Engine component fault prediction system based on fusion of multiple inertial sensors

Abstract

The invention discloses an engine component fault prediction system based on fusion of multiple inertial sensors, and particularly relates to the technical field of engine state monitoring and fault prediction, comprising the following steps: collecting multi-dimensional inertial motion data of a key component; receiving multi-dimensional inertial motion data, and dynamically adjusting a sensor array data acquisition strategy based on real-time working condition parameters of the engine; generating a fusion feature vector; outputting a state feature vector of the health state of the engine component; and deducing a future health state degradation track of the engine component through probability simulation. According to the invention, through a multi-inertial-sensor acquisition module, a self-adaptive synchronous acquisition module, a spatial-temporal feature fusion module, a component state agent model module and a fault evolution reasoning module, a multi-inertial-sensor fused engine component fault prediction system is constructed; the problems that the comprehensiveness and synchronism of data collection are insufficient, the effectiveness of feature fusion is insufficient, and the refinement degree of state evaluation is insufficient are solved.

Description

Technical Field

[0001] This invention relates to the field of engine condition monitoring and fault prediction technology, and more specifically, to an engine component fault prediction system based on multi-inertial sensor fusion. Background Technology

[0002] As a core power source in aviation, shipbuilding, and automotive industries, the operational status of key components (such as rotating parts, transmission parts, and reciprocating parts) directly determines the reliability and safety of the entire engine. With the development of industrial intelligence, engine component fault prediction has become one of the core technologies for ensuring continuous equipment operation and reducing maintenance costs. Related technical solutions mainly revolve around the core processes of data acquisition, feature processing, condition assessment, and fault prediction.

[0003] However, in practical use, it still has some shortcomings, such as insufficient comprehensiveness and synchronization of data acquisition: a single sensor and a small number of measurement points cannot cover the core area of ​​component movement and force transmission path nodes, making it difficult to capture multi-dimensional inertial motion data (acceleration, angular velocity, angular acceleration); the fixed sampling frequency cannot adapt to the characteristic frequency changes of the engine under different operating conditions, resulting in the loss of high-frequency fault signals or low-frequency data redundancy, affecting the accuracy of subsequent feature fusion and state assessment.

[0004] The scientific rigor and effectiveness of feature fusion are lacking: the physical field coupling effect and fault signal transmission path characteristics are not considered. Relying only on a single-dimensional basic feature makes it difficult to reveal the essence of the fault (such as bearing wear faults, which can simultaneously cause changes in vibration-stress coupling and changes in the transmission path attenuation coefficient), resulting in unreasonable feature vector dimensions and insufficient dimensions leading to the loss of key information.

[0005] The granularity of the condition assessment is insufficient: the training dataset covers a limited range of working conditions and states, the model has weak generalization ability and is difficult to adapt to the complex and ever-changing actual operating scenarios of the engine; the working condition parameters and physical features are not integrated into the input, which makes the model unable to distinguish the different health states corresponding to the same feature values ​​under different working conditions; the single output result cannot meet the granular assessment needs of "accurately locating the state, quantifying the risk level, and predicting the degradation trend" in industrial scenarios.

[0006] The lack of foresight and quantification in fault prediction: fixed threshold judgment is a post-event alarm and cannot predict potential faults in advance; traditional prediction methods do not combine real-time status data to dynamically update model parameters, and the prediction results deviate greatly from the actual degradation pattern of components; the lack of probabilistic extrapolation of future health status leads to either delayed maintenance decisions or failure to achieve the core goal of predictive maintenance.

[0007] Therefore, there is an urgent need for an engine component fault prediction system that can achieve accurate multi-dimensional data collection, scientific feature fusion, refined condition assessment, and quantitative fault prediction, in order to overcome the above-mentioned shortcomings of existing technologies. Summary of the Invention

[0008] To overcome the aforementioned deficiencies of the prior art, the present invention provides an engine component fault prediction system based on multi-inertial sensor fusion, which solves the problems mentioned in the background art through the following scheme.

[0009] To achieve the above objectives, the present invention provides the following technical solution: an engine component fault prediction system based on multi-inertial sensor fusion, comprising:

[0010] Multi-inertial sensor acquisition module: A multi-inertial sensor array is arranged on the measurement points of key components of the engine to collect multi-dimensional inertial motion data of the key components;

[0011] Adaptive synchronous acquisition module: connected to the multi-inertial sensor array module, used to receive multi-dimensional inertial motion data and dynamically adjust the sensor array data acquisition strategy based on the engine's real-time operating parameters;

[0012] Spatiotemporal feature fusion module: Connected to the adaptive synchronous acquisition module, it preprocesses multi-dimensional inertial motion data to generate a fused feature vector containing physical field coupling feature vector and transmission path feature vector;

[0013] Component State Agent Model Module: Connects the spatiotemporal feature fusion module and the adaptive synchronous acquisition module, receives the fused feature vector and the engine's operating parameters, and outputs the state feature vector of the engine component health status. The state feature vector includes health index, potential failure probability and deterioration trend parameters.

