A method and system for monitoring the operation of a fuel pump test bench

Abstract

The present application relates to the technical field of test bench operation monitoring, in particular to a fuel pump test bench operation monitoring method and system. The method comprises: obtaining monitoring time series data of the fuel pump test bench operation and performing feature extraction, constructing a coupling prediction model of a spatiotemporal graph attention network-unscented Kalman filter, predicting the process noise covariance matrix and the measurement noise covariance matrix according to the prediction process, performing state estimation using the unscented Kalman filter to obtain the vector posterior probability distribution, constructing a fault evolution trajectory manifold, calculating the Mahalanobis distance between the vector posterior probability distribution and the fault evolution trajectory manifold, and taking the Mahalanobis distance as a monitoring index. That is, the scheme of the present application can extract deep features from multi-source data that can better reflect the inherent nonlinearity and complex dynamics of the system, improve the accuracy of the filter in estimating the potential health status of the system, and overcome the defects of fixed model parameters and difficulty in capturing gradual faults in traditional methods.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of test bench operation monitoring. More particularly, the present application relates to a fuel pump test bench operation monitoring method and system. BACKGROUND

[0002] As a core power component in high-end equipment such as aviation, aerospace and automobile engines, the reliability and safety of the fuel pump directly affect the success or failure of the entire system. The fuel pump test bench is a key equipment specially used for performance testing, durability testing and fault diagnosis of the fuel pump. Accurate and real-time monitoring of its operating state is of great significance to ensure product quality and prevent catastrophic accidents. Traditional fuel pump test bench monitoring methods mainly rely on setting fixed threshold alarms for key parameters such as pressure, flow rate and speed, or performing simple time or frequency domain analysis on vibration signals, such as using Fourier transform, short-time Fourier transform, etc. However, such methods have significant limitations: first, they usually analyze single sensor signals in isolation, ignoring the complex nonlinear coupling relationship between different physical quantities; second, they are not sensitive to early weak fault characteristics, especially for the gradual performance degradation process of the test bench from normal to failure, which can easily lead to false negatives; finally, the fixed threshold or simple feature-based judgment logic is difficult to adapt to the dynamic changes of the test bench, and has poor generalization ability and robustness.

[0003] To improve the accuracy and intelligence level of monitoring, some advanced signal processing and machine learning methods have been introduced into the field of fault monitoring. For example, wavelet packet transform, empirical mode decomposition and other techniques are used to extract deep features of signals, and support vector machines, neural networks and other models are used for fault classification. In addition, state estimation-based monitoring methods, especially Kalman filter and its derivative algorithms (such as unscented Kalman filter), have attracted attention because they can effectively handle noise in nonlinear systems and estimate the potential state of the system. Although these methods have improved monitoring performance to some extent, they still face challenges: on the one hand, existing machine learning models have limited ability to mine the spatio-temporal correlation between multi-source heterogeneous monitoring data, making it difficult to fully utilize the complementary information between data; on the other hand, when applying state estimation algorithms such as unscented Kalman filter, the core parameters - process noise covariance matrix and measurement noise covariance matrix, usually need to be preset based on experience and remain constant during operation. This static setting cannot adapt to the dynamic changes in system noise characteristics when the state of the fuel pump test bench changes, severely restricting the accuracy and reliability of state estimation. Therefore, there is an urgent need for a new monitoring method that can deeply integrate multi-source data, adaptively optimize model parameters and finely quantify the evolution process of the health state. SUMMARY

[0004] The application aims to provide a fuel pump test bench operation monitoring method and system to solve the technical problems of fixed model parameters and difficulty in capturing gradual failure in the prior art.

[0005] In the first aspect, the application provides a fuel pump test bench operation monitoring method, comprising:

