CN-121327466-B - Vehicle engine running state monitoring method, device, controller and storage medium

CN121327466BCN 121327466 BCN121327466 BCN 121327466BCN-121327466-B

Abstract

The invention provides a method, a device, a controller and a storage medium for monitoring the running state of a vehicle engine, wherein the method comprises the steps of acquiring a monitoring signal of the vehicle engine in the running state in real time, carrying out sliding window slicing and normalization processing on the monitoring signal to generate a window sequence, inputting the window sequence into a feature extraction model which is trained completely, outputting an instantaneous latent vector, wherein the feature extraction model is used for capturing the slight change of early weak faults of the engine in time domain, frequency domain and phase, the instantaneous latent vector is used for compressing and representing the current health state feature of the engine, outputting early fault risk values through an early warning module, and outputting early warning signals when a plurality of continuous early fault risk values are larger than a dynamic threshold value. The invention realizes the propulsion of the detection granularity from the second-level integral energy to the millisecond-level transient state form, greatly reduces the detectable signal-to-noise ratio threshold, and remarkably improves the capturing capability of early weak faults of the engine.

Inventors

CHEN ZHENGQUAN
XIAN ZHIGANG

Assignees

广州万震达动力科技有限公司

Dates

Publication Date: 20260508
Application Date: 20251011

Claims (8)

