A piston ring-cylinder liner friction pair state evaluation method, system and model processing method

Abstract

The application provides a piston ring-cylinder liner friction pair state evaluation method and system and a model processing method. The model processing method comprises: establishing a machine learning data set comprising multi-modal signals and corresponding real wear annotations, the multi-modal signals comprising vibration signals, acoustic emission signals, cylinder exhaust side acoustic signals and oil particle signals; constructing a state evaluation model comprising a multi-modal fusion neural network model, the multi-modal fusion neural network model comprising a multi-modal encoder, a cross-modal fusion layer, a time series modeling layer and a multi-task output head connected in sequence; the multi-task output head at least comprising a first task output head for outputting wear grade recognition results and a second task output head for outputting wear depth prediction results; based on the machine learning data set, the state evaluation model is subjected to multi-task joint fine-tuning training to obtain a trained state evaluation model.

Description

Technical Field

[0001] This application relates to the field of internal combustion engine technology, specifically to a method, system, and model processing method for evaluating the condition of piston ring-cylinder liner friction pairs. Background Technology

[0002] Piston-type internal combustion engines, represented by diesel engines, are core power units in transportation, construction machinery, and power generation. The piston rings and cylinder liners form the core moving pair, and the wear state of this friction pair directly determines the engine's airtightness, lubrication performance, and combustion efficiency, thus affecting the overall engine power and fuel economy. Therefore, accurate real-time monitoring of the wear state of the piston ring-cylinder liner friction pair and prediction of its remaining service life are core technical requirements for ensuring reliable operation of internal combustion engines, implementing predictive maintenance, and reducing total life-cycle maintenance costs. Current piston ring-cylinder liner wear detection technologies are mainly divided into two categories: Firstly, while stopping the machine for disassembly and measurement can directly obtain high-precision wear data, it suffers from drawbacks such as high testing costs and long testing cycles. It cannot achieve real-time monitoring of the equipment's operating status, nor can it meet the needs for early fault warning.

[0003] Secondly, online monitoring relies heavily on the analysis of single physical signals, such as vibration signals, acoustic signals, or abrasive particles in lubricating oil. These approaches have significant technical limitations: Firstly, the operating conditions of internal combustion engines are complex and variable, and the wear mechanism evolves dynamically with these conditions. A single signal source is susceptible to fluctuations in operating conditions and interference from background noise, making it difficult to comprehensively characterize the complex wear process. The accuracy and reliability of the monitoring results are insufficient, leading to false alarms and missed alarms. Secondly, remaining service life prediction often uses statistical models or simplified physical models based on historical data, lacking the ability to comprehensively analyze the dynamic changes in real-time operating parameters and multi-source monitoring data. This makes it difficult to capture the nonlinear evolution trend of wear states, resulting in low prediction accuracy and poor generalization ability. It cannot provide individualized and accurate prediction results for different equipment, thus limiting its guiding value for maintenance decisions.

[0004] In view of this, this application proposes a method, system and model processing method for assessing the condition of piston ring-cylinder liner friction pairs, in order to solve one or more of the above-mentioned technical problems. Summary of the Invention

[0005] The first aspect of this application is to provide a method for processing a piston ring-cylinder liner friction pair condition assessment model.

[0006] The second aspect of this application is to provide a method for evaluating the condition of a piston ring-cylinder liner friction pair.

[0007] A third aspect of this application is to provide a piston ring-cylinder liner friction pair condition assessment system.

[0008] The piston ring-cylinder liner friction pair condition assessment model processing method according to the first aspect of this application includes: Multimodal signals during the operation of an internal combustion engine are acquired, including vibration signals, acoustic emission signals, cylinder exhaust side acoustic signals, and oil particle signals. The actual wear markings of the piston ring-cylinder liner friction pair of the internal combustion engine are obtained, including wear depth and geometric shape change data. A machine learning dataset including the multimodal signals and the corresponding actual wear markings is established. A state evaluation model is constructed, comprising a multimodal fusion neural network model. This model includes a multimodal encoder, a cross-modal fusion layer, a temporal modeling layer, and a multi-task output head connected sequentially. The multimodal encoder comprises a vibration encoder, an acoustic encoder, and an oil particle encoder connected in parallel. The vibration encoder is used for feature extraction from the vibration signal; the acoustic encoder is used for feature extraction from the acoustic emission signal and the cylinder exhaust acoustic signal; and the oil particle encoder is used for feature extraction from the oil particle signal. The cross-modal fusion layer adopts a Transformer encoder structure. Self-attention computation is performed to fuse the features output by the multimodal encoder into fused features; the temporal modeling layer includes a feature-level linear modulation unit and a temporal model. The feature-level linear modulation unit is used to perform feature-by-feature adaptive modulation on the fused features within each working cycle or time window. The temporal model adopts a Transformer encoder structure and is used to perform temporal modeling on the temporal feature sequence composed of the fused features of multiple consecutive working cycles or time windows after modulation; the multi-task output head includes at least a first task output head for outputting wear level identification results and a second task output head for outputting wear depth prediction results. Based on the machine learning dataset, the state evaluation model is jointly fine-tuned and trained using multiple tasks to obtain the trained state evaluation model.

