Engine Fault Prediction Methods, Programs, Products, Media, and Vehicles
By using a CNN-LSTM cascaded architecture model to perform deep learning on engine operating data, the problems of difficulty in capturing early and subtle signs and modeling difficulties in existing technologies have been solved, enabling early prediction and accurate warning of latent engine faults.
Patent Information
- Application Number
- CN202610993591.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-06
- Publication Date
- 2026-11-17
AI Technical Summary
Existing engine fault diagnosis methods struggle to capture early, subtle signs. Traditional threshold methods are prone to missed or false alarms, while physical modeling methods are difficult to model and lack adaptability to operating conditions, making it difficult to predict latent faults in advance.
The CNN-LSTM cascaded architecture model is used to perform deep learning on engine operating data, integrates multi-source features, automatically extracts deep temporal patterns, and has the ability to adapt to operating conditions and trace fault sources. The trained model outputs the fault probability and remaining life for early warning.
It enables early prediction of latent engine faults, reduces false alarm rate, and improves the accuracy of fault prediction and adaptability to different operating conditions.
Smart Images

Figure CN123115992A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of engine health management, and in particular relates to an engine fault prediction method, program product, medium and vehicle. Background Technology
[0002] As the core component of various power equipment, the engine's operational reliability directly affects the equipment's safety and economy. Mechanical faults such as piston ring wear and abnormal valve clearance have a progressive development characteristic, making them difficult to detect through conventional threshold alarms in their early stages. Once they reach a severe stage, they can lead to catastrophic consequences such as cylinder scoring and valve burning. The early detection and prediction of these "hidden faults" is a technical challenge in the field of intelligent engine operation and maintenance.
[0003] Existing technical solutions are mainly based on sensor fault diagnosis using observers. The core idea of this method is to construct an observer library using two nonlinear adaptive observers to estimate key parameters of the engine such as air-fuel ratio, temperature, and pressure, and to achieve fault detection through residual analysis between actual measured values and estimated values.
[0004] Existing engine fault diagnosis methods mainly suffer from the following problems: traditional threshold methods are difficult to capture early and weak signs; physical model methods have difficulty modeling complex wear mechanisms; and the warning thresholds are fixed, lacking the ability to adapt to operating conditions. Summary of the Invention
[0005] Embodiments of this application provide an engine fault prediction method, program product, medium, and vehicle, which can at least improve the ability to predict engine faults to a certain extent.
[0006] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part from practice of this application.
[0007] According to a first aspect of the embodiments of this application, an engine fault prediction method is provided. The method includes: acquiring engine operating data within a preset operating cycle; inputting the engine operating data into a trained first model, such that the first model outputs a first fault probability, a second fault probability, and the remaining engine life; wherein the first model is a CNN-LSTM cascaded architecture; acquiring the current operating condition of the engine; obtaining a warning interval corresponding to the current operating condition; outputting warning information for the first fault when the first fault probability and the remaining engine life fall within the warning interval; and outputting warning information for the second fault when the second fault probability and the remaining engine life fall within the warning interval.
[0008] As one possible implementation of the first aspect, the aforementioned engine operating data includes: exhaust temperature fluctuation data and oil consumption rate.
[0009] As one possible implementation of the first aspect, the exhaust temperature fluctuation data includes: inter-cylinder deviation data, deviation standard data, and frequency domain energy data.
[0010] As one possible implementation of the first aspect, the oil consumption rate is calculated based on an oil level sensor and oil filling records.
[0011] As one possible implementation of the first aspect, the first model includes at least two one-dimensional convolutional layers and at least two LSTM layers connected after the at least two one-dimensional convolutional layers.
[0012] As one possible implementation of the first aspect, obtaining the warning interval corresponding to the current operating condition based on the current operating condition includes: modifying a preset warning interval based on statistical characteristics of engine speed and load to obtain the warning interval corresponding to the current operating condition. The statistical characteristics include one or more of the following: mean, variance, and quantiles.
[0013] As one possible implementation of the first aspect, the method further includes: calculating the contribution of the engine operating data to the first fault and the second fault based on the attention weights of the first model; and obtaining a fault heatmap based on the contribution of the engine operating data to the first fault and the second fault.
[0014] According to a second aspect of the embodiments of this application, a computer program product is provided, the computer program product including computer instructions stored in a computer-readable storage medium and adapted to be read and executed by a processor to cause a computer device having the processor to perform the method as described in any of the embodiments of the first aspect above.
[0015] According to a third aspect of the present application, a computer-readable storage medium is provided, wherein at least one computer program instruction is stored therein, the at least one computer program instruction being loaded and executed by a processor to perform the operation as described in any of the embodiments of the first aspect above.
