Diesel engine fault diagnosis method and device based on autonomous fidelity digital twin model, and storage medium

Abstract

The invention discloses a diesel engine fault diagnosis method and device based on an autonomous fidelity digital twinborn model and a storage medium, and belongs to the technical field of digital twinborn technologies and fault diagnosis. Firstly, a digital twin model of the diesel engine is constructed; secondly, updating parameters of the numerical simulation model in real time by using a chaos adaptive whale swarm optimization algorithm in combination with a space-time composite fidelity evaluation index; then, according to the fault mechanism of the diesel engine, setting a fault mode of the diesel engine in the numerical simulation model so as to generate a complete simulation fault sample; and finally, training a time sequence classification network of a multi-stage mixed structure by using the simulation fault samples, so that unknown measurement fault sample types can be identified. According to the method, the problem that a traditional intelligent diesel engine fault diagnosis method fails due to lack of an actual measurement fault sample is solved, and the fault type of an unknown measurement fault sample can be accurately recognized under the condition that the actual measurement fault sample is lacked.

Description

Technical Field

[0001] The present invention belongs to the field of digital twin technology and fault diagnosis technology, and specifically relates to a diesel engine fault diagnosis method, equipment and storage medium based on an autonomous fidelity digital twin model. Background Art

[0002] Marine diesel engines, as the "heart" of a ship, provide critical power support for navigation, and their operating conditions are crucial to the entire ship. The reliable operation of diesel engines not only affects the operational efficiency and safety of the ship, but also directly determines the successful execution of various maritime missions. By using historical data to construct a diesel engine fault diagnosis model offline, and combining it with real-time monitoring data for online fault detection, potential faults can be discovered in a timely manner. However, due to the special operating environment of marine diesel engines and the high requirements for safety, it is usually difficult to obtain sufficient measured historical data to meet the needs of model training. The scarcity of fault samples limits the application of diagnostic methods based on artificial intelligence (AI) in actual engineering.

[0003] Therefore, it is crucial to build an autonomous and fidelity digital twin model of diesel engines, which can provide the necessary training samples for the AI ​​model and ensure the high accuracy and robustness of the AI ​​model in actual marine diesel engine fault diagnosis. Researching and developing an autonomous and fidelity digital twin model for marine diesel engines not only has theoretical value, but also has significant practical application significance. By building such a model, the dynamic response of the diesel engine under various fault conditions can be predicted, thereby providing accurate training samples for AI-based fault diagnosis methods. This enables the AI ​​model to predict unknown fault types at an early stage, avoiding safety accidents or failures in mission execution caused by diesel engine failures, thereby improving the safety and efficiency of ship operations. Therefore, it is extremely important to study marine diesel engine fault diagnosis methods driven by autonomous and fidelity digital twin models. Summary of the Invention

[0004] This paper proposes a marine diesel engine fault diagnosis method driven by an autonomous, high-fidelity digital twin model. The core of this method is to construct a self-calibrating digital twin model of the diesel engine, which is then used to generate simulated fault samples. These samples are used to train a probabilistic neural network, enabling it to identify fault types in actual, unknown measurement signals. This system can proactively identify early diesel engine faults, thereby preventing accidents and ensuring the safety of the ship.

[0005] The present invention provides a marine diesel engine fault diagnosis method driven by an autonomous fidelity digital twin model, comprising:

[0006] Step 1: Based on the basic structure and relevant parameters of the marine diesel engine, a one-dimensional simulation model of the diesel engine is constructed;

[0007] Step 2: A chaotic adaptive whale swarm optimization algorithm is introduced based on a one-dimensional simulation model of a diesel engine to optimize the one-dimensional simulation model for model calibration. The chaotic adaptive whale swarm optimization algorithm includes a search and hunting phase. The search and hunting phase introduces an adaptive step size to search the solution space over a large range, and a hunting strategy is collaboratively selected based on a probability mechanism and a coefficient vector. The spatiotemporal composite fidelity evaluation index between the measured data and the simulation data is used as an objective function to update the cylinder combustion model parameters, intake and exhaust system parameters, turbocharger model parameters, and intercooler model parameters of the one-dimensional simulation model.

[0008] Step 3: By adjusting the relevant parameters in the input simulation model, various fault states of the diesel engine are simulated, and the dynamic response characteristics under various fault states are obtained through simulation calculation to generate simulated fault samples;

[0009] Acquire simulated fault samples to construct a data set, preprocess each simulated fault data in the data set; construct a training set with the preprocessed simulated fault data, where each sample in the training set includes simulated fault data and a corresponding fault category; the fault category is obtained based on the dynamic response characteristics in the simulated fault sample data;

[0010] Step 4: Use the training set to train the time series classification network with a multi-level hybrid structure;

[0011] Step 5: Input the actual fault data to be tested into the pre-trained multi-level hybrid structure time series classification network to output the fault mode of the diesel engine.

[0012] Furthermore, the basic structure of the marine diesel engine includes: a cylinder model, an intake and exhaust pipe model, an intake and exhaust valve model, a supercharger, and an intercooler; the relevant parameters include operating parameters, combustion parameters, valve timing parameters, and supercharging system parameters.

[0013] Furthermore, the step 2 specifically includes the following steps:

[0014] Step 2.1: Initialization phase; using chaos mapping to optimize the population initialization in the chaotic adaptive whale swarm optimization algorithm

[0015] Step 2.2: Optimal position storage stage; calculate the spatiotemporal composite fidelity evaluation index, select m minimum spatiotemporal composite fidelity evaluation indexes, and store the current global optimal position w * (t) and the positions of other excellent individuals are stored in the elite memory bank for subsequent calls to prevent falling into local optimality.

[0016] Step 2.3: Search and hunting phase; introduce adaptive step size α(t) and coefficient vector A(t) to dynamically adjust the search range. When hunting, the adaptive coefficient A(t) and random probability P∈[0,1] are combined to coordinate and determine the predation strategy.

