Vehicle engine operation state monitoring method and device, controller and storage medium

Abstract

The application provides a method and device for monitoring the running state of a vehicle engine, a controller and a storage medium, wherein the method comprises: acquiring a monitoring signal of the vehicle engine in a running state in real time, performing sliding window slicing and normalization processing on the monitoring signal to generate a window sequence; inputting the window sequence into a trained feature extraction model to output an instantaneous latent vector, the feature extraction model being used to capture subtle changes of early weak faults of the engine in time domain, frequency domain and phase, and the instantaneous latent vector being used to compress and represent the current health state characteristics of the engine; outputting an early fault risk value through an early warning module, and outputting an early warning signal when the continuous multiple early fault risk values are all greater than a dynamic threshold. The application realizes the promotion of the detection granularity from the second-level overall energy to the millisecond-level transient state, greatly reduces the detectable signal-to-noise ratio threshold, and significantly improves the capture ability of early weak faults of the engine.

Description

Technical Field

[0001] This invention relates to the field of vehicle engine monitoring technology, and more specifically, to a method, device, controller, and storage medium for monitoring the operating status of a vehicle engine. Background Technology

[0002] Traditional vehicle engine operating status monitoring systems mostly rely on threshold alarms or single frequency domain characteristics. When faced with early, minor faults, such as microcracks in turbine blades, early pitting of bearings, or local misfires in the combustion chamber, their signal-to-noise ratio is extremely low, and the fault characteristics are often submerged in operating condition fluctuations, sensor drift, and environmental noise.

[0003] Even more serious is that these subtle anomalies manifest as transient impacts in the time domain, lasting only a few milliseconds; in the frequency domain, they exhibit diffuse spectra with "wideband, low amplitude," highly overlapping with combustion noise and gear meshing harmonics; and in the phase domain, they are modulated by speed fluctuations and sampling clock jitter, further weakening their detectability. As a result, even with the most refined threshold adjustments, existing systems can only trigger alarms at the obvious fault stage where the signal-to-noise ratio is greater than 10 dB. By this time, the crack has already expanded to a critical size, the repair window is significantly compressed, and it may even lead to catastrophic failure. Summary of the Invention

[0004] The main objective of this invention is to provide a method, device, controller, and storage medium for monitoring the operating status of a vehicle engine, so as to significantly improve the ability to detect early and minor engine faults.

[0005] In a first aspect, to achieve the aforementioned objective, the present invention provides a method for monitoring the operating status of a vehicle engine, comprising:

[0006] Real-time acquisition of monitoring signals from the vehicle engine during operation;

[0007] The monitoring signal is sliced ​​using a sliding window and normalized to generate a window sequence;

[0008] The window sequence is input into the trained feature extraction model, which outputs an instantaneous latent vector. The feature extraction model is composed of a lightweight convolutional neural network and a Transformer encoder, which is used to capture subtle changes in the time domain, frequency domain and phase of early engine minor faults. The instantaneous latent vector is a high-order abstract representation used to compress and characterize the current health status features of the engine.

[0009] The instantaneous latent vector is input into the early warning module, which outputs an early fault risk value. When the early fault risk values ​​corresponding to multiple consecutive window sequences are all greater than the dynamic threshold, an early warning signal is output. The early warning module includes an LSTM network and a fully connected layer, which are used to convert the instantaneous latent vector into an early fault risk value.

[0010] Furthermore, before inputting the window sequence into the already trained feature extraction model, the process further includes:

[0011] Multiple monitoring signal samples are acquired, and a window sequence sample corresponding to each monitoring signal sample is generated. Under the condition of no manual labeling, a preset proportion of the window sequence samples are randomly occluded and reconstructed from the context to obtain the target window sequence sample.

[0012] The Transformer encoder is used to map each target window sequence sample into a low-dimensional feature vector to obtain the first instantaneous latent vector;

[0013] The first instantaneous latent vectors arranged in chronological order in the same running record are sequentially fed into the convolutional neural network to generate the second instantaneous latent vector. The second instantaneous latent vector contains the historical information of all target window sequence samples before each target window sequence sample.

[0014] Select the second instantaneous latent vector of the current target window sequence sample and the first instantaneous latent vector of the subsequent target window sequence sample that is adjacent in time to form a positive sample pair. Select the first instantaneous latent vector of the target window sequence sample that is not adjacent or originates from different running records within the same batch and form a negative sample pair with the current second instantaneous latent vector.

[0015] For each sample pair, the similarity score between the second instantaneous latent vector and the first instantaneous latent vector is calculated, and the parameters of the encoder and convolutional neural network are updated through backpropagation until the similarity score of the positive sample pair is maximized and the similarity score of the negative sample pair is minimized, thus generating a feature extraction model that has been trained.

[0016] Furthermore, after generating the trained feature extraction model, the process further includes:

[0017] When each new batch window sequence arrives, the third instantaneous latent vector is predicted using the current weight vector of the convolutional neural network, and the difference between the prediction and the actual observed fourth instantaneous latent vector is calculated to obtain the observation residual.

[0018] The residual covariance is calculated based on the observed residuals and cross-referenced with the current system state covariance of the convolutional neural network to obtain the optimal correction amount.

[0019] The optimal correction is applied to the current weight vector of the convolutional neural network to obtain the updated weight vector, and the current system state covariance is updated to reflect the new confidence level of the weight vector estimation.

[0020] The updated weight vector is immediately written back to the convolutional neural network, and the next batch of window sequences is processed.

[0021] Preferably, the calculation of the similarity score between the second instantaneous latent vector and the first instantaneous latent vector includes:

[0022] Both the second instantaneous latent vector and the first instantaneous latent vector are fed into the phase channel and the energy channel, and the first phase latent vector and the first energy latent vector corresponding to the second instantaneous latent vector are output, as well as the second phase latent vector and the second energy latent vector corresponding to the first instantaneous latent vector. The phase channel is used to perform Hilbert transform on the second instantaneous latent vector or the first instantaneous latent vector to extract the instantaneous phase sequence, and obtain the phase latent vector through a causal dilated convolutional network. The energy channel is used to perform continuous wavelet transform on the second instantaneous latent vector or the first instantaneous latent vector to extract the energy spectrum, and obtain the energy latent vector through an asymmetric convolutional neural network.

