Self-aligning roller bearing fault diagnosis method and system

Abstract

The invention belongs to the technical field of bearing fault diagnosis, and discloses a self-aligning roller bearing fault diagnosis method and system, and the method comprises the steps: obtaining operation data and sensor data, carrying out the simulation through a digital twin model, calculating the deviation between model prediction and actual measurement, and carrying out the mode recognition according to the deviation. Digital twin model parameters are dynamically calibrated in a normal mode, fault type identification and bearing positioning are carried out in combination with the calibrated model in an abnormal mode, accurate diagnosis of new installation and bearing faults in a running-in period is realized, dynamic calibration of the digital twin model parameters is realized, normal deviation and abnormal deviation are distinguished, and fault diagnosis accuracy is improved. The accuracy of fault diagnosis in the new installation and running-in period is improved, and fault type recognition and fault bearing positioning are achieved.

Description

Technical Field

[0001] The present application relates to the technical field of bearing fault diagnosis, and in particular to a spherical roller bearing fault diagnosis method and system. Background Art

[0002] Large ball mills are core production equipment in mining machinery, and their stable and reliable operation is crucial. The mill's massive cylinder and the material inside generate extremely high loads, primarily supported by its main bearings. Typically, a large ball mill is equipped with multiple large spherical roller bearings as main bearings to share and support the enormous radial and axial loads. These main bearings are one of the equipment's weak links, and their health directly impacts the mill's operating efficiency and safety. To enable predictive maintenance of these critical components and reduce unplanned downtime, fault diagnosis methods based on digital twin technology have been introduced in recent years. This approach aims to construct a virtual digital model of the bearing and, combined with real-time operating data, simulate the bearing's operating status and degradation process, thereby enabling early warning and diagnosis of faults.

[0003] In the prior art, digital twin-based bearing fault diagnosis typically begins with constructing an idealized virtual digital twin model of a single spherical roller bearing based on its design drawing parameters, material properties, internal structure, and ideal assembly relationship with the main shaft and bearing seat. To obtain real-world bearing operating data, various sensors are deployed near each operating bearing or on its seat, such as a three-axis accelerometer for collecting vibration signals and a temperature sensor for monitoring temperature. These sensors collect key operating data, such as vibration and temperature, from the bearing under actual operating conditions in real time and transmit the data to the digital twin platform. After receiving the real-time sensor data, the digital twin platform maps it to the corresponding virtual digital model and drives the model to perform simulation calculations. By comparing the model simulation results (such as the predicted vibration spectrum, internal stress, and temperature field) with the actual sensor data, it is hoped that the actual operating state of the bearing can be reflected. If there is a deviation between the model's predicted state and the actual state, a fault can be determined.

[0004] However, in actual industrial applications, this digital twin-based approach faces significant accuracy challenges throughout the bearing's lifecycle, particularly during the early stages of a newly installed bearing. Specifically, after the equipment is installed and equipped with a brand-new spherical roller bearing, it enters the initial operation phase. Even if the manufacturing process for a newly installed spherical roller bearing adheres to strict tolerances and the installation process follows strict specifications, the cumulative effects of manufacturing tolerances and subtle variations that are difficult to completely avoid during actual installation (such as the tightness of the bearing-to-shaft journal and housing bore, the bearing's position within the housing, and the amount and distribution of grease) can lead to inherent differences between the digital twin model constructed based on ideal design parameters in key parameters such as the initial bearing internal clearance, the actual contact angle between the roller and raceway, the roller load distribution, and the actual fit between the bearing, housing, and spindle. Because these initial parameters fail to accurately reflect the initial state of the actual bearing, the simulation results of the digital twin model during the initial operation of the new bearing, such as the predicted vibration spectrum characteristics, internal bearing stress distribution, or temperature field distribution, can deviate significantly from the data collected by actual sensors. This deviation is not caused by a fault, but rather by a mismatch between the initial modeled state and the true physical state.

[0005] After the equipment is put into operation, the bearings will go through a running-in period. During the running-in period, the contact surfaces of the rollers, raceways and cages will undergo microscopic plastic deformation and slight wear, causing the contact surface morphology to change and possibly leading to slight adjustments in parameters such as the internal clearance of the bearing. These microscopic changes occur dynamically, and their process and results are affected by a variety of operating factors such as the actual load, speed, lubrication conditions and ambient temperature borne by the bearing. It is a gradual and nonlinear process. The running-in process will cause the actual internal geometric parameters (such as clearance, contact angle) and surface characteristics (such as roughness) of the bearing to change gradually until a relatively stable state is reached. Existing digital twin models are usually based on idealized physical models or macroscopic degradation models. Their internal parameters are often statically set, or only consider long-term accumulated macroscopic wear. It is difficult to accurately capture the microscopic and dynamic parameter changes that occur during the running-in period, and how these changes affect the internal dynamic behavior and external signal characteristics of the bearing. Therefore, during the running-in period, as the actual state of the bearing changes rapidly, the deviation between the prediction results of the static digital twin model and the actual sensor data will further evolve, resulting in a greater deviation between the virtual state predicted by the model and the actual state of the real bearing.

[0006] In addition, a large ball mill usually has multiple main bearings working together to support the huge cylinder and material loads. These bearings may have individual differences in initial state deviation, running-in process, and subsequent degradation rate. To make it more complicated, these bearings are connected and coupled to each other through rigid structural components such as the main shaft and cylinder. The state change of one bearing (such as changes in internal stiffness due to wear or impact caused by early failure) may affect the force distribution and vibration response of other bearings through structural transmission. The original digital twin model may model each bearing independently, or only consider a simple system coupling relationship. It is difficult to accurately simulate the mutual influence between multiple bearings through complex structural coupling, as well as the synergistic effect of different bearings at different degradation stages. When one of the bearings has an early minor fault, the weak signal it generates may be superimposed on the vibration signals of other bearings through structural transmission, making the fault signal characteristics complicated, making it difficult to accurately locate which bearing has a problem, and it is also difficult to distinguish the degradation of an individual bearing from the overall response of the system.

[0007] Traditional fault diagnosis methods or diagnoses based on static digital twin models have difficulty distinguishing between the normal running-in state of a new bearing during its initial operation and early signs of faults. Microscopic changes in the bearing surface during the running-in period may produce certain signal characteristics (such as high-frequency impact components), which may be similar in some aspects to the signals produced by early pitting or spalling, easily leading to false alarms, increased maintenance costs, and reduced system trust. At the same time, due to inaccurate initial model parameters, insufficient capture of dynamic changes during the running-in period, and the complexity of the coupling of multi-bearing systems, digital twin models cannot accurately reflect the health status of the actual bearing in the early stages of the bearing life cycle, especially the newly installed running-in period and the initial degradation stage. This reduces the reliability and timeliness of early fault diagnosis and may even lead to missed or false alarms.

[0008] In order to improve the accuracy of the digital twin model throughout the entire life cycle of the bearing, especially in the early stages, a mechanism is needed that can dynamically calibrate the initial parameters and model structure of the digital twin model using sensor data from the new installation and running-in period. This calibration needs to be able to identify and distinguish the impact of initial manufacturing and installation errors, parameter changes caused by normal running-in, and abnormal changes caused by early faults. Considering that actual industrial field data may be noisy and occasionally missing, and the resources for edge computing may be limited, how to design an efficient and robust calibration algorithm that can quickly and accurately adjust the digital twin model under data uncertainty and limited computing resources, so that it can adapt to the state changes of the bearing from new to old throughout its life cycle, and improve the sensitivity and diagnostic accuracy of early and minor faults, especially accurately distinguishing the source of faults in multi-bearing coupling systems, is a key technical challenge in achieving high-precision predictive maintenance of large ball mill main bearings.

[0009] In view of the above problems, the existing technology is in urgent need of improvement. Summary of the Invention

[0010] The purpose of this application is to provide a spherical roller bearing fault diagnosis method and system, which can realize dynamic calibration of digital twin model parameters, distinguish normal deviations from abnormal deviations, improve the accuracy of fault diagnosis during new installation and running-in period, and realize fault type identification and fault bearing location.

[0011] In a first aspect, the present application provides a method for diagnosing spherical roller bearing faults, which is used for diagnosing spherical roller bearing faults in a large ball mill multi-spherical roller bearing system during the new installation stage and the running-in period. The method comprises the following steps:

[0012] A1. Obtain the ball mill operating parameters and the multi-channel raw data collected by the multi-channel sensor, and preprocess them to obtain the preprocessed ball mill operating parameters and the preprocessed multi-channel measurement data;

[0013] A2. Input the preprocessed ball mill operating parameters into the pre-established system-level digital twin model for simulation calculation to obtain multi-channel prediction data predicted by the model;

[0014] A3. Calculate multi-channel deviation data based on the multi-channel prediction data and the preprocessed multi-channel measurement data;

[0015] A4. Extract time-domain features, frequency-domain features, and spatial correlation features from the multi-channel deviation data to obtain multi-channel deviation features;

[0016] A5. Based on the multi-channel deviation characteristics, deviation pattern recognition is performed to identify whether the multi-channel deviation data belongs to a normal deviation pattern or an abnormal deviation pattern;

[0017] A6. If a normal deviation pattern is identified, dynamically calibrate the key parameters of the system-level digital twin model using a parameter identification algorithm to obtain a calibrated system-level digital twin model.