[0017] Step 2.4: If the value of the spatiotemporal composite fidelity evaluation index does not change when the number of iterations exceeds 100, the optimization of the default model is converged or the number of iterations exceeds the maximum number of iterations M, and the optimal position w is output. * (t); otherwise, return to step 2.2.

[0018] Furthermore, the initialization phase of step 2.1 includes the following steps:

[0019] Step 2.1.1: Set the population size N, the maximum number of iterations T, and the parameter search space [w min,d , w max,d ], adaptive step range [α min ,α max ];

[0020] Step 2.1.2: Chaos initialization position: Use Logistic chaotic mapping to generate chaotic sequence {x n}, then the generated chaotic sequence {x n} is mapped to each parameter dimension search space [w min,d , w max,d ], initialize the whale's individual position w i,d ;

[0021] x n+1 =μx n (1-x n )

[0022] w i,d =w min,d +x n (w max,d -w min,d )

[0023] Among them, x n ∈(0,1); μ is a parameter, set to 4; w i,d is the initial position of the i-th whale in the d-th dimension; w max,d and w min,d are the upper and lower bounds of the search space in the d-th dimension.

[0024] Furthermore, the search and hunting phase in step 2.3 is specifically as follows:

[0025] The adaptive step size α(t):

[0026]

[0027] The coefficient vector A(t):

[0028] A(t)=2α(t)r-α(t)

[0029] Among them, t is the current iteration number; r is a random vector in the range [0,1]; α max and α min are the initial maximum and minimum values, respectively;

[0030] The adaptive coefficient A(t) and the random probability P select the hunting strategy for the current iteration:

[0031] If the random constant P<0.5, the bubble net attack strategy is adopted; based on the current optimal position w * (t), the individual whale performs local search along the spiral trajectory:

[0032] w(t+1)=|w * (t)-w(t)|·ec ·n cos(2πn)+w * (t)

[0033] Where w(t) is the current individual position; c is the spiral amplitude parameter, and n is a random number in the range of [-1,1].

[0034] If the random constant P ≥ 0.5, adopt the strategy of surrounding the prey and determine whether |A(t)| is less than 1;

[0035] If |A(t)|<1, adopt the local search strategy of surrounding the prey; at the current best position w * (t), the position is updated by the adaptive coefficient A(t):

[0036] w(t+1)=w * (t)-A(t)|C(t)·w * (t)-w(t)|

[0037] Where C(t) is the random perturbation vector;

[0038] If |A(t)|>1, adopt the global search strategy of surrounding the prey; take the current best individual position w * (t) is used as the basis, and the prey search strategy is used to update the individual whale position w(t+1)

[0039] w(t+1)=w * (t)-g·A(t)|u·w * (t)-w(t)|

[0040] Among them, g is a linearly decreasing parameter, and u is a random vector in [0,1];

[0041] According to one of the bubble net attack strategy, the local search strategy for surrounding prey, and the global search strategy for surrounding prey, the updated position w(t+1) of the whale individual is obtained and the spatiotemporal composite fidelity evaluation index of the whale individual is calculated. If there is a spatiotemporal composite fidelity evaluation index of the updated position w(t+1) of the whale individual that is less than the global optimal solution w * (t), then update the global optimal solution w * (t).

[0042] Furthermore, the spatiotemporal composite fidelity evaluation index is a weighted superposition of the time dimension comprehensive error and the space dimension comprehensive error through weight coefficients w1 and w2;

[0043] F total =w1·F time +w2·F space

[0044] Among them, F total is the spatiotemporal composite fidelity evaluation index; w1 and w2 are the weight coefficients of temporal and spatial errors in the overall evaluation respectively; F time is the comprehensive error in time dimension; F space is the comprehensive error of spatial dimension.

[0045] Furthermore, the evaluation index of the time dimension selects the key variables of speed, torque, fuel injection amount, and boost pressure that change with time. The time dimension error is calculated as follows:

[0046] (1) Root mean square error RMSE:

[0047]

[0048] Among them, y sim (t i ) is the simulation signal at time point t i The value of y meas (t i ) is the measured signal at time point t i The value of , N is the total number of sampling points;

[0049] (2) Waveform correlation coefficient ρ(X,Y):

[0050]

[0051] Among them, x i =y sim (t i ), y i =y meas (t i ), and are the average values ​​of the simulated signal and the measured signal respectively; the closer the correlation coefficient ρ is to 1, the more similar the waveforms are.

[0052] (3) Mean deviation MD:

[0053]

[0054] (4) Time dimension comprehensive error F time :

[0055] F time =α1·RMSE+α2·(1-ρ)+α3·MD

[0056] Among them, α1, α2, and α3 are the weight coefficients of the root mean square error, waveform correlation coefficient, and average deviation, respectively.

[0057] Furthermore, the evaluation indicators of the spatial dimension are selected from the cylinder partition error, the supercharger performance mapping deviation, and the pipeline segmentation error variables. The spatial dimension error is calculated as follows:

[0058] (1) Cylinder partition error E cyl :

[0059]

[0060] Among them, p sim (r i ) and p meas (r i ) represent the spatial coordinates r i The simulated and measured pressure values ​​at , M is the total number of spatial sampling points;

[0061] (2) Turbocharger performance mapping deviation E tc :

[0062]

[0063] Among them, η sim (q k ) and η meas (q k ) represent the flow rate q k The simulated and measured supercharger efficiency at , K is the total number of flow sampling points;

[0064] (3) Pipeline segmentation error E pipe :

[0065]

[0066] Where ΔP sim (l) and ΔP meas(l) represents the simulated and measured pressure losses in the lth section of the pipeline, respectively, where L is the number of pipeline sections;

[0067] (4) Time dimension comprehensive error F space :

[0068] F space =β1·E cyl +β2·E tc +β3·E pipe

[0069] Among them, β1, β2, and β3 are the weight coefficients of cylinder partition error, supercharger performance mapping deviation, and pipeline segmentation error, respectively.