1. A method of monitoring an operating state of an engine of a vehicle, comprising: acquiring monitoring signals of a vehicle engine in a running state in real time; sliding window slicing and normalization processing are carried out on the monitoring signals, and window sequences are generated; Inputting the window sequence into a feature extraction model which is trained, and outputting an instantaneous latent vector, wherein the feature extraction model is formed by cascading a lightweight convolutional neural network and a transducer encoder and is used for capturing the subtle changes of early weak faults of an engine in time domain, frequency domain and phase, and the instantaneous latent vector is a high-order abstract representation and is used for compressing and representing the current health state features of the engine; Inputting the instantaneous latent vector into an early warning module, outputting an early fault risk value, and outputting an early warning signal when the early fault risk values corresponding to a plurality of continuous window sequences are all larger than a dynamic threshold value, wherein the early warning module comprises an LSTM network and a full-connection layer and is used for converting the instantaneous latent vector into the early fault risk value; wherein before the window sequence is input into the feature extraction model which is already trained, the method further comprises: Acquiring a plurality of monitoring signal samples, generating window sequence samples corresponding to each monitoring signal sample, randomly shielding window sequence samples of a preset proportion under the condition of no artificial tag, and reconstructing the window sequence samples by context to obtain target window sequence samples; mapping each target window sequence sample into a low-dimensional feature vector by using a transducer encoder to obtain a first instantaneous latent vector; Sequentially sending the first instantaneous latent vectors arranged in time sequence in the same operation record into a convolutional neural network to generate a second instantaneous latent vector, wherein the second instantaneous latent vector contains the history information of all target window sequence samples before each target window sequence sample; Selecting a second instantaneous latent vector of a current target window sequence sample and a first instantaneous latent vector of a time adjacent subsequent target window sequence sample to form a positive sample pair, and selecting a first instantaneous latent vector of a target window sequence sample which is non-adjacent or is derived from different operation records in the same batch and a current second instantaneous latent vector to form a negative sample pair; For each sample pair, calculating the similarity score of the second instantaneous latent vector and the first instantaneous latent vector, updating the parameters of the encoder and the convolutional neural network through back propagation until the similarity score of the positive sample pair is maximized and the similarity score of the negative sample pair is minimized, and generating a feature extraction model after training; Wherein said calculating a similarity score for the second instantaneous latent vector and the first instantaneous latent vector comprises: The method comprises the steps of sending a second instantaneous latent vector and a first instantaneous latent vector into a phase channel and an energy channel, outputting a first phase latent vector and a first energy latent vector corresponding to the second instantaneous latent vector, and a second phase latent vector and a second energy latent vector corresponding to the first instantaneous latent vector, wherein the phase channel is used for carrying out Hilbert transformation on the second instantaneous latent vector or the first instantaneous latent vector, extracting an instantaneous phase sequence, obtaining the phase latent vector through a causal expansion convolution network, and the energy channel is used for carrying out continuous wavelet transformation on the second instantaneous latent vector or the first instantaneous latent vector, extracting an energy spectrogram, and obtaining the energy latent vector through an asymmetric convolution neural network; calculating a first cosine distance between the first phase latent vector and the second phase latent vector, and calculating a second cosine distance between the first energy latent vector and the second energy latent vector; And carrying out weighted summation on the first cosine distance and the second cosine distance to obtain a similarity score.
2. The method for monitoring the running state of the vehicle engine according to claim 1, further comprising, after the generating the feature extraction model that has completed training: when each new batch of window sequences arrives, predicting a third instantaneous latent vector by using the current weight vector of the convolutional neural network, and calculating a difference between the third instantaneous latent vector and a fourth instantaneous latent vector actually observed to obtain an observation residual; Calculating residual covariance according to the observation residual, and intersecting with the current system state covariance of the convolutional neural network to obtain an optimal correction amount; The optimal correction amount is acted on the current weight vector of the convolutional neural network to obtain an updated weight vector, and meanwhile, the current system state covariance is updated to reflect the new confidence coefficient of the weight vector estimation; and immediately writing the updated weight vector back to the convolution neural network, and continuing to process the window sequence of the next batch.
3. The method for monitoring the running state of the vehicle engine according to claim 1, wherein the sliding window slicing and normalization process is performed on the monitoring signal to generate a window sequence, comprising: Performing mirror image overturning on the monitoring signal to generate a symmetrical waveform, obtaining a drift component according to quadratic curve fitting of local extremum points of the symmetrical waveform, and superposing the drift component on the symmetrical waveform to obtain an extension signal; calculating the instantaneous energy envelope of the extended signal, detecting sparse pulses of the instantaneous energy envelope, and marking the instantaneous energy envelope as a suspected signal band if an instantaneous energy envelope interval which is lower than a global median and has a sparse pulse density which is greater than a neighborhood mean exists; Determining the complexity of a local signal according to the local gradient density of the instantaneous energy envelope, taking a variable-length primitive window as a seed, and elongating or shortening the suspected signal band based on the complexity of the local signal to obtain a target signal band; Splitting each window corresponding to the target signal band into a front section, a middle section and a rear section, performing maximum or minimum normalization based on local extremum on the front section and the rear section to inhibit endpoint drift, performing mean variance normalization based on instantaneous energy envelope on the middle section to highlight fault related fluctuation, and splicing the results of the three normalization sections into a complete window; And arranging each normalized window according to a time sequence, and inserting a transition frame between adjacent windows to obtain a window sequence, wherein the transition frame is generated based on weighted fusion of the front window and the rear window.