[0018] A7. If an abnormal deviation pattern is identified, the multi-channel deviation data is used to analyze the multi-channel characteristics of the abnormal deviation, and combined with the latest system-level digital twin model, the fault type is identified and the faulty bearing is located.

[0019] Preferably, step A1 includes:

[0020] A101. Obtain the ball mill's cylinder speed, material filling rate, ambient temperature, and lubricating oil temperature during startup, constant speed operation, and shutdown to obtain the ball mill's operating parameters.

[0021] A102. Using triaxial acceleration sensors and temperature sensors installed on multiple spherical roller bearing seats, we collect vibration acceleration and temperature data from each bearing seat in three directions, generating multi-channel raw vibration acceleration data and multi-channel raw temperature data.

[0022] A103. Preprocess the ball mill operating parameters and multi-channel raw data to obtain preprocessed ball mill operating parameters and preprocessed multi-channel measurement data; the preprocessing includes time synchronization processing, denoising processing and outlier removal processing.

[0023] Preferably, step A3 includes:

[0024] A301. For vibration acceleration, calculate the root mean square error and spectrum amplitude difference between the preprocessed multi-channel measurement data and the preprocessed multi-channel prediction signal to obtain vibration deviation data;

[0025] A302. Calculate the difference between the preprocessed multi-channel measurement data and the corresponding data in the preprocessed multi-channel prediction data for temperature, cylinder speed, and material filling rate, respectively, to obtain temperature deviation data, speed deviation data, and material filling rate deviation data;

[0026] A303. The vibration deviation data, temperature deviation data, speed deviation data and material filling rate deviation data are normalized to obtain normalized vibration deviation data, normalized temperature deviation data, normalized speed deviation data and normalized material filling rate deviation data;

[0027] A305. Integrate the normalized vibration deviation data, the normalized temperature deviation data, the normalized speed deviation data, and the normalized material filling rate deviation data to obtain multi-channel deviation data.

[0028] Preferably, step A4 includes:

[0029] A401. For multi-channel deviation data, use the Fast Fourier Transform algorithm to calculate the spectrum of each channel deviation data and extract the spectrum amplitude, spectrum energy, spectrum centroid, and spectrum variance as frequency domain features.

[0030] A402. For multi-channel deviation data, calculate the mean, variance, root mean square, kurtosis, and skewness of each channel deviation data as time domain features;

[0031] A403. For the multi-channel deviation data, calculate the Pearson correlation coefficient, mutual information, and dynamic time warping distance between the deviation data of each channel as spatial correlation features;

[0032] A404. Fuse the extracted frequency domain features, time domain features, and spatial correlation features to obtain multi-channel deviation features.

[0033] Preferably, step A5 includes:

[0034] A501. Obtain a pre-trained initial deviation pattern classification model; the initial deviation pattern classification model includes a multi-layer stacked autoencoder and takes multi-channel deviation features as input;

[0035] A502. Use the dynamic time warping algorithm to calculate the similarity between the real-time multi-channel deviation features and the historical normal deviation pattern features. If the similarity falls below a preset similarity threshold, the operating condition is determined to have changed. The initial deviation pattern classification model is then fine-tuned using a transfer learning algorithm to obtain a deviation pattern classification model that is adaptive to the operating condition.

[0036] A503. Calculate the coupling strength of each channel deviation feature using the mutual information algorithm, construct a coupling strength matrix, and decouple the multi-channel deviation features using a graph convolutional neural network to obtain the decoupled multi-channel deviation features.

[0037] A504. Input the decoupled multi-channel deviation features into the working condition-adaptive deviation pattern classification model, and use the Softmax classifier to calculate the probability that the multi-channel deviation data belongs to an abnormal deviation pattern. If the probability of the abnormal deviation pattern is greater than the preset probability threshold, it is judged as an abnormal deviation pattern; otherwise, it is judged as a normal deviation pattern.

[0038] Preferably, step A503 includes:

[0039] The mutual information algorithm is used to calculate the mutual information value between every two channel deviation features in each channel deviation feature to construct the initial coupling strength matrix;

[0040] For the initial coupling strength matrix, a coupling strength threshold is set, and elements smaller than the coupling strength threshold are set to zero to obtain a sparse coupling strength matrix;

[0041] According to the sparse coupling strength matrix, the adjacency matrix of the graph convolutional neural network is constructed. The elements of the adjacency matrix represent whether there is a coupling relationship between the corresponding channel deviation features. If there is a coupling relationship, the element value is 1, otherwise it is 0.

[0042] The multi-channel deviation features and adjacency matrix are input into the graph convolutional neural network, and the graph convolutional neural network is used to perform multi-layer convolution operations on the multi-channel deviation features to achieve feature decoupling and obtain the decoupled multi-channel deviation features.

[0043] Preferably, step A6 includes:

[0044] A601. For the system-level digital twin model, establish a set of parameters to be identified, including bearing internal clearance, roller-raceway contact angle, damping coefficient, and equivalent stiffness of structural connections.

[0045] A602. Construct a parameter identification objective function aimed at minimizing the error between the model's predicted output and the actual measured data; wherein the error is the weighted root mean square error of the deviation data for each channel, and the weights of different channels are adaptively adjusted based on the signal-to-noise ratio of the corresponding channel deviation data to suppress the influence of low signal-to-noise ratio channels on parameter identification;

[0046] A603. Using the particle swarm optimization algorithm, within the preset parameter range, solve the parameter identification objective function to obtain the optimal solution for the parameter set to be identified;

[0047] A604. For each parameter in the optimized solution, calculate its confidence interval. If the confidence interval contains the parameter's initial value, the parameter is considered insensitive to bias and is removed from the set of parameters to be identified to reduce the computational complexity of model calibration.

[0048] A605. Update the optimized solutions of the remaining parameter sets to be identified to the system-level digital twin model to obtain a calibrated system-level digital twin model.

[0049] Preferably, step A7 includes:

[0050] A701. Extract time-domain, frequency-domain, and spatial correlation features from multi-channel deviation data to construct a multidimensional feature vector. Then, use principal component analysis to reduce the dimensionality of the multidimensional feature vector to obtain the reduced-dimensional abnormal deviation feature vector.

[0051] A702. Input the reduced-dimensional abnormal deviation feature vector into a pre-trained multi-classification support vector machine to identify a set of candidate fault types that match the reduced-dimensional abnormal deviation feature vector;

[0052] A703. For each fault type in the candidate fault type set, simulate the multi-channel vibration response of that fault type under the current operating conditions based on the latest system-level digital twin model to obtain the fault characteristics predicted by the model.

[0053] A704. Calculate the cosine similarity between the reduced-dimensional abnormal deviation feature vector and each fault feature predicted by the model, and determine the fault type with the highest cosine similarity as the final fault type.

[0054] A705. Based on the determined fault type and combined with the load distribution of each bearing seat in the system-level digital twin model, locate the bearing seat with the largest load change and identify it as the faulty bearing seat.

[0055] Preferably, step A703 includes:

[0056] For each fault type in the candidate fault type set, setting a fault parameter corresponding to the fault type; wherein the fault parameter includes crack size, wear amount and imbalance amount;

[0057] Import the set fault parameters into the latest system-level digital twin model, simulate the multi-channel vibration response of the fault type under the current working conditions, and obtain the multi-channel time-domain vibration signal predicted by the model;

[0058] The multi-channel time-domain vibration signal predicted by the model is subjected to short-time Fourier transform to extract the sensitive frequency components corresponding to the fault type and bearing structure parameters, and to construct the fault features predicted by the model.

[0059] In a second aspect, the present application provides a spherical roller bearing fault diagnosis system for diagnosing spherical roller bearing faults in a large ball mill multi-spherical roller bearing system during the new installation stage and running-in period. The system includes a data acquisition system, an edge computing unit, and a cloud computing platform.

[0060] The data acquisition system is used to obtain the ball mill operating parameters, collect multi-channel raw data through a multi-channel sensor, and send the ball mill operating parameters and the multi-channel raw data to the edge computing unit;

[0061] The edge computing unit is used to preprocess the ball mill operating parameters and multi-channel raw data to obtain preprocessed ball mill operating parameters and preprocessed multi-channel measurement data, input the preprocessed ball mill operating parameters into a pre-established system-level digital twin model, perform simulation calculations, and obtain multi-channel prediction data predicted by the model; calculate multi-channel deviation data based on the multi-channel prediction data and the preprocessed multi-channel measurement data, extract time domain features, frequency domain features, and spatial correlation features from the multi-channel deviation data to obtain multi-channel deviation features, and send the multi-channel deviation data and the multi-channel deviation features to the cloud computing platform;

[0062] The cloud computing platform is used to perform deviation pattern recognition based on multi-channel deviation characteristics, and identify whether the multi-channel deviation data belongs to a normal deviation pattern or an abnormal deviation pattern. If it is recognized as a normal deviation pattern, the key parameters of the system-level digital twin model are dynamically calibrated through a parameter identification algorithm to obtain a calibrated system-level digital twin model, and the calibrated system-level digital twin model is sent to the edge computing unit for model updating. If it is recognized as an abnormal deviation pattern, the multi-channel characteristics of the abnormal deviation are analyzed using the multi-channel deviation data, and combined with the latest system-level digital twin model, fault type identification and fault bearing location are performed.