[0014] Fault evolution reasoning module: Connected to the component state proxy model, it receives the state feature vector and predicts the probability of failure by probabilistic simulation to infer the future health state degradation trajectory of the engine component.

[0015] The technical effects and advantages of this invention are as follows:

[0016] 1. This invention employs a multi-inertial sensor array arranged at multiple measurement points on key engine components, covering the core motion area of ​​the components and force transmission path nodes, realizing comprehensive acquisition of multi-dimensional inertial motion data (such as acceleration, angular velocity, and angular acceleration), improving the comprehensiveness and synchronization of data acquisition, and significantly enhancing the ability to capture fault characteristics;

[0017] 2. In the feature extraction stage, this invention innovatively introduces physical field coupling features (such as vibration-stress correlation coefficient, temperature-vibration coupling stiffness parameter) and transmission path features (such as transfer function, attenuation coefficient, path stiffness parameter), which deepens the feature characterization capability from the perspective of multi-physical field coupling and signal transmission mechanism, and improves the accuracy of fault identification.

[0018] 3. This invention constructs a Gaussian process regression (GPR) model and inputs the fused feature vector and real-time operating parameters to achieve a refined assessment of the health status of engine components under different operating conditions, enabling refined and multi-dimensional status assessment and supporting precise maintenance decisions;

[0019] 4. This invention, through a fault evolution reasoning module, combined with online Bayesian updates and Monte Carlo simulation, enables probabilistic extrapolation of the future health degradation trajectory of components, achieving forward-looking and probabilistic fault prediction and promoting the implementation of predictive maintenance. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall system structure of the present invention.

[0021] Figure 2 This is a schematic diagram of the multi-inertial sensor acquisition module of the present invention.

[0022] Figure 3 This is a schematic diagram of the adaptive synchronous acquisition module of the present invention.

[0023] Figure 4 This is a schematic diagram of the spatiotemporal feature fusion module of the present invention.

[0024] Figure 5 This is a schematic diagram of the component state proxy model structure of the present invention. Detailed Implementation

[0025] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Please see Figures 1-5 As shown, this embodiment of the invention provides an engine component fault prediction system based on multi-inertial sensor fusion, including:

[0027] Multi-inertial sensor acquisition module: A multi-inertial sensor array is arranged on the measurement points of key components of the engine to collect multi-dimensional inertial motion data of the key components;

[0028] Adaptive synchronous acquisition module: connected to the multi-inertial sensor array module, used to receive multi-dimensional inertial motion data and dynamically adjust the sensor array data acquisition strategy based on the engine's real-time operating parameters;

[0029] Spatiotemporal feature fusion module: Connected to the adaptive synchronous acquisition module, it preprocesses multi-dimensional inertial motion data to generate a fused feature vector containing physical field coupling feature vector and transmission path feature vector;

[0030] Component State Agent Model Module: Connects the spatiotemporal feature fusion module and the adaptive synchronous acquisition module, receives the fused feature vector and the engine's operating parameters, and outputs the state feature vector of the engine component health status. The state feature vector includes health index, potential failure probability and deterioration trend parameters.

[0031] Fault evolution reasoning module: Connected to the component state proxy model, it receives the state feature vector and predicts the probability of failure by probabilistic simulation to infer the future health state degradation trajectory of the engine component.

[0032] The multi-inertial sensor acquisition module, consisting of a multi-inertial sensor array, is arranged on the measurement points of key components of the engine to collect multi-dimensional inertial motion data of the key components.

[0033] A101: Screening of Key Engine Components and Measurement Points

[0034] Based on engine structure analysis and fault statistics, key components with high failure rates and impacting the overall engine operational safety are identified, including but not limited to:

[0035] Rotating components: crankshaft, connecting rod, turbine blades, compressor impeller; Transmission components: bearings (spindle bearings, connecting rod bearings), gear sets; Reciprocating components: pistons, valve mechanisms; Fixed key components: cylinder block, engine base;

[0036] Measurement point location planning: Arrange 3 to 6 measurement points for each key component, covering the core area of ​​engine motion and force transmission path nodes:

[0037] Crankshaft: front flange face, middle main journal corresponding to the cylinder block position, rear output end;

[0038] Turbine blades: Two measuring points are evenly distributed circumferentially at the blade root (hub) and blade tip;

[0039] Formula for calculating the distance between measuring points: ,in For the distance between measuring points, For the length of the critical motion range of the component, The number of measuring points should be determined to ensure that the measuring points cover the area of ​​component movement and deformation.

