Marine diesel engine fault processing method based on thermodynamics-assisted random forest

Abstract

The invention relates to the technical field of ship power fault identification and processing, and in particular to a ship diesel engine fault processing method based on thermodynamics-assisted random forest, which can improve the efficiency and accuracy of ship power main engine fault analysis and processing, and comprises the following steps: step 1: constructing a one-dimensional thermodynamic model to simulate the working process of the ship diesel engine, and calibrating the model with the help of data collected by a ship main engine data acquisition system DCM; step 2: simulating five typical faults of a combustion chamber component of a ship diesel engine by calibrating key system parameters, and preliminarily outputting thermal parameters related to the faults; step 3: using random forest to pre-identify faults, calculating the SHAP value of each thermal parameter, and screening out the most influential parameters, and the screening results of the parameters will be analyzed in combination with the one-dimensional thermodynamic model; step 4: the screened parameters are re-input into the random forest for training and verification.

Description

Technical Field

[0001] The present invention relates to the technical field of ship power fault identification and processing, and in particular to a ship diesel engine fault processing method based on thermodynamics-assisted random forest, which can improve the efficiency and accuracy of ship power main engine fault analysis and processing. Background Art

[0002] The key component in the ship propulsion system is the diesel engine, which has a complex internal structure and many components. For example, the combustion chamber, as an important part of the diesel engine, includes the cylinder head, cylinder liner, piston and some other components. Due to the harsh working environment, the combustion chamber has become one of the main parts of diesel engine failure. Combustion chamber failure will directly affect the operating efficiency and safety of the ship's main engine diesel engine. Therefore, fault diagnosis is crucial in diesel engine management. By monitoring thermal parameters, the fault diagnosis model can accurately locate and identify faults, providing strong support for maintenance timing prediction, plan formulation and cost estimation. However, in actual ship operation, fault diagnosis faces many challenges, such as small number of fault samples, long monitoring cycle, high collection cost, etc., which makes diagnosis challenging.

[0003] At present, the fault diagnosis methods of marine diesel engines are mainly divided into three categories: model-based, data-based, and the combined drive of the two. Such methods need to use physical laws, engineering experience, and the working principle of diesel engines to build a mathematical model of the diesel engine. The model is simulated by simulation software to analyze the behavior of the diesel engine under normal and fault conditions. However, under the existing conditions, it is very difficult to obtain an accurate physical mathematical model of marine diesel. Establishing an accurate mathematical model requires a lot of prior knowledge, and the model is often very complex, making it difficult to handle the coupling relationship between nonlinearity and multiple variables.

[0004] Data-driven fault diagnosis methods do not require accurate modeling of the physical characteristics of complex systems such as ship propulsion engines. However, data-driven models mainly rely on learning feature relationships from data rather than being based on clear physical laws, so it is difficult to explain the decision-making basis of the model, especially in complex models such as deep learning. The interaction of complex nonlinear relationships and high-dimensional features often turns the model into a black box, and the decision-making process and reasons are difficult to understand. Data and model-driven methods usually first establish a physical mathematical model of the fault, then use experimental data to calibrate the model, and finally diagnose the simulated output data or mixed data. However, these methods do not conduct in-depth analysis of the model's decision-making process in combination with the physical model. The physical model plays a more important role in fault simulation and supplementing the data set. Summary of the invention

[0005] In view of the problems that traditional data-driven models lack knowledge support for fault mechanisms and have poor interpretability, the present invention proposes a marine diesel engine fault processing method based on thermodynamics-assisted random forest, which can improve the efficiency and accuracy of ship power main engine fault analysis and processing.

[0006] The present invention is achieved by the following measures:

[0007] A method for troubleshooting a marine diesel engine based on thermodynamics-assisted random forests, comprising the following steps:

[0008] Step 1: Construct a one-dimensional thermodynamic model to simulate the working process of the marine diesel engine, and calibrate the model with the data collected by the ship's main engine data acquisition system DCM;

[0009] Step 2: By calibrating the key system parameters, five typical faults of the combustion chamber components of a marine diesel engine are simulated, and the thermal parameters related to the faults are preliminarily output;

[0010] Step 3: Use random forest to pre-identify faults, calculate the SHAP value of each thermal parameter, and screen out the most influential parameters. The screening results of the parameters will be analyzed in combination with the one-dimensional thermodynamic model;

[0011] Step 4: The selected parameters are re-input into the random forest for training and validation.

[0012] The marine diesel engine involved in constructing the one-dimensional thermodynamic model in step 1 of the present invention is a low-speed, six-cylinder, two-stroke marine diesel engine equipped with a turbocharger and an intercooler. The one-dimensional thermodynamic model includes a system boundary, an intake and exhaust manifold, a turbocharger TC1, an intercooler CO1, and six cylinders C1 to C6. Six monitoring points are also set up, respectively denoted as MP1 to MP6, for monitoring key engine parameters. Pipeline 1 represents the compressor intake pipe, pipeline 2 refers to the intercooler intake pipe, pipeline 3 is the intercooler exhaust pipe, pipelines 4 to 9 correspond to the cylinder intake pipes, and pipelines 10 to 15 are the cylinder exhaust pipes. Pipeline 16 is connected to the turbocharger intake pipe, and pipeline 17 is the turbocharger exhaust pipe.

[0013] In step 1 of the present invention, a single Weber model is used to establish a combustion heat release model of a diesel engine, and the expression is as follows:

[0014] (1), where: Q It is the heat energy released by the combustion of fuel mixture in one cycle, kJ; θ is the crankshaft angle, (°); θ 0 is the angle of the crankshaft when the fuel mixture in the cylinder starts to burn (°); a and bare the parameters of the Weibull function, which determine the duration and shape of the combustion process,

[0015] By integrating the Weibull function, we can get the mass fraction of fuel consumed from the start of combustion to a certain time, that is, the mass fraction X :

[0016] (2) The Woschni1978 model is also used to estimate the convection heat transfer coefficient of the combustion gas in the cylinder, and the expression is as follows:

[0017] (3), where: h is the heat transfer coefficient, W / m²K; p is the instantaneous pressure in the cylinder, MPa; T g is the instantaneous temperature of the gas, K; v is the characteristic velocity of the gas, m / s; C 1 is an empirical constant, characteristic velocity v The calculation is based on the movement in the cylinder, and the calculation formula given by Woschni is as follows:

[0018] (4), where: V d is the displacement of the cylinder in m³; V is the cylinder volume during combustion, m³; p m is the average pressure in the cylinder, MPa; v p is the speed of the piston, m / s; C 2 and C 3 is an empirical constant, and its specific value depends on the engine design and operating conditions.