[0016] According to a fourth aspect of the present application, a vehicle is provided, the vehicle including one or more processors and one or more memories, the one or more memories storing at least one computer program instruction, the at least one computer program instruction being loaded and executed by the one or more processors to implement the method described in any of the embodiments of the first aspect above.
[0017] The technical solution provided in this application, through deep learning of the evolution law of latent faults using a CNN-LSTM hybrid model, can capture weak signs and predict engine faults in their early stages. Furthermore, the CNN module effectively suppresses the interference of measurement noise and short-term operating condition fluctuations on feature extraction, reducing the false alarm rate.
[0018] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description
[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. In the drawings: Figure 1 A schematic diagram of the CNN-LSTM cascaded architecture model provided in an embodiment of this application is shown; Figure 2 A flowchart of the engine fault prediction method in an embodiment of this application is shown; Figure 3 A detailed flowchart of the engine fault prediction method in an embodiment of this application is shown; Figure 4 A block diagram of an engine fault prediction device according to an embodiment of this application is shown; Figure 5 A schematic diagram of the vehicle structure in an embodiment of this application is shown. Detailed Implementation
[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0021] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.
[0022] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.
[0023] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily need to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.
[0024] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "multiple" means two or more.
[0025] To enable those skilled in the art to better understand this application, the application scenarios involved in this application will be briefly described first.
[0026] As the core component of various power equipment, the engine's operational reliability directly affects the equipment's safety and economy. Mechanical faults such as piston ring wear and abnormal valve clearance have a progressive development characteristic, making them difficult to detect through conventional threshold alarms in their early stages. Once they reach a severe stage, they can lead to catastrophic consequences such as cylinder scoring and valve burning. The early detection and prediction of these "hidden faults" is a technical challenge in the field of intelligent engine operation and maintenance.
[0027] Existing technical solutions are mainly based on sensor fault diagnosis using observers. The core idea of this method is to construct an observer library using two nonlinear adaptive observers to estimate key parameters of the engine such as air-fuel ratio, temperature, and pressure, and to achieve fault detection through residual analysis between actual measured values and estimated values.
[0028] For example, some solutions can predict engine faults using a threshold mechanism based on residual variance analysis and the Euclidean norm of the sliding window vector, thus reducing the false alarm rate. However, this method mainly targets sensor faults and has limited ability to detect latent faults in mechanical components such as piston ring wear and excessive valve clearance; moreover, it relies on a precise mathematical model of the engine, making modeling difficult.
[0029] Existing engine fault diagnosis methods have the following main shortcomings: (1) Traditional threshold method is difficult to capture early weak signs: Conventional methods are based on fixed thresholds of a single sensor (such as exhaust temperature, oil pressure) for alarm. However, the signal deviation caused by latent faults in the early stage is small and the rate of change is slow. It is easily overwhelmed by operating condition fluctuations and measurement noise, resulting in "missed alarms" or "false alarms".
[0030] (2) Physical modeling methods are difficult to model complex wear mechanisms: Piston ring wear involves multiple fields of thermo-mechanical-friction coupling, and valve clearance changes are affected by multiple factors such as thermal expansion and mechanical wear. It is extremely difficult to establish an accurate analytical model, and the model parameters drift with engine aging, which limits its practicality.
[0031] (3) Fixed warning threshold and lack of adaptive capability: The normal state parameter range of the engine varies significantly under different speed and load conditions. Using a uniform threshold will inevitably lead to the contradiction of false alarms under high operating conditions and missed alarms under low operating conditions.
[0032] In summary, there is an urgent need for a method to predict latent engine faults that can integrate multi-source features, automatically extract deep temporal patterns, and possess adaptive operating conditions and fault tracing capabilities.
[0033] In view of this, some embodiments of this application provide an engine fault prediction method. In this method, a fault analysis module can acquire engine operating data within a preset operating cycle. Next, the fault analysis module inputs the engine operating data into a trained first model, so that the first model outputs a first fault probability, a second fault probability, and the remaining engine life. This method can fuse multi-source features and automatically extract deep temporal patterns. Furthermore, this method possesses operating condition adaptive capability and fault tracing capability.
[0034] Before introducing the engine fault prediction method provided in the embodiments of this application, the CNN-LSTM cascaded architecture model provided in the embodiments of this application will be explained first.
[0035] The CNN-LSTM cascaded architecture is a hybrid deep learning network that is serially fused. "Cascaded" means that the CNN and LSTM are connected sequentially, with data passing in one direction and no parallel branches. The original input data flows completely through the CNN module to complete preprocessing and feature extraction. The feature sequence output by the CNN is then fed into the LSTM module as the sole input. The two types of networks perform their respective functions and complete feature mining step by step, which is different from the fusion scheme where CNN and LSTM extract features in parallel and then concatenate them.