[0040] A102: Selection and Installation of Multiple Inertial Sensors

[0041] A1021: Uses a MEMS triaxial inertial sensor with integrated accelerometer and gyroscope to support the acquisition of acceleration, angular velocity and angular acceleration in multi-dimensional inertial motion data;

[0042] The range parameters of the MEMS triaxial inertial sensor are calculated based on engine operating conditions; acceleration range. ,in , Maximum engine speed , The recommended measuring range is the distance from the measuring point to the center of rotation. ;

[0043] Gyroscope range , For the highest angular velocity of the component, the recommended range is [range value]. ;

[0044] Angular acceleration is calculated by differentiating the angular velocity data output from the gyroscope. The specific difference formula is as follows: ,in For the first angular velocity at time, The sampling interval;

[0045] A1022: Multiple inertial sensor installation; select the installation method according to the component's operating conditions.

[0046] Static or low-vibration components (cylinder body, base): use high-temperature epoxy adhesive for bonding and fixing, with the adhesive layer thickness controlled between 0.1 and 0.3 mm. After evenly applying the adhesive, press the sensor and allow it to cure for more than 24 hours. The curing environment temperature should be controlled at 20℃±5℃.

[0047] Dynamic or high-vibration components (crankshaft, connecting rod, turbine blades, bearings, pistons, valve mechanism): are fixed with M3 threads, and the sensor base has a pre-drilled threaded hole for locking with an anti-loosening nut;

[0048] Installation attitude calibration involves adjusting the sensor attitude using a high-precision level to ensure the accelerometer... The shaft and components move in the same direction. The crankshaft measuring point X-axis is parallel to the crankshaft axis, Y-axis is perpendicular to the axis, and Z-axis is perpendicular to the mounting surface. The attitude deviation is controlled within ±1 degree.

[0049] Cable arrangement: Cables are fixed along the surface of components with high-temperature resistant clips at intervals of ≤10cm to avoid interference with moving parts. Cable joints are sealed with heat shrink tubing to prevent oil mist intrusion.

[0050] A103: Array Synchronization Calibration

[0051] Array synchronization calibration is performed based on the PTP protocol. Clock synchronization configuration: The adaptive synchronization acquisition module is used as the PTP master node, and the multi-inertial sensor array is used as the slave node. PTP protocol parameters are configured with a synchronization period of 10ms.

[0052] Synchronization error detection: Trigger the PTP master node to send a synchronization signal, collect the timestamps of the slave nodes, and calculate the synchronization error. ,in master node timestamp, For the node timestamp, the requirements are: ;

[0053] Data synchronization verification: The sinusoidal vibration signal output by the standard vibration table triggers the synchronous acquisition of multiple inertial sensor arrays. The timestamps and amplitudes of the data from each sensor are compared. The timestamp deviation is controlled within one acquisition node, and the amplitude deviation is controlled within ±2%, ensuring the consistency of the data from multiple inertial sensors.

[0054] The adaptive synchronous acquisition module is connected to the multi-inertial sensor array module and is used to receive multi-dimensional inertial motion data and dynamically adjust the sensor array data acquisition strategy based on the engine's real-time operating parameters.

[0055] B101: Selection of main body for adaptive acquisition module:

[0056] A multi-channel data acquisition card supporting the PTP clock synchronization protocol is selected. Specifically, the number of channels is greater than the total number of channels in the sensor array. Based on the 3 acceleration and 3 angular velocity channels of each MEMS triaxial sensor, 6 initialization channels are designed. If there are 32 sensors, then 192 channels are required.

[0057] The sampling frequency range is controlled between 100Hz and 10kHz and supports dynamic adjustment; it supports SPI / I2C protocol matching with sensor array signal interface and collects engine operating parameters through CAN bus interface;

[0058] PTP synchronization module configuration: Selects standard PTP synchronization accuracy. It supports master-slave mode switching, and this module acts as the master node of PTP;

[0059] Operating condition parameter acquisition interface selection: Connect to the engine ECU electronic control unit via CAN bus interface to acquire speed, load, and exhaust temperature data. The communication baud rate is set to 250kbps, which conforms to the industrial engine ECU communication standard.

[0060] B102: Physical connection between modules, connecting multiple inertial sensor array modules.