[0070] Furthermore, the multi-stage hybrid structure time series classification network includes a multi-scale convolutional feature extraction layer, a Transformer encoding module, a gated recurrent unit integration unit, and a fusion output layer;

[0071] First, the original time series in the simulated fault dataset is input into the multi-scale convolutional layer to extract local patterns at different time scales and fuse them in the channel dimension. Then, the fused feature sequence is sent to the Transformer encoding module to capture global dependencies and key feature interactions through the self-attention mechanism. The output of the Transformer encoding module is further input into the gated recurrent unit for temporal memory and integration of accumulated information. Finally, combined with the fault category label information of the training set, the probability prediction of the fault category is completed by fusing the fully connected mapping and Softmax function of the output layer. Iterative optimization is performed with the goal of minimizing the cross-entropy loss, and the collaborative representation of the features of each layer of the network is converted into the final classification result.

[0072] The present invention also provides a computer device / equipment / system, comprising a memory, a processor, and a computer program stored in the memory. When the processor executes the computer program, it implements the steps of any of the above-mentioned diesel engine fault diagnosis methods based on the digital twin model.

[0073] The present invention also provides a computer-readable storage medium having a computer program / instruction stored thereon, which, when executed by a processor, implements the steps of any of the above-mentioned diesel engine fault diagnosis methods based on a digital twin model.

[0074] The present invention also provides a computer program product, comprising a computer program / instruction, which, when executed by a processor, implements the steps of any of the above-mentioned diesel engine fault diagnosis methods based on a digital twin model.

[0075] The beneficial effects of the present invention are:

[0076] The present invention addresses the problem of unsatisfactory fault diagnosis results due to the lack of actual measured fault samples during actual marine diesel engine fault diagnosis. A method is provided for using a digital twin model of a diesel engine with autonomous fidelity to generate simulated fault samples to replace actual measured fault samples. The chaotic adaptive whale swarm optimization algorithm is combined with a spatiotemporal composite fidelity evaluation index to update the numerical simulation model parameters in real time to ensure the high fidelity of the model. This method effectively solves the problem of failure of intelligent fault diagnosis methods in actual engineering applications due to missing samples, and realizes the fault type identification of unknown measured fault samples of marine diesel engines. The method of the present invention is a new tool and method for fault diagnosis of marine diesel engines, which is expected to promote the development of related technologies and has important academic and engineering value. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 This is a one-dimensional simulation model diagram of a diesel engine according to the present invention;

[0078] Figure 2 This is the calculation flow chart of the chaotic adaptive whale optimization algorithm of the present invention;

[0079] Figure 3 This is a working process diagram of the diesel engine of the present invention;

[0080] Figure 4 This is a time series classification network structure diagram of the multi-level hybrid structure of the present invention;

[0081] Figure 5 This is a flow chart of the marine diesel engine fault diagnosis method driven by the autonomous fidelity digital twin model of the present invention. DETAILED DESCRIPTION

[0082] The present invention will be further described below with reference to the accompanying drawings.

[0083] This paper proposes a marine diesel engine fault diagnosis method driven by an autonomous, high-fidelity digital twin model. The core of this method is to construct a self-calibrating digital twin model of the diesel engine, which is then used to generate simulated fault samples. These samples are used to train a probabilistic neural network, enabling it to identify fault types in actual, unknown measurement signals. This system can proactively identify early diesel engine faults, thereby preventing accidents and ensuring the safety of the ship.

[0084] The present invention discloses a marine diesel engine fault diagnosis method driven by an autonomous fidelity digital twin model, comprising:

[0085] Step 1: Based on the basic structure and relevant parameters of a certain diesel engine, a one-dimensional simulation model of the diesel engine was constructed in AVL-BOOST software. The basic structure includes the cylinder model, intake and exhaust pipe model, intake and exhaust valve model, turbocharger, and intercooler. The cylinder model includes: cylinder diameter, stroke, compression ratio, and connecting rod length; the intake and exhaust pipe model includes: pipe diameter, pipe length, roughness, and wall temperature; the intake and exhaust valve model includes: valve diameter, lift curve, and valve opening and closing angle; the turbocharger includes: turbine efficiency, compressor efficiency, and compression ratio; the intercooler includes: heat exchange efficiency and outlet air temperature; relevant parameters include operating parameters, combustion parameters, valve timing parameters, and supercharging system parameters; operating parameters include: rated speed, load, and intake and exhaust conditions; combustion parameters include: injection pattern, injection advance angle, and fuel characteristics; valve timing parameters include: valve lift curve and valve timing phase; and supercharging system parameters include: supercharging efficiency and bypass valve setting parameters.

[0086] Step 2: After constructing the one-dimensional diesel engine simulation model, the next key step is to configure the various components within the model in detail. This includes global parameters, turbocharger characteristics, cylinder status, and intercooler performance. To ensure that the simulation model accurately reflects the actual operating conditions of the diesel engine, the initial parameters of each component must be precisely set, and the system boundary conditions must be set to simulate the actual operating environment of the diesel engine. Furthermore, to ensure the self-correcting capabilities of the simulation model, a chaotic adaptive whale swarm optimization algorithm was introduced. This algorithm uses the spatiotemporal hybrid evaluation index error between the measured and simulated data as a fidelity assessment metric and utilizes an optimization algorithm to update the parameters of the diesel engine numerical simulation model, thereby maintaining the high fidelity of the diesel engine digital simulation model.

[0087] Among them, the chaotic adaptive whale swarm optimization algorithm introduces adaptive step size α(t) and search coefficient vector A(t) to search the solution space over a large range on the basis of retaining the original search and hunting stages; and provides three hunting strategies including bubble net attack strategy, local search strategy of surrounding prey, and global search strategy of surrounding prey; when hunting, it adaptively and effectively selects the predation strategy according to the probability mechanism and coefficient vector, thereby expanding the search space and improving the overall convergence efficiency.

