Ship diesel engine fault monitoring method based on LSTM-Iforest algorithm
Patent Information
- Application Number
- CN202611041088.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-14
- Publication Date
- 2026-10-30
AI Technical Summary
[0004]本发明提供一种基于LSTM-Iforest算法的船舶柴油机故障监测方法，以克服现有技术连续监测性能差、因诊断精准度受限导致增加误判概率、单一数据异常无法判定故障类型、因早期故障的参数微弱变化导致无法进行早期预警的问题
[0011]有益效果：本发明一种基于LSTM-Iforest算法的船舶柴油机故障监测方法，通过基于LSTM-Iforest算法构建的故障监测模型，仅利用正常运行数据进行无监督训练，无需大量带标注的故障样本，有效克服了实际工程中故障类型多样、标签稀缺的难题，降低了模型部署门槛，具有良好的泛化性和可推广性；
Smart Images
Figure CN123012864A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of fault diagnosis of marine power systems, and in particular to a fault monitoring method for marine diesel engines based on the LSTM-Iforest algorithm. Background Technology
[0002] As the primary vehicle for global trade, the reliability, safety, and economy of marine engines—the core power source of ships—directly affect navigation safety, operational efficiency, and environmental protection. With increasingly stringent emission regulations from the International Maritime Organization (IMO) and the trend towards intelligent and unmanned ships, unprecedentedly high standards have been placed on the health management of marine machinery. Marine engines, as the "heart" of a ship, have complex structures and operate under harsh conditions, enduring long-term high temperatures, high pressures, and high vibration loads, making them a high-risk component for ship malfunctions. These malfunctions often exhibit gradual, insidious, and interconnected characteristics; a major failure can lead not only to significant direct economic losses (such as repair costs and downtime) but also potentially to serious maritime safety accidents and environmental pollution incidents. Therefore, achieving early warning and accurate diagnosis of marine engine malfunctions is of paramount urgency.
[0003] Currently, the main methods for fault monitoring of marine diesel engines are as follows: The manual experience-based detection method mainly relies on the daily engine room inspections by staff. This method is highly subjective and depends entirely on the professional experience and operational skills of maintenance personnel, resulting in inconsistent judgment standards. It cannot achieve continuous monitoring around the clock and is prone to problems such as missed fault detection and misjudgment, making it difficult to meet the early warning requirements for the safe and stable operation of modern marine diesel engines. Vibration signal monitoring methods collect mechanical vibration signals in real time and extract fault characteristic parameters by combining time-domain and frequency-domain analysis to diagnose typical problems such as bearing wear, valve failure, connecting rod loosening, and abnormal cylinder impact. However, the dense equipment in the ship's engine room and various interferences can easily mask weak fault characteristic signals, resulting in a low signal-to-noise ratio. Diesel engines are variable-condition operating devices, and frequent changes in speed and load can cause vibration spectrum shifts, making fixed diagnostic thresholds less applicable. At the same time, different faults often produce similar vibration characteristics, making fault location and tracing difficult. Minor early fault characteristics are easily covered by background signals, limiting diagnostic accuracy. The oil physicochemical monitoring method uses diesel engine lubricating oil and fuel as the detection medium. By periodically extracting oil samples, it analyzes changes in oil viscosity, moisture, acidity, and the composition, morphology, and content of metal abrasive particles to determine the wear degree of friction pairs such as pistons, cylinder liners, and bearings, as well as the aging state of the lubrication system. However, this method mainly relies on offline periodic sampling and cannot provide real-time early warning of sudden faults. The monitoring scope is limited, and it can only effectively diagnose friction and wear problems, failing to accurately identify non-wear faults. Furthermore, the harsh operating conditions of ships can easily cause oil contamination, fuel dilution, and additive failure, further interfering with the test results and increasing the probability of misjudgment. The performance parameter monitoring method relies on the ship's automated control system to collect engine thermal and operating parameters in real time. By setting safety thresholds, alarms are triggered for abnormal fluctuations in parameters. However, various operating parameters of diesel engines are interdependent, and a single parameter anomaly cannot directly determine the fault type. When multiple faults occur concurrently, signal confusion is very likely to occur. The equipment operates in extreme environments for a long time, and temperature and pressure sensors are prone to zero-point drift and performance aging, resulting in distorted monitoring data. Most early faults only cause slight changes in parameters, which are difficult to exceed the alarm threshold, making it impossible to achieve early warning. False alarms and invalid alarms are prominent problems in long-term operation. Summary of the Invention
[0004] This invention provides a fault monitoring method for marine diesel engines based on the LSTM-Iforest algorithm, which overcomes the problems of poor continuous monitoring performance, increased probability of misjudgment due to limited diagnostic accuracy, inability to determine fault type due to single data anomalies, and inability to provide early warning due to slight changes in parameters of early faults.