[0006] S1: obtaining monitoring time series data of the fuel pump test bench operation, the monitoring time series data comprising pressure, flow, vibration and rotation speed data; S2: performing feature extraction on the monitoring time series data; processing vibration data by using fractional wavelet packet transform to obtain time-frequency energy entropy features; calculating the correlation dimension and maximum Lyapunov exponent of pressure data and flow data as dynamic characteristic features based on phase space reconstruction theory; fusing the time-frequency energy entropy features, dynamic characteristic features and the rotation speed data into a time series feature vector sequence; S3: constructing and training a coupled prediction model of a spatiotemporal graph attention network-universal Kalman filter offline using historical monitoring data; inputting the time series feature vector sequence obtained in S2 into the coupled prediction model, wherein the spatiotemporal graph attention network is used to mine the spatiotemporal correlation between different monitoring data and adaptively predict the process noise covariance matrix and measurement noise covariance matrix of the universal Kalman filter; S4: performing state estimation on the time series feature vector sequence by using the universal Kalman filter configured with the predicted measurement noise covariance matrix to obtain a vector posterior probability distribution representing the potential health state of the fuel pump test bench; S5: constructing a fault evolution trajectory manifold in the health state space in advance according to historical data covering the whole process from normal to failure by using the method of steps S1 to S4; calculating the Mahalanobis distance between the vector posterior probability distribution obtained in step S4 and the fault evolution trajectory manifold, and taking the Mahalanobis distance as a monitoring index representing the current operation state.

[0007] Preferably, the time-frequency energy entropy features of the vibration data obtained by using the fractional wavelet packet transform comprise: performing multi-layer fractional wavelet packet transform on the vibration time series data to obtain coefficients of multiple frequency bands; reconstructing signals of each frequency band; calculating the energy of each frequency band signal, and calculating the time-frequency energy entropy based on the proportion of the energy of each frequency band signal in the total energy, and taking the entropy value as the time-frequency energy entropy feature of the vibration data.

[0008] Preferably, the calculation of the correlation dimension and maximum Lyapunov exponent of the pressure data and flow data as dynamic characteristic features based on the phase space reconstruction theory comprises: determining the delay time and embedding dimension m of the phase space reconstruction of the pressure time series data; calculating the correlation dimension and maximum Lyapunov exponent of the pressure data and flow data according to the delay time and embedding dimension m, reconstructing the one-dimensional pressure time series data into a trajectory in a high-dimensional phase space; based on the reconstructed phase space trajectory, calculating the correlation dimension and the maximum Lyapunov exponent of the pressure data; repeating the above steps for the flow data to obtain the correlation dimension and the maximum Lyapunov exponent of the flow data.

[0009] Preferably, when reconstructing the phase space, the mutual information method is used to calculate the optimal delay time, and the delay time is obtained. Preferably, the pseudo-nearest neighbor method is used to determine the minimum embedding dimension, and the delay bit number m is obtained.

[0010] Preferably, the spatiotemporal graph attention network is used to mine the spatiotemporal correlation between different monitoring data, including: at each time, taking each dimension of the time series feature vector as a node of a graph, and constructing a graph structure; calculating the mutual influence weight between nodes through a spatial attention module; inputting the node features weighted by the spatial attention into a recurrent neural network unit to capture the time evolution law of the features.

[0011] Preferably, the recurrent neural network unit is a gated recurrent unit.

[0012] Preferably, the process noise covariance matrix and the measurement noise covariance matrix of the adaptive prediction unscented Kalman filter include: after the output layer of the spatiotemporal graph attention network, two independent fully connected layers are arranged in parallel; the feature vector containing spatiotemporal correlation extracted by the spatiotemporal graph attention network is simultaneously input into the two fully connected layers; the output of one fully connected layer is reshaped into the process noise covariance matrix , the output of the other fully connected layer is reshaped into the measurement noise covariance matrix , and an activation function is applied to the two outputs to ensure that the generated covariance matrix is a positive definite matrix.

[0013] Preferably, the state estimation of the time series feature vector sequence by the unscented Kalman filter configured with the predicted measurement noise covariance matrix includes: generating a set of Sigma points and performing prediction according to the posterior state estimation at time k-1, to obtain a predicted state mean and an enhanced covariance matrix from the predicted process noise covariance matrix ; calculating the Kalman gain according to the predicted observation value, the actual observation value, and the enhanced observation covariance matrix from the predicted measurement noise covariance matrix ; combining the predicted state mean and the Kalman gain to update the state estimation, to obtain the posterior state estimation at time k.

[0014] Preferably, the fault evolution trajectory manifold in the health state space is constructed, comprising: dividing historical data covering the whole process from normal to failure in a sliding window manner; for the data in each window, applying steps S1 to S4 to calculate a vector posterior probability distribution representing the health state of the window; taking the mean vector of the posterior probability distribution of all data windows as a point in the high-dimensional health state space to form a fault evolution trajectory; and projecting the fault evolution trajectory into a low-dimensional space using a manifold learning algorithm to form a fault evolution trajectory manifold.