[0036] The convolutional neural network at the front end of the architecture is responsible for mining local features of the data. It relies on the sliding convolutional kernels to traverse the data, capturing short-scale information such as local changes and abrupt changes within a window. Combined with pooling layers, it performs resampling and dimensionality reduction, compressing data dimensions and filtering out random noise. Simultaneously, it reduces the number of network parameters and computational overhead through a weight-sharing mechanism. After multiple layers of convolution and pooling operations, the original high-dimensional, noisy raw data is transformed into a concise, effective, and structured feature sequence.
[0037] The feature sequences refined by CNN are fed into the backend Long Short-Term Memory (LSTM) network. LSTM has a built-in gating structure, which can solve the gradient vanishing problem of long sequences in traditional recurrent neural networks, and has the ability to remember, forget, and update temporal information. This module no longer focuses on local instantaneous features, but rather sorts out the correlation between features along the time dimension, capturing the changing patterns and temporal dependencies of data over a long period of time.
[0038] The CNN-LSTM cascade architecture achieves complementarity through division of labor: the CNN performs noise reduction, dimensionality reduction, and local feature selection in the front, providing high-quality low-dimensional input for the LSTM; the LSTM mines long-term temporal dependencies in the back, while retaining the core capabilities of both types of networks. It takes into account both local detail extraction and long-term temporal relationship modeling, and is a classic basic hybrid model for processing temporal grid data.
[0039] For example, see Figure 1 The CNN-LSTM cascaded architecture model (hereinafter referred to as the first model) provided in this application embodiment includes the following structure: input layer, convolutional layer (CNN), long short-term memory (LSTM) layer, and output layer.
[0040] Input layer: Receives multidimensional feature sequences with a fixed time window (e.g., the most recent 30 days), with the dimension being (time step × feature dimension).
[0041] Convolutional Layer (CNN): Employs a two-layer one-dimensional convolutional neural network (Conv1D). The first layer has a kernel size of 3, and the second layer has a kernel size of 2. Each layer is followed by a ReLU activation function and batch normalization. The CNN module is used to automatically extract local temporal patterns, such as short-term trends in exhaust temperature fluctuations and abrupt changes in inter-cylinder differences.
[0042] Long Short-Term Memory (LSTM) layer: The high-level feature sequences extracted by the CNN are input into a two-layer LSTM network, with 64-128 hidden units in each layer. LSTM captures long-range dependencies across time steps through a gating mechanism, learning the complete evolution of latent faults from their inception to development.
[0043] Output layer: A fully connected layer with Softmax / Sigmoid activation function is used to output the predicted probabilities of two types of latent faults and the predicted estimated remaining useful lifetime (RUL).
[0044] It should be understood that the above models are typically used after training. The training phase of the CNN-LSTM model provided in the embodiments of this application is described below.
[0045] As one possible embodiment of the technical solution provided in this application, the training phase uses a historical dataset with clearly labeled data for model iteration. The dataset contains normal operating condition samples and various types of fault samples. All samples undergo uniform preprocessing, and the data scale is standardized and divided into training and validation sets. The training uses a weighted cross-entropy loss function. To address the sample imbalance problem caused by the uneven distribution of normal and fault samples in the dataset, differentiated weights are assigned to different categories of samples, amplifying the impact of a few fault samples on the loss value, preventing the model from favoring the category with a higher proportion of samples, and improving the balance of fault identification.
[0046] Throughout the training process, validation set metrics are monitored synchronously, and an early stopping strategy is introduced to suppress overfitting. After each complete training round, the model is fed into the validation set to calculate loss and recognition accuracy. The trend of validation set loss is continuously tracked. If the validation set loss no longer decreases or even shows an upward trend after several consecutive rounds, the model is considered overfitted, and iterative training is automatically terminated, retaining the weights of the model with the best validation performance. This entire training process balances sample distribution through weighted loss and constrains generalization error through the early stopping strategy, ensuring recognition accuracy for various types of samples while improving the model's adaptability to unknown data.
[0047] Understandably, the training device can train the initial first model and obtain the trained first model after the training is completed.
[0048] The criteria for determining training completion can be any one of the following: the number of training iterations reaches a preset number, an early stopping condition is triggered, or the loss function converges to a stable interval.
[0049] For ease of explanation, the first model trained as described above can be referred to as the fault prediction model in the following text.
[0050] The engine fault prediction method provided in the embodiments of this application will be described below with reference to the accompanying drawings.
[0051] For example, see Figure 2 The engine fault prediction method provided in this application embodiment may include steps S200-S203.
[0052] Step S200: Obtain engine operating data within a preset operating cycle.
[0053] The preset operating cycle can be one month, two months, three months, etc. The engine operating data includes exhaust temperature of each cylinder, oil pressure, and oil consumption.