[0088] The optimization algorithm updates the parameters of the diesel engine numerical simulation model, including cylinder combustion model parameters, intake and exhaust system parameters, turbocharger model parameters, and intercooler model parameters; cylinder combustion model parameters: combustion heat release rate, combustion duration, and injection advance angle; intake and exhaust system parameters: pipe wall heat transfer coefficient, intake and exhaust pipe pressure drop, and pipe roughness; turbocharger model parameters: inertia parameters of the turbine and compressor, efficiency curve, and response delay; intercooler model parameters: heat transfer characteristics and dynamic temperature response characteristics.

[0089] Step 1: Initialization phase

[0090] Step 1.1: Set algorithm parameters

[0091] Set the population size N, the maximum number of iterations T, and the parameter search space [w min,d ,w max,d ], adaptive step range [α min ,α max ].

[0092] Step 1.2: Chaos initialization position

[0093] Use Logistic chaotic map to generate chaotic sequence {x n}:

[0094] x n+1 =μx n (1-x n ),μ=4,x n ∈(0,1)

[0095] The chaotic sequence {x n} is mapped to the search space of each parameter dimension and the individual whale position w is initialized i,d :

[0096] w i,d =w min,d +x n (w max,d -w min,d )

[0097] Among them, w i,d is the initial position of the i-th whale in the d-th dimension, w max,d and w min,d They represent the upper and lower bounds of the search space of the d-th dimension respectively.

[0098] Step 2: Fitness calculation and optimal solution recording

[0099] Step 2.1: Calculate the spatiotemporal composite fidelity evaluation index and determine the current global optimal position w * (t).

[0100] Step 2.2: Store the current optimal position and some excellent individuals in the elite memory bank for subsequent use to prevent falling into local optimality.

[0101] Step 3: Search and Hunt Phase

[0102] Step 3.1: Adaptive step size control

[0103] Adaptively calculate the step size based on the current number of iterations t

[0104]

[0105] Calculate the search coefficient vector A(t) using step size α(t)

[0106] A(t)=2α(t)r-α(t)

[0107] Where r is a random vector in the range [0,1].

[0108] Step 3.2: Hunting strategy selection and position update

[0109] Select the hunting strategy for the current iteration according to the random probability P∈[0,1]:

[0110] If P<0.5, the bubble net attack strategy is adopted. Based on the current optimal position w * (t), the individual whale performs local search along the spiral trajectory:

[0111] w(t+1)=|w * (t)-w(t)|·e c·n cos(2πn)+w * (t)

[0112] Where c is the spiral amplitude parameter and n is a random number in the range of [-1,1].

[0113] If P≥0.5, adopt the strategy of surrounding the prey.

[0114] If yes, then the strategy of surrounding the prey (local search) is adopted. Near the current best position w*(t), the position is updated by the adaptive coefficient A(t):

[0115] w(t+1)=w * (t)-A(t)|C(t)·w * (t)-w(t)|

[0116] Where C(t) is the random perturbation vector.

[0117] If not, a global search strategy is adopted. Take the current best individual position w * (t) is used as the basis, and the prey search strategy is used to update the individual whale position w(t+1)

[0118] w(t+1)=w * (t)-g·A(t)|u·w * (t)-w(t)|

[0119] Among them, w(t) is the current individual position, g is a linearly decreasing parameter, and u is a random vector in [0,1].

[0120] Step 4: Update the optimal solution

[0121] Calculate the spatiotemporal composite fidelity evaluation index of each individual after the position update; if the new individual spatiotemporal composite fidelity evaluation index is less than the global optimal solution, that is, the individual is better than the currently recorded optimal solution, then update the global optimal solution w * (t).

[0122] Step 5: Terminate the judgment

[0123] If the value of the optimal result does not change after more than 100 iterations, the optimization of the default model is converged or the number of iterations exceeds the maximum number of iterations M, the algorithm terminates and outputs the optimal position w * (t).

[0124] In the diesel engine parameter update process, the chaotic adaptive whale swarm optimization algorithm uses the spatiotemporal composite evaluation index objective function between the simulation data generated by the diesel engine simulation model and the actual measurement data. The spatiotemporal objective function evaluation index is calculated as follows:

[0125] The evaluation index of the time dimension selects the key variables of speed, torque, fuel injection amount, and boost pressure that change over time, and calculates the error using the following formula:

[0126] (1) Root mean square error:

[0127]

[0128] Among them, y sim (t i ) is the simulation signal at time point t i The value of y meas (t i ) is the measured signal at time point t i The value of , N is the total number of sampling points.

[0129] (2) Waveform correlation coefficient

[0130]

[0131] Among them, x i =y sim (t i ), y i =y meas (t i ), and are the average values ​​of the simulated signal and the measured signal, respectively. The closer the correlation coefficient ρ is to 1, the more similar the waveforms are.

[0132] (3) Average deviation

[0133]

[0134] The above time dimension errors are weighted and superimposed by weight coefficients α1, α2, and α3 to obtain the time dimension comprehensive error F time :

[0135] F time =α1·RMSE+α2·(1-ρ)+α3·MD

[0136] The evaluation indicators of the spatial dimension are selected as cylinder partition error, supercharger performance mapping deviation, and pipeline segmentation error variables, and the error is calculated using the following formula:

[0137] (4) Cylinder partition error:

[0138]

[0139] Among them, p sim (r i ) and p meas (r i ) represent the spatial coordinates r i The simulated and measured pressure values ​​at , M is the total number of spatial sampling points.

[0140] (5) Turbocharger performance mapping deviation:

[0141]

[0142] Among them, η sim (q k ) and η meas (q k ) represent the flow rate q k The simulated and measured supercharger efficiency at , K is the total number of flow sampling points.