[0005] To achieve the above objectives, the technical solution of the present invention is as follows: A method for fault monitoring of marine diesel engines based on the LSTM-Iforest algorithm includes: S1. Construct and calibrate a simulation model of a marine dual-fuel turbocharged engine; set multiple load conditions and collect normal operation data and fault operation data of the calibrated marine dual-fuel turbocharged engine simulation model under each load condition; S2. Based on the normal operation data and the fault operation data, through preprocessing operations, obtain the three-dimensional time series sample dataset of normal operation and the three-dimensional time series sample dataset of fault operation. S3. Construct a fault monitoring model based on the LSTM-Iforest algorithm; use the normal operating three-dimensional time series sample dataset as the training set and the faulty operating three-dimensional time series sample dataset as the test set; train the fault monitoring model based on the LSTM-Iforest algorithm to obtain the trained fault monitoring model based on the LSTM-Iforest algorithm. S4. Obtain the operating data of the marine dual-fuel turbocharged engine to be detected and preprocess it to obtain the three-dimensional time-series sample dataset to be detected; input the three-dimensional time-series sample dataset to be detected into the trained fault monitoring model based on the LSTM-Iforest algorithm to obtain fault labels, and issue fault warnings based on the fault labels.
[0006] Furthermore, the fault monitoring model based on the LSTM-Iforest algorithm includes an LSTM temporal autoencoder, a dynamic adaptive threshold module, a feature-enhanced isolated forest network, and a joint judgment module. The LSTM temporal autoencoder is used to obtain a reconstruction error sequence based on a three-dimensional temporal sample dataset. The dynamic adaptive threshold module is used to obtain an adaptive threshold based on the reconstructed error sequence; The feature-enhanced isolated forest network is used to obtain eight types of enhanced statistical features through a time-series window based on a three-dimensional time-series sample dataset, a reconstructed error sequence, and an adaptive threshold; and to obtain anomaly discrimination labels based on the eight types of enhanced statistical features through the isolated forest algorithm. The joint judgment module is used to obtain fault labels based on the reconstructed error sequence, adaptive threshold, and anomaly discrimination label.
[0007] Furthermore, the LSTM timing autoencoder includes an encoder, an intermediate layer, and a decoder; The encoder is used to obtain temporal deep features by sequentially passing through a first LSTM layer and a second LSTM layer based on a three-dimensional temporal sample dataset; and to obtain a one-dimensional hidden state vector by passing through a 16-neuron fully connected layer based on the temporal deep features. The intermediate layer is used to obtain the temporal dimension vector from the one-dimensional hidden state vector through the repeated vector layer; The decoder is used to obtain reconstructed temporal features by passing the temporal dimension vector through the third LSTM layer and the fourth LSTM layer in sequence; and to obtain the reconstructed error sequence by passing the reconstructed temporal features through a time-distributed fully connected layer.
[0008] Furthermore, the dynamic adaptive threshold module includes a Z-score algorithm submodule, an EWMA exponentially weighted mean algorithm submodule, an IQR quartile algorithm submodule, and a threshold selection strategy submodule; The Z-score algorithm submodule is used to obtain the first threshold by using the Z-score algorithm based on the reconstruction error sequence; The EWMA exponentially weighted mean algorithm submodule is used to obtain the second threshold based on the reconstructed error sequence using the EWMA exponentially weighted mean algorithm. The IQR quartile algorithm submodule is used to obtain the third threshold based on the reconstructed error sequence using the IQR quartile algorithm. The threshold selection strategy submodule obtains the minimum value among the three thresholds (first threshold, second threshold, and third threshold) and applies upper and lower limit constraints to obtain an adaptive threshold.
