Diesel engine misfire fault diagnosis method based on GRU and PINN

Abstract

The invention provides a diesel engine misfire fault diagnosis method based on a GRU and a PINN. The diesel engine misfire fault diagnosis method comprises the following steps of diesel engine multi-source data acquisition, data preprocessing and model sample generation; designing a gating circulation unit, introducing a physical constraint embedding module, constructing a misfire fault diagnosis model, and fusing the misfire fault diagnosis model with a long-short term memory model and a physical neural network; performing joint training and optimization on the misfire fault diagnosis model in the step S2 by using the samples in the step S1; and deploying a misfire fault diagnosis model and performing real-time diagnosis. The method has the beneficial effects that the time sequence data-physical model dual-drive diagnosis of the misfire fault of the diesel engine is realized; the accuracy of misfire diagnosis is improved; and the false alarm rate of complex working conditions is effectively reduced.

Description

Technical Field

[0001] The invention belongs to the technical field of diesel engine fault diagnosis, and in particular relates to a diesel engine misfire fault diagnosis method based on GRU and PINN. Background Art

[0002] Diesel engines are the core power source for ships, generator sets, and heavy machinery. Their operational reliability directly impacts the lifespan and safety of the equipment. In-cylinder misfire is a typical diesel engine fault, caused by abnormal fuel injection, insufficient compression, or ignition failure. It can lead to serious consequences such as sudden power drop, deteriorating emissions, and even mechanical damage. Traditional misfire diagnosis techniques rely on threshold rules or single-modal signal analysis, which has significant limitations. Existing technologies use a threshold for the cylinder head vibration signal amplitude to make judgments. However, the actual operating conditions of diesel engines are complex and variable, and fixed thresholds are prone to false alarms. The vibration amplitude of normal high-load combustion may overlap with low-load misfire conditions, resulting in an increased false positive rate. Another type of technology uses crankshaft speed fluctuation as a criterion, but the speed signal response has a cycle delay, making it difficult to detect early signs of misfire, delaying troubleshooting.

[0003] With the application of deep learning technologies, data-driven models based on RNNs and CNNs have gradually been introduced into the field of fault diagnosis. However, they face significant challenges in diesel engine scenarios. Firstly, misfire samples are scarce in actual diesel engine operation and maintenance, and data-driven diagnostic models require large amounts of sample data, resulting in insufficient accuracy with small sample sizes. Secondly, purely data-driven models ignore the physical constraints of the combustion process and are prone to misjudgment under unknown operating conditions and noise interference. Furthermore, multi-information fusion and real-time performance are limited. Diesel engine misfire characteristics exhibit millisecond-level transients and cross-cycle correlations, making it difficult for traditional fixed-size convolution kernels to effectively capture the sudden changes in non-stationary signals.

[0004] Therefore, existing diesel engine misfire diagnosis technologies face bottlenecks in small-sample generalization, the depth of physical mechanism integration, and the accuracy of multi-source time series modeling. To address these issues, this paper proposes a diesel engine misfire diagnosis method based on the fusion of a gated recurrent unit (GRU) and a physical information neural network (PINN). By combining multimodal data-driven feature extraction with bidirectional optimization of physical equation constraints, supplemented by an attention mechanism, this method overcomes the limitations of traditional methods, achieving high-precision misfire diagnosis in small sample conditions and improving the algorithm's interpretability and adaptability to operating conditions. Summary of the Invention

[0005] In view of this, the present invention aims to propose a diesel engine misfire fault diagnosis method based on GRU and PINN to solve the bottleneck problems of existing diesel engine misfire diagnosis technology in small sample generalization ability, physical mechanism fusion depth, and multi-source time series modeling accuracy.

[0006] To achieve the above object, the technical solution of the present invention is achieved as follows: A diesel engine misfire fault diagnosis method based on GRU and PINN includes the following steps: S1. Diesel engine multi-source data collection, data preprocessing, and model sample generation; S2. Design a gated recurrent unit, introduce a physical constraint embedding module, and build a misfire fault diagnosis model that integrates the long short-term memory model and the physical neural network. S3, using the samples in step S1 to jointly train and optimize the misfire fault diagnosis model in step S2; S4. Deployment of misfire fault diagnosis model and real-time diagnosis.