[0143] (6) Pipeline segmentation error:

[0144]

[0145] Where ΔP sim (l) and ΔP meas (l) represents the simulated and measured pressure losses in the first section of the pipeline, and L is the number of pipeline sections. The above spatial dimension errors are weighted and superimposed by weight coefficients β1, β2, and β3 to obtain the spatial dimension comprehensive error F space :

[0146] F space =β1·E cyl +β2·E tc +β3·E pipe

[0147] The spatiotemporal composite fidelity metric is to superimpose the time dimension and space dimension errors by weight coefficients w1 and w2 to form a comprehensive fidelity evaluation index F total :

[0148] F total =w1·F time +w2·F space

[0149] Among them, w1 and w2 are the weight coefficients of temporal and spatial errors in the overall evaluation, respectively.

[0150] Step 3: By adjusting the relevant parameters in the simulation model, various typical fault conditions that may occur in the diesel engine are simulated, and then the operating scenarios of the diesel engine under different fault conditions are reproduced. The dynamic response characteristics under each fault condition are obtained through simulation calculation, including key performance indicators such as cylinder pressure, cylinder temperature, combustion heat release rate, dynamic characteristics of the intake and exhaust system, power, torque and fuel consumption rate, boost performance and efficiency analysis.

[0151] When selecting characteristic diesel engine operating parameters as a basis for fault diagnosis, it is important to ensure that the parameters are directly related to the engine's operating status or fault mode. For example, parameters that clearly reflect changes in engine performance, such as fuel consumption, power output, and emissions, are ideal diagnostic criteria. Furthermore, the parameters must be highly sensitive to the occurrence and changes of faults, quickly and significantly reflecting fault information to facilitate timely diagnosis. Furthermore, the selected parameters must remain stable during normal engine operation, and the measurement method must be reliable to reduce misdiagnosis due to parameter fluctuations or measurement errors. Finally, the parameters must be easily accessible and monitored via sensors or instruments. This not only simplifies the diagnostic process but also effectively improves diagnostic efficiency.

[0152] Step 4: Select key fault features, such as temperature and pressure, from the simulated fault sample data. These features should accurately reflect the diesel engine's operating conditions and failure modes. The data is then normalized or standardized to ensure consistent dimensionality across different features, facilitating subsequent model training.

[0153] Step 5: Initialize hyperparameters for the multi-stage hybrid time series classification network, including the size of the multi-scale convolution kernels, the number of attention heads and feedforward network size of the Transformer, and the hidden layer dimensions and activation function of the gated recurrent unit. To meet the requirements for high-precision feature extraction and classification of time series signals, the following multi-stage hybrid time series classification network is proposed for identifying fault characteristics and status patterns, specifically:

[0154] (1) Input processing unit: normalize, denoise or segment the original time series signal to obtain The input sequence is used to reduce noise interference and take into account sequence data of different sampling rates or lengths.

[0155] (2) Multi-scale convolution feature extraction layer: By using convolution operations with different convolution kernel sizes or expansion rates on parallel branches, short-term fluctuations and longer period features are obtained; after the outputs of each branch are spliced ​​in the channel dimension, channel fusion is performed through a 1×1 convolution or fully connected layer to generate a multi-scale convolution feature X conv .

[0156] (3) Transformer encoding module: multi-scale convolution output X conv It is regarded as a vector sequence and input to the Transformer encoder, which uses self-attention to obtain global dependencies. The Transformer encoder consists of two parts: multi-head attention and feedforward network, and cooperates with residual connection and layer normalization to output a feature sequence X containing global correlation information. trans .

[0157] (4) Gated recurrent unit integration unit: Transformer encoder output X trans It is further sent to GRU to retain the accumulated memory information that is more sensitive to time sequence dependence, and the temporal features are finally compressed to obtain the hidden state h rnn .

[0158] (5) Fusion output layer: Based on the above feature extraction, the following two strategies can be used to achieve classification: 1) For h rnn

[0159] Apply a fully connected (FC) mapping and obtain the category probability through the Softmax layer; 2) or combine the residual gating mechanism to perform feature fusion between multi-scale convolution, Transformer and GRU, and then perform Softmax output on the final fused vector.

[0160] (6) Loss function and training strategy: Cross entropy loss and stochastic gradient descent Adam are used for parameter optimization; if the model robustness needs to be enhanced, Dropout or regularization terms can be added at each stage to suppress overfitting.

[0161] In the above structure, the multi-scale convolutional layer captures local patterns for subsequences of different lengths, the Transformer encoding module focuses on global dependencies and feature interactions, and the gated recurrent unit integration unit strengthens temporal sequence and accumulated information, thereby forming a hybrid classification network that combines local and global, short-term and long-term dependencies, achieving high-precision identification of marine diesel engine fault types.

[0162] Step 6: Input the actual fault data to be tested into the pre-trained multi-level hybrid structure time series classification network to output the fault mode of the diesel engine.

[0163] Example 1

[0164] A marine diesel engine fault diagnosis method driven by an autonomous high-fidelity digital twin model, including:

[0165] Step 1: Based on the basic structure and related parameters of a certain type of hovercraft diesel engine, a one-dimensional simulation model of the diesel engine is built in the AVL-BOOST software. Figure 1 As shown in the figure, SB1, SB2, and SB3 are system boundaries; CO1 is the intercooler; PL1 is the intake and exhaust manifold; TC1 is the turbocharger; VP1 to VP6 are 6 crankcases, C1 to C6 are 6 cylinders, MP1 to MP8 are the arranged measurement points; and 1 to 29 are connecting pipes.

[0166] Step 2: After constructing the one-dimensional diesel engine simulation model, the next key step is to configure the detailed settings for each component in the model. This includes global parameters, turbocharger characteristics, cylinder status, and intercooler performance. To ensure that the simulation model accurately reflects the actual operating conditions of the diesel engine, the initial parameters of each component must be precisely set, and the system boundary conditions must be set to simulate the actual operating environment of the diesel engine. Furthermore, to ensure the simulation model has self-correcting capabilities, a chaotic adaptive whale swarm optimization algorithm was introduced. This algorithm uses the difference between measured and simulated data as a fidelity evaluation metric and utilizes an optimization algorithm to update the parameters of the diesel engine numerical simulation model, thereby maintaining the high fidelity of the diesel engine digital simulation model.