[0009] The model processing method provided in this application is used to provide a state assessment model for piston ring-cylinder liner friction pairs. It uses multimodal signals, including four types of signals: vibration, acoustic emission, cylinder exhaust acoustics, and oil particle signals, as input sources. It can characterize the wear process of the friction pair from multiple physical dimensions such as structural vibration, material acoustic emission, exhaust acoustics, and oil particle abrasion, which helps to overcome the shortcomings of single-type signals being susceptible to operating condition fluctuations and background noise interference. A modal-specific encoder is set up in conjunction with Transformer self-attention cross-modal fusion, which can dynamically weight and align the features of each modality according to the samples. When the quality of a certain modal signal deteriorates, the weights are adaptively redistributed to reliable modes, improving the robustness of feature representation. After cross-modal fusion, a feature-level linear modulation unit is set up in conjunction with temporal Transformer modeling, which can normalize the fused features under different operating conditions to a comparable feature space and capture the nonlinear temporal evolution law of wear state. A multi-task output head, such as wear level classification and wear depth regression, is set up for joint fine-tuning training, which enables multiple tasks to share feature representations and mutually regularize constraints, improving data efficiency and prediction accuracy under the condition of limited real labeled data.

[0010] In some embodiments, obtaining the multimodal signal and the actual wear markings through a friction and wear test includes: Multiple acceleration sensors are arranged axially on the outer wall of the cylinder of the piston internal combustion engine to obtain the vibration signal; an acoustic emission sensor is arranged on the outer wall of the cylinder to obtain the acoustic emission signal; a microphone is arranged at the exhaust position of the cylinder to obtain the acoustic signal on the exhaust side of the cylinder; and an oil particle size analyzer is arranged in the oil return path of the cylinder to sample the oil return to obtain the oil particle signal.

[0011] In some embodiments, the multi-tasking output head further includes a third task head, which is configured to obtain the health status of the piston ring-cylinder liner friction pair and, based on the health status, obtain a predicted remaining service life; wherein... The health status is represented as follows: , In the formula, for Health at time , The wear depth prediction result, For the predetermined initial depth, The predetermined failure threshold; The health status is compared with a predetermined alarm threshold to predict the remaining service life; the health status and the alarm threshold are dynamically adjusted based on the false alarm rate and false negative rate of the wear level identification result of the first task output head and the wear depth prediction result of the second task output head.

[0012] In some embodiments, the state assessment model employs a deep ensemble structure, comprising multiple independently trained sub-models, and is configured to: infer in parallel from the multiple sub-models, using the mean of the predicted samples as the predicted value and the variance of the predicted samples as the prediction uncertainty, and obtain the prediction results of the wear depth and / or the remaining service life with confidence intervals based on the uncertainty; and, for samples with uncertainty higher than a predetermined threshold, conservatively shifting the benchmark alarm threshold, and / or using the prediction uncertainty as a screening criterion for high-risk data segments.

[0013] In some embodiments, the model processing method further includes: compressing the state evaluation model by at least one of knowledge distillation, network pruning, and weight and activation low-bit quantization, and then deploying it on a device-side edge computing node to perform real-time inference.

[0014] In some embodiments, the operation of the feature-level linear modulation unit is represented as follows: , In the formula, FiLM() is the feature-level linear modulation unit, X is the fused feature tensor output by the cross-modal fusion layer, γ is the scaling factor, β is the offset factor, and ⊙ is the Hadamard product.

[0015] In some embodiments, the model processing method further includes: performing leave-one-out cross-validation on the state evaluation model based on a plurality of piston internal combustion engines to obtain the state evaluation model that has passed leave-one-out cross-validation.

[0016] In some embodiments, the processing method further includes: after the state assessment model is deployed, performing online incremental updates on the state assessment model based on the collected actual operating data, and monitoring the wear prediction accuracy and false alarm rate of the updated version on the validation set; when the wear prediction accuracy is lower than a predetermined accuracy threshold or the false alarm rate is higher than a predetermined false alarm rate threshold, rolling back the parameters of the state assessment model to the version before the update.

[0017] In some embodiments, the model processing method specifically includes the steps of pre-training, supervised fine-tuning, domain adaptation, and deployment of the state evaluation model.

[0018] According to a piston ring-cylinder liner friction pair condition assessment method of the second aspect of this application, the method uses a condition assessment model obtained by the model processing method described in the first aspect to assess the condition of the piston ring-cylinder liner friction pair of a target internal combustion engine, including: The multimodal signals of the target piston internal combustion engine are acquired in real time during operation. The collected multimodal signals are input into the state assessment model deployed on the target piston internal combustion engine for real-time inference to obtain the wear level identification result and wear depth prediction result of the piston ring-cylinder liner friction pair.

[0019] In some embodiments, the health level, remaining useful life, and uncertainty obtained from model inference are used as the criteria for screening high-risk data; when the health level or the remaining useful life is lower than a set threshold, or the uncertainty is higher than a set threshold, the multimodal data fragments and their inference results within the corresponding time window are uploaded as high-risk data.