[0009] Furthermore, the eight types of enhanced statistical features include the mean of thermodynamic parameters in the three-dimensional time-series sample dataset, the standard deviation of thermodynamic parameters in the three-dimensional time-series sample dataset, the maximum value of thermodynamic parameters in the three-dimensional time-series sample dataset, the minimum value of thermodynamic parameters in the three-dimensional time-series sample dataset, the skewness of thermodynamic parameters in the three-dimensional time-series sample dataset, the kurtosis of thermodynamic parameters in the three-dimensional time-series sample dataset, the reconstruction error, and the ratio of the reconstruction error to the adaptive threshold.
[0010] Furthermore, the fault operation data includes single-cylinder fuel supply reduction fault, turbocharger efficiency reduction fault, fuel injection advance angle lag fault, air cooler efficiency reduction fault, and exhaust valve clearance expansion fault.
[0011] Beneficial effects: The present invention provides a fault monitoring method for marine diesel engines based on the LSTM-Iforest algorithm. The fault monitoring model built based on the LSTM-Iforest algorithm only uses normal operation data for unsupervised training, without the need for a large number of labeled fault samples. It effectively overcomes the problem of diverse fault types and scarce labels in actual engineering, reduces the threshold for model deployment, and has good generalization and scalability. LSTM networks can effectively capture long-term dependencies in time-series data and extract high-level time-series features; the IFOres algorithm has excellent anomaly sensitivity in high-dimensional space; the combination of the two can suppress false alarms caused by fluctuations in normal operating conditions and quickly locate abnormal patterns, significantly improving the accuracy and real-time performance of fault monitoring. Attached Figure Description
[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0013] Figure 1 This is a flowchart illustrating the fault monitoring method of the present invention; Figure 2 This is a normal data reconstruction error in the embodiments of the present invention; Figure 3These are some of the predicted fault points and actual fault points in the embodiments of the present invention; Figure 4 This is a comparison of the fault reconstruction error and dynamic threshold of single-cylinder fuel supply reduction in an embodiment of the present invention. Figure 5 This is a comparison of the turbocharger efficiency reduction fault reconstruction error and the dynamic threshold in an embodiment of the present invention. Figure 6 This is a comparison of the reconstruction error of the injection advance angle lag fault with the dynamic threshold in an embodiment of the present invention; Figure 7 This is a comparison of the fault reconstruction error and dynamic threshold of the air cooler efficiency reduction in this embodiment of the invention. Figure 8 This is a comparison result of the fault reconstruction error of the exhaust valve gap expansion and the dynamic threshold in an embodiment of the present invention. Detailed Implementation
[0014] 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 embodiments of the present invention, 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.
[0015] This embodiment provides a method for monitoring marine diesel engine faults based on the LSTM-Iforest algorithm, such as... Figure 1 As shown, it includes: S1. Construct and calibrate a simulation model of a marine dual-fuel turbocharged engine; set multiple load conditions and collect normal operation data and fault operation data of the calibrated marine dual-fuel turbocharged engine simulation model under each load condition; In this embodiment, the Wärtsilä 9L34DF dual-fuel turbocharged engine is taken as the research object. A thermodynamic simulation model of the engine is built in the AVL-BOOST engine simulation software. The model includes major components such as the intake system, turbocharging system, cylinder combustion chamber, exhaust system, and air cooler. First, the operational data from the actual ship bench test was obtained as the calibration benchmark. The output simulation results of the simulation model under four typical load conditions of 25%, 50%, 75%, and 100% were compared and verified with the actual ship bench test results. The results are shown in Table 1. The errors between the simulation model and the actual ship bench test data in seven key performance parameters, including power, mean effective pressure (BMEP), air cooler outlet temperature, turbocharger outlet pressure, cylinder exhaust temperature, intake air flow, and single-cylinder fuel consumption, are all less than 5%, which meets the accuracy requirements of fault simulation and can truly reflect the actual operating status of the marine diesel engine. Table 1 Actual and simulated values of various performance parameters under different operating conditions in fuel mode
[0016] After the model calibration is completed, different fault conditions are set in the AVL-BOOST software to collect normal engine operation data and five typical fault operation data.