[0167] The chaotic adaptive whale swarm optimization algorithm includes exploration and hunting phases. The search and hunting phases introduce adaptive step sizes to search the solution space over a large range, and select hunting strategies based on the probability mechanism and coefficient vector. This method reduces the risk of falling into a local optimal solution, expands the search space, and improves overall efficiency. The parameter optimization process is as follows: Figure 2 shown.

[0168] Step 3: Set global parameters in the simulation model, change relevant model parameters, and simulate the parameter changes that trigger diesel engine failures, thereby simulating the occurrence of various faults. Based on the experience of experts and field personnel, nine typical diesel engine faults were selected for simulation, represented by codes F1 to F9, and the normal state is represented by F0. Each fault is assigned four fault states, ranging from minor to severe. The specific settings are shown in Table 1.

[0169] Table 1 Typical thermodynamic fault simulation parameter settings

[0170]

[0171] Combined with the selection criteria mentioned above and the actual conditions, select Figure 3 The 10 thermal parameters shown are used as fault characteristics. Figure 3 The thermal parameters represented by each symbol and the diesel engine state that can be reflected by selecting each thermal parameter are shown in Table 2.

[0172] Table 2 Selection of fault characteristics of hovercraft diesel engine

[0173]

[0174] Based on the fault parameter settings in Table 2, simulations were performed for nine fault states (F1-F9), generating 500 data sets for each state. Simultaneously, 2000 simulations were performed for the diesel engine in the normal state (F0). Ten thermal parameters from the resulting simulation data were extracted and normalized according to the settings in Table 2, forming an experimental dataset totaling 6500 samples, consisting of 4500 sets of fault data and 2000 sets of normal data.

[0175] Step 4: Next, extract key features from the simulated fault data, such as temperature and pressure. These features should accurately reflect the operating conditions and fault modes of the diesel engine, as shown in Table 1. The data is then normalized or standardized to ensure that the dimensions of different features remain consistent, facilitating subsequent model training.

[0176] Step 5: Initialize the hyperparameters of the time series classification network with a multi-level hybrid structure, including the size of the multi-scale convolution kernel, the number of attention heads of the Transformer and the scale of the feedforward network, as well as the hidden layer dimension and activation function of the gated recurrent unit; Next, input the original time series in the simulated fault dataset into the multi-scale convolution layer, extract the local patterns at different time scales, and fuse them in the channel dimension; Then, send the fused feature sequence into the Transformer encoding module, and capture the global dependency and key feature interaction through the self-attention mechanism; On this basis, the Transformer output is further input into the GRU for temporal memory and integration of accumulated information; Finally, combined with the training set label information, the probability prediction of the fault category is completed through the fully connected mapping and Softmax function of the output layer, and iterative optimization is performed with the cross entropy loss as the goal, converting the collaborative representation of the features of each layer of the network into the final classification result. The schematic diagram of its probabilistic neural network for identifying the fault type of marine diesel engine is shown below. Figure 4 shown.

[0177] 600 sets of fault-free data were selected, along with 600 sets each for nine fault conditions, including compressor fault, crankcase blowby, injection delay, and intercooler water-side blockage, for a total of 6,000 sets of data. This fault simulation data was used to train a multi-stage hybrid time series classification network. The network primarily consists of multi-scale convolutional layers, a Transformer encoder module, a GRU, and an output layer. The multi-scale convolutional layers use kernel sizes of [3, 5, 7] [3, 5, 7] [3, 5, 7] and dilation rates of [1, 2, 4] [1, 2, 4] [1, 2, 4] to extract features at different time scales. The output channels are 64, and the Reinforced Luminaire (ReLU) activation function is used. The Transformer encoder module employs eight attention heads, a feedforward network width of 256, and a four-layer stacking mechanism. A dropout ratio of 0.1 is used to enhance model robustness. The GRU layer has a hidden unit dimension of 128, is stacked in two layers, and employs a bidirectional architecture to capture forward and reverse temporal dependencies. The output layer maps features to classification results through a fully connected layer with a width of 128 and a softmax activation function. The network is trained using an optimizer with a learning rate of 0.001, a dropout ratio of 0.2 for regularization, and a cross-entropy loss function as the target to ensure the model's efficient classification of time series data.

[0178] In summary, the present invention addresses the problem of unsatisfactory fault diagnosis results during actual marine diesel engine fault diagnosis due to a lack of actual measured fault samples. A method is provided for replacing actual measured fault samples with simulated fault samples generated using a digital twin model of a diesel engine with autonomous fidelity. The present invention first constructs a numerical simulation model of the diesel engine and installs sensors on the actual diesel engine to collect operating data in real time, thereby constructing the digital twin model of the diesel engine. Secondly, a chaotic adaptive whale swarm optimization algorithm is used in conjunction with a spatiotemporal composite fidelity evaluation index to update the numerical simulation model parameters in real time, ensuring the high fidelity of the model. Next, based on the diesel engine's fault mechanism, the diesel engine's fault mode is set in the numerical simulation model to generate complete simulated fault samples. Finally, these simulated fault samples are used to train a multi-level hybrid structured time series classification network, enabling it to identify the types of unknown measured fault samples. This method addresses the problem of traditional intelligent diesel engine fault diagnosis methods failing due to the lack of actual measured fault samples, and can accurately identify the fault type of unknown measured fault samples in the absence of actual measured fault samples. The method of the present invention is a new tool and method for fault diagnosis of marine diesel engines, which promotes the development of related technologies and has important academic and engineering value.

