Intelligent vehicle-mounted LNG (liquefied natural gas) and CNG (compressed natural gas) multi-path electromagnetic valve monitoring system and method

Abstract

The invention relates to the technical field of automobiles, in particular to an intelligent vehicle-mounted LNG (Liquefied Natural Gas) and CNG (Compressed Natural Gas) multi-path electromagnetic valve monitoring system and method. The method comprises the following steps: extracting electrical characteristics and spectral characteristics representing electromagnetic valve response from multi-channel synchronous time sequence data, and fusing temperature parameters to form a high-dimensional feature vector; inputting the high-dimensional feature vector into a lightweight Transform-CNN (Convolutional Neural Network) hybrid network model to obtain the confidence coefficient of the state of the valve body; and performing fault early warning decision based on the confidence coefficient of the valve body state, uploading vehicle end operation data and a decision result to a cloud platform through a federated learning framework to trigger global incremental updating of the hybrid network model, issuing updated model parameters to the vehicle end to complete self-evolution, and establishing a'data-model-reasoning-decision 'closed loop to complete self-evolution of the hybrid network model. Accurate monitoring and intelligent early warning of the state of the vehicle-mounted LNG and CNG multi-way electromagnetic valve are achieved.

Description

Technical Field

[0001] This invention relates to the field of automotive technology, and in particular to an intelligent on-board LNG and CNG multi-channel solenoid valve monitoring system and method. Background Technology

[0002] Traditional onboard LNG and CNG supply systems often employ parallel solenoid valve assemblies for fuel delivery control, relying primarily on simple voltage or current threshold judgment methods for status monitoring. However, under complex real-world conditions such as cold starts, electromagnetic interference, significant power supply voltage fluctuations, and drastic coil temperature changes, the mechanical movement of the valve core and the electrical response of the coil exhibit strong nonlinear characteristics, leading to high false alarm and false negative rates for monitoring methods based on fixed thresholds. Furthermore, the mutual coupling of signals during multi-valve parallel operation further complicates accurate status identification. Existing solutions generally lack self-learning and evolution capabilities, cannot adapt to device aging and operating condition drift, and require manual calibration and maintenance. Their reliability, real-time performance, and intelligence levels are insufficient to meet the high safety and reliability requirements of current commercial vehicles. Therefore, there is an urgent need for an intelligent monitoring system capable of accurately modeling the dynamic response characteristics of solenoid valves, possessing online learning capabilities, and meeting the requirements for automotive-grade embedded deployment. Summary of the Invention

[0003] This invention overcomes the shortcomings of the prior art and provides an intelligent on-board LNG and CNG multi-channel solenoid valve monitoring system and method.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The first aspect of this invention discloses a method for monitoring intelligent on-board LNG and CNG multi-channel solenoid valves, comprising the following steps: The coil voltage, current and temperature signals of each parallel solenoid valve are collected synchronously, and the signals are cleaned and aligned to generate multi-channel synchronous timing data. From the multi-channel synchronous time-series data, electrical and spectral characteristics characterizing the response of the solenoid valve are extracted, and a high-dimensional feature vector is formed by fusing temperature parameters. The high-dimensional feature vector is input into a lightweight Transformer-CNN hybrid network model. The CNN module captures the local details of the electrical characteristics, and the Transformer module models the long-range dependencies between multi-channel signals to obtain the valve body state confidence. Fault warning decisions are made based on the confidence level of the valve body state, and the vehicle-side operating data and decision results are uploaded to the cloud platform via a federated learning framework to trigger a global incremental update of the hybrid network model. The updated model parameters are then sent back to the vehicle to complete self-evolution.

[0005] Preferably, the coil voltage, current, and temperature signals of each parallel solenoid valve are synchronously acquired, and the signals are cleaned and aligned to generate multi-channel synchronous timing data, specifically: Synchronously acquire signals from various sensors, including the voltage and current signals of the solenoid valve coil and the temperature signal acquired by the temperature sensor, to obtain the original synchronization signal sequence; The original synchronization signal sequence is filtered and denoised to generate a preprocessed signal sequence; The preprocessed signal sequence is subjected to outlier detection and removal, and missing data is filled by interpolation to form a clean signal sequence; A unified time reference is assigned to each of the clean signal sequences and time alignment is performed to generate a time-aligned multi-channel signal sequence. The time-aligned multi-channel signal sequence is reassembled according to the sampling time to construct multi-channel synchronous time-series data.