[0019] In step 2 of the present invention, five typical faults of the combustion chamber components of a marine diesel engine are simulated in the following manner: there are six health states of the combustion chamber components, represented by codes F0 to F6: normal state (F0), cylinder head crack (F1), piston top ablation (F2), cylinder liner wear (F3), piston ring wear (F4), piston ring adhesion (F5), and the parameter calibration method of the five fault modes (F1-5) is:

[0020] Cylinder head crack (F1): Cylinder head crack fault is simulated by setting the surface temperature of the cylinder head. The cylinder head is subjected to high temperature and pressure during operation, especially the temperature near the combustion chamber is extremely high. When cracks appear, the structural strength of these areas is reduced, resulting in increased local thermal stress concentration. The cooling effect of the cylinder head will be affected by the presence of cylinder head cracks, resulting in local surface overheating;

[0021] Piston top ablation (F2): By setting the piston surface temperature and blowby mass flow rate, the piston top ablation fault is simulated. As the piston top surface material is ablated, the heat distribution on the piston surface is uneven, and the local temperature of the ablation area increases significantly. In addition, the loss of surface material caused by ablation will affect the sealing between the piston and the cylinder wall, thereby causing blowby;

[0022] Cylinder liner wear (F3): By setting the cylinder diameter and blowby mass flow rate, the cylinder liner wear fault is simulated. Failure of the air filter or poor sealing can cause foreign matter to enter the combustion chamber, thus causing abnormal wear of the cylinder liner. The worn cylinder liner will cause roundness error, increase the local cylinder diameter, and affect the sealing performance of the cylinder, which will lead to the occurrence of blowby;

[0023] Piston ring wear (F4): Set the blowby mass flow rate to simulate the cylinder head crack failure. The piston ring may be worn due to factors such as impurities in the cylinder, resulting in the end face being out of round, the cylinder sealing performance being reduced, and eventually causing blowby;

[0024] Piston ring adhesion (F5): By setting the blowby mass flow rate, cylinder diameter and piston surface temperature, the piston ring adhesion failure is simulated. When the engine has too much carbon deposits, insufficient lubrication or sludge accumulation, the piston ring movement may be restricted and lose its sealing function. Under normal operating conditions, the piston ring can transfer the heat from the top of the piston to the cylinder wall, thereby effectively dissipating heat. However, adhesion failure significantly weakens this heat conduction process, resulting in a significant increase in the piston surface temperature. In addition, adhesion destroys the normal sliding between the piston ring and the cylinder liner, and the friction in the local contact area increases sharply, further exacerbating the wear of the cylinder liner.

[0025] Step 3 of the present invention specifically includes:

[0026] Step 3-1: Screen out potentially relevant thermal parameters from the diesel engine fault simulation output to form an original data set;

[0027] Step 3-2: Input the data set into the random forest for pre-identification. Based on the pre-identification results, the Tree SHAP method is used to calculate the SHAP value of each parameter, and the model pre-identification process is visualized;

[0028] Step 3-3: Combine the parameter density scatter plot and feature interaction plot to perform an explanatory analysis of the importance assessment results;

[0029] Step 3-4: Screen out key thermal parameters based on the importance assessment results, generate an optimized data set, and re-input the optimized data set into the random forest model for further processing.

[0030] In step 3-1 of the present invention, in the process of calculating the SHAP value, the contribution value is used to describe the influence of the thermal parameters on the model output. The marginal contribution of the thermal parameter i after adding it under a certain parameter combination is It is expressed by the following formula: (5), where: S is the set of other parameters that does not include parameter i; f(S) is the contribution value generated by the set S; f (S {i}) is the contribution value after parameter i is added. After considering all possible parameter combinations, the average contribution value increment of parameter i, that is, the SHAP value φi, is calculated as follows:

[0031] (6),

[0032] Where: N is the set of all parameters; S is a subset of the parameters used in the model; p Representation Subset S The number of parameters in ;

[0033] For a certain parameter combination S When performing an overall evaluation, the SHAP value of each parameter is added to obtain the evaluation result of the combination. Whenever a parameter is added to the parameter combination, the output is the original SHAP value plus the SHAP value of the newly added parameter. S The final evaluation result is the algebraic sum of all parameter SHAP values, that is: (7), where: φ 0 is the baseline value, which means the average output of the model without any parameters; The SHAP value is the sum of the SHAP values ​​predicted by each feature pair model. The SHAP interaction value is used to further explore the potential relationship between thermal parameters. i and j The interaction value φ i,j Defined as: (8) When i ≠ j and: (9) ， By combining formulas (8-9), the SHAP interaction value between parameters can be calculated.

[0034] In step 3-2 of the present invention, Tree SHAP only calculates the marginal contribution along the path that meets the current input conditions, thereby avoiding enumerating all parameter combinations and greatly improving the calculation efficiency. i , the SHAP value calculation formula based on TreeSHAP is as follows: (10), where: T is the number of decision trees; Pj For the j The path set of a tree; S is a subset of the features used in the model; p For subset S The number of features in .

[0035] In step 3-2 of the present invention, random forest is used as the basic model for fault identification, and SHAP and thermodynamic model are combined to perform interpretability analysis on random forest. The calculation of SHAP value is based on the decision process of random forest. Considering that thermal parameters are continuous variables, these parameters need to be discretized to meet the input requirements of decision tree model. The decision tree splitting method for thermal parameters is as follows: Decision tree is the basis of random forest algorithm, and prediction is made by dividing feature space. For a sample set D , assuming that k The proportion of class samples is p k , then the information entropy of D is defined as: (11), assuming the use of discrete attributes For samples D Dividing it will produce V branch node, in v In the branch node, the attribute a The upper value is a v The sample is recorded as D v , then we can calculate according to formula (9) D v Considering that different branch nodes contain different numbers of samples, weights are assigned to branch nodes. , you can calculate the attribute a For the sample set D The information gain obtained by partitioning is: (12) The larger the information gain, the more helpful the attribute is in distinguishing data categories and can more effectively reduce the uncertainty of data. Therefore, when the information gain of a feature is large, it is more likely to be selected as a split node in the decision tree;

[0036] For the thermal parameters of ship diesel engines, such as temperature, pressure and speed, since the values ​​of these parameters are continuous, it is impossible to directly divide the nodes according to their values, so it is necessary to discretize the continuous attributes: given a fault sample D and thermal parameters P , assuming P exist D There is nDifferent values, arranged from small to large, are , set a partition point t , will be in the parameter P The upper value is not greater than t The sample is marked as D - t , and the value is greater than t The sample is marked as D + t , obviously, for P i and P i+1 As far as the value of t is concerned, as long as it falls within the interval In the above example, the division results are consistent. Therefore, for thermal parameters P , the interval The median point As its candidate partition point, it contains the set of all candidate partition points T P It can be expressed as:

[0037] (13), therefore, the thermal parameters P For fault samples D The information gain obtained by partitioning is Gain( D , P ) can be expressed as:

[0038] (14), when Gain( D , P ) is relatively large, the decision tree will be more inclined to select thermal parameters P As a division node. In step 3-3 of the present invention, the parameter density scatter plot and the feature interaction diagram are combined to perform an explanatory analysis on the importance assessment results, wherein the distribution changes of the sample points caused by the changes in thermal parameters are given by the parameter density scatter plot. Each row in the parameter density scatter plot represents a parameter, and the horizontal axis represents the SHAP value. The parameters are sorted according to the average absolute SHAP value, and the most important parameters appear at the top. Each point corresponds to a sample, and the color of the sample point reflects the size of the parameter value. A SHAP value less than 0 indicates a negative impact on the prediction, while a SHAP value greater than 0 indicates a positive impact. The horizontal distribution of the points reflects the size of the impact. The wider the distribution, the greater the impact.

[0039] In order to deeply study the interaction between a certain thermal parameter and other parameters, the SHAP interaction value of each parameter was further calculated. The influence trend of the target parameter on the model output and its interaction with other parameters were intuitively displayed through the characteristic interaction diagram. In the figure, the X-axis represents the actual value of the parameter, and the Y-axis corresponds to its SHAP value. Each point represents a data sample, and the color of the point reflects the size of the interaction parameter value.