[0063] Beneficial effects: The present application provides a spherical roller bearing fault diagnosis method and system, which obtains operating data and sensor data, uses a digital twin model for simulation, calculates the deviation between the model prediction and the actual measurement, and performs pattern recognition based on the deviation. In normal mode, the digital twin model parameters are dynamically calibrated, and in abnormal mode, the calibrated model is combined to identify the fault type and locate the bearing, thereby achieving accurate diagnosis of bearing faults during new installation and the running-in period. It has the function of realizing dynamic calibration of digital twin model parameters, distinguishing normal deviations from abnormal deviations, improving the accuracy of fault diagnosis during new installation and the running-in period, and achieving fault type identification and faulty bearing location. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is a flow chart of a spherical roller bearing fault diagnosis method provided in an embodiment of the present application.

[0065] Figure 2 A schematic structural diagram of a spherical roller bearing fault diagnosis system provided in an embodiment of the present application.

[0066] Explanation of numbers: 1. Data acquisition system; 2. Edge computing unit; 3. Cloud computing platform. DETAILED DESCRIPTION

[0067] The technical solutions in this application will be clearly and completely described below in conjunction with the drawings in this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all of the embodiments. The components of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application for which protection is claimed, but merely represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of this application.

[0068] It should be noted that similar reference numerals and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings. At the same time, in the description of this application, the terms "first", "second", etc. are only used to distinguish the description and should not be understood as indicating or implying relative importance.

[0069] refer to Figure 1 This application proposes a spherical roller bearing fault diagnosis method for large ball mill multi-spherical roller bearing system during the new installation stage and running-in period. The method comprises the following steps:

[0070] A1. Obtain the ball mill operating parameters and the multi-channel raw data collected by the multi-channel sensor, and preprocess them to obtain the preprocessed ball mill operating parameters and the preprocessed multi-channel measurement data;

[0071] A2. Input the preprocessed ball mill operating parameters into the pre-established system-level digital twin model for simulation calculation to obtain multi-channel prediction data predicted by the model;

[0072] A3. Calculate multi-channel deviation data based on the multi-channel prediction data and the preprocessed multi-channel measurement data;

[0073] A4. Extract time-domain features, frequency-domain features, and spatial correlation features from the multi-channel deviation data to obtain multi-channel deviation features;

[0074] A5. Based on the multi-channel deviation characteristics, deviation pattern recognition is performed to identify whether the multi-channel deviation data belongs to a normal deviation pattern or an abnormal deviation pattern;

[0075] A6. If a normal deviation pattern is identified, dynamically calibrate the key parameters of the system-level digital twin model using a parameter identification algorithm to obtain a calibrated system-level digital twin model.

[0076] A7. If an abnormal deviation pattern is identified, the multi-channel characteristics of the abnormal deviation are analyzed using multi-channel deviation data, combined with the latest system-level digital twin model (if the system-level digital twin model has been calibrated, the latest system-level digital twin model refers to the calibrated system-level digital twin model; otherwise, it refers to the initial system-level digital twin model) to identify the fault type and locate the faulty bearing.

[0077] This method uses the discrepancies between digital twin model predictions and actual measurement data for fault diagnosis. By calculating the deviation between the model and the actual system and performing pattern recognition on these deviations, it can distinguish between the model and actual conditions under normal conditions and those under abnormal conditions. This approach addresses the technical challenges of distinguishing normal from abnormal conditions, identifying faults, and locating faults during new installations and the run-in period. The model calibration step enables the digital twin model to adapt to dynamic changes in the system during new installations and the run-in period, supporting subsequent fault diagnosis.

[0078] Specifically, the method first acquires ball mill operating parameters and raw data collected by multi-channel sensors, such as drum speed, material filling rate, ambient temperature, lubricating oil temperature, and bearing seat vibration acceleration and temperature. These raw data undergo preprocessing, including time synchronization, denoising, and outlier removal, to improve data quality. This results in preprocessed operating parameters and measurement data. The preprocessed operating parameters are then input into a pre-established system-level digital twin model, which simulates the dynamic behavior of the ball mill and its multi-bearing system. The model performs simulation calculations based on the input operating parameters and outputs model-predicted multi-channel data, such as predicted bearing vibration and temperature. The model-predicted data are compared with the preprocessed actual measurement data, and the difference between the two is calculated to generate multi-channel deviation data. This deviation data reflects the degree of mismatch between the digital twin model and the actual system. To quantify the nature of the deviation, time-domain, frequency-domain, and spatial correlation features are extracted from the multi-channel deviation data. Time-domain features describe the temporal variation of the deviation, frequency-domain features describe the distribution of the deviation at different frequencies, and spatial correlation features describe the interrelationships between the deviations in different channels. These features are then integrated into a multi-channel deviation feature vector. Deviation feature vectors are used for pattern recognition to determine whether the current deviation is a fluctuation within the normal range (normal deviation pattern) or caused by an abnormal situation (abnormal deviation pattern). If the deviation is identified as normal, it indicates that the difference between the model and the actual system is within the normal range, possibly due to running-in or environmental changes. At this time, the parameter identification algorithm is activated, and the key parameters of the digital twin model are dynamically adjusted and calibrated using the current measurement data to make the model more closely match the current state of the actual system. The calibrated model is used for subsequent simulation and diagnosis. If the deviation is identified as abnormal, it indicates that the deviation exceeds the normal range and a fault may exist. At this time, the multi-channel characteristics of the abnormal deviation are combined with the latest digital twin model to identify the fault type and determine the possible fault type, such as roller or raceway damage. Furthermore, combined with the system information provided by the digital twin model, such as the load distribution of each bearing, the faulty bearing is located. Through the above steps, this method can effectively distinguish between normal conditions and early faults during new installation and running-in periods, and realize fault identification and location.

[0079] In some specific embodiments, ball mill operating parameters may include cylinder speed, material fill rate, ambient temperature, and lubricating oil temperature. Multi-channel sensors may include triaxial accelerometers and temperature sensors mounted on each bearing housing. Preprocessing may include aligning the different sensor data to a common time base, applying filters to remove noise, and identifying and removing data points outside a reasonable range. The system-level digital twin model may be a multibody dynamics model that includes modeling of the cylinder, main shaft, multiple spherical roller bearings, and bearing housings, taking into account their interconnections and interactions. Simulation calculations can calculate the force, motion, and temperature response of each bearing based on parameters such as the current cylinder speed and material fill rate. Multi-channel deviation data can be calculated as the difference or error between the actual measurement and the model's predicted value, such as the root mean square error (RMS) of a vibration signal or the difference in spectral amplitude, or the direct difference in temperature. Time domain features may include mean, variance, RMS, etc. Frequency domain features may include spectral amplitude, energy, etc. Spatial correlation features may include the correlation coefficient between deviations from different channels. Deviation pattern recognition can utilize machine learning classifiers, such as support vector machines or neural networks, which can be trained to distinguish between normal and abnormal deviation patterns. Parameter identification algorithms can employ optimization algorithms, such as particle swarm optimization or genetic algorithms, to minimize the deviation between model predictions and actual measurements, adjusting key model parameters such as bearing internal clearance, contact stiffness, or damping coefficient. Fault type identification can utilize abnormal deviation signatures to match those generated by model simulations of different fault types. Fault bearing location can be determined by combining the load or stress distribution of each bearing calculated by the model to determine which bearing is most likely to fail first due to load changes.

[0080] In some embodiments, step A1 comprises:

[0081] A101. Obtain the ball mill's cylinder speed, material filling rate, ambient temperature, and lubricating oil temperature during startup, constant speed operation, and shutdown to obtain the ball mill's operating parameters.

[0082] A102. Using triaxial acceleration sensors and temperature sensors installed on multiple spherical roller bearing seats, we collect vibration acceleration and temperature data from each bearing seat in three directions, generating multi-channel raw vibration acceleration data and multi-channel raw temperature data.

[0083] A103. Preprocess the ball mill operating parameters and multi-channel raw data to obtain preprocessed ball mill operating parameters and preprocessed multi-channel measurement data; the preprocessing includes time synchronization processing, denoising processing and outlier removal processing.

[0084] Specifically, this technical solution addresses data quality issues faced during fault diagnosis of multi-spherical roller bearing systems in large ball mills during the new installation and run-in phases by standardizing data collection content and implementing necessary preprocessing steps. First, key operating parameters such as drum speed, material fill rate, ambient temperature, and lubricating oil temperature are clearly collected. These parameters directly reflect the ball mill's operating status and external environment, providing accurate operating condition inputs for the subsequent digital twin model, ensuring that model simulation results more closely reflect actual operating conditions. Furthermore, triaxial accelerometers and temperature sensors are deployed on multiple bearing seats to comprehensively collect vibration and temperature data from each bearing seat, which are direct physical indicators of bearing health. Multi-channel data collection ensures monitoring of all critical support bearings in the system, providing essential data for analyzing inter-bearing interactions and locating faults. The collected raw data may contain time asynchrony, outliers due to environmental noise, or sensor anomalies, which can seriously affect the accuracy of subsequent data analysis and model calibration. Therefore, preprocessing is performed on the collected operating parameters and multi-channel raw data. Time synchronization aligns the data collection time points of different sensors to ensure that when comparing model predictions with actual measurement data, both correspond to the system state at the same moment. Denoising filters out random interference in the signal, improves the signal-to-noise ratio of the signal, and makes the features extracted from the vibration and temperature signals clearer and more reliable. Outlier removal removes obviously erroneous data points to prevent them from misleading statistical calculations and model training. Through these preprocessing steps, the raw data is converted into high-quality, reliable input data, which significantly improves the accuracy and robustness of subsequent digital twin model simulation, deviation calculation, deviation pattern recognition, and fault type identification and location. Especially under the challenges of complex and changeable data characteristics during new installations and running-in periods, improving data quality is crucial to distinguishing normal running-in from early faults.