[0023] Calculate the first cosine distance between the first phase latent vector and the second phase latent vector, and calculate the second cosine distance between the first energy latent vector and the second energy latent vector;

[0024] The similarity score is obtained by weighted summation of the first cosine distance and the second cosine distance.

[0025] Preferably, the step of performing sliding window slicing and normalization processing on the monitoring signal to generate a window sequence includes:

[0026] The monitoring signal is mirrored to generate a symmetrical waveform. The drift component is obtained by fitting a quadratic curve based on the local extreme points of the symmetrical waveform. The drift component is then superimposed on the symmetrical waveform to obtain the extended signal.

[0027] Calculate the instantaneous energy envelope of the extended signal, detect the sparse pulses of the instantaneous energy envelope, and if there is an instantaneous energy envelope interval that is lower than the global median and has a sparse pulse density greater than the neighborhood mean, then mark it as a suspected signal band.

[0028] The complexity of the local signal is determined based on the local gradient density of the instantaneous energy envelope. Using a variable-length primitive window as a seed, the suspected signal band is lengthened or shortened based on the complexity of the local signal to obtain the target signal band.

[0029] Each window corresponding to the target signal band is divided into three segments: front, middle, and back. Max or minimum normalization based on local extrema is performed on the front and back segments to suppress endpoint drift. Mean and variance normalization based on instantaneous energy envelope is performed on the middle segment to highlight fault-related fluctuations. The normalized results of the three segments are then spliced ​​together to form a complete window.

[0030] Each normalized window is arranged in chronological order, and a transition frame is inserted between adjacent windows to obtain a window sequence. The transition frame is generated based on the weighted fusion of the two consecutive windows.

[0031] Furthermore, before arranging each normalized window in chronological order, the process also includes:

[0032] If the offset of the current window from the previous window at the centroid of the spectrum exceeds an empirical threshold, then perform mild smoothing on the current window.

[0033] If the deviation of the current window from the historical window in terms of sparse pulse characteristics is greater than a preset deviation, a short-term buffer is triggered to avoid false alarms caused by occasional pulses.

[0034] Furthermore, after outputting the early warning signal, the method further includes:

[0035] Map the early fault risk value corresponding to the early warning signal to the corresponding original window sequence;

[0036] The original window sequence is fed in reverse into the reverse path of the phase encoder and the energy encoder. The phase encoder is used to calculate the first gradient magnitude of the instantaneous phase sequence layer by layer, and the energy encoder is used to calculate the second gradient magnitude of the energy spectrum layer by layer.

[0037] The first gradient magnitude and the second gradient magnitude are normalized channel by channel. If the first gradient magnitude or the second gradient magnitude is greater than the preset gradient magnitude in the same channel and at the same time point, it is marked that there is an anomaly in the frequency band of the channel.

[0038] If only the magnitude of the first gradient is greater than the preset gradient magnitude, it is marked as a phase abnormality but the energy is normal, indicating a mechanical phase drift fault.

[0039] If only the second gradient amplitude is greater than the preset gradient amplitude, it is marked as an energy anomaly but a normal phase, indicating a combustion efficiency-related fault.

[0040] Use natural language templates to convert the tagged results into readable exception explanations.

[0041] Secondly, the present invention also provides a vehicle engine operating status monitoring device, comprising:

[0042] The acquisition module is used to acquire monitoring signals of the vehicle engine in real time during its operation.

[0043] The generation module is used to perform sliding window slicing and normalization processing on the monitoring signal to generate a window sequence;

[0044] The feature output module is used to input the window sequence into the trained feature extraction model and output an instantaneous latent vector. The feature extraction model is composed of a lightweight convolutional neural network and a Transformer encoder cascaded together. It is used to capture subtle changes in the time domain, frequency domain and phase of early engine minor faults. The instantaneous latent vector is a high-order abstract representation used to compress and characterize the current health status features of the engine.

[0045] The early warning output module is used to input the instantaneous latent vector into the early warning module and output the early fault risk value. When the early fault risk values ​​corresponding to multiple consecutive window sequences are all greater than the dynamic threshold, an early warning signal is output. The early warning module includes an LSTM network and a fully connected layer, which is used to convert the instantaneous latent vector into an early fault risk value.

[0046] Thirdly, the present invention also provides a storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the vehicle engine operating status monitoring method as described in any of the preceding claims.

[0047] Fourthly, the present invention also provides a controller, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the vehicle engine operating status monitoring method described above.

[0048] The present invention provides a method, device, controller, and storage medium for monitoring the operating status of a vehicle engine. Through the combined effects of real-time synchronous sampling, sliding window, normalization, feature distillation using a convolutional neural network and Transformer encoder, trend tracking via an LSTM network, and dynamic threshold alarm, the detection granularity is advanced from second-level overall energy to millisecond-level transient morphology. This significantly lowers the detectable signal-to-noise ratio threshold, enabling the detection of subtle anomalies such as turbine blade microcracks, early bearing pitting, and localized combustion chamber misfires at a low energy and minute morphology stage. Furthermore, because the warning time is significantly advanced, the maintenance window extends from near-failure to early sub-health conditions, thereby avoiding catastrophic failure and addressing the pain point of traditional systems where alarms are immediately at a critical point. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating a method for monitoring the operating status of a vehicle engine according to an embodiment of the present invention.

[0050] Figure 2 This is a schematic block diagram of a vehicle engine operating status monitoring device according to an embodiment of the present invention;

[0051] Figure 3 This is a schematic block diagram of the controller according to an embodiment of the present invention.

[0052] The realization of the objectives, functional features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0054] refer to Figure 1 As shown, in one embodiment, the present invention provides a method for monitoring the operating status of a vehicle engine, the method comprising:

[0055] S11. Real-time acquisition of monitoring signals of the vehicle engine during operation;

[0056] S12. Perform sliding window slicing and normalization processing on the monitoring signal to generate a window sequence;

[0057] S13. Input the window sequence into the trained feature extraction model and output the instantaneous latent vector. The feature extraction model is composed of a lightweight convolutional neural network and a Transformer encoder cascaded together. It is used to capture the subtle changes in the time domain, frequency domain and phase of early engine minor faults. The instantaneous latent vector is a high-order abstract representation used to compress and characterize the current health status features of the engine.