[0179] In particular, in some preferred embodiments of the present invention, a computer device is also provided, including a memory and a processor and a computer program stored on the memory. When the processor executes the computer program, the steps of the marine diesel engine fault diagnosis method driven by the autonomous high-fidelity digital twin model described in any of the above embodiments are implemented.

[0180] In other preferred embodiments of the present invention, a computer-readable storage medium is also provided, on which a computer program / instructions are stored. When the computer program is executed by a processor, the steps of the marine diesel engine fault diagnosis method driven by the autonomous high-fidelity digital twin model described in any of the above embodiments are implemented.

[0181] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment method can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it may include the process of the embodiment of the marine diesel engine fault diagnosis method driven by the above-mentioned autonomous fidelity digital twin model, which will not be repeated here.

[0182] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment method can be implemented by instructing related hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the method of using bionic mechanical fish to identify and track aquatic biological communities as described above.

[0183] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples without contradiction.

[0184] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of technical features indicated. Thus, a feature specified as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "N" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0185] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code comprising one or more executable instructions for implementing a custom logical function or step of a process, and the scope of the preferred embodiments of the invention includes alternative implementations in which functions may be performed out of the order shown or discussed, including performing functions in a substantially simultaneous manner or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of the invention pertain.

[0186] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or N wires (electronic devices), a portable computer disk cartridge (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and programmable read-only memory (EPROM or flash memory), fiber optic devices, and a portable compact disc read-only memory (CDROM). Furthermore, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting or processing it in another suitable manner if necessary, and then storing it in a computer memory.

[0187] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiment, the N steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having logic gate circuits for implementing logic functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0188] Those skilled in the art will understand that all or part of the steps in the method of the above embodiment can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiment.

[0189] In addition, the functional units in the various embodiments of the present invention may be integrated into a single processing module, or each unit may exist physically separately, or two or more units may be integrated into a single module. The aforementioned integrated modules may be implemented in the form of hardware or in the form of software functional modules. If the integrated modules are implemented in the form of software functional modules and sold or used as independent products, they may also be stored in a computer-readable storage medium.

[0190] The storage medium mentioned above may be a read-only memory, a magnetic disk, or an optical disk, etc. Although the embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and are not to be construed as limiting the present invention. Persons skilled in the art may make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A marine diesel engine fault diagnosis method driven by an autonomous fidelity digital twin model, characterized in that: include: Step 1: Based on the basic structure and relevant parameters of the marine diesel engine, a one-dimensional simulation model of the diesel engine is constructed; Step 2: A chaotic adaptive whale swarm optimization algorithm is introduced based on a one-dimensional simulation model of a diesel engine to optimize the one-dimensional simulation model for model calibration. The chaotic adaptive whale swarm optimization algorithm includes a search and hunting phase. The search and hunting phase introduces an adaptive step size to search the solution space over a large range, and a hunting strategy is collaboratively selected based on a probability mechanism and a coefficient vector. The spatiotemporal composite fidelity evaluation index between the measured data and the simulation data is used as an objective function to update the cylinder combustion model parameters, intake and exhaust system parameters, turbocharger model parameters, and intercooler model parameters of the one-dimensional simulation model. Step 3: By adjusting the relevant parameters in the input simulation model, various fault states of the diesel engine are simulated, and the dynamic response characteristics under various fault states are obtained through simulation calculation to generate simulated fault samples; Acquire simulated fault samples to build a data set, and preprocess each simulated fault data in the data set; The pre-processed simulated fault data is used to construct a training set, where each sample in the training set includes simulated fault data and the corresponding fault category; The fault category is obtained based on the dynamic response characteristics in the simulated fault sample data; Step 4: Use the training set to train the time series classification network with a multi-level hybrid structure; Step 5: Input the actual fault data to be tested into the pre-trained multi-level hybrid structure time series classification network to output the fault mode of the diesel engine.

2. The marine diesel engine fault diagnosis method driven by the autonomous fidelity digital twin model according to claim 1 is characterized in that: The basic structure of the marine diesel engine includes: a cylinder model, an intake and exhaust pipe model, an intake and exhaust valve model, a supercharger, and an intercooler; the relevant parameters include operating parameters, combustion parameters, valve timing parameters, and supercharging system parameters.

3. The marine diesel engine fault diagnosis method driven by the autonomous high-fidelity digital twin model according to claim 1 is characterized in that: The step 2 specifically includes the following steps: Step 2.1: Initialization phase; using chaos mapping to optimize the population initialization in the chaotic adaptive whale swarm optimization algorithm Step 2.2: Optimal position storage stage; calculate the spatiotemporal composite fidelity evaluation index, select m minimum spatiotemporal composite fidelity evaluation indexes, and store the current global optimal position w * (t) and the positions of other excellent individuals are stored in the elite memory bank for subsequent calls to prevent falling into local optimality. Step 2.3: Search and hunting phase; introduce adaptive step size α(t) and coefficient vector A(t) to dynamically adjust the search range. When hunting, the adaptive coefficient A(t) and random probability P∈[0,1] are combined to coordinate and determine the predation strategy. Step 2.4: If the value of the spatiotemporal composite fidelity evaluation index does not change when the number of iterations exceeds 100, the optimization of the default model is converged or the number of iterations exceeds the maximum number of iterations M, and the optimal position w is output. * (t); otherwise, return to step 2.

2.

4. The marine diesel engine fault diagnosis method driven by the autonomous fidelity digital twin model according to claim 3 is characterized in that: The initialization phase of step 2.1 includes the following steps: Step 2.1.1: Set the population size N, the maximum number of iterations T, and the parameter search space [w min,d , w max,d ], adaptive step range [α min ,α max ]; Step 2.1.2: Chaos initialization position: Use Logistic chaotic mapping to generate chaotic sequence {x n }, then the generated chaotic sequence {x n } is mapped to each parameter dimension search space [w min,d , w max,d ], initialize the whale's individual position w i,d ; x n+1 =μx n (1-x n ) w i,d =w min,d +x n (w max,d -w mid,d ) Among them, x n ∈(0,1); μ is a parameter, set to 4; w i,d is the initial position of the i-th whale in the d-th dimension; w max,d and w min,d are the upper and lower bounds of the search space in the d-th dimension.

