicates the fatigue index of a material. t Indicates the first t The remaining life of the slewing bearing is estimated by calculating the total amount of fatigue damage generated under different operating conditions. m and These are material property parameters, which are usually obtained through experiments.

[0055] In practical applications, It is a dynamic quantity that changes continuously with the actual operating state of the slewing bearing. To match it with the actual workload, this system incorporates a machine learning model to... Further optimization is performed by learning from historical operating conditions and load data to obtain more accurate fatigue damage predictions.

[0056] The lifetime prediction formula after combining the Lundberg-Palmgren model with the LSTM network is as follows: in, Indicates the first t Remaining lifespan at any given moment This represents the adjustment factor, used to correct for the impact of fatigue damage based on actual working conditions. These parameters, obtained through historical data and experiments, are characteristic parameters of slewing bearings under specific working environments. By providing a dynamic, time-varying prediction of remaining life, the influence of working conditions on fatigue damage is considered. By combining the Lundberg-Palmgren fatigue damage model with an LSTM network, more accurate fault classification and probabilistic prediction of remaining life can be achieved.

[0057] Cumulative fatigue damage The main result of the physical model is that it quantifies the health status of the slewing bearing. Meanwhile, the recurrent neural network (RNN) models the time-series data, captures potential nonlinear features, and predicts the remaining life. The dynamic changes. The output of a recurrent neural network is represented as:

[0058] Among them, X t Indicates the first t Sensor input characteristics at any given time, The parameters of a recurrent neural network, This represents the remaining life predicted based on historical data learned by the network. The output of the recurrent neural network reflects the prediction of the remaining life of the slewing bearing under various operating conditions through a data-driven model.

[0059] The outputs of the two models are weighted and fused, and the specific process is as follows: in, This indicates the predicted remaining lifetime after fusion. This represents the remaining lifetime output by the physical model. This represents the predicted output of the recurrent neural network. w 1 and w2 represents the weighting coefficients of the physical model and the recurrent neural network output, respectively.

[0060] Weight w 1 and w 2. During training, these parameters are dynamically adjusted using optimization algorithms, typically gradient descent. The training process is as follows: Initialize parameters, including all model parameters, such as the material fatigue strength in the physical model. and fatigue index m And the weights and biases of the neural network, initializing the weighting coefficients. w 1 and w 2; Data preparation: Collect and preprocess time-series data of the slewing bearing, including multi-modal sensor data such as stress, load, temperature, and vibration, and prepare the actual remaining life of the slewing bearing as a label for training purposes. The output of the physical model is calculated, and the cumulative fatigue damage at each time step is calculated using the physical model (Lundberg-Palmgren theory). And further calculate the remaining lifetime. The output is based on the stress history and material properties of sensor data; The output of a recurrent neural network (RNN) is calculated by inputting time-series data into the RNN to determine the model's predicted remaining lifetime. The neural network predicts the remaining life of the slewing bearing by learning patterns in time-series data. The outputs of the weighted fusion physical model and the recurrent neural network (RNN) are used, with weighting coefficients. w 1 and w 2. Output the physical model and recurrent neural network output Weighted fusion is performed to obtain the final remaining lifetime prediction. ; Calculate the loss function and define the loss function. Mean squared error (MSE) is used to measure the difference between predicted and true values: in, The label indicates the actual remaining useful life. Indicates the number of samples; Calculate the gradient of the loss function with respect to the parameters, and then calculate the loss function using the backpropagation algorithm. L Regarding network parameters and physical model parameters, the network parameters include the weighting coefficients of the physical model. w 1 and w 2. And the weights and biases of the recurrent neural network (RNN), and the physical model parameters including the material's fatigue index.m and material fatigue strength The gradient.