[0006] Preferably, electrical and spectral characteristics characterizing the solenoid valve response are extracted from the multi-channel synchronous timing data, and a high-dimensional feature vector is formed by fusing temperature parameters, specifically: Empirical mode decomposition is performed on the voltage and current signals in the synchronous time-series data to obtain a set of intrinsic mode functions. The Hilbert marginal spectral energy of each mode is then extracted as the nonlinear spectral energy distribution feature. Calculate the coherence coefficient of the transfer function of the current signal relative to the voltage signal, and combine the coherence coefficient of the transfer function with the differential inductance parameter to form the frequency domain impedance characteristic set of the solenoid valve; The temperature signal and voltage signal are coupled and analyzed to extract temperature-voltage joint modulation characteristic parameters; The high-dimensional feature vector is generated by fusing the nonlinear spectral energy distribution characteristics, frequency domain impedance feature set, and temperature-voltage joint modulation feature parameters through multi-domain feature fusion.

[0007] Specifically, the coherence coefficient of the transfer function of the current signal relative to the voltage signal is calculated, and the coherence coefficient of the transfer function, together with the differential inductance parameter, constitutes the frequency domain impedance characteristic set of the solenoid valve. Complex analytic envelope extraction is performed on the voltage and current signals to obtain voltage envelope signals and current envelope signals; Based on the voltage envelope signal and the current envelope signal, the instantaneous phase difference is calculated by Hilbert transform to generate a phase coherence index; The current signal is subjected to high-order differential operation to extract the current change acceleration feature, and the current change acceleration feature is nonlinearly mapped to the instantaneous rate of change of the voltage signal to generate differential inductance parameters; Based on the phase coherence index and differential inductance parameters, the time-series characteristics are converted into a two-dimensional feature image through Gram angle field transformation to generate an impedance feature spectrum. The impedance feature map is input into a lightweight convolutional network for deep feature extraction to obtain a frequency domain impedance feature set.

[0008] Preferably, the high-dimensional feature vector is input into a lightweight Transformer-CNN hybrid network model. The CNN module captures the local details of the electrical characteristics, and the Transformer module models the long-range dependencies between multi-channel signals to obtain the valve body state confidence. Specifically: The high-dimensional feature vector is adaptively sampled through a deformable one-dimensional convolutional layer to generate a geometrically adaptive local feature map. The local feature map is input into the multi-head causal convolution module to model temporal causal relationships and outputs a spatiotemporal correlation feature tensor. After encoding the spatiotemporal correlation feature tensor with relative position, it is input into the gated attention unit to calculate the causal attention weights between channels and generate attention-weighted context features. After residual connection of the context features and the original local features, feature filtering is performed through gated linear units to generate refined feature representations. The refined feature representation is input into the time series focusing module to calculate the feature importance weights at different time steps, and the feature importance weights at different time steps are used to perform probability divergence calculation based on the reference distribution to generate focusing coefficients. When the focusing coefficient exceeds a preset coefficient threshold, key timing features are enhanced, and finally, the valve body state confidence is generated through the sigmoid activation function.

[0009] Specifically, the refined feature representation is input into the temporal focusing module to calculate the feature importance weights at different time steps. Then, using these feature importance weights, a probability divergence calculation based on a reference distribution is performed to generate focusing coefficients. Temporal autocorrelation analysis is performed on the refined feature representation to extract the feature activation saliency at each time step and generate a temporal saliency mapping. Based on the aforementioned temporal saliency mapping, the feature importance weights for each time step are obtained through information entropy weighted calculation. The JS divergence of the feature importance weights and the preset feature importance prior distribution is used to obtain the time-series distribution difference measure. The time-series distribution difference metric is adaptively scaled and converted into a focusing coefficient.