[0054] One possible approach is for the vehicle to acquire raw data via sensors. Next, the raw data undergoes outlier removal, missing value imputation, filtering and noise reduction, and is resampled at fixed time windows (e.g., daily) to obtain the transmitter's operational data.
[0055] As an example, the engine operating data mentioned above includes: exhaust temperature fluctuation characteristics and oil consumption rate characteristics.
[0056] The aforementioned exhaust temperature fluctuation characteristics were obtained by calculating the inter-cylinder deviation and standard deviation of the exhaust temperature for each cylinder, and extracting the spectral energy distribution characteristics of the exhaust temperature signal. Piston ring wear leads to a decrease in cylinder sealing, which can cause abnormal fluctuations in the exhaust temperature of that cylinder; excessive valve clearance can cause valves to open too early or close too late, which will also manifest as a specific pattern of fluctuation in the exhaust temperature signal.
[0057] The oil consumption rate characteristic was obtained by calculating the oil consumption rate (unit: g / kWh) based on the oil level sensor or oil fill records. Piston ring wear is the direct cause of abnormally increased oil consumption, and this characteristic is highly specific for piston ring failures.
[0058] Step S201: Input the engine operation data into the trained first model.
[0059] The trained first model is the fault prediction model. Based on its input, the fault prediction model outputs the probability of a first fault, the probability of a second fault, and the engine's remaining life (RUL). As one possible implementation, the fault prediction model can also output the probability of a first fault and the engine's remaining life. Alternatively, it can output the probability of a second fault and the engine's remaining life. The first fault probability characterizes the likelihood of the engine experiencing a first fault within a preset future timeframe, and the second fault probability characterizes the likelihood of the engine experiencing a second fault within the same preset timeframe; this preset timeframe can be set to any duration such as one week, one month, two months, or six months. In some possible embodiments, the first fault could be piston ring wear, and the second fault could be excessive valve clearance.
[0060] It should be understood that the aforementioned exhaust temperature fluctuation characteristics and oil consumption rate characteristics are correlated to some extent with the first and second faults. For example, as shown in Table 1 below... Table 1. Correlation between faults and characteristics
[0061] Step S202: Obtain the current operating condition of the engine.
[0062] The aforementioned current operating conditions include: speed signal, load signal, and instantaneous fuel consumption.
[0063] Understandably, hardware tolerance limits and safety risk levels vary under different operating conditions. During high-load operation, component thermal loads and the risk of oil and gas leaks are higher, requiring narrower warning ranges and earlier intervention protection. During idling or shutdown, the system load is low, allowing for greater margin for normal fluctuations. Setting ranges for different operating conditions allows for matching the safety management requirements of those conditions.
[0064] Step S203: Obtain the warning interval corresponding to the current working condition based on the current working condition.
[0065] It's understandable that the combustion intensity, fuel evaporation rate, fuel pressure in the pipeline, and the amount of vapor generated in the fuel tank are completely different under various operating conditions, such as engine idling, low speed and light load, high speed and heavy load, and engine shutdown. The normal operating baseline ranges for each parameter are not uniform. If a uniform warning threshold is used, normal parameter fluctuations caused by different operating conditions will be judged as abnormal.
[0066] Therefore, the embodiments of this application can obtain the warning range corresponding to the current working condition according to different working conditions, so as to achieve the effect of adaptively correcting the warning threshold according to the real-time working condition of the engine, thereby balancing sensitivity and specificity in the entire working condition range.
[0067] As one possible implementation, step S203 includes: adaptively offsetting the preset warning threshold based on the statistical distribution characteristics (mean, variance, quantiles, etc.) of the engine's current speed and load. The preset warning threshold is raised under high load conditions and lowered under low load conditions.
[0068] For example, the preset warning thresholds are shown in Table 2 below.
[0069] Table 2 Preset Warning Thresholds
[0070] For example, adjustment coefficients can be obtained based on the aforementioned statistical characteristics. Then, based on the preset warning interval and the aforementioned adjustment coefficients, a warning interval corresponding to the current operating condition is obtained. Subsequently, based on the aforementioned warning interval corresponding to the current operating condition, and according to the first fault probability, the second fault probability, and the remaining engine life output by the aforementioned fault prediction model, warning information can be generated.
[0071] For example, if the probability of failure and the remaining engine life fall within the range corresponding to the first level mentioned above, the vehicle records the event and prompts the driver to increase monitoring.
[0072] For example, if the probability of a first failure and the remaining engine life fall within the range corresponding to the second level mentioned above, the vehicle will push maintenance suggestions for the first failure and prompt the vehicle's repair shop to prepare spare parts for the first failure.