[0015] In a second aspect, a fuel pump test bench operation monitoring system comprises:

[0016] a processor; a memory storing computer instructions for fuel pump test bench operation monitoring, which, when executed by the processor, causes the system to perform the fuel pump test bench operation monitoring method described above.

[0017] The present application has the following beneficial effects: the present application provides a fuel pump test bench operation monitoring method, which has the advantage that by combining fractional wavelet packet transform and phase space reconstruction theory, deeper features that can better reflect the inherent nonlinearity and complex dynamics of the system can be extracted from multi-source data such as vibration, pressure, and flow; the use of a spatiotemporal graph attention network can effectively mine the spatiotemporal relationships between different monitoring data and use this correlation to determine the key parameters of an unscented Kalman filter, thereby improving the accuracy of the filter in estimating the potential health state of the system; by constructing a health state evolution trajectory manifold covering the whole process from normal to failure and calculating the Mahalanobis distance between the current state and the trajectory manifold, a monitoring index that can finely quantify the degree of performance degradation is formed, which is more sensitive to early weak faults and enhances the predictability and reliability of the monitoring, overcoming the defects of fixed model parameters and difficulty in capturing gradual faults in traditional methods. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 A step flowchart of the fuel pump test bench operation monitoring method in the present embodiment is schematically shown;

[0019] Figure 2 A structural block diagram of the fuel pump test bench operation monitoring system in the present embodiment is schematically shown. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.

[0021] As Figure 1 shown, a fuel pump test bench operation monitoring method in the present embodiment comprises the following steps:

[0022] Step S1, obtaining monitoring time series data of the fuel pump test bench running, the monitoring time series data including pressure, flow, vibration and rotating speed data.

[0023] Specifically, acceleration sensors, pressure sensors, flow meters and rotating speed sensors are respectively installed at key positions such as pump body shell, oil inlet and outlet of the fuel pump test bench, voltage signals of the sensors are synchronously collected by a NI data acquisition card at a sampling frequency of 20 kHz, and continuous signal streams are divided into time windows with a length of 2048 data points as unit data segments for subsequent processing.

[0024] Step S2, feature extraction is performed on the monitoring time series data: time-frequency energy entropy features are obtained by processing vibration data by using fractional wavelet packet transform; based on phase space reconstruction theory, correlation dimension and maximum Lyapunov exponent of pressure data and flow data are calculated as dynamic characteristic features; the time-frequency energy entropy features, the dynamic characteristic features and the rotating speed data are fused into a time series feature vector sequence.

[0025] The time-frequency energy entropy features of the vibration data are obtained by processing the vibration data by using fractional wavelet packet transform, including: performing multi-layer fractional wavelet packet transform on the vibration time series data to obtain coefficients of multiple frequency bands; reconstructing signals of the frequency bands; calculating energy of the signals of the frequency bands, and based on the proportion of energy of each frequency band in the total energy, calculating time-frequency energy entropy, and taking the entropy value as the time-frequency energy entropy feature of the vibration data.

[0026] Specifically, taking the bearing vibration signal of a certain water pump as an example, the vibration signal is sampled at a frequency of 10240 Hz. The fractional order is selected as 0.5, and three-layer fractional wavelet packet decomposition is performed. This decomposition divides the frequency band of the original signal into eight equal-width sub-bands, i.e. from frequency band zero to frequency band seven. Each sub-band contains a set of wavelet packet coefficients reflecting the characteristics of the signal in the frequency band.

[0027] The signals of each sub-band are reconstructed using these wavelet packet coefficients to obtain eight independent time-domain signals. Then the energy of each reconstructed signal is calculated, for example, the energy of the third frequency band signal is E3, the energy of the fifth frequency band signal is E5, and so on, and the total energy E of all eight signals is calculated. Through the formula p i equal to E i divided by E, the proportion of the energy of each frequency band in the total energy is calculated. Finally, using the Shannon entropy formula, i.e. the negative sum of p i times the logarithm of p i with base 2, a single entropy value is calculated, such as 2.15. The entropy value is the time-frequency energy entropy feature of the vibration data, which can quantify the complexity and uncertainty of the distribution of signal energy in different frequency bands, and it is particularly sensitive to energy transfer caused by early weak faults.