[0010] Preferably, the fault warning decision is made based on the confidence level of the valve body state, specifically as follows: An energy integral operation is performed on the confidence level of the valve body state using a sliding time window to generate a confidence level energy sequence; the Mahalanobis distance between the confidence level energy sequence and the normal operation reference template is calculated to generate a confidence level deviation index. The confidence deviation index is processed by nonlinear mapping and combined with real-time operating parameters for adaptive sensitivity adjustment, and the fault probability estimate is output. Based on the estimated fault probability, a multimodal decision fusion mechanism is used to generate the final hierarchical early warning instruction. The graded warning instructions are encoded into a standard CAN message format and sent to the main control ECU via the CAN vehicle bus. The main control ECU then triggers the corresponding audible and visual alarm response based on the message content.

[0011] Preferably, the vehicle-side operational data and decision results are uploaded to the cloud platform via a federated learning framework to trigger a global incremental update of the hybrid network model. The updated model parameters are then distributed back to the vehicle to complete its self-evolution. Specifically: Local differential privacy processing is performed on the multi-channel synchronous time-series data and corresponding valve body state decision results in the vehicle-side operation data to generate a desensitized feature dataset. The local update gradient of the lightweight hybrid network model is calculated based on the desensitized feature dataset, and the local update gradient is pruned and quantized to generate an encrypted model update fragment. The encrypted model update fragments from multiple vehicle terminals are uploaded to the cloud platform and fused using a secure aggregation algorithm to generate a global model update gradient. The global model update gradient is used to incrementally update the cloud-based baseline model, generating parameters for a new generation of lightweight hybrid network models. After differential compression and signature verification of the new generation model parameters, they are distributed to each vehicle via OTA through the vehicle communication module to complete the self-evolution and update of the vehicle model.

[0012] The second aspect of this invention discloses an intelligent on-board LNG and CNG multi-channel solenoid valve monitoring system. The intelligent on-board LNG and CNG multi-channel solenoid valve system includes a memory and a processor. The memory stores an intelligent on-board LNG and CNG multi-channel solenoid valve method program. When the intelligent on-board LNG and CNG multi-channel solenoid valve method program is executed by the processor, the steps of the intelligent on-board LNG and CNG multi-channel solenoid valve method described in any one of the claims are implemented.

[0013] This invention addresses the technical deficiencies in the prior art and has the following beneficial effects: By simultaneously acquiring multi-source signals and fusing deep features, this invention improves the accuracy of identifying the true opening and closing state of the solenoid valve, effectively suppressing false alarms caused by voltage fluctuations, temperature changes, and electromagnetic interference. The lightweight hybrid model constructed achieves low-latency and high-reliability inference on automotive-grade hardware. At the same time, by leveraging the federated learning framework, the system has the ability to continuously evolve, adapting to different operating conditions and device aging, ultimately forming a highly reliable, highly accurate, and self-optimizing intelligent monitoring closed-loop system under all operating conditions. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained from these drawings without creative effort.

[0015] Figure 1 This is a flowchart of the intelligent on-board LNG and CNG multi-channel solenoid valve monitoring method. Figure 2 This is a system block diagram of the intelligent on-board LNG and CNG multi-channel solenoid valve monitoring system. Detailed Implementation

[0016] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0017] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0018] like Figure 1 As shown, the first aspect of this invention discloses an intelligent on-board LNG and CNG multi-channel solenoid valve monitoring method, comprising the following steps: S1. Synchronously acquire the coil voltage, current and temperature signals of each parallel solenoid valve, and perform signal cleaning and alignment to generate multi-channel synchronous timing data; S2. Extract the electrical and spectral characteristics that characterize the response of the solenoid valve from the multi-channel synchronous timing data, and form a high-dimensional feature vector by fusing the temperature parameters. S3. Input the high-dimensional feature vector into a lightweight Transformer-CNN hybrid network model. The CNN module captures the local details of the electrical characteristics, and the Transformer module models the long-range dependencies between multi-channel signals to obtain the valve body state confidence. S4. Based on the confidence level of the valve body state, make a fault warning decision, and upload the vehicle-side operation data and decision results to the cloud platform through the federated learning framework to trigger the global incremental update of the hybrid network model. The updated model parameters are then sent back to the vehicle to complete self-evolution.