[0061] For weighting coefficients w 1 and w 2. Calculate the weighting coefficients respectively. w 1 and w The corresponding gradient of 2 and For the parameters of a neural network, such as the weights and biases of an RNN, the chain rule is used for calculation. For the physical model parameters, the corresponding gradients are calculated using numerical or analytical methods;

[0062] Update the physical model parameters by performing gradient updates on the physical model parameters, assuming the gradient of the physical model parameters is... and The update rule is as follows: in, Indicates the learning rate. This indicates the current material fatigue strength. Indicates the previous material fatigue strength. This indicates the current material fatigue index. This indicates the previous material fatigue index.

[0063] Update the network parameters of the recurrent neural network (RNN) using gradient descent to update the weights and biases. The update rule is as follows:

[0064] in, The parameters of the neural network, This represents the parameters of the recurrent neural network (RNN) before this update. This indicates that the recurrent neural network (RNN) has obtained new parameters after this gradient descent update.

[0065] Update the weighting coefficients using the gradient descent method. w 1 and w 2. To better balance the output contributions of the physical model and the recurrent neural network (RNN), the update rule is as follows:

[0066] in, Indicates the current weighting coefficient w 1, Indicates the weighting coefficient of the previous time.w 1, Indicates the current weighting coefficient w 2, Indicates the weighting coefficient of the previous time. w 2.

[0067] Repeat the training until convergence, and repeat the steps of calculating the loss function and updating the weighting coefficients until the loss function reaches the predetermined convergence criterion or the training reaches the maximum number of rounds, thereby obtaining a well-trained hybrid prediction and diagnosis model.

[0068] In this embodiment, by acquiring and preprocessing multimodal raw signals and then performing adaptive feature fusion, compared to methods based on a single signal or manually selected features, the true operating state of the slewing bearing can be more comprehensively characterized. Adaptive feature fusion can dynamically adjust the weights of each modal feature according to changes in operating conditions, reducing redundant information and noise interference, thereby improving the discriminative ability of the health status feature vector. Furthermore, by writing standardized feature data into a time-series database and writing historical operating data and equipment raw data into a relational database, a unified full life-cycle data system is constructed, making the operating status of the slewing bearing from commissioning to decommissioning recordable, queryable, and traceable. At the same time, the hybrid prediction and diagnostic model organically combines the physical model and the data-driven model. The physical model provides mechanistic constraints and boundary conditions, while the data-driven model captures complex nonlinear features and degradation laws. The two complement each other to overcome the limitations of a single model. By outputting the failure mode probability of the fused health status feature vector, it can not only identify whether there is a failure, but also distinguish the specific failure type, improving the diagnostic precision and robustness.

[0069] In one example, such as Figure 3 As shown, a hierarchical full life cycle health monitoring architecture corresponding to the slewing bearing full life cycle health monitoring method based on multi-source data fusion is presented.

[0070] exist Figure 3 In this system, the perception layer is used to capture raw data of the multi-physics state during the operation of the slewing bearing, outputting multimodal raw signals or their edge-preprocessed feature data. The data generated by the perception layer serves as the foundational input for subsequent full lifecycle data center construction and intelligent analysis.

[0071] The data layer includes a time-series database, a relational database, a unified data center, and a data fusion management module. The time-series database stores high-frequency / continuous time-series data or standardized feature data packages from the perception layer to support efficient time-series queries and trend analysis. The relational database stores static or semi-static full lifecycle data for the slewing bearing, including design parameters, assembly records, historical operating events, maintenance work orders, and replacement files. The unified data center performs unified modeling, identification, association, and consistency maintenance of time-series and structured data, enabling data from different stages to be interconnected according to the same equipment identifier and timeline. The data fusion management module performs data access parsing, cleaning and verification, alignment and association, synchronization, and access control, and provides directly accessible fused data services to the core analysis layer.