[0085] In some embodiments, step A3 comprises:

[0086] A301. For vibration acceleration, calculate the root mean square error and spectrum amplitude difference between the preprocessed multi-channel measurement data and the preprocessed multi-channel prediction signal to obtain vibration deviation data;

[0087] A302. Calculate the difference between the preprocessed multi-channel measurement data and the corresponding data in the preprocessed multi-channel prediction data for temperature, cylinder speed, and material filling rate, respectively, to obtain temperature deviation data, speed deviation data, and material filling rate deviation data;

[0088] A303. The vibration deviation data, temperature deviation data, speed deviation data and material filling rate deviation data are normalized to obtain normalized vibration deviation data, normalized temperature deviation data, normalized speed deviation data and normalized material filling rate deviation data;

[0089] A305. Integrate the normalized vibration deviation data, the normalized temperature deviation data, the normalized speed deviation data, and the normalized material filling rate deviation data to obtain multi-channel deviation data.

[0090] Among them, for vibration acceleration data, the root mean square error and spectrum amplitude difference are used to quantify the deviation between the measured data and the predicted data (the calculation method of the root mean square error and spectrum amplitude difference is the existing technology and will not be described in detail here). The root mean square error reflects the difference in the overall energy of the signal, and the spectrum amplitude difference reflects the difference in the frequency domain components of the signal. In this way, the deviation characteristics of the vibration signal in the time domain and frequency domain can be captured. For data such as temperature, cylinder speed and material filling rate, difference calculation is used to quantify the deviation between the measured data and the predicted data. Difference calculation can reflect the degree of numerical deviation. Furthermore, the deviation data calculated from different physical quantities are normalized. The normalization process maps the deviation data of different channels to a unified numerical range, eliminates the dimensional difference, and makes the deviation data of different channels comparable. Finally, the various normalized deviation data are integrated to form a unified multi-channel deviation data.

[0091] Specifically, this method addresses the problem of multi-channel data containing different physical quantities, units, and dimensions. First, different methods are used to calculate the deviation between measured and predicted data for different physical quantities. For dynamic signals such as vibration acceleration, the root mean square error and spectral amplitude difference are calculated to fully reflect the differences in the time and frequency domains. For scalar or slowly varying signals such as temperature, rotational speed, and material fill rate, simple differences are calculated. This method generates raw deviation data reflecting the degree of deviation of different physical quantities. However, due to the different units and dimensions of these raw deviation data, the numerical ranges vary greatly. Directly combining these raw deviation data can cause some deviations to have a disproportionately large impact on the overall data. Therefore, these raw deviation data are normalized. Normalization scales the deviation data from different channels to a uniform scale, for example, mapping them to a distribution between 0 and 1 or with a mean of 0 and a variance of 1. This eliminates the influence of dimension and numerical range, making the deviation data from different channels comparable. Finally, these normalized deviation data are combined to form a unified multi-channel deviation dataset. This dataset contains the difference information between the system's predictions and measurements on each measurement channel. This information has been standardized and has a unified scale, which facilitates subsequent deviation feature extraction and pattern recognition, overcomes the limitations of simple deviation calculations, and lays the foundation for subsequent diagnostic steps.

[0092] In some embodiments, step A4 comprises:

[0093] A401. For multi-channel deviation data, use the Fast Fourier Transform algorithm to calculate the spectrum of each channel deviation data and extract the spectrum amplitude, spectrum energy, spectrum centroid, and spectrum variance as frequency domain features.

[0094] A402. For multi-channel deviation data, calculate the mean, variance, root mean square, kurtosis, and skewness of each channel deviation data as time domain features;

[0095] A403. For the multi-channel deviation data, calculate the Pearson correlation coefficient, mutual information, and dynamic time warping distance between the deviation data of each channel as spatial correlation features;

[0096] A404. Fuse the extracted frequency domain features, time domain features, and spatial correlation features to obtain multi-channel deviation features.

[0097] Specifically, this method aims to address the challenge of effectively extracting features that accurately reflect system status and potential fault information from complex multi-channel deviation data. A comprehensive feature set is constructed by performing time-domain, frequency-domain, and spatial-correlation feature extraction on the multi-channel deviation data. First, a fast Fourier transform is performed on the deviation data of each channel to calculate the spectrum. The spectral amplitude, spectral energy, spectral centroid, and spectral variance are extracted as frequency-domain features to capture the distribution and intensity of the deviation data at different frequencies. Next, the mean, variance, root mean square (RMS), kurtosis, and skewness of the deviation data of each channel are calculated as time-domain features to describe the statistical characteristics of the deviation data in the time series. Furthermore, the Pearson correlation coefficient, mutual information, and dynamic time warping distance between the deviation data of each channel are calculated as spatial-correlation features to reflect the interrelationship and synchronization of the deviation data between different channels. Finally, the extracted time-domain, frequency-domain, and spatial-correlation features are fused (i.e., these features are used to form a feature set) to obtain multi-channel deviation features. The fusion of different types of features provides richer and more comprehensive information, making up for the shortcomings of a single type of feature, thereby more accurately describing the characteristics of multi-channel deviation data and providing a basis for subsequent deviation pattern recognition and fault diagnosis.

[0098] In some specific embodiments, the collected multi-channel deviation data is processed. For example, for the multi-channel deviation data within a certain time period, first apply fast Fourier transform to the data segment of each channel, calculate its power spectrum density, and extract the total energy, the amplitude corresponding to the main peak frequency, the energy centroid frequency, and the energy variance from the power spectrum density as frequency domain features. At the same time, calculate the arithmetic mean, standard deviation, root mean square value, kurtosis coefficient, and skewness coefficient of each channel data segment as time domain features. Further, calculate the Pearson correlation coefficient between any two channel data segments to measure linear correlation; calculate mutual information to measure nonlinear correlation; calculate dynamic time warping distance to measure the similarity of time series and obtain a spatial correlation feature matrix. Finally, the time domain features, frequency domain features, and spatial correlation features between channels of all channels are spliced ​​or combined to form a high-dimensional multi-channel deviation feature vector for subsequent analysis.

[0099] In some embodiments, step A5 comprises:

[0100] A501. Obtain a pre-trained initial deviation pattern classification model; the initial deviation pattern classification model includes a multi-layer stacked autoencoder and takes multi-channel deviation features as input;

[0101] A502. Use the dynamic time warping algorithm to calculate the similarity between the real-time multi-channel deviation features and the historical normal deviation pattern features. If the similarity falls below a preset similarity threshold, the operating condition is determined to have changed. The initial deviation pattern classification model is then fine-tuned using a transfer learning algorithm to obtain a deviation pattern classification model that is adaptive to the operating condition.

[0102] A503. Calculate the coupling strength of each channel deviation feature using the mutual information algorithm, construct a coupling strength matrix, and decouple the multi-channel deviation features using a graph convolutional neural network to obtain the decoupled multi-channel deviation features.

[0103] A504. Input the decoupled multi-channel deviation features into the working condition-adaptive deviation pattern classification model, and use the Softmax classifier to calculate the probability that the multi-channel deviation data belongs to an abnormal deviation pattern. If the probability of the abnormal deviation pattern is greater than the preset probability threshold, it is judged as an abnormal deviation pattern; otherwise, it is judged as a normal deviation pattern.

[0104] This can be achieved through offline training. Using a historical multi-channel deviation feature dataset containing labels for normal and abnormal deviation patterns, a classification model based on a multi-layer stacked autoencoder is trained to learn feature representations under different patterns. A multi-layer stacked autoencoder is a deep learning model that, through layer-by-layer encoding and decoding, learns low-dimensional representations of input features that can be used for classification tasks.

[0105] In step A502, the currently acquired multi-channel deviation feature sequence is compared with the stored multi-channel deviation feature sequence under historical normal operating conditions to quantify their time series similarity. This can be achieved using a dynamic time warping (DTW) algorithm. The DTW algorithm calculates the optimal matching path between two time series on different time axes, thereby obtaining a distance or similarity value reflecting sequence similarity. The historical normal deviation pattern features can be stored as a feature template under typical normal operating conditions or as a collection of multiple normal samples.

[0106] Using the transfer learning algorithm to fine-tune the initial deviation pattern classification model means that after determining the working condition change, a small amount of labeled or unlabeled data under the current working condition is used to adjust the parameters of the pre-trained initial model to make it better adapt to the characteristic distribution of the current working condition. Specifically, this can be achieved by using the fine-tuning technology in transfer learning. For example, some bottom-level parameters of the multi-layer stacked autoencoder are fixed, and only the top-level or some high-level parameters are trained and adjusted, or a lower learning rate is used to train the entire model.