[0040] The present invention introduces a thermodynamic model into the SHAP analysis results of the random forest to improve the credibility of the model. The thermodynamic model provides a more in-depth explanation for fault diagnosis in the following three aspects: (1) Fault cause: By explaining the physical change mechanism of each parameter under a specific fault, the fault cause behind the statistical information reflected by the SHAP value is revealed. Since the essence of the SHAP value is to reflect the correlation between parameters rather than the causal relationship, the specific physical change mechanism of each parameter when the fault occurs can be further clarified through the auxiliary analysis of the thermodynamic model, thereby enhancing the explanatory depth of the model. (2) Important parameters: By analyzing the key driving parameters that affect the occurrence of the fault, the physical meaning of the important parameters is revealed by combining the SHAP value and the thermodynamic model. The SHAP value can identify the parameters that contribute more to the model prediction results, and the introduction of the thermodynamic model can deeply analyze the physical driving effect of these key parameters on the fault process, helping to understand their practical significance for fault diagnosis. (3) Parameter relationship: By revealing the interaction between parameters and their physical background, the interpretation of the model prediction results is further improved. The SHAP value can not only quantify the importance of a single feature, but also describe the interaction between features and their joint influence on the prediction results. With the help of thermodynamic models, we can deeply analyze the correlation between these parameters and their dynamic mechanisms. By analyzing the interaction of parameters, we can fully understand the dynamic nature of the fault process, thus providing a scientific basis for more accurate fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Attached Figure 1 It is a flow chart of the present invention.

[0042] Attached Figure 2 It is a schematic diagram of the thermodynamic model of the marine diesel engine in the present invention.

[0043] Attached Figure 3 It is a parameter screening flow chart based on SHAP in the present invention.

[0044] Attached Figure 4 It is a schematic diagram of the marginal contribution of parameter i in the present invention.

[0045] Attached Figure 5 It is a schematic diagram of the additivity of SHAP values ​​in the present invention.

[0046] Attached Figure 6is a comparison diagram of SHAP and Tree SHAP in an embodiment of the present invention, wherein Figure 6 (a) is the flow chart of direct processing of SHAP data; Figure 6 (b) is the TreeSHAP data processing flow chart.

[0047] Attached Figure 7 It is a structural schematic diagram of the DCM signal acquisition system in the present invention.

[0048] Attached Figure 8 is a diagram of the thermodynamic fault simulation result in the present invention, wherein Figure 8 (a) is the simulation result diagram of the in-cylinder pressure fault, (b) is the simulation result diagram of the in-cylinder temperature fault, (c) is the simulation result diagram of the exhaust pressure fault, (d) is the simulation result diagram of the exhaust temperature fault, (e) is the simulation result diagram of the piston surface heat flow fault, (f) is the simulation result diagram of the cylinder head surface heat flow fault, (g) is the simulation result diagram of the cylinder liner surface heat flow fault, (h) is the simulation result diagram of the blowby heat flow fault, (i) is the simulation result diagram of the blowby mass flow rate fault, (j) is the simulation result diagram of the turbocharger efficiency fault, (k) is the simulation result diagram of the turbocharger intake pressure fault, (l) is the simulation result diagram of the turbocharger intake temperature fault, (m) is the simulation result diagram of the turbocharger exhaust pressure fault, and (n) is the simulation result diagram of the turbocharger exhaust temperature fault.

[0049] Attached Fig. 9 is a schematic diagram of the importance of thermal parameters based on SHAP values ​​in the present invention, where Fig. 9 (a) is a stacked bar chart of importance. Fig. 9 (b) is the importance percentage graph.

[0050] Attached Fig.10 The evaluation process of the fault type of a certain sample point in the embodiment of the present invention, wherein Fig.10 (a) is a schematic diagram of the normal state. Fig.10 (b) is a schematic diagram of the cylinder head crack assessment process, (c) is a schematic diagram of the piston top ablation assessment process, Fig.10 (d) is a schematic diagram of the cylinder liner wear assessment process. Fig.10 (e) is a schematic diagram of the piston ring wear assessment process. Fig.10 (f) is a schematic diagram of the piston ring adhesion evaluation process.

[0051] Attached Fig.11 is a parameter density scatter plot based on SHAP values ​​in an embodiment of the present invention, wherein Fig.11 (a) is the normal state; (b) is the cylinder head crack; (c) is the piston top burnout; (d) is the cylinder liner wear; (e) is the piston ring wear; (f) is the piston ring adhesion.

[0052] Attached Fig.12 is the SHAP interaction value of each parameter in the embodiment of the present invention, where Fig.12 (a) is the normal state; (b) is the cylinder head crack; (c) is the piston top burnout; (d) is the cylinder liner wear; (e) is the piston ring wear; (f) is the piston ring adhesion.

[0053] Attached Fig.13 Schematic diagram of the interaction between various parameter features in the embodiment of the present invention, where Fig.13 (a) is the normal state; (b) is the cylinder head crack; (c) is the piston top burnout.

[0054] Attached Fig.14 Schematic diagram of the interaction between various parameter features in the embodiment of the present invention, where Fig.14 (a) shows cylinder liner wear; (b) shows piston ring wear; (c) shows piston ring adhesion.

[0055] Attached Fig.15 Schematic diagram of experimental results of the embodiment of the present invention, wherein Fig.15 (a) is the performance of KNN on the original data set; (b) is the performance of KNN on the filtered data set; (c) is the performance of SVM on the original data set.

[0056] Attached Fig.16 Schematic diagram of experimental results of the embodiment of the present invention, wherein Fig.16 (a) shows the performance of SVM on the filtered data set; (b) shows the performance of KNN on the original data set; (c) shows the performance of KNN on the filtered data set. DETAILED DESCRIPTION

[0057] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0058] Example:

[0059] The process of the explainable fault diagnosis method for marine diesel engines based on thermodynamics-assisted random forests proposed in this example is as follows: Figure 1As shown. First, a one-dimensional thermodynamic model was constructed to simulate the working process of the diesel engine, and the model was calibrated with the help of data collected by the ship's main engine data acquisition system (DCM) to ensure its accuracy and reliability. Then, by calibrating the key parameters of the system, five typical faults of the combustion chamber components were simulated, and the thermal parameters related to the faults were preliminarily output. Subsequently, random forests were used for fault pre-identification, the SHAP values ​​of each thermal parameter were calculated, and the most influential parameters were screened out. The screening results of the parameters will be explained in depth in combination with the one-dimensional thermodynamic model. Finally, the screened parameters are re-input into the random forest for training and verification. In addition, the present invention also compares random forests with traditional machine learning methods such as support vector machines (SVM) and K nearest neighbors (KNN), further verifying the superiority and adaptability of random forests in such fault diagnosis.

[0060] The one-dimensional thermodynamic model of a marine diesel engine in this example discusses a low-speed, six-cylinder, two-stroke marine diesel engine equipped with a turbocharger and an intercooler. The structure of the one-dimensional thermodynamic model of a diesel engine is as follows: Figure 2 As shown, it includes system boundaries (SB1, SB2), intake and exhaust manifolds (PL1, PL2), turbochargers (TC1), intercoolers (CO1) and six cylinders (C1-C6). In addition, six monitoring points (MP1 to MP6) are set up to monitor key engine parameters. In terms of pipelines, pipeline 1 represents the compressor intake pipe, pipeline 2 refers to the intercooler intake pipe, and pipeline 3 is the intercooler exhaust pipe. Pipes 4 to 9 correspond to the intake pipes of the cylinders, while pipelines 10 to 15 are the exhaust pipes of the cylinders. Pipe 16 is connected to the intake pipe of the turbocharger, and pipeline 17 is the exhaust pipe of the turbocharger.