[0020] According to the piston ring-cylinder liner friction pair condition assessment system of the third aspect of this application, it includes a multimodal signal acquisition unit and a model inference unit; the model inference unit is used to perform condition assessment based on vibration signals, acoustic emission signals, cylinder exhaust side acoustic signals and oil particle signals from the multimodal signal acquisition unit, using the condition assessment method as described in the second aspect, to obtain the wear level identification result and wear depth prediction result of the piston ring-cylinder liner friction pair of the target piston internal combustion engine; the multimodal signal acquisition unit is used to acquire the vibration signals, acoustic emission signals, cylinder exhaust side acoustic signals and oil particle signals of the target piston internal combustion engine in real time.

[0021] In some embodiments, the multimodal signal acquisition unit includes: a plurality of acceleration sensors distributed axially on the outer wall of the cylinder for acquiring the vibration signal; an acoustic emission sensor disposed on the outer wall of the cylinder for acquiring the acoustic emission signal; a microphone disposed on the exhaust side of the cylinder for acquiring the acoustic signal on the exhaust side of the cylinder; and an oil particle size analyzer for sampling the oil return flow path of the cylinder to obtain the oil particle signal. Attached Figure Description

[0022] The above and other features, properties and advantages of this application will become more apparent from the following description taken in conjunction with the accompanying drawings and embodiments, in which the same reference numerals always denote the same features. It should be noted that these drawings are merely illustrative and are not drawn to scale, and should not be construed as limiting the scope of protection actually claimed by this application.

[0023] Figure 1 This is a schematic flowchart of a piston ring-cylinder liner condition assessment model processing and detection method according to one embodiment.

[0024] Figure 2 This is a schematic diagram of the structure of a multimodal fusion neural network model used in a piston ring-cylinder liner wear detection method according to one embodiment.

[0025] Figure 3 The figure schematically illustrates the wear depth prediction results obtained by using the piston ring-cylinder liner wear detection method of this disclosure in an example. The horizontal axis represents the running time (in hours) and the vertical axis represents the wear depth (in micrometers). Detailed Implementation

[0026] Reference will now be made in detail to various embodiments of this application, examples of which are shown in the accompanying drawings and described below. Although this application will be described in conjunction with exemplary embodiments, it should be understood that this application is not intended to be limited to those exemplary embodiments. Rather, this application is intended to cover not only these exemplary embodiments, but also various alternatives, modifications, equivalents, and other embodiments that may be included within the spirit and scope of this application as defined by the appended claims.

[0027] This application uses specific terms to describe embodiments of the application. For example, "an embodiment" and / or "one embodiment" refers to a particular feature, structure, or characteristic associated with at least one embodiment of the application. Therefore, it should be emphasized and noted that "an embodiment" or "one embodiment" mentioned twice or more in different locations in this application does not necessarily refer to the same embodiment. Furthermore, certain features, structures, or characteristics in one or more embodiments of the application can be appropriately combined.

[0028] In this application, the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying positional relationships or order of importance.

[0029] In the following description, the orientation or positional relationship indicated by terms such as "upper," "lower," "inner," "outer," "front," "back," or other directional terms is based on the orientation or positional relationship shown in the accompanying drawings and is only for the convenience of describing this application and simplifying the description, and is not intended to indicate or imply that the device or component referred to must have a specific orientation, be constructed and / or be implemented in a specific orientation, and therefore should not be construed as a limitation of this application.

[0030] In the following description, unless otherwise expressly specified or limited, the terms "installation," "connection," "linking," "fixing," "contact," etc., should be interpreted broadly; for example, they can refer to fixed connections or movable connections; they can refer to direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of these terms in the embodiments of this application according to the specific circumstances.

[0031] This application uses flowcharts to illustrate the operations performed according to embodiments of this application. It should be understood that, depending on the actual situation, the steps shown in the diagrams are not necessarily performed sequentially, and other operations may be added to these processes, or one or more steps may be removed from these processes.

[0032] It is understood that the model processing method, wear detection method and system of the condition assessment model provided in this application are particularly suitable for wear detection of piston ring-cylinder liner friction pairs in diesel engines, and can also be applied to other applicable occasions where wear detection of piston ring-cylinder liner friction pairs needs to be carried out, without limitation.

[0033] like Figures 1 to 3 As shown, this application provides a model processing method for evaluating the state of a piston ring-cylinder liner friction pair, including: Multimodal signals during the operation of a piston internal combustion engine are acquired, including vibration signals, acoustic emission signals, cylinder exhaust side acoustic signals, and oil particle signals. Correspondingly, the actual wear markings of the piston ring-cylinder liner friction pair of the piston internal combustion engine are obtained, including wear depth and geometric shape change data. A machine learning dataset including the multimodal signals and the corresponding actual wear markings is established. A state evaluation model is constructed, comprising a multimodal fusion neural network model, wherein the multimodal fusion neural network model includes a multimodal encoder, a cross-modal fusion layer, a temporal modeling layer, and a multi-task output head connected in sequence; Based on the machine learning dataset, the state assessment model is subjected to multi-task joint fine-tuning training, including at least the wear level identification task and wear depth prediction, to obtain the trained state assessment model for state assessment of the piston ring-cylinder liner friction pair of an internal combustion engine.

[0034] The wear depth is the amount of thinning of the mating surface at a characteristic location relative to its initial state, used to characterize the wear level.

