A cylinder on-line life prediction system and method

Abstract

This invention proposes an online life prediction system and method for hydraulic cylinders. The system includes: a data acquisition module, a data preprocessing module, a seal life prediction module, a bearing life prediction module, a coating life prediction module, a model fusion module, an early warning judgment module, and an overall early warning output module. Using this invention, hydraulic cylinder operating parameters can be collected in real time. After preprocessing, the input to each prediction model is processed, and after fusion, the overall predicted life value of the hydraulic cylinder and early warning information are output, providing accurate basis for hydraulic cylinder operation and maintenance, avoiding over-maintenance or under-maintenance, reducing operation and maintenance costs, and improving the safety of hydraulic cylinder operation.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance technology of hydraulic cylinders, specifically relating to an online life prediction system and method for hydraulic cylinders. Background Technology

[0002] As a core power component of marine engineering and heavy infrastructure equipment, the ultra-large hydraulic cylinder has a complex structure and operates under extreme conditions. It must withstand heavy loads of 5,000 tons, high salt spray corrosion, and frequent start-stop impacts. The operating status of core components such as seals, bearings, and coatings directly determines the safety of equipment construction and the progress of the project.

[0003] Currently, the industry generally adopts a traditional operation and maintenance model that combines regular offline inspection with experience-based maintenance. This not only requires interrupting the construction process but also suffers from serious problems of over-maintenance or under-maintenance, and faces many insurmountable technical bottlenecks: the status perception of key components is lacking, and faults such as seal leakage, bearing wear, and coating peeling are mostly discovered after the fact without an early warning mechanism, which can easily lead to major safety accidents such as cylinder explosion and pile frame instability; the data collection dimension is limited, only monitoring macroscopic parameters such as pressure and displacement, and lacking monitoring of key indicators such as sealing temperature, bearing swing angle, and microscopic defects in coating, which cannot comprehensively reflect the operating status of the cylinder.

[0004] While imported digital twin operation and maintenance systems can achieve some intelligent monitoring functions, the cost of a single unit is high, and their compatibility with domestically produced ultra-large hydraulic cylinders with a diameter ≥1.6m and a stroke ≥21m is poor. They cannot perform virtual simulation and fault tracing, and fault cause analysis relies on manual experience, resulting in low efficiency and insufficient accuracy. Lifespan prediction lacks scientific model support, relying solely on periodic replacement based on usage time. This leads to the replacement of some components before their lifespan has expired, while some overdue components are not replaced in time, causing failures. Data management is fragmented, with design, manufacturing, and operation and maintenance data independent of each other, lacking a complete traceability chain and failing to provide data support for hydraulic cylinder iterative optimization. Furthermore, the core algorithm is closed, incompatible with the manufacturing process data of domestically produced hydraulic cylinders, and poses risks of data security leakage and delayed after-sales response (≥72h). Moreover, there is no complete digital twin model to support it. From a sensor perspective, existing domestic monitoring technologies lack dedicated high-precision sensors adapted to the complex structure of hydraulic cylinders. From a model perspective, they only reach the geometric modeling stage, failing to accurately map the mechanical characteristics and operational performance of hydraulic cylinders. From an algorithm perspective, life prediction models mostly use single-parameter fitting, neglecting the coupled effects of corrosion, wear, and temperature under marine conditions, resulting in low prediction accuracy. From a data perspective, data transmission latency is large, failing to meet real-time monitoring requirements. From an adaptability perspective, existing systems are mostly developed for small and medium-sized hydraulic cylinders, unable to accommodate the large size, heavy load, and complex operating conditions of ultra-large hydraulic cylinders. While imported systems offer more complete functions, they suffer from technological blockades, high costs, and data security risks, making it difficult to meet the needs of large-scale industrial applications in China. Summary of the Invention

[0005] To address the core technical challenges of real-time status monitoring of critical components in ultra-large hydraulic cylinders, such as low accuracy in lifespan prediction, lack of digital twin models, non-closed-loop data throughout the entire lifecycle, and insufficient intelligent operation and maintenance, this invention proposes an online lifespan prediction system and method for hydraulic cylinders.

[0006] An online life prediction system for hydraulic cylinders, which achieves one of the objectives of this invention, includes: The data acquisition module is used to collect sealing-related parameters, bearing-related parameters, coating-related parameters, and coating defect images during the operation of the hydraulic cylinder. The sealing-related parameters include the sealing cavity temperature, sealing leakage, and the concentration of oil contaminants inside the sealing cavity. The bearing-related parameters include the operating load, bearing swing angle, bearing working surface wear, bearing impact frequency, and bearing operating temperature. The coating-related parameters include the coating surface temperature of the outer surface of the hydraulic cylinder barrel and piston rod. The coating defect images are images of the coating on the outer surface of the hydraulic cylinder barrel and piston rod acquired by a linear scanning camera, used to identify the defect area ratio and corrosion depth. The data preprocessing module is used to perform Kalman filtering for noise reduction, outlier removal and completion, and normalization on the sealing-related parameters, bearing-related parameters, and coating-related parameters to generate standardized data. The sealing life prediction module is used to input the sealing-related parameters in the standardized data into the sealing life prediction model based on BP neural network and output the remaining service life of the seal; normalization processing refers to mapping the sealing-related parameters, bearing-related parameters, and coating-related parameters to the [0,1] interval; The bearing life prediction module is used to input the bearing-related parameters in the standardized data into the bearing life prediction model based on CNN-BP hybrid neural network, and output the remaining service life of the bearing. The coating lifetime prediction module is used to input coating defect images and coating-related parameters from the standardized data into a coating lifetime prediction model based on YOLOv8-CNN hybrid neural network, and output the remaining lifetime of the coating. The model fusion module is used to perform fusion calculations on the remaining service life of the seal, the remaining service life of the bearing, and the remaining service life of the coating through a dynamic weighted fusion method, so as to obtain the overall online life prediction value of the hydraulic cylinder.

[0007] Furthermore, the dynamic weighted fusion method includes: first, presetting the basic weights of each component; then, dynamically adjusting the weights according to the component's early warning level and the model's prediction accuracy, and performing normalization processing; finally, using the normalized weights to perform a weighted summation of the remaining service life of the seals, bearings, and coatings to obtain the overall remaining service life of the hydraulic cylinder.

[0008] Furthermore, the model fusion module calculates the overall online life prediction value of the hydraulic cylinder using the following formula: L 总 =a×L 密 +b×L 轴 +c×L 涂 ; Where L 总 L is the predicted online lifespan of the hydraulic cylinder. 密 L 轴 L 涂 These represent the remaining service life of the seal, the remaining service life of the bearing, and the remaining service life of the coating, respectively; a, b, and c represent the weights of the seal, bearing, and coating, respectively.

[0009] Furthermore, it also includes a warning determination module, used to determine the warning level based on the remaining service life of the seal, the remaining service life of the bearing, and the remaining service life of the coating, specifically including: When the remaining service life of the seal is lower than the set secondary warning threshold, the seal is determined to trigger a secondary warning; when the remaining service life of the seal is lower than the set primary warning threshold but not lower than the secondary warning threshold, the seal is determined to trigger a primary warning. When the remaining service life of the bearing is lower than the set bearing secondary warning threshold, the bearing is determined to trigger a secondary warning. When the remaining service life of the bearing is lower than the set bearing primary warning threshold but not lower than the bearing secondary warning threshold, the bearing is determined to trigger a primary warning. When the remaining service life of the coating is lower than the set secondary warning threshold, the coating is determined to trigger a secondary warning. When the remaining service life of the coating is lower than the set primary warning threshold but not lower than the secondary warning threshold, the coating is determined to trigger a primary warning.

[0010] Furthermore, the model fusion module also includes a weight allocation unit, which is used to execute the following dynamic adjustment strategy for the weights of seals, bearings, and coatings: Set the base weights for seals, bearings, and coatings. For each component among seals, bearings, and coatings, if a component triggers a level 2 warning, the corresponding weight adjustment is set to 0.1; if the component triggers a level 1 warning, the corresponding weight adjustment is set to 0.05; if no warning is triggered, the corresponding weight adjustment is set to 0. Add the base weights and corresponding adjustment values ​​of each component to obtain the unnormalized weights of each component. Then normalize the unnormalized weights of each component to obtain the final dynamic weights of each component.

[0011] It should be noted that in the dynamic weight adjustment strategy of the present invention, the remaining service life of the early warning component is already low. Increasing its weight can make the total service life prediction value more sensitive to the deterioration trend of the component, thereby triggering the overall early warning in a timely manner.