[0073] For example, if the probability of failure and the remaining engine life fall within the range corresponding to the three levels mentioned above, the vehicle will issue an audible and visual alarm and recommend that the user of the vehicle have it inspected.
[0074] For example, if the second failure probability and the remaining engine life fall within the range corresponding to the first level mentioned above, the vehicle records the event and prompts the driver to increase monitoring.
[0075] For example, if the probability of a second failure and the remaining engine life fall within the range corresponding to the second level mentioned above, the vehicle will push maintenance suggestions for the second failure and prompt the vehicle's repair shop to prepare spare parts for the second failure.
[0076] For example, if the second failure probability and the remaining engine life fall within the range corresponding to the three levels mentioned above, the vehicle will issue an audible and visual alarm and recommend that the user of the vehicle have it inspected.
[0077] Optionally, after step S203 above, steps S204 and / or S205 may be performed.
[0078] For example, if the first fault probability and engine life output by the fault prediction model fall within the warning range corresponding to the current operating condition, step S204 is executed. If the second fault probability and engine life output by the fault prediction model fall within the warning range corresponding to the current operating condition, step S205 is executed.
[0079] Step S204: Output early warning information for the first fault.
[0080] For example, if the probability of failure and the remaining engine life fall within the range corresponding to the three levels mentioned above, the vehicle's dashboard will display a warning sign, and the vehicle's display screen will show a prompt suggesting that the vehicle user have the vehicle inspected.
[0081] Step S205: Output early warning information for the second fault.
[0082] For example, if the probability of a second failure and the remaining engine life fall within the range corresponding to the second level mentioned above, the vehicle's dashboard will display a warning sign, the vehicle's display screen will push repair suggestions for the second failure, and the vehicle's repair shop will be prompted to prepare spare parts for the second failure.
[0083] As one possible implementation, the engine fault prediction method provided in this application embodiment can also be applied to the maintenance process after a vehicle has been sent for repair. For example, the engine fault prediction method provided in this application embodiment can generate a fault heat map. The fault heat map can guide the actions of maintenance personnel, that is, maintenance personnel can use the fault heat map to quickly locate the fault and repair it.
[0084] For example, the engine fault prediction method provided in the embodiments of this application may further include steps S300-S301.
[0085] S300. Based on the attention weights of the fault prediction model, calculate the contribution of engine operating data to the first and second faults.
[0086] Understandably, the fault prediction model employs a CNN-LSTM cascaded architecture, with an attention layer embedded within the model. Real-time engine runtime sequence data is input to complete the forward inference, and the weight matrix output by the attention layer is read. The weight values represent the degree of attention paid to the model's output by operational features at each time step and in each dimension; a higher weight indicates a greater influence of that feature on fault determination.
[0087] For example, for the first fault prediction branch: extract the attention weights corresponding to the first fault prediction results output by the model, accumulate and normalize the weights of the same feature over the entire time series to obtain each engine operating parameter. For the second fault prediction branch: similarly extract the attention weights specific to the second fault, normalize them, and then calculate the quantitative contribution of each operating parameter to the second fault. Normalization rule: the sum of the contributions of all features under a single fault type is 1, with a value range of 0 to 1, intuitively distinguishing key influencing features from irrelevant features.
[0088] S301. A fault heatmap is obtained based on the contribution of engine operating data to the first fault and the contribution of engine operating data to the second fault.
[0089] For example, a fault heatmap can be created by using engine operating characteristics as the horizontal axis and fault types as the vertical axis, with matrix cells filled with the contribution values of the corresponding features to the fault. The horizontal axis of the fault heatmap includes dimensions such as engine speed, load, fuel temperature, fuel tank pressure, and valve opening / closing time. The vertical axis can include dimensions such as first fault, second fault, etc. In the fault heatmap, the higher the contribution value, the darker the cell color; a contribution value close to 0 is close to the background color.
[0090] Heat maps can intuitively show the differences in the core causes of the two types of faults, making it easy to quickly locate the key monitoring parameters of different faults and support adaptive calibration of early warning intervals and fault source analysis.
[0091] Next, taking the inline six-cylinder plug-in hybrid engine in a passenger car as the monitoring object, we will exemplarily illustrate the complete implementation process of early warning of piston ring wear based on the above-mentioned fault prediction model.
[0092] The test process continuously collected historical time-series data of the engine under all operating conditions for 6 consecutive months. The collected dataset fully covered the engine's normal steady-state operation without wear, the stage of slight piston ring wear development, and the stage of severe piston ring wear near failure. All data samples were manually labeled with faults. Each sample synchronously recorded multi-dimensional operating parameters such as engine speed, load, exhaust temperature of each cylinder, oil filling amount, fuel consumption, and running time. At the same time, the engine overhaul disassembly record was retained as a benchmark for fault true value comparison.