[0035] The geometric morphology change data refers to at least one of the deviations of the macroscopic dimensions, macroscopic shape, and microscopic surface morphology of the piston ring-cylinder liner friction pair mating surfaces from their initial state. This data characterizes the spatial distribution pattern and failure mode of wear. Examples include one or more of the following: radial thickness reduction of the piston ring, increase in opening clearance, wear profile of the working surface, roundness, cylindricity, taper, and local wear profile depth of the cylinder liner, and the evolution of surface roughness and waviness of the mating surfaces. The macroscopic dimensions and shape parameters of the cylinder liner can be measured by a coordinate measuring machine, and the wear profile and microscopic surface morphology can be measured by a profilometer; these are not limited to these methods.

[0036] Using wear depth and geometric shape changes together as the supervision ground value of a multimodal fusion neural network can distinguish different failure modes such as uniform wear, uneven wear out-of-roundness, and cylinder scoring at similar wear depths, thereby enhancing the sufficiency of training supervision.

[0037] A state evaluation model is constructed, which includes a multimodal fusion neural network model, comprising a multimodal encoder, a cross-modal fusion layer, a temporal modeling layer, and a multi-task output head connected in sequence.

[0038] The multimodal encoder includes a vibration encoder and an oil particle encoder connected in parallel. The vibration encoder is used to extract features from the vibration signal, the acoustic encoder is used to extract features from the acoustic emission signal and the cylinder exhaust side acoustic signal, and the oil particle encoder is used to extract features from the oil particle signal. The network (e.g., a one-dimensional residual network, a two-dimensional convolutional network, and a multilayer perceptron) that matches the physical characteristics of each signal is used for feature extraction to avoid the representation loss caused by the unified processing of heterogeneous signals.

[0039] The cross-modal fusion layer employs a Transformer encoder structure to perform self-attention computation. It receives modal features output from each modal encoder and dynamically calculates the correlation weights between these features using a self-attention mechanism, fusing the features output from the multimodal encoders into a fused feature. The cross-modal fusion layer dynamically weights and aligns the modal features based on the samples. When the quality of a particular modality degrades or is even lost due to fluctuations in operating conditions or background noise, it can adaptively reallocate weights to reliable modes, thereby overcoming the shortcomings of single signal sources, such as susceptibility to interference and susceptibility to false alarms and missed alarms.

[0040] The temporal modeling layer includes a feature-level linear modulation unit and a temporal model. The feature-level linear modulation unit receives the fused features from each working cycle or time window output by the cross-modal fusion layer and adaptively modulates them feature by feature. Through feature-level linear modulation unit scaling and offsetting operations, the fused features under different operating states are normalized to a comparable feature space, providing a stable input for subsequent temporal modeling. The temporal model employs a Transformer encoder structure to perform temporal modeling on the feature sequences of the fused features from multiple consecutive working cycles or time windows after modulation, capturing the nonlinear temporal evolution of wear.

[0041] The multi-task output head includes at least a first task output head (a classification head) for outputting wear level identification results and a second task output head (a regression head) for outputting wear depth prediction results, enabling multiple tasks to share feature representations and act as regularization constraints on each other, thereby improving data efficiency and prediction accuracy under the condition of limited labeled data.

[0042] The piston ring-cylinder liner friction pair state assessment model processing method provided in this application is used to provide a state assessment model. By using multimodal signals including four types of signals—vibration, acoustic emission, cylinder exhaust side acoustics, and oil particle signals—as input sources, it can characterize the wear process of the friction pair from multiple physical dimensions, including structural vibration, material acoustic emission, exhaust acoustics, and oil particle abrasion, overcoming the shortcomings of single signal sources being susceptible to operating condition fluctuations and background noise interference. By setting up a modal-specific encoder in conjunction with Transformer self-attention cross-modal fusion, it can dynamically weight and align the features of each modality based on the samples, and adaptively adjust the signal quality when the quality of a certain modality deteriorates. By reallocating weights to reliable modes, the robustness of feature representation is improved. By setting feature-level linear modulation units in conjunction with temporal Transformer modeling after cross-modal fusion, the fused features under different operating conditions can be normalized to a comparable feature space, and the nonlinear temporal evolution law of wear state can be captured. By setting up wear level classification and wear depth regression multi-task output heads for joint fine-tuning training, multiple tasks can share feature representations and be mutually regularized constraints, improving data efficiency and prediction accuracy under limited labeled data conditions, so as to ultimately achieve accurate and reliable detection of the wear state of the piston ring-cylinder liner friction pair of internal combustion engine.

[0043] like Figure 1 As shown, the model processing method may include at least one of the following steps: self-supervised pre-training, supervised fine-tuning, domain adaptation, and deployment of the state evaluation model.

[0044] Specifically, in some embodiments, the model processing method includes one or more of the following steps S1 to S8.

[0045] S1. Test preparation steps, including sensor installation, wiring and hardware synchronization, that is, installing accelerometers, acoustic emission sensors, broadband acoustic microphones, oil particle online meters, etc. for multimodal signal acquisition on the piston internal combustion engine, and using crankshaft angle encoders to achieve multi-channel synchronous acquisition of the high-speed acquisition system through hardware triggering or precise timestamps, to ensure that the multimodal signals are time-synchronized (e.g., error less than 100 microseconds).