[0012] Furthermore, the sealing life prediction model based on a BP neural network in the sealing life prediction module is a four-layer feedforward neural network structure, specifically including: Input layer: Used to receive normalized sealing-related parameters, match the data dimension, and then pass them to the first hidden layer; the matching data dimension refers to converting the discrete sealing-related parameters into a vector format consistent with the number of neurons in the input layer, so that the input data can be recognized by the network; First hidden layer: used to extract local degradation feature vectors from sealing-related parameters, the local degradation features including sealing cavity temperature change trend, leakage fluctuation characteristics, and contaminant concentration accumulation characteristics; ReLU activation function is used to nonlinearly activate the local degradation features; The second hidden layer is used to receive the activated local degradation feature vector output by the first hidden layer, propose a global nonlinear mapping feature vector, and perform nonlinear activation on the global nonlinear mapping feature vector using the ReLU activation function; the global nonlinear mapping feature vector represents the mapping relationship between the remaining service life of the seal and various seal-related parameters; The ReLU activation function effectively solves the gradient vanishing problem in deep network training, improves the model convergence speed and training stability, and the two-layer stacked structure significantly enhances the model's nonlinear mapping capability, providing comprehensive and reliable feature support for the output layer to finally output accurate remaining service life of the seal. Output layer: Used to receive the activated global nonlinear mapping feature vector output from the second hidden layer, map it to the first preset interval using the Sigmoid activation function, and output the remaining service life of the seal; During the training process of the sealing life prediction model based on the BP neural network, the Adam gradient descent method is used to backpropagate and iteratively update the network weights and biases, and minimize the mean squared error loss function. The learning rate is fixed at 0.01, and the loss function is the mean squared error. Training stops when the loss function value converges to no more than 0.001.

[0013] Furthermore, the bearing life prediction model based on a CNN-BP hybrid neural network in the bearing life prediction module includes: Input layer: Used to concatenate normalized bearing-related parameters into time-series data using a time-series concatenation method, and then pass the data to the CNN feature extraction layer after matching the data dimensions; CNN Feature Extraction Layer: Used to learn multi-scale temporal features of the bearing degradation process from the time-series data to extract degradation characteristics strongly correlated with the remaining service life of the bearing; it includes two convolutional layers and one pooling layer, wherein the first convolutional layer uses 3×1 convolutional kernels with 16 kernels, the second convolutional layer uses 3×1 convolutional kernels with 32 kernels, and the pooling layer uses 2×1 max pooling; the first convolutional layer is used to extract local degradation feature maps of the bearing from the time-series data, the local degradation feature maps including bearing swing angle fluctuation patterns, load impact waveform features, and wear change trends; the second convolutional layer is used to perform deep feature extraction on the local degradation feature maps and output a deep degradation feature map; the pooling layer is used to reduce the dimensionality of the deep degradation feature map and output a pooled degradation feature map, and then flatten the pooled degradation feature map into a one-dimensional feature vector before outputting it to the BP prediction layer; BP prediction layer: used to map the one-dimensional feature vector to a predicted value of the bearing's remaining service life; it includes a first hidden layer, a second hidden layer, and an output layer. The first hidden layer contains 16 neurons with a ReLU activation function, used to receive the one-dimensional feature vector, extract nonlinear combination features of the bearing degradation process, and output the combination feature vector to the second hidden layer. The second hidden layer contains 16 neurons with a ReLU activation function, used to receive the combination feature vector and map it to a global nonlinear relationship feature vector before outputting it to the output layer. The global nonlinear relationship feature vector represents the complex mapping relationship between the bearing's remaining service life and various bearing-related parameters. Output layer: Used to map the output of the second hidden layer to the second preset interval using the Sigmoid activation function, and output the remaining service life of the bearing; During the training process of the bearing life prediction model based on CNN-BP hybrid neural network, the optimizer adopts Adam gradient descent method with an initial learning rate of 0.01, which decreases by 0.1 every 100 iterations. The loss function adopts mean squared error and introduces L2 regularization with a regularization coefficient of 0.001.

[0014] Furthermore, the coating lifetime prediction model based on the YOLOv8-CNN hybrid neural network in the coating lifetime prediction module includes: YOLOv8 Defect Recognition Layer: Used to detect defects in coating defect images, outputting the defect area ratio and corrosion depth to the CNN lifetime prediction layer; the defect area ratio represents the extent of coating damage, and the corrosion depth represents the severity of coating damage; CNN lifetime prediction layer: used to map the defect area ratio, corrosion depth, and coating surface temperature to predicted values ​​of the remaining lifetime of the coating; includes a first convolutional layer, a second convolutional layer, a pooling layer, a fully connected layer, and an output layer, wherein: The first convolutional layer uses 16 3×3 convolutional kernels with ReLU activation function to extract local features of defect area ratio, corrosion depth, and coating surface temperature, and outputs the first convolutional feature map; the first convolutional feature map encodes the initial coupling relationship between coating damage parameters and temperature. The second convolutional layer uses 3×3 convolutional kernels with a total of 32 kernels and ReLU activation function. It is used to extract deep features from the first convolutional feature map and output the second convolutional feature map. The second convolutional feature map represents the nonlinear evolution law in the coating degradation process. Pooling layer: used to reduce the dimensionality of the second convolutional feature map and output a pooled feature map; the pooled feature map retains the main features of coating degradation while reducing computational parameters; Fully connected layer: Contains 16 neurons, used to map the pooling feature map into a feature vector, which is then output to the output layer. The feature vector integrates global degradation information coupled with coating damage and temperature. The fully connected layer uses a linear transformation to map the pooling feature map into a feature vector: first, the pooling feature map is flattened into a one-dimensional vector, and then a one-dimensional feature vector is obtained through linear calculation using a pre-trained weight matrix and bias. This feature vector serves as the input to the output layer. The linear transformation of the fully connected layer is a well-known technique in the art.

[0015] Output layer: The Sigmoid activation function is used to map the feature vector to a third preset interval, and the remaining service life of the coating is output.

[0016] The YOLOv8-CNN hybrid neural network was trained in the following way: Multiple coating defect images were selected and divided into training, validation and test sets in an 8:1:1 ratio. The defect area ratio and corrosion depth were labeled on the images, and the corresponding coating temperature data were matched and associated with the coating's remaining service life label. The remaining service life label of the coating is determined by accelerated aging tests in the laboratory and is used to record the actual failure time of the coating under different defect levels. The initial learning rate of the YOLOv8 defect detection layer is 0.001. The optimizer uses stochastic gradient descent with a momentum of 0.937 and a weight decay of 0.0005. The loss function is the cross-intersection over union loss. The learning rate of the CNN lifetime prediction layer is 0.01, the optimizer uses the Adam gradient descent method, the batch size is 32, the number of training iterations is no less than 300, the accuracy is verified using the validation set every 10 iterations, and the loss function is the mean squared error. Training is stopped when the defect identification accuracy is not less than 98%, the defect area ratio detection error does not exceed the set ratio of 0.5%, the corrosion depth detection error does not exceed the set threshold of 0.05mm, and the lifetime prediction loss function value converges to not more than 0.001. The test set images were input into the YOLOv8-CNN model to verify the accuracy of defect identification and lifetime prediction. The network weights were adjusted to ensure that the accuracy of coating lifetime prediction reproduction was not less than 92% and the corrosion depth detection error was not more than 0.05 mm.

[0017] Furthermore, the YOLOv8 defect identification layer is implemented using a lightweight YOLOv8n network structure. Specifically, this layer includes: C2f module: used to receive standardized coating defect images, which are 640×640 pixels in size in this embodiment. It extracts image features through a cross-stage local connection structure and outputs a first feature map. SPPF and PAN-FPN modules: SPPF is a fast spatial pyramid pooling structure, and PAN-FPN is a path aggregation network feature pyramid. The two work together to receive the first feature map, perform multi-scale feature fusion, and output the second feature map. Decoupling head: Used to receive the second feature map and output the defect area ratio and corrosion depth through independent regression branches.

[0018] Furthermore, the Kalman filtering noise reduction in the data preprocessing module is implemented in the following way: Initialize the system state equation and observation equation, set the process noise covariance Q=1e-4 and the observation noise covariance R=1e-3, and calculate the optimal estimate in real time based on the state estimate of the previous time step and the observation value of the current time step. Furthermore, the outlier removal in the data preprocessing module is achieved in the following way: Based on the rated measurement accuracy of each sensor, calculate the standard deviation σ of multiple consecutive sets (e.g., 100 sets) of valid data for a single sensor, and judge data with a deviation exceeding 3σ as abnormal data; Furthermore, the completion in the data preprocessing module is achieved in the following way: For single-point abnormal data, the mean of multiple adjacent (e.g., 5) valid data is used to complete the data. For multiple consecutive sets (e.g., 3 or more) of abnormal data, the data is marked as sensor fault and a local audible and visual warning is triggered.