[0061] B1021: Connects to a multi-inertial sensor array module

[0062] For signal connection, the SPI / I2C channel of the B101 acquisition module is matched with the signal output terminal of the sensor array, and each sensor is assigned a unique channel address;

[0063] Power connection, outputting 3.3V / 5V low ripple power, connected in parallel with the power supply circuit of the sensor array to ensure power supply consistency;

[0064] Grounding is handled using a single-point grounding method, with the grounding terminal sharing a common ground with the sensor array grounding terminal and the engine block. The grounding resistance... .

[0065] B1022: Connection for acquiring operating parameters of the engine ECU

[0066] Connect the CAN interface of the B101 acquisition module to the CAN output interface of the engine ECU via the CAN bus data cable. The shield of the data cable needs to be grounded. Configure the CAN bus parameters, baud rate 250kbps, data frame format according to the ECU protocol, and parse the data frame ID and byte bits corresponding to speed, load, and exhaust temperature.

[0067] B1023: PTP Synchronization Link Deployment

[0068] Connect the B101 acquisition module's PTP master node and the sensor array's PTP slave nodes via Ethernet interface to form a star-shaped synchronous network; check the network link: measure the link delay with a network latency tester to ensure that the one-way delay is controlled within [specific parameters]. To ensure the accuracy of PTP synchronous acquisition;

[0069] PTP clock synchronization protocol configuration, master node configuration: Start the B101 acquisition module, enable PTP master node mode through configuration, and set the clock source to the local high-precision crystal oscillator;

[0070] Slave node synchronization adaptation: Through the configuration interface of the sensor array, set all sensors to PTP mode as slave nodes, specify the IP address of the B101 acquisition module, and synchronize with the master node.

[0071] B103: Implementation and Debugging of Dynamic Adjustment Strategies

[0072] Operating Parameter Acquisition and Analysis: Start the engine to acquire operating parameters. Receive operating data frames sent by the ECU via the CAN bus and analyze them to obtain the real-time engine speed. ,load Exhaust temperature ;

[0073] Analysis and verification: The raw data from the ECU was read using a CAN bus analyzer and compared with the acquired operating parameters to determine the speed error. Load error Exhaust temperature error If a parsing error occurs, check the data frame ID and byte bit mapping relationship, and reconfigure the parsing parameters.

[0074] B104: Execution and Optimization of Dynamic Data Acquisition Strategies

[0075] Dynamic adjustment of sampling frequency: Real-time calculation of characteristic frequencies under current operating conditions. ,in Indicates real-time rotation speed, Indicate the specific operating point, and according to Automatically adjust sampling frequency;

[0076] Example: Engine speed 5000r / min (high speed, full load). hour, sampling frequency Rounded to 1.7kHz; Boundary limits: Sampling frequency not lower than 500Hz and not higher than 10kHz, to avoid exceeding hardware capabilities.

[0077] The transmission bandwidth and caching strategy are dynamically adapted. Based on the real-time sampling frequency and number of channels, the transmission bandwidth is automatically allocated to ensure that the bandwidth utilization rate is ≥90% and there is no packet loss.

[0078] Strategy switching: When the exhaust temperature is ≥650℃ and the load is ≥100N・m, it automatically switches to the cache priority mode. When the cached data reaches 80%, it triggers batch transmission to the subsequent spatiotemporal feature fusion module.

[0079] B105: Data Interaction with the Spatiotemporal Feature Fusion Module

[0080] The adaptive synchronous acquisition module encapsulates the synchronized multi-dimensional inertial motion data in the format of sensor ID, timestamp, acceleration, angular velocity, and angular acceleration, and transmits it to the spatiotemporal feature fusion module.

[0081] Use a data analyzer to check the integrity of the transmitted data, ensuring there is no missing data or format errors, and that the transmission delay is controlled within 50ms.

[0082] The spatiotemporal feature fusion module is connected to the adaptive synchronous acquisition module, which preprocesses the multi-dimensional inertial motion data to generate a fused feature vector containing physical field coupling feature vector and transmission path feature vector.

[0083] C101: Input Data Format and Protocol Confirmation

[0084] Define the input data type: Receive two types of data transmitted from the adaptive synchronous acquisition module and parse them according to a fixed format:

[0085] Multi-dimensional inertial motion data: [sensor ID, timestamp, acceleration X / Y / Z, angular velocity X / Y / Z, angular acceleration X / Y / Z], where the timestamp is based on the PTP synchronization protocol and has high accuracy. ;

[0086] Engine operating parameters: [timestamp, real-time speed] ,load Exhaust temperature Aligned with the inertial data timestamp, with the alignment timestamp deviation controlled within one acquisition node;

[0087] Data transmission protocol adaptation: TCP / IP protocol is used to receive data, and the data frame header identifier is set to 0x5A and the frame tail identifier is set to 0xA5 to avoid data frame loss or misalignment; the transmission baud rate is dynamically matched with the transmission bandwidth of the adaptive synchronous acquisition module.