[0028] Based on the theory of phase space reconstruction, the correlation dimension and the largest Lyapunov exponent of the pressure data and the flow data are calculated as the dynamic characteristic features, including: for the pressure time series data, the delay time and the embedding dimension m of the phase space reconstruction are determined According to the delay time and the embedding dimension m, the one-dimensional pressure time series data is reconstructed into a trajectory in the high-dimensional phase space; based on the reconstructed phase space trajectory, the correlation dimension and the largest Lyapunov exponent of the pressure data are calculated; the above steps are repeated for the flow data to obtain the correlation dimension and the largest Lyapunov exponent of the flow data.

[0029] Specifically, a piece of device outlet pressure time series data containing 2000 data points is processed. In order to reconstruct its phase space, the mutual information method is used to calculate the optimal delay time, and the delay time is equal to 12. This means that every 12 data points are taken to construct a new dimension. Then, the pseudo-nearest neighbor method is used to determine the minimum embedding dimension, and the result is m equal to 5. This indicates that a five-dimensional phase space needs to be constructed to fully unfold the dynamic structure of the original one-dimensional signal.

[0030] Based on and m equal to 5, the original pressure data points p(t) are converted into a series of five-dimensional vectors, such as the first vector is p(1), p(13), p(25), p(37), p(49), the second vector is p(2), p(14), p(26), p(38), p(50), and so on. These vectors form a trajectory in the five-dimensional space. The G-P algorithm (Grassberger-Procaccia algorithm) is applied to this phase space trajectory, and the correlation dimension is calculated to be 2.78, which reflects the fractal dimension of the system attractor. At the same time, the wolf algorithm (Wolf algorithm) is applied to calculate the largest Lyapunov exponent, which is 0.09. This positive value indicates that the system has chaotic characteristics. The same set of processes is applied to the flow data of the device, so that the correlation dimension and the largest Lyapunov exponent of the flow data are also obtained, which are used as dynamic characteristic features together.

[0031] Step S3, a spatiotemporal graph attention network-unscented Kalman filter coupled prediction model is constructed and trained offline using historical monitoring data; the time series feature vector sequence obtained in S2 is input into the coupled prediction model, wherein the spatiotemporal graph attention network is used to mine the spatiotemporal correlation between different monitoring data and adaptively predict the process noise covariance matrix and the measurement noise covariance matrix of the unscented Kalman filter.

[0032] The spatio-temporal graph attention network is used to mine the spatio-temporal correlation between different monitoring data, including: at each time point, taking each dimension of the time series feature vector as a node of a graph, and constructing a graph structure; the mutual influence weight between nodes is calculated through a spatial attention module; the node features weighted by the spatial attention are input into a recurrent neural network unit to capture the time evolution law of the features.

[0033] Specifically, it is assumed that at time point k, a five-dimensional feature vector is obtained, including vibration entropy, pressure correlation dimension, pressure Lyapunov exponent, flow correlation dimension and flow Lyapunov exponent. The five features are regarded as five nodes of a graph, and then a fully connected graph is constructed, in which each node is connected to all other nodes, representing that there may be mutual influence between all features.

[0034] Next, the spatial attention module starts to calculate the dependence strength between nodes. For example, for the vibration entropy node, the spatial attention module calculates that the influence weight of the pressure correlation dimension on it is 0.5, and the influence weight of the flow correlation dimension on it is 0.2. This indicates that at the current time k, the dynamic change of the pressure has a greater influence on the vibration state than the flow. All nodes are subjected to such weighting calculation, generating a new feature vector which not only contains the original feature values, but also integrates the mutual influence information between the features. Finally, the feature vector weighted by the spatial attention is sent to a gated recurrent unit. The gated recurrent unit learns how the features and their relationships evolve over time by combining the hidden state at the previous time k-1, thereby effectively capturing the dynamic spatio-temporal evolution pattern of the device state.

[0035] The process noise covariance matrix and the measurement noise covariance matrix of the adaptive predictive unscented Kalman filter, including: after the output layer of the spatio-temporal graph attention network, two independent fully connected layers are arranged in parallel; the feature vector containing spatio-temporal correlation extracted by the spatio-temporal graph attention network is simultaneously input into the two fully connected layers; the output of one fully connected layer is reshaped into a process noise covariance matrix , and the output of the other fully connected layer is reshaped into a measurement noise covariance matrix , and an activation function is applied to the two outputs to ensure that the generated covariance matrix is positive definite.