[0058] S14. Input the instantaneous latent vector into the early warning module and output the early fault risk value. When the early fault risk values ​​corresponding to multiple consecutive window sequences are all greater than the dynamic threshold, output the early warning signal. The early warning module includes an LSTM network and a fully connected layer, which is used to convert the instantaneous latent vector into the early fault risk value.

[0059] Traditional monitoring systems often only collect a single physical quantity or use a fixed sampling pattern. Once operating conditions fluctuate or sensors drift, the signal baseline shifts as a whole, masking weak faults. The sampling strategy in this embodiment maintains dynamic synchronization with operating conditions such as engine speed and load, ensuring that the signal is always in a comparable relative coordinate system. This avoids baseline drift caused by sudden changes in operating conditions, thus preserving the true transient anomaly energy for subsequent monitoring.

[0060] The monitoring signals may include multi-channel high sampling rate signals such as vibration acceleration (turbine, bearing, cylinder block), cylinder pressure, instantaneous speed, injection pressure, and exhaust temperature.

[0061] The impact of a weak fault is extremely short-lived. If the entire signal is fed directly into the model, the transient energy will be averaged out by the long-term energy. The sliding window in this embodiment can cut the signal into short segments, locking the millisecond-level impact completely within a window; normalization smooths out the amplitude differences between different operating conditions and different sensors, allowing the model to focus only on shape rather than absolute amplitude. In this way, the window sequence becomes a series of local snapshots of the same scale, which amplifies the transient characteristics while shielding the overall drift of the operating conditions, solving the dilemma of false alarms when the threshold is adjusted and false alarms when it is not adjusted in the traditional threshold method.

[0062] Furthermore, traditional frequency domain features rely on manual frequency band selection, which becomes ineffective once the fault spectrum line overlaps with the noise spectrum line. In this embodiment, the local receptive field of the convolutional neural network excels at capturing the sharp form of the impact in the time domain; the self-attention mechanism of the Transformer encoder captures phase coupling relationships across channels and frequencies globally. After cascading, the model can simultaneously resolve the time-domain impact, the frequency-domain diffuse spectrum line, and the phase modulation in a single forward propagation, adjusting the early, weak fault from a chaotic background of "wideband, low amplitude" into a high-order instantaneous latent vector. This instantaneous latent vector is no longer bound by specific physical units but is instead an abstract measure of the healthy / sub-healthy state, thus completely eliminating the dependence on fixed frequency bands or fixed thresholds.

[0063] Furthermore, a single point of anomaly is insufficient to indicate that a fault is evolving. In this embodiment, the LSTM network performs temporal modeling of the latent vector sequence of consecutive windows, memorizing and tracking the evolution trend of fault characteristics; the fully connected layer maps the hidden state of the LSTM network to a continuously changing risk value. An alarm is triggered only when the early fault risk values ​​of multiple consecutive windows are all higher than a dynamic threshold. The "dynamic" aspect here is reflected in the fact that the threshold is adjusted in real time according to operating conditions and environmental noise, ensuring consistent early warning sensitivity under different speeds and loads. This avoids false alarms caused by traditional fixed thresholds when noise is slightly high, and also prevents missed alarms when the threshold is too high, leading to cracks expanding to critical sizes.

[0064] This embodiment utilizes a combination of real-time synchronous sampling, sliding windowing, normalization, feature distillation using a convolutional neural network and Transformer encoder, trend tracking via LSTM network, and dynamic threshold alarms to advance the detection granularity from second-level overall energy to millisecond-level transient morphology. This significantly lowers the detectable signal-to-noise ratio threshold, enabling the detection of subtle anomalies such as turbine blade microcracks, early bearing pitting, and localized combustion chamber misfires at a low energy and minute morphology stage. Furthermore, because the warning time is significantly advanced, the maintenance window extends from near-failure to early sub-health conditions, thus avoiding catastrophic failures and addressing the pain point of traditional systems where alarms are immediately at the critical point.

[0065] In one embodiment, before inputting the window sequence into the already trained feature extraction model, the method further includes:

[0066] Multiple monitoring signal samples are acquired, and a window sequence sample corresponding to each monitoring signal sample is generated. Under the condition of no manual labeling, a preset proportion of the window sequence samples are randomly occluded and reconstructed from the context to obtain the target window sequence sample.

[0067] The Transformer encoder is used to map each target window sequence sample into a low-dimensional feature vector to obtain the first instantaneous latent vector;

[0068] The first instantaneous latent vectors arranged in chronological order in the same running record are sequentially fed into the convolutional neural network to generate the second instantaneous latent vector. The second instantaneous latent vector contains the historical information of all target window sequence samples before each target window sequence sample.

[0069] Select the second instantaneous latent vector of the current target window sequence sample and the first instantaneous latent vector of the subsequent target window sequence sample that is adjacent in time to form a positive sample pair. Select the first instantaneous latent vector of the target window sequence sample that is not adjacent or originates from different running records within the same batch and form a negative sample pair with the current second instantaneous latent vector.

[0070] For each sample pair, the similarity score between the second instantaneous latent vector and the first instantaneous latent vector is calculated, and the parameters of the encoder and convolutional neural network are updated through backpropagation until the similarity score of the positive sample pair is maximized and the similarity score of the negative sample pair is minimized, thus generating a feature extraction model that has been trained.

[0071] This embodiment can segment a large number of historical monitoring signal samples into continuous small windows, forming a window sequence sample. For any given sample, without any manual annotation, a portion of the segment is randomly selected for masking, essentially removing the vibration or pressure pulse of the engine at a certain instant. The model then relies solely on the remaining context to infer the removed content. In this way, the network must learn to understand the temporal coherence, local morphology, and global structure of the signal, rather than simply memorizing numerical values.

[0072] The masked target window sequence samples are input into the Transformer encoder. The encoder uses a self-attention mechanism to compress the relationships between the sampling points within the window into a compact low-dimensional feature vector, called the first instantaneous latent vector, which records the key features of the window in the time domain, frequency domain, and phase domain, but does not contain any artificial labels.

[0073] The first instantaneous latent vectors, arranged chronologically from the same run record, are sequentially fed into a convolutional neural network. The convolutional kernel slides along the time dimension, gradually accumulating information from past vectors into the current vector, outputting the second instantaneous latent vector. At this point, each second instantaneous latent vector not only contains the instantaneous features of the current window but also incorporates the evolutionary trends of all previous windows, essentially giving the model a memory.