5. The marine diesel engine fault diagnosis method driven by the autonomous fidelity digital twin model according to claim 3 is characterized in that: The search and hunting phase of step 2.3 is specifically as follows: The adaptive step size α(t): The coefficient vector A(t): A(t)=2α(t)r-α(t) Among them, t is the current iteration number; r is a random vector in the range [0,1]; α max and α min are the initial maximum and minimum values, respectively; The adaptive coefficient A(t) and the random probability P select the hunting strategy for the current iteration: If the random constant P<0.5, the bubble net attack strategy is adopted; based on the current optimal position w * (t), the individual whale performs local search along the spiral trajectory: w(t+1)=|w * (t)-w(t)|·e c·n cos(2πn)+w * (t) Where w(t) is the current individual position; c is the spiral amplitude parameter, and n is a random number in the range of [-1,1]. If the random constant P ≥ 0.5, adopt the strategy of surrounding the prey and determine whether |A(t)| is less than 1; If |A(t)|<1, adopt the local search strategy of surrounding the prey; at the current best position w * (t), the position is updated by the adaptive coefficient A(t): w(t+1)=w * (t)-A(t)|C(t)·w * (t)-w(t)| Where C(t) is the random perturbation vector; If |A(t)|>1, adopt the global search strategy of surrounding the prey; take the current best individual position w * (t) is used as the basis, and the prey search strategy is used to update the individual whale position w(t+1) w(t+1)=w * (t)-g·A(t)|u·w * (t)-w(t)| Among them, g is a linearly decreasing parameter, and u is a random vector in [0,1]; According to one of the bubble net attack strategy, the local search strategy for surrounding prey, and the global search strategy for surrounding prey, the updated position w(t+1) of the whale individual is obtained and the spatiotemporal composite fidelity evaluation index of the whale individual is calculated. If there is a spatiotemporal composite fidelity evaluation index of the updated position w(t+1) of the whale individual that is less than the global optimal solution w * (t), then update the global optimal solution w * (t).

6. The marine diesel engine fault diagnosis method driven by the autonomous fidelity digital twin model according to claim 3 is characterized in that: The spatiotemporal composite fidelity evaluation index is a weighted superposition of the time dimension comprehensive error and the space dimension comprehensive error by weight coefficients w1 and w2; f total =w1·F time +w2·F space Among them, F total is the spatiotemporal composite fidelity evaluation index; w1 and w2 are the weight coefficients of temporal and spatial errors in the overall evaluation respectively; F time is the comprehensive error in time dimension; F space is the comprehensive error of spatial dimension.

7. The marine diesel engine fault diagnosis method driven by the autonomous fidelity digital twin model according to claim 6 is characterized in that: The evaluation index of the time dimension is selected as the key variables of speed, torque, fuel injection amount, and boost pressure that change over time. The time dimension error is calculated as follows: (1) Root mean square error RMSE: Among them, y sim (t i ) is the simulation signal at time point t i The value of y meas (t i ) is the measured signal at time point t i The value of , N is the total number of sampling points; (2) Waveform correlation coefficient ρ(X,Y): Among them, x i =y sim (t i ), y i =y meas (t i ), and are the average values ​​of the simulated signal and the measured signal respectively; the closer the correlation coefficient ρ is to 1, the more similar the waveforms are. (3) Mean deviation MD: (4) Time dimension comprehensive error F time : F time =α1·RMSE+α2·(1-ρ)+α3·MD Among them, α1, α2, and α3 are the weight coefficients of the root mean square error, waveform correlation coefficient, and average deviation, respectively. The evaluation indicators of the spatial dimension are selected from the cylinder partition error, the supercharger performance mapping deviation, and the pipeline segmentation error variables. The spatial dimension error is calculated as follows: (1) Cylinder partition error E cyl : Among them, p sim (r i ) and p meas (r i ) represent the spatial coordinates r i The simulated and measured pressure values ​​at , M is the total number of spatial sampling points; (2) Turbocharger performance mapping deviation E tc : Among them, η sim (q k ) and η meas (q k ) represent the flow rate q k The simulated and measured supercharger efficiency at , K is the total number of flow sampling points; (3) Pipeline segmentation error E pipe : Where ΔP sim (l) and ΔP meas (l) represents the simulated and measured pressure losses in the lth section of the pipeline, respectively, where L is the number of pipeline sections; (4) Time dimension comprehensive error F space : F space =β1·E cyl +β2·E tc +β3·E pipe Among them, β1, β2, and β3 are the weight coefficients of cylinder partition error, supercharger performance mapping deviation, and pipeline segmentation error, respectively.

8. The marine diesel engine fault diagnosis method driven by the autonomous high-fidelity digital twin model according to claim 1 is characterized in that: The multi-level hybrid structure time series classification network includes a multi-scale convolutional feature extraction layer, a Transformer encoding module, a gated recurrent unit integration unit, and a fusion output layer; First, the original time series in the simulated fault dataset is input into the multi-scale convolutional layer to extract local patterns at different time scales and fuse them in the channel dimension. Then, the fused feature sequence is sent to the Transformer encoding module to capture global dependencies and key feature interactions through the self-attention mechanism. The output of the Transformer encoding module is further input into the gated recurrent unit for temporal memory and integration of accumulated information. Finally, combined with the fault category label information of the training set, the probability prediction of the fault category is completed by fusing the fully connected mapping and Softmax function of the output layer. Iterative optimization is performed with the goal of minimizing the cross-entropy loss, and the collaborative representation of the features of each layer of the network is converted into the final classification result.

9. A computer device / apparatus / system comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images