[0046] Preferably, an accelerometer is fixed to the cylinder head near the target cylinder and to the outer side of the cylinder block near the base, using bolts and rigid coupling washers to ensure a rigid connection, for detecting high-frequency vibration signals; an acoustic emission sensor is attached to the cylinder head surface and fixed using a high-temperature coupling agent and clamps, for detecting acoustic emission signals; a broadband acoustic microphone is positioned near the exhaust port or exhaust pipe and mounted using a vibration-damping bracket to reduce external noise coupling, for detecting acoustic signals on the cylinder exhaust side; an online oil particle detector is installed at the return port of the lubricating oil return pipeline and a bypass sampling interface is provided to achieve offline oil sample collection, for detecting oil particle signals, including the number and size of abrasive particles in the lubricating oil, to characterize wear products and wear mechanisms. All sensor signal lines are shielded and grounded to ensure system electromagnetic compatibility.

[0047] S2. Wear test and true value acquisition steps: Wear progression test is carried out on a test bench with functions such as load control and lubrication adjustment. At each wear progression stage of the test (for example, divided into five stages: initial stage, light wear, moderate wear, heavy wear and critical wear), the wear depth and geometric morphology change data of the piston ring-cylinder liner friction pair at the corresponding time are collected and recorded using a coordinate measuring machine, profilometer and other equipment.

[0048] S3. Data preprocessing and sample construction steps, namely, data preprocessing and sample construction of multimodal experimental data.

[0049] Preferably, the original signal is sliced ​​according to the crankshaft angle, and statistical and time-frequency features are extracted through processing such as filtering, envelope analysis, short-time Fourier transform and continuous wavelet transform. Then, normalization processing is performed to finally construct a machine learning dataset including a training set, a validation set and a test set.

[0050] Among them, short-time Fourier transform, continuous wavelet transform and envelope analysis are mainly applicable to waveform signals such as vibration, acoustic emission and cylinder exhaust side acoustics; oil particle signals, because they are low-rate counting and particle size data, their preprocessing is mainly trend extraction, counting and concentration statistics and normalization.

[0051] Crankshaft angle slicing is used to achieve phase alignment of each working cycle, to locate the phase window where wear and cylinder scoring are most severe, such as the top dead center commutation zone, and to suppress the influence of speed fluctuations; filtering is used to remove out-of-band interference such as DC drift, low-frequency engine vibration, and combustion excitation, and to improve the signal-to-noise ratio of weak signals from friction and acoustic emission; envelope analysis uses Hilbert transform demodulation to demodulate the amplitude modulation components of high-frequency carrier waves caused by abrasive scratches, rough peak collisions, and friction impacts; short-time Fourier transform is used to characterize the quasi-stationary time-frequency evolution within a single working cycle; continuous wavelet transform is used to perform high-resolution time-frequency localization of short-time broadband transient signals such as acoustic emission, making up for the shortcomings of short-time Fourier transform in transient capture. On this basis, statistical and time-frequency features such as root mean square value, kurtosis, peak factor, acoustic emission count, and energy are extracted; normalization is used to eliminate dimensional and magnitude differences between modes, stabilize model training, and improve the generalization ability across working conditions and engine models.

[0052] S4. Build a self-supervised multi-task deep learning model, that is, build a multimodal fusion neural network model that includes vibration, acoustic and oil particle information as described above.

[0053] Preferably, the vibration encoder adopts a one-dimensional residual network (1D-ResNet) structure to adapt to the feature extraction of one-dimensional vibration time-series signals; the acoustic encoder adopts a two-dimensional convolutional network or residual network (2D-CNN / ResNet18) structure to uniformly extract features from the two-dimensional time-frequency spectra obtained by time-frequency transformation of the acoustic emission signal and the cylinder exhaust side acoustic signal. Since the acoustic emission signal and the cylinder exhaust side acoustic signal belong to the same acoustic waveform signal, their time-frequency spectra have similar two-dimensional time-frequency structures, so they can share the same acoustic encoder architecture to reduce the number of parameters and improve the feature generalization ability across acoustic modes; the oil particle encoder adopts a multi-layer perceptron (MLP) structure to adapt to the processing of low-rate statistical features such as oil particle count and concentration, so as to realize the feature extraction of vibration, acoustic emission, cylinder exhaust side acoustic and oil particle signals using networks that match their physical characteristics, avoiding the representation loss caused by the unified processing of heterogeneous signals.

[0054] Preferably, the multimodal encoder also reserves a working condition encoder interface, and adopts an MLP structure to access working condition signals such as speed and load as auxiliary inputs when needed, providing an explicit working condition reference for the cross-modal fusion layer, and assisting the feature-level linear modulation unit to more accurately complete the feature distribution alignment under different working conditions, further improving the inference accuracy of the model in variable working condition operation scenarios.

[0055] Preferably, the operation of the feature-level linear modulation unit can be expressed as follows: , In the formula, FiLM() is the feature-level linear modulation unit, X is the fusion feature tensor output by the cross-modal fusion layer, γ is the scaling factor generated adaptively by the network, β is the offset factor generated adaptively by the network, and ⊙ is the Hadamard product.

[0056] By employing a feature-level linear modulation mechanism, the network adaptively generates scaling and offset coefficients based on the current fused features. The fused features of each work cycle are then subjected to affine transformation channel by channel, aligning the feature distributions under different operating conditions such as speed and load to a unified feature space. This eliminates the influence of operating condition fluctuations on the offset of feature distributions, and improves the stability of subsequent time series modeling and its cross-operating condition generalization ability.