[0093] In the feature engineering phase, the raw operating data collected daily from the vehicle controller is standardized and preprocessed to eliminate differences in the dimensions of different parameters. Then, two types of key temporal features characterizing piston ring wear are extracted. The first type is the statistical characteristics of exhaust temperature in each cylinder: the mean and standard deviation of the exhaust temperature samples from the six cylinders are calculated daily, and the maximum difference between the six cylinder exhaust temperatures is further determined. Under normal, wear-free engine conditions, the piston sealing performance is intact, the combustion state in each cylinder is uniform, and the maximum deviation of the inter-cylinder exhaust temperature is consistently below 30℃. When piston ring wear occurs and the sealing gap widens, the combustion of the air-fuel mixture in the cylinder is incomplete, the combustion consistency between cylinders deteriorates significantly, and as the wear continues to deepen, the maximum deviation of the inter-cylinder exhaust temperature gradually increases. In the later stages of the fault development, this value stabilizes in the range of 60℃ to 80℃. The second type is the engine oil consumption rate index, calculated by converting daily fuel consumption and oil replenishment amounts to obtain the oil consumption rate per unit of power. When the piston ring seal is intact, the amount of oil entering the combustion chamber and participating in combustion is extremely low, and the oil consumption rate remains stable near the first value. As piston ring wear intensifies, oil burning increases continuously, and the oil consumption rate rises to a second value higher than the initial value during the near-failure stage. A two-dimensional time-series feature sequence is constructed by combining the daily calculated maximum inter-cylinder exhaust temperature deviation and oil consumption rate, serving as the basic input data for the model.
[0094] Subsequently, during the model training phase, a labeled historical dataset was used to iteratively train the aforementioned fault prediction model. This dataset contained a large number of normal operating condition samples and piston ring light, moderate, and severe wear fault samples. A weighted cross-entropy loss function was employed to alleviate the model bias problem caused by the imbalance between the number of normal and fault samples. Simultaneously, an early stopping method was introduced, along with real-time monitoring of model loss changes using a validation set, to suppress overfitting and ensure the model's generalization ability to recognize unknown operating conditions. The front-end one-dimensional convolutional layer slides along the time window to extract short-term trends in exhaust temperature deviation and oil consumption rate. A pooling layer performs data resampling, noise reduction, and dimensionality reduction. The back-end LSTM layer mines the long-term temporal evolution patterns of features over multiple days, ultimately outputting two predictive indicators: the probability of piston ring wear fault occurrence and the remaining service life (RUL).
[0095] Next, in the online real-time warning inference phase, the engine's exhaust temperature deviation and oil consumption rate time-series feature sequences extracted from the engine over the past 30 consecutive days are retrieved and input into the successfully trained and converged fault prediction model for forward inference calculations. The model outputs a 68% probability of piston ring wear failure and simultaneously predicts a remaining fault-free operating time (RUL) of 12 days. According to the preset fault classification rules, the failure probability is in the 50%–75% range, and the remaining lifespan is 7–15 days, meeting the trigger conditions for a level-two warning. After receiving the model prediction results, the vehicle controller immediately pushes targeted maintenance prompts through the vehicle's central control instrument panel and remote vehicle networking platform. The prompts may include, for example, "The probability of piston ring wear is high. It is recommended to focus on checking cylinders 2 and 4, and to prepare spare piston ring repair parts in advance for scheduled maintenance at the factory."
[0096] Then, during the real-vehicle performance verification phase, on the 12th day after the system issued a level-two warning, the driver reported that the vehicle was experiencing typical oil burning faults, such as difficulty starting the engine when cold and continuous blue smoke from the exhaust pipe during acceleration. The repair personnel went to the shop to disassemble the engine cylinder head and piston assembly for testing. The disassembly test results showed that the wear of the piston rings of the 2nd and 4th cylinders of the vehicle far exceeded the design allowable threshold, and the piston ring seals failed, which was completely consistent with the faulty cylinder and fault type given by the model prediction.
[0097] Therefore, the fault warning method provided in this embodiment can output early warning information before the piston ring wear failure fully erupts and obvious abnormal driving symptoms appear, allowing sufficient vehicle maintenance period, avoiding major secondary failures such as cylinder scoring and cylinder block damage caused by piston ring failure, and significantly reducing vehicle maintenance costs and road driving safety hazards.
[0098] In another optional embodiment, a family hybrid vehicle equipped with an inline four-cylinder naturally aspirated gasoline engine was selected as the test vehicle to conduct predictive and early warning verification work on the latent fault of gradually widening valve clearance after long-term engine operation. After the vehicle was put into use, it was driven under mixed urban and highway conditions for a long time. With the accumulation of mileage, the valve rocker arms and tappets showed slight wear, and the valve clearance gradually exceeded the factory calibration standard value. The vehicle slowly showed progressive fault manifestations such as abnormal metallic noise from the valves during cold starts and power loss during high-speed acceleration.