[0036] Specifically, after the spatio-temporal graph attention network processes the input features at the current time, it outputs a hidden state vector containing rich spatio-temporal information, for example, a vector of length 128. This vector is simultaneously input into two different neural network branches. The first branch is a fully connected layer, which is tasked with predicting the process noise covariance matrix . If the state vector of the system is five-dimensional, this fully connected layer will output a vector of length 25.

[0037] Likewise, the second branch is also a fully connected layer, receiving the same 128-dimensional hidden state vector and outputting a vector of length 25 as well, which is used to predict the measurement noise covariance matrix . These two 25-dimensional output vectors are reshaped into matrices, respectively. To ensure that these two matrices are legitimate covariance matrices, i.e. they must be positive semi-definite, a special activation function is applied to them. For example, the output of the fully connected layer can be considered as the elements of a lower triangular matrix L, and then the product of L and its transpose L is computed, resulting in a positive semi-definite matrix, to which a small positive number is added to the diagonal, thus ensuring that the final generated and matrices are positive definite. In turn, the noise matrix can be dynamically adjusted according to the characteristics of the current data.

[0038] In step S4, a state estimation is performed on the sequence of time-series feature vectors using an unscented Kalman filter configured with the predicted measurement noise covariance matrix, to obtain a vector posterior distribution representing the potential health state of the fuel pump test bench.

[0039] The state estimation on the sequence of time-series feature vectors using the unscented Kalman filter configured with the predicted measurement noise covariance matrix comprises: generating a set of Sigma points and performing one-step prediction according to the posterior state estimation at k-1 time, to obtain a predicted state mean and an enhanced covariance matrix from the predicted process noise covariance matrix ; calculating a Kalman gain according to the predicted observation value, the actual observation value and an enhanced observation covariance matrix from the predicted measurement noise covariance matrix

[0040] ; and updating the state estimation by combining the predicted state mean and the Kalman gain, to obtain the posterior state estimation at k time. Specifically, at k-1 time, a posterior estimation mean vector and a covariance matrix have been obtained about the health state of the equipment. The unscented transformation first selects a set of Sigma points around this mean vector, for example, eleven Sigma points are selected for a five-dimensional state vector. These points are then propagated through the nonlinear state transition function representing the system dynamics to obtain a set of predicted Sigma points. The weighted average of these predicted points is the predicted state mean at k time, and their weighted covariance plus the process noise covariance predicted by the neural network constitutes the predicted state covariance

[0041] An update step is entered. The actual observed feature vector at time k is taken as the measurement. The predicted Sigma points are transformed to the measurement space by the measurement function to obtain the predicted measurement. The difference between the predicted measurement and the actual measurement constitutes the innovation. Using the predicted state covariance and the measurement noise covariance predicted by the neural network , the Kalman gain is calculated. The innovation is multiplied by the Kalman gain to correct the predicted state mean, thereby obtaining the final, more accurate posterior state estimation mean and covariance at time k.

[0042] Step S5, according to the historical data covering the whole process from normal to failure, the method of steps S1 to S4 is used to construct the fault evolution trajectory manifold in the health state space; the Mahalanobis distance between the vector posterior probability distribution obtained in step S4 and the fault evolution trajectory manifold is calculated, and the Mahalanobis distance is taken as a monitoring index representing the current operating state.

[0043] The fault evolution trajectory manifold in the health state space is constructed, including: the historical data covering the whole process from normal to failure is segmented in a sliding window manner; for the data in each window, steps S1 to S4 are applied to calculate a vector posterior probability distribution representing the health state of the window; the mean vector of the posterior probability distribution obtained by all data windows is taken as a point in the high-dimensional health state space, forming a fault evolution trajectory; a manifold learning algorithm is used to project the fault evolution trajectory into a low-dimensional space to form a fault evolution trajectory manifold.

[0044] Specifically, assume that there is a complete monitoring data of an equipment from brand-new operation to occurrence of a serious wear failure for 2000 hours. A sliding window with a width of 100 hours is set, and the window is slid by 10 hours each time. The first window contains data from 0 to 100 hours, the second window contains data from 10 to 110 hours, and so on, and a total of 191 data windows are generated.