[0057] Preferably, the multi-task output head further includes a third task output head (a regression head) for obtaining health status and predicting remaining useful life (RUL), and the three task output heads correspond to the wear level identification, wear depth prediction and health status and remaining useful life prediction tasks, respectively.

[0058] The health level can be represented by a normalized health index ranging from 0 to 1. A health level of 1 indicates that the friction pair is in a wear-free state, while a health level of 0 indicates that the friction pair has reached a predetermined failure threshold. This is expressed as: , In the formula, Represents the health status at time t. , , , These are the predicted wear depth, predetermined initial depth, and predetermined failure threshold output by the multimodal fusion neural network model, respectively.

[0059] The remaining service life can be obtained by extrapolating the health degradation curve: based on the historical window, the decline trend of health over time is estimated, and the health is extrapolated to the operating time corresponding to its decline to 0 (i.e., the predicted wear depth increases to the failure threshold). The difference between this operating time and the current time is taken as the remaining service life. The true value of the remaining service life is marked by the actual operating time experienced by the friction pair from the current state to the wear depth reaching the failure threshold in the bench wear propagation test.

[0060] S5. Multi-task supervised fine-tuning step, which uses the acquired machine learning dataset to perform multi-task joint fine-tuning training, and simultaneously optimizes tasks such as wear level identification and wear depth prediction.

[0061] Preferably, step S5 further includes independently training multiple sub-models to jointly constitute a deep ensemble structure of the state assessment model, thereby introducing uncertainty estimation for model inference using a deep ensemble method. During state assessment model inference, each sub-model performs forward computation in parallel, taking the mean of its wear depth, health status, and / or remaining service life prediction samples as the final prediction value, taking the variance of the prediction samples as the prediction uncertainty, and outputting the prediction result with confidence interval based on the uncertainty.

[0062] Furthermore, when the uncertainty rises to a predetermined threshold, the upper bound of the confidence interval for wear depth and the lower bound of the confidence interval for remaining service life and health are used for conservative judgment to reduce false alarm rate and false negative rate; and / or, uncertainty is used as the basis for triggering conservative maintenance and manual review, the basis for screening high-risk data segments, and at least one of the following gating conditions for online incremental updates and rollback after model deployment.

[0063] S6. Model verification and threshold adjustment steps, namely, using multiple internal combustion engines to perform leave-one-out cross-validation experiments, and dynamically adjusting the health and alarm thresholds based on the false alarm rate and false negative rate of wear and life prediction performance.

[0064] Figure 3 The figure shows the wear depth prediction results obtained by using an example of the condition assessment method provided in this application. The predicted value and the actual value in the figure maintain a high degree of agreement throughout the entire life cycle, indicating that the method can achieve accurate prediction of wear depth.

[0065] Step S7: Edge deployment and online monitoring step, which involves deploying the state assessment model on the device-side edge computing node, such as an in-vehicle or other edge computing node, to achieve real-time feature extraction and inference without relying on cloud network connection.

[0066] Preferably, a lightweight state evaluation model is obtained by compression processing through at least one of knowledge distillation, network pruning, and low-bit quantization of weights and activations. The compressed model performs real-time feature extraction and inference locally on an embedded computing platform, such as a vehicle or other edge computing node.

[0067] Preferably, the health status, remaining useful life, and prediction uncertainty obtained through local inference are used as the criteria for screening high-risk data segments. When the health status or remaining useful life is lower than a set threshold, or the prediction uncertainty is higher than a set threshold, the multimodal data segment and its inference results for the corresponding time are triggered to be uploaded to the cloud platform as high-risk data segments, so as to retain high-risk data samples with low prediction confidence, sudden changes in wear status, or close to failure threshold.

[0068] Furthermore, for the remaining data segments that do not belong to the high-risk data segments, no data is uploaded or only key statistical summaries are uploaded, thereby reducing communication bandwidth and cloud computing and storage overhead while selectively preserving rare degradation samples with critical wear and abnormal mutations.

[0069] Step S8: Online fine-tuning step, which is to perform small-step online incremental fine-tuning based on the actual operating data of the target internal combustion engine after model deployment, so that the model can continuously adapt to the individual characteristics and operating environment changes of the target internal combustion engine, thereby improving the accuracy of individualized prediction.

[0070] Preferably, it also includes a safety rollback step, which involves continuously monitoring performance changes on the validation set. When the prediction accuracy of the validation set decreases or the false alarm rate increases significantly, exceeding a predetermined threshold, the model parameters are automatically rolled back to the version before the update. This ensures that the update process is safe and controllable and avoids the degradation of model performance after the update from affecting the normal wear detection function.

[0071] This application also provides a condition assessment method for the piston ring-cylinder liner friction pair of an internal combustion engine (such as a diesel engine). This method uses a condition assessment model obtained from the model processing method described above to perform wear detection and life prediction on the piston ring-cylinder liner friction pair of the target piston-type internal combustion engine, including: The multimodal signals of the target piston internal combustion engine are acquired in real time during operation. The collected multimodal signals are input into the state assessment model deployed on the target piston internal combustion engine for real-time inference to obtain the wear level identification result and wear depth prediction result of the piston ring-cylinder liner friction pair.