[0017] Preferably, the fault operation data includes single-cylinder fuel supply reduction fault, turbocharger efficiency reduction fault, fuel injection advance angle lag fault, air cooler efficiency reduction fault, and exhaust valve clearance expansion fault.
[0018] Specifically, the normal operating data includes 15 thermodynamic parameters such as effective power, fuel consumption power, average effective pressure, maximum combustion pressure, maximum pressure rise and fall, compressor outlet pressure, compressor outlet temperature, intake air volume, intercooler intake temperature, intercooler outlet temperature, intercooler outlet pressure, cylinder exhaust temperature, intake manifold pressure, exhaust manifold pressure, and turbine mass flow rate. Some original sample data are shown in Table 2. Table 2 Partial Original Sample Data
[0019] This embodiment collected a total of 3,000 time series data points, including 2,000 data points under normal operating conditions and 1,000 data points under fault conditions (200 data points for each type of fault).
[0020] S2. Based on the normal operation data and the fault operation data, through preprocessing operations, obtain the three-dimensional time series sample dataset of normal operation and the three-dimensional time series sample dataset of fault operation. Specifically, the preprocessing operations include data noise reduction, data standardization, and data windowing; The data noiseification involves introducing uniformly distributed, small random fluctuations into the effective power to simulate load changes in a real-world environment. The remaining 14 thermodynamic parameters are not directly subjected to independent random noise, but are dynamically derived based on the laws of thermodynamics; for example: The compressor outlet temperature is calculated by coupling the isentropic compression relationship with the pressure. Exhaust manifold pressure is constrained by turbine mass flow rate and exhaust temperature; There is a theoretical correlation between cylinder burst pressure and fuel injection quantity, intake air temperature, and pressure. This physically constrained data augmentation method ensures that the generated dataset contains both the micro-state fluctuations in real operation and maintains the intrinsic correlation between thermodynamic parameters, avoiding the distortion of physical laws due to the introduction of unreasonable noise, and can also prevent model overfitting to a certain extent. The data noise reduction described herein is a prior art technique and will not be elaborated upon here.
[0021] The data standardization is achieved by normalizing the data using the Z-score standardization method. The Z-score standardization formula is:
[0022] In the formula, Standardized data; This is the original data; The mean of the features in the dataset; The standard deviation of the features in the dataset; After Z-score standardization, the data distribution of each feature has a mean of 0 and a standard deviation of 1, exhibiting a standard normal distribution.
[0023] The data windowing uses a sliding window technique to divide continuous time series data into fixed-length subsequence samples. The specific parameter settings are as follows: The window size (window_size) is set to 10, meaning that each sample contains data from 10 consecutive time points; The sliding step size is set to 1, meaning that a new sample is generated by sliding forward 1 time point each time. After windowing, the data dimension is transformed from two-dimensional (number of time points × 15 features) to three-dimensional (number of samples × 10 × 15), where: the first dimension is the number of samples; the second dimension is the length of the time window (10 time steps); and the third dimension is the feature dimension (15 thermodynamic parameters). This three-dimensional time-series sample can provide temporal context information for the LSTM network, enabling the model to learn "the evolution of each feature over 10 consecutive time steps" rather than just "the static relationship of 15 features at a certain moment"; the window size of 10 is determined based on optimization experiments, which can effectively capture the short-term dynamic features of diesel engine parameters without causing insufficient sample quantity due to an excessively long window. The windowing process described is existing technology, and the specific process will not be elaborated here.
[0024] S3. Construct a fault monitoring model based on the LSTM-Iforest algorithm; use the normal operating three-dimensional time series sample dataset as the training set and the faulty operating three-dimensional time series sample dataset as the test set; train the fault monitoring model based on the LSTM-Iforest algorithm to obtain the trained fault monitoring model based on the LSTM-Iforest algorithm. Among them, LSTM stands for Long Short-Term Memory, which is a special type of recurrent neural network capable of learning long-term dependency information; Iforest stands for Isolation Forest, which is an unsupervised anomaly detection algorithm based on a tree structure.