[0061] This example uses a single Weber model to establish a combustion heat release model for a diesel engine. In the combustion heat release model, the single Weber model is an empirical model commonly used for combustion process modeling. The model has a simple structure and only requires a small number of parameters to describe the combustion process. In the fault simulation involving multi-parameter adjustment of the present invention, the single Weber model has higher computational efficiency. Through parameterized settings, the model can control variables such as the combustion start angle and heat release duration, and its expression is as follows:

[0062] (1),

[0063] Where: Q It is the heat energy released by the combustion of fuel mixture in one cycle, kJ; θ is the crankshaft angle, (°); θ 0 is the angle of the crankshaft when the fuel mixture in the cylinder starts to burn (°); a and bare the parameters of the Weber function, which determine the duration and shape of the combustion process. By integrating the Weber function, we can get the mass fraction of fuel consumed from the start of combustion to a certain time, that is, the mass fraction X :

[0064] (2),

[0065] The present invention uses the Woschni 1978 model to estimate the convection heat transfer coefficient of the combustion gas in the cylinder. The model comprehensively considers key factors such as turbulent flow and temperature in the cylinder, thereby improving the accuracy of heat conduction calculation. Its advantage is that it can accurately describe the heat transfer changes under different combustion chamber component failure conditions. The expression of the model is as follows:

[0066] (3),

[0067] Where: h is the heat transfer coefficient, W / m²K; p is the instantaneous pressure in the cylinder, MPa; T g is the instantaneous temperature of the gas, K; v is the characteristic velocity of the gas, m / s; C 1 is an empirical constant.

[0068] Characteristic speed v The calculation is based on the movement in the cylinder, and the calculation formula given by Woschni is as follows:

[0069] (4),

[0070] Where: V d is the displacement of the cylinder in m³; V is the cylinder volume during combustion, m³; p m is the average pressure in the cylinder, MPa; v p is the speed of the piston, m / s; C 2 and C 3 is an empirical constant, and its specific value depends on the engine design and operating conditions.

[0071] The six combustion chamber component health states established in this example are represented by codes F0 to F6: normal state (F0), cylinder head crack (F1), piston top ablation (F2), cylinder liner wear (F3), piston ring wear (F4), and piston ring adhesion (F5). The parameter calibration method for the five failure modes (F1~5) of the combustion chamber components is as follows:

[0072] Cylinder head crack (F1): Cylinder head crack fault is simulated by setting the surface temperature of the cylinder head. The cylinder head is subjected to high temperature and pressure during operation, especially the temperature near the combustion chamber is extremely high. When cracks appear, the structural strength of these areas is reduced, resulting in local thermal stress concentration. The cooling effect of the cylinder head will be affected by the presence of cylinder head cracks, resulting in local surface overheating.

[0073] Piston top ablation (F2): By setting the piston surface temperature and the blowby mass flow rate, the piston top ablation fault is simulated. As the piston top surface material is ablated, the heat distribution on the piston surface is uneven, and the local temperature of the ablation area increases significantly. In addition, the loss of surface material caused by ablation will affect the sealing between the piston and the cylinder wall, thereby causing blowby.

[0074] Cylinder liner wear (F3): By setting the cylinder diameter and blowby mass flow rate, the cylinder liner wear fault is simulated. Failure of the air filter or poor sealing can cause foreign matter to enter the combustion chamber, causing abnormal wear of the cylinder liner. The worn cylinder liner will cause roundness error, increase the local cylinder diameter, and affect the sealing performance of the cylinder, which will lead to the occurrence of blowby;

[0075] Piston ring wear (F4): Simulates cylinder head crack failure by setting the blowby mass flow rate. The piston ring may be worn due to factors such as impurities in the cylinder, resulting in the end face being out of round, the cylinder's sealing performance being reduced, and eventually causing blowby;

[0076] Piston ring sticking (F5): By setting the blowby mass flow rate, cylinder diameter and piston surface temperature, the piston ring sticking failure is simulated. When the engine has too much carbon deposits, insufficient lubrication or sludge accumulation, the piston ring movement may be restricted and lose its sealing function. Under normal operating conditions, the piston ring can transfer the heat from the top of the piston to the cylinder wall, thereby effectively dissipating heat. However, the sticking failure significantly weakens this heat transfer process, resulting in a significant increase in the piston surface temperature. In addition, the sticking destroys the normal sliding between the piston ring and the cylinder liner, and the friction in the local contact area increases sharply, further aggravating the wear of the cylinder liner.

[0077] This example builds a diesel engine thermal parameters screening process based on SHAP values, such as Figure 3As shown in the figure. First, the potentially relevant thermal parameters are screened out from the diesel engine fault simulation output to form the original data set. Then, the data set is input into the random forest for pre-identification. Based on the pre-identification results, the Tree SHAP method is used to calculate the SHAP value of each parameter, and the model pre-identification process is visualized. Then, the importance assessment results are interpreted by combining the parameter density scatter plot and feature interaction diagram. Finally, the key thermal parameters are screened out according to the importance assessment results, and the optimized data set is generated and re-input into the random forest model for further processing.

[0078] Compared with traditional feature selection methods (such as Relief, LVW, etc.), Tree SHAP can not only provide the global importance of thermal parameters, but also provide a deeper explanation from multiple perspectives, including the positive and negative effects of parameters, the interaction between parameters, and the distribution of sample points. This provides strong data support for combining thermodynamic models to analyze physical mechanisms and reveal the potential correlation between parameters.

[0079] The following is a detailed description of the calculation method of the thermal parameter SHAP value:

[0080] SHAP is a feature selection and model interpretability method widely used in the field of machine learning. In the calculation process of SHAP value, contribution value is used to describe the influence of thermal parameters on model output. i Marginal contribution after adding a certain parameter combination (like Figure 4 As shown), it can be expressed by the following formula:

[0081] (5), where: S Does not contain parameters i A collection of other parameters; f ( S ) is a set S The contribution value generated; f ( S { i}) is a parameter i Contribution value after joining.

[0082] After considering all possible parameter combinations, the parameters can be calculated i The average contribution increment, that is, the SHAP value φ i , as shown below:

[0083] (6), where: N is the set of all parameters; S is a subset of the parameters used in the model;p Representation Subset S The number of parameters in a parameter combination. S When performing an overall evaluation, the SHAP value of each parameter is added to get the evaluation result of the combination. Whenever a parameter is added to the parameter combination, the output is the original SHAP value plus the SHAP value of the newly added parameter, such as Figure 5 Parameter combination S The final evaluation result is the algebraic sum of all parameter SHAP values, that is:

[0084] (7), where: φ 0 is the baseline value, which means the average output of the model without any parameters; It is the sum of the SHAP values ​​predicted by the model for each feature pair.

[0085] This example uses SHAP interaction values ​​to further explore the potential relationship between thermal parameters: i and j The interaction value φ i,j Defined as:

[0086] (8) When i ≠ j and:

[0087] (9), by combining formulas (8-9), the SHAP interaction value between parameters can be calculated.

[0088] Tree SHAP is an efficient SHAP value calculation method specifically for tree structure models (such as random forests, etc.). Unlike directly calculating the SHAP value of each parameter, Tree SHAP uses the structural characteristics of the tree model to significantly improve the calculation efficiency: Figure 6 As shown in (a), directly calculating the SHAP value requires traversing all possible combinations of thermal parameters, which will lead to a huge amount of calculations, and the computational complexity increases exponentially as the number of features increases. In the tree model, Tree SHAP only follows the path that meets the current input conditions (i.e. Figure 6 (b) satisfies i 1>5) to calculate the marginal contribution, thus avoiding the enumeration of all parameter combinations and greatly improving the calculation efficiency.