[0088] C102: Preprocessing of multi-dimensional inertial motion data

[0089] Data noise reduction and noise type identification: Frequency domain analysis is performed using FFT transformation to identify electromagnetic interference and vibration background noise in the engine compartment;

[0090] Layered noise reduction algorithm, using A low-pass filter with a cutoff frequency of 5kHz and an order of 4 has the following filtering formula:

[0091]

[0092] in , Indicates cutoff frequency, Indicates the filter order; wavelet thresholding is used for noise reduction, with db4 wavelet selected, decomposition level 3, and the threshold set to the default value. , For noise standard deviation, For data length;

[0093] Spatiotemporal alignment processing: Time alignment: Using the PTP master node timestamp of the adaptive synchronous acquisition module as a reference, the inertial data and operating parameters of each sensor are matched with timestamps, and data with a timestamp deviation of 1 acquisition node are deleted; Interpolation completion: Missing data is completed using linear interpolation, with the specific formula as follows:

[0094]

[0095] in Indicates the valid timestamps before and after the missing data. This represents valid data corresponding to the timestamp.

[0096] Spatial alignment: Based on the coordinates of the measurement points of the multi-inertial sensor array module, the inertial data of each sensor is mapped to a unified spatial coordinate system;

[0097] Spatial interpolation is used to complete spatial data for the motion core area of ​​key components, such as the area at the tip of turbine blades where no measuring points are arranged, by using the Kriging interpolation method to ensure the integrity of spatial features.

[0098] The spatiotemporally aligned multidimensional inertial data (acceleration, angular velocity, angular acceleration) is subjected to Z-score normalization, using the following formula: ,in For the data mean, The standard deviation of the data is 0, and the standard deviation of the data is 1. This avoids the impact of differences in the amount of data in different dimensions on the feature fusion effect.

[0099] C103: Core Feature Extraction Stage, Physical Field Coupling and Transmission Path Features

[0100] C1031: Calculation of physical field coupling feature vectors, based on the coupling mechanism of engine vibration field, stress field, and temperature field, extracting two types of core features:

[0101] Engine vibration-stress correlation coefficient calculation, stress data derivation, and indirect stress calculation using acceleration data: ,in This indicates the elastic modulus of engine component materials, such as steel E=206GPa. Indicates strain, Represents the standardized acceleration, This represents the actual acceleration; the Pearson correlation coefficient is used to measure the linear correlation between vibration and stress, and the calculation formula is:

[0102]

[0103] in For the covariance of acceleration and stress, , representing the variances of acceleration and stress, respectively, with values ​​ranging from [-1, 1]. The closer to ±1, the stronger the coupling.

[0104] Engine temperature-vibration coupled stiffness parameter calculation, temperature data correlation: engine exhaust temperature Mapped to component surface temperature Mapping based on the heat transfer formula: ,in To represent heat transfer efficiency, take , Ambient temperature;

[0105] Coupling stiffness calculation: Modal stiffness of vibration signals at different temperatures is extracted through frequency domain analysis. Formula: ,in Equivalent mass of component (kg) The vibration angular frequency (rad / s) For vibration amplitude, This represents the acceleration amplitude.

[0106] Eigenvector construction, physical field coupling eigenvectors are The dimensions are expanded according to the component measurement points; for example, 32 measurement points would result in a 64-dimensional vector.

[0107] C1032: Calculation of transmission path feature vectors, based on the fault signal transmission mechanism, extracting 3 types of core features:

[0108] Transfer function calculation: The ratio of the frequency domain amplitude of the vibration signal at a certain measuring point (signal receiver) to that at the fault source measuring point (signal transmitter), formula: ,in The frequency domain amplitude of the vibration signal at the fault source measurement point, The amplitude of the vibration signal at the receiving end measurement point in the frequency domain is taken as the core feature, and the amplitude of the transfer function at the characteristic frequency point is taken as the core feature.

[0109] Attenuation coefficient calculation: The rate attenuation of the fault signal's amplitude along the transmission path, specifically calculated using the following formula: , where d is the distance between the two measuring points, in dB / m;

[0110] Path stiffness parameter calculation:

[0111] Derivation based on the relationship between stress and displacement ,in For the cross-sectional area of ​​the component, The relative displacement between the two measuring points, Obtained by integral of acceleration ;

[0112] Feature vector construction, the path feature vector is The vector can be expanded according to the dimension of the measurement points; for example, 32 measurement points would result in a 96-dimensional vector.