[0074] For any current target window sequence sample, take its second instantaneous latent vector, and then take the first instantaneous latent vector of its next adjacent target window sequence sample to form a positive sample pair—they are continuous in physical time and should be similar to each other.

[0075] Meanwhile, within the same batch, target window sequence samples with long time intervals or from completely different running records are randomly selected, and their first instantaneous latent vector is paired with the second instantaneous latent vector to form negative sample pairs—they are physically unrelated and should not be similar.

[0076] The similarity score for each pair of latent vectors is calculated, and a loss function is used to maximize the score of positive sample pairs and minimize the score of negative sample pairs. Backpropagation updates all parameters of the Transformer encoder and convolutional network, forcing the model to learn to distinguish between naturally continuous and randomly unrelated signal segments. After multiple iterations, the model automatically extracts a set of feature representations that are highly sensitive to engine health status but robust to operating condition fluctuations and environmental noise, forming the final feature extraction model.

[0077] This embodiment can learn features that are highly sensitive to weak anomalies through random masking, reconstruction, and contrastive learning, without any manual fault labeling. The convolutional neural network accumulates historical information layer by layer into the current latent vector, so that the evolution trend of early weak faults is encoded in the features, avoiding single-point misjudgment. In addition, the design of positive and negative sample pairs forces the model to pay attention to local and global consistency, and has natural immunity to sensor drift and changes in operating conditions. After training, the entire feature extraction model can be seamlessly connected to the operation monitoring pipeline, outputting instantaneous latent vectors in real time, directly supporting subsequent early warning or health assessment modules without the need for re-labeling or fine-tuning.

[0078] In one embodiment, after randomly occluding a preset proportion of window sequence samples and reconstructing the target window sequence samples from the context, the method may further include:

[0079] According to the real-time engine operating conditions, phase drift and energy decay are injected into each window sequence sample to generate pseudo-fault window sequence samples with consistent operating conditions. The pseudo-fault window sequence samples are added to the target window sequence samples to form a training set. The phase drift is used to superimpose a micro-offset on the instantaneous phase, and the energy decay is used to introduce narrowband attenuation on the energy spectrum.

[0080] The system synchronously acquires engine operating condition information such as current speed, load, and temperature, which serves as an operating condition fingerprint. A micro-offset corresponding to the operating condition fingerprint is superimposed on the instantaneous phase of the same window sequence samples to simulate phase jitter caused by early cracks, pitting, etc. The offset is adaptively adjusted according to changes in operating conditions to simulate random fluctuations generated under real-world conditions. Several narrow bands are selected on the energy spectrum, and their amplitude is reduced according to the attenuation coefficient given by the operating condition fingerprint to reproduce the weak energy leakage caused by early bearing wear and localized misfires in the combustion chamber. The generated pseudo-fault window sequence samples are mixed with the original target window sequence samples to form a large-scale unlabeled training set with consistent operating conditions and controllable fault characteristics.

[0081] In this embodiment, both drift and decay are coupled to real-time operating conditions, ensuring that pseudo-faults exist under the same background noise and modulation conditions as the real operating environment, thus avoiding mismatch between the training distribution and the online distribution. A large number of early, weak fault samples are actively inserted into the training pool, which originally contained only healthy samples. During the contrastive learning phase, the model is forced to learn to distinguish subtle differences between healthy and sub-healthy conditions, significantly improving its sensitivity to real early faults. Furthermore, since the injected drift and decay follow a condition-driven stochastic strategy, the model can maintain stable recognition capabilities even when faced with unprecedented faults after going live, without the need for repeated retraining or manual parameter tuning.

[0082] In one embodiment, after generating the trained feature extraction model, the method further includes:

[0083] When each new batch window sequence arrives, the third instantaneous latent vector is predicted using the current weight vector of the convolutional neural network, and the difference between the prediction and the actual observed fourth instantaneous latent vector is calculated to obtain the observation residual.

[0084] The residual covariance is calculated based on the observed residuals and cross-referenced with the current system state covariance of the convolutional neural network to obtain the optimal correction amount.

[0085] The optimal correction is applied to the current weight vector of the convolutional neural network to obtain the updated weight vector, and the current system state covariance is updated to reflect the new confidence level of the weight vector estimation.

[0086] The updated weight vector is immediately written back to the convolutional neural network, and the next batch of window sequences is processed.

[0087] When a new batch of window sequences arrives, a forward prediction is first performed using the weight vector currently stored in the convolutional neural network of the feature extraction model that has been trained, to obtain the third instantaneous latent vector; at the same time, the fourth instantaneous latent vector that is actually observed is generated. The former is regarded as the model prediction and the latter is regarded as the actual observation. The difference between the two can be used to obtain the observation residual.

[0088] The residual covariance is obtained by multiplying the observed residual by itself. This residual covariance quantifies the spatial distribution of the prediction error of the current weights on this batch of data. The larger the residual, the further the weights deviate from the true distribution.

[0089] By performing a cross-operation between the residual covariance and the system state covariance maintained internally by the convolutional neural network, an optimal correction amount that fuses prediction and observation confidence is obtained. When observation noise is low and residuals are large, the correction amount is amplified; when the system state covariance itself is already quite deterministic, the correction amount is reduced, thereby avoiding oscillations.

[0090] The system state covariance is a quantitative representation of the uncertainty in the current estimate of the model weight vector. By straightening out all the weights of the convolutional neural network, it can be viewed as a high-dimensional state vector. The system state covariance is a symmetric matrix, where each element on the diagonal represents the estimated variance of the corresponding weight—the larger the value, the less confident the convolutional neural network is in the true value of that weight. The off-diagonal elements represent the correlation between the estimation errors of different weights. During the recursive update process, it is cross-operated with the observation residuals to determine the magnitude and direction of the optimal correction.

[0091] If the system state covariance is large (due to uncertainty about the weights), the convolutional neural network is more willing to believe new observations, resulting in a larger correction amount.

[0092] If the system state covariance is small (the weights are very certain), the convolutional neural network is more conservative and the correction amount becomes smaller.

[0093] Each time the weights are updated, the system state covariance is also updated synchronously.

[0094] Finally, the current weight vector is fine-tuned using the optimal correction amount to obtain the updated weight vector. At the same time, the system state covariance is contracted or expanded using the same correction amount, which represents a reassessment of the confidence in the weight estimation. The updated weight vector immediately covers the original weights of the convolutional neural network. When the next batch of window sequences arrives, the above steps are repeated to form a closed-loop online adaptive system.