4. A method of monitoring the operating state of a vehicle engine according to claim 3, wherein before the time-sequentially arranging each window after normalization, further comprising: if the offset of the current window and the previous window at the center of gravity position of the frequency spectrum exceeds an empirical threshold, performing mild smoothing on the current window; If the deviation between the current window and the history window in the sparse pulse characteristic is larger than the preset deviation, triggering the short-time cache, and avoiding false alarm caused by accidental pulses.
5. The method for monitoring the operating state of the vehicle engine according to claim 1, further comprising, after the outputting of the early warning signal: mapping an early fault risk value corresponding to the early warning signal to a corresponding original window sequence; reversely sending the original window sequence into a reverse path of a phase encoder and an energy encoder, calculating a first gradient amplitude of an instantaneous phase sequence layer by using the phase encoder, and calculating a second gradient amplitude of an energy spectrogram layer by using the energy encoder; Normalizing the first gradient amplitude and the second gradient amplitude channel by channel, and marking that the frequency band of the channel is abnormal if the first gradient amplitude or the second gradient amplitude is larger than a preset gradient amplitude under the same channel and the same time point; If only the first gradient amplitude is larger than the preset gradient amplitude, marking as abnormal phase but normal energy, and prompting mechanical phase drift faults; If only the second gradient amplitude is larger than the preset gradient amplitude, marking as abnormal energy but normal phase, and prompting a combustion efficiency fault; The marked results are converted into readable anomaly interpretations using natural language templates.
6. An operation state monitoring device of a vehicle engine, characterized by comprising: the acquisition module is used for acquiring monitoring signals of the vehicle engine in a running state in real time; the generation module is used for carrying out sliding window slicing and normalization processing on the monitoring signals to generate window sequences; the feature output module is used for inputting the window sequence into a feature extraction model which is trained, and outputting an instantaneous latent vector, wherein the feature extraction model is formed by cascading a lightweight convolutional neural network and a transducer encoder and is used for capturing the fine changes of early weak faults of an engine in time domain, frequency domain and phase, and the instantaneous latent vector is a high-order abstract representation and is used for compressing and representing the current health state features of the engine; The early warning output module is used for inputting the instantaneous latent vector into the early warning module, outputting an early fault risk value, and outputting an early warning signal when the early fault risk values corresponding to a plurality of continuous window sequences are all larger than a dynamic threshold value, wherein the early warning module comprises an LSTM network and a full connection layer and is used for converting the instantaneous latent vector into the early fault risk value; wherein before the window sequence is input into the feature extraction model which is already trained, the method further comprises: Acquiring a plurality of monitoring signal samples, generating window sequence samples corresponding to each monitoring signal sample, randomly shielding window sequence samples of a preset proportion under the condition of no artificial tag, and reconstructing the window sequence samples by context to obtain target window sequence samples; mapping each target window sequence sample into a low-dimensional feature vector by using a transducer encoder to obtain a first instantaneous latent vector; Sequentially sending the first instantaneous latent vectors arranged in time sequence in the same operation record into a convolutional neural network to generate a second instantaneous latent vector, wherein the second instantaneous latent vector contains the history information of all target window sequence samples before each target window sequence sample; Selecting a second instantaneous latent vector of a current target window sequence sample and a first instantaneous latent vector of a time adjacent subsequent target window sequence sample to form a positive sample pair, and selecting a first instantaneous latent vector of a target window sequence sample which is non-adjacent or is derived from different operation records in the same batch and a current second instantaneous latent vector to form a negative sample pair; For each sample pair, calculating the similarity score of the second instantaneous latent vector and the first instantaneous latent vector, updating the parameters of the encoder and the convolutional neural network through back propagation until the similarity score of the positive sample pair is maximized and the similarity score of the negative sample pair is minimized, and generating a feature extraction model after training; Wherein said calculating a similarity score for the second instantaneous latent vector and the first instantaneous latent vector comprises: The method comprises the steps of sending a second instantaneous latent vector and a first instantaneous latent vector into a phase channel and an energy channel, outputting a first phase latent vector and a first energy latent vector corresponding to the second instantaneous latent vector, and a second phase latent vector and a second energy latent vector corresponding to the first instantaneous latent vector, wherein the phase channel is used for carrying out Hilbert transformation on the second instantaneous latent vector or the first instantaneous latent vector, extracting an instantaneous phase sequence, obtaining the phase latent vector through a causal expansion convolution network, and the energy channel is used for carrying out continuous wavelet transformation on the second instantaneous latent vector or the first instantaneous latent vector, extracting an energy spectrogram, and obtaining the energy latent vector through an asymmetric convolution neural network; calculating a first cosine distance between the first phase latent vector and the second phase latent vector, and calculating a second cosine distance between the first energy latent vector and the second energy latent vector; And carrying out weighted summation on the first cosine distance and the second cosine distance to obtain a similarity score.
7. A storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the operation state monitoring method of the vehicle engine according to any one of claims 1 to 5.
8. A controller for a vehicle, which is configured to control a controller, characterized by comprising the following steps: A processor; A memory; Wherein the memory stores a computer program, the processor implementing the method of monitoring the operating state of the vehicle engine according to any one of claims 1 to 5 when executing the computer program.