[0007] Furthermore, in step S1, data preprocessing includes: S11. Deploy a vibration sensor to collect vibration signals, a cylinder pressure sensor to collect combustion pressure fluctuations, an oil pressure sensor to collect high-pressure oil pipe oil pressure pulsation signals, and a speed sensor to collect instantaneous speed signals; S12. Adaptively decompose the vibration signal using variational modal decomposition, optimize the number of modes and penalty factors based on the minimum Pearson correlation coefficient principle, and separate noise and fault characteristics; S13. Calculate the time domain parameters of the vibration signal, combine the time domain signals of the cylinder pressure and oil pressure pulsation signals and the instantaneous speed in different periods, and perform multimodal feature extraction through time-frequency domain fusion; S14. Taking the crankshaft top dead center signal as a reference, intercepting a time series segment of a certain length, constructing a time series aligned multi-dimensional feature vector, and generating a data set sample after normalization; S15. Repeat steps S11 to S14 to perform fault injection experiments in different cylinders, record data and annotate labels to generate model samples.

[0008] Furthermore, in step S15, a model sample is generated, including: Sample A is used for model training; Sample B is used for model validation; Sample C is for model testing.

[0009] Furthermore, in step S2, a misfire fault diagnosis model is constructed, including: S21: Design the input layer, GRU structure and output layer; S22: At the input stage of the model, the attention mechanism is used to weight the input and assign different weights to different input elements according to their importance; S23: Design a physical constraint embedding module to analyze the in-cylinder combustion physics equation. The expression of the in-cylinder pressure differential equation is as follows: ; Where, is the specific heat ratio, is the transformation function of cylinder volume with crankshaft angle, is the cylinder pressure, and the heat release rate under normal working conditions Through the inverse calculation of actual measurement, the fault is caused by Linear attenuation; The crankshaft dynamic equation is expressed as follows: ; Where M, C, and K are the equivalent mass, damping, and stiffness matrices of the crankshaft system, respectively. is the load torque, is the torque loss term caused by misfire; S24: Combining the GRU network structure of step S21 and step S22 with the physical constraint embedding module of step S23 to jointly construct a diesel engine misfire diagnosis model.

[0010] Furthermore, in step S21, the input layer, GRU structure and output layer are designed, including: S211, receiving 128×4 dimensional data features as input layer; S212, GRU uses two core gating units and a candidate hidden state calculation module to reset the gate to control the influence of the hidden state h(t-1) of the previous time step on the current candidate hidden state, and update the gate to determine the historical information and new information retained by the hidden state h(t-1); S213, the candidate hidden state generates new information based on the current input and the state adjusted by the reset gate; S214,Bi-GRU processes the sequence through two independent GRUs, forward and reverse, and concatenates the outputs.

[0011] Furthermore, in step S3, joint training and optimization include: S31, design data driven loss part; In S32, in the PINN branch, the physical constraint loss is constructed according to S23, and the fused physical residual is designed by combining the crankshaft dynamics equation and the cylinder pressure differential equation. The expression is as follows: ; S33. Combine data-driven and physical constraints to design the total joint loss function, which is expressed as follows: ; Avoid single loss dominating the optimization process by dynamically weighting the average contribution of data and physical constraints; S34, train the GRU branch, freeze the PINN parameters, optimize only the GRU network parameters in the first 100 rounds, and use focal loss to fit the fault features; the initial learning rate is 0.001, and the decay coefficient is 0.2 every 50 rounds; S35: After 100 rounds of training, the joint fine-tuning phase begins. The PINN parameters are unfrozen, the weight parameters of the GRU and PINN are jointly adjusted, and the weight parameters are updated synchronously through the Adam optimizer. S36, dynamic parameter optimization, setting gradient thresholds to prevent gradient explosion caused by violent fluctuations in physical residuals, inserting Dropout between GRU connection layers, and adaptively adjusting constraint weights based on changes such as speed fluctuations; S37, using the early stopping method, if the accuracy of the validation set does not improve in ten consecutive rounds, terminate the training early, otherwise train directly to 300 rounds and end the training; S38. Solidify the fire diagnosis model parameters and save the optimal model.