[0107] To quantify the degree of interdependence between any two channel features in a multi-channel deviation feature, a mutual information algorithm can be used to calculate the mutual information value between each channel deviation feature as the coupling strength. A larger mutual information value indicates a stronger correlation between the two channel features and a higher degree of coupling. The calculated coupling strengths between the channel deviation features are organized into a matrix, specifically an N×N matrix, where N is the number of channels and the elements of the matrix represent the coupling strength between the corresponding channel features.

[0108] Using graph convolutional neural networks to decouple multi-channel deviation features means treating multi-channel deviation features as node features in the graph structure, and using graph convolution operations to learn and extract independent feature representations that are not affected by inter-channel coupling. Specifically, this can be achieved using a graph convolutional neural network (GCN), converting the coupling strength matrix into a graph adjacency matrix (for example, by setting a threshold to set weak coupling relationships to zero and constructing a sparse graph), and then inputting the multi-channel deviation features as node features into the GCN. Through multi-layer graph convolution operations, the GCN can aggregate the feature information of neighboring nodes and reduce the redundancy and coupling between channels during the learning process to obtain a decoupled feature representation (that is, the feature vector output after processing by the graph convolutional neural network).

[0109] The decoupled features are used as input data and fed into a classification model that has been adjusted for the current operating conditions. A Softmax function is used at the output layer of the adaptive deviation pattern classification model to convert the model output into a probability distribution representing the input data belonging to each deviation pattern (normal or abnormal). The probability value of the input feature belonging to the abnormal deviation pattern is then extracted.

[0110] Specifically, this solution performs deviation pattern recognition based on multi-channel deviation features. First, a pre-trained initial deviation pattern classification model is obtained. This model is built based on a multi-layer stacked autoencoder and takes the multi-channel deviation features as input. Subsequently, a dynamic time warping algorithm is used to calculate the similarity between the real-time multi-channel deviation features and the historical normal deviation pattern features. This similarity is used to determine whether the current operating conditions have changed. If the similarity falls below a preset threshold, indicating that the operating conditions may have changed, a transfer learning algorithm is then used to fine-tune the initial model parameters to adapt them to the current operating conditions, resulting in a condition-adaptive deviation pattern classification model. Furthermore, to address the complex coupling relationships that may exist between multi-channel deviation features, a mutual information algorithm is used to calculate the coupling strength of each channel feature, construct a coupling strength matrix, and decouple the multi-channel deviation features using a graph convolutional neural network to obtain decoupled features. Finally, these decoupled features are input into the condition-adapted classification model. A softmax classifier is used to calculate the probability that the data belongs to an abnormal deviation pattern. If the calculated abnormal probability exceeds a preset probability threshold, the data is classified as an abnormal deviation pattern; otherwise, it is classified as a normal deviation pattern. In this way, this solution combines working condition adaptive model adjustment and feature decoupling processing, so that the deviation pattern recognition process can effectively cope with the challenges brought by working condition diversity and feature coupling in practical applications, improve the accuracy and robustness of the recognition results, and more effectively distinguish normal deviations from abnormal deviations.

[0111] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0112] Obtain an initial deviation pattern classification model pre-trained on historical data under various typical operating conditions. This model uses a stacked autoencoder structure consisting of three encoder layers and three decoder layers. Each encoder layer uses a fully connected layer and a Reluctant Unit (ReLU) activation function. The decoder structure is symmetrical to the encoder, and the input layer receives multi-channel deviation features.

[0113] After acquiring new multi-channel deviation features in real time, the dynamic time warping algorithm is used to calculate the DTW distance between the real-time feature sequence and the stored normal operating condition feature template sequence. A similarity threshold is set. For example, if the DTW distance is greater than a preset value (the corresponding similarity is lower than the DTW threshold), the operating condition is determined to have changed. At this point, a transfer learning method is used to fine-tune the top two fully connected layers of the initial model using a small amount of normal or abnormal sample data collected under the current operating conditions at a low learning rate (for example, 0.001), resulting in a deviation pattern classification model that is adaptive to the operating conditions.

[0114] The mutual information algorithm is used to calculate the mutual information between any two channel features in the current multi-channel deviation feature, constructing an N×N mutual information matrix, where N is the number of channels. A coupling strength threshold is set, for example, 0.5, and elements in the mutual information matrix that are less than the threshold are set to zero, resulting in a sparse coupling strength matrix. Based on this sparse matrix, the adjacency matrix of the graph convolutional neural network is constructed. If the mutual information between two channel features is greater than or equal to the threshold, the corresponding element in the adjacency matrix is ​​1, otherwise it is 0. The multi-channel deviation features and the constructed adjacency matrix are input into a GCN model containing two graph convolutional layers for feature decoupling, resulting in the decoupled multi-channel deviation features.

[0115] The decoupled multi-channel deviation features are input into a fine-tuned condition-adaptive deviation pattern classification model. The final layer of the model is connected to a Softmax classifier to calculate the probability that the input features belong to an abnormal deviation pattern. A preset probability threshold, such as 0.7, is set. If the calculated abnormal probability is greater than 0.7, the current multi-channel deviation data is considered to be in an abnormal deviation pattern; otherwise, it is considered to be in a normal deviation pattern.

[0116] Preferably, step A503 may include:

[0117] The mutual information algorithm is used to calculate the mutual information value between every two channel deviation features in each channel deviation feature to construct the initial coupling strength matrix;

[0118] For the initial coupling strength matrix, a coupling strength threshold is set, and elements smaller than the coupling strength threshold are set to zero to obtain a sparse coupling strength matrix;

[0119] According to the sparse coupling strength matrix, the adjacency matrix of the graph convolutional neural network is constructed. The elements of the adjacency matrix represent whether there is a coupling relationship between the corresponding channel deviation features. If there is a coupling relationship, the element value is 1, otherwise it is 0.

[0120] The multi-channel deviation features and adjacency matrix are input into the graph convolutional neural network, and the graph convolutional neural network is used to perform multi-layer convolution operations on the multi-channel deviation features to achieve feature decoupling and obtain the decoupled multi-channel deviation features.

[0121] The mutual information value can be calculated using the mutual information calculation formula for discrete or continuous variables. For example, for discrete variables, the mutual information can be calculated by the expectation of the logarithmic ratio of the joint probability distribution and the marginal probability distribution. For continuous variables, it can be calculated using the probability density function, or approximated using methods based on kernel density estimation or K-nearest neighbor. The calculation results form a symmetric matrix, where each element of the matrix represents the mutual information value between the corresponding row and column channel features.

[0122] Among them, for the initial coupling strength matrix, a coupling strength threshold is set, and elements less than the coupling strength threshold are set to zero. The purpose of obtaining a sparse coupling strength matrix is ​​to remove weak correlations and noise, thereby screening the calculated initial coupling strength matrix. Specifically, this can be achieved by setting a fixed threshold, such as setting a mutual information value based on experience; or by using an adaptive threshold method, such as determining the threshold based on the statistical distribution of matrix elements (such as mean, standard deviation) or based on graph theory indicators (such as connectivity). Mutual information values ​​less than the threshold are forced to zero, resulting in a sparse matrix containing a large number of zero elements, which only retains the strong coupling relationship between channels.

[0123] Among them, constructing the adjacency matrix of the graph convolutional neural network based on the sparse coupling strength matrix means converting the non-zero elements in the sparse coupling strength matrix into binary information representing the connection relationship to construct a graph structure suitable for the graph convolutional neural network. Specifically, this can be achieved by setting the elements greater than zero in the sparse coupling strength matrix to 1 and the elements equal to zero to 0. In this way, a binary adjacency matrix is ​​obtained, where the matrix element is 1, indicating that there is a significant coupling relationship between the corresponding channel features (i.e., the edge in the graph), and the element is 0, indicating that the coupling relationship is not significant (i.e., there is no edge in the graph).

[0124] In this approach, multi-channel deviation features and an adjacency matrix are input into a graph convolutional neural network (GCNN). The GCNN then performs multi-layer convolution operations on the multi-channel deviation features to achieve feature decoupling. Decoupled multi-channel deviation features are obtained by using the constructed graph structure to guide feature processing. This can be achieved using standard GCNN layers, such as GCN, GraphSAGE, or GAT. Multi-channel deviation features serve as graph node feature inputs, and the adjacency matrix defines the connectivity between nodes. The GCNN layer aggregates the feature information of each node and its neighbors and performs nonlinear transformations to learn the interaction patterns between channel features. By stacking multiple layers of graph convolution, the network can capture more complex inter-channel dependencies and separate mutually coupled features, ultimately outputting decoupled multi-channel deviation features. This process leverages the structured coupling between channels, making the decoupling more targeted. This technique, combined with the previously mentioned techniques for extracting multi-channel deviation features, leverages spatial correlation information between features (quantified by mutual information) to guide feature decoupling. This reduces redundancy and coupling between channels at the feature level, providing more independent input for subsequent deviation pattern recognition. This method of building a graph structure based on the coupling strength between features and performing graph convolution decoupling can more effectively handle complex spatial correlations in multi-channel data and improve the expressiveness and discrimination of features.