[0019] It should be noted that this invention addresses the problems of high false alarm rates and insufficient reliability in traditional on-board LNG and CNG solenoid valve monitoring schemes under conditions such as parallel multi-valve operation, voltage fluctuations, and coil temperature rise. By constructing a closed loop of "data-model-reasoning-decision," it achieves accurate identification of the actual opening and closing status of the solenoid valve. In this invention, LNG stands for Liquefied Natural Gas (LNG) in Chinese and Compressed Natural Gas (CNG) in English. The technical solution of this invention is applicable to the monitoring method of compressed gas solenoid valves, including those for on-board LNG and CNG.

[0020] Preferably, the coil voltage, current, and temperature signals of each parallel solenoid valve are synchronously acquired, and the signals are cleaned and aligned to generate multi-channel synchronous timing data, specifically: Synchronously acquire signals from various sensors, including the voltage and current signals of the solenoid valve coil and the temperature signal acquired by the temperature sensor, to obtain the original synchronization signal sequence; The original synchronization signal sequence is filtered and denoised to generate a preprocessed signal sequence; The preprocessed signal sequence is subjected to outlier detection and removal, and missing data is filled by interpolation to form a clean signal sequence; A unified time reference is assigned to each of the clean signal sequences and time alignment is performed to generate a time-aligned multi-channel signal sequence. The time-aligned multi-channel signal sequence is reassembled according to the sampling time to construct multi-channel synchronous time-series data.

[0021] It should be noted that outlier detection and removal (e.g., statistical methods based on Z-score or interquartile range) and interpolation imputation (e.g., using linear interpolation or spline interpolation) are performed on the preprocessed signal to eliminate the negative impact of abnormal data points such as instantaneous false alarms from sensors and packet loss during signal transmission on model training.

[0022] Preferably, electrical and spectral characteristics characterizing the solenoid valve response are extracted from the multi-channel synchronous timing data, and a high-dimensional feature vector is formed by fusing temperature parameters, specifically: Empirical mode decomposition is performed on the voltage and current signals in the synchronous time-series data to obtain a set of intrinsic mode functions. The Hilbert marginal spectral energy of each mode is then extracted as the nonlinear spectral energy distribution feature. Calculate the coherence coefficient of the transfer function of the current signal relative to the voltage signal, and combine the coherence coefficient of the transfer function with the differential inductance parameter to form the frequency domain impedance characteristic set of the solenoid valve; The temperature signal and voltage signal are coupled and analyzed to extract temperature-voltage joint modulation characteristic parameters; The high-dimensional feature vector is generated by fusing the nonlinear spectral energy distribution characteristics, frequency domain impedance feature set, and temperature-voltage joint modulation feature parameters through multi-domain feature fusion.

[0023] Specifically, the calculation of the transfer function coherence coefficient of the current signal relative to the voltage signal, and the combination of the transfer function coherence coefficient and the differential inductance parameter to form the frequency domain impedance feature set of the solenoid valve, involves: performing complex analytical envelope extraction on the voltage and current signals to obtain voltage envelope signals and current envelope signals; calculating the instantaneous phase difference based on the voltage and current envelope signals using Hilbert transform to generate a phase coherence index; performing high-order differential operations on the current signal to extract the current change acceleration feature, and nonlinearly mapping the current change acceleration feature to the instantaneous rate of change of the voltage signal to generate a differential inductance parameter; converting the time-series features into a two-dimensional feature image using Gram angle field transform to generate an impedance feature spectrum; and inputting the impedance feature spectrum into a lightweight convolutional network for deep feature extraction to obtain the frequency domain impedance feature set.

[0024] It should be noted that, due to the non-stationary and nonlinear characteristics of the coil's electrical response during the opening and closing of the solenoid valve, traditional Fourier transform is difficult to analyze effectively. Empirical Mode Decomposition (EMD) can adaptively decompose complex signals into a set of steady-state eigenmode functions with frequencies ranging from high to low, thereby extracting the wave components at different scales in the signal. From these, the Hilbert marginal spectrum energy of each mode can be extracted to quantify the cumulative energy distribution of different frequency components over the entire time range, thus forming a nonlinear spectral energy distribution characteristic. This effectively characterizes the unique patterns of electromagnetic energy release under different operating conditions (such as normal engagement, start-up jamming, and insufficient voltage).