[0095] This embodiment eliminates the need for offline retraining or checkpoint saving. The network continuously evolves with the data stream, ensuring the monitoring system maintains up-to-date identification capabilities over the long term. Each update directly targets the prediction error, with precise weight adjustments to avoid oscillations or forgetting caused by blind gradient descent. Furthermore, the system state covariance reflects confidence in the weight estimation in real time. When data quality deteriorates or sudden changes in operating conditions lead to increased residuals, the covariance automatically amplifies to prevent over-reliance on unreliable observations; conversely, it results in faster convergence.

[0096] In one embodiment, calculating the similarity score between the second instantaneous latent vector and the first instantaneous latent vector includes:

[0097] Both the second instantaneous latent vector and the first instantaneous latent vector are fed into the phase channel and the energy channel, and the first phase latent vector and the first energy latent vector corresponding to the second instantaneous latent vector are output, as well as the second phase latent vector and the second energy latent vector corresponding to the first instantaneous latent vector. The phase channel is used to perform Hilbert transform on the second instantaneous latent vector or the first instantaneous latent vector to extract the instantaneous phase sequence, and obtain the phase latent vector through a causal dilated convolutional network. The energy channel is used to perform continuous wavelet transform on the second instantaneous latent vector or the first instantaneous latent vector to extract the energy spectrum, and obtain the energy latent vector through an asymmetric convolutional neural network.

[0098] Calculate the first cosine distance between the first phase latent vector and the second phase latent vector, and calculate the second cosine distance between the first energy latent vector and the second energy latent vector;

[0099] The similarity score is obtained by weighted summation of the first cosine distance and the second cosine distance.

[0100] In this embodiment, the second instantaneous latent vector and the first instantaneous latent vector can each be copied and sent to two independent channels respectively:

[0101] Phase channel: First, perform Hilbert transform to obtain the instantaneous phase sequence; then use a causal dilated convolutional network to expand the receptive field while maintaining temporal causality, compressing the phase sequence into a phase latent vector.

[0102] Energy Channel: First, perform continuous wavelet transform to generate an energy spectrum; then use an asymmetric convolutional neural network (with irregular kernel size and left-right asymmetry) to capture local strength distributions on the frequency and time planes to obtain the energy latent vector.

[0103] In this way, each latent vector is split into two complementary sub-vectors: one focusing on phase dynamics and the other focusing on energy distribution.

[0104] The first cosine distance is calculated for the two phase latent vectors—only the directional difference is considered, the magnitude is ignored, and the degree of phase alignment is emphasized. At the same time, the second cosine distance is calculated for the two energy latent vectors—again, only the shape difference of the energy distribution is considered, and it is not affected by the overall amplitude.

[0105] The first cosine distance and the second cosine distance are linearly added with adjustable weights to obtain the final similarity score. The weights can be adaptively learned during training or manually set according to business needs, thus flexibly balancing the contributions of phase consistency and energy consistency to the overall similarity.

[0106] In this embodiment, phase and energy are separated, allowing the convolutional neural network to utilize both high-resolution phase information and wideband energy information simultaneously, avoiding mutual interference. The cosine distance is insensitive to amplitude changes, ensuring stable measurement of true similarity even with sensor gain drift or overall energy variations. By adjusting the weights, it's possible to intuitively control whether phase synchronization or energy spectrum shape is prioritized, making the similarity score more closely reflect the specific fault mode.

[0107] In one embodiment, the step of performing sliding window slicing and normalization processing on the monitoring signal to generate a window sequence includes:

[0108] The monitoring signal is mirrored to generate a symmetrical waveform. The drift component is obtained by fitting a quadratic curve based on the local extreme points of the symmetrical waveform. The drift component is then superimposed on the symmetrical waveform to obtain the extended signal.

[0109] Calculate the instantaneous energy envelope of the extended signal, detect the sparse pulses of the instantaneous energy envelope, and if there is an instantaneous energy envelope interval that is lower than the global median and has a sparse pulse density greater than the neighborhood mean, then mark it as a suspected signal band.

[0110] The complexity of the local signal is determined based on the local gradient density of the instantaneous energy envelope. Using a variable-length primitive window as a seed, the suspected signal band is lengthened or shortened based on the complexity of the local signal to obtain the target signal band.

[0111] Each window corresponding to the target signal band is divided into three segments: front, middle, and back. Max or minimum normalization based on local extrema is performed on the front and back segments to suppress endpoint drift. Mean and variance normalization based on instantaneous energy envelope is performed on the middle segment to highlight fault-related fluctuations. The normalized results of the three segments are then spliced ​​together to form a complete window.

[0112] Each normalized window is arranged in chronological order, and a transition frame is inserted between adjacent windows to obtain a window sequence. The transition frame is generated based on the weighted fusion of the two consecutive windows.

[0113] This embodiment mirrors the original monitoring signal left and right to form a symmetrical waveform. Local extrema are found on the symmetrical waveform, and a smooth drift component is obtained by fitting these extrema with a quadratic curve. The drift component is then superimposed back onto the symmetrical waveform to generate an extended signal that is longer than the original signal. This process extrapolates the signal trend in the time domain, eliminating the need for manual truncation in subsequent processing and reducing boundary distortion.

[0114] The instantaneous energy envelope of the extended signal is calculated—that is, the smoothed curve of the square of the instantaneous amplitude of the time-domain waveform. The instantaneous energy envelope is scanned, and if an interval is found where the amplitude is lower than the global median, and the density of sparse pulses (short-duration high-amplitude spikes) is greater than the neighborhood mean, this interval is marked as a suspected signal band to select regions with low energy but occasional pulses.

[0115] The local gradient density of the instantaneous energy envelope is calculated as an indicator of local signal complexity. A variable-length primitive window is used as a seed: the primitive window is shortened in regions of high complexity to preserve fine-grained fault transients; the primitive window is lengthened in regions of low complexity to smooth redundant fluctuations. The primitive window slides and adaptively expands and contracts on the suspected signal band, ultimately forming the target signal band. Thus, the window boundary is no longer fixed but dynamically adjusted according to the signal content.

[0116] Divide each window within the target signal band into three segments: front, middle, and back.

[0117] Front and back segments: Local extrema are used for maximum or minimum normalization to compress endpoint drift to a uniform scale and prevent jumps during window splicing.