[0113] C104: Feature fusion and optimization stage, generating highly recognizable fused feature vectors.

[0114] Feature vector concatenation involves concatenating the physical field coupling feature vector and the transmission path feature vector according to the measurement point order to form an initial fused feature vector. Example: The initial fused feature vector for 32 measurement points has a dimension of 32×2 (physical field coupling) + 32×3 (transmission path) = 160 dimensions.

[0115] The feature dimensions are optimized and feature redundancy is eliminated. The mutual information method is used to calculate the mutual information value between each feature dimension. The redundancy elimination rule is that if the mutual information value between two feature dimensions is ≥0.8 (indicating extremely high redundancy), the dimension with higher information gain is retained. Finally, the feature vector is fused and output. After eliminating redundancy, the feature dimensions are controlled between 50 and 100.

[0116] The component state proxy model module is connected to the spatiotemporal feature fusion module and the adaptive synchronous acquisition module. It receives the fused feature vector and the engine's operating parameters, and outputs the state feature vector of the engine component health status. The state feature vector includes health index, potential failure probability and deterioration trend parameters.

[0117] D101: Receiving fused feature vectors and engine operating parameters

[0118] Receive fused feature vectors from the spatiotemporal feature fusion module. The format is [measurement point ID, feature dimension identifier, physical field coupling feature value, transmission path feature value], with a dimension of 50 to 100 and a data precision of float32.

[0119] Receive real-time engine operating parameters from the adaptive synchronous acquisition module, in the format of [timestamp, real-time speed]. ,load Exhaust temperature The timestamp of the fused feature vector is aligned with the timestamp, with an alignment error of ≤1 sampling point.

[0120] D102: Construction of Component State Proxy Model

[0121] D1021: Training dataset construction and preprocessing, with collection scenarios including: engine idling / medium speed load / high speed full load conditions, covering four states: component health, minor degradation, moderate fault, and severe fault;

[0122] Collect ≥5000 samples for each condition, with a total sample size of ≥20000 samples. Sample format: [fused feature vector, operating condition parameters, labeled values ​​(health index / failure probability / deterioration trend)];

[0123] Further explanation is needed regarding the labeled values, which are based on the engine's historical operating conditions. When engine components are in healthy condition: the health index is... Potential faults are The degradation trend parameter is ;

[0124] When engine components show slight deterioration: the health index is Potential faults are The degradation trend parameter is ;

[0125] When engine components are in a moderate fault condition: the health index is Potential faults are The degradation trend parameter is ;

[0126] When engine components are in serious condition: the health index is Potential faults are The degradation trend parameter is ;

[0127] Dataset preprocessing: The dataset is divided into training, validation, and test sets in a 7:2:1 ratio. Normalization: The fused feature vector and working condition parameters are subjected to Min-Max normalization, mapped to [0,1]. The criterion is to remove outlier samples that deviate from the mean by three times the standard deviation.

[0128] D1022: Design of Gaussian Process Regression (GPR) Model (Simplified Kernel Function + Dimensionality Reduction):

[0129] Input layer: fuse feature vectors (50 to 100 dimensions) and working condition parameters (3 dimensions), and concatenate them into a 53 to 103-dimensional input;

[0130] Kernel function selection balances accuracy and computational efficiency: Radial basis function (RBF) and white noise kernel are selected. White noise kernel function formula:

[0131] in For signal variance, For length scale, For noise variance, It is the Dirac function, which is 1 only when i=j;

[0132] The output layer uses three regression paths to output health index, potential failure probability, and degradation trend parameters, avoiding model complexity caused by multi-task coupling.

[0133] Model simplification strategy to adapt to real-time industrial scenarios: Feature dimensionality reduction, using principal component analysis (PCA) algorithm on the input layer, retaining principal components with a cumulative variance contribution rate of ≥95%, reducing the input dimension from 53 to 103 dimensions to 20 to 30 dimensions.

[0134] D103: State feature vector of engine component health status

[0135] For the fused feature vector of the synchronous timestamp input and the engine's real-time operating parameters, a Gaussian process regression inference process is performed:

[0136] D1031: Health Index Calculation

[0137] Will Assuming it is the regression target, the GPR model provides the predicted distribution: ,in This represents the fused feature vector of the synchronous timestamp input and the engine's real-time operating parameters;

[0138] The health index output is the predicted mean. As a point estimate of the current health index, ensure that it falls within [the range of the index]. Interval, using Functions for mapping and scaling:

[0139]

[0140] in The predicted mean of the health baseline This represents the scaling factor, calibrated based on historical degradation data.