[0025] The coherence coefficient of the transfer function of the current signal relative to the voltage signal is calculated and combined with the differential inductance parameter to form the frequency domain impedance characteristic set of the solenoid valve. This aims to deeply explore the frequency domain impedance characteristics of the coil, which are core physical parameters characterizing the dynamic response of the solenoid valve. Specifically: Complex analytical envelope extraction is performed on the voltage and current signals to obtain envelopes that reflect the trends of signal amplitude and phase changes, removing carrier interference and focusing on modulation information. Based on the voltage and current envelope signals, the instantaneous phase difference is calculated using Hilbert transform to reflect the degree of mismatch between the voltage and current waveforms. Its coherence index can effectively characterize whether the inductive characteristics of the coil deviate due to valve core movement, temperature rise, or power supply disturbance, and is extremely sensitive to faults. High-order differential operations are performed on the current signal to extract the current change acceleration characteristics, and this is nonlinearly mapped to the instantaneous rate of change of the voltage signal to generate differential inductance parameters. Differential inductance (dL / dt or dI / dt) is better at capturing small inductance changes caused by dynamic processes such as core movement than static inductance, and is a key feature for identifying the initial motion state of the valve core. The above phase coherence index and differential inductance parameters are converted into a two-dimensional feature image through Gram Angular Field Transform (GAF). Gaussian Image Processing (GAF) encodes one-dimensional time series into two-dimensional matrices, preserving their temporal dependencies (such as time correlations) and transforming these key impedance time-series features into impedance feature maps. These impedance feature maps are then input into a lightweight convolutional network (such as using a SqueezeNet or MobileNet architecture) for deep feature extraction. Leveraging the powerful image feature extraction capabilities of CNNs, the network automatically learns deep, abstract patterns related to state within the map, ultimately obtaining a highly condensed and representative frequency-domain impedance feature set. This transforms the challenging problem of time-series analysis into a more manageable image recognition problem, improving the efficiency and effectiveness of feature representation.

[0026] Furthermore, it should be noted that the temperature and voltage signals are coupled and analyzed to extract temperature-voltage joint modulation characteristic parameters. The change in coil resistance with temperature is one of the main causes of electrical characteristic drift and false alarms. By analyzing the rate of voltage change required to maintain the same current at a specific temperature, or calculating the equivalent voltage after temperature compensation, the modulation effect of temperature on electrical parameters is quantified, thus providing the model with key inputs for adaptive compensation and enhancing the model's stability across the entire temperature range (-40℃ to 85℃). The nonlinear spectral energy distribution characteristics, frequency domain impedance feature set, and temperature-voltage joint modulation characteristic parameters are then fused across multiple domains to generate a high-dimensional feature vector that characterizes the dynamic spectrum, represents the impedance essence, and includes temperature compensation information.

[0027] Preferably, the high-dimensional feature vector is input into a lightweight Transformer-CNN hybrid network model. The CNN module captures the local details of the electrical characteristics, and the Transformer module models the long-range dependencies between multi-channel signals to obtain the valve body state confidence. Specifically: The high-dimensional feature vector is adaptively sampled through a deformable one-dimensional convolutional layer to generate a geometrically adaptive local feature map. The local feature map is input into the multi-head causal convolution module to model temporal causal relationships and outputs a spatiotemporal correlation feature tensor. After encoding the spatiotemporal correlation feature tensor with relative position, it is input into the gated attention unit to calculate the causal attention weights between channels and generate attention-weighted context features. After residual connection of the context features and the original local features, feature filtering is performed through gated linear units to generate refined feature representations. The refined feature representation is input into the time series focusing module to calculate the feature importance weights at different time steps, and the feature importance weights at different time steps are used to perform probability divergence calculation based on the reference distribution to generate focusing coefficients. When the focusing coefficient exceeds a preset coefficient threshold, key timing features are enhanced, and finally, the valve body state confidence is generated through the sigmoid activation function.