[0012] Furthermore, in step S31, the data-driven loss part is designed, including: S311 and GRU branches use focal loss; S312. Let the input be multi-source time series features, extract instantaneous features and long-term dependency features through Bi-GRU, and output the failure probability.

[0013] Furthermore, in step S4, the misfire fault diagnosis model deployment and real-time diagnosis include: S41. The diesel engine misfire diagnosis model is lightweighted into an edge-adaptive diagnosis model. The model size is compressed while roughly maintaining model accuracy through knowledge distillation and quantitative pruning methods. S42, selecting a main control unit and an acquisition card to deploy an embedded fire fault diagnosis model; S43, collecting diesel engine sensor observation information in real time, obtaining real-time measurement data streams according to steps S11 to S15, aligning and inputting the data into the deployed misfire diagnosis model; S44: Perform delay testing and resource occupancy monitoring tests on the edge device diagnostic capability. If the resources exceed the limit or do not meet the preset accuracy, return to step S41 or step S42 for re-optimization.

[0014] Compared with the prior art, the diesel engine misfire fault diagnosis method based on GRU and PINN described in the present invention has the following beneficial effects: (1) The present invention realizes the dual-driven diagnosis of diesel engine misfire faults based on time series data and physical model by integrating GRU time series modeling and PINN physical equation constraints; (2) Combining lightweight models with real-time edge computing responses to improve the accuracy of fire diagnosis in small sample sizes; (3) Relying on the vibration-pressure-speed multi-modal feature fusion mechanism, the false alarm rate of complex working conditions can be effectively reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The accompanying drawings, which constitute part of the present invention, are intended to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are intended to explain the present invention and do not constitute an undue limitation of the present invention. In the accompanying drawings: Figure 1 A schematic diagram of the method described in an embodiment of the present invention; Figure 2 This is a schematic diagram of a specific model structure according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the specific collection, training, and prediction processes described in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0017] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, features defined as "first", "second", etc. may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0018] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; mechanical connections or electrical connections; direct connections or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0019] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0020] like Figures 1 to 3As shown, a diesel engine misfire fault diagnosis method based on GRU and PINN includes the following steps: S1. Diesel engine multi-source data collection, data preprocessing, and model sample generation; S2. Design a gated recurrent unit, introduce a physical constraint embedding module, and build a misfire fault diagnosis model that integrates the long short-term memory model and the physical neural network. S3, using the samples in step S1 to jointly train and optimize the misfire fault diagnosis model in step S2; S4. Deployment of misfire fault diagnosis model and real-time diagnosis.

[0021] The specific implementation is as follows: like Figure 1 The diagram shows the connection structure of this diesel engine misfire fault diagnosis method, which is generally divided into a training phase and a prediction phase. The training phase includes data collection through fault injection experiments, the construction of time-aligned data samples, the generation of a misfire diagnosis dataset, the design of a misfire diagnosis model integrating GRU and PINN, joint training, and embedded deployment. Fault injection is used to collect multi-source sensor signals under different misfire conditions and improve dataset diversity. The construction of time-aligned data samples aims to segment continuous time series signals into aligned model input data segments. The generation of the misfire diagnosis dataset aims to divide the overall sample into training, validation, and test sets. The design of the misfire diagnosis model integrating GRU and PINN is the core of this method. Its function is to fully extract multidimensional features under small sample conditions and improve misfire detection accuracy under the influence of data-driven and physical quantity constraints. The joint training process adjusts model parameters and optimizes and iterates the specific model. Embedded deployment deploys the trained model to edge devices through pruning and lightweighting methods for real-time prediction. The prediction phase also requires denoising and alignment of the real-time collected data according to the training strategy, and then inputs it into the deployed misfire diagnosis model for online fault diagnosis.