[0141] D1032: Calculation of Potential Failure Probability

[0142] The failure probability is not directly regressed, but rather calculated based on a comprehensive assessment of prediction uncertainty and the degree of deviation from the baseline, defining an anomaly score. The larger the value, the more abnormal the state and the higher the confidence level; calculate the failure probability: map the abnormal score to the [0,1] interval using the cumulative distribution function; for example, assume an abnormal score in a healthy state. If it follows a chi-square distribution, then:

[0143]

[0144] in It is the cumulative distribution function of the chi-square distribution. The degree of freedom and probability are used to quantify the likelihood that the current observation does not belong to a healthy distribution.

[0145] D1033: Degradation Trend Parameter Calculation

[0146] Trend calculation, within a 24-hour sliding time window, affects the health index. The time series is linearly fitted or exponentially smoothed to obtain... Slope as a function of time Trend parameter output: slope The percentage degradation per 100 hours is used as a trend parameter: ; A negative value indicates a decrease in health, while a positive value indicates recovery.

[0147] The fault evolution reasoning module is connected to the component state proxy model. It is used to receive the state feature vector and predict the probability of failure by probabilistically simulating the future health state degradation trajectory of the engine component.

[0148] This embodiment is the core of the system's predictive maintenance decision-making. By probabilistically extrapolating the dynamic evolution of the engine component health status, it achieves a quantitative assessment of future risks. Specific steps include:

[0149] E101: Online Bayesian Update

[0150] Real-time data reception, receiving the real-time state feature vector stream from the component state proxy model module. ,in ;

[0151] The likelihood function is constructed by taking the observed... By combining real-time changes with the degradation process prediction of the Gaussian process regression (GPR) model, a likelihood function is constructed. ;

[0152] Posterior distribution sampling, Bayesian update: The Markov chain Monte Carlo (MCMC) method is used, leveraging newly observed... The posterior likelihood function of the model parameters Real-time updates and sampling enable the model parameters to learn adaptively and fit the actual degradation data of the currently monitored engine components.

[0153] E102: Future Trajectory Projection Based on Monte Carlo Simulation

[0154] Multi-track parallel simulation, from the updated parameter likelihood function In the process, N sets of parameter samples are randomly selected;

[0155] For each set of parameter samples, combined with future operating condition inputs, starting from the current time t, a forward numerical integration is performed on the degradation process model to simulate a path from t to the future prediction. A deterministic degradation trajectory;

[0156] Mapping from trajectory to health index: The simulated degradation trajectory is transformed back into the future trajectory of the health index HI through the inverse mapping relationship defined in the surrogate model module. .

[0157] E103: Extraction and Output of Key Forecasting Indicators

[0158] Calculate the distribution statistics of the health index at each future time point on all simulated trajectories; output the predicted median trajectory and the 90% prediction interval trajectory to show the possible range of future health status.

[0159] FTT distribution and fault occurrence probability: Based on the D1021 annotation value, a fault threshold is determined. For each simulated trajectory, the first time its HI value falls below a certain threshold is calculated. The probability distribution of the simulated trajectory over time;

[0160] Generate maintenance decision recommendations: trigger different levels of alerts such as continued monitoring, planned maintenance, or immediate inspection.

[0161] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0162] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An engine component fault prediction system based on multi-inertial sensor fusion, characterized in that, include: Multi-inertial sensor acquisition module: A multi-inertial sensor array is arranged on the measurement points of key components of the engine to collect multi-dimensional inertial motion data of the key components; Adaptive synchronous acquisition module: connected to the multi-inertial sensor array module, used to receive multi-dimensional inertial motion data and dynamically adjust the sensor array data acquisition strategy based on the engine's real-time operating parameters; Spatiotemporal feature fusion module: Connected to the adaptive synchronous acquisition module, it preprocesses multi-dimensional inertial motion data to generate a fused feature vector containing physical field coupling feature vector and transmission path feature vector; Component State Agent Model Module: Connects the spatiotemporal feature fusion module and the adaptive synchronous acquisition module, receives the fused feature vector and the engine's operating parameters, and outputs the state feature vector of the engine component health status. The state feature vector includes health index, potential failure probability and deterioration trend parameters. Fault evolution reasoning module: Connected to the component state proxy model, it receives the state feature vector and predicts the probability of failure by probabilistic simulation to infer the future health state degradation trajectory of the engine component.