[0118] Middle section: By normalizing the mean and variance of the instantaneous energy envelope, energy fluctuations are amplified to a uniform variance level, highlighting potential fault-related transients.

[0119] After normalization of the three segments, they are reassembled into a complete window, taking into account both endpoint stability and mid-segment sensitivity.

[0120] Transition frames are inserted between the normalized windows arranged in chronological order. Each transition frame is formed by weighted fusion of the end of the previous window and the beginning of the next window, with the weights transitioning smoothly over time. In this way, the window sequence presents a "soft connection" in the time domain, eliminating abrupt changes caused by window sliding and ensuring that the sequence received by the subsequent model is statistically continuous and smooth.

[0121] In this embodiment, mirror flipping and drift compensation eliminate endpoint distortion caused by hard truncation, allowing early weak transients to be fully preserved. Dual gating based on sparse impulses and energy envelopes ensures that only low-energy segments truly carrying fault information are retained. A complexity-driven variable primitive window allows the window length to scale with the signal content, avoiding both excessive smoothing and excessive fragmentation. Furthermore, a three-segment strategy suppresses drift while amplifying fault fluctuations, ensuring the subsequent feature extraction model always faces high-contrast input. Transition frames eliminate step transitions between windows, eliminating the need for additional handling of discontinuities during training or inference, thus improving stability and generalization ability.

[0122] In one embodiment, before arranging each normalized window in chronological order, the method further includes:

[0123] If the offset of the current window from the previous window at the centroid of the spectrum exceeds an empirical threshold, then perform mild smoothing on the current window.

[0124] If the deviation of the current window from the historical window in terms of sparse pulse characteristics is greater than a preset deviation, a short-term buffer is triggered to avoid false alarms caused by occasional pulses.

[0125] When detecting the shift in the spectral centroid, a short-time Fourier transform is first performed on the current window to calculate its spectral centroid. This is then compared to the spectral centroid of the previous window to obtain the shift in the centroid's position. If this shift exceeds an empirical threshold, it indicates a significant jump in the frequency energy distribution between the two windows. At this point, a mild smoothing is triggered: locally weighted regression is used to slightly smooth the amplitude curve of the current window, suppressing high-frequency spikes while preserving the main contour of the fault transient.

[0126] When performing deviation detection on sparse pulses, the sparse pulse features (pulse amplitude, density, and time interval distribution) of the current window are first extracted and compared with the rolling statistical baseline constructed from historical windows to calculate the deviation. If the deviation exceeds a preset deviation, a short-term buffer is triggered: the current window is not enqueued but placed in a sliding buffer; subsequent windows continue to be received, and the buffer is released as a whole after the pulse features in the buffer re-match the baseline or after several consecutive windows have stabilized. During the buffering period, the model does not respond immediately to sudden pulses to prevent occasional impacts from being misjudged as faults.

[0127] In this embodiment, mild smoothing intervenes only when the center of gravity shift exceeds the limit, eliminating false alarms caused by abnormal frequency band abrupt changes without excessively sacrificing effective frequency information. Short-term buffering isolates sporadic pulses outside the decision chain, making the system more sensitive to the continuous evolution of real early faults and more immune to one-off shocks. In addition, both judgments are based on real-time sliding statistics, eliminating the need for manual threshold reset and automatically maintaining stable performance under different operating conditions.

[0128] In one embodiment, after outputting the early warning signal, the method further includes:

[0129] Map the early fault risk value corresponding to the early warning signal to the corresponding original window sequence;

[0130] The original window sequence is fed in reverse into the reverse path of the phase encoder and the energy encoder. The phase encoder is used to calculate the first gradient magnitude of the instantaneous phase sequence layer by layer, and the energy encoder is used to calculate the second gradient magnitude of the energy spectrum layer by layer.

[0131] The first gradient magnitude and the second gradient magnitude are normalized channel by channel. If the first gradient magnitude or the second gradient magnitude is greater than the preset gradient magnitude in the same channel and at the same time point, it is marked that there is an anomaly in the frequency band of the channel.

[0132] If only the magnitude of the first gradient is greater than the preset gradient magnitude, it is marked as a phase abnormality but the energy is normal, indicating a mechanical phase drift fault.

[0133] If only the second gradient amplitude is greater than the preset gradient amplitude, it is marked as an energy anomaly but a normal phase, indicating a combustion efficiency-related fault.

[0134] Use natural language templates to convert the tagged results into readable exception explanations.

[0135] The early warning signal, carrying an early fault risk value, is directly mapped back to the original window sequence that triggered that risk value, ensuring accurate subsequent tracing. This original window sequence is then fed back into the phase encoder and energy encoder along the original path. Both encoders retain intermediate feature maps from each layer during forward propagation for gradient calculation.

[0136] The phase encoder takes the instantaneous phase sequence as input and calculates the derivative layer by layer in reverse along the network layer to obtain the first gradient magnitude; the larger the value, the more sensitive the layer is to the phase change at that time point.

[0137] The energy encoder takes the energy spectrum as input and calculates the derivative layer by layer in reverse along the network to obtain the second gradient amplitude; the larger the value, the more sensitive the layer is to energy changes in that frequency band.

[0138] The two gradient magnitude sequences are normalized channel by channel to eliminate dimensional differences; then a preset gradient magnitude is set. If the first or second gradient magnitude is greater than the preset gradient magnitude in the same channel and at the same time point, an anomaly is marked in the frequency band of that channel.

[0139] If only the magnitude of the first gradient is greater than the preset gradient magnitude, it is marked as a phase abnormality but the energy is normal, indicating a mechanical phase drift fault.

[0140] If only the second gradient amplitude is greater than the preset gradient amplitude, it is marked as an energy anomaly but a normal phase, indicating a combustion efficiency-related fault.

[0141] Using a preset template, the above-marked results are automatically filled into the "Time-Frequency-Abnormality Type-Possible Cause" field to generate a readable natural language explanation, such as "Around a certain second, an energy abnormality occurs in the high-frequency band but the phase is normal, which is suspected to be a decrease in combustion efficiency."

[0142] This embodiment uses inverse gradient descent to reconstruct abstract early fault risk values ​​to specific time and frequency bands, allowing maintenance personnel to directly locate suspicious components or operating conditions. Simultaneously, through independent dual-channel judgment using phase encoders and energy encoders, automatic classification of early fault types is achieved, avoiding the need for engineers to troubleshoot item by item. Furthermore, natural language templates translate technical details into plain language, enabling field personnel to understand and perform subsequent maintenance without requiring a deep learning background.