[0045] For each 100-hour data window, the whole process of feature extraction, spatiotemporal graph attention network processing and adaptive unscented Kalman filtering is performed. The final output of the process is the posterior probability distribution of the device health state within the time window, and the mean vector is taken as a single data point representing the device health state in this window. After processing all 191 windows, 191 five-dimensional data points are obtained. These points are connected in chronological order in a five-dimensional space to form a fault evolution trajectory. Since the five-dimensional space cannot be directly displayed, the t-distributed neighborhood embedding algorithm is used to project the trajectory from the five-dimensional space to a two-dimensional plane. On the generated two-dimensional graph, a manifold composed of points can be clearly seen. The trajectory starts from a dense area representing the health state, gradually and smoothly migrates to another area representing the fault state over time, intuitively presenting the health degradation process of the entire device.

[0046] The present application also provides a fuel pump test bench operation monitoring system. As shown in Figure 2 The system includes a processor and a memory, and the memory stores computer program instructions that, when executed by the processor, implement the fuel pump test bench operation monitoring method according to the present application.

[0047] The system also includes a communication bus and a communication interface, as well as other components known to those skilled in the art, the settings and functions of which are known in the art, and therefore will not be described here.

[0048] In this description, the term "application" also means any computer program product storing such a program for use with or in connection with a computer system, apparatus or device. The program can be stored on any apparatus-readable medium, for example, but not limited to, any volatile memory or non-volatile memory. In this description, the term "memory" also means any computer program product storing such a program for use with or in connection with a computer system, apparatus or device. The program can be stored on any apparatus-readable medium, for example, but not limited to, any volatile memory or non-volatile memory. In this description, the term "computer-readable medium" means any tangible medium that stores, communicates, or otherwise provides data that can be used by an instruction execution system, apparatus or device. The computer-readable medium can be any suitable magnetic storage medium or magneto-optical storage medium, such as, for example, resistive random access memory (RRAM), dynamic random access memory (DRAM), static random access memory (SRAM), enhanced dynamic random access memory (EDRAM), high-bandwidth memory (HBM), hybrid memory cube (HMC), and the like, or any other medium that can be used to store the desired information and that can be accessed by an application, module, or both. Any such computer storage media can be part of the device or accessible or connectable thereto. Any application or module described in this description can be implemented by computer-readable / executable instructions stored or otherwise held by such computer-readable media.

[0049] In the description of the present description, the meaning of "a plurality of" is at least two, for example, two, three or more, and the like, unless otherwise explicitly specified.

[0050] Although the present description has shown and described a number of embodiments of the present application, it will be apparent to those skilled in the art that many modifications, variations, and alternatives to the embodiments described herein can be made in light of the teachings herein.

Claims

1. A method for monitoring the operation of a fuel pump test bench, characterized in that, Includes the following steps: S1: Acquire the monitoring time series data of the fuel pump test bench operation, including pressure, flow rate, vibration and speed data; S2: Feature extraction of the monitoring time series data: The vibration data is processed using fractional wavelet packet transform to obtain time-frequency energy entropy features; Based on phase space reconstruction theory, the correlation dimension and the maximum Lyapunov exponent of pressure data and flow data are calculated as dynamic characteristic features; the time-frequency energy entropy features, dynamic characteristic features and the speed data are fused into a time-series feature vector sequence. S3: Construct and use historical monitoring data to train a coupled prediction model of a spatiotemporal graph attention network and an unscented Kalman filter offline; input the temporal feature vector sequence obtained in S2 into the coupled prediction model. The spatiotemporal graph attention network is used to mine the spatiotemporal correlation between different monitoring data and adaptively predict the process noise covariance matrix and measurement noise covariance matrix of the unscented Kalman filter. S4: Using an unscented Kalman filter configured with a predicted measurement noise covariance matrix, state estimation is performed on the time-series feature vector sequence to obtain a vector posterior probability distribution characterizing the potential health state of the fuel pump test bench. S5: Based on historical data covering the entire process from normal to fault, construct the fault evolution trajectory manifold in the health state space using the methods in steps S1 to S4. The Mahalanobis distance between the vector posterior probability distribution obtained in step S4 and the fault evolution trajectory manifold is calculated, and the Mahalanobis distance is used as a monitoring indicator to characterize the current operating state.