[0125] Through the above technical solution, the present application can effectively handle the complex coupling relationship between multi-channel deviation features. By calculating the mutual information and setting the threshold, strong correlations with practical significance between channels can be identified and retained, while filtering out weak correlations that may be caused by noise. A sparse graph structure is constructed based on the strong correlations after screening, which provides the graph convolutional neural network with effective topological information that can reflect the dependencies between key channels. The graph convolutional neural network uses this sparse graph structure to perform targeted decoupling of multi-channel deviation features, separate the mutually coupled features, and thus obtain decoupled features with stronger independence between channels. This decoupled feature can more accurately reflect the true state of each channel, reduce the interference of inter-channel coupling on subsequent deviation pattern recognition, and improve the accuracy and robustness of pattern recognition.

[0126] In some embodiments, step A6 includes:

[0127] A601. For the system-level digital twin model, establish a set of parameters to be identified, including bearing internal clearance, roller-raceway contact angle, damping coefficient, and equivalent stiffness of structural connections.

[0128] A602. Construct a parameter identification objective function aimed at minimizing the error between the model's predicted output and the actual measured data; wherein the error is the weighted root mean square error of the deviation data for each channel, and the weights of different channels are adaptively adjusted based on the signal-to-noise ratio of the corresponding channel deviation data to suppress the influence of low signal-to-noise ratio channels on parameter identification;

[0129] A603. Using the particle swarm optimization algorithm, within the preset parameter range, solve the parameter identification objective function to obtain the optimal solution for the parameter set to be identified;

[0130] A604. For each parameter in the optimized solution, calculate its confidence interval. If the confidence interval contains the parameter's initial value, the parameter is considered insensitive to bias and is removed from the set of parameters to be identified to reduce the computational complexity of model calibration.

[0131] A605. Update the optimized solutions of the remaining parameter sets to be identified to the system-level digital twin model to obtain a calibrated system-level digital twin model.

[0132] Specifically, when the multi-channel deviation data is identified as belonging to a normal deviation pattern, it indicates that the discrepancy between the current model prediction and actual measurement is primarily due to a mismatch between the model parameters and the actual state, rather than a fault. At this point, the parameter identification process is initiated to calibrate the system-level digital twin model. First, key model parameters that require adjustment are identified. These parameters directly affect the bearing's dynamic response and overall system behavior, such as the bearing's internal clearance, the actual contact angle between the roller and raceway, the damping characteristics of the bearing and structure, and the equivalent stiffness of the structural connection. These parameters constitute the parameter set to be identified. To quantify the discrepancy between the model prediction and actual measurement, an objective function is constructed whose goal is to minimize this discrepancy. This discrepancy is measured using the weighted root mean square error (RMS) of the deviation data for each channel. To account for noise in the actual measurement data and variations in data quality across sensor channels, weighting factors are introduced. These weighting factors are adaptively adjusted based on the signal-to-noise ratio (SNR) of the deviation data for each channel. Data from channels with higher SNRs are given greater weight, thereby reducing the impact of channels with lower SNRs (higher noise) on the parameter identification results and improving the robustness of the identification. The parameter identification problem is transformed into finding the minimum value of the objective function to find the optimal parameter set. This is solved using the particle swarm optimization algorithm, which searches for the global optimal solution in a multidimensional parameter space to find the parameter combination that minimizes the error between the model prediction and the actual measurement. After obtaining the optimized solution for the parameter set, the sensitivity of each parameter to the current model bias is further evaluated. By calculating the confidence interval of each parameter's optimized solution (for example, using the bootstrap method or likelihood ratio test), it is determined whether a small change in that parameter significantly affects the model output bias. If a parameter's confidence interval contains its initial setting, the parameter is considered insensitive to the current model bias, meaning that adjusting it will have little effect on reducing the error. These insensitive parameters are removed from the set of parameters to be identified, and only the bias-sensitive parameters are retained for update. This step effectively reduces the computational complexity of model calibration, avoids unnecessary adjustments to insensitive parameters, and improves the efficiency and stability of the calibration process. Finally, the optimized solutions of the selected bias-sensitive parameters to be identified are updated to the system-level digital twin model, completing the dynamic calibration of the model. The calibrated digital twin model can more accurately reflect the actual operating status of the current ball mill multi-bearing system, providing a more precise benchmark for subsequent deviation analysis and possible fault diagnosis.

[0133] In some embodiments, step A7 includes:

[0134] A701. Extract time-domain, frequency-domain, and spatial correlation features from multi-channel deviation data to construct a multidimensional feature vector. Then, use principal component analysis to reduce the dimensionality of the multidimensional feature vector to obtain the reduced-dimensional abnormal deviation feature vector.

[0135] A702. Input the reduced-dimensional abnormal deviation feature vector into a pre-trained multi-classification support vector machine to identify a set of candidate fault types that match the reduced-dimensional abnormal deviation feature vector;

[0136] A703. For each fault type in the candidate fault type set, simulate the multi-channel vibration response of that fault type under the current operating conditions based on the latest system-level digital twin model to obtain the fault characteristics predicted by the model.

[0137] A704. Calculate the cosine similarity between the reduced-dimensional abnormal deviation feature vector and each fault feature predicted by the model, and determine the fault type with the highest cosine similarity as the final fault type.

[0138] A705. Based on the determined fault type and combined with the load distribution of each bearing seat in the system-level digital twin model, locate the bearing seat with the largest load change and identify it as the faulty bearing seat.

[0139] In step A701, principal component analysis (PCA) is used to reduce the dimensionality of the multidimensional feature vectors to obtain the reduced abnormal deviation feature vectors. This can be achieved by calculating the covariance matrix, solving for eigenvalues ​​and eigenvectors, selecting the main eigenvectors, and projecting them. This is done to remove redundant information from the features, retain the main abnormal information, reduce the complexity of subsequent calculations, and improve the processing efficiency of the model.

[0140] In step A702, a multi-classification SVM model can be pre-built using a one-to-one or one-to-many strategy and trained using historical fault data. Using this pre-trained SVM model, pattern recognition can be performed based on the input abnormal feature vector. A set of candidate fault types matching the reduced abnormal deviation feature vector is identified. This means that based on the current abnormal features, the model identifies one or more possible fault types, narrowing the scope of fault diagnosis.

[0141] In step A703, fault parameters corresponding to a specific fault type, such as crack size and wear, can be set in the digital twin model, and then the model can be run to simulate the system response under the current ball mill operating parameters (such as speed and load). The fault characteristics predicted by the model can be characteristics obtained by processing the simulated time-domain vibration signal, such as the amplitude of the sensitive frequency component extracted by short-time Fourier transform. By utilizing the digital twin model, the impact of a specific fault type on the entire multi-bearing system under the current actual operating conditions can be simulated to generate theoretical fault characteristics.

[0142] In step A704, the degree of match between the observed anomaly signature and the various candidate fault signatures predicted by the model is compared. The fault type with the highest cosine similarity is determined as the final fault type. The highest similarity is chosen as the final result because the fault signature simulated by the digital twin model most closely matches the actual anomaly signature.

[0143] In step A705, the digital twin model can provide information about the system's internal state, such as the load distribution of each bearing. Specific fault types often cause changes in the internal load distribution of the damaged bearing. By analyzing the load changes caused by that fault type in the model, the specific location of the fault can be pinpointed. The bearing seat identified as the faulty bearing seat is then identified. Identifying the bearing seat with the greatest load change as the faulty bearing is a physical-based location-based approach.

[0144] Specifically, after identifying abnormal deviation patterns, the solution first performs feature extraction and dimensionality reduction on the multi-channel deviation data in step A701, extracting the core information reflecting the abnormal state from the complex raw signal. These feature vectors contain information in the time, frequency, and spatial domains, comprehensively describing the characteristics of the abnormal signal and its correlations across different channels. Dimensionality reduction removes redundancy and noise, improving feature validity. Next, in step A702, the reduced feature vectors are input into a pre-trained multi-classification support vector machine. Based on the learned patterns, the support vector machine makes a preliminary assessment of the possible fault types associated with the abnormal features and outputs a set of candidate fault types. This step leverages the pattern recognition capabilities of machine learning to rapidly narrow the range of possible faults. Then, in step A703, each fault type in the candidate set is simulated using the latest system-level digital twin model. The digital twin model simulates the impact of a specific fault type on the entire multi-bearing system under current operating conditions. Taking into account complex factors such as structural coupling, it generates a theoretical multi-channel vibration response and extracts the fault characteristics predicted by the model. This step is key to the solution, providing a fault signature benchmark based on the physical model that more closely resembles the actual complex system response. In step A704, the cosine similarity between the observed reduced-dimensional anomaly feature vector and each fault signature predicted by the model is calculated. By comparing these similarities, the degree of match between the actual anomaly and various theoretical fault patterns is quantified. The fault type with the highest similarity is selected as the final diagnosis. This matching approach based on model-simulated features overcomes the limitations of relying solely on experience or single features, improving the accuracy of fault type identification, particularly in scenarios with multiple coupled bearings and complex signals. Finally, in step A705, the faulty bearing is located based on the determined fault type and the load distribution of each bearing seat in the digital twin model. The digital twin model provides information on the load changes of each bearing in the system at the time of the fault. Certain fault types typically result in load changes on the damaged bearing. By analyzing the load changes predicted by the model, the bearing seat with the largest load change can be identified as the faulty seat. This physics-based location approach, combining the advantages of the digital twin model, improves fault location accuracy and addresses the difficulty in distinguishing the fault source in multi-bearing systems. The entire process utilizes data-driven feature analysis, machine learning pattern recognition, and physical model-based simulation and analysis capabilities to collaboratively solve the difficult problems of fault type identification and bearing positioning in the multi-spherical roller bearing system of a large ball mill.