[0019] Furthermore, the normalization process in the data preprocessing module maps the original data to the 0-1 interval, eliminating the dimensional differences between different parameters; its calculation formula includes: normalized data = (original data - minimum value of all samples of the parameter) / (maximum value of all samples of the parameter - minimum value of all samples of the parameter).

[0020] Furthermore, the data acquisition module includes multiple dedicated sensors, which are deployed as follows: three monitoring points are deployed on the cylinder, four monitoring points are deployed on the piston rod, two monitoring points are deployed on the sealing cavity, and two monitoring points are deployed on the bearing seat. The installation angle of each sensor is 45° to 60°, and it is fixed by a stainless steel bracket. The wiring terminals are sealed with waterproof aviation plugs.

[0021] Furthermore, it also includes a data acquisition and transmission control module, which adopts an industrial Ethernet gateway that supports the Profinet protocol and has a communication rate of no less than 100Mbps, and is configured with an integrated hardware and software acquisition module that integrates the AD7606 data acquisition chip. The sampling frequency is set to 10Hz, and the data transmission adopts the AES-256 encryption protocol, and is uploaded to the local redundant storage server and cloud database in real time.

[0022] A method for predicting the online lifespan of a hydraulic cylinder, which achieves the second objective of this invention, includes: Collect sealing-related parameters, bearing-related parameters, coating-related parameters, and coating defect images during the operation of the hydraulic cylinder; Kalman filtering, outlier removal and completion, and normalization are performed on the sealing-related parameters, bearing-related parameters, and coating-related parameters to generate standardized data. The sealing-related parameters in the standardized data are input into the sealing life prediction model based on BP neural network, and the remaining service life of the seal is output. The bearing-related parameters in the standardized data are input into the bearing life prediction model based on CNN-BP hybrid neural network, and the remaining service life of the bearing is output. The coating defect image and coating-related parameters from the standardized data are input into a coating lifetime prediction model based on a YOLOv8-CNN hybrid neural network, and the remaining lifetime of the coating is output. The remaining service life of the seal, the remaining service life of the bearing, and the remaining service life of the coating are calculated by dynamically weighted fusion to obtain the overall online life prediction value of the hydraulic cylinder.

[0023] A computer program product for achieving the third objective of the present invention includes a computer program / instruction, which, when executed by a processor, implements the steps of the online life prediction method for hydraulic cylinders. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the system described in this invention. Detailed Implementation

[0025] The following detailed embodiments are provided to explain the technical solutions of the present invention, so that those skilled in the art can understand the present invention. The scope of protection of the present invention is not limited to the following specific embodiments. Any modifications or improvements made by those skilled in the art that incorporate the technical solutions of the present invention but differ from the following detailed embodiments are also within the scope of protection of the present invention.

[0026] Example 1 A hydraulic cylinder online life prediction system This embodiment provides an online life prediction system for hydraulic cylinders, applicable to domestically produced ultra-large hydraulic cylinders with a diameter ≥ 1.6m and a stroke ≥ 21m, meeting the extreme operating conditions required for marine engineering and heavy infrastructure equipment. The system includes a data acquisition module, a data preprocessing module, a seal life prediction module, a bearing life prediction module, a coating life prediction module, a model fusion module, an early warning judgment module, and an overall early warning output module, such as... Figure 1 As shown.

[0027] The data acquisition module is used to collect sealing-related parameters, bearing-related parameters, coating-related parameters, and coating defect images during the operation of the hydraulic cylinder. Specifically, the following dedicated sensors are deployed in this embodiment: Laser displacement sensor: measuring range 0~25m, accuracy ±0.1mm, deployed at the bearing housing and piston rod, used to collect bearing swing angle and bearing wear data; Infrared thermal imager: Temperature measurement range -20~250℃, resolution 0.1℃, deployed around the sealed cavity and bearing housing to collect temperature data of the sealed cavity and bearing operating temperature; Linear scanning camera: Deployed on the outer wall of the cylinder and the surface of the piston rod, used to acquire images of the coating surface; Oil contamination sensor: with a particle size detection accuracy of ≥2μm, deployed in the oil circulation pipeline of the sealed cavity to collect contaminant concentration data in the oil of the sealed cavity; Leakage sensor: measuring range 0~20mL / h, accuracy ±0.05mL / h, deployed at the bottom of the sealing cavity, used to collect leakage data of the sealing component; Load sensor: measuring range 0~6000t, accuracy ±0.1%FS, deployed below the bearing housing, used to collect data on the hydraulic cylinder's operating load and impact frequency.

[0028] The sensor deployment plan is as follows: 3 points on the cylinder, 4 points on the piston rod, 2 points in the sealing cavity, and 2 points on the bearing housing, with installation angles ranging from 45° to 60° to avoid motion interference areas and stress concentration points. All sensors are fixed with stainless steel brackets, and the wiring terminals are sealed with waterproof aviation plugs to adapt to high-salt-spray marine environments.

[0029] The data acquisition system uses a Siemens SCALANCE XB-200 industrial Ethernet gateway, supporting the Profinet protocol with a communication rate ≥100Mbps, and unifies the access of all sensor data. It is configured with an integrated hardware and software acquisition module incorporating the AD7606 data acquisition chip, with the sampling frequency set to 10Hz based on the single extension / retraction cycle of the hydraulic cylinder. Data transmission uses the AES-256 encryption protocol and is uploaded in real-time to a local redundant storage server and a cloud database.

[0030] The cumulative collection of valid operational data is no less than 30,000 sets, covering different operating conditions such as heavy load, light load, and frequent swaying, as well as different environmental conditions such as temperature -20~120℃ and pollutant concentration 0~50ppm.

[0031] The data preprocessing module performs Kalman filtering for noise reduction, outlier removal and completion, and normalization on sealing-related parameters (sealing cavity temperature, leakage rate, oil contaminant concentration), bearing-related parameters (bearing angle, operating load, wear rate, impact frequency, bearing operating temperature), and coating-related parameters (coating surface temperature), generating standardized data. Coating defect images are not processed by this module and are directly transmitted to the coating life prediction module.

[0032] (1) Kalman filtering for noise reduction The system state equation and observation equation are initialized, and the process noise covariance Q=1e-4 and the observation noise covariance R=1e-3 are set. These parameters are determined through multiple experiments, taking into account the sensor's rated accuracy and the intensity of marine electromagnetic interference. The optimal estimate is calculated in real time based on the state estimate of the previous moment and the observation value of the current moment, effectively eliminating marine electromagnetic interference and sensor noise.

[0033] (2) Outlier removal and completion Based on the rated measurement accuracy of each sensor, the standard deviation σ of 100 consecutive sets of valid data for a single sensor is calculated. Data with a deviation exceeding 3σ is identified as abnormal data. For single-point abnormal data, the mean of five adjacent valid data points is used to complete the data. For three or more consecutive sets of abnormal data, the data is marked as sensor fault and a local audible and visual warning is triggered to ensure data integrity.

[0034] (3) Normalization The denoised and completed effective data is then subjected to min-max normalization to map the data to the [0,1] interval, eliminating the influence of dimensions. The normalization formula is: x'=(xx min ) / (x max -x min ), where x is the original data, x min x max These are the minimum and maximum values ​​of the parameter in the sample data, respectively. Standardized data are obtained after normalization.

[0035] The reliability of the preprocessed data is no less than 99%, and the single-parameter processing error does not exceed ±0.1%. The effective preprocessed data is divided into training set, validation set, and test set in a ratio of 8:1:1.

[0036] The seal life prediction module is used to input the seal-related parameters from the standardized data into the seal life prediction model based on the BP neural network, and output the remaining service life of the seal.