2. The engine component fault prediction system based on multi-inertial sensor fusion according to claim 1, characterized in that, The multi-inertial sensor acquisition module includes: The measurement point arrangement of the multi-inertial sensor array is based on the key components selected from engine structural analysis and fault statistics data. The key components include rotating components, transmission components, reciprocating components, and fixed key components. Each key component has 3 to 6 measurement points, covering the core area of ​​component movement and force transmission path nodes.

3. The engine component fault prediction system based on multi-inertial sensor fusion according to claim 2, characterized in that, The multi-inertial sensor acquisition module includes sensor selection and installation, which includes: MEMS triaxial inertial sensors integrating accelerometers and gyroscopes are selected. Static or low-vibration components are fixed with high-temperature epoxy adhesive, with the adhesive layer thickness controlled between 0.1 and 0.3 mm, the curing time greater than 24 hours, and the curing environment temperature controlled at 20℃±5℃. Dynamic or high-vibration components are fixed with M3 threads and locked with anti-loosening nuts; during installation, the sensor attitude is adjusted by a high-precision level to ensure that the accelerometer axis is consistent with the direction of component movement, and the attitude deviation is controlled within ±1 degree.

4. The engine component fault prediction system based on multi-inertial sensor fusion according to claim 1, characterized in that, The adaptive synchronous acquisition module includes: Calibration is performed based on the PTP protocol, with the adaptive synchronous acquisition module as the PTP master node and the multi-inertial sensor array as the slave node, and the synchronization period is 10ms. A multi-channel data acquisition card supporting the PTP clock synchronization protocol is selected, with the number of channels greater than the total number of channels in the sensor array. The sampling frequency range is controlled between 100Hz and 10kHz. It supports SPI / I2C protocol matching with the sensor array signal interface and acquires engine operating parameters through the CAN bus interface. The PTP synchronization accuracy is [value missing], and master-slave mode switching is supported. The CAN bus communication baud rate is set to 250kbps.

5. The engine component fault prediction system based on multi-inertial sensor fusion according to claim 1, characterized in that, The spatiotemporal feature fusion module includes: Preprocessing includes data denoising, using a low-pass filter with a cutoff frequency of 5kHz and an order of 4, and wavelet threshold denoising with db4 wavelet and a decomposition level of 3; spatiotemporal alignment is based on the PTP master node timestamp, missing data is filled in using linear interpolation, and spatial data is filled in using kriging interpolation; the data is Z-score normalized, and a fused feature vector is generated after preprocessing.

6. The engine component fault prediction system based on multi-inertial sensor fusion according to claim 5, characterized in that, The spatiotemporal feature fusion module fuses feature vectors including: The physical field coupling feature vector includes: vibration-stress correlation coefficient and temperature-vibration coupling stiffness parameter; the transfer path feature vector includes: transfer function, attenuation coefficient and path stiffness parameter, where the transfer function is the frequency domain amplitude ratio of the vibration signal at the fault source measurement point to that at the receiving end measurement point; The physical field coupling feature vector and the transmission path feature vector are concatenated in the order of measurement points to form an initial fused feature vector. The mutual information method is used to remove redundant features. When the mutual information value of the two feature dimensions is ≥0.8, the information gain dimension is retained. Finally, the dimension of the fused feature vector is controlled between 50 and 100.

7. The engine component fault prediction system based on multi-inertial sensor fusion according to claim 1, characterized in that, The component state proxy model module includes: The training dataset for the component state proxy model was collected under engine idling, medium-speed load, and high-speed full-load conditions, covering four states: component health, minor degradation, moderate fault, and severe fault. 5,000 samples were collected for each state, for a total of 20,000 samples. The dataset was divided into training set, validation set, and test set in a 7:2:1 ratio. The Gaussian process regression (GPR) model is adopted. The input layer consists of the spliced ​​fused feature vector and operating parameters. Radial basis function (RBF) and white noise kernel are selected as kernel functions. Principal component analysis (PCA) algorithm is used for dimensionality reduction to retain principal components with a cumulative variance contribution rate of ≥95%. The output layer outputs health index, potential failure probability, and degradation trend parameters through three-way regression.

8. The engine component fault prediction system based on multi-inertial sensor fusion according to claim 1, characterized in that, The fault evolution reasoning module includes: The model parameters are updated and sampled in real time by constructing a likelihood function through online Bayesian updates and using the Markov chain Monte Carlo (MCMC) method. Monte Carlo simulation randomly selects N sets of parameter samples from the updated likelihood function, combines them with future operating condition inputs for forward numerical integration, simulates the degradation trajectory, and converts it into the future trajectory of the health index.

Images