[0072] The core analysis layer is the core of intelligent algorithms and model operation. It includes an adaptive feature fusion module, a hybrid prediction and diagnosis module, an online adaptive learning engine, and a digital twin model. The adaptive feature fusion module dynamically weights and deeply fuses features from different sensors / modalities to form a fused feature vector representing the current health status. The hybrid prediction and diagnosis module outputs the failure mode probability and remaining life prediction results based on the fused feature vector. The digital twin model is used to characterize the state evolution mechanism of the slewing bearing under specific structural parameters and operating conditions, providing mechanistic constraints and comparisons for diagnosis, prediction, and interpretation. When the application layer sends back maintenance verification data or end-of-life data, the online adaptive learning engine triggers incremental learning / Bayesian updates to the fused model and the prediction and diagnosis model, enabling the model to continuously evolve as the equipment operates.

[0073] At the application level, the system supports operational decision-making and visualization, including a health assessment dashboard, a real-time early warning system, a lifespan prediction curve, maintenance suggestion generation, and a closed-loop feedback interface. The health assessment dashboard displays health scores, key indicators, and status trends; the real-time early warning system outputs tiered alarms based on fault probability and threshold strategies; the lifespan prediction curve displays the remaining lifespan prediction and its changing trends over time; the maintenance suggestion generation system comprehensively considers fault risk, lifespan reserve, and operating condition information to output actionable maintenance windows and handling suggestions; and the closed-loop feedback interface writes structured maintenance execution information, fault confirmation results, and component replacement records back to a unified data center, triggering an online adaptive learning engine to complete closed-loop updates.

[0074] Specifically, the data flow enters the data layer (time-series database / relational database) from the perception layer, and after being associated and merged in the unified data center and data fusion management module, it is called by the core analysis layer. The core analysis results enter the application layer for display, early warning and maintenance decision-making. The application layer sends back "true value data" such as maintenance verification and end-of-life data to the data layer and core analysis layer through the closed-loop feedback interface, driving online learning to realize closed-loop operation and maintenance of "monitoring-analysis-decision-feedback-re-optimization".

[0075] Furthermore, such as Figure 4 As shown, the adaptive feature fusion module is implemented by constructing a deep learning fusion model. The front end of this model uses a one-dimensional convolutional layer to extract the local correlations of feature sequences from each sensor, while the back end connects to a multi-head self-attention layer. The attention mechanism dynamically generates a set of weight coefficients by calculating the correlations between features, adaptively enhancing the signal features most relevant to the current state. For example, in the early micro-pitting stage, it automatically increases the weight of acoustic emission features while suppressing interference from noise or irrelevant features. The final output is a high-dimensional, information-dense fused health state feature vector.

[0076] The hybrid prediction and diagnosis module comprises two sub-modules: fault diagnosis and remaining life prediction. The fault diagnosis sub-module inputs a fused health status feature vector into a deep neural network-based classifier. This classifier outputs a probability distribution vector, corresponding to various preset fault modes such as "normal," "raceway fatigue spalling," "poor lubrication," and "loose fasteners." When the probability of any fault mode exceeds a preset warning threshold, the module generates a warning signal with the corresponding level. The remaining life prediction sub-module is a hybrid of a physical model and a data-driven model, and its architecture is shown in the attached figure. Figure 4 As shown, based on the classical life theory of rolling bearing contact fatigue, the theoretical remaining life L is calculated using the current and historical equivalent load P and the basic rated dynamic load C of the bearing as inputs. Simultaneously, a recurrent neural network is constructed, with the input being a sequence of historical "fused health state feature vectors". This network is trained to learn and predict the complex nonlinear deviation between the actual equipment degradation trajectory and the physical model baseline. The final individualized remaining life prediction is obtained by correcting the physical baseline using the output of the data-driven model. The system also outputs the probability distribution of the life prediction to quantify the uncertainty of the prediction. The online adaptive learning engine is implemented by using all multi-source data from a period of time back to the point of abnormal condition of the slewing bearing as new training samples, with the confirmed fault type as the ground truth label, to incrementally learn the fault diagnosis classifier. Furthermore, if the slewing bearing reaches the end of its lifespan, its complete full-cycle data sequence will be used to fine-tune and validate the parameters of the data-driven model in the remaining life prediction submodule, thereby achieving continuous evolution of the overall diagnostic prediction performance of the system.