[0037] In one embodiment, the seal life prediction model based on a BP neural network is a four-layer feedforward neural network structure, specifically including: Input layer: 3 neurons, corresponding to the three input parameters of sealing cavity temperature, sealing leakage, and oil contaminant concentration. A linear transfer function is used to ensure that the input data is transmitted without distortion. The first hidden layer contains 18 neurons and uses the ReLU activation function to extract local degradation feature vectors from sealing-related parameters. The local degradation features include the temperature change trend of the sealing cavity, leakage fluctuation characteristics, and pollutant concentration accumulation characteristics. The second hidden layer contains 18 neurons and uses the ReLU activation function. It receives the local degradation feature vector output by the first hidden layer and outputs a global nonlinear relationship feature vector to the output layer. The global nonlinear relationship feature vector represents the mapping relationship between the remaining service life of the seal and various seal-related parameters. Output layer: 1 neuron, the activation function is the Sigmoid function, the output of the second hidden layer is mapped to the first preset interval [0,1], 0 represents exhaustion, 1 represents new; then mapped to the actual lifespan interval of the seal. In this embodiment, the maximum design lifespan of the seal is 10,000 hours. The normalized lifespan value is converted into hours, and the remaining lifespan of the seal is output in hours.

[0038] The model training process is as follows: Select multiple sets (e.g., 10,000 sets) of preprocessed sealing-related data and divide them into training, validation, and test sets in a ratio of 8:1:1; the output layer corresponds to the remaining service life label of the seal.

[0039] The network weights are initialized using a random normal distribution, with values ​​ranging from -0.1 to 0.1. The initial bias is 0.01. The optimizer uses Adam gradient descent, with a fixed learning rate of 0.01, a batch size of 32, and at least 500 training iterations. The loss function is the mean squared error (MSE) = 1 / n × Σ(y i -ŷ i Training is stopped when the loss function value converges to no greater than 0.001, or when the accuracy of the validation set shows no improvement for 20 consecutive iterations and the fluctuation does not exceed 0.5%.

[0040] The bearing life prediction module is used to input bearing-related parameters from standardized data into a bearing life prediction model based on a CNN-BP hybrid neural network, and output the remaining service life of the bearing.

[0041] Considering the temporal and correlational nature of bearing operating parameters, this embodiment employs a bearing life prediction model based on a CNN-BP hybrid neural network. Convolutional and pooling layers extract temporal features, while the BP layer performs regression prediction. The specific model structure includes: Input layer: The five bearing-related parameters are concatenated into time series data using a time series concatenation method. In this embodiment, the time series window length is 10 sampling points, and the data is output to the CNN feature extraction layer.

[0042] CNN Feature Extraction Layer: Used to learn multi-scale temporal features of the bearing degradation process from the time-series data to extract degradation characteristics strongly correlated with the remaining service life of the bearing; it includes two convolutional layers and one pooling layer, wherein the first convolutional layer uses 3×1 convolutional kernels with 16 kernels, the second convolutional layer uses 3×1 convolutional kernels with 32 kernels, and the pooling layer uses 2×1 max pooling; the first convolutional layer is used to extract local degradation feature maps of the bearing from the time-series data, the local degradation feature maps including bearing swing angle fluctuation patterns, load impact waveform features, and wear change trends; the second convolutional layer is used to perform deep feature extraction on the local degradation feature maps and output a deep degradation feature map; the pooling layer is used to reduce the dimensionality of the deep degradation feature map and output a pooled degradation feature map, and then flatten the pooled degradation feature map into a one-dimensional feature vector before outputting it to the BP prediction layer.

[0043] The BP prediction layer comprises a first hidden layer, a second hidden layer, and an output layer. The first hidden layer contains 16 neurons with the ReLU activation function. It receives the one-dimensional feature vector, extracts the nonlinear combination features of the bearing degradation process, and outputs the combination feature vector to the second hidden layer. The second hidden layer contains 16 neurons with the ReLU activation function. It receives the combination feature vector, maps it to a global nonlinear relationship feature vector, and outputs it to the output layer. The global nonlinear relationship feature vector represents the complex mapping relationship between the remaining service life of the bearing and various bearing-related parameters. The output layer uses the Sigmoid activation function to map the global nonlinear relationship feature vector to a second preset interval [0,1], corresponding to the maximum design life of the bearing, which is 8000 hours in this embodiment, and outputs the remaining service life of the bearing.

[0044] The bearing life prediction model based on the CNN-BP hybrid neural network was trained in the following way: Select multiple sets (e.g., 8000 sets) of preprocessed bearing-related data and divide them into training, validation, and test sets in an 8:1:1 ratio; set the label for the remaining service life of the bearing corresponding to the output layer; The weights of the convolutional layers were initialized using a He normal distribution, and the weights of the backpropagation layers were initialized using a random normal distribution, with an initial bias value of 0.01. The optimizer used the Adam gradient descent method with an initial learning rate of 0.01, which decreased by 0.1 every 100 iterations. The batch size was 32, and the number of training iterations was no less than 600. The loss function was the mean squared error, and L2 regularization was introduced to suppress overfitting, with the regularization coefficient set to 0.001. Training was stopped when the loss function value converged to no greater than 0.0008, or when the accuracy on the validation set showed no improvement for 25 consecutive iterations and the fluctuation did not exceed 0.4%.

[0045] The coating lifetime prediction module is used to input coating defect images and coating-related parameters from standardized data into a coating lifetime prediction model based on YOLOv8-CNN hybrid neural network, and output the remaining lifetime of the coating.

[0046] In this embodiment, the coating lifetime prediction model based on the YOLOv8-CNN hybrid neural network is divided into two stages: the YOLOv8 defect identification layer and the CNN lifetime prediction layer.

[0047] (1) YOLOv8 Defect Identification Layer: A lightweight YOLOv8n network structure is used to receive standardized coating defect images. In this embodiment, the image size is 640×640 pixels. The defect area ratio and corrosion depth are output to the CNN lifetime prediction layer.

[0048] The specific internal structure is as follows. It should be noted that this internal structure is a preferred implementation, and those skilled in the art can use other equivalent structures: C2f module: Receives coating defect images, extracts image features through cross-stage local connectivity structures, and outputs the first feature map.

[0049] SPPF and PAN-FPN modules: SPPF is a fast spatial pyramid pooling structure, and PAN-FPN is a path aggregation network feature pyramid. The two work together to receive the first feature map, perform multi-scale feature fusion, and output the second feature map.

[0050] Decoupling Head: Receives the second feature map and outputs the defect area ratio and corrosion depth through independent regression branches. The defect area ratio is the proportion of the defect area to the total coating area, expressed as a percentage, and the corrosion depth is the maximum depth of the corrosion pits on the coating surface, expressed in millimeters.

[0051] (2) CNN lifetime prediction layer: The system receives three input parameters: defect area ratio, corrosion depth, and coating surface temperature. A 1D convolutional neural network is used for feature extraction and regression prediction. The specific structure includes: The first convolutional layer uses 16 3×3 convolutional kernels with the ReLU activation function to extract local features of the defect area ratio, corrosion depth, and coating surface temperature, and outputs the first convolutional feature map. The first convolutional feature map encodes the initial coupling relationship between coating damage parameters and temperature.

[0052] The second convolutional layer uses 3×3 convolutional kernels, with a total of 32 kernels and the ReLU activation function. It is used to extract deep features from the first convolutional feature map and output the second convolutional feature map. The second convolutional feature map represents the nonlinear evolution law in the coating degradation process.

[0053] Pooling layer: used to reduce the dimensionality of the second convolutional feature map and output a pooled feature map; the pooled feature map retains the main features of coating degradation while reducing computational parameters.

[0054] The fully connected layer contains 16 neurons and is used to map the pooling feature map into a feature vector, which is then output to the output layer. This feature vector integrates global degradation information coupled with coating damage and temperature. Specifically, the pooling feature map is first flattened into a one-dimensional vector, and then a linear transformation (output vector = weight matrix × input vector + bias) is performed to obtain a one-dimensional feature vector, which is then output to the output layer. This linear transformation is a well-known technique in the art.

[0055] Output layer: The Sigmoid activation function is used to receive the feature vector and map it to the third preset interval [0,1], and then it is denormalized to the interval between 0 and the maximum design life of the coating of 6000 hours, and the remaining life of the coating is output.

[0056] The model training process is as follows: 12,000 images of coating defects were selected (9,600 for training, 1,200 for validation, and 1,200 for testing). All images were labeled with the defect area percentage and corrosion depth. Corresponding coating surface temperature data were also matched to form a joint dataset. The output layer is labeled with the remaining service life of the coating. The YOLOv8 defect recognition layer uses pre-trained weights based on the COCO dataset with an initial learning rate of 0.001. The optimizer uses stochastic gradient descent (SGD) with a momentum of 0.937 and weight decay of 0.0005, and the loss function is Complete IoU Loss (CIoU). The CNN life prediction layer has a learning rate of 0.01, uses the Adam optimizer, has a batch size of 32, and undergoes at least 300 training iterations. Accuracy is validated using the validation set every 10 iterations, and the loss function is mean squared error. Training stops when the defect identification accuracy is not less than 98%, the error in defect area and corrosion depth detection does not exceed 0.05 mm, and the lifetime prediction loss function value converges to not more than 0.001.