[0143] refer to Figure 2 As shown in the figure, this embodiment of the invention also provides a vehicle engine operating status monitoring device, the device comprising:

[0144] The acquisition module 11 is used to acquire the monitoring signals of the vehicle engine in real time when it is running.

[0145] The generation module 12 is used to perform sliding window slicing and normalization processing on the monitoring signal to generate a window sequence;

[0146] The feature output module 13 is used to input the window sequence into the trained feature extraction model and output the instantaneous latent vector. The feature extraction model is composed of a lightweight convolutional neural network and a Transformer encoder cascaded together. It is used to capture the subtle changes in the time domain, frequency domain and phase of early engine minor faults. The instantaneous latent vector is a high-order abstract representation used to compress and characterize the current health status features of the engine.

[0147] The early warning output module 14 is used to input the instantaneous latent vector into the early warning module and output the early fault risk value. When the early fault risk values ​​corresponding to multiple consecutive window sequences are all greater than the dynamic threshold, an early warning signal is output. The early warning module includes an LSTM network and a fully connected layer, which is used to convert the instantaneous latent vector into an early fault risk value.

[0148] As described above, it can be understood that each component of the vehicle engine operating status monitoring device proposed in this invention can realize the function of any of the vehicle engine operating status monitoring methods described above, and the specific structure will not be described in detail.

[0149] refer to Figure 3 As shown in the figure, an embodiment of the present invention also provides a controller, the internal structure of which can be as follows: Figure 3 As shown, the controller includes a processor, memory, network interface, and database connected via a system bus. The processor is designed to provide computing and control capabilities. The controller's memory includes a storage medium and internal memory. The storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the storage medium. The controller's database stores relevant data for the vehicle engine's operating status monitoring method. The controller's network interface is used for communication with external electronic devices via a network connection. When the computer program is executed by the processor, it implements a vehicle engine operating status monitoring method.

[0150] In one embodiment of the present invention, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements a method for monitoring the operating status of a vehicle engine.

[0151] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media provided and used in the present invention and embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0152] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0153] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for monitoring the operating status of a vehicle engine, characterized in that, include: Real-time acquisition of monitoring signals from the vehicle engine during operation; The monitoring signal is sliced ​​using a sliding window and normalized to generate a window sequence; The window sequence is input into the trained feature extraction model, which outputs an instantaneous latent vector. The feature extraction model is composed of a lightweight convolutional neural network and a Transformer encoder, which is used to capture subtle changes in the time domain, frequency domain and phase of early engine minor faults. The instantaneous latent vector is a high-order abstract representation used to compress and characterize the current health status features of the engine. The instantaneous latent vector is input into the early warning module, which outputs an early fault risk value. When the early fault risk values ​​corresponding to multiple consecutive window sequences are all greater than the dynamic threshold, an early warning signal is output. The early warning module includes an LSTM network and a fully connected layer, which are used to convert the instantaneous latent vector into an early fault risk value. Before inputting the window sequence into the already trained feature extraction model, the method further includes: Multiple monitoring signal samples are acquired, and a window sequence sample corresponding to each monitoring signal sample is generated. Under the condition of no manual labeling, a preset proportion of the window sequence samples are randomly occluded and reconstructed from the context to obtain the target window sequence sample. The Transformer encoder is used to map each target window sequence sample into a low-dimensional feature vector to obtain the first instantaneous latent vector; The first instantaneous latent vectors arranged in chronological order in the same running record are sequentially fed into the convolutional neural network to generate the second instantaneous latent vector. The second instantaneous latent vector contains the historical information of all target window sequence samples before each target window sequence sample. Select the second instantaneous latent vector of the current target window sequence sample and the first instantaneous latent vector of the subsequent target window sequence sample that is adjacent in time to form a positive sample pair. Select the first instantaneous latent vector of the target window sequence sample that is not adjacent or originates from different running records within the same batch and form a negative sample pair with the current second instantaneous latent vector. For each sample pair, the similarity score between the second instantaneous latent vector and the first instantaneous latent vector is calculated, and the parameters of the encoder and convolutional neural network are updated through backpropagation until the similarity score of the positive sample pair is maximized and the similarity score of the negative sample pair is minimized, thus generating a feature extraction model that has been trained. The calculation of the similarity score between the second instantaneous latent vector and the first instantaneous latent vector includes: Both the second instantaneous latent vector and the first instantaneous latent vector are fed into the phase channel and the energy channel, and the first phase latent vector and the first energy latent vector corresponding to the second instantaneous latent vector are output, as well as the second phase latent vector and the second energy latent vector corresponding to the first instantaneous latent vector. The phase channel is used to perform Hilbert transform on the second instantaneous latent vector or the first instantaneous latent vector to extract the instantaneous phase sequence, and obtain the phase latent vector through a causal dilated convolutional network. The energy channel is used to perform continuous wavelet transform on the second instantaneous latent vector or the first instantaneous latent vector to extract the energy spectrum, and obtain the energy latent vector through an asymmetric convolutional neural network. Calculate the first cosine distance between the first phase latent vector and the second phase latent vector, and calculate the second cosine distance between the first energy latent vector and the second energy latent vector; The similarity score is obtained by weighted summation of the first cosine distance and the second cosine distance.

2. The method for monitoring the operating status of a vehicle engine according to claim 1, characterized in that, After generating the trained feature extraction model, the process further includes: When each new batch window sequence arrives, the third instantaneous latent vector is predicted using the current weight vector of the convolutional neural network, and the difference between the prediction and the actual observed fourth instantaneous latent vector is calculated to obtain the observation residual. The residual covariance is calculated based on the observed residuals and cross-referenced with the current system state covariance of the convolutional neural network to obtain the optimal correction amount. The optimal correction is applied to the current weight vector of the convolutional neural network to obtain the updated weight vector, and the current system state covariance is updated to reflect the new confidence level of the weight vector estimation. The updated weight vector is immediately written back to the convolutional neural network, and the next batch of window sequences is processed.