[0077] In one example, a pre-defined weighted fusion rule is used to comprehensively calculate the failure probability, remaining life, and current operating load of the slewing bearing to obtain a real-time comprehensive health score. Real-time fault warning information, remaining life prediction trends, and comprehensive health scores of each slewing bearing are displayed via the web or mobile terminal.

[0078] A visual decision support and closed-loop feedback layer is established. This layer provides a human-computer interaction interface based on web technology, which can dynamically display health assessment dashboards, real-time warnings, and lifespan prediction curves, and automatically generate graded maintenance recommendations based on the prediction results. All executed maintenance operations and their results are structured and recorded through this layer, and automatically fed back as key data to the data center and analysis core, thus forming a complete "monitoring-analysis-decision-optimization" monitoring closed loop, driving the continuous iteration and improvement of the entire system.

[0079] The real-time comprehensive health score is calculated using a weighted fusion model, specifically: in, w 3 indicates the failure probability weight. w 4 indicates the remaining lifetime weight. w 5 indicates the load weight of the operating condition, which satisfies... The initial value was determined through optimization using historical fault data and maintenance costs, and subsequently fine-tuned dynamically by an online adaptive learning engine. This represents the highest probability of the fault mode output by the hybrid prediction and diagnosis module (value 0-1). This indicates a dynamic prediction of remaining lifetime. Indicates the design rated life. Indicates real-time operating load. This represents the rated load. The scoring dimensions are logically integrated to align with the closed-loop characteristics of the system's entire lifecycle data: the failure probability is used to represent the current level of health and safety; the ratio of remaining lifespan to rated lifespan reflects the lifespan reserve level; and the inverse ratio of actual load to rated load reflects the safety of operating conditions. By balancing the core health indicators through preset weights, a comprehensive score of 0-100 is finally output, with higher scores indicating better health status. This achieves the transformation of multi-source monitoring data into an intuitive, single score.

[0080] When a high-level warning is generated and the predicted remaining lifespan is lower than the preset safe operating threshold, the intelligent decision engine automatically and comprehensively assesses maintenance resources and costs, generating one or more actionable recommended maintenance suggestions. Finally, maintenance personnel feed back structured information such as maintenance operation details, the actual condition observed, and the serial numbers of replaced parts to the system via their terminals. This serves as crucial closed-loop data, automatically collected into the corresponding equipment file in the full lifecycle data center, and simultaneously triggers the online adaptive learning engine. This completes the entire closed loop from status perception, intelligent analysis, scientific decision-making to knowledge feedback, truly achieving continuous optimization of equipment health management.

[0081] like Figure 5 As shown, Figure 5 This is a simulation of the remaining life prediction trend curve. The vertical axis represents the percentage scale of health indicators, and the horizontal axis represents time. Within the historical interval before the "current" mark, a solid line represents the historical health indicator sequence, reflecting the gradual decline in the equipment's health status over time during operation. The dot at the "current" mark represents the observed health indicator value at the current moment. From the "current" mark onwards, the future health status is extrapolated based on the prediction model, and a dashed line represents the predicted central trend curve of the health indicators. The shaded area represents the range of prediction uncertainty (e.g., confidence intervals or upper and lower bounds), used to characterize possible fluctuations in the future degradation path.

[0082] Based on the above method, this application discloses a slewing bearing full life cycle health monitoring system based on multi-source data fusion, referencing... Figure 6 The slewing bearing full life cycle health monitoring system 1 includes a data acquisition module 11, a data processing module 12, and a life prediction module 13, wherein... Data acquisition module 11 is used to acquire multimodal raw signals of each slewing bearing during machine operation, and to preprocess the multimodal raw signals to obtain standardized feature data packets; Data processing module 12 is used to write standardized feature data packets into the time series database, and write the historical operating data and original equipment data corresponding to the slewing bearing into the relational database. Based on the data stored in the time series database and the relational database, it constructs the full life cycle data corresponding to the slewing bearing. The life prediction module 13 is used to perform adaptive feature fusion based on the multimodal features in the full life cycle data to obtain a fused health status feature vector, and input the fused health status feature vector into the hybrid prediction and diagnosis model to obtain the failure mode probability and remaining life prediction results of the slewing bearing. The hybrid prediction and diagnosis model includes a physical model and a data-driven model.