[0057] The early warning determination module determines the early warning level for each component based on the remaining service life of the seals, the remaining service life of the bearings, and the remaining service life of the coating: Seals: When the remaining service life of the seal is less than 30 days below the secondary warning threshold, a secondary warning is triggered; when it is less than 60 days below the primary warning threshold but not less than the secondary warning threshold, a primary warning is triggered.

[0058] Bearings: A level 2 warning is triggered when the remaining service life of the bearing is less than 30 days below the level 2 warning threshold; a level 1 warning is triggered when the remaining service life is less than 60 days below the level 1 warning threshold but not less than the level 2 warning threshold.

[0059] Coating: When the remaining service life of the coating is less than 30 days below the level 2 warning threshold, a level 2 warning is triggered; when it is less than 60 days below the level 1 warning threshold but not less than the level 2 warning threshold, a level 1 warning is triggered.

[0060] The above threshold values ​​are for illustrative purposes only and can be adjusted according to the actual design life and operating conditions.

[0061] The model fusion module uses an attention-weighted fusion method to calculate the remaining service life of the seals, bearings, and coatings, outputting a predicted overall online life of the hydraulic cylinder. The fusion formula is: L 总 =a×L 密 +b×L 轴 +c×L 涂 Where L 总 L is the predicted online lifespan of the hydraulic cylinder. 密 L 轴 L 涂 These represent the remaining service life of the seal, the remaining service life of the bearing, and the remaining service life of the coating, respectively; a, b, and c represent the weights of the seal, bearing, and coating, respectively.

[0062] In this embodiment, the model fusion module further includes a weight allocation unit, used to execute the following dynamic weight adjustment strategy: The base weights for seals, bearings, and coatings are set to 0.35, 0.4, and 0.25, respectively.

[0063] For each component, if a Level 2 warning is triggered, the adjustment amount is set to 0.1; if a Level 1 warning is triggered, the adjustment amount is set to 0.05; if there is no warning, the adjustment amount is set to 0.

[0064] The unnormalized weights are obtained by adding the base weights of each component to their corresponding adjustment amounts.

[0065] Normalize all unnormalized weights (divide each unnormalized weight by the sum of the three) so that the sum of the three is 1, and obtain the final dynamic weights a, b, c.

[0066] For example: When only a single component triggers the warning, assuming the bearing triggers a level two warning, its remaining life L 轴 =20 days, other components are normal, L 密 For 100 days, L 涂 The duration is 100 days; assuming basic weights a=0.35, b=0.4, c=0.25; L 总 =0.35×100+0.4×20+0.25×100=35+8+25=68 days.

[0067] After adjustment, the bearing weights are increased by 0.1, and the weights become: a'=0.35, b'=0.4+0.1=0.5, c'=0.25; the sum is 1.1; the normalized weights are: a=0.35 / 1.1≈0.318, b=0.5 / 1.1≈0.455, c=0.25 / 1.1≈0.227.

[0068] L 总 =0.318×100+0.455×20+0.227×100=31.8+9.1+22.7=63.6 days.

[0069] Total life expectancy decreased from 68 days to 63.6 days.

[0070] When multiple components trigger warnings simultaneously, assuming both the bearing and the coating trigger a level two warning, L 轴 =20, L 涂 =15, L 密 =100. Before adjustment: L 总 =0.35×100+0.4×20+0.25×15=35+8+3.75=46.75 days; the adjusted temporary weights are a=0.35, b=0.4+0.1=0.5, c=0.25+0.1=0.35, summing to 1.2; the normalized weights are: a=0.35 / 1.2≈0.292, b=0.5 / 1.2≈0.417, c=0.35 / 1.2≈0.292. L 总 =0.292×100+0.417×20+0.292×15=29.2+8.34+4.38=41.92 days.

[0071] The overall early warning output module determines the comprehensive early warning level based on the output of the early warning judgment module: When any component, bearing, or coating triggers a Level 2 warning, it is classified as a Level 2 overall warning, representing the highest level of urgency. When none of the components trigger a Level 2 warning but at least one component triggers a Level 1 warning, it is classified as a Level 1 overall warning. When none of the components trigger any warnings, the system is considered to be in a normal state. Overall warning signals can be pushed through a web-based or mobile visual maintenance platform to guide maintenance personnel in taking appropriate measures; for example, a Level 2 overall warning requires immediate shutdown and inspection.

[0072] During system operation, the data acquisition module collects relevant parameters of seals, bearings, and coatings, as well as images of coating defects, in real time; the data preprocessing module denoises, completes, and normalizes the parameter data to generate standardized data; the seal, bearing, and coating life prediction modules respectively use their trained neural network models to output the remaining service life of each component; the early warning judgment module determines the early warning level of each component based on the remaining service life; the model fusion module dynamically adjusts the weights according to the early warning level and calculates the overall online life prediction value of the hydraulic cylinder; and the overall early warning output module outputs the comprehensive early warning level.

[0073] Example 2 A method for predicting the online lifespan of a hydraulic cylinder includes: S1. Selection of Dedicated Sensors S101. Based on the actual operating conditions and monitoring requirements of the ultra-large hydraulic cylinder, the following sensors are selected: Laser displacement sensor: measuring range 0~25m, accuracy ±0.1mm, deployed at the bearing housing and piston rod, used to collect bearing swing angle and bearing wear data; Infrared thermal imager: Temperature measurement range -20~250℃, resolution 0.1℃, deployed around the sealed cavity and bearing housing to collect data on the temperature of the sealed cavity and the operating temperature of the bearing; Linear scanning camera: Defect recognition accuracy of 0.1mm, deployed on the outer wall of cylinder and piston rod surface, used to acquire images of coating surface; Oil contamination sensor: with a particle size detection accuracy of ≥2μm, deployed in the oil circulation pipeline of the sealed cavity to collect contaminant concentration data in the oil of the sealed cavity; Leakage sensor: measuring range 0~20mL / h, accuracy ±0.05mL / h, deployed at the bottom of the sealing cavity, used to collect leakage data of the sealing component; Load sensor: measuring range 0~6000t, accuracy ±0.1%FS, deployed below the bearing housing, used to collect data on hydraulic cylinder operating load and impact frequency.

[0074] S102. Sensors are deployed according to a plan of 3 points on the cylinder barrel, 4 points on the piston rod, 2 points on the sealing cavity, and 2 points on the bearing seat. The points are positioned to avoid interference areas and stress concentration points of the cylinder movement. The installation angle is 45°-60° to ensure no blind spots in monitoring. The sensors are fixed by stainless steel brackets, and the wiring terminals are sealed with waterproof aviation plugs to adapt to extreme marine environments.

[0075] S103. A data acquisition system is built using a Siemens SCALANCE XB-200 industrial Ethernet gateway. This gateway supports the Profinet protocol and has a communication rate of at least 100Mbps, allowing unified access to data from all sensors. An integrated hardware and software acquisition module with an integrated AD7606 data acquisition chip is configured, with a sampling frequency set to 10Hz. This frequency is determined based on the single extension / retraction cycle of the hydraulic cylinder, ensuring complete capture of parameter changes. Data transmission uses the AES-256 encryption protocol and is uploaded in real-time to a local redundant storage server and a cloud database. The local redundant storage server is an Advantech IPC-610L, with a storage capacity of at least 1TB and support for RAID1 redundancy backup, ensuring no data loss.

[0076] S104. Accumulate no less than 30,000 sets of valid operational data, covering different operating conditions such as heavy load, light load, and frequent sway angles, as well as different environmental conditions with temperatures ranging from -20 to 120℃ and contaminant concentrations ranging from 0 to 50 ppm. Collected parameters include core parameters such as sealing cavity temperature, leakage rate, oil contaminant concentration, bearing sway angle, operating load, wear rate, impact frequency, coating defect images, defect area, and corrosion depth, providing sufficient samples for subsequent data preprocessing and model training. S2, Data Preprocessing The collected raw monitoring data undergoes denoising, outlier removal, completion, and standardization to eliminate noise interference and data bias, ensuring the reliability and standardization of the data input to the model and providing a high-quality data source for model training. Specifically, this includes: The Kalman filter algorithm is used to denoise the acquired raw data. The system state equation and observation equation are initialized, and the process noise covariance Q = 1e⁻⁴ and the observation noise covariance R = 1e⁻³ are set. These parameters are determined through multiple experimental fittings, taking into account the sensor's rated accuracy and the intensity of marine electromagnetic interference. Based on the state estimate from the previous moment and the observed value at the current moment, the optimal estimate is calculated in real time, which can effectively eliminate the influence of marine electromagnetic interference and the sensor's own noise.