2. The fuel pump test bench operation monitoring method according to claim 1, characterized in that, The process of obtaining time-frequency energy entropy features from vibration data using fractional wavelet packet transform includes: Multi-level fractional wavelet packet transform is performed on the vibration time series data to obtain coefficients for multiple frequency bands; Reconstruct the signals of each frequency band; The energy of each frequency band signal is calculated, and the time-frequency energy entropy is calculated based on the proportion of the energy of each frequency band signal to the total energy. The entropy value is then used as the time-frequency energy entropy characteristic of the vibration data.

3. The fuel pump test bench operation monitoring method according to claim 1, characterized in that, The method, based on phase space reconstruction theory, calculates the correlation dimension and maximum Lyapunov exponent of pressure and flow data as dynamic characteristic features, including: For pressure time series data, determine the delay time for phase space reconstruction. and embedding dimension m; According to the delay time With an embedding dimension m, one-dimensional pressure time series data is reconstructed into trajectories in a high-dimensional phase space; Based on the reconstructed phase space trajectory, the correlation dimension and the maximum Lyapunov exponent of the pressure data are calculated. Repeat the above steps for the traffic data to obtain the correlation dimension and the maximum Lyapunov exponent of the traffic data.

4. The fuel pump test bench operation monitoring method according to claim 3, characterized in that, When reconstructing the phase space, the mutual information method is used to calculate the optimal delay time, and the delay time is obtained. The minimum embedding dimension is determined by using the pseudo-nearest neighbor method, and the delay bits m are obtained.

5. The fuel pump test bench operation monitoring method according to claim 1, characterized in that, The spatiotemporal graph attention network is used to mine the spatiotemporal correlations between different monitoring data, including: At each time step, each dimension of the temporal feature vector is used as a node of the graph, and the graph structure is constructed. The spatial attention module is used to calculate the mutual influence weights between nodes. The node features, which have undergone spatial attention weighting, are input into a recurrent neural network unit to capture the temporal evolution of the features.

6. The fuel pump test bench operation monitoring method according to claim 5, characterized in that, The recurrent neural network unit is a gated recurrent unit.

7. The fuel pump test bench operation monitoring method according to claim 1, characterized in that, The process noise covariance matrix and measurement noise covariance matrix of the adaptive predictive unscented Kalman filter include: Two independent fully connected layers are set in parallel after the output layer of the spatiotemporal graph attention network; The feature vector containing spatiotemporal correlation extracted by the spatiotemporal graph attention network is simultaneously input into these two fully connected layers. Reshape the output of a fully connected layer into the process noise covariance matrix. The output of the other fully connected layer is reshaped into the measurement noise covariance matrix. An activation function is applied to both outputs to ensure that the generated covariance matrix is ​​a positive definite matrix.

8. The fuel pump test bench operation monitoring method according to claim 1, characterized in that, The process of estimating the state of the time-series feature vector sequence using an unscented Kalman filter configured with a predicted measurement noise covariance matrix includes: Based on the posterior state estimate at time k-1, a set of Sigma points are generated and predicted to obtain the predicted state mean and the predicted process noise covariance matrix. Enhanced covariance; Based on the predicted observations, the actual observations, and the predicted measurement noise covariance matrix Calculate the Kalman gain by enhancing the observation covariance; By combining the predicted state mean and Kalman gain, the state estimate is updated to obtain the posterior state estimate at time k.

9. The fuel pump test bench operation monitoring method according to claim 1, characterized in that, The construction of the fault evolution trajectory manifold in the healthy state space includes: Historical data covering the entire process from normal to fault is segmented using a sliding window method; For the data within each window, steps S1 to S4 are applied to calculate a vector posterior probability distribution representing the health status of the window. The mean vector of the posterior probability distribution obtained from all data windows is used as a point in the high-dimensional health state space to form the fault evolution trajectory. The fault evolution trajectory is projected into a low-dimensional space using a manifold learning algorithm to form a fault evolution trajectory manifold.

10. A fuel pump test bench operation monitoring system, characterized in that, include: processor; A memory storing computer instructions for monitoring the operation of a fuel pump test bench, which, when executed by the processor, cause the system to perform the fuel pump test bench operation monitoring method according to any one of claims 1-9.

Images