[0028] Specifically, the implementation process of the lightweight Transformer-CNN hybrid network model is as follows: First, the high-dimensional feature vector generated in the previous steps is input into a deformable one-dimensional convolutional layer (64 kernels, kernel length 5, dilation rate 2). The convolutional layer adaptively learns the sampling offset to perceive key feature points and generate a geometrically adaptive local feature map. Subsequently, the local feature map is input into a multi-head causal convolutional module with four heads (kernel lengths of 3 / 5 / 7 / 9 respectively). While ensuring temporal causality, it captures multi-scale local patterns and outputs a spatiotemporal correlation feature tensor with a dimension of 256. Then, after encoding the feature tensor with sine-cosine relative position, it is input into a gated attention unit with 128 hidden units. The inter-channel causal attention weights are calculated (4 attention heads, 64 key-value pair dimensions), generating attention-weighted contextual features. These contextual features are then residually concatenated with the original local feature maps, and feature filtering is performed through a gated linear unit (SiLU activation function, 0.1 dropout rate), outputting a refined 128-dimensional feature representation. Finally, this feature representation is input into the temporal focusing module, where autocorrelation analysis is used to calculate the feature importance weights at each time step (200 time steps, 128 feature dimensions). The JS divergence is then calculated with a preset normal prior distribution. When the generated focusing coefficient exceeds a threshold of 0.35, the key time step features are enhanced by a factor of 3. Finally, the valve state confidence is output through a sigmoid activation function.

[0029] Specifically, the refined feature representation is input into the temporal focusing module to calculate the feature importance weights at different time steps. Then, using these feature importance weights, a probability divergence calculation based on a reference distribution is performed to generate focusing coefficients. Temporal autocorrelation analysis is performed on the refined feature representation to extract the feature activation saliency at each time step and generate a temporal saliency mapping. Based on the aforementioned temporal saliency mapping, the feature importance weights for each time step are obtained through information entropy weighted calculation. The JS divergence of the feature importance weights and the preset feature importance prior distribution is used to obtain the time-series distribution difference measure. The time-series distribution difference metric is adaptively scaled and converted into a focusing coefficient.

[0030] It should be noted that by calculating the autocorrelation coefficients of the feature sequences at different time delays, the repetitive and periodic patterns of features at each time step within their context are quantified, thereby extracting the feature activation salience of each time step. This effectively captures strongly correlated latent fault symptoms, such as the minute sub-twists of the current characteristics at the moment of solenoid valve startup, generating a more reliable time-series salience mapping. Subsequently, based on this mapping, the feature importance weight of each time step is obtained through information entropy weighting. Information entropy measures the uncertainty or information content of the feature value at that time step. The lower the entropy value, the more stable and abnormal the feature pattern at that point (e.g., the feature pattern during a fault is usually different from the normal stable operation mode), and therefore a higher weight is assigned, thus transforming subjective experience weighting into objective quantitative weighting based on data-driven information. The calculated feature importance weights are then compared with the preset feature importance prior distribution using JS divergence calculation. The prior distribution is derived from a large number of normal samples and represents "how the importance weights of each time step should be distributed under normal conditions". JS divergence is used to measure the difference between the current actual weight distribution and the normal prior distribution, resulting in a temporal distribution difference metric. A larger temporal distribution difference metric indicates an abnormal focal point in the current sequence that differs from the normal pattern (e.g., abnormally high weights at a certain time step), corresponding to the time of fault occurrence. Finally, this temporal distribution difference metric is adaptively scaled (e.g., using a sigmoid function to map it to a preset coefficient range), converting it into the final focusing coefficients. This ensures that the valve body state confidence score output by the model accurately reflects the true state of the solenoid valve, especially showing significant probability changes in the early stages of a fault, effectively reducing the false negative rate.