[0077] Based on the rated measurement accuracy of each sensor, the standard deviation σ of 100 consecutive sets of valid data for a single sensor is calculated. Data with a deviation exceeding 3σ is identified as abnormal data. For single-point abnormal data, the mean of five adjacent valid data points is used to complete the data. For three or more consecutive sets of abnormal data, the data is marked as sensor fault and a local audible and visual warning is triggered to ensure data integrity.

[0078] The denoised and completed valid data are subjected to min-max normalization. This valid data includes sealing cavity temperature, leakage rate, contaminant concentration, bearing swing angle, load, wear, coating defect area, and corrosion depth. Normalization maps the data to the [0,1] interval, effectively eliminating the influence of dimensions. The normalization formula is as follows: x '=( x - x min ) / ( x max - x min ), x This is the original data. x min , x max These are the minimum and maximum values ​​of the parameter, respectively, which are derived from the collected sample data.

[0079] The preprocessed valid data was divided into training, validation, and test sets in an 8:1:1 ratio. For coating defect image data, annotation tools were used to select and label defect areas in the images, clearly defining the defect area and corrosion depth, forming an annotated image dataset. The reliability of the preprocessed data was no less than 99%, and the single-parameter processing error did not exceed ±0.1%, making it directly usable for subsequent model training.

[0080] S3, Hydraulic Cylinder Life Prediction The cylinder life prediction in this invention includes seal life prediction, bearing life prediction, and coating life prediction; specifically as follows: S301, Predicted Sealing Life A backpropagation (BP) neural network is used to construct a seal life prediction model. The input parameters of the model are the core parameters related to the seal, and the output parameter is the remaining service life of the seal. The seal life prediction model adopts a four-layer feedforward neural network structure, and the specific parameters are as follows: Input layer: 3 neurons, corresponding to the pre-processed sealing cavity temperature, sealing leakage, and oil contaminant concentration. A linear transfer function is used to ensure that the input data is transmitted without distortion.

[0081] Hidden layers: 2 layers, 18 neurons per layer. The ReLU activation function is used to address the vanishing gradient problem and improve the model's training convergence speed. Its function expression is: f ( x )=max(0, x ).

[0082] Output layer: 1 neuron, with the output parameter being the remaining service life of the seal in hours. The activation function is the Sigmoid function, mapping the output value to the interval [0, 10000], which corresponds to the maximum design life of the seal. The Sigmoid function expression is as follows: f ( x )=1 / (1+e^(- x )).

[0083] When initializing the BP neural network structure, the network weights and biases are initialized using a random normal distribution. The weight values ​​range from [-0.1, 0.1], and the initial bias value is 0.01.

[0084] The model training process includes: 10,000 sets of preprocessed sealing-related data were selected, including 8,000 sets for training, 1,000 sets for validation, and 1,000 sets for testing. This sealing-related data includes sealing cavity temperature, sealing leakage, and oil contaminant concentration. The input layer receives the preprocessed sealing-related data, and the output layer outputs a label corresponding to the remaining service life of the seal. This label is determined through laboratory failure tests and is used to record the actual time from operation to failure of the seal under different parameter combinations.

[0085] The optimizer uses Adam gradient descent with a fixed learning rate of 0.01, a batch size of 32, and at least 500 training iterations. The loss function is mean squared error (MSE). MSE measures the deviation between predicted and actual values, and its expression is MSE = 1 / n × Σ(y i -ŷ i ) 2 n is the number of samples, y i For the actual remaining seal life, ŷ i These are the model's predicted values.

[0086] Training is stopped when the loss function value converges to no more than 0.001, or when the accuracy of the validation set does not improve for 20 consecutive iterations and the accuracy fluctuation does not exceed 0.5%, in order to avoid overfitting or underfitting of the model.

[0087] After training, the test set data is input into the model, the deviation between the predicted and actual values ​​is calculated, and the network weights and biases are adjusted to ensure that the model's prediction accuracy is not less than 90%, the mean absolute error is not more than 2.5 days, and the leakage-related prediction error is not more than 0.1 mL / h.

[0088] In one embodiment, the system further includes setting two levels of warnings: a Level 1 warning for sealing life corresponds to a remaining life of 30 to 60 days and a leakage rate of 2 to 5 mL / h; and a Level 2 warning for sealing life corresponds to a leakage rate of not less than 5 mL / h or a remaining life of not more than 30 days.

[0089] S302, Bearing Life Prediction Considering the temporal and correlational nature of bearing operating parameters, a CNN-BP hybrid neural network is used to construct a bearing life prediction model. This model employs a hybrid structure of CNN feature extraction layers and BP prediction layers, balancing feature extraction with accurate prediction. Specific parameters include: Input layer: 5 neurons, corresponding to the standardized operating load, bearing swing angle, standardized wear amount, impact frequency, and bearing operating temperature. The input data adopts a time-series splicing method to preserve the parameter change trend.

[0090] CNN Feature Extraction Layer: Contains 2 convolutional layers and 1 pooling layer. Convolutional layer 1 uses 3×1 convolutional kernels, with a total of 16 kernels and the activation function is ReLU. Convolutional layer 2 uses 3×1 convolutional kernels, with a total of 32 kernels and the activation function is ReLU. The pooling layer uses 2×1 max pooling to extract temporal features of the parameters, reduce data dimensionality, and avoid overfitting.

[0091] Backpropagation (BP) prediction layer: It consists of 2 hidden layers and 1 output layer. Each hidden layer has 16 neurons with ReLU activation function; the output layer has 1 neuron with Sigmoid activation function, which is used to map the features extracted by the CNN feature extraction layer to the remaining lifespan of the bearing.

[0092] Network initialization: Convolutional layer weights are initialized using a He normal distribution, and BP layer weights are initialized using a random normal distribution. The initial bias value is 0.01 for both.

[0093] The model training process includes: Eighty-thousand sets of pre-processed bearing-related operating loads, swing angles, wear amounts, impact frequencies, and temperatures were selected, including 6,400 sets for training, 800 sets for validation, and 800 sets for testing. The input layer receives the pre-processed operating load, swing angle, wear amount, impact frequency, and temperature data, while the output layer outputs a label corresponding to the bearing's remaining service life. This label, determined through laboratory failure tests, is used to record the actual bearing failure duration under different operating conditions.

[0094] The optimizer employs Adam gradient descent with a dynamically adjusted learning rate. The initial learning rate is 0.01, decreasing by 0.1 every 100 iterations. The batch size is 32, and the training iterations are at least 600. The loss function is mean squared error. L2 regularization with a coefficient of 0.001 is also introduced to suppress overfitting.

[0095] Training is stopped when the loss function value converges to no greater than 0.0008, or when the accuracy of the validation set does not improve for 25 consecutive iterations and the accuracy fluctuation does not exceed 0.4%.

[0096] Input the test set data into the model, calculate the prediction bias, adjust the convolution kernel size and BP layer weights to ensure that the model's prediction accuracy is not less than 91%, the prediction bias is not more than 8%, and the swing angle related detection error is not more than 0.5°.

[0097] In one embodiment, the system further includes setting two levels of warnings: the first level of bearing life warning corresponds to a remaining life of 30 to 60 days and a swing angle between ±20° and ±30°; the second level of bearing life warning corresponds to a swing angle exceeding the limit by ±30°, wear amount not less than 0.3mm, or remaining life not exceeding 30 days.

[0098] S303, Coating Life Prediction Combining coating defect image features and quantization parameters, a YOLOv8-CNN hybrid neural network is used to first identify coating defects and then predict the remaining lifetime. This involves a hybrid structure of a YOLOv8 defect identification layer and a CNN lifetime prediction layer, divided into two stages: defect identification and lifetime prediction. Specific parameters are as follows: The YOLOv8 defect recognition layer adopts the lightweight YOLOv8n network structure. The input is a standardized coating defect image of 640×640 pixels. The backbone layer uses the C2f module to extract image features. The neck layer completes multi-scale feature fusion through SPPF and PAN-FPN. The head layer is a decoupled head structure. The output is the defect area ratio and corrosion depth.