[0089] For a parameter in a decision tree i , the SHAP value calculation formula based on Tree SHAP is as follows:

[0090] (10), where: T is the number of decision trees;P j For the j The set of paths of a tree; S is a subset of the features used in the model; p For subset S The number of features in .

[0091] This example uses random forest as the basic model for fault identification, and combines SHAP with the thermodynamic model to perform interpretability analysis on random forest. The calculation of SHAP value is based on the decision-making process of random forest. Considering that thermal parameters are continuous variables, this example discretizes these parameters to meet the input requirements of the decision tree model. Finally, by introducing a thermodynamic model, the SHAP analysis results are further physically verified to ensure the consistency of the model prediction results with the laws of thermodynamics, thereby enhancing the interpretability of random forest. The decision tree splitting method for thermal parameters is as follows: Decision tree, as the basis of the random forest algorithm, makes predictions by splitting the feature space. For a sample set D , assuming that k The proportion of class samples is p k , then the information entropy of D is defined as:

[0092] (11),

[0093] Assuming the use of discrete attributes For samples D Dividing it will produce V branch node. v In the branch node, the attribute a The upper value is a v The sample is recorded as D v , then we can calculate according to formula (9) D v Considering that different branch nodes contain different numbers of samples, weights are assigned to branch nodes. , you can calculate the attribute a For the sample set D The information gain obtained by partitioning is:

[0094] (12),

[0095] The larger the information gain, the more helpful the attribute is in distinguishing data categories and the more effective it is in reducing data uncertainty. Therefore, when a feature has a larger information gain, it is more likely to be selected as a split node in the decision tree.

[0096] For the thermal parameters of ship diesel engines (such as temperature, pressure, and speed), since the values ​​of these parameters are continuous, it is impossible to directly divide the nodes according to their values, so it is necessary to discretize the continuous attributes. D and thermal parameters P , assuming P exist D There is n Different values, arranged from small to large, are . Set a split point t , will be in the parameter P The upper value is not greater than t The sample is marked as D - t , and the value is greater than t The sample is marked as D + t Obviously, for P i and P i+1 As far as the value of t is concerned, as long as it falls within the interval Therefore, for thermal parameters P , the interval The median point As its candidate partition point, it contains the set of all candidate partition points T P It can be expressed as:

[0097] (13),

[0098] Therefore, the thermal parameters P For fault samples D The information gain obtained by partitioning is Gain( D , P ) can be expressed as:

[0099] (14), when Gain( D , P ) is relatively large, the decision tree will be more inclined to select thermal parameters P as a partition node.

[0100] Random forest is an ensemble learning algorithm that has been widely used in the field of fault diagnosis. Random forest improves the performance of the model by building multiple independent decision trees and combining their results.

[0101] For the classification results of random forests, SHAP can reveal how certain parameters affect decisions and provide support for the interpretability of the model. However, the parameter importance provided by SHAP relies on statistical data and does not consider the underlying physical mechanism. For diesel engine systems, this data-driven explanation lacks practical physical meaning. In addition, SHAP cannot reveal the causal relationship between parameters, but can only show the direction and degree of influence of parameters on model output, which has great limitations in explaining diesel engine combustion chamber failures.

[0102] In view of the above problems, the present invention introduces a thermodynamic model into the SHAP analysis results of random forest to improve the credibility of the model. The thermodynamic model provides a more in-depth explanation for fault diagnosis in the following three aspects: (1) Fault cause: By explaining the physical change mechanism of each parameter under a specific fault, the fault cause behind the statistical information reflected by the SHAP value is revealed. Since the essence of the SHAP value is to reflect the correlation between parameters rather than the causal relationship, the specific physical change mechanism of each parameter when the fault occurs can be further clarified through the auxiliary analysis of the thermodynamic model, thereby enhancing the explanatory depth of the model. (2) Important parameters: By analyzing the key driving parameters that affect the occurrence of the fault, the physical meaning of the important parameters is revealed by combining the SHAP value and the thermodynamic model. The SHAP value can identify the parameters that contribute more to the model prediction results, and the introduction of the thermodynamic model can deeply analyze the physical driving effect of these key parameters on the fault process, helping to understand its practical significance for fault diagnosis. (3) Parameter relationship: By revealing the interaction between parameters and their physical background, the interpretation of the model prediction results is further improved. The SHAP value can not only quantify the importance of a single feature, but also describe the interaction between features and their joint influence on the prediction results. With the help of thermodynamic models, we can deeply analyze the correlation between these parameters and their dynamic mechanisms. By analyzing the interaction of parameters, we can fully understand the dynamic nature of the fault process, thus providing a scientific basis for more accurate fault diagnosis.

[0103] In this example, during the simulation of the thermodynamic model, the test data is obtained as follows: the ship test data is collected by the ship host data acquisition system, and its structure is as follows: Figure 7As shown. The system collects the operating data of the main engine and its accessories from the ship's main engine control system, the fast signal unit and the main engine exhaust treatment device, obtains the operating data of the auxiliary machinery related to the main engine from the centralized monitoring system, and collects the ship's navigation status data from the navigation equipment. After being processed by the internal module, these data will be encrypted and transmitted to the shore-based receiving center via maritime satellite. The DCM system records the slow data of all internal and external equipment every 10 seconds, and stores it in the local database on the ship side. The average calculation is performed every 60 seconds and written to the database. The DCM system collects a total of 56,912 sample points under different operating conditions of ship diesel. The present invention selects 1,000 sample points under stable main engine speed as the test data set for model verification.

[0104] The main specification parameters of the diesel engine used in the present invention are shown in Table 1. After the diesel engine simulation model is constructed, it is necessary to input parameters for each component in the model. The accurate input of these component parameters is crucial to the performance of the whole machine because they directly affect the accuracy of the simulation calculation model. Table 2 shows the main input parameter settings of the diesel engine simulation model.

[0105] Table 1 Diesel engine specifications

[0106]

[0107] Table 2 Model setting details

[0108]

[0109] The model validation is as follows:

[0110] It can be seen from Table 3 that under rated conditions, the deviation between the main diesel engine parameters output by the diesel engine model constructed by the present invention and the test data is within 5%. Therefore, it is feasible to carry out fault simulation and diesel engine performance parameter analysis under fault mode based on this model.

[0111] Table 3 Comparison of simulation and test data

[0112]

[0113] Based on the existing research results, the present invention calibrates the relevant parameters under various faults to improve the reliability of the simulation. However, since there may still be a certain gap between the simulation results and the actual situation, the setting of these parameter ranges needs to be further optimized. In future research, the parameter settings in the thermal fault simulation process will be continuously adjusted and improved, and the range of fault parameters will be gradually calibrated to improve the accuracy and effectiveness of the fault simulation. Table 4 shows the details of the calibration of each fault parameter.

[0114] Table 4 Fault parameter calibration details

[0115]

[0116] The fault simulation results are as follows:

[0117] The present invention preliminarily selects 14 possibly relevant thermal parameters for monitoring, and the details of the thermal parameters are shown in Table 5; Figure 8 (a)-(n) show the changing trends of 14 thermal parameters under 6 healthy conditions, with the horizontal axis representing the crankshaft angle and the vertical axis representing the parameter amplitude. For ship monitoring systems, the cylinder pressure and temperature are the most intuitive key parameters that reflect the health status of the combustion chamber. However, under the five fault conditions studied, the changes in cylinder pressure are mainly concentrated in the intake and exhaust stages, and the amplitude is small (see Figure 8 In contrast, the cylinder temperature in all fault conditions is higher than that in normal conditions, such as Figure 8 As shown in (b) in the figure. For exhaust temperature and pressure (see Figure 8 In (c)-(d)), the changes caused by fault types F1 and F3 are particularly significant.