[0145] Preferably, step A703 may include:

[0146] For each fault type in the candidate fault type set, setting a fault parameter corresponding to the fault type; wherein the fault parameter includes crack size, wear amount and imbalance amount;

[0147] Import the set fault parameters into the latest system-level digital twin model, simulate the multi-channel vibration response of the fault type under the current working conditions, and obtain the model-predicted multi-channel time-domain vibration signal (i.e., the vibration acceleration in the model prediction result);

[0148] The multi-channel time-domain vibration signal predicted by the model is subjected to short-time Fourier transform to extract the sensitive frequency components corresponding to the fault type and bearing structure parameters, and to construct the fault features predicted by the model.

[0149] For each fault type in the candidate fault type set, setting the fault parameters corresponding to that fault type means defining a set of parameters that can quantify the physical characteristics of each fault type based on predetermined possible bearing fault types, such as inner ring cracks, outer ring wear, rolling element spalling, or bearing imbalance. Specifically, for crack faults, dimensional parameters such as the length, width, and depth of the crack can be set to characterize the severity and morphology of the fault; for wear faults, parameters such as the wear depth, area, or surface roughness change can be set to characterize the degree of material loss; and for imbalance faults, the size of the unbalanced mass and its radial and angular position relative to the center of rotation can be set to characterize the degree of imbalance. These parameters are the inputs to the digital twin model for fault simulation, and they adjust the model's physical characteristics to a state that contains the specific fault.

[0150] Among them, when performing the simulation, the parameter values ​​set above, which represent the specific fault type, are updated to the corresponding parameter positions in the digital twin model. The current operating condition refers to the current operating state of the ball mill, including the cylinder speed, material filling rate, load distribution, etc. After importing the fault parameters and setting the current operating conditions, the digital twin model is run to perform dynamic simulation calculations. The simulation process simulates the dynamic response of the bearing with a specific fault and the entire ball mill system in which it is located under the current operating conditions. The vibration acceleration signals of multiple sensor positions (multi-channel) that change with time are extracted from the simulation output, that is, the multi-channel time domain vibration signals predicted by the model.

[0151] A short-time Fourier transform (STFT) is applied to the time-domain vibration signal of each channel obtained from the simulation. STFT is a time-frequency analysis method that represents how the signal's frequency components change over time by sliding a window across the signal and calculating the Fourier transform within each window. Bearing faults, particularly pitting, spalling, or cracks, typically produce energy concentrations at specific frequencies (called bearing fault characteristic frequencies, such as BPFI, BPFO, BSF, and FTF) and their harmonics or sidebands. These characteristic frequencies are calculated based on the bearing's geometric structural parameters (such as the number of rollers, inner and outer ring diameters, and contact angle) and bearing speed and are associated with specific fault types (inner ring, outer ring, rolling element, and cage). Sensitive frequency components refer to these characteristic frequencies and their surrounding frequency components that are directly related to the fault type and bearing structural parameters. Information such as the amplitude, energy, or phase of these sensitive frequency components is extracted from the STFT results. The information extracted from the sensitive frequency components of all channels is combined to form a vector or matrix, which constructs the fault signature predicted by the model. This eigenvector represents the typical characteristics of the system in the frequency domain and time-frequency domain after simulating a specific fault type in the digital twin model.

[0152] Through the above technical solution, the present application provides a specific method for simulating different candidate fault types and generating features for fault type identification using a digital twin model after identifying an abnormal deviation pattern. This method accurately imports the physical characteristics of the fault into the latest system-level digital twin model for simulation by setting physical parameters corresponding to specific fault types. By performing a short-time Fourier transform on the simulated time-domain vibration signal and specifically extracting sensitive frequency components related to the characteristic frequency of the bearing fault, a model prediction fault feature that has physical meaning and can effectively distinguish different fault types is constructed. The accuracy and recognition of the generated model prediction fault feature are ensured. As a result, the accuracy and reliability of subsequent fault type identification are improved, and the problem of inaccurate or lack of recognition of model prediction features due to the lack of specific and targeted simulation and feature extraction methods in the context of the dynamic characteristics of bearing faults and the complexity of vibration signals is solved.

[0153] refer to Figure 2 , the present application provides a spherical roller bearing fault diagnosis system for spherical roller bearing fault diagnosis of a large ball mill multi-spherical roller bearing system during the new installation stage and running-in period, the system includes a data acquisition system 1, an edge computing unit 2 and a cloud computing platform 3;

[0154] The data acquisition system 1 is used to obtain the ball mill operating parameters, and collect multi-channel raw data through a multi-channel sensor, and send the ball mill operating parameters and the multi-channel raw data to the edge computing unit 2;

[0155] The edge computing unit 2 is used to preprocess the ball mill operating parameters and multi-channel raw data to obtain preprocessed ball mill operating parameters and preprocessed multi-channel measurement data, input the preprocessed ball mill operating parameters into a pre-established system-level digital twin model, perform simulation calculations, and obtain multi-channel prediction data predicted by the model; based on the multi-channel prediction data and the preprocessed multi-channel measurement data, calculate and obtain multi-channel deviation data; extract time domain features, frequency domain features, and spatial correlation features from the multi-channel deviation data to obtain multi-channel deviation features (for the specific process, please refer to steps A1 to A4 above), and send the multi-channel deviation data and the multi-channel deviation features to the cloud computing platform 3;

[0156] The cloud computing platform 3 is used to perform deviation pattern recognition based on multi-channel deviation characteristics, and identify whether the multi-channel deviation data belongs to a normal deviation pattern or an abnormal deviation pattern. If it is recognized as a normal deviation pattern, the key parameters of the system-level digital twin model are dynamically calibrated through a parameter identification algorithm to obtain a calibrated system-level digital twin model, and the calibrated system-level digital twin model is sent to the edge computing unit 2 for model updating. If it is recognized as an abnormal deviation pattern, the multi-channel characteristics of the abnormal deviation are analyzed using the multi-channel deviation data, and combined with the latest system-level digital twin model, fault type recognition and fault bearing location are performed (the specific process can be referred to steps A5-A7 above).

[0157] Specifically, this system uses a distributed architecture to address the challenges of processing the massive amount of multi-channel sensor data from large ball mills, and the high computing resource requirements for digital twin simulation and advanced analytics. Data acquisition system 1 acquires ball mill operating parameters and raw sensor data and sends this data to edge computing unit 2. Edge computing unit 2 receives the data close to the data source and performs preprocessing to obtain clean measurement data and operating parameters. Using a pre-established digital twin model, the edge unit inputs the operating parameters for simulation, generating predicted data. By comparing the predicted data with the measured data, it calculates multi-channel deviation data. The edge unit further extracts time-domain, frequency-domain, and spatial correlation features from the deviation data and sends them to a cloud computing platform. Edge computing unit 2 handles tasks with high real-time requirements and large data volumes, alleviating the burden on the cloud platform and reducing data transmission latency. Cloud computing platform 3 receives the deviation data and features and performs deviation pattern recognition. If a normal deviation pattern is identified, the cloud platform executes a parameter identification algorithm to dynamically calibrate key parameters of the digital twin model, adapting the model to the changing state of the bearing during new installation and run-in. The calibrated model is distributed to the edge unit for updating, ensuring it uses the latest model for simulation. If an abnormal deviation pattern is identified, the cloud platform analyzes the abnormal characteristics using the deviation data and, combined with the latest digital twin model, identifies the fault type and locates the faulty bearing. The cloud platform handles computationally intensive tasks requiring global information, improving the accuracy and capabilities of fault diagnosis. This distributed architecture, combining edge and cloud computing, enables efficient and accurate fault diagnosis of spherical roller bearings in large ball mills.

[0158] The above description is merely an embodiment of the present application and is not intended to limit the scope of protection of the present application. For those skilled in the art, various modifications and variations of the present application are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A method for diagnosing faults of spherical roller bearings, used for diagnosing faults of spherical roller bearings in a large ball mill multi-spherical roller bearing system during the new installation stage and the running-in period, characterized in that: The steps of the method include: A1. Obtain the ball mill operating parameters and the multi-channel raw data collected by the multi-channel sensor, and preprocess them to obtain the preprocessed ball mill operating parameters and the preprocessed multi-channel measurement data; A2. Input the preprocessed ball mill operating parameters into the pre-established system-level digital twin model for simulation calculation to obtain multi-channel prediction data predicted by the model; A3. Calculate multi-channel deviation data based on the multi-channel prediction data and the preprocessed multi-channel measurement data; A4. Extract time-domain features, frequency-domain features, and spatial correlation features from the multi-channel deviation data to obtain multi-channel deviation features; A5. Based on the multi-channel deviation characteristics, deviation pattern recognition is performed to identify whether the multi-channel deviation data belongs to a normal deviation pattern or an abnormal deviation pattern; A6. If a normal deviation pattern is identified, dynamically calibrate the key parameters of the system-level digital twin model using a parameter identification algorithm to obtain a calibrated system-level digital twin model. A7. If an abnormal deviation pattern is identified, the multi-channel deviation data is used to analyze the multi-channel characteristics of the abnormal deviation, and combined with the latest system-level digital twin model, the fault type is identified and the faulty bearing is located.

2. A spherical roller bearing fault diagnosis method according to claim 1, characterized in that: Step A1 includes: A101. Obtain the ball mill's cylinder speed, material filling rate, ambient temperature, and lubricating oil temperature during startup, constant speed operation, and shutdown to obtain the ball mill's operating parameters. A102. Using triaxial acceleration sensors and temperature sensors installed on multiple spherical roller bearing seats, we collect vibration acceleration and temperature data from each bearing seat in three directions, generating multi-channel raw vibration acceleration data and multi-channel raw temperature data. A103. Preprocess the ball mill operating parameters and multi-channel raw data to obtain preprocessed ball mill operating parameters and preprocessed multi-channel measurement data; the preprocessing includes time synchronization processing, denoising processing and outlier removal processing.

3. A spherical roller bearing fault diagnosis method according to claim 2, characterized in that: Step A3 includes: A301. For vibration acceleration, calculate the root mean square error and spectrum amplitude difference between the preprocessed multi-channel measurement data and the preprocessed multi-channel prediction signal to obtain vibration deviation data; A302. Calculate the difference between the preprocessed multi-channel measurement data and the corresponding data in the preprocessed multi-channel prediction data for temperature, cylinder speed, and material filling rate, respectively, to obtain temperature deviation data, speed deviation data, and material filling rate deviation data; A303. The vibration deviation data, temperature deviation data, speed deviation data and material filling rate deviation data are normalized to obtain normalized vibration deviation data, normalized temperature deviation data, normalized speed deviation data and normalized material filling rate deviation data; A305. Integrate the normalized vibration deviation data, the normalized temperature deviation data, the normalized speed deviation data, and the normalized material filling rate deviation data to obtain multi-channel deviation data.

4. A spherical roller bearing fault diagnosis method according to claim 1, characterized in that: Step A4 includes: A401. For multi-channel deviation data, use the Fast Fourier Transform algorithm to calculate the spectrum of each channel deviation data and extract the spectrum amplitude, spectrum energy, spectrum centroid, and spectrum variance as frequency domain features. A402. For multi-channel deviation data, calculate the mean, variance, root mean square, kurtosis, and skewness of each channel deviation data as time domain features; A403. For the multi-channel deviation data, calculate the Pearson correlation coefficient, mutual information, and dynamic time warping distance between the deviation data of each channel as spatial correlation features; A404. Fuse the extracted frequency domain features, time domain features, and spatial correlation features to obtain multi-channel deviation features.

5. A spherical roller bearing fault diagnosis method according to claim 4, characterized in that: Step A5 includes: A501. Obtain a pre-trained initial deviation pattern classification model; the initial deviation pattern classification model includes a multi-layer stacked autoencoder and takes multi-channel deviation features as input; A502. Use the dynamic time warping algorithm to calculate the similarity between the real-time multi-channel deviation features and the historical normal deviation pattern features. If the similarity falls below a preset similarity threshold, the operating condition is determined to have changed. The initial deviation pattern classification model is then fine-tuned using a transfer learning algorithm to obtain a deviation pattern classification model that is adaptive to the operating condition. A503. Calculate the coupling strength of each channel deviation feature using the mutual information algorithm, construct a coupling strength matrix, and decouple the multi-channel deviation features using a graph convolutional neural network to obtain the decoupled multi-channel deviation features. A504. Input the decoupled multi-channel deviation features into the working condition-adaptive deviation pattern classification model, and use the Softmax classifier to calculate the probability that the multi-channel deviation data belongs to an abnormal deviation pattern. If the probability of the abnormal deviation pattern is greater than the preset probability threshold, it is judged as an abnormal deviation pattern; otherwise, it is judged as a normal deviation pattern.

6. A spherical roller bearing fault diagnosis method according to claim 5, characterized in that: Step A503 includes: The mutual information algorithm is used to calculate the mutual information value between every two channel deviation features in each channel deviation feature to construct the initial coupling strength matrix; For the initial coupling strength matrix, a coupling strength threshold is set, and elements smaller than the coupling strength threshold are set to zero to obtain a sparse coupling strength matrix; According to the sparse coupling strength matrix, the adjacency matrix of the graph convolutional neural network is constructed. The elements of the adjacency matrix represent whether there is a coupling relationship between the corresponding channel deviation features. If there is a coupling relationship, the element value is 1, otherwise it is 0. The multi-channel deviation features and adjacency matrix are input into the graph convolutional neural network, and the graph convolutional neural network is used to perform multi-layer convolution operations on the multi-channel deviation features to achieve feature decoupling and obtain the decoupled multi-channel deviation features.

7. A spherical roller bearing fault diagnosis method according to claim 1, characterized in that: Step A6 includes: A601. For the system-level digital twin model, establish a set of parameters to be identified, including bearing internal clearance, roller-raceway contact angle, damping coefficient, and equivalent stiffness of structural connections. A602. Construct a parameter identification objective function aimed at minimizing the error between the model's predicted output and the actual measured data; wherein the error is the weighted root mean square error of the deviation data for each channel, and the weights of different channels are adaptively adjusted based on the signal-to-noise ratio of the corresponding channel deviation data to suppress the influence of low signal-to-noise ratio channels on parameter identification; A603. Using the particle swarm optimization algorithm, within the preset parameter range, solve the parameter identification objective function to obtain the optimal solution for the parameter set to be identified; A604. For each parameter in the optimized solution, calculate its confidence interval. If the confidence interval contains the parameter's initial value, the parameter is considered insensitive to bias and is removed from the set of parameters to be identified to reduce the computational complexity of model calibration. A605. Update the optimized solutions of the remaining parameter sets to be identified to the system-level digital twin model to obtain a calibrated system-level digital twin model.

8. A spherical roller bearing fault diagnosis method according to claim 4, characterized in that: Step A7 includes: A701. Extract time-domain, frequency-domain, and spatial correlation features from multi-channel deviation data to construct a multidimensional feature vector. Then, use principal component analysis to reduce the dimensionality of the multidimensional feature vector to obtain the reduced-dimensional abnormal deviation feature vector. A702. Input the reduced-dimensional abnormal deviation feature vector into a pre-trained multi-classification support vector machine to identify a set of candidate fault types that match the reduced-dimensional abnormal deviation feature vector; A703. For each fault type in the candidate fault type set, simulate the multi-channel vibration response of that fault type under the current operating conditions based on the latest system-level digital twin model to obtain the fault characteristics predicted by the model. A704. Calculate the cosine similarity between the reduced-dimensional abnormal deviation feature vector and each fault feature predicted by the model, and determine the fault type with the highest cosine similarity as the final fault type. A705. Based on the determined fault type and combined with the load distribution of each bearing seat in the system-level digital twin model, locate the bearing seat with the largest load change and identify it as the faulty bearing seat.

9. A spherical roller bearing fault diagnosis method according to claim 8, characterized in that: Step A703 includes: For each fault type in the candidate fault type set, setting a fault parameter corresponding to the fault type; wherein the fault parameter includes crack size, wear amount and imbalance amount; Import the set fault parameters into the latest system-level digital twin model, simulate the multi-channel vibration response of the fault type under the current working conditions, and obtain the multi-channel time-domain vibration signal predicted by the model; The multi-channel time-domain vibration signal predicted by the model is subjected to short-time Fourier transform to extract the sensitive frequency components corresponding to the fault type and bearing structure parameters, and to construct the fault features predicted by the model.

10. A spherical roller bearing fault diagnosis system for large ball mill multi-spherical roller bearing systems during the new installation stage and running-in period, characterized by: The system includes a data acquisition system, an edge computing unit, and a cloud computing platform; The data acquisition system is used to obtain the ball mill operating parameters, collect multi-channel raw data through a multi-channel sensor, and send the ball mill operating parameters and the multi-channel raw data to the edge computing unit; The edge computing unit is used to preprocess the ball mill operating parameters and multi-channel raw data to obtain preprocessed ball mill operating parameters and preprocessed multi-channel measurement data, input the preprocessed ball mill operating parameters into a pre-established system-level digital twin model, perform simulation calculations, and obtain multi-channel prediction data predicted by the model; calculate multi-channel deviation data based on the multi-channel prediction data and the preprocessed multi-channel measurement data, extract time domain features, frequency domain features, and spatial correlation features from the multi-channel deviation data to obtain multi-channel deviation features, and send the multi-channel deviation data and the multi-channel deviation features to the cloud computing platform; The cloud computing platform is used to perform deviation pattern recognition based on multi-channel deviation characteristics, and identify whether the multi-channel deviation data belongs to a normal deviation pattern or an abnormal deviation pattern. If it is recognized as a normal deviation pattern, the key parameters of the system-level digital twin model are dynamically calibrated through a parameter identification algorithm to obtain a calibrated system-level digital twin model, and the calibrated system-level digital twin model is sent to the edge computing unit for model updating. If it is recognized as an abnormal deviation pattern, the multi-channel characteristics of the abnormal deviation are analyzed using the multi-channel deviation data, and combined with the latest system-level digital twin model, fault type identification and fault bearing location are performed.

Images