[0099] The input parameters for the CNN lifetime prediction layer are the standardized defect area ratio and corrosion depth identified by the YOLOv8 defect recognition layer, and the standardized coating surface temperature acquired by an infrared thermal imager, totaling three input neurons. The CNN layer contains two convolutional layers, one pooling layer, and one fully connected layer. The convolutional layers use 3×3 convolutional kernels, with 16 and 32 kernels respectively, and the activation function is ReLU for both. The fully connected layer has 16 neurons with the activation function ReLU. The output layer has one neuron with the activation function Sigmoid, used to output the remaining lifetime of the coating.

[0100] Network initialization: The YOLOv8 layers use pre-trained weights based on the COCO (Common Objects in Context) dataset, and the CNN layer weights are initialized using the He normal distribution with an initial bias value of 0.01.

[0101] The model training process includes: 12,000 images of coating defects were selected, including 9,600 images for training, 1,200 images for validation, and 1,200 images for testing. All images were labeled with the defect area ratio and corrosion depth. Simultaneously, corresponding coating temperature data were matched to form a joint dataset of images and quantified parameters, outputting a label for the remaining service life of the corresponding coating layer. This label was determined through accelerated aging tests in the laboratory and is used to record the actual failure time of the coating under different defect levels.

[0102] The YOLOv8 defect detection layer has an initial learning rate of 0.001, uses SGD as the optimizer, has a momentum of 0.937, a weight decay of 0.0005, and employs CIoU as the loss function. The CNN lifetime prediction layer has a learning rate of 0.01, uses Adam as the optimizer, has a batch size of 32, and undergoes at least 300 training iterations. Accuracy is validated using a validation set every 10 iterations, and mean squared error is used as the loss function.

[0103] Training is stopped when the defect identification accuracy is not less than 98%, the error in defect area and corrosion depth detection does not exceed 0.05 mm, and the lifetime prediction loss function value converges to no more than 0.001.

[0104] The test set images and quantization parameters were input into the model to verify the accuracy of defect identification and lifetime prediction. The network weights were adjusted to ensure that the coating lifetime prediction reproduction accuracy was not less than 92% and the corrosion depth detection error was not more than 0.05 mm.

[0105] The trained model receives real-time coating images, identifies defect information, extracts quantification parameters, and combines this with coating temperature data to output the remaining service life of the coating. Simultaneously, two warning thresholds are set: Level 1 warning corresponds to a defect area of ​​3% to 5% and a corrosion depth of 0.05 to 0.1 mm; Level 2 warning corresponds to a defect area of ​​no less than 5% or a corrosion depth of no less than 0.1 mm.

[0106] S304. Model Fusion An attention-weighted fusion method is used to integrate the life prediction results of the three major components to obtain the overall online life prediction value of the hydraulic cylinder. Compared with the traditional fixed-weighted fusion method, this fusion method is more in line with actual working conditions and has higher prediction accuracy. The specific fusion process includes: An attention mechanism is introduced to dynamically adjust the weights of each model's output based on the real-time operating status and prediction accuracy of each component. The core logic is to increase the weight of components that trigger secondary warnings and components with higher prediction accuracy, ensuring that the fusion results are more targeted.

[0107] The weighting rules are as follows: The basic weights are set as follows: Based on laboratory failure test statistics, the basic weights are determined according to the degree of influence of each component on the overall life of the cylinder. The basic weight of the seal is a0=0.35, the basic weight of the bearing is b0=0.4, and the basic weight of the coating is c0=0.25. The sum of the three is 1.

[0108] The dynamic weight adjustment strategy includes: First, if a component triggers a level 2 warning, its weight increases by 0.1 from the base weight, while the weights of other components decrease accordingly, with the decrease distributed proportionally to the base weights; Second, if a component triggers a level 1 warning, its weight increases by 0.05 from the base weight, while the weights of other components decrease accordingly; Third, if a component does not trigger a warning, its weight remains unchanged from the base weight; Fourth, based on the prediction accuracy of each model, for every 1% increase in accuracy, the weight increases by 0.01, ensuring that models with higher accuracy have a greater impact on the fusion results.

[0109] The fusion computing formula includes: L 总 =a×L 密 +b×L 轴 +c×L涂 Where a, b, and c are the weights of the seal, bearing, and coating, respectively, and the sum of the three is 1, L 密 L 轴 L 涂 These are the real-time remaining service lives output by the seal life prediction model, bearing life prediction model, and coating life prediction model, respectively, in hours. The calculated L... 总 The value is rounded to one decimal place and used as the overall online life prediction value for the hydraulic cylinder.

[0110] Based on the warning status of each component and the fused life prediction value, there are two overall cylinder warning systems: Level 1 and Level 2. A Level 1 overall cylinder warning occurs when none of the components trigger their corresponding Level 2 warnings (sealing, bearing, or coating), but at least one component triggers its corresponding Level 1 warning (sealing, bearing, or coating), or the overall cylinder life prediction value is between 30 and 60 days. A Level 2 overall cylinder warning occurs when any component triggers its corresponding Level 2 warning (sealing, bearing, or coating), or the overall cylinder life prediction value does not exceed 30 days. If none of the components trigger a warning and the overall cylinder life prediction value exceeds 60 days, it is considered a normal state.

[0111] After a Level 1 warning for the hydraulic cylinder is triggered, maintenance personnel must conduct a comprehensive inspection of the corresponding component within 24 hours, focusing on verifying the changing trends of the warning-related parameters of the component, investigating potential faults, and increasing the real-time monitoring frequency of the component to ensure that the parameters do not fluctuate abnormally. After a Level 2 warning for the entire hydraulic cylinder is triggered, the heavy-load operation of the hydraulic cylinder must be stopped immediately and switched to a light-load or shutdown state. Maintenance personnel must conduct emergency repairs within 4 hours, inspecting, repairing, or replacing the component that triggered the Level 2 warning. The hydraulic cylinder can only be restarted for normal operation after the component's condition returns to normal and the warning is lifted, in order to avoid the entire hydraulic cylinder failure due to component failure.

[0112] To further verify the feasibility and reliability of the above attention-weighted fusion method and ensure that the overall life prediction result of the fused cylinder can meet the actual application requirements, this embodiment conducts effect verification through laboratory simulation. 300 sets of actual cylinder life data are selected and compared one by one with the fused prediction value. During the verification process, the core indicators are strictly controlled to ensure that the overall life prediction deviation of the cylinder does not exceed 7%, the early warning response time does not exceed 1 second, and there are no missed or false alarms. This fully demonstrates that the fusion method can accurately output the overall life prediction value of the cylinder and is fully adapted to the actual needs of online life prediction.

[0113] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0114] This invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the various steps of the method described in this invention.

[0115] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A hydraulic cylinder online life prediction system, characterized in that, include: The data acquisition module is used to collect sealing-related parameters, bearing-related parameters, coating-related parameters, and coating defect images during the operation of the hydraulic cylinder. The data preprocessing module is used to perform Kalman filtering for noise reduction, outlier removal and completion, and normalization on the sealing-related parameters, bearing-related parameters, and coating-related parameters to generate standardized data. The sealing life prediction module is used to input the sealing-related parameters in the standardized data into the sealing life prediction model based on BP neural network, and output the remaining service life of the seal. The bearing life prediction module is used to input the bearing-related parameters in the standardized data into the bearing life prediction model based on CNN-BP hybrid neural network, and output the remaining service life of the bearing. The coating lifetime prediction module is used to input coating defect images and coating-related parameters from the standardized data into a coating lifetime prediction model based on YOLOv8-CNN hybrid neural network, and output the remaining lifetime of the coating. The model fusion module is used to perform fusion calculations on the remaining service life of the seal, the remaining service life of the bearing, and the remaining service life of the coating through a dynamic weighted fusion method, so as to obtain the overall online life prediction value of the hydraulic cylinder.

2. The online life prediction system for hydraulic cylinders as described in claim 1, characterized in that, The model fusion module calculates the overall online life prediction value of the hydraulic cylinder using the following formula: L 总 =a×L 密 +b×L 轴 +c×L 涂 ; Where L 总 L is the predicted online lifespan of the hydraulic cylinder. 密 L 轴 L 涂 These represent the remaining service life of the seal, the remaining service life of the bearing, and the remaining service life of the coating, respectively; a, b, and c represent the weights of the seal, bearing, and coating, respectively.