3. The method for monitoring the operating status of a vehicle engine according to claim 1, characterized in that, The step of performing sliding window slicing and normalization processing on the monitoring signal to generate a window sequence includes: The monitoring signal is mirrored to generate a symmetrical waveform. The drift component is obtained by fitting a quadratic curve based on the local extreme points of the symmetrical waveform. The drift component is then superimposed on the symmetrical waveform to obtain the extended signal. Calculate the instantaneous energy envelope of the extended signal, detect the sparse pulses of the instantaneous energy envelope, and if there is an instantaneous energy envelope interval that is lower than the global median and has a sparse pulse density greater than the neighborhood mean, then mark it as a suspected signal band. The complexity of the local signal is determined based on the local gradient density of the instantaneous energy envelope. Using a variable-length primitive window as a seed, the suspected signal band is lengthened or shortened based on the complexity of the local signal to obtain the target signal band. Each window corresponding to the target signal band is divided into three segments: front, middle, and back. Max or minimum normalization based on local extrema is performed on the front and back segments to suppress endpoint drift. Mean and variance normalization based on instantaneous energy envelope is performed on the middle segment to highlight fault-related fluctuations. The normalized results of the three segments are then spliced ​​together to form a complete window. Each normalized window is arranged in chronological order, and a transition frame is inserted between adjacent windows to obtain a window sequence. The transition frame is generated based on the weighted fusion of the two consecutive windows.

4. The method for monitoring the operating status of a vehicle engine according to claim 3, characterized in that, Before arranging each normalized window in chronological order, the process also includes: If the offset of the current window from the previous window at the centroid of the spectrum exceeds an empirical threshold, then perform mild smoothing on the current window. If the deviation of the current window from the historical window in terms of sparse pulse characteristics is greater than a preset deviation, a short-term buffer is triggered to avoid false alarms caused by occasional pulses.

5. The method for monitoring the operating status of a vehicle engine according to claim 1, characterized in that, After outputting the early warning signal, the system also includes: Map the early fault risk value corresponding to the early warning signal to the corresponding original window sequence; The original window sequence is fed in reverse into the reverse path of the phase encoder and the energy encoder. The phase encoder is used to calculate the first gradient magnitude of the instantaneous phase sequence layer by layer, and the energy encoder is used to calculate the second gradient magnitude of the energy spectrum layer by layer. The first gradient magnitude and the second gradient magnitude are normalized channel by channel. If the first gradient magnitude or the second gradient magnitude is greater than the preset gradient magnitude in the same channel and at the same time point, it is marked that there is an anomaly in the frequency band of the channel. If only the magnitude of the first gradient is greater than the preset gradient magnitude, it is marked as a phase abnormality but the energy is normal, indicating a mechanical phase drift fault. If only the second gradient amplitude is greater than the preset gradient amplitude, it is marked as an energy anomaly but a normal phase, indicating a combustion efficiency-related fault. Use natural language templates to convert the tagged results into readable exception explanations.

6. A device for monitoring the operating status of a vehicle engine, characterized in that, include: The acquisition module is used to acquire monitoring signals of the vehicle engine in real time during its operation. The generation module is used to perform sliding window slicing and normalization processing on the monitoring signal to generate a window sequence; The feature output module is used to input the window sequence into the trained feature extraction model and output an instantaneous latent vector. The feature extraction model is composed of a lightweight convolutional neural network and a Transformer encoder cascaded together. It is used to capture subtle changes in the time domain, frequency domain and phase of early engine minor faults. The instantaneous latent vector is a high-order abstract representation used to compress and characterize the current health status features of the engine. The early warning output module is used to input the instantaneous latent vector into the early warning module and output the early fault risk value. When the early fault risk values ​​corresponding to multiple consecutive window sequences are all greater than the dynamic threshold, an early warning signal is output. The early warning module includes an LSTM network and a fully connected layer, which is used to convert the instantaneous latent vector into an early fault risk value. Before inputting the window sequence into the already trained feature extraction model, the method further includes: Multiple monitoring signal samples are acquired, and a window sequence sample corresponding to each monitoring signal sample is generated. Under the condition of no manual labeling, a preset proportion of the window sequence samples are randomly occluded and reconstructed from the context to obtain the target window sequence sample. The Transformer encoder is used to map each target window sequence sample into a low-dimensional feature vector to obtain the first instantaneous latent vector; The first instantaneous latent vectors arranged in chronological order in the same running record are sequentially fed into the convolutional neural network to generate the second instantaneous latent vector. The second instantaneous latent vector contains the historical information of all target window sequence samples before each target window sequence sample. Select the second instantaneous latent vector of the current target window sequence sample and the first instantaneous latent vector of the subsequent target window sequence sample that is adjacent in time to form a positive sample pair. Select the first instantaneous latent vector of the target window sequence sample that is not adjacent or originates from different running records within the same batch and form a negative sample pair with the current second instantaneous latent vector. For each sample pair, the similarity score between the second instantaneous latent vector and the first instantaneous latent vector is calculated, and the parameters of the encoder and convolutional neural network are updated through backpropagation until the similarity score of the positive sample pair is maximized and the similarity score of the negative sample pair is minimized, thus generating a feature extraction model that has been trained. The calculation of the similarity score between the second instantaneous latent vector and the first instantaneous latent vector includes: Both the second instantaneous latent vector and the first instantaneous latent vector are fed into the phase channel and the energy channel, and the first phase latent vector and the first energy latent vector corresponding to the second instantaneous latent vector are output, as well as the second phase latent vector and the second energy latent vector corresponding to the first instantaneous latent vector. The phase channel is used to perform Hilbert transform on the second instantaneous latent vector or the first instantaneous latent vector to extract the instantaneous phase sequence, and obtain the phase latent vector through a causal dilated convolutional network. The energy channel is used to perform continuous wavelet transform on the second instantaneous latent vector or the first instantaneous latent vector to extract the energy spectrum, and obtain the energy latent vector through an asymmetric convolutional neural network. Calculate the first cosine distance between the first phase latent vector and the second phase latent vector, and calculate the second cosine distance between the first energy latent vector and the second energy latent vector; The similarity score is obtained by weighted summation of the first cosine distance and the second cosine distance.

7. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for monitoring the operating status of a vehicle engine as described in any one of claims 1 to 5.

8. A controller, characterized in that, include: processor; Memory; The memory stores a computer program, and when the processor executes the computer program, it implements the vehicle engine operating status monitoring method according to any one of claims 1 to 5.

Images