[0118] Changes in heat flow to combustion chamber components can often reflect component failure characteristics. Figure 8 As shown in (e), under the F2 fault state, the heat flux on the piston surface is significantly lower than that of other fault types, which is consistent with the phenomenon of ablation on the piston top. The heat flux on the cylinder head surface shows a significant change during the combustion chamber working stage, which is consistent with the result of cracks in the cylinder head (see Figure 8 In addition, different types of faults will weaken the airtightness and cooling effect to a certain extent, thus causing changes in the heat flow on the cylinder liner surface, such as Figure 8 As shown in (g).

[0119] Blow-by gas heat flow and blow-by gas mass flow rate (see Figure 8 (h)-(i)) are key thermal parameters that reflect the leakage behavior of the combustion chamber. In the fault setting, the blowby heat flow and blowby mass flow rate of fault types F0 and F1 are both zero, so no blowby occurs. For the other three fault types, the blowby degree is consistent with the fault setting.

[0120] Figure 8 (j)-(k) are the relevant thermal parameters of the turbocharger. Under different fault conditions, the intake pressure, intake temperature and exhaust temperature of the turbocharger have changed significantly. In addition, the turbocharger power and exhaust pressure are highly sensitive to fault type F3.

[0121] Table 5 Thermodynamic model output parameters

[0122]

[0123] Experimental results and analysis

[0124] There are many kinds of thermal parameters in diesel engines, and it is unrealistic to fully monitor all parameters for fault diagnosis. Therefore, it is necessary to effectively screen the parameters to find the key parameters with the most diagnostic value. The number of monitored parameters should be reduced as much as possible without reducing the ability to distinguish faults. This can not only avoid information redundancy and the complexity of data processing, but also effectively improve the efficiency and accuracy of fault diagnosis.

[0125] Fig. 9 (a) shows a stacked diagram of the contribution rate of each thermal parameter under different fault conditions, which intuitively reflects the importance of these parameters in the model. The horizontal axis represents the average absolute value of the SHAP value, and the vertical axis represents the 14 thermal parameters sorted from high to low according to importance. Fig. 9 (b) further presents the distribution of the importance scores of each parameter under different fault conditions. Based on the analysis results, the present invention selects the top 8 thermal parameters with the highest total contribution rate as the core parameters for fault identification, including: P14 (turbocharger exhaust temperature), P05 (cylinder liner surface heat flow), P06 (blow-by heat flow), P07 (blow-by mass flow rate), P11 (turbocharger intake pressure), P12 (turbocharger intake temperature), P03 (piston surface heat flow) and P04 (cylinder head surface heat flow).

[0126] Single sample evaluation process analysis: In the process of model importance evaluation, the present invention uses the waterfall plot in the SHAP technology for visualization. Fig.10 (a)-(f) show the identification process of a sample point in a normal healthy state, and evaluate the score of the sample point in five fault states. The waterfall chart can be used to visually observe the specific contribution (positive or negative) of each input feature to the prediction result. Among them, the red bar indicates that the feature positively promotes the prediction result, and the blue bar indicates that the feature negatively promotes the prediction result. In the normal state (see Fig.10 In (a), parameters P05 and P14 have a significant positive contribution to the prediction results, making the final prediction score reach 5.859. Compared with the evaluation results of this sample in other fault modes, this output value is significantly higher than the predicted value of other classification results, indicating that the model can successfully identify this sample as a normal state.

[0127] Importance analysis of thermal parameters:

[0128] This example summarizes the evaluation results of all sample points and displays them through a parameter density scatter plot (see Fig.11) gives the distribution changes of sample points caused by changes in thermal parameters. At the same time, the present invention combines the average SHAP value of each parameter on the basis of this figure, which can more intuitively feel the importance distribution of each parameter under different faults. Each row in the figure represents a parameter, and the horizontal axis represents the SHAP value. The parameters are sorted according to the average absolute SHAP value, and the most important parameters appear at the top. Each point corresponds to a sample, and the color of the sample point reflects the size of the parameter value. A SHAP value less than 0 indicates a negative impact on the prediction, while a value greater than 0 indicates a positive impact. The horizontal distribution of points reflects the size of the impact, and the wider the distribution, the greater the impact. The parameter density scatter plot can provide a detailed analysis of the importance of each thermal parameter under a specific fault, and show the positive and negative effects of each parameter. The distribution of parameter importance for different faults is interpreted as follows:

[0129] When the combustion chamber components are in normal condition (see Fig.11 In (a), high values ​​of P05 and P06, and low values ​​of P12 and P14, usually significantly increase the SHAP value. Under normal conditions, the blowby mass flow rate in the thermodynamic model is zero, that is, no blowby occurs in the combustion chamber. However, when blowby is observed, accompanied by an increase in the temperature of the supercharger inlet and exhaust ports, this indicates that there may be an abnormality in the combustion process.

[0130] Cylinder head crack failure (see Fig.11 In (b), the increase in the values ​​of P06 and P04 has a significant positive impact on the model output. This phenomenon can be attributed to the adjustment of the initial temperature by the thermodynamic model under the condition of cylinder head crack, which leads to an increase in the surface temperature of the cylinder head, thereby enhancing the system's sensitivity to parameters such as P04. At the same time, compared with other fault types, the blowby mass flow rate of cylinder head crack in the thermodynamic model is zero, so there is no obvious blowby phenomenon. This feature makes P06 an effective indicator for distinguishing cylinder head cracks.

[0131] Piston ablation failure (see Fig.11 In (c), P14, P03, and P07 are identified as key parameters, and the increase in their values ​​significantly increases the SHAP value. This phenomenon reflects the strong influence of piston ablation on the heat flow of the piston wall. Compared with other fault types, the thermodynamic model adjusts the initial piston wall temperature to a higher level for piston ablation, so the system shows a higher response sensitivity to the change of the P03 parameter.

[0132] In case of cylinder liner wear (see Fig.11In (d), P11 and P14 are identified as key influencing parameters. Cylinder liner wear leads to decreased air tightness in the cylinder, which reduces the overall engine performance. In order to maintain the output power, the turbocharger usually adopts over-boosting to compensate for the decrease in combustion efficiency, resulting in increased intake pressure. At the same time, compared with other faults, the low value of P14 has a higher specificity for distinguishing cylinder liner wear faults. In addition, in the thermodynamic fault model, the blowby mass flow rate under cylinder liner wear is not zero, so P06 and P07 have a dividing line near the zero point, which provides a certain basis for the identification of this fault.

[0133] In the case of piston ring wear, P12, P11 and P07 are identified as key influencing parameters. Due to the weakening of cylinder sealing caused by piston ring wear, the blowby phenomenon is aggravated. The over-boosting of the turbocharger causes a significant increase in intake temperature and pressure. In addition, the change of P07 reflects the blowby phenomenon caused by piston ring wear, thus becoming an important indicator for diagnosing this fault.

[0134] In the case of piston ring adhesion, the system shows higher sensitivity to the fluctuations of parameters P14, P05 and P07. This sensitivity is mainly due to the local high temperature phenomenon caused by piston ring adhesion. Local high temperature will cause the piston ring to adhere to the cylinder liner, further aggravating the wear of the cylinder liner and also providing conditions for the occurrence of blowby.