3. The online life prediction system for hydraulic cylinders as described in claim 1, characterized in that, It also includes a warning determination module, used to determine the warning level based on the remaining service life of the seal, the remaining service life of the bearing, and the remaining service life of the coating, specifically including: When the remaining service life of the seal is lower than the set secondary warning threshold, the seal is determined to trigger a secondary warning; when the remaining service life of the seal is lower than the set primary warning threshold but not lower than the secondary warning threshold, the seal is determined to trigger a primary warning. When the remaining service life of the bearing is lower than the set bearing secondary warning threshold, the bearing is determined to trigger a secondary warning; when the remaining service life of the bearing is lower than the set bearing primary warning threshold but not lower than the bearing secondary warning threshold, the bearing is determined to trigger a primary warning. When the remaining service life of the coating is lower than the set secondary warning threshold, the coating is determined to trigger a secondary warning; when the remaining service life of the coating is lower than the set primary warning threshold but not lower than the secondary warning threshold, the coating is determined to trigger a primary warning.

4. The online life prediction system for hydraulic cylinders as described in claim 3, characterized in that, The model fusion module further includes a weight allocation unit, which executes the following dynamic adjustment strategy for the weights of seals, bearings, and coatings: Set the base weights for seals, bearings, and coatings; for each component among seals, bearings, and coatings, if a component triggers a level 2 warning, then adjust the weight of that component to 0.

1. If a component triggers a Level 1 warning, the corresponding weight adjustment amount for that component is set to 0.05; the base weight of each component is added to the corresponding adjustment amount to obtain the unnormalized weight of each component; then the unnormalized weight of each component is normalized to obtain the final weight of each component.

5. The online life prediction system for hydraulic cylinders as described in claim 1, characterized in that, The sealing life prediction module uses a BP neural network-based sealing life prediction model, which is a four-layer feedforward neural network structure, specifically including: Input layer: Used to receive normalized sealing-related parameters, match the data dimensions, and then pass them to the first hidden layer; First hidden layer: used to extract local degradation feature vectors from sealing-related parameters, the local degradation features including sealing cavity temperature change trend, leakage fluctuation characteristics, and contaminant concentration accumulation characteristics; ReLU activation function is used to nonlinearly activate the local degradation features; The second hidden layer is used to receive the activated local degradation feature vector output by the first hidden layer, propose a global nonlinear mapping feature vector, and perform nonlinear activation on the global nonlinear mapping feature vector using the ReLU activation function; the global nonlinear mapping feature vector represents the mapping relationship between the remaining service life of the seal and various seal-related parameters; Output layer: Used to receive the activated global nonlinear mapping feature vector output from the second hidden layer, map it to the first preset interval using the Sigmoid activation function, and output the remaining service life of the seal; During the training process of the sealing life prediction model based on the BP neural network, the Adam gradient descent method is used to backpropagate and iteratively update the network weights and biases, and minimize the mean square error loss function. Training stops when the loss function value converges to no greater than a set value.

6. The online life prediction system for hydraulic cylinders as described in claim 1, characterized in that, The bearing life prediction model based on a CNN-BP hybrid neural network in the bearing life prediction module includes: Input layer: Used to concatenate normalized bearing-related parameters into time-series data using a time-series concatenation method, and then pass the data to the CNN feature extraction layer after matching the data dimensions; CNN Feature Extraction Layer: Used to learn multi-scale temporal features of the bearing degradation process from the time-series data to extract degradation characteristics strongly correlated with the remaining service life of the bearing; it includes two convolutional layers and one pooling layer; the first convolutional layer is used to extract local degradation feature maps of the bearing from the time-series data, the local degradation feature maps including bearing swing angle fluctuation patterns, load impact waveform features, and wear change trends; the second convolutional layer is used to perform deep feature extraction on the local degradation feature maps and output a deep degradation feature map; the pooling layer is used to reduce the dimensionality of the deep degradation feature map and output a pooled degradation feature map, and then flatten the pooled degradation feature map into a one-dimensional feature vector before outputting it to the BP prediction layer; BP prediction layer: used to map the one-dimensional feature vector to a predicted value of the bearing's remaining service life; it includes a first hidden layer, a second hidden layer, and an output layer. The activation function of the first hidden layer is ReLU, used to receive the one-dimensional feature vector, extract the nonlinear combination features of the bearing degradation process, and output the combination feature vector to the second hidden layer. The activation function of the second hidden layer is ReLU, used to receive the combination feature vector and map it to a global nonlinear relationship feature vector before outputting it to the output layer. The global nonlinear relationship feature vector represents the complex mapping relationship between the bearing's remaining service life and various bearing-related parameters. Output layer: used to map the global nonlinear relationship feature vector to a second preset interval using the Sigmoid activation function, and output the remaining service life of the bearing; During the training process of the bearing life prediction model based on CNN-BP hybrid neural network, the optimizer adopts the Adam gradient descent method, and the loss function adopts the mean squared error.

7. The online life prediction system for hydraulic cylinders as described in claim 1, characterized in that, The coating lifetime prediction model based on the YOLOv8-CNN hybrid neural network in the coating lifetime prediction module includes: YOLOv8 Defect Recognition Layer: Used to detect defects in coating defect images, outputting the defect area ratio and corrosion depth to the CNN lifetime prediction layer; the defect area ratio represents the extent of coating damage, and the corrosion depth represents the severity of coating damage; CNN lifetime prediction layer: used to map the defect area ratio, corrosion depth, and coating surface temperature to predicted values ​​of the remaining lifetime of the coating; includes a first convolutional layer, a second convolutional layer, a pooling layer, a fully connected layer, and an output layer, wherein: The first convolutional layer uses ReLU as the activation function to extract local features of the defect area ratio, corrosion depth, and coating surface temperature, and outputs the first convolutional feature map. The first convolutional feature map encodes the initial coupling relationship between coating damage parameters and temperature. The second convolutional layer uses ReLU as the activation function to extract deep features from the first convolutional feature map and outputs a second convolutional feature map. The second convolutional feature map represents the nonlinear evolution law in the coating degradation process. Pooling layer: used to reduce the dimensionality of the second convolutional feature map and output a pooled feature map; the pooled feature map retains the main features of coating degradation while reducing computational parameters; Fully connected layer: used to map the pooled feature map into a feature vector and output it to the output layer; the feature vector integrates global degradation information under the coupling of coating damage and temperature; Output layer: The Sigmoid activation function is used to map the feature vector to a third preset interval, and the remaining lifespan of the coating is output.

8. The online life prediction system for hydraulic cylinders as described in claim 7, characterized in that, The YOLOv8-CNN hybrid neural network was trained in the following way: Multiple coating defect images were selected and divided into training, validation and test sets in an 8:1:1 ratio. The defect area ratio and corrosion depth were labeled on the images, and the corresponding coating temperature data were matched and associated with the coating's remaining service life label. The YOLOv8 defect identification layer uses stochastic gradient descent as the optimizer and full intersection-over-union loss as the loss function. The CNN lifetime prediction layer uses Adam gradient descent as the optimizer and mean squared error as the loss function. Training stops when the defect identification accuracy is not lower than the set ratio, the defect area and corrosion depth detection errors do not exceed the corresponding set thresholds, and the lifetime prediction loss function value converges to less than the set value. The test set images were input into the YOLOv8-CNN model to verify the accuracy of defect identification and lifetime prediction. The network weights were adjusted to ensure that the accuracy of coating lifetime prediction reproduction was not lower than the set ratio and that the corrosion depth detection error did not exceed the corresponding set threshold.

9. A method for predicting the online lifespan of a hydraulic cylinder based on the system described in claim 1, characterized in that, include: Collect sealing-related parameters, bearing-related parameters, coating-related parameters, and coating defect images during the operation of the hydraulic cylinder; Kalman filtering, outlier removal and completion, and normalization are performed on the sealing-related parameters, bearing-related parameters, and coating-related parameters to generate standardized data. The sealing-related parameters in the standardized data are input into the sealing life prediction model based on BP neural network, and the remaining service life of the seal is output. The bearing-related parameters in the standardized data are input into the bearing life prediction model based on CNN-BP hybrid neural network, and the remaining service life of the bearing is output. The coating defect image and coating-related parameters from the standardized data are input into a coating lifetime prediction model based on a YOLOv8-CNN hybrid neural network, and the remaining lifetime of the coating is output. The remaining service life of the seal, the remaining service life of the bearing, and the remaining service life of the coating are calculated by dynamically weighted fusion to obtain the overall online life prediction value of the hydraulic cylinder.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the online life prediction method for hydraulic cylinders as described in claim 9.

Images