[0022] Figure 2The figure shows a deep learning-based diesel engine misfire diagnosis model, including the GRU structure and the bidirectional GRU model structure designed by the present invention with an attention mechanism. Input data is fed into the forward GRU and the backward GRU, respectively, through the attention mechanism. The bidirectional feature vectors are merged into a one-dimensional vector by merging the intermediate output layers. A multilayer perceptron with a [256, 128, 64] structure is then used to jointly train a more accurate diesel engine misfire diagnosis model, driven by both data and physical constraints. The GRU consists of a reset gate and an update gate. Its introduction can better capture the long-term dependencies of sequence data. Compared to LSTM, it offers higher processing efficiency, easier embedded deployment, and improved real-time performance. The introduction of the PINN enhances small-sample learning capabilities, improves data efficiency, and enhances interpretability, providing a physical model basis for the misfire diagnosis model.

[0023] Figure 3 The data acquisition process of the present invention is shown as follows. The data acquisition part of the present invention collects diesel engine vibration signals, cylinder pressure signals, rail pressure, instantaneous speed, and other signals, and performs normal signal acquisition and fault injection experiments respectively. Each set of signals is expanded through signal noise reduction, time series alignment, feature extraction, and normalization operations. Multiple sets of sensor signals are collected and processed according to the data acquisition preprocessing process. Finally, the training set, validation set, and test set are divided into 80%, 10%, and 10% of the total number. The model training part of the present invention is trained according to the backpropagation algorithm of deep learning. After the data is input into the model, dynamic parameter optimization is performed. When the training epoch is less than 100, the model is mainly driven by data. When the epoch is greater than 100, the physical constraint loss is added to the loss function for joint fine-tuning. Through multiple rounds of iteration and weight update, the training process is stopped when the training epoch exceeds 300 or the accuracy has not improved after ten rounds. The model parameters are solidified and the optimal model file is saved. The real-time prediction part of the present invention embeds the solidified optimal model in the edge device and inputs the preprocessed real-time signal into the edge device to predict the fire diagnosis result.

[0024] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A diesel engine misfire fault diagnosis method based on GRU and PINN, characterized by: The following steps are involved: S1. Diesel engine multi-source data collection, data preprocessing, and model sample generation; S2. Design a gated recurrent unit, introduce a physical constraint embedding module, and build a misfire fault diagnosis model that integrates the long short-term memory model and the physical neural network. S3, using the samples in step S1 to jointly train and optimize the misfire fault diagnosis model in step S2; S4. Deployment of misfire fault diagnosis model and real-time diagnosis.

2. The diesel engine misfire fault diagnosis method based on GRU and PINN according to claim 1 is characterized in that: In step S1, data preprocessing includes: S11. Deploy a vibration sensor to collect vibration signals, a cylinder pressure sensor to collect combustion pressure fluctuations, an oil pressure sensor to collect high-pressure oil pipe oil pressure pulsation signals, and a speed sensor to collect instantaneous speed signals; S12. Adaptively decompose the vibration signal using variational modal decomposition, optimize the number of modes and penalty factors based on the minimum Pearson correlation coefficient principle, and separate noise and fault characteristics; S13. Calculate the time domain parameters of the vibration signal, combine the time domain signals of the cylinder pressure and oil pressure pulsation signals and the instantaneous speed in different periods, and perform multimodal feature extraction through time-frequency domain fusion; S14. Taking the crankshaft top dead center signal as a reference, intercepting a time series segment of a certain length, constructing a time series aligned multi-dimensional feature vector, and generating a data set sample after normalization; S15. Repeat steps S11 to S14 to perform fault injection experiments in different cylinders, record data and annotate labels to generate model samples.

3. The diesel engine misfire fault diagnosis method based on GRU and PINN according to claim 2 is characterized in that: In step S15, a model sample is generated, including: Sample A is used for model training; Sample B is used for model validation; Sample C is for model testing.