Description

Vehicle engine running state monitoring method, device, controller and storage medium Technical Field The invention relates to the technical field of vehicle engine monitoring, in particular to a method and a device for monitoring the running state of a vehicle engine, a controller and a storage medium. Background The conventional vehicle engine operation state monitoring system mostly depends on threshold value alarming or single frequency domain characteristics, and when facing early weak faults, such as turbine blade microcracks, early bearing pitting or local combustion chamber fire, the signal to noise ratio is extremely low, and the fault characteristics are often submerged in working condition fluctuation, sensor drift and environmental noise. More seriously, the weak anomalies are expressed as transient impact in the time domain and have a duration of only a few milliseconds, and the frequency domain is expressed as a 'wide-band and low-amplitude' dispersive spectrum line which is highly overlapped with combustion noise and gear meshing harmonic wave, and is modulated by rotating speed fluctuation and sampling clock jitter in phase, so that the detectability is further weakened. As a result, existing systems, even if the threshold is tuned finer, can only trigger an alarm at significant failure phases with signal to noise ratios greater than 10 db, where the crack has propagated to critical dimensions, the repair window is greatly compressed, and even evolved into catastrophic failure. Disclosure of Invention The invention mainly aims to provide a method, a device, a controller and a storage medium for monitoring the running state of a vehicle engine so as to remarkably improve the capturing capability of early weak faults of the engine. In order to achieve the above object, the present invention provides a method for monitoring an operating state of a vehicle engine, comprising: acquiring monitoring signals of a vehicle engine in a running state in real time; sliding window slicing and normalization processing are carried out on the monitoring signals, and window sequences are generated; Inputting the window sequence into a feature extraction model which is trained, and outputting an instantaneous latent vector, wherein the feature extraction model is formed by cascading a lightweight convolutional neural network and a transducer encoder and is used for capturing the subtle changes of early weak faults of an engine in time domain, frequency domain and phase, and the instantaneous latent vector is a high-order abstract representation and is used for compressing and representing the current health state features of the engine; The instantaneous latent vector is input into an early warning module, an early fault risk value is output, and when the early fault risk values corresponding to a plurality of continuous window sequences are all larger than a dynamic threshold value, an early warning signal is output, wherein the early warning module comprises an LSTM network and a full connection layer and is used for converting the instantaneous latent vector into the early fault risk value. Further, before inputting the window sequence into the feature extraction model after training, the method further comprises: Acquiring a plurality of monitoring signal samples, generating window sequence samples corresponding to each monitoring signal sample, randomly shielding window sequence samples of a preset proportion under the condition of no artificial tag, and reconstructing the window sequence samples by context to obtain target window sequence samples; mapping each target window sequence sample into a low-dimensional feature vector by using a transducer encoder to obtain a first instantaneous latent vector; Sequentially sending the first instantaneous latent vectors arranged in time sequence in the same operation record into a convolutional neural network to generate a second instantaneous latent vector, wherein the second instantaneous latent vector contains the history information of all target window sequence samples before each target window sequence sample; Selecting a second instantaneous latent vector of a current target window sequence sample and a first instantaneous latent vector of a time adjacent subsequent target window sequence sample to form a positive sample pair, and selecting a first instantaneous latent vector of a target window sequence sample which is non-adjacent or is derived from different operation records in the same batch and a current second instantaneous latent vector to form a negative sample pair; And for each sample pair, calculating the similarity score of the second instantaneous latent vector and the first instantaneous latent vector, updating the parameters of the encoder and the convolutional neural network through back propagation until the similarity score of the positive sample pair is maximized and the similarity score of the negative sample pair is minimized, and gene