[0083] In one example, the data acquisition module 11 is used to perform bandpass filtering on the vibration signal in the multimodal raw signal to obtain a standard vibration signal; to perform bandpass filtering on the acoustic emission signal in the multimodal raw signal for typical damage frequency bands to obtain a standard acoustic emission signal; to calculate the time-domain statistical characteristics and frequency-domain characteristics of each sensor within a fixed time window; to package the time-domain statistical characteristics, frequency-domain characteristics, standard vibration signal and standard acoustic emission signal into a standardized feature data package; and to attach a timestamp and device identification code to each standardized feature data in the standardized feature data package.

[0084] In one example, a static attribute table and a dynamic maintenance event table are created for each slewing bearing in a relational database. The static attribute table includes the design model, geometric dimensions, rated load, serial number, and initial installation clearance measurement value. The dynamic maintenance event table includes the type and dosage of grease used for each lubrication, bolt preload verification value, and component replacement history.

[0085] In one example, the lifespan prediction module 13 is used to convert multimodal features from different sensors in the full lifespan data into a multimodal feature matrix in the form of a time window; input the multimodal feature matrix into a front-end feature extraction network containing a one-dimensional convolutional layer to obtain the feature representation corresponding to the multimodal feature matrix; input the feature representation into a multi-head self-attention layer based on a scaled dot product attention mechanism to obtain the attention weight of each feature representation in the current time window; perform weighted summation and deep fusion operations on the feature representations of different sensors according to the attention weights to obtain fused features, and map the fused features through a multi-layer fully connected network to output a fused health status feature vector representing the current operating health status of the slewing bearing.

[0086] In one example, the life prediction module 13 is used to input the fused health status feature vector into the classifier so that the classifier outputs the failure mode probabilities corresponding to normal, raceway fatigue spalling, poor lubrication and fastener loosening, and generates the corresponding graded early warning signal when the failure mode probability is greater than the preset failure threshold.

[0087] In one example, the training process for the hybrid prediction and diagnostic model includes: Multimodal data is collected and preprocessed, and the actual remaining life labels corresponding to the multimodal data are obtained to construct a training sample set. The multimodal data includes stress data, load data, temperature data, and vibration data. Based on Lundberg-Palmgren fatigue life theory, stress data, load data and material parameters in the training sample set are used as inputs to obtain the first remaining life prediction output of the physical model, and multimodal data are input into the neural network to obtain the second remaining life prediction output of the data-driven model. The first remaining lifetime prediction output and the second remaining lifetime prediction output are weighted and fused with corresponding weighting coefficients to obtain the remaining lifetime prediction result of the hybrid prediction and diagnostic model output. The mean squared error between the remaining lifetime prediction result and the corresponding actual remaining lifetime label is used as the loss function to calculate the loss value of the current batch of training sample set. The weighting coefficients and the corresponding model parameters of the physical model and the data-driven model are updated jointly through backpropagation and gradient descent until the loss function converges or the number of training rounds reaches the preset value.

[0088] In one example, the method also includes: Using preset weighted fusion rules, the failure probability, remaining life, and current operating load of the slewing bearing are comprehensively calculated to obtain a real-time comprehensive health score. Real-time fault warning information, remaining life prediction trends, and comprehensive health scores for each slewing bearing can be displayed via the web or mobile terminal.

[0089] Please see Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 7 As shown, the electronic device 3 may include: at least one processor 31, at least one network interface 34, user interface 33, memory 35, and at least one communication bus 32.