4. The diesel engine misfire fault diagnosis method based on GRU and PINN according to claim 1, characterized in that: In step S2, a misfire fault diagnosis model is constructed, including: S21: Design the input layer, GRU structure and output layer; S22: At the input stage of the model, the attention mechanism is used to weight the input and assign different weights to different input elements according to their importance; S23: Design a physical constraint embedding module to analyze the in-cylinder combustion physics equation. The expression of the in-cylinder pressure differential equation is as follows: ； Where, is the specific heat ratio, is the transformation function of cylinder volume with crankshaft angle, is the cylinder pressure, heat release rate under normal working conditions Through the inverse calculation of actual measurement, the fault is caused by Linear attenuation; The crankshaft dynamic equation is expressed as follows: ； Where M, C, and K are the equivalent mass, damping, and stiffness matrices of the crankshaft system, respectively. is the load torque, is the torque loss term caused by misfire; S24: Combining the GRU network structure of step S21 and step S22 with the physical constraint embedding module of step S23 to jointly construct a diesel engine misfire diagnosis model.

5. The diesel engine misfire fault diagnosis method based on GRU and PINN according to claim 4 is characterized in that: In step S21, the input layer, GRU structure and output layer are designed, including: S211, receiving 128×4 dimensional data features as input layer; S212, GRU uses two core gating units and a candidate hidden state calculation module to reset the gate to control the influence of the hidden state h(t-1) of the previous time step on the current candidate hidden state, and update the gate to determine the historical information and new information retained by the hidden state h(t-1); S213, the candidate hidden state generates new information based on the current input and the state adjusted by the reset gate; S214,Bi-GRU processes the sequence through two independent GRUs, forward and reverse, and concatenates the outputs.

6. The diesel engine misfire fault diagnosis method based on GRU and PINN according to claim 1, characterized in that: In step S3, joint training and optimization are performed, including: S31, design data driven loss part; In S32, in the PINN branch, the physical constraint loss is constructed according to S23, and the fused physical residual is designed by combining the crankshaft dynamics equation and the cylinder pressure differential equation. The expression is as follows: ； S33. Combine data-driven and physical constraints to design the total joint loss function, which is expressed as follows: ； Avoid single loss dominating the optimization process by dynamically weighting the average contribution of data and physical constraints; S34, train the GRU branch, freeze the PINN parameters, optimize only the GRU network parameters in the first 100 rounds, and use focal loss to fit the fault features; the initial learning rate is 0.001, and the decay coefficient is 0.2 every 50 rounds; S35: After 100 rounds of training, the joint fine-tuning phase begins. The PINN parameters are unfrozen, the weight parameters of the GRU and PINN are jointly adjusted, and the weight parameters are updated synchronously through the Adam optimizer. S36, dynamic parameter optimization, setting gradient thresholds to prevent gradient explosion caused by violent fluctuations in physical residuals, inserting Dropout between GRU connection layers, and adaptively adjusting constraint weights based on changes such as speed fluctuations; S37, using the early stopping method, if the accuracy of the validation set does not improve in ten consecutive rounds, terminate the training early, otherwise train directly to 300 rounds and end the training; S38. Solidify the fire diagnosis model parameters and save the optimal model.

7. The diesel engine misfire fault diagnosis method based on GRU and PINN according to claim 6, characterized in that: In step S31, the data-driven loss part is designed, including: S311 and GRU branches use focal loss; S312. Let the input be multi-source time series features, extract instantaneous features and long-term dependency features through Bi-GRU, and output the failure probability.

8. The diesel engine misfire fault diagnosis method based on GRU and PINN according to claim 1 is characterized in that: In step S4, the misfire fault diagnosis model is deployed and diagnosed in real time, including: S41. The diesel engine misfire diagnosis model is lightweighted into an edge-adaptive diagnosis model. The model size is compressed while roughly maintaining model accuracy through knowledge distillation and quantitative pruning methods. S42, selecting a main control unit and an acquisition card to deploy an embedded fire fault diagnosis model; S43, collecting diesel engine sensor observation information in real time, obtaining real-time measurement data streams according to steps S11 to S15, aligning and inputting the data into the deployed misfire diagnosis model; S44: Perform delay testing and resource occupancy monitoring tests on the edge device diagnostic capability. If the resources exceed the limit or do not meet the preset accuracy, return to step S41 or step S42 for re-optimization.

Images