[0031] Preferably, the fault warning decision is made based on the confidence level of the valve body state, specifically as follows: An energy integral operation is performed on the confidence level of the valve body state using a sliding time window to generate a confidence level energy sequence; the Mahalanobis distance between the confidence level energy sequence and the normal operation reference template is calculated to generate a confidence level deviation index. The confidence deviation index is processed by nonlinear mapping and combined with real-time operating parameters for adaptive sensitivity adjustment, and the fault probability estimate is output. It should be noted that an improved S-shaped function is used to perform a nonlinear mapping on the confidence deviation index D. The function expression is: [equation missing in original text], where the slope parameter k is initially set to 2.5, and the center point parameter D0 is initially set to 3.0 (corresponding to a Mahalanobis distance of 3 standard deviations). This maps the deviation index to a basic failure probability estimate P between 0 and 1. baseThen, sensitivity adaptive adjustment is performed by monitoring key operating parameters in real time, including the current power supply voltage, coil temperature, and vehicle speed. The parameters of the nonlinear mapping function are adjusted according to a preset sensitivity adjustment coefficient table. For example, when the power supply voltage is detected to be below 24V and the coil temperature below -20°C, it is determined to be a severe "low temperature and low pressure" condition. The slope parameter k is increased to 3.2 to increase sensitivity and ensure effective fault detection even when the electrical signal is weak. When the vehicle is traveling at high speed (e.g., above 80km / h), to avoid false alarms that could distract the driver, the center point parameter D0 is increased to 3.5 to appropriately reduce sensitivity. Finally, the final fault probability estimate is calculated based on the real-time adjusted parameters k1 and D1. This value is output as a continuous quantity for subsequent hierarchical decision-making. Where k1 is the slope parameter adaptively adjusted according to real-time operating parameters (voltage, temperature, vehicle speed); D1 is the center point parameter adaptively adjusted according to real-time operating parameters; and e is the natural constant.

[0032] Based on the estimated fault probability, a multimodal decision fusion mechanism is used to generate the final hierarchical early warning instruction. It should be noted that setting multiple probability threshold ranges, for example, setting the estimated failure probability P... fault The range [0, 0.3) is defined as "normal", [0.3, 0.7) as "attention", [0.7, 0.9) as "warning", and [0.9, 1.0] as "severe". Simultaneously, additional modal information is continuously monitored and incorporated for decision fusion, including the temporal trend of the current fault probability (e.g., the rate of increase of the probability value within the most recent 200ms window), the real-time operating status of the vehicle (e.g., whether the engine is starting or the vehicle is accelerating), and the historical alarm frequency records of the valve. The final decision is executed by a rule engine: if the current P... fault If the value reaches the "Caution" level, but the vehicle is in a stable idling condition and the probability value trend is stable, then no warning will be issued and monitoring will continue; if P fault If the value reaches the "Warning" level, and a rapid increase in the probability value is detected within a short period of time (e.g., an increase of more than 0.2 within 50ms), the decision result will be upgraded from "Warning" to "Severe"; conversely, if P... fault Even if the value is at the "warning" level, if the vehicle is under heavy acceleration and has a good historical record, the decision will be downgraded to the "attention" level. By integrating multimodal information such as real-time probability values, dynamic trends, operating condition context, and historical data, a comprehensive and graded warning instruction is finally output. This instruction reflects the risk level, effectively avoiding misjudgment based on a single probability value under complex operating conditions and improving the reliability of the warning system.

[0033] The graded warning instructions are encoded into a standard CAN message format and sent to the main control ECU via the CAN vehicle bus. The main control ECU then triggers the corresponding audible and visual alarm response based on the message content.

[0034] Preferably, the vehicle-side operational data and decision results are uploaded to the cloud platform via a federated learning framework to trigger a global incremental update of the hybrid network model. The updated model parameters are then distributed back to the vehicle to complete its self-evolution. Specifically: Local differential privacy processing is performed on the multi-channel synchronous time-series data and corresponding valve body state decision results in the vehicle-side operation data to generate a desensitized feature dataset. The local update gradient of the lightweight hybrid network model is calculated based on the desensitized feature dataset, and the local update gradient is pruned and quantized to generate an encrypted model update fragment. The encrypted model update fragments from multiple vehicle terminals are uploaded to the cloud platform and fused using a secure aggregation algorithm to generate a global model update gradient. The global model update gradient is used to incrementally update the cloud-based baseline model, generating parameters for a new generation of lightweight hybrid network models. After differential compression and signature verification of the new generation model parameters, they are distributed to each vehicle via OTA through the vehicle communication module to complete the self-evolution and update of the vehicle model.