[0135] In summary, P14 is the most critical thermal parameter for distinguishing various types of faults. At the same time, P05 and P07 also play an important role in distinguishing fault types.

[0136] The interaction relationship analysis of thermal parameters is as follows: In order to further study the interaction between a certain thermal parameter and other parameters, the present invention further calculates the SHAP interaction value of each parameter. The relevant results are as follows: Fig.12 At the same time, the present invention focuses on analyzing the part with the highest interaction value in each fault category (i.e., the situation where the interaction between this parameter and other parameters has the most significant impact on the model output). Fig.13 and Fig.14 It provides detailed information on feature interactions, intuitively showing the impact trend of the target parameter on the model output and its interactive relationship with other parameters. In the figure, the X-axis represents the actual value of the parameter, the Y-axis corresponds to its SHAP value, each point represents a data sample, and the color of the point reflects the size of the interactive parameter value.

[0137] from Fig.12 and Fig.13As can be seen from (a)-(c) in the figure, there is an obvious correlation between parameters P03, P04 and P05. Among them, P03 is positively correlated with P04, while P05 is negatively correlated with both. This correlation can be attributed to the heat transfer between the combustion chamber components. Under the F0 fault mode, the SHAP value of P05 tends to be stable after its value exceeds -0.75 and no longer increases significantly. For P03 and P04, their SHAP values ​​show a stable upward trend as the absolute value of the parameters increases under the F1 and F2 fault modes. Fig.12 and Fig.14 (a) in Figure 2 shows that under fault F3, P14 and P12 have strong interaction, and when the value of P14 is close to 700, its SHAP value changes significantly. In contrast, parameter P11 shows high independence. Fig.12 and Fig.14 As can be seen in (b), there is no obvious correlation between P11 and other parameters. Fig.12 and Fig.14 The result in (c) shows that under fault F5, the interaction between P12 and P14 is still significant, which is consistent with the law of F3 fault mode.

[0138] Model performance comparison:

[0139] The present invention extracts samples from a complete working cycle of a diesel engine simulation model as experimental objects, wherein each health state contains 120 sets of data, totaling 720 sets of samples. Based on these samples, the present invention constructs an original data set and a screening data set. All samples are normalized and randomly divided into a training set (70%) and a test set (30%) in proportion. Due to the small number of samples, the present invention uses 5-fold cross validation to determine the recognition accuracy of the model, as shown in Table 6. In addition, the performance of the three models of SVM, KNN and RF on the two data sets is shown in Table 6. Fig.14 shown.

[0140] like Fig.15 As shown in (a)-(b), under the given hyperparameters, KNN performs poorly in the classification of the six health states. After parameter screening, its average accuracy slightly decreases. SVM performs better in classification performance and can better distinguish the first three fault states from the last three fault states. The accuracy remains basically stable on the screened data set, as shown in Figure 1. Fig.15 (c) and Fig.16 In contrast, RF has a strong ability to identify the six fault states, but it is slightly insufficient in identifying the F4 and F5 faults, as shown in (a). Fig.16As shown in (b)-(c). However, both faults are related to the piston ring and have a relatively concentrated impact on the system. In general, the RF based on the tree model shows greater advantages than the other two types of algorithms on the fault data set constructed by the present invention.

[0141] Table 6 Average model accuracy

[0142]

[0143] Aiming at the problems faced by pure data-driven and physical model fault diagnosis methods, the present invention studies a fault diagnosis model based on thermodynamics-assisted random forest. In terms of model simulation, the overall system output under five combustion chamber component faults is quickly evaluated. In terms of fault diagnosis, the present invention adopts the SHAP algorithm and deeply combines the thermodynamic model for auxiliary fault diagnosis. The proposed method aims to provide support for the fault diagnosis and result interpretability of the combustion chamber. The effectiveness of the constructed one-dimensional thermodynamic model has been verified by the actual operation data of the diesel engine. The test data comes from the ship main engine data acquisition system (DCM). The experimental results show that random forest has the best adaptability under the diesel engine simulation model constructed by the present invention. In addition, the present invention also studies the contribution rate of thermal parameters and parameter interaction under different faults, and interprets the selection of features and model decision results. The experimental results show that under the model established by the present invention, the exhaust temperature of the turbocharger, the heat flow of the blowby gas and the heat flow of the cylinder liner surface are the key thermal parameters affecting the fault diagnosis results.