[0099] During the raw data acquisition process, the vehicle controller synchronously collects real-time, multi-dimensional operational data, including exhaust temperature of all four cylinders, coolant temperature during cold starts, engine speed, intake pressure, and mileage. This comprehensively records the vehicle's operational data from its initial, fault-free state to when valve clearance exceeds limits. Simultaneously, the measured valve clearance calibration data from the repair shop is recorded as a true fault label. In the feature extraction phase, the focus is on analyzing the temporal fluctuation characteristics of exhaust temperature in each cylinder under cold start conditions. Excessive valve clearance exhibits a fixed pattern: when the valve clearance exceeds limits in a cold engine state, the intake and exhaust valves close insufficiently, and the valves open prematurely, causing unburned combustion gases to be directly expelled from the cylinder, resulting in a significant decrease in cylinder exhaust temperature. After the engine warms up and runs for a period, the metal components expand due to heat, the valve clearance slightly decreases, and the exhaust temperature gradually returns to the normal range. Based on this physical mechanism, the average exhaust temperature and fluctuation amplitude of cylinders 1 and 3 within the first 90 seconds of each cold start phase are extracted as core fault features to construct a continuous temporal feature sequence.
[0100] A cold start exhaust temperature time-series dataset of multiple labeled vehicles of the same model was input into the initial fault prediction model for training. The training process employed weighted cross-entropy loss to balance the data distribution of normal samples and valve clearance fault samples, and combined with validation set monitoring and an early shutdown strategy to optimize the model's generalization ability. The model relies on convolutional layers to capture the instantaneous fluctuation features of exhaust temperature during a single cold start, and LSTM layers to mine the long-term changing trends of cold start exhaust temperature features over multiple days, thereby achieving probability prediction of valve clearance abnormal faults and assessment of remaining service life.
[0101] During online monitoring, the time-series data of the four-cylinder exhaust temperature during the daily cold start phase of the vehicle is continuously collected. This data is then spliced with the feature sequences from the past 30 days and input into the model for inference calculations. The model outputs a 55% probability of valve clearance exceeding the safe operating range (RUL) and predicts a remaining normal operating time (RUL) of 20 days. Based on the preset fault classification standards, the fault probability is in the 30%–50% range, and the remaining service life is 15–30 days, triggering a Level 1 warning mechanism. The onboard system simultaneously records this fault event, and the backend maintenance platform marks the vehicle as a key monitoring target, continuously tracking subsequent changes in its operational characteristics.
[0102] During the effectiveness verification phase, 22 days after the system issued a Level 1 warning, the driver brought the vehicle in for repair due to abnormal noise and insufficient power. The repair personnel used a professional feeler gauge to measure the valve clearance of each cylinder. The test results showed that the valve clearance of cylinder 1 and cylinder 3 exceeded the factory standard value by 0.12mm, which completely corresponded to the faulty cylinders identified by the model. After the repair personnel readjusted the valve clearance, the abnormal noise and power reduction problems during cold starts of the vehicle were completely eliminated.
[0103] This embodiment demonstrates that the fault warning scheme of this application can identify the trend of hidden faults in advance before the valve clearance fault develops into obvious abnormal noises and power loss symptoms that can be perceived by the driver, and continuously track the changes in characteristics. This not only avoids pushing maintenance reminders too early and causing trouble for users, but also locks in potential faults in advance, allowing users to plan reasonable maintenance time, and taking into account both vehicle operation safety and user experience.
[0104] It is understandable that the engine fault prediction method proposed in this application does not require pre-entry and calibration of the inherent physical parameters specific to various power units, and has good cross-engine and cross-power type transfer and adaptation capabilities. Based on the general modeling logic of time-series feature mining and CNN-LSTM fault prediction model, after model training is completed using only historical equipment operating data, it can be directly transferred and applied to online fault monitoring and early warning scenarios of various reciprocating and rotary power units such as gasoline engines, diesel engines, and gas turbines. There is no need to redesign the underlying judgment logic for different power units, which greatly reduces the development, calibration, and iteration costs of multi-engine diagnostic systems.
[0105] See Figure 4 The diagram shows a block diagram of the engine fault prediction device in an embodiment of this application.
[0106] like Figure 4 As shown, the engine fault prediction device 400 according to an embodiment of this application includes: a first acquisition unit 401, a first processing unit 402, a second acquisition unit 403, a second processing unit 404, a first output unit 405, and a second output unit 406.
[0107] The first acquisition unit 401 is used to acquire engine operating data within a preset operating cycle; the first processing unit 402 is used to input the engine operating data into a trained first model so that the first model outputs a first fault probability, a second fault probability, and the remaining engine life; wherein, the first model is a CNN-LSTM cascaded architecture; the second acquisition unit 403 is used to acquire the current operating condition of the engine; the second processing unit 404 is used to obtain a warning interval corresponding to the current operating condition based on the current operating condition; the first output unit 405 is used to output warning information for the first fault when the first fault probability and the remaining engine life belong to the warning interval; the second output unit 406 is used to output warning information for the second fault when the second fault probability and the remaining engine life belong to the warning interval.
[0108] Based on the same inventive concept, embodiments of this application also provide a computer program product, the computer program product including computer instructions stored in a computer-readable storage medium and adapted to be read and executed by a processor to cause a computer device having the processor to perform the method described above.
[0109] Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium storing at least one computer program instruction, which is loaded and executed by a processor to perform the operation as described above.
[0110] Based on the same inventive concept, this application also provides a vehicle, see reference. Figure 5 The diagram shows a structural schematic of a vehicle according to an embodiment of this application. The vehicle includes one or more memories 1004, one or more processors 1002, and at least one computer program (computer program instruction) stored in the memory 1004 and executable on the processor 1002. When the processor 1002 executes the computer program, it implements the method described above.
[0111] Among them, Figure 5In this document, a bus architecture (represented by bus 1000) is used. Bus 1000 may include any number of interconnected buses and bridges, linking various circuits including one or more processors represented by processor 1002 and memory represented by memory 1004. Bus 1000 may also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. Bus interface 1005 provides an interface between bus 1000 and receiver 1001 and transmitter 1003. Receiver 1001 and transmitter 1003 may be the same element, i.e., a transceiver, providing a unit for communicating with various other devices over a transmission medium. Processor 1002 is responsible for managing bus 1000 and general processing, while memory 1004 may be used to store data used by processor 1002 during operation.
[0112] The functions described herein may be implemented in hardware, software executed by a processor, firmware, or any combination thereof. If implemented in software executed by a processor, the functions may be stored as one or more instructions or codes on or transmitted via a computer-readable medium. Other examples and embodiments are within the scope and spirit of this application and the appended claims. For example, due to the nature of software, the functions described above may be implemented using software executed by a processor, hardware, firmware, hardwired, or any combination thereof. Furthermore, the functional units may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit.
[0113] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.
[0114] The units described as separate components may or may not be physically separate. Similarly, the components of the control device may or may not be physical units; they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.
[0115] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing computer program instructions, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.
[0116] The above description is merely an embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.
Claims
1. An engine failure prediction method characterized by, The method includes: Acquire engine operating data within a preset operating cycle; The engine operating data is input into a trained first model so that the first model outputs a first failure probability, a second failure probability, and the remaining engine life; wherein, the first model is a CNN-LSTM cascaded architecture; Obtain the current operating condition of the engine; Based on the current operating condition, a warning interval corresponding to the current operating condition is obtained; If the first failure probability and the remaining engine life fall within the warning range, output warning information for the first failure. If the second failure probability and the remaining engine life fall within the warning range, a warning message for the second failure is output.
2. The method of claim 1, wherein, The engine operating data includes: exhaust temperature fluctuation data and oil consumption rate.
3. The method according to claim 2, characterized in that, The exhaust temperature fluctuation data includes: cylinder deviation data, deviation standard data, and frequency domain energy data.
4. The method according to claim 2, characterized in that, The oil consumption rate is calculated based on the oil level sensor and oil filling records.
5. The method according to any one of claims 1-4, characterized in that, The first model includes at least two convolutional layers and at least two LSTM layers connected after the at least two convolutional layers.
6. The method according to claim 5, characterized in that, The process of obtaining the warning interval corresponding to the current operating condition based on the current operating condition includes: Based on the statistical characteristics of engine speed and load, the preset warning interval is corrected to obtain the warning interval corresponding to the current operating condition. The statistical characteristics include one or more of the following: mean, variance, and quantiles.
7. The method according to claim 6, characterized in that, The method further includes: Based on the attention weights of the first model, the contribution of the engine operating data to the first and second faults is calculated. A fault heatmap is obtained based on the contribution of the engine operating data to the first and second faults.
8. A computer program product, characterized in that, The computer program product includes computer instructions stored in a computer-readable storage medium and adapted to be read and executed by a processor to cause a computer device having the processor to perform the method as described in any one of claims 1 to 7.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions that are loaded and executed by a processor to perform the operations described in any one of claims 1 to 7.
10. A vehicle comprising a processor and a memory, characterized in that, The memory stores computer program instructions that can be executed by the processor, and when the processor executes the computer program instructions, it implements the instructions of the method as described in any one of claims 1 to 7.