[0072] This condition assessment method can directly use the multimodal fusion condition assessment model provided above to perform real-time inference on the multimodal signals collected in real time from the target internal combustion engine. It can obtain prediction results such as wear level classification and continuous wear depth without disassembling the equipment, thus achieving real-time and accurate monitoring of the wear state of the friction pair.

[0073] This application also provides a condition assessment system for piston ring-cylinder liner friction pairs in internal combustion engines (such as diesel engines), comprising a multimodal signal acquisition unit and a model inference unit; the model inference unit is used to perform real-time inference based on vibration signals, acoustic emission signals, cylinder exhaust side acoustic signals, and oil particle signals from the multimodal signal acquisition unit, using the condition assessment model obtained by the processing method described in the first aspect, to obtain wear level identification results and wear depth prediction results for the piston ring-cylinder liner friction pairs of the target piston internal combustion engine; the multimodal signal acquisition unit is used to acquire the vibration signals, acoustic emission signals, cylinder exhaust side acoustic signals, and oil particle signals of the target piston internal combustion engine in real time.

[0074] Specifically, the multimodal signal acquisition unit includes: multiple acceleration sensors distributed along the axial direction of the cylinder on the outer wall of the cylinder for acquiring the vibration signal; an acoustic emission sensor arranged on the outer wall of the cylinder for acquiring the acoustic emission signal; a microphone arranged on the exhaust side of the cylinder for acquiring the acoustic signal from the exhaust side of the cylinder; and an oil particle size analyzer for sampling the oil return flow path of the cylinder to obtain the oil particle signal. The installation and wiring of these sensors are basically the same as those described in the test preparation step S1 above, and are adapted to the actual operating environment of the internal combustion engine, and will not be elaborated here.

[0075] In summary, the beneficial technical effects of this application include, but are not limited to, at least one of the following: The piston ring-cylinder liner friction pair state assessment model processing method provided in this application is used to provide a state assessment model. By using multimodal signals including four types of signals—vibration, acoustic emission, cylinder exhaust side acoustics, and oil particle signals—as input sources, it can characterize the wear process of the friction pair from multiple physical dimensions such as structural vibration, material acoustic emission, exhaust acoustics, and oil particle abrasion, overcoming the shortcomings of single signal sources being susceptible to operating condition fluctuations and background noise interference. By setting up a modal-specific encoder in conjunction with Transformer self-attention cross-modal fusion, it can dynamically weight and align the features of each modality according to the samples, and adaptively adjust the signal when the quality of a certain modality signal deteriorates. Reliable modal weight redistribution enhances the robustness of feature representation. By setting a feature-level linear modulation unit in conjunction with temporal Transformer modeling after cross-modal fusion, the fused features under different operating conditions can be normalized to a comparable feature space, and the nonlinear temporal evolution of wear state can be captured. By setting up a multi-task output head for wear level classification and wear depth regression for joint fine-tuning training, multiple tasks can share feature representations and act as regular constraints on each other, improving data efficiency and prediction accuracy under limited labeled data conditions, so as to ultimately achieve accurate and reliable detection of the wear state of the piston ring-cylinder liner friction pair in internal combustion engines.

[0076] While this application discloses preferred embodiments as described above, it is not intended to limit the scope of this application. Any changes and modifications can be made by those skilled in the art without departing from the spirit and scope of this application. Therefore, any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of this application, without departing from the content of the technical solution of this application, fall within the protection scope defined by the claims of this application.

Claims

1. A method for processing a piston ring-cylinder liner friction pair condition assessment model, characterized in that, include: Multimodal signals during the operation of an internal combustion engine are acquired, including vibration signals, acoustic emission signals, cylinder exhaust side acoustic signals, and oil particle signals. The actual wear markings of the piston ring-cylinder liner friction pair of the internal combustion engine are obtained, including wear depth and geometric shape change data. A machine learning dataset including the multimodal signals and the corresponding actual wear markings is established. A state evaluation model is constructed, comprising a multimodal fusion neural network model, wherein the multimodal fusion neural network model includes a multimodal encoder, a cross-modal fusion layer, a temporal modeling layer, and a multi-task output head connected in sequence; wherein, The multimodal encoder includes a vibration encoder, an acoustic encoder, and an oil particle encoder connected in parallel. The vibration encoder is used to extract features from the vibration signal, the acoustic encoder is used to extract features from the acoustic emission signal and the cylinder exhaust side acoustic signal, and the oil particle encoder is used to extract features from the oil particle signal. The cross-modal fusion layer uses a Transformer encoder structure to perform self-attention calculation, which is used to fuse the features output by the multimodal encoder into fused features; The timing modeling layer includes a feature-level linear modulation unit and a timing model. The feature-level linear modulation unit is used to perform feature-by-feature adaptive modulation on the fused features within each working cycle or time window. The timing model adopts a Transformer encoder structure and is used to perform timing modeling on the timing feature sequence composed of the fused features of multiple consecutive working cycles or time windows after modulation. The multi-task output head includes at least a first task output head for outputting wear level identification results and a second task output head for outputting wear depth prediction results; Based on the machine learning dataset, the state evaluation model is jointly fine-tuned and trained using multiple tasks to obtain the trained state evaluation model.