[0025] In this embodiment, after windowing processing, 3000 time series data points (including 2000 normal operating condition data points and 1000 fault operating condition data points) are converted into a three-dimensional time series sample dataset and divided into: Training set 3D time series sample dataset: contains 3D time series samples corresponding to normal operation data, used for model training; Test set 3D time series sample dataset: contains 3D time series samples corresponding to normal operation data and fault operation data, used for model validation; During the training phase, the model is unsupervised learning using only normally functioning 3D time-series sample datasets (training set), enabling the model to learn the data distribution characteristics and time-series evolution patterns under normal operating conditions. The three-dimensional time-series sample dataset (test set) from the faulty operation is not used in the training process, but only for subsequent model performance verification and evaluation.
[0026] Preferably, the fault monitoring model based on the LSTM-Iforest algorithm includes an LSTM temporal autoencoder, a dynamic adaptive threshold module, a feature-enhanced isolated forest network, and a joint judgment module; The LSTM temporal autoencoder is used to obtain a reconstruction error sequence based on a three-dimensional temporal sample dataset. Preferably, the LSTM timing autoencoder includes an encoder, an intermediate layer, and a decoder; The encoder is used to obtain temporal deep features by sequentially passing through a first LSTM layer and a second LSTM layer based on a three-dimensional temporal sample dataset; and to obtain a one-dimensional hidden state vector by passing through a 16-neuron fully connected layer based on the temporal deep features. Specifically, the first LSTM layer has 64 LSTM neurons to output complete temporal features based on a three-dimensional temporal sample dataset; The second LSTM layer has 32 LSTM neurons to output deep temporal features based on complete temporal features; The 16-neuron fully connected layer has 16 neurons and the activation function is ReLU, which further compresses the temporal deep features output by LSTM into a one-dimensional hidden state vector. The intermediate layer is used to obtain the temporal dimension vector from the one-dimensional hidden state vector through the repeated vector layer; Specifically, the intermediate layer uses a RepeatVector layer to copy the one-dimensional hidden state vector (dimension 16) output by the encoder multiple times to generate a three-dimensional tensor with dimensions (batch_size, 10, 16) to adapt to the input requirements of the decoder. The RepeatVector layer is a standard layer type in the Keras deep learning framework. Its function is to repeat a one-dimensional vector a specified number of times in the time dimension, thereby expanding the compressed hidden state information into a temporal representation with the same length as the input temporal window.
[0027] The decoder is used to obtain reconstructed temporal features by passing the temporal dimension vector through the third LSTM layer and the fourth LSTM layer in sequence; and to obtain the reconstructed error sequence by passing the reconstructed temporal features through a time-distributed fully connected layer.
[0028] Specifically, the third LSTM layer has the same structure as the second LSTM layer, with 32 LSTM neurons; The structure of the fourth LSTM layer is the same as that of the first LSTM layer, and it has 64 LSTM neurons. The time-distributed fully connected layer is applied independently to each time step, with an output dimension of 15 (corresponding to 15 thermodynamic parameters), realizing time-step reconstruction.
[0029] In this embodiment, the LSTM temporal autoencoder is trained unsupervised using 2000 normal operating condition data points by employing the Adam optimizer, MSE loss function, and early stopping strategy; the maximum number of training iterations is set to 50. During training, the LSTM autoencoder learns the temporal patterns and inherent correlations of normal data; after training, the model can achieve high-precision reconstruction of normal data (with small reconstruction errors), while the reconstruction error of faulty data that deviates from the normal pattern increases significantly; for example... Figure 2 The reconstruction error distribution of normal data; The unsupervised learning described is an existing technology and will not be elaborated upon here.
[0030] The dynamic adaptive threshold module is used to obtain an adaptive threshold based on the reconstructed error sequence; Preferably, the dynamic adaptive threshold module includes a Z-score algorithm submodule, an EWMA exponentially weighted mean algorithm submodule, an IQR quartile algorithm submodule, and a threshold selection strategy submodule; The Z-score algorithm submodule is used to obtain the first threshold by using the Z-score algorithm based on the reconstruction error sequence; Specifically, the first threshold is calculated and obtained by calculating the mean and standard deviation of all elements in the reconstructed error sequence based on the mean and standard deviation. The EWMA exponentially weighted mean algorithm submodule is used to obtain the second threshold based on the reconstructed error sequence using the EWMA exponentially weighted mean algorithm. Specifically, by assigning decreasing weights to historical errors in the reconstructed error sequence, with higher weights to recent data, the latest trend value of the error distribution is captured; and the second threshold is calculated based on the trend value. The IQR quartile algorithm submodule is used to obtain the third threshold based on the reconstructed error sequence using the IQR quartile algorithm. Specifically, calculate the quartiles Q1 (25th quartile), Q3 (75th quartile), and interquartile range IQR = Q3 - Q1 of the error queue; and calculate the third threshold based on Q3 and IQR. The Z-score algorithm, EWMA exponentially weighted mean algorithm, and IQR quartile algorithm are existing technologies and will not be described in detail here. The threshold selection strategy submodule obtains the minimum value among the three thresholds (first threshold, second threshold, and third threshold) and applies upper and lower limit constraints to obtain an adaptive threshold.
[0031] The threshold selection strategy submodule updates the error distribution characteristics in real time through a fixed-length error sliding queue to dynamically adjust the upper and lower limit constraint ranges. The fixed-length error sliding queue is a sliding error queue (deque data structure) with a length of 100, which stores the reconstruction error values of the most recent 100 samples in real time. The queue adopts a first-in-first-out (FIFO) mechanism, and automatically removes the oldest error value each time a new error value is added, so as to keep the queue length constant. The lower limit constraint is 50% of the initial static threshold, and the upper limit constraint is twice the initial static threshold; the initial static threshold is based on the 95th percentile (Q) of the reconstruction error of the normal data in the training set. 95 )Sure.
[0032] The feature-enhanced isolated forest network is used to obtain eight types of enhanced statistical features through a time-series window based on a three-dimensional time-series sample dataset, a reconstructed error sequence, and an adaptive threshold; and to obtain anomaly discrimination labels based on the eight types of enhanced statistical features through the isolated forest algorithm. In this embodiment, during the online detection phase, for each new time-series window sample processed, the following steps are performed: calculate the reconstruction error of the sample; add the error value to the sliding queue; recalculate the dynamic threshold based on the updated queue; and use the new threshold to determine whether the current sample is abnormal. If the current sample is abnormal, then remove the current sample; otherwise, extract 8 types of enhanced statistical features for each time window. Preferably, the eight types of enhanced statistical features include the mean of thermodynamic parameters in the three-dimensional time-series sample dataset, the standard deviation of thermodynamic parameters in the three-dimensional time-series sample dataset, the maximum value of thermodynamic parameters in the three-dimensional time-series sample dataset, the minimum value of thermodynamic parameters in the three-dimensional time-series sample dataset, the skewness of thermodynamic parameters in the three-dimensional time-series sample dataset, the kurtosis of thermodynamic parameters in the three-dimensional time-series sample dataset, the reconstruction error, and the ratio of the reconstruction error to the adaptive threshold. Each time-series window sample can be extracted from the above 8 types of features, and each type of feature can be further subdivided to obtain the following results: The mean values of thermodynamic parameters in the three-dimensional time-series sample dataset form a 15-dimensional feature vector. The standard deviations of thermodynamic parameters in the three-dimensional time-series sample dataset form a 15-dimensional feature vector. The maximum values of thermodynamic parameters in the three-dimensional time-series sample dataset form a 15-dimensional feature vector. The minimum values of thermodynamic parameters in the three-dimensional time-series sample dataset form a 15-dimensional feature vector. The skewness of thermodynamic parameters in a three-dimensional time-series sample dataset forms a one-dimensional feature vector. The kurtosis of thermodynamic parameters in a three-dimensional time-series sample dataset forms a one-dimensional feature vector. The reconstruction error forms a 1-dimensional feature vector. The ratio of reconstruction error to adaptive threshold forms a 1-dimensional feature vector; The final result is a high-dimensional (64-dimensional) statistical feature vector; these features characterize the statistical distribution characteristics of the time series window from different perspectives, which is more conducive to the discovery of anomaly patterns in isolated forests compared with the original time series data.
[0033] In this embodiment, the contamination rate of the isolated forest is adaptively determined based on the actual proportion of high-error samples in the training set. At the same time, the isolated forest model is trained by using the extracted augmented statistical features. After training, the isolated forest can identify anomalies in the augmented features of new samples and output anomaly scores and binary labels (normal / abnormal). The training process and isolated forest described are existing technologies and will not be elaborated upon here.
[0034] The joint judgment module is used to obtain fault labels based on the reconstructed error sequence, adaptive threshold, and anomaly discrimination label.
[0035] Specifically, the joint judgment module, in order to reduce the false alarm rate and false negative rate of a single algorithm, uses a dual-condition joint judgment strategy based on the reconstructed error sequence, adaptive threshold, and anomaly discrimination label. A fault state is determined when one of the following two conditions is met: Condition 1: The reconstruction error calculated by the LSTM autoencoder is greater than the dynamic adaptive threshold. Condition 2: Isolated forest discrimination result = anomaly When either the time-series reconstruction anomaly (LSTM determination) or the high-dimensional feature anomaly (isolated forest determination) occurs, a fault warning is triggered, a fault label is output, and an alarm signal is issued. This dual verification mechanism effectively integrates the advantages of temporal deep learning and traditional machine learning, significantly improving the accuracy and robustness of fault detection under complex working conditions.
[0036] S4. Obtain the operating data of the marine dual-fuel turbocharged engine to be detected and preprocess it to obtain the three-dimensional time-series sample dataset to be detected; input the three-dimensional time-series sample dataset to be detected into the trained fault monitoring model based on the LSTM-Iforest algorithm to obtain fault labels, and issue fault warnings based on the fault labels.
[0037] In a specific embodiment, the test set is input into a fault monitoring model based on the LSTM-Iforest algorithm for anomaly monitoring. For each test sample, the following steps are executed sequentially: Step 1: Input the test sample into the LSTM autoencoder to obtain the reconstruction error; Step 2: Add the reconstruction error to the sliding error queue, and the dynamic threshold module calculates the current dynamic threshold based on the updated queue; Step 3: Extract 8 types of enhanced statistical features from the sample, input them into the isolated forest model for anomaly detection, and obtain the classification label; Step 4: Apply dual-condition joint judgment logic: If the judgment condition is met, output a fault label and issue an early warning; otherwise, determine it as a normal state. The final result is as follows Figures 3 to 8 As shown, Figure 3 The comparison between some predicted fault points and actual fault points is shown, demonstrating that the present invention can accurately identify the fault moment. Figures 4 to 8 The reconstruction error and dynamic threshold comparison curves under five types of fault conditions are shown respectively. After analysis, it can be seen that the fault monitoring model based on LSTM-Iforest algorithm has an accuracy of 97.23%, a precision of 96.65%, and an F1 score of 0.98. Compared with traditional algorithm models, the standard isolated forest algorithm has an accuracy of only 85%, and the standard autoencoder has an accuracy of 96%. This shows that the performance of our model is almost optimal, and the diagnostic accuracy is significantly higher than that of traditional algorithm models. This further demonstrates that the fault monitoring model based on the LSTM-Iforest algorithm has high recognition accuracy, is effective in monitoring faults in marine diesel engines, and has high practicality.
[0038] The present invention has the following beneficial effects: This invention provides a fault monitoring method for marine diesel engines based on the LSTM-Iforest algorithm. The fault monitoring model built based on the LSTM-Iforest algorithm uses only normal operation data for unsupervised training, without the need for a large number of labeled fault samples. This effectively overcomes the difficulties of diverse fault types and scarce labels in actual engineering, lowers the threshold for model deployment, and has good generalization and scalability. LSTM networks can effectively capture long-term dependencies in time-series data and extract high-level time-series features; the IFOres algorithm has excellent anomaly sensitivity in high-dimensional space; the combination of the two can suppress false alarms caused by fluctuations in normal operating conditions and quickly locate abnormal patterns, significantly improving the accuracy and real-time performance of fault monitoring.
[0039] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for fault monitoring of marine diesel engines based on the LSTM-Iforest algorithm, characterized in that, include: S1. Construct and calibrate a simulation model of a marine dual-fuel turbocharged engine; Multiple load conditions were set up, and normal operation data and fault operation data of the calibrated marine dual-fuel turbocharged engine simulation model under each load condition were collected. S2. Based on the normal operation data and the fault operation data, through preprocessing operations, obtain the three-dimensional time series sample dataset of normal operation and the three-dimensional time series sample dataset of fault operation. S3. Construct a fault monitoring model based on the LSTM-Iforest algorithm; The normal operating three-dimensional time series sample dataset is used as the training set, and the faulty operating three-dimensional time series sample dataset is used as the test set; the fault monitoring model based on the LSTM-Iforest algorithm is trained to obtain the trained fault monitoring model based on the LSTM-Iforest algorithm. S4. Obtain the operating data of the marine dual-fuel turbocharged engine to be detected and preprocess it to obtain the three-dimensional time-series sample dataset to be detected; input the three-dimensional time-series sample dataset to be detected into the trained fault monitoring model based on the LSTM-Iforest algorithm to obtain fault labels, and issue fault warnings based on the fault labels.
2. The method for monitoring marine diesel engine faults based on the LSTM-Iforest algorithm according to claim 1, characterized in that, The fault monitoring model based on the LSTM-Iforest algorithm includes an LSTM temporal autoencoder, a dynamic adaptive threshold module, a feature-enhanced isolated forest network, and a joint judgment module. The LSTM temporal autoencoder is used to obtain a reconstruction error sequence based on a three-dimensional temporal sample dataset. The dynamic adaptive threshold module is used to obtain an adaptive threshold based on the reconstructed error sequence; The feature-enhanced isolated forest network is used to obtain eight types of enhanced statistical features through a time-series window based on a three-dimensional time-series sample dataset, a reconstructed error sequence, and an adaptive threshold. Based on eight types of enhanced statistical features, anomaly detection labels are obtained using the isolated forest algorithm; The joint judgment module is used to obtain fault labels based on the reconstructed error sequence, adaptive threshold, and anomaly discrimination label.
3. The method for monitoring marine diesel engine faults based on the LSTM-Iforest algorithm according to claim 2, characterized in that, The LSTM timing autoencoder includes an encoder, an intermediate layer, and a decoder; The encoder is used to obtain deep temporal features by sequentially passing the data through a first LSTM layer and a second LSTM layer based on a three-dimensional temporal sample dataset. Based on the deep temporal features, a one-dimensional hidden state vector is obtained through a 16-neuron fully connected layer; The intermediate layer is used to obtain the temporal dimension vector from the one-dimensional hidden state vector through the repeated vector layer; The decoder is used to obtain reconstructed temporal features by passing the temporal dimension vector through the third LSTM layer and the fourth LSTM layer in sequence; and to obtain the reconstructed error sequence by passing the reconstructed temporal features through a time-distributed fully connected layer.
4. The method for monitoring marine diesel engine faults based on the LSTM-Iforest algorithm according to claim 2, characterized in that, The dynamic adaptive threshold module includes a Z-score algorithm submodule, an EWMA exponentially weighted mean algorithm submodule, an IQR quartile algorithm submodule, and a threshold selection strategy submodule. The Z-score algorithm submodule is used to obtain the first threshold by using the Z-score algorithm based on the reconstruction error sequence; The EWMA exponentially weighted mean algorithm submodule is used to obtain the second threshold based on the reconstructed error sequence using the EWMA exponentially weighted mean algorithm. The IQR quartile algorithm submodule is used to obtain the third threshold based on the reconstructed error sequence using the IQR quartile algorithm. The threshold selection strategy submodule obtains the minimum value among the three thresholds (first threshold, second threshold, and third threshold) and applies upper and lower limit constraints to obtain an adaptive threshold.
5. The method for monitoring marine diesel engine faults based on the LSTM-Iforest algorithm according to claim 2, characterized in that, The eight types of enhanced statistical features include the mean of thermodynamic parameters in the three-dimensional time series sample dataset, the standard deviation of thermodynamic parameters in the three-dimensional time series sample dataset, the maximum value of thermodynamic parameters in the three-dimensional time series sample dataset, the minimum value of thermodynamic parameters in the three-dimensional time series sample dataset, the skewness of thermodynamic parameters in the three-dimensional time series sample dataset, the kurtosis of thermodynamic parameters in the three-dimensional time series sample dataset, the reconstruction error, and the ratio of the reconstruction error to the adaptive threshold.
6. The method for monitoring marine diesel engine faults based on the LSTM-Iforest algorithm according to claim 1, characterized in that, The fault operation data includes single-cylinder fuel supply reduction fault, turbocharger efficiency reduction fault, fuel injection advance angle lag fault, air cooler efficiency reduction fault, and exhaust valve clearance expansion fault.