[0035] It should be noted that, on the vehicle side, local differential privacy processing is performed on the locally stored multi-channel synchronous time-series data (such as voltage, current, and temperature waveform segments from the most recent 24 hours) and the corresponding valve status decision results (such as "normal" or "valve not open" labels). Specifically, a Laplace mechanism is used to add ε-differential privacy-satisfying noise (e.g., setting a privacy budget ε=2.0) to the statistics of the high-dimensional feature vector, generating an anonymized feature dataset. This ensures that the uploaded data cannot be used to infer the original information of any single vehicle, protecting user privacy and data security from the source.

[0036] Based on this desensitized feature dataset, a lightweight hybrid network model is trained locally on the vehicle-mounted MCU (training period is 1 epoch), and the local update gradient of the model parameters is calculated. To adapt to the limited computing power of automotive-grade chips and reduce communication overhead, the local update gradient is subjected to gradient pruning (e.g., limiting the gradient norm to within 1.5) and quantization encoding (uniformly quantizing the 32-bit floating-point gradient value into 8-bit integers). Finally, encrypted model update fragments are generated to compress the huge model update information into extremely small data packets (e.g., only 50kB per update), saving wireless communication traffic.

[0037] Then, encrypted model update fragments from multiple vehicles (e.g., more than 1,000 vehicles) are uploaded to the cloud platform via cellular network. The cloud uses a secure aggregation algorithm to perform weighted averaging and fusion of thousands of update fragments without decrypting the individual data packets from each vehicle, generating a global model update gradient. Like a "blind box" operation, it only aggregates the quantized results, ensuring the indistinguishability of update information for individual vehicles.

[0038] The gradient is updated using this global model, and an incremental update is performed on the baseline model stored in the cloud (using the Adam optimizer with a learning rate of 0.001) to generate parameters for a new generation of lightweight hybrid network model. This process leverages diverse operating conditions across the entire fleet (such as temperature differences in different regions and different drivers' operating habits) to enable the model to continuously evolve and enhance its generalization ability.

[0039] Finally, the parameters of the new generation model are differentially compared with the older version parameters widely used in the current vehicle terminals to generate a differential compressed package (usually less than 80kB). A digital signature algorithm (such as ECDSA) is then used to verify the signature of this compressed package, ensuring that it is not tampered with during transmission. Ultimately, this signed differential compressed package is distributed to each vehicle terminal via OTA (Over-The-Air). After verifying the signature, the vehicle terminal automatically updates and replaces the model parameters, achieving self-evolution and updating of the vehicle model.

[0040] In this embodiment, it also includes: Based on the valve body state confidence and failure probability estimate, a Lyapunov energy function is constructed with the response deviation and its first derivative as joint state variables. Based on the time derivative of the Lyapunov energy function, a stability criterion containing the nonlinear characteristics of the solenoid valve is derived. A radial basis function neural network is used to approximate the unmodeled dynamic characteristics in the stability criterion online, and the weight adjustment amount and real-time approximation error of the neural network are output. The weight adjustment amount is coupled and fused with the real-time acquired coil current differential signal and flux change amount through multi-physics field coupling to generate an adaptive control signal containing nonlinear compensation characteristics. The phase and amplitude of the reference pulse width modulation carrier are reconstructed using the adaptive control signal. By adjusting the rising slope and falling delay characteristics of the carrier signal, a driving pulse sequence with frequency and duty cycle co-modulation is generated. The driving pulse sequence is applied to the solenoid valve coil via a power amplifier circuit to form a closed-loop control based on Lyapunov stability to actively suppress response deviation.

[0041] It should be noted that the valve body state confidence level (p) and the estimated failure probability (P) obtained from the previous steps... faultUsing the response deviation (e=1-p) and its first derivative (ė) as input, a Lyapunov energy function V(e,ė) is constructed with the response deviation (e=1-p) and its first derivative (ė) as joint state variables. This function (e.g., in the form V(e,ė)=(1 / 2)*e²+(1 / 2)*ė²) quantifies the current system state from the ideal steady state (p=1,P) in terms of energy. fault The difference between =0).

[0042] Based on the time derivative of the Lyapunov energy function, a stability criterion incorporating nonlinear characteristics of the solenoid valve (such as hysteresis and eddy current effects) is derived. To compensate for these difficult-to-model dynamic characteristics online in real time, a radial basis function (RBF) neural network is used to approximate the unmodeled dynamic characteristics in the stability criