A method and system for monitoring the rubbing condition of sliding bearings

Abstract

This invention discloses a method and system for monitoring the rubbing condition of sliding bearings, relating to the field of sliding bearing monitoring technology. The method includes: performing time-frequency domain joint analysis on the operating state parameters of the sliding bearing to construct a multi-dimensional fusion feature matrix; constructing a rubbing condition identification model, identifying the rubbing condition on the multi-dimensional fusion feature matrix, and obtaining the rubbing condition identification result; based on the real-time operating parameters of the sliding bearing, constructing a dynamic adjustment model for the early warning threshold, and combining a gradient descent algorithm to iteratively update the weights of the feature extraction layer in the rubbing condition identification model to obtain a dynamic early warning threshold adapted to the current operating condition; based on the rubbing condition identification result and the dynamic early warning threshold, combining multi-level hash verification, constructing a monitoring state record chain, and storing it on an industrial blockchain. This invention achieves high-precision identification of the rubbing condition of sliding bearings, overcoming the bottlenecks of existing technologies in terms of operating condition adaptability, model persistence, and data reliability.

Description

Technical Field

[0001] This invention relates to the field of sliding bearing monitoring technology, and more specifically, to a method and system for monitoring the rubbing condition of sliding bearings. Background Technology

[0002] As critical support components in major rotating machinery such as steam turbines, hydroelectric generators, and ship propulsion systems, sliding bearings operate under complex conditions of high speed, heavy load, and extreme temperature differences for extended periods. They are prone to abnormal rubbing failures between the journal and the bearing bush due to factors such as lubrication failure, shaft misalignment, component wear, or changes in assembly clearances. If such failures are not identified and addressed promptly, they can rapidly escalate into serious accidents such as localized bearing failure, increased shaft vibration, or even equipment shutdown, resulting in significant economic losses and safety risks. Therefore, constructing a high-precision, robust, and traceable sliding bearing rubbing condition monitoring system has become an urgent requirement for ensuring the safe and reliable operation of high-end equipment.

[0003] Current sliding bearing condition monitoring mainly relies on threshold alarms or trend analysis of sensor signals such as vibration, temperature, and oil film pressure. These early warning criteria are largely based on empirical settings or laboratory calibration, lacking adaptability to dynamic changes in actual operating conditions. Under changing conditions, fault characteristics are easily drowned out by background noise or drift, leading to frequent false alarms or missed alarms with fixed threshold methods. Although deep learning technology has been widely applied in fault diagnosis in recent years, achieving automatic extraction and classification of nonlinear fault characteristics by constructing models such as CNN (Convolutional Neural Network) and LSTM (Long Short-Term Memory Network), most existing methods employ offline training and fixed deployment. Once model parameters are determined, they are no longer updated, making it difficult to cope with long-term evolution processes such as equipment performance degradation and environmental disturbances, resulting in a gradual decline in diagnostic performance over time.

[0004] Furthermore, existing monitoring systems generally employ a centralized data storage architecture, with status data, diagnostic results, and early warning records stored centrally on local servers or in monitoring centers. This poses risks of data tampering, loss, or single points of failure, making it difficult to meet the high reliability, auditability, and non-repudiation requirements of industries such as power, energy, and rail transportation for critical equipment operational data. Especially in the event of equipment failure or safety incidents, the lack of a legally valid chain of evidence hinders the ability to trace responsibility and conduct causal analysis.

[0005] Therefore, existing sliding bearing rubbing condition monitoring technologies still face core challenges such as poor adaptability to operating conditions, rigid early warning mechanisms, lack of model update capabilities, and insufficient data reliability.

[0006] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0007] In view of the problems in related technologies, the present invention proposes a method and system for monitoring the rubbing condition of sliding bearings, so as to overcome the above-mentioned technical problems existing in the existing related technologies.

[0008] Therefore, the specific technical solution adopted by the present invention is as follows: According to one aspect of the present invention, a method for monitoring the rubbing condition of a sliding bearing is provided, the method comprising the following steps: S1. Perform time-frequency domain joint analysis on the operating parameters of the sliding bearing, extract the rubbing-sensitive feature set, and construct a multi-dimensional fusion feature matrix; S2. Combining multi-head self-attention mechanism and hybrid neural network, construct a collision state recognition model, and use the collision state recognition model to perform collision state recognition on multi-dimensional fused feature matrix to obtain collision state recognition results. S3. Based on the real-time operating parameters of the sliding bearing, a dynamic adjustment model for the early warning threshold is constructed. Combined with the gradient descent algorithm, the weights of the feature extraction layer in the collision state recognition model are iteratively updated to obtain a dynamic early warning threshold that adapts to the current operating conditions. S4. Based on the collision status recognition results and dynamic early warning threshold, generate corresponding early warning instruction information, combine multi-level hash verification to construct a monitoring status record chain, and distribute and store the monitoring status record chain on the industrial blockchain.

[0009] Furthermore, the operating state parameters of the sliding bearing are subjected to joint time-frequency domain analysis to extract the rubbing-sensitive feature set and construct a multi-dimensional fusion feature matrix, including the following steps: S11. Obtain the operating state parameters of the sliding bearing, wherein the operating state parameters include dynamic signal parameters and steady-state signal parameters; S12. Using the ensemble empirical mode decomposition algorithm, the dynamic signal parameters are preprocessed to obtain the preprocessed dynamic signal parameters. S13. Based on the outlier detection algorithm, the steady-state signal parameters are preprocessed to obtain the preprocessed steady-state signal parameters. S14. Using a time-frequency domain joint analysis algorithm, the collision-sensitive features of the preprocessed dynamic signal parameters and the preprocessed steady-state signal parameters are extracted and filtered to obtain a collision-sensitive feature set. S15. Organize the rubbing-sensitive feature set in a structured manner according to a preset dimension, with time-series samples as rows and feature indicators as columns. Add timestamps, equipment numbers, and operating parameters as auxiliary information, and label the rubbing status to obtain a multi-dimensional fusion feature matrix.

[0010] Furthermore, a joint time-frequency domain analysis algorithm is employed to extract rubbing-sensitive features from the preprocessed dynamic signal parameters and the preprocessed steady-state signal parameters, and then filter them to obtain a rubbing-sensitive feature set, including the following steps: S131. Perform time-frequency transformation on the preprocessed dynamic signal parameters and the preprocessed steady-state signal parameters respectively, and combine the short-time Fourier transform algorithm to convert the time domain into the time-frequency domain to obtain the dynamic time-frequency distribution map and the steady-state time-frequency distribution map. S132. Extract time-domain, frequency-domain, and time-frequency-domain features from the dynamic time-frequency distribution map and the steady-state time-frequency distribution map to obtain the dynamic multi-domain feature set and the steady-state multi-domain feature set; S133. Align and fuse the dynamic multi-domain feature set with the steady-state multi-domain feature set to construct an initial high-dimensional feature pool; S134. A sparse feature selection algorithm is used to filter the high-dimensional feature pool to obtain a sensitive feature subset, and the sensitive feature subset is normalized to obtain the rubbing sensitive feature set.

[0011] Furthermore, by combining a multi-head self-attention mechanism and a hybrid neural network, a collision state recognition model is constructed. This model is then used to identify the collision state from a multi-dimensional fused feature matrix. The resulting collision state recognition results include the following steps: S21. A hybrid neural network model is constructed by using graph convolutional neural networks and gated recurrent units as the core and integrating multi-head self-attention mechanism. S22. Obtain the historical multi-dimensional fusion feature matrix as a training sample set, and train the hybrid neural network model based on the training sample set to obtain the rubbing state recognition model. S23. Use the collision state recognition model to perform collision state recognition on the multi-dimensional fusion feature matrix to obtain collision state recognition results including collision state category, recognition confidence, timestamp and contribution features.

[0012] Furthermore, the historical multi-dimensional fusion feature matrix is ​​obtained as a training sample set. Based on the training sample set, the hybrid neural network model is trained to obtain the rubbing state recognition model, including the following steps: S221. The weight initialization algorithm is used to initialize the parameters of the hybrid neural network model, and the gradient descent optimization algorithm, cross-entropy loss function, learning rate and its decay strategy are configured. Regularization layer and batch normalization layer are introduced to improve the stability and generalization ability of model training. S222. Obtain the historical multi-dimensional fusion feature matrix as the training sample set, normalize the sample set, and divide it into training set, validation set and test set according to the preset ratio. S223. The training set responds to the hybrid neural network model, calculates the output probability distribution through forward propagation, calculates the loss value based on the true label, and updates the parameters of the hybrid neural network model based on the backpropagation algorithm. S224. Based on the recognition accuracy and comprehensive evaluation index on the validation set, the parameters of the updated hybrid neural network model are updated a second time. S225. The performance of the trained model is verified using the test set. The hybrid neural network model after the second update is evaluated by confusion matrix, receiver operation feature curve and confidence analysis. The collision state recognition model is optimized based on the evaluation results.

[0013] Furthermore, based on the real-time operating parameters of the sliding bearing, a dynamic adjustment model for the early warning threshold is constructed. Combined with the gradient descent algorithm, the weights of the feature extraction layer in the rubbing condition identification model are iteratively updated to obtain a dynamic early warning threshold adapted to the current operating condition. This includes the following steps: S31. Perform data cleaning and normalization on the real-time operating parameters of the acquired sliding bearings, and construct an operating condition mapping database by combining historical monitoring data. S32. Using a fully connected neural network as the core, a dynamic adjustment model for the warning threshold is constructed. The input of the dynamic adjustment model for the warning threshold is the current working condition parameter vector, and the output of the dynamic adjustment model for the warning threshold is the confidence adjustment coefficient and feature channel weighted vector of each collision state category. S33. Based on the working condition mapping database, and combined with the preset loss function and gradient descent algorithm, train the early warning threshold dynamic adjustment model. S34. Based on the pre-trained warning threshold, dynamically adjust the feature weighting vector output by the model, update the feature extraction layer weights of the collision state recognition model, and combine the updated output probability distribution of the collision state recognition model with the working condition mapping database to dynamically calculate the classification decision boundary under the current working condition, thereby obtaining a dynamic pre-warning threshold adapted to the current working condition.

[0014] Furthermore, based on the trained warning threshold, the feature weighting vector output by the model is dynamically adjusted to update the weights of the feature extraction layer of the collision condition recognition model. Then, combining the updated output probability distribution of the collision condition recognition model with the working condition mapping database, the classification decision boundary under the current working condition is dynamically calculated to obtain a dynamic warning threshold adapted to the current working condition. This includes the following steps: S341. Dynamically adjust the feature channel weighted vector output by the trained warning threshold and apply it to the input of the feature extraction layer in the collision state recognition model to perform channel weighted fusion on the multi-dimensional fusion feature matrix to obtain the weighted feature input. S342. Input the weighted feature input into the collision state recognition model, perform forward inference, obtain the initial state recognition result and the corresponding output probability distribution, and calculate the prediction loss of the collision state recognition model by combining the real-time annotation information. S343. Based on the prediction loss of the bump and rub state recognition model, the gradient descent algorithm is used to perform backpropagation to adjust the convolution kernel weights and gating parameters in the feature extraction layer of the bump and rub state recognition model, so as to obtain the updated bump and rub state recognition model. S344. Reapply the updated collision state recognition model to the weighted feature input to obtain the optimized output probability distribution. S345. Extract the historical normal sample output probability distribution of the corresponding working condition interval from the working condition mapping database, calculate the corresponding mean and variance, and obtain the confidence tolerance interval. S346. Based on the optimized output probability distribution and confidence tolerance interval, dynamically calculate the classification decision boundary of each collision level under the current state to obtain a dynamic early warning threshold adapted to the current working condition.

[0015] Furthermore, based on the collision status recognition results and dynamic early warning thresholds, corresponding early warning instruction information is generated, and combined with multi-level hash verification, a monitoring status record chain is constructed and distributed and stored in the industrial blockchain, including the following steps: S41. Compare the dynamic warning threshold with the collision state identification result, determine the current warning level according to the preset graded warning rules, and obtain the corresponding warning instruction information. The warning instruction information includes the warning level, identification state category, confidence value, trigger timestamp, and suggested handling measures. S42. Extract the multi-dimensional fusion feature matrix, real-time operating parameters and corresponding early warning instructions within the current monitoring period, and associate them with the equipment number to construct a status data unit; S43. Based on the state data unit, construct a monitoring state record chain that coordinates the metadata chain and the data digest chain, and combine digital signature, improved consensus mechanism and multi-level hash verification to distribute and store the monitoring state record chain on the industrial blockchain.

[0016] Furthermore, based on state data units, a monitoring state recording chain is constructed that coordinates the metadata chain and the data digest chain. Combined with digital signatures, an improved consensus mechanism, and multi-level hash verification, the monitoring state recording chain is distributed and stored on an industrial blockchain, including the following steps: S431. Based on the time-series correlation of the state data unit, construct a metadata chain and a data digest chain respectively, and combine the hash pointer to associate the metadata chain and the data digest chain to obtain the monitoring state record chain; S432. Perform multi-level hash operations on the monitoring status record chain to obtain multi-level hash values; S433. Use the device's private key to digitally sign the multi-level hash value, and verify and reach consensus on the digitally signed multi-level hash value based on an improved consensus mechanism to obtain a new block; S434. Store the new block to the industrial blockchain to achieve distributed evidence storage of the monitoring status record chain.

[0017] According to another aspect of the present invention, a sliding bearing rubbing condition monitoring system is provided, the system comprising: a multi-dimensional fusion feature matrix construction module, a rubbing condition identification module, a model threshold collaborative optimization module, and a monitoring condition record chain construction and storage module; The multi-dimensional fusion feature matrix construction module is used to perform time-frequency domain joint analysis and processing of the operating state parameters of sliding bearings, extract the rubbing-sensitive feature set, and construct a multi-dimensional fusion feature matrix. The collision state recognition module is used to combine multi-head self-attention mechanism and hybrid neural network to build collision state recognition model, and use the collision state recognition model to perform collision state recognition on multi-dimensional fused feature matrix to obtain collision state recognition result. The model threshold collaborative optimization module is used to construct a dynamic adjustment model for the early warning threshold based on the real-time operating parameters of the sliding bearing, and combined with the gradient descent algorithm, iteratively update the weights of the feature extraction layer in the collision and rubbing state recognition model to obtain a dynamic early warning threshold that adapts to the current operating conditions. The monitoring status record chain is constructed and stored in a storage module. Based on the contact status identification results and dynamic warning thresholds, it generates corresponding warning instruction information. Combined with multi-level hash verification, it constructs the monitoring status record chain and distributes the monitoring status record chain on the industrial blockchain.

[0018] The beneficial effects of this invention are as follows: 1. This invention achieves high-precision and robust identification of sliding bearing rubbing conditions. Furthermore, through innovative mechanisms such as dynamic threshold adjustment, online model optimization, and blockchain-based trusted evidence storage, it constructs an adaptive, evolvable, traceable, and highly secure intelligent monitoring system. This effectively overcomes the bottlenecks of existing technologies in terms of working condition adaptability, model continuity, and data reliability, demonstrating outstanding technological advancement and broad application prospects.

[0019] 2. This invention extracts multi-dimensional rubbing-sensitive features by employing a time-frequency domain joint analysis algorithm, and combines a multi-head self-attention mechanism with a hybrid neural network model for state recognition. It effectively integrates local feature extraction, temporal dependency modeling, and key segment focusing capabilities, significantly enhancing the sensitivity and classification accuracy of early faults such as minor rubbing and intermittent rubbing. It also solves the problem of misjudgment and missed judgment in low signal-to-noise ratio environments using traditional methods.

[0020] 3. This invention constructs a dynamic adjustment model for the early warning threshold and optimizes the model by combining it with the gradient descent algorithm. This enables the early warning threshold to adapt to changes in the operating conditions of the sliding bearing in real time, thereby improving the accuracy and response speed of the early warning and reducing the possibility of false alarms and missed alarms.

[0021] 4. This invention uses multi-level hash verification and industrial blockchain technology to construct a monitoring status record chain, realizing secure data storage and tamper-proofing. Through distributed evidence storage, it ensures the integrity and traceability of monitoring data, providing a reliable data management solution. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of a sliding bearing rubbing condition monitoring method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a sliding bearing rubbing condition monitoring system according to an embodiment of the present invention.

[0024] In the picture: 1. Multi-dimensional fusion feature matrix construction module; 2. Touch and rubbing state recognition module; 3. Model threshold collaborative optimization module; 4. Monitoring state record chain construction and storage module. Detailed Implementation

[0025] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0026] According to an embodiment of the present invention, a method and system for monitoring the rubbing condition of sliding bearings are provided.

[0027] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, the sliding bearing rubbing condition monitoring method according to an embodiment of the present invention includes the following steps: S1. Perform time-frequency domain joint analysis on the operating parameters of the sliding bearing, extract the rubbing-sensitive feature set, and construct a multi-dimensional fusion feature matrix.

[0028] It should be explained that by combining time-frequency domain joint analysis algorithms, the complex characteristics of sliding bearings in the early stage of rubbing, such as weak impact, non-stationary vibration and local temperature rise, are fully captured, overcoming the limitation of single-domain analysis being susceptible to interference from operating condition fluctuations. The constructed multi-dimensional fusion feature matrix retains the spatiotemporal correlation of the original signal, significantly improving the sensitivity and robustness of rubbing fault identification.

[0029] In this optional embodiment, the operating state parameters of the sliding bearing are subjected to joint time-frequency domain analysis to extract the rubbing-sensitive feature set and construct a multi-dimensional fusion feature matrix in the following steps: S11. Obtain the operating state parameters of the sliding bearing, including dynamic signal parameters and steady-state signal parameters; Specifically, dynamic signal parameters include vibration acceleration signals and acoustic emission signals; steady-state signal parameters include bearing surface temperature, lubricating oil pressure, and lubricating oil flow rate.

[0030] S12. Using the ensemble empirical mode decomposition algorithm, the dynamic signal parameters are preprocessed to obtain the preprocessed dynamic signal parameters. S13. Based on the outlier detection algorithm, the steady-state signal parameters are preprocessed to obtain the preprocessed steady-state signal parameters. S14. Using a time-frequency domain joint analysis algorithm, the collision-sensitive features of the preprocessed dynamic signal parameters and the preprocessed steady-state signal parameters are extracted and filtered to obtain a collision-sensitive feature set. S15. Organize the rubbing-sensitive feature set in a structured manner according to a preset dimension, with time-series samples as rows and feature indicators as columns. Add timestamps, equipment numbers, and operating parameters as auxiliary information, and label the rubbing status to obtain a multi-dimensional fusion feature matrix.

[0031] It should be explained that a two-dimensional matrix of the monitoring sample sequence and the sensitive feature dimension set is used as the core data structure. The row and column dimensions of the matrix correspond to the time-series monitoring samples and the optimized sensitive features, respectively, so as to achieve the unity of temporal correlation and dimensional integrity of the collision features.

[0032] Based on continuous monitoring of the sliding bearing's operating status, feature data is acquired at a sampling frequency of 100ms / set. Every 10 sets of continuously acquired feature data constitute one monitoring sample sequence, which serves as one row of the matrix. By arranging the sample sequences in a temporal sequence, the matrix row dimension contains the temporal evolution information of the rubbing features, adapting to the temporal feature extraction requirements of the subsequent GRU module.

[0033] The selected and optimized impact-sensitive features include time-domain peak factor, frequency-domain spectral kurtosis, time-frequency domain wavelet packet energy entropy, and bearing temperature deviation. These impact-sensitive features are used as matrix columns, with the number of features controlled to 8-12. This preserves key characterization information of impact-related faults while avoiding increased computational load due to dimensional redundancy.

[0034] The column dimensions of the two-dimensional matrix are expanded with auxiliary information fields, including identification fields, time-series identification fields, operating condition association fields, and status label annotations. The identification field includes a unique equipment identifier, enabling the binding and traceability of the feature matrix to specific sliding bearings. The time-series identification field includes data acquisition timestamps, ensuring the time-series accuracy of the monitoring data. The operating condition association field includes real-time load, speed, and other operating condition parameters, establishing a mapping relationship between features and operating conditions. The status label annotation includes a sliding bearing monitoring sample matrix for known operating states, annotated with status labels using one-hot encoding. The labels are divided into four categories: normal, slight rubbing, moderate rubbing, and severe rubbing. The annotated sample matrix serves as a training dataset for subsequent parameter training and performance optimization of the sliding bearing rubbing condition recognition model.

[0035] In this optional embodiment, a joint time-frequency domain analysis algorithm is used to extract rubbing-sensitive features from the preprocessed dynamic signal parameters and the preprocessed steady-state signal parameters, and then filter them to obtain a rubbing-sensitive feature set, including the following steps: S131. Perform time-frequency transformation on the preprocessed dynamic signal parameters and the preprocessed steady-state signal parameters respectively, and combine the short-time Fourier transform algorithm to convert the time domain into the time-frequency domain to obtain the dynamic time-frequency distribution map and the steady-state time-frequency distribution map. S132. Extract time-domain, frequency-domain, and time-frequency-domain features from the dynamic time-frequency distribution map and the steady-state time-frequency distribution map to obtain the dynamic multi-domain feature set and the steady-state multi-domain feature set; Specifically, time-domain features are extracted from the preprocessed vibration and acoustic emission dynamic signals to reflect the intensity and regularity of the impact; frequency-domain features are extracted using Fast Fourier Transform (FFT) to identify the characteristic frequency shifts caused by the impact; and time-frequency domain features are extracted using Wavelet Packet Transform (WPT) to capture the time-frequency coupling characteristics of instantaneous impact. For the preprocessed steady-state signals such as bearing temperature and lubricating oil pressure, trend features are extracted using sliding window analysis to reflect the gradual change in steady-state parameters caused by continuous impact.

[0036] S133. Align and fuse the dynamic multi-domain feature set with the steady-state multi-domain feature set to construct an initial high-dimensional feature pool; S134. A sparse feature selection algorithm is used to filter the high-dimensional feature pool to obtain a sensitive feature subset, and the sensitive feature subset is normalized to obtain the rubbing sensitive feature set.

[0037] S2. Combining multi-head self-attention mechanism and hybrid neural network, a collision state recognition model is constructed, and the collision state recognition model is used to perform collision state recognition on multi-dimensional fused feature matrix to obtain collision state recognition results.

[0038] It should be explained that by combining CNN (Graph Convolutional Neural Network), GRU (Gated Recurrent Unit), and multi-head self-attention mechanism, a collision and rubbing condition recognition model is constructed. This model can not only effectively extract the spatial local patterns and temporal dynamic features of the feature matrix, but also automatically weight important feature dimensions, significantly improving the accuracy and robustness of collision and rubbing condition recognition. At the same time, the model parameters are iteratively optimized based on historical data to ensure that it still has good adaptability and stability under varying working conditions, providing reliable technical support for the health monitoring of sliding bearings.

[0039] In this optional embodiment, a collision state recognition model is constructed by combining a multi-head self-attention mechanism and a hybrid neural network. The collision state recognition model is then used to identify the collision state of the multi-dimensional fused feature matrix to obtain the collision state recognition result, which includes the following steps: S21. A hybrid neural network model is constructed by using graph convolutional neural networks and gated recurrent units as the core and integrating multi-head self-attention mechanism. Specifically, the hybrid neural network model includes a CNN feature extraction layer, a GRU temporal modeling layer, a multi-head self-attention layer, and a classification output layer connected in sequence. The CNN feature extraction layer uses three one-dimensional convolutional layers (with kernel sizes of 3×3, 3×3, and 2×2, and a stride of 1) and two max pooling layers (with a pooling kernel of 2×2) to extract spatial correlation features from a multi-dimensional fused feature matrix. The GRU temporal modeling layer has two hidden layers (with 128 and 64 neurons respectively) and uses reset and update gate mechanisms to capture the temporal evolution of rubbing signals, including the duration and frequency of rubbing impacts. The multi-head self-attention layer is configured with four parallel attention heads and uses scaled dot product attention calculation to strengthen the weights of rubbing-sensitive features and suppress interfering features. The classification output layer is connected to a fully connected layer (with 32 neurons) and a Softmax activation function to output four probability values: normal, mild rubbing, moderate rubbing, and severe rubbing. The layers are matched in dimension and the information is transferred completely through fully connected mapping.

[0040] S22. Obtain the historical multi-dimensional fusion feature matrix as a training sample set, and train the hybrid neural network model based on the training sample set to obtain the rubbing state recognition model. S23. Use the collision state recognition model to perform collision state recognition on the multi-dimensional fusion feature matrix to obtain collision state recognition results including collision state category, recognition confidence, timestamp and contribution features.

[0041] In this optional embodiment, obtaining a historical multi-dimensional fusion feature matrix as a training sample set, and training a hybrid neural network model based on the training sample set to obtain a rubbing state recognition model includes the following steps: S221. The weight initialization algorithm is used to initialize the parameters of the hybrid neural network model, and the gradient descent optimization algorithm, cross-entropy loss function, learning rate and its decay strategy are configured. Regularization layer and batch normalization layer are introduced to improve the stability and generalization ability of model training. Specifically, hybrid neural network models need to match initialization algorithms to the characteristics of different sub-networks. CNN convolutional layers use He normal initialization and combine it with the ReLU activation function to avoid gradient vanishing. GRU recurrent layers use Xavier uniform initialization and combine it with the tanh activation function to ensure the numerical stability of forward propagation and backward gradient. Fully connected layers use orthogonal initialization to reduce redundant correlations between parameters and improve feature representation ability.

[0042] S222. Obtain the historical multi-dimensional fusion feature matrix as the training sample set, normalize the sample set, and divide it into training set, validation set and test set according to the preset ratio. S223. The training set responds to the hybrid neural network model, calculates the output probability distribution through forward propagation, calculates the loss value based on the true label, and updates the parameters of the hybrid neural network model based on the backpropagation algorithm. S224. Based on the recognition accuracy and comprehensive evaluation index on the validation set, the parameters of the updated hybrid neural network model are updated a second time. S225. The performance of the trained model is verified using the test set. The hybrid neural network model after the second update is evaluated by confusion matrix, receiver operation feature curve and confidence analysis. The collision state recognition model is optimized based on the evaluation results.

[0043] It should be explained that after the training set sample set is input, the CNN layer extracts spatial features through 3×3 convolution kernels, the GRU layer captures temporal features through a gating mechanism, and the fully connected layer maps the fused features into the probability distribution of four types of collision states.

[0044] S3. Based on the real-time operating parameters of the sliding bearing, a dynamic adjustment model for the early warning threshold is constructed. Combined with the gradient descent algorithm, the weights of the feature extraction layer in the collision state recognition model are iteratively updated to obtain a dynamic early warning threshold that adapts to the current operating conditions.

[0045] It should be explained that by setting up a closed loop of working condition perception, feature weighting, parameter fine-tuning and threshold resetting, the early warning threshold can be dynamically evolved. This not only improves the robustness of the model under changing working conditions, but also avoids the maintenance costs caused by frequent offline retraining.

[0046] In this optional embodiment, based on the feature weighted vector output by the dynamically adjusted warning threshold model after training, a dynamic warning threshold adjustment model is constructed for the real-time operating parameters based on the sliding bearing. Combined with the gradient descent algorithm, the weights of the feature extraction layer in the collision and rubbing state recognition model are iteratively updated to obtain a dynamic warning threshold adapted to the current operating condition. This includes the following steps: S31. Perform data cleaning and normalization on the real-time operating parameters of the acquired sliding bearings, and construct an operating condition mapping database by combining historical monitoring data. S32. Using a fully connected neural network as the core, a dynamic adjustment model for the warning threshold is constructed. The input of the dynamic adjustment model for the warning threshold is the current working condition parameter vector, and the output of the dynamic adjustment model for the warning threshold is the confidence adjustment coefficient and feature channel weighted vector of each collision state category. S33. Based on the working condition mapping database, and combined with the preset loss function and gradient descent algorithm, train the early warning threshold dynamic adjustment model. S34. Based on the pre-trained warning threshold, dynamically adjust the feature weighting vector output by the model, update the feature extraction layer weights of the collision state recognition model, and combine the updated output probability distribution of the collision state recognition model with the working condition mapping database to dynamically calculate the classification decision boundary under the current working condition, thereby obtaining a dynamic pre-warning threshold adapted to the current working condition.

[0047] It should be explained that the model threshold co-optimization includes the following steps: Step 1: Real-time operating condition parameter acquisition and feature construction.

[0048] Real-time operating parameters of the sliding bearing during operation are collected, including spindle speed, load torque, lubricating oil temperature and oil supply pressure. The parameters are preprocessed and normalized to construct a standardized real-time operating feature vector.

[0049] Step 2: Historical data analysis and initial threshold setting.

[0050] A working condition-state mapping database is established based on historical monitoring data. The distribution characteristics of rubbing-sensitive features are statistically analyzed according to different working condition intervals. The dynamic baseline and fluctuation range of each feature under normal conditions are determined as a reference for the initial warning threshold.

[0051] Step 3: Construction of an adaptive weight adjustment mechanism for operating conditions. The real-time operating condition feature vector is input into a lightweight fully connected network or an interpolation lookup table is used to generate feature weighting coefficients for the corresponding operating conditions. This dynamically adjusts the input weights of the feature extraction layer in the collision and friction state recognition model, enhancing its responsiveness to sensitive feature channels under the current operating conditions.

[0052] Step 4: Online fine-tuning and optimization of model parameters.

[0053] During model inference, the gradient of the feature extraction layer parameters is calculated by backpropagation using the gradient descent algorithm, combined with the current recognition result and the preset loss function with state misjudgment penalty term. The CNN convolution kernel weights and GRU gate parameters are updated incrementally with small steps to achieve continuous optimization of feature extraction capability. Step 5: Calculation and output of dynamic early warning threshold.

[0054] Based on the state probability distribution output by the updated model, the classification boundaries of various collision states are recalculated. Combining the confidence interval and sliding window statistical method, a dynamic early warning threshold matching the current working condition is generated.

[0055] In this optional embodiment, the feature weighting vector output by the model is dynamically adjusted based on the trained warning threshold to update the weights of the feature extraction layer of the collision condition recognition model. The updated output probability distribution of the collision condition recognition model is then combined with the working condition mapping database to dynamically calculate the classification decision boundary under the current working condition, thus obtaining a dynamic warning threshold adapted to the current working condition. This includes the following steps: S341. Dynamically adjust the feature channel weighted vector output by the trained warning threshold and apply it to the input of the feature extraction layer in the collision state recognition model to perform channel weighted fusion on the multi-dimensional fusion feature matrix to obtain the weighted feature input. S342. Input the weighted feature input into the collision state recognition model, perform forward inference, obtain the initial state recognition result and the corresponding output probability distribution, and calculate the prediction loss of the collision state recognition model by combining the real-time annotation information. S343. Based on the prediction loss of the bump and rub state recognition model, the gradient descent algorithm is used to perform backpropagation to adjust the convolution kernel weights and gating parameters in the feature extraction layer of the bump and rub state recognition model, so as to obtain the updated bump and rub state recognition model. S344. Reapply the updated collision state recognition model to the weighted feature input to obtain the optimized output probability distribution. S345. Extract the historical normal sample output probability distribution of the corresponding working condition interval from the working condition mapping database, calculate the corresponding mean and variance, and obtain the confidence tolerance interval. S346. Based on the optimized output probability distribution and confidence tolerance interval, dynamically calculate the classification decision boundary of each collision level under the current state to obtain a dynamic early warning threshold adapted to the current working condition.

[0056] S4. Based on the collision status recognition results and dynamic early warning threshold, generate corresponding early warning instruction information, combine multi-level hash verification to construct a monitoring status record chain, and distribute and store the monitoring status record chain on the industrial blockchain.

[0057] It should be explained that, based on the closed-loop operation of intelligent identification, dynamic criteria, hierarchical response and trusted evidence storage, the system realizes full-process management from fault identification to early warning response and data traceability; the status record chain combined with blockchain technology ensures the authenticity and non-repudiation of key monitoring data, meets the compliance audit and accountability traceability needs in industrial scenarios, and provides trusted data support for equipment health management.

[0058] In this optional embodiment, based on the rubbing status recognition results and dynamic early warning thresholds, corresponding early warning instruction information is generated, and combined with multi-level hash verification, a monitoring status record chain is constructed and distributed and stored in the industrial blockchain, including the following steps: S41. Compare the dynamic warning threshold with the collision state identification result, determine the current warning level according to the preset graded warning rules, and obtain the corresponding warning instruction information. The warning instruction information includes the warning level, identification state category, confidence value, trigger timestamp, and suggested handling measures. Specifically, the structured recognition results output by the sliding bearing rubbing state recognition model are obtained, namely the state label and corresponding confidence level. The dynamic early warning threshold under the current working condition is retrieved, and a comparison relationship between the recognition results and the threshold is established.

[0059] When the condition is determined to be normal, a Level 0 warning is triggered, generating a continuous monitoring instruction and recording only the current status to the monitoring log. When the condition is determined to be minor contact, a Level 1 warning is triggered, generating a maintenance inspection prompt instruction and pushing the equipment number, current operating parameters, and sensitive characteristic values ​​to the maintenance terminal, prompting a targeted inspection to be completed within 24 hours. When the condition is determined to be moderate contact, a Level 2 warning is triggered, generating an equipment load reduction control instruction and sending a load reduction signal to the sliding bearing control system via the industrial bus, while simultaneously pushing the warning information to the maintenance center. When the condition is determined to be severe contact, a Level 3 warning is triggered, generating an emergency shutdown linkage instruction, immediately cutting off the equipment's power circuit, activating the audible and visual alarm device, and automatically dialing the emergency phone number of the maintenance manager.

[0060] The warning level, trigger timestamp, action to be performed, and device unique identifier (SN code) are encoded according to a preset protocol (such as JSON format) to form a structured warning instruction, ensuring that the instruction can be parsed and executed by the terminal device.

[0061] S42. Extract the multi-dimensional fusion feature matrix, real-time operating parameters and corresponding early warning instructions within the current monitoring period, and associate them with the equipment number to construct a status data unit; S43. Based on the state data unit, construct a monitoring state record chain that coordinates the metadata chain and the data digest chain, and combine digital signature, improved consensus mechanism and multi-level hash verification to distribute and store the monitoring state record chain on the industrial blockchain.

[0062] In this optional embodiment, based on state data units, a monitoring state recording chain is constructed that coordinates the metadata chain and the data digest chain. Combined with digital signatures, an improved consensus mechanism, and multi-level hash verification, the monitoring state recording chain is distributed and stored on an industrial blockchain, including the following steps: S431. Based on the time-series correlation of the state data unit, construct a metadata chain and a data digest chain respectively, and combine the hash pointer to associate the metadata chain and the data digest chain to obtain the monitoring state record chain; S432. Perform multi-level hash operations on the monitoring status record chain to obtain multi-level hash values; S433. Use the device's private key to digitally sign the multi-level hash value, and verify and reach consensus on the digitally signed multi-level hash value based on an improved consensus mechanism to obtain a new block; S434. Store the new block to the industrial blockchain to achieve distributed evidence storage of the monitoring status record chain.

[0063] It should be explained that building a distributed monitoring status record chain and storing it on an industrial blockchain includes the following steps: Step 1: Based on the status data units organized according to the time sequence of the early warning trigger timestamp and the equipment operation cycle number, construct a metadata chain and a data summary chain respectively. The metadata chain stores the core metadata and the hash value of the previous node, while the data summary chain standardizes and encapsulates the status data units and calculates the hash value of the full data. The two chains are bound by the unique device identifier and the cycle number, and the chain association is realized by combining the hash pointer to form a collaborative monitoring status record chain. Step 2: Implement a multi-level hash verification mechanism for the state data unit, and calculate its core field hash (h_core), full data hash (h_full), and total hash (h_total) in sequence. Use h_total as the representative hash value of the data digest chain for this period, which will be used for subsequent signing and on-chain verification. Step 3: Use the private key of the blockchain wallet pre-registered on the device to digitally sign h_total and generate a signature value sig; submit h_total, sig, timestamp and node identity information to the industrial consortium blockchain network, and the consensus node will perform verification based on the improved consensus mechanism; Step 4: The improved consensus mechanism dynamically adjusts participation weights based on node reputation scores, comprehensively evaluates the completeness of submitted data, the legality of signatures, and the compliance of node behavior, and generates a new block to be uploaded to the blockchain after reaching a consensus. Step 5: Add the new block to the industrial blockchain main chain to complete the distributed notarization of the monitoring status record chain and synchronize it to each participating node in the network to achieve immutable and traceable data storage.

[0064] like Figure 2 As shown, according to another embodiment of the present invention, a sliding bearing rubbing condition monitoring system is provided. The system includes: a multi-dimensional fusion feature matrix construction module 1, a rubbing condition identification module 2, a model threshold collaborative optimization module 3, and a monitoring condition record chain construction and storage module 4. Among them, the multi-dimensional fusion feature matrix construction module 1, the rubbing state recognition module 2, the model threshold collaborative optimization module 3, and the monitoring state record chain construction and storage module 4 are connected in sequence. The multi-dimensional fusion feature matrix construction module 1 is used to perform time-frequency domain joint analysis and processing on the operating state parameters of the sliding bearing, extract the rubbing-sensitive feature set, and construct the multi-dimensional fusion feature matrix. The rubbing state recognition module 2 is used to combine a multi-head self-attention mechanism and a hybrid neural network to build a rubbing state recognition model, and to use the rubbing state recognition model to perform rubbing state recognition on a multi-dimensional fused feature matrix to obtain the rubbing state recognition result. Model threshold collaborative optimization module 3 is used to construct a dynamic adjustment model for early warning threshold based on the real-time operating parameters of sliding bearings, and combined with the gradient descent algorithm, iteratively update the weights of the feature extraction layer in the collision state recognition model to obtain a dynamic early warning threshold adapted to the current operating conditions. The monitoring status record chain construction storage module 4 is used to generate corresponding early warning instruction information based on the contact status identification results and dynamic early warning thresholds. Combined with multi-level hash verification, the monitoring status record chain is constructed and distributed and stored on the industrial blockchain.

[0065] In summary, by utilizing the above-mentioned technical solutions of this invention, the present invention achieves high-precision and robust identification of sliding bearing rubbing conditions. Furthermore, through innovative mechanisms such as dynamic threshold adjustment, online model optimization, and blockchain-based trusted evidence storage, it constructs an adaptive, evolvable, traceable, and highly secure intelligent monitoring system. This effectively overcomes the bottlenecks of existing technologies in terms of operating condition adaptability, model persistence, and data reliability, demonstrating significant technological advancement and broad application prospects. This invention employs a time-frequency domain joint analysis algorithm to extract multi-dimensional rubbing-sensitive features and combines a multi-head self-attention mechanism with a hybrid neural network model for state identification. This effectively integrates local feature extraction, temporal dependency modeling, and key segment focusing capabilities, significantly enhancing the sensitivity and classification accuracy for identifying early faults such as minor and intermittent rubbing, and solving the problems of misjudgment and missed detection in low signal-to-noise ratio environments using traditional methods. This invention constructs a dynamic adjustment model for early warning thresholds and optimizes the model using a gradient descent algorithm, enabling the early warning thresholds to adapt to changes in the operating conditions of the sliding bearing in real time. This improves the accuracy and response speed of early warnings and reduces the possibility of false alarms and missed detections. This invention uses multi-level hash verification and industrial blockchain technology to construct a monitoring status record chain, realizing secure data storage and tamper-proofing. Through distributed evidence storage, it ensures the integrity and traceability of monitoring data, providing a reliable data management solution.

[0066] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for monitoring the rubbing condition of a sliding bearing, characterized in that, The method includes the following steps: S1. Perform time-frequency domain joint analysis on the operating parameters of the sliding bearing, extract the rubbing-sensitive feature set, and construct a multi-dimensional fusion feature matrix; S2. Combining multi-head self-attention mechanism and hybrid neural network, construct a collision state recognition model, and use the collision state recognition model to perform collision state recognition on multi-dimensional fused feature matrix to obtain collision state recognition results. S3. Based on the real-time operating parameters of the sliding bearing, a dynamic adjustment model for the early warning threshold is constructed. Combined with the gradient descent algorithm, the weights of the feature extraction layer in the rubbing condition identification model are iteratively updated to obtain a dynamic early warning threshold adapted to the current operating condition. Specifically, this includes: S31. Perform data cleaning and normalization on the real-time operating parameters of the acquired sliding bearings, and construct an operating condition mapping database by combining historical monitoring data. S32. Using a fully connected neural network as the core, a dynamic adjustment model for the warning threshold is constructed. The input of the dynamic adjustment model for the warning threshold is the current working condition parameter vector, and the output of the dynamic adjustment model for the warning threshold is the confidence adjustment coefficient and feature channel weighted vector of each collision state category. S33. Based on the working condition mapping database, and combined with the preset loss function and gradient descent algorithm, train the early warning threshold dynamic adjustment model. S34. Based on the trained warning threshold, dynamically adjust the feature weighting vector output by the model to update the weights of the feature extraction layer of the collision condition recognition model. Combine the updated output probability distribution of the collision condition recognition model with the working condition mapping database to dynamically calculate the classification decision boundary under the current working condition, thus obtaining a dynamic warning threshold adapted to the current working condition. Specifically, this includes: S341. Dynamically adjust the feature channel weighted vector output by the trained warning threshold and apply it to the input of the feature extraction layer in the collision state recognition model to perform channel weighted fusion on the multi-dimensional fusion feature matrix to obtain the weighted feature input. S342. Input the weighted feature input into the collision state recognition model, perform forward inference, obtain the initial state recognition result and the corresponding output probability distribution, and calculate the prediction loss of the collision state recognition model by combining the real-time annotation information. S343. Based on the prediction loss of the bump and rub state recognition model, the gradient descent algorithm is used to perform backpropagation to adjust the convolution kernel weights and gating parameters in the feature extraction layer of the bump and rub state recognition model, so as to obtain the updated bump and rub state recognition model. S344. Reapply the updated collision state recognition model to the weighted feature input to obtain the optimized output probability distribution. S345. Extract the historical normal sample output probability distribution of the corresponding working condition interval from the working condition state mapping database, calculate the corresponding mean and variance, and obtain the confidence tolerance interval. S346. Based on the optimized output probability distribution and confidence tolerance interval, dynamically calculate the classification decision boundary of each collision level under the current state to obtain a dynamic early warning threshold adapted to the current working condition. S4. Based on the collision status recognition results and dynamic early warning threshold, generate corresponding early warning instruction information, combine multi-level hash verification to construct a monitoring status record chain, and distribute and store the monitoring status record chain on the industrial blockchain.

2. The method for monitoring the rubbing condition of a sliding bearing according to claim 1, characterized in that, The process of performing joint time-frequency domain analysis on the operating parameters of the sliding bearing, extracting the rubbing-sensitive feature set, and constructing a multi-dimensional fused feature matrix includes the following steps: S11. Obtain the operating state parameters of the sliding bearing, wherein the operating state parameters include dynamic signal parameters and steady-state signal parameters; S12. Using the ensemble empirical mode decomposition algorithm, the dynamic signal parameters are preprocessed to obtain the preprocessed dynamic signal parameters. S13. Based on the outlier detection algorithm, the steady-state signal parameters are preprocessed to obtain the preprocessed steady-state signal parameters. S14. Using a time-frequency domain joint analysis algorithm, the collision-sensitive features of the preprocessed dynamic signal parameters and the preprocessed steady-state signal parameters are extracted and filtered to obtain a collision-sensitive feature set. S15. Organize the rubbing-sensitive feature set in a structured manner according to a preset dimension, with time-series samples as rows and feature indicators as columns. Add timestamps, equipment numbers, and operating parameters as auxiliary information, and label the rubbing status to obtain a multi-dimensional fusion feature matrix.

3. The method for monitoring the rubbing condition of a sliding bearing according to claim 2, characterized in that, The method employs a joint time-frequency domain analysis algorithm to extract and filter the rubbing-sensitive features from the preprocessed dynamic signal parameters and the preprocessed steady-state signal parameters, resulting in a rubbing-sensitive feature set. This process includes the following steps: S131. Perform time-frequency transformation on the preprocessed dynamic signal parameters and the preprocessed steady-state signal parameters respectively, and combine the short-time Fourier transform algorithm to convert the time domain into the time-frequency domain to obtain the dynamic time-frequency distribution map and the steady-state time-frequency distribution map. S132. Extract time-domain, frequency-domain, and time-frequency-domain features from the dynamic time-frequency distribution map and the steady-state time-frequency distribution map to obtain the dynamic multi-domain feature set and the steady-state multi-domain feature set; S133. Align and fuse the dynamic multi-domain feature set with the steady-state multi-domain feature set to construct an initial high-dimensional feature pool; S134. A sparse feature selection algorithm is used to filter the high-dimensional feature pool to obtain a sensitive feature subset, and the sensitive feature subset is normalized to obtain the rubbing sensitive feature set.

4. The method for monitoring the rubbing condition of a sliding bearing according to claim 1, characterized in that, The process of constructing a collision state recognition model by combining a multi-head self-attention mechanism and a hybrid neural network, and then using this model to identify the collision state of a multi-dimensional fused feature matrix to obtain the collision state recognition result includes the following steps: S21. A hybrid neural network model is constructed by using graph convolutional neural networks and gated recurrent units as the core and integrating multi-head self-attention mechanism. S22. Obtain the historical multi-dimensional fusion feature matrix as a training sample set, and train the hybrid neural network model based on the training sample set to obtain the rubbing state recognition model. S23. Use the collision state recognition model to perform collision state recognition on the multi-dimensional fusion feature matrix to obtain collision state recognition results including collision state category, recognition confidence, timestamp and contribution features.

5. The method for monitoring the rubbing condition of a sliding bearing according to claim 4, characterized in that, The steps of obtaining the historical multi-dimensional fusion feature matrix as a training sample set, and training the hybrid neural network model based on the training sample set to obtain the rubbing state recognition model include the following: S221. The weight initialization algorithm is used to initialize the parameters of the hybrid neural network model, and the gradient descent optimization algorithm, cross-entropy loss function, learning rate and its decay strategy are configured. Regularization layer and batch normalization layer are introduced to improve the stability and generalization ability of model training. S222. Obtain the historical multi-dimensional fusion feature matrix as the training sample set, normalize the sample set, and divide it into training set, validation set and test set according to the preset ratio. S223. The training set responds to the hybrid neural network model, calculates the output probability distribution through forward propagation, calculates the loss value based on the true label, and updates the parameters of the hybrid neural network model based on the backpropagation algorithm. S224. Based on the recognition accuracy and comprehensive evaluation index on the validation set, the parameters of the updated hybrid neural network model are updated a second time. S225. The performance of the trained model is verified using the test set. The hybrid neural network model after the second update is evaluated by confusion matrix, receiver operation feature curve and confidence analysis. The collision state recognition model is optimized based on the evaluation results.

6. The method for monitoring the rubbing condition of a sliding bearing according to claim 1, characterized in that, The process of generating corresponding early warning instructions based on the rubbing status recognition results and dynamic early warning thresholds, and constructing a monitoring status record chain by combining multi-level hash verification and distributing it on the industrial blockchain includes the following steps: S41. Compare the dynamic warning threshold with the collision state identification result, determine the current warning level according to the preset graded warning rules, and obtain the corresponding warning instruction information. The warning instruction information includes the warning level, identification state category, confidence value, trigger timestamp, and suggested handling measures. S42. Extract the multi-dimensional fusion feature matrix, real-time operating parameters and corresponding early warning instructions within the current monitoring period, and associate them with the equipment number to construct a status data unit; S43. Based on the state data unit, construct a monitoring state record chain that coordinates the metadata chain and the data digest chain, and combine digital signature, improved consensus mechanism and multi-level hash verification to distribute and store the monitoring state record chain on the industrial blockchain.

7. The method for monitoring the rubbing condition of a sliding bearing according to claim 6, characterized in that, The process of constructing a monitoring status record chain based on state data units, coordinating a metadata chain and a data digest chain, and distributing the monitoring status record chain on an industrial blockchain using digital signatures, an improved consensus mechanism, and multi-level hash verification includes the following steps: S431. Based on the time-series correlation of the state data unit, construct a metadata chain and a data digest chain respectively, and combine the hash pointer to associate the metadata chain and the data digest chain to obtain the monitoring state record chain; S432. Perform multi-level hash operations on the monitoring status record chain to obtain multi-level hash values; S433. Use the device's private key to digitally sign the multi-level hash value, and verify and reach consensus on the digitally signed multi-level hash value based on an improved consensus mechanism to obtain a new block; S434. Store the new block to the industrial blockchain to achieve distributed evidence storage of the monitoring status record chain.

8. A sliding bearing rubbing condition monitoring system, used to implement the sliding bearing rubbing condition monitoring method according to any one of claims 1-7, characterized in that, The system includes: a multi-dimensional fusion feature matrix construction module, a collision state recognition module, a model threshold collaborative optimization module, and a monitoring state record chain construction and storage module; The multi-dimensional fusion feature matrix construction module is used to perform time-frequency domain joint analysis and processing on the operating state parameters of the sliding bearing, extract the rubbing sensitive feature set, and construct a multi-dimensional fusion feature matrix. The collision state recognition module is used to combine a multi-head self-attention mechanism and a hybrid neural network to construct a collision state recognition model, and to use the collision state recognition model to perform collision state recognition on a multi-dimensional fused feature matrix to obtain the collision state recognition result. The model threshold collaborative optimization module is used to construct a dynamic adjustment model for the early warning threshold based on the real-time operating parameters of the sliding bearing, and to iteratively update the weights of the feature extraction layer in the collision state recognition model by combining the gradient descent algorithm, so as to obtain a dynamic early warning threshold that is adapted to the current operating conditions. The monitoring status record chain construction and storage module is used to generate corresponding early warning instruction information based on the contact status identification results and dynamic early warning thresholds, and to construct the monitoring status record chain by combining multi-level hash verification, and to distribute and store the monitoring status record chain on the industrial blockchain.

Images