[0090] The communication bus 32 is used to enable communication between these components.

[0091] The user interface 33 may include a display screen and a camera. Optionally, the user interface 33 may also include a standard wired interface and a wireless interface.

[0092] The network interface 34 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0093] The processor 31 may include one or more processing cores. The processor 31 connects to various parts of the server using various interfaces and lines, and performs various server functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 35, and by calling data stored in the memory 35. Optionally, the processor 31 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 31 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 31 and may be implemented as a separate chip.

[0094] The memory 35 may include random access memory (RAM) or read-only memory. Optionally, the memory 35 may include non-transitory computer-readable storage medium. The memory 35 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 35 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 35 may also be at least one storage device located remotely from the aforementioned processor 31. Figure 7 As shown, the memory 35, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an application program for a slewing bearing full life-cycle health monitoring method based on multi-source data fusion.

[0095] exist Figure 7In the electronic device 3 shown, the user interface 33 is mainly used to provide an input interface for the user and to obtain the user input data; while the processor 31 can be used to call the application stored in the memory 35, which is a method for full life cycle health monitoring of slewing bearings based on multi-source data fusion. When executed by one or more processors, the electronic device executes one or more methods as described in the above embodiments.

[0096] A non-transitory computer-readable storage medium stores instructions that, when executed by one or more processors, cause a computer to perform one or more methods as described in the above embodiments.

[0097] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0098] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0099] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the shown or discussed mutual couplings or direct couplings or communication connections may be through some service interfaces; indirect couplings or communication connections between apparatuses or units may be electrical or other forms.

[0100] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0101] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0102] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0103] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for full life-cycle health monitoring of slewing bearings based on multi-source data fusion, characterized in that, The method includes: Collect multimodal raw signals of each slewing bearing during machine operation, and preprocess the multimodal raw signals to obtain standardized feature data packets; The standardized feature data package is written into the time series database, and the historical operating data and original equipment data corresponding to the slewing bearing are written into the relational database. Based on the data stored in the time series database and the relational database, the full life cycle data corresponding to the slewing bearing is constructed. Adaptive feature fusion is performed based on the multimodal features in the full life cycle data to obtain a fused health status feature vector. The fused health status feature vector is then input into a hybrid prediction and diagnostic model to obtain the failure mode probability and remaining life prediction results corresponding to the slewing bearing. The hybrid prediction and diagnostic model includes a physical model and a data-driven model.

2. The method for full life-cycle health monitoring of slewing bearings based on multi-source data fusion as described in claim 1, characterized in that, The preprocessing of the multimodal raw signal to obtain a standardized feature data packet specifically includes: The vibration signal in the original multimodal signal is bandpass filtered to obtain a standard vibration signal; The acoustic emission signal in the original multimodal signal is bandpass filtered for typical damaged frequency bands to obtain a standard acoustic emission signal; Within a fixed time window, the time-domain statistical characteristics and frequency-domain characteristics of each sensor are calculated. The time-domain statistical characteristics, the frequency-domain characteristics, the standard vibration signal, and the standard acoustic emission signal are packaged into a standardized feature data package. A timestamp and device identification code are added to each standardized feature data in the standardized feature data package.

3. The method for full life-cycle health monitoring of slewing bearings based on multi-source data fusion as described in claim 1, characterized in that, In the relational database, a static attribute table and a dynamic maintenance event table are established for each slewing bearing. The static attribute table includes the design model, geometric dimensions, rated load, factory serial number, and initial installation clearance measurement value. The dynamic maintenance event table includes the type and dosage of grease used for each lubrication, bolt preload verification value, and component replacement history.

