A Method and System for Predicting the Remaining Service Life of Gear Pumps Based on Spatiotemporal Graph Modeling

Abstract

This invention provides a method and system for predicting the remaining service life (RUL) of gear pumps based on spatiotemporal graph modeling. The invention employs a technical approach of multi-source data acquisition, dynamic graph modeling, spatiotemporal feature learning, and RUL probability prediction. It characterizes the time-varying characteristics of parameter coupling through a dynamic graph learning mechanism, and mines deep degradation features using a spatiotemporal graph convolutional network. This constructs a dual-driven RUL prediction model based on both mechanism and data. The time-varying graph structure adaptively characterizes the evolutionary laws of parameter correlation. This invention efficiently integrates spatial correlation and temporal evolution information of parameters, improving the feature learning efficiency and dynamic operating condition generalization ability of high-dimensional noisy data. Ultimately, it achieves high-precision prediction of the RUL of gear pumps throughout their entire lifecycle, providing a quantitative reliability assessment basis for maintenance decisions.

Description

Technical Field

[0001] This invention relates to the field of gear pump reliability assessment technology, and more specifically, to a method and system for predicting the remaining service life of gear pumps based on spatiotemporal graph modeling. Background Technology

[0002] As the core power source of a ship's hydraulic transmission system, the reliability of gear pumps directly determines the operational accuracy and service safety of the main equipment. During long-term service under complex conditions, gear pumps face the coupled effects of multiple physical failure modes, including tooth surface adhesive wear, bearing contact fatigue, and seal aging failure. This manifests as degradation characteristics such as decreased volumetric efficiency, increased internal leakage, and abnormal casing vibration amplitude, ultimately leading to system downtime. According to statistics from the *Chinese Journal of Construction Machinery* in 2024, gear pump failures accounted for 42% of all hydraulic system failures, with direct economic losses from a single unplanned downtime reaching 18,000 RMB per hour. Therefore, constructing a high-precision RUL (Reliability, Upgrade, and Duration) prediction model has significant engineering and academic value for implementing predictive maintenance and reducing the total lifecycle maintenance cost.

[0003] Existing gear pump RUL prediction technologies face two major bottlenecks: First, the challenge of dynamically coupled modeling multiple performance parameters. Traditional methods often employ stochastic process models such as Wiener and Gamma processes to model the degradation of a single performance index (e.g., volumetric efficiency), failing to characterize the time-varying correlation mechanisms between multiple parameters, such as tooth surface wear and vibration characteristics, and oil metal particle concentration. Although some studies have attempted multivariate modeling (e.g., the multivariate t-distribution degradation model proposed by Liu et al., and the multi-source state-space model developed by Feng et al.), they all assume that the parameter correlation matrix is ​​constant, neglecting the evolutionary characteristics of coupling strength caused by fault propagation effects in the later stages of degradation. Second, the low efficiency of feature learning for high-dimensional non-Euclidean data. While deep learning methods (e.g., Long Short-Term Memory Networks (LSTM) and Convolutional Neural Networks (CNN)) can extract temporal or local features, they are limited to the feature extraction paradigm of Euclidean space data and cannot adapt to the inherent non-Euclidean geometric structure characteristics of multi-source monitoring data, resulting in a significant decrease in generalization performance under dynamic operating conditions.

[0004] The development of multi-source monitoring technologies such as infrared thermal imaging, vibration spectrum analysis, and oil physicochemical testing has provided data support for accurate assessment. However, existing technologies have failed to achieve deep integration of data structure and degradation mechanism: the vibration image coding method proposed by Guan et al. only focuses on single-mode feature extraction and does not explore cross-modal parameter correlation information. Therefore, a method and system for predicting the remaining service life of gear pumps based on spatiotemporal graph modeling is proposed. Summary of the Invention

[0005] The purpose of this invention is to address the problems raised in the existing background technology. To achieve the above-mentioned objective, this invention provides the following technical solution: a method for predicting the remaining service life of a gear pump based on spatiotemporal graph modeling, comprising the following steps: S1 Multi-source degradation data acquisition and preprocessing: acquiring gear pump performance parameters (outlet flow rate, outlet pressure, input speed, input torque), housing radial / axial vibration signals, and oil analysis data (metal particle content, viscosity, acid value), and after standardization and time scale alignment, constructing a multi-source data matrix with dimension N×M, where N is the time step and M is the total number of monitored parameters; S2 Dynamic Graph Structure Construction: The initial graph structure is constructed based on the dual constraints of data-driven and physical mechanisms. The initial graph construction method in S2 is as follows: the data-driven adjacency matrix A1 is obtained by fusing Pearson correlation coefficient (linear association) and mutual information (non-linear association), and the prior adjacency matrix A2 is constructed by combining the gear pump degradation mechanism. The initial adjacency matrix A = 0.6A1 + 0.4A2 is obtained by weighted fusion, and significant associated edges are retained by using a threshold τ = 0.1. A time-varying adjacency matrix is ​​generated by an encoder-decoder architecture, and an adaptive update of the graph structure is achieved by combining a two-layer alternating optimization strategy, resulting in a time-varying graph sequence G(t)=(V,E(t),X(t)), where V is the parameter node set, E(t) is the edge set at time t, and X(t) is the node feature matrix. S3 Spatiotemporal Feature Deep Learning: It uses a 3-layer Chebyshev polynomial approximation graph convolution to extract spatial correlation features, captures temporal evolution information through gated recurrent units (GRU) and temporal attention mechanism, and obtains global degradation features after multi-scale feature fusion. S4 Remaining Useful Life Prediction: A hybrid density network (MDN) is used to model the probability distribution of remaining useful life (RUL) and output the expected value of RUL and the 90% confidence interval.

[0006] As a preferred technical solution of the present invention, the vibration signal preprocessing in S1 includes: sampling frequency of 10kHz, extracting 12 features such as root mean square, kurtosis, and meshing frequency amplitude, downsampling to 1Hz to align with performance parameters; missing values ​​are filled by local weighted regression interpolation, and performance parameters and vibration features are standardized by z-score.

[0007] As a preferred technical solution of the present invention, the dynamic graph generation in S2 adopts an encoder-decoder architecture: the encoder is a 3-layer perceptron, which maps the node features to a 64-dimensional latent space to obtain the node embedding h; after normalization, a symmetric non-negative dynamic adjacency matrix A(t) is obtained.

[0008] As a preferred technical solution of the present invention, the spatial graph convolution in S3 is approximately implemented using a K=3rd order Chebyshev polynomial to achieve spectral domain convolution, combined with residual connections: ,in, It is a non-linear activation function. For summation, These are the training parameters for the convolutional layer. Indicates performing an operation on the Laplace matrix. Polynomial transformation of order 1; For the normalized graph Laplace matrix, The node feature matrix is ​​constructed using three convolutional layers with hidden dimensions of 64→128→64 respectively.

[0009] As a preferred technical solution of the present invention, the multi-scale fusion in S3 adopts three parallel spatiotemporal convolution branches, which process the input data with time steps of 2, 4 and 8 respectively, and extract features F1, F2 and F3 at different scales.

[0010] The gear pump remaining service life prediction system based on spatiotemporal graph modeling includes: a multi-source data acquisition module, which consists of an acceleration sensor, a pressure sensor, a flow sensor, and an oil sampling device, supporting synchronous acquisition and data transmission of multiple parameters; and a dynamic graph modeling module, which runs the graph generation and optimization algorithm of step S2 in claim 1 and outputs a time-varying graph structure sequence G(t). Spatiotemporal feature learning module: equipped with a 3-layer Chebyshev graph convolution and GRU network, performing spatiotemporal feature extraction in step S3 of claim 1; Prediction output module: runs a hybrid density network, outputting the expected value of remaining lifetime (RUL), 90% confidence interval and alarm signal.

[0011] As a preferred technical solution of the present invention, the data storage and management module uses a MySQL database to store historical monitoring data and model parameters, and supports retrospective querying of the evolution process of degradation curves and graph structures.

[0012] As a preferred technical solution of the present invention, the multi-source data acquisition module has an accelerometer with a range of ±50g and an accuracy of ±0.5%, a pressure sensor with a range of 0-30MPa and an accuracy of 0.2, supports PoE power supply and Ethernet transmission, and has a delay of <100ms; the prediction output module supports RS485 / Modbus protocol access to the SCADA system.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: Excellent accuracy in dynamic coupling modeling: The time-varying graph structure can adaptively characterize the evolution of parameter correlations. The learned graph density dynamically increases from 0.23 in the healthy state to 0.46 in the near-failure state, which is highly consistent with the physical mechanism of gear pump failure propagation. On the NASA C-MAPSS FD001 dataset, the RUL prediction RMSE is 9.42, which is 5.6% and 20.6% lower than MTGNN (RMSE=9.98) and GCN-LSTM (RMSE=11.87), respectively. On the FD004 complex working condition dataset, the RMSE is 16.28, which is a significant improvement in generalization performance compared with existing methods. Excellent spatiotemporal feature learning efficiency: The fusion architecture of 3-layer graph convolution and GRU compresses the feature dimension by 60% and improves the model training convergence speed by 40%; the multi-scale fusion mechanism effectively suppresses dynamic working condition interference. In complex scenarios with speed fluctuations of ±10% and load fluctuations of ±20%, the prediction accuracy decreases by ≤5%, which is significantly better than the static graph model (accuracy decrease of 15%). High reliability of uncertainty quantification: The actual coverage of the 90% confidence interval output by the MDN prediction head reaches 88.7%, and the Pearson correlation coefficient between the prediction standard deviation and the actual error is 0.82, providing a quantitative basis for risk decision-making; the fault identification rate in the early stage of degradation (first 10% of the life cycle) reaches 92.3%, and the false negative rate is <3%, meeting the needs of early warning engineering. Highly practical for engineering applications: The multi-source data acquisition module has an IP67 protection rating and is suitable for harsh industrial environments ranging from -20℃ to 85℃; the core hardware cost is less than 12,000 yuan, which is 85% lower than imported similar systems (more than 80,000 yuan); the data processing latency of the entire process is less than 1.5s, it supports network monitoring of 100 devices, and the annual maintenance cost of a single gear pump is reduced by 35%, demonstrating significant engineering application value. Attached Figure Description

[0014] Figure 1 A flowchart of the method provided by this invention; Figure 2 The system module data block diagram provided by this invention; Figure 3 The hardware data block diagram provided for this invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention.

[0016] Therefore, the following detailed description of the embodiments of the present invention is not intended to limit the scope of the claimed invention, but merely illustrates some embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention. It should be noted that, in the absence of conflict, the embodiments and features and technical solutions in the embodiments of the present invention can be combined with each other. It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0017] Example 1: A method for predicting the remaining service life of a gear pump based on spatiotemporal graph modeling, including the following steps: S1 Multi-source degradation data acquisition and preprocessing: Acquire gear pump performance parameters (outlet flow rate, outlet pressure, input speed, input torque), radial / axial vibration signals of the housing, and oil analysis data (metal particle content, viscosity, acid value). After standardization and time scale alignment, construct a multi-source data matrix with dimension N×M, where N is the time step and M is the total number of monitored parameters; S2 Dynamic Graph Structure Construction: The initial graph structure is constructed based on the dual constraints of data-driven and physical mechanisms. The initial graph construction method in S2 is as follows: the data-driven adjacency matrix A1 is obtained by fusing Pearson correlation coefficient (linear association) and mutual information (non-linear association), and the prior adjacency matrix A2 is constructed by combining the gear pump degradation mechanism. The initial adjacency matrix A = 0.6A1 + 0.4A2 is obtained by weighted fusion, and significant associated edges are retained by using a threshold τ = 0.1. A time-varying adjacency matrix is ​​generated by an encoder-decoder architecture, and an adaptive update of the graph structure is achieved by combining a two-layer alternating optimization strategy, resulting in a time-varying graph sequence G(t)=(V,E(t),X(t)), where V is the parameter node set, E(t) is the edge set at time t, and X(t) is the node feature matrix. S3 Spatiotemporal Feature Deep Learning: It uses a 3-layer Chebyshev polynomial approximation graph convolution to extract spatial correlation features, captures temporal evolution information through gated recurrent units (GRU) and temporal attention mechanism, and obtains global degradation features after multi-scale feature fusion. S4 Remaining Useful Life Prediction: A hybrid density network (MDN) is used to model the probability distribution of remaining useful life (RUL) and output the expected value of RUL and the 90% confidence interval.

[0018] The vibration signal preprocessing in S1 includes: sampling frequency of 10kHz, extracting 12 features such as root mean square, kurtosis, and meshing frequency amplitude, downsampling to 1Hz to align with performance parameters; missing values ​​are filled by local weighted regression interpolation, and performance parameters and vibration features are standardized using z-score.

[0019] In S2, dynamic graph generation adopts an encoder-decoder architecture: the encoder is a 3-layer perceptron that maps node features to a 64-dimensional latent space to obtain node embedding h; after normalization, a symmetric non-negative dynamic adjacency matrix A(t) is obtained.

[0020] In S3, spatial graph convolution is approximated using a K=3rd order Chebyshev polynomial, which is combined with residual connections. ,in, It is a non-linear activation function. For summation, These are the training parameters for the convolutional layer. Indicates performing an operation on the Laplace matrix. Polynomial transformation of order 1; For the normalized graph Laplace matrix, The node feature matrix is ​​constructed using three convolutional layers with hidden dimensions of 64→128→64 respectively.

[0021] In S3, multi-scale fusion uses three parallel spatiotemporal convolutional branches, which process the input data with time steps of 2, 4, and 8 respectively, and extract features F1, F2, and F3 at different scales.

[0022] The gear pump remaining service life prediction system based on spatiotemporal graph modeling includes: a multi-source data acquisition module, which consists of an acceleration sensor, a pressure sensor, a flow sensor, and an oil sampling device, supporting synchronous acquisition and data transmission of multiple parameters; and a dynamic graph modeling module, which runs the graph generation and optimization algorithm of step S2 in claim 1 and outputs a time-varying graph structure sequence G(t). Spatiotemporal feature learning module: equipped with a 3-layer Chebyshev graph convolution and GRU network, performing spatiotemporal feature extraction in step S3 of claim 1; Prediction output module: runs a hybrid density network, outputting the expected value of remaining lifetime (RUL), 90% confidence interval and alarm signal.

[0023] Data storage and management module: It uses a MySQL database to store historical monitoring data and model parameters, and supports retrospective querying of the evolution process of degradation curves and graph structures.

[0024] The multi-source data acquisition module includes an accelerometer with a range of ±50g and an accuracy of ±0.5%, a pressure sensor with a range of 0-30MPa and an accuracy of 0.2%, supports PoE power supply and Ethernet transmission, and has a latency of <100ms; the predictive output module supports RS485 / Modbus protocol access to the SCADA system.

[0025] Experimental Example: The experimental object was a CB-B100 external gear pump (rated pressure 16MPa, displacement 100mL / r, rated speed 1500r / min). An integrated power-loading-monitoring test platform was built: power system (15kW variable frequency motor, speed adjustment range 500-2000r / min, steady-state accuracy ±0.5%); loading system (electromagnetic overflow valve, loading range 0-16MPa, response time <50ms); monitoring system (3 sets of IEPE acceleration sensors installed on the front and rear end covers and the middle of the pump body, sampling frequency 10kHz; pressure / flow sensors installed on the inlet and outlet pipelines, accuracy classes 0.2 and 0.5 respectively; oil sampling cycle 24h). The model's generalization ability was verified using the NASA C-MAPSS turbine engine degradation dataset (FD001-FD004). This dataset contains 14 valid sensor parameters, where FD001 represents a single-condition single-fault mode, and FD004 represents six-condition dual-fault modes.

[0026] Data Acquisition and Preprocessing: Gear pump experiment: The rated operating conditions of 1500 r / min and 10 MPa pressure were set, and data were collected once per hour until the volumetric efficiency dropped to 70% (failure threshold). The full life cycle data of 5 pumps (about 120 time steps per pump) were obtained; the speed (1200 r / min, 1800 r / min) and pressure (8 MPa, 14 MPa) were adjusted to construct a dynamic operating condition dataset; C-MAPSS data processing: 7 constant sensors were removed, and 14 degradation-related parameters were retained; min-max normalization was used to [0,1], and piecewise linear RUL labels were set (RUL>125 was set to 125); the training / validation / test sets were divided into 8:1:1 to ensure that data from a single device did not cross sets.

[0027] Model training and parameter setting: Training environment configuration: Intel Core i7-12700 processor (20 cores, 28 threads), NVIDIA RTX 3060 graphics card (12GB VRAM), 16GB DDR5 memory, developed based on PyTorch 1.12 deep learning framework; Model hyperparameters were determined through grid search optimization: 3 graph convolutional layers, 64 GRU hidden dimensions, 30 time window length, 32 batch size, initial learning rate 0.001 (decreasing using cosine annealing strategy), 200 training epochs (early stopping strategy: training is terminated if the validation set loss does not decrease for 3 consecutive epochs); Dynamic graph parameters: 64 node embedding dimensions, 4 attention heads, sparsity coefficient λ1=0.01, smoothness coefficient λ2=0.005.

[0028] Analysis of experimental results: Prediction accuracy comparison: The prediction accuracy comparison experiment used statistical results from 5 independent repeated experiments. On the FD001 dataset, the RUL prediction RMSE of the model of this invention was 9.42±0.35, which was reduced by 41.6%, 20.6%, and 5.6% compared with LSTM (16.14±0.62), GCN-LSTM (11.87±0.48), and MTGNN (9.98±0.32), respectively. On the FD004 complex working condition dataset, the model RMSE was 16.28±0.52, which was reduced by 18.1% and 9.2% compared with DA-RNN (19.87±0.71) and TempGCN (17.92±0.45), respectively. In the gear pump physical experiment, the mean absolute error (MAE) of RUL prediction under rated working conditions was 4.8±0.6h, and the MAE under dynamic working conditions was 6.2±0.8h, both of which met the industrial-grade prediction accuracy requirements (MAE≤8h).

[0029] Dynamic graph structure verification: The evolution characteristics of the dynamic graph structure are highly consistent with the degradation mechanism of the gear pump: In the healthy state (RUL > 100h), the graph density is 0.23 and the average node degree is 3.21, retaining only parameter connections within the same functional module (such as the coupling relationship between outlet pressure and flow), reflecting the relatively independent operating characteristics of each subsystem in the healthy state; in the mid-degradation stage (50 < RUL ≤ 100h), the graph density increases to 0.35 and the average node degree is 4.86, with a significant increase in cross-module connections (such as the correlation between vibration characteristics and oil metal particle content), corresponding to the enhanced parameter coupling caused by fault propagation; in the near-failure stage (RUL ≤ 50h), the graph density reaches 0.46 and the average path length decreases to 1.65, forming a densely associated network, reflecting the typical characteristics of cascading failure of the system. The critical node importance analysis shows that the attention weight of tooth surface wear-related parameters such as outlet pressure increases from 0.09 to 0.18, consistent with the evolution law of the dominant failure factors.

[0030] Ablation experiments: The ablation experiments used the controlled variable method to quantify the contribution of each core component. After removing the dynamic graph learning module, the model RMSE increased to 11.87±0.51 (performance decreased by 26.0%), and the prediction error in the later stages of degradation increased significantly (RMSE difference > 40%), verifying the necessity of time-varying graph structures for characterizing coupled evolution. After removing the temporal attention mechanism, RMSE=10.56±0.43 (performance decreased by 12.1%), and the model's response delay to degradation inflection points increased by about 8 time steps. After removing the multi-scale fusion module, RMSE=10.26±0.39 (performance decreased by 8.6%), and the robustness to high-frequency noise was significantly reduced. After removing the physical mechanism constraints, RMSE=10.83±0.47 (performance decreased by 15.1%), and the learned graph structure showed physical and logical conflicts. Network depth experiments show that the 3-layer graph convolutional network has the best performance (RMSE=9.42), while the 4-layer and above networks exhibit oversmoothing (the cosine similarity of node features increases from 0.42 to 0.71).

[0031] The above embodiments are only used to illustrate the present invention and are not intended to limit the technical solutions described herein. Although the present invention has been described in detail with reference to the above embodiments, the present invention is not limited to the specific embodiments described above. Therefore, any modifications or equivalent substitutions to the present invention, as well as all technical solutions and improvements that do not depart from the spirit and scope of the invention, are covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the remaining service life of a gear pump based on spatiotemporal graph modeling, characterized in that, Includes the following steps: S1 Multi-source Degradation Data Acquisition and Preprocessing: Gear pump performance parameters (outlet flow rate, outlet pressure, input speed, input torque), housing radial / axial vibration signals, and oil analysis data (metal particle content, viscosity, acid value) are collected. After standardization and time scale alignment, a multi-source data matrix with dimension N×M is constructed, where N is the time step and M is the total number of monitored parameters. S2 Dynamic Graph Structure Construction: The initial graph structure is constructed based on the dual constraints of data-driven and physical mechanisms. The initial graph construction method in S2 is as follows: the data-driven adjacency matrix A1 is obtained by fusing Pearson correlation coefficient (linear association) and mutual information (non-linear association), and the prior adjacency matrix A2 is constructed by combining the gear pump degradation mechanism. The initial adjacency matrix A = 0.6A1 + 0.4A2 is obtained by weighted fusion, and significant associated edges are retained by using a threshold τ = 0.

1. A time-varying adjacency matrix is ​​generated by an encoder-decoder architecture, and an adaptive update of the graph structure is achieved by combining a two-layer alternating optimization strategy, resulting in a time-varying graph sequence G(t)=(V,E(t),X(t)), where V is the parameter node set, E(t) is the edge set at time t, and X(t) is the node feature matrix. S3 Spatiotemporal Feature Deep Learning: It uses a 3-layer Chebyshev polynomial approximation graph convolution to extract spatial correlation features, captures temporal evolution information through gated recurrent units (GRU) and temporal attention mechanism, and obtains global degradation features after multi-scale feature fusion. S4 Remaining Useful Life Prediction: A hybrid density network (MDN) is used to model the probability distribution of remaining useful life (RUL) and output the expected value of RUL and the 90% confidence interval.

2. The method for predicting the remaining service life of a gear pump based on spatiotemporal graph modeling according to claim 1, characterized in that, The vibration signal preprocessing in S1 includes: sampling frequency of 10kHz, extracting 12 features such as root mean square, kurtosis, and meshing frequency amplitude, downsampling to 1Hz to align with performance parameters; missing values ​​are filled by local weighted regression interpolation, and performance parameters and vibration features are standardized using z-score.

3. The method for predicting the remaining service life of a gear pump based on spatiotemporal graph modeling according to claim 1, characterized in that, In S2, dynamic graph generation adopts an encoder-decoder architecture: the encoder is a 3-layer perceptron that maps node features to a 64-dimensional latent space to obtain node embedding h; after normalization, a symmetric non-negative dynamic adjacency matrix A(t) is obtained.

4. The method for predicting the remaining service life of a gear pump based on spatiotemporal graph modeling according to claim 1, characterized in that, In S3, spatial graph convolution is approximated using a K=3rd order Chebyshev polynomial, combined with residual connections: ,in, It is a non-linear activation function. For summation, These are the training parameters for the convolutional layer. Indicates performing an operation on the Laplace matrix. Polynomial transformation of order 1; For the normalized graph Laplace matrix, The node feature matrix is ​​constructed using three convolutional layers with hidden dimensions of 64→128→64 respectively.

5. The method for predicting the remaining service life of a gear pump based on spatiotemporal graph modeling according to claim 1, characterized in that, In S3, multi-scale fusion uses three parallel spatiotemporal convolutional branches, which process the input data with time steps of 2, 4, and 8 respectively, and extract features F1, F2, and F3 at different scales.

6. A gear pump remaining service life prediction system based on spatiotemporal graph modeling, characterized in that, include: Multi-source data acquisition module: composed of an accelerometer, a pressure sensor, a flow sensor, and an oil sampling device, supporting synchronous acquisition and data transmission of multiple parameters; Dynamic graph modeling module: runs the graph generation and optimization algorithm of step S2 as described in claim 1, and outputs the time-varying graph structure sequence G(t); Spatiotemporal feature learning module: equipped with a 3-layer Chebyshev graph convolution and GRU network, performing spatiotemporal feature extraction in step S3 of claim 1; Prediction output module: runs a hybrid density network, outputting the expected value of remaining lifetime (RUL), 90% confidence interval and alarm signal.

7. The gear pump remaining service life prediction system based on spatiotemporal graph modeling according to claim 6, characterized in that, Data storage and management module: It uses a MySQL database to store historical monitoring data and model parameters, and supports retrospective querying of degradation curves and graph structure evolution processes.

8. The gear pump remaining service life prediction system based on spatiotemporal graph modeling according to claim 7, characterized in that, The multi-source data acquisition module includes an accelerometer with a range of ±50g and an accuracy of ±0.5%, a pressure sensor with a range of 0-30MPa and an accuracy of 0.2%, supports PoE power supply and Ethernet transmission, and has a latency of <100ms; the predictive output module supports RS485 / Modbus protocol access to the SCADA system.

Images