2. The model processing method according to claim 1, characterized in that, The multimodal signals and the actual wear marks are obtained through friction and wear tests, including: Multiple acceleration sensors are arranged axially on the outer wall of the cylinder of the piston internal combustion engine to obtain the vibration signal; an acoustic emission sensor is arranged on the outer wall of the cylinder to obtain the acoustic emission signal; a microphone is arranged at the exhaust position of the cylinder to obtain the acoustic signal on the exhaust side of the cylinder; and an oil particle size analyzer is arranged in the oil return path of the cylinder to sample the oil return to obtain the oil particle signal.

3. The model processing method according to claim 1, characterized in that, The multi-task output head also includes a third task head, which is configured to obtain the health status of the piston ring-cylinder liner friction pair and, based on the health status, obtain a predicted remaining service life result; wherein... The health status is represented as follows: ， In the formula, for Health at time , The wear depth prediction result, For the predetermined initial depth, The predetermined failure threshold; The health status is compared with a predetermined alarm threshold to predict the remaining lifespan; The multi-task joint fine-tuning training also includes synchronously optimizing the remaining service life prediction task, and dynamically adjusting the alarm thresholds of the health status and remaining service life based on the false alarm rate and false negative rate of the wear level identification task and the wear depth estimation wear depth prediction task.

4. The model processing method according to claim 3, characterized in that, The condition assessment model adopts a deep ensemble structure, including multiple independently trained sub-models, and is configured to: infer in parallel from the multiple sub-models, using the mean of the predicted samples as the predicted value and the variance of the predicted samples as the prediction uncertainty, and obtain the prediction results of the wear depth and / or the remaining service life with confidence intervals based on the uncertainty; and, based on the uncertainty, take a conservative judgment on at least one of the predicted results of health, remaining service life and wear depth, and / or use the prediction uncertainty as the screening criterion for high-risk data segments.

5. The model processing method according to claim 1, characterized in that, Also includes: The state evaluation model is compressed using at least one of knowledge distillation, network pruning, and low-bit quantization of weights and activations, and then deployed on the device-side edge computing node to perform real-time inference.

6. The model processing method according to claim 1, characterized in that, The operation process of the feature-level linear modulation unit is represented as follows: ， In the formula, FiLM() is the feature-level linear modulation unit, X is the fused feature tensor output by the cross-modal fusion layer, γ is the scaling factor, β is the offset factor, and ⊙ is the Hadamard product.

7. The model processing method according to claim 1, characterized in that, The model processing method further includes: performing leave-one-out cross-validation on the state evaluation model based on multiple piston internal combustion engines to obtain the state evaluation model that has passed leave-one-out cross-validation.

8. The model processing method according to claim 1, characterized in that, Also includes: After the state assessment model is deployed, it is updated online based on the collected actual operating data, and the wear prediction accuracy and false alarm rate of the updated version on the validation set are monitored. When the wear prediction accuracy is lower than a predetermined accuracy threshold or the false alarm rate is higher than a predetermined false alarm rate threshold, the parameters of the state assessment model are rolled back to the version before the update.

9. The model processing method according to claim 1, characterized in that, Specifically, this includes the pre-training, supervised fine-tuning, domain adaptation, and deployment steps of the state evaluation model.

10. A method for evaluating the condition of a piston ring-cylinder liner friction pair, characterized in that, The state assessment model obtained by the model processing method as described in any one of claims 1 to 9 is used to assess the state of the piston ring-cylinder liner friction pair of the target piston internal combustion engine, including: The multimodal signals of the target piston internal combustion engine are acquired in real time during operation. The collected multimodal signals are input into the state assessment model deployed on the target piston internal combustion engine for real-time inference to obtain the wear level identification result and wear depth prediction result of the piston ring-cylinder liner friction pair.

11. The condition assessment method according to claim 10, characterized in that, The health level, remaining useful life, and uncertainty obtained from model inference are used as the criteria for screening high-risk data. When the health level or the remaining useful life is lower than a set threshold, or the uncertainty is higher than a set threshold, the multimodal data segment and its inference results within the corresponding time window are uploaded as high-risk data.

12. A piston ring-cylinder liner friction pair condition assessment system, characterized in that, Includes a multimodal signal acquisition unit and a model inference unit; The model inference unit is used to perform state assessment based on the vibration signal, acoustic emission signal, cylinder exhaust side acoustic signal and oil particle signal from the multimodal signal acquisition unit, using the state assessment method as described in claim 10 or 11, to obtain the wear level identification result and wear depth prediction result of the piston ring-cylinder liner friction pair of the target piston internal combustion engine. The multimodal signal acquisition unit is used to acquire in real time the vibration signal, the acoustic emission signal, the cylinder exhaust side acoustic signal, and the oil particle signal of the target piston internal combustion engine.

13. The condition assessment system according to claim 12, characterized in that, The multimodal signal acquisition unit includes: Multiple acceleration sensors are distributed axially on the outer wall of the cylinder to collect the vibration signals; An acoustic emission sensor is disposed on the outer wall of the cylinder for collecting the acoustic emission signal; A microphone is positioned on the exhaust side of the cylinder to collect acoustic signals from the exhaust side of the cylinder. An oil particle size analyzer is used to sample the oil return flow path of the cylinder to obtain the oil particle signal.

Images