4. The method for full life-cycle health monitoring of slewing bearings based on multi-source data fusion as described in claim 1, characterized in that, The adaptive feature fusion based on the multimodal features in the full lifecycle data specifically includes: The multimodal features from different sensors in the full lifecycle data are converted into a multimodal feature matrix in the form of a time window. The multimodal feature matrix is ​​input into a front-end feature extraction network containing a one-dimensional convolutional layer to obtain the feature representation corresponding to the multimodal feature matrix; The feature representations are input into a multi-head self-attention layer based on a scaled dot product attention mechanism to obtain the attention weights of each feature representation in the current time window. Based on the attention weights, the feature representations of different sensors are weighted, summed, and deeply fused to obtain fused features. These fused features are then mapped through a multi-layer fully connected network to output a fused health status feature vector representing the current operating health status of the slewing bearing.

5. The method for full life-cycle health monitoring of slewing bearings based on multi-source data fusion as described in claim 4, characterized in that, After outputting the fused health status feature vector representing the current operating health status of the slewing bearing, the method further includes: The fused health status feature vector is input into the classifier so that the classifier outputs the probability of the fault mode corresponding to normal, raceway fatigue spalling, poor lubrication and fastener loosening, and generates the corresponding graded early warning signal when the probability of the fault mode is greater than the preset fault threshold.

6. The method for full life-cycle health monitoring of slewing bearings based on multi-source data fusion as described in claim 1, characterized in that, The training process of the hybrid prediction and diagnosis model includes: Multimodal data is collected and preprocessed, and the actual remaining lifetime labels corresponding to the multimodal data are obtained to construct a training sample set. The multimodal data includes stress data, load data, temperature data, and vibration data. Based on Lundberg-Palmgren fatigue life theory, the stress data, load data and material parameters in the training sample set are used as inputs to obtain the first remaining life prediction output of the physical model, and the multimodal data is input into the neural network to obtain the second remaining life prediction output of the data-driven model. The first remaining lifetime prediction output and the second remaining lifetime prediction output are weighted and fused with corresponding weighting coefficients to obtain the remaining lifetime prediction result output by the hybrid prediction and diagnostic model. The mean squared error between the predicted remaining lifetime and the corresponding actual remaining lifetime label is used as the loss function to calculate the loss value of the current batch of training samples. The weighting coefficients and the corresponding model parameters of the physical model and the data-driven model are updated jointly through backpropagation and gradient descent until the loss function converges or the number of training rounds reaches a preset value.

7. The method for full life-cycle health monitoring of slewing bearings based on multi-source data fusion as described in claim 1, characterized in that, The method further includes: Using preset weighted fusion rules, the failure probability, remaining life, and current operating load of the slewing bearing are comprehensively calculated to obtain a real-time comprehensive health score. Real-time fault warning information, remaining life prediction trends, and comprehensive health scores for each slewing bearing can be displayed via the web or mobile terminal.

8. A slewing bearing full life-cycle health monitoring system based on multi-source data fusion, characterized in that, The slewing bearing full life cycle health monitoring system (1) includes a data acquisition module (11), a data processing module (12), and a life prediction module (13), wherein, The data acquisition module (11) is used to acquire the multimodal raw signals of each slewing bearing during machine operation, and to preprocess the multimodal raw signals to obtain standardized feature data packets; The data processing module (12) is used to write the standardized feature data packet into the time series database, and write the historical operating data and original equipment data corresponding to the slewing bearing into the relational database. Based on the data stored in the time series database and the relational database, the full life cycle data corresponding to the slewing bearing is constructed. The life prediction module (13) is used to perform adaptive feature fusion based on the multimodal features in the full life cycle data to obtain a fused health status feature vector, and input the fused health status feature vector into the hybrid prediction and diagnosis model to obtain the failure mode probability and remaining life prediction results corresponding to the slewing bearing. The hybrid prediction and diagnosis model includes a physical model and a data-driven model.

9. An electronic device, characterized in that, The device includes a processor (31), a memory (35), a user interface (33), and a network interface (34). The memory (35) is used to store instructions. The user interface (33) and the network interface (34) are used to communicate with other devices. The processor (31) is used to execute the instructions stored in the memory (35) to cause the electronic device (3) to perform the method as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-7.

Images