Claims

1. A method for handling marine diesel engine faults based on thermodynamics-assisted random forest, characterized in that: The following steps are involved: Step 1: Construct a one-dimensional thermodynamic model to simulate the working process of the marine diesel engine, and calibrate the model with the data collected by the ship's main engine data acquisition system DCM; Step 2: By calibrating the key system parameters, five typical faults of the combustion chamber components of a marine diesel engine are simulated, and the thermal parameters related to the faults are preliminarily output; Step 3: Use random forest to pre-identify faults, calculate the SHAP value of each thermal parameter, and select the most influential parameters. The screening results of the parameters will be analyzed in combination with the one-dimensional thermodynamic model, including: Step 3-1: Screen out potentially relevant thermal parameters from the diesel engine fault simulation output to form an original data set; Step 3-2: Input the data set into the random forest for pre-identification. Based on the pre-identification results, the Tree SHAP method is used to calculate the SHAP value of each parameter, and the model pre-identification process is visualized; Step 3-3: Combine the parameter density scatter plot and feature interaction plot to perform an explanatory analysis of the importance assessment results; Step 3-4: Screen out key thermal parameters based on the importance assessment results, generate an optimized data set, and re-input the optimized data set into the random forest model for further processing; Step 4: The selected parameters are re-input into the random forest for training and validation; The marine diesel engine involved in the one-dimensional thermodynamic model constructed in step 1 is a low-speed, six-cylinder, two-stroke marine diesel engine equipped with a turbocharger and an intercooler. The one-dimensional thermodynamic model includes the system boundary, the intake and exhaust manifolds, the turbocharger TC1, the intercooler CO1, and the six cylinders C1 to C6. Six monitoring points are also set up, which are respectively recorded as MP1 to MP6 for monitoring key engine parameters. Pipeline 1 represents the compressor intake pipe, pipe 2 refers to the intercooler intake pipe, pipe 3 is the intercooler exhaust pipe, pipes 4 to 9 correspond to the cylinder intake pipe, and pipes 10 to 15 are the cylinder exhaust pipes. Pipeline 16 is connected to the turbocharger intake pipe, and pipe 17 is the turbocharger exhaust pipe. In step 1, the single Weber model is used to establish the combustion heat release model of the diesel engine, and the expression is as follows: Where: Q is the heat energy released by the combustion of the fuel mixture in one cycle, kJ; θ is the crankshaft angle, (°); θ0 is the angle of the crankshaft when the fuel mixture in the cylinder starts to burn (°); a and b are the parameters of the Weber function, which determine the duration and shape of the combustion process. By integrating the Weber function, we can get the mass fraction of fuel consumed from the start of combustion to a certain time, that is, the mass fraction X: The Woschni1978 model is also used to estimate the convection heat transfer coefficient of the combustion gas in the cylinder, and the expression is as follows: Where: h is the heat transfer coefficient, W / m 2 K; p is the instantaneous pressure in the cylinder, MPa; T g is the instantaneous temperature of the gas, K; v is the characteristic velocity of the gas, m / s; C1 is an empirical constant. The calculation of the characteristic velocity v is based on the movement in the cylinder. The calculation formula given by Woschni is as follows: Where: V d is the displacement of the cylinder in m 3 ; V is the cylinder volume during combustion, m 3 ;p m is the average pressure in the cylinder, MPa; v p is the speed of the piston, m / s; C2 and C3 are empirical constants, the specific values ​​of which depend on the design and operating conditions of the engine; In step 2, five typical faults of the combustion chamber components of a marine diesel engine are simulated in the following way: there are six health states of the combustion chamber components, represented by codes F0 to F6: normal state (F0), cylinder head crack (F1), piston top ablation (F2), cylinder liner wear (F3), piston ring wear (F4), piston ring adhesion (F5), and the parameter calibration method of the five fault modes (F1-5) is: Cylinder head crack (F1): The cylinder head crack failure is simulated by setting the surface temperature of the cylinder head. The cylinder head is subjected to high temperature and high pressure during operation, especially the temperature near the combustion chamber is extremely high. When cracks appear, the structural strength of these areas is reduced, resulting in increased local thermal stress concentration. The cooling effect of the cylinder head will be affected by the presence of the cylinder head crack, resulting in local surface overheating; Piston top ablation (F2): The piston top ablation failure is simulated by setting the piston surface temperature and the blowby gas mass flow rate. As the piston top surface material is ablated, the heat distribution on the piston surface is uneven, and the local temperature of the ablation area increases significantly. In addition, the loss of surface material caused by ablation will affect the sealing between the piston and the cylinder wall, thereby causing blowby; Cylinder liner wear (F3): The cylinder liner wear failure is simulated by setting the cylinder diameter size and the blowby gas mass flow rate. Failure of the air filter or poor sealing can cause foreign matter to enter the combustion chamber, thereby causing abnormal wear of the cylinder liner. ; The worn cylinder liner will cause roundness error, increase the local cylinder diameter, and affect the sealing performance of the cylinder, which will lead to the occurrence of blowby; Piston ring wear (F4): Set the blowby mass flow rate to simulate the cylinder head crack failure. The piston ring may be worn due to factors such as impurities in the cylinder, resulting in the end face being out of round, the cylinder sealing performance is reduced, and finally causing blowby; Piston ring adhesion (F5): By setting the blowby mass flow rate, cylinder diameter size and piston surface temperature, the piston ring adhesion failure is simulated. When the engine has too much carbon deposits, insufficient lubrication or sludge accumulation, the piston ring movement may be restricted and lose its sealing function. Under normal working conditions, the piston ring can transfer the heat from the top of the piston to the cylinder wall, thereby effectively dissipating heat; however, adhesion failure significantly weakens this heat conduction process, resulting in a significant increase in the piston surface temperature. In addition, adhesion destroys the normal sliding between the piston ring and the cylinder liner, and the friction force in the local contact area increases sharply, further aggravating the wear of the cylinder liner; In step 3-1, in the calculation process of SHAP value, contribution value is used to describe the influence of thermal parameters on model output. The marginal contribution Δ of thermal parameter i after adding it to a certain parameter combination is val (i,S) is expressed by the following formula: Δ val (i,S)=f(S∪{i})-f(S)(5), where: S is the set of other parameters that does not contain parameter i; f(S) is the contribution value generated by set S; f(S∪{i}) is the contribution value after parameter i is added. After considering all possible parameter combinations, the average contribution value increment of parameter i, that is, SHAP value φi, is calculated as follows: Where: N is the set of all parameters; S is the subset of parameters used in the model; p represents the number of parameters in the subset S; when evaluating a parameter combination S as a whole, the evaluation result of the combination is obtained by adding the SHAP value of each parameter. Whenever a parameter is added to the parameter combination, the output is the original SHAP value plus the SHAP value of the newly added parameter. The final evaluation result of the parameter combination S is the algebraic sum of the SHAP values ​​of all parameters, that is: Where: φ0 is the reference value, which represents the average output of the model without any parameters; The SHAP value is the sum of the SHAP values ​​predicted by each feature pair model. The SHAP interaction value is used to further explore the potential relationship between thermal parameters. The interaction value φ between parameters i and j i,j Defined as: Δ val (S,i,j)=f(S∪{i,j})-f(S∪{i})-f(S∪{j})+f(S)(9), by combining formulas (8-9), we can calculate the SHAP interaction value between parameters; In step 3-2, Tree SHAP only calculates the marginal contribution along the path that meets the current input conditions, thereby avoiding enumerating all parameter combinations and greatly improving the calculation efficiency. For a parameter i in a decision tree, the SHAP value calculation formula based on Tree SHAP is as follows: Where: T is the number of decision trees; P j is the path set of the jth tree; S is the subset of features used in the model; p is the number of features in subset S; In step 3-2, random forest is used as the basic model for fault identification, and SHAP and thermodynamic model are combined to perform interpretability analysis on random forest. The calculation of SHAP value is based on the decision process of random forest. Considering that thermal parameters are continuous variables, the parameters need to be discretized to meet the input requirements of decision tree model. The decision tree splitting method for thermal parameters is as follows: Decision tree is the basis of random forest algorithm, which makes predictions by dividing feature space. For a sample set D, assuming that the proportion of samples in the kth category is p k (k=1,2,…,|y|), then the information entropy of D is defined as: Assume that we use discrete attribute a = {a 1 ,a 2 ,…,a n } Dividing the sample D will generate V branch nodes, and the value of attribute a in the vth branch node is a v The sample is denoted as D v , then according to formula (9) D is calculated v The information entropy of , considering that different branch nodes contain different numbers of samples, gives weights to branch nodes | D v | / |D|, we can calculate the information gain of attribute a dividing the sample set D: The larger the information gain, the more helpful the attribute is in distinguishing data categories and can more effectively reduce the uncertainty of data. Therefore, when the information gain of a feature is large, it is more likely to be selected as a split node in the decision tree. For the thermal parameters of ship diesel engines, continuous attributes need to be discretized: Given a fault sample D and a thermal parameter P, assume that P has n different values ​​on D, arranged from small to large as {P 1 ,P 2 ,…,P n }, set a partition point t, and mark the samples whose values ​​on parameter P are not greater than t as D - t , and samples with values ​​greater than t are marked as D + t , for P i With P i+1 As far as the value of t is concerned, as long as it falls within the interval [P i ,P i+1 ], the division results are consistent. Therefore, for the thermal parameter P, the interval [P i ,P i+1 The median point (P i +P i+1 ) / 2 as its candidate partition point, then the set T containing all candidate partition points P It can be expressed as: Therefore, the information gain Gain(D,P) obtained by dividing the fault sample D by the thermal parameter P can be expressed as: When Gain(D,P) is relatively large, the decision tree will be more inclined to select thermal parameter P as the partition node; In step 3-3, the parameter density scatter plot and feature interaction diagram are combined to perform an explanatory analysis of the importance assessment results. The parameter density scatter plot is used to give the distribution changes of sample points caused by changes in thermal parameters. Each row in the parameter density scatter plot represents a parameter, and the horizontal axis represents the SHAP value. The parameters are sorted according to the average absolute SHAP value, and the most important parameters appear at the top. Each point corresponds to a sample, and the color of the sample point reflects the size of the parameter value. A SHAP value less than 0 indicates a negative impact on the prediction, while a SHAP value greater than 0 indicates a positive impact. The horizontal distribution of the points reflects the size of the impact. The wider the distribution, the greater the impact. The feature interaction diagram intuitively shows the impact trend of the target parameter on the model output and its interactive relationship with other parameters. In the figure, the X-axis represents the actual value of the parameter, and the Y-axis corresponds to its SHAP value. Each point represents a data sample, and the color of the point reflects the size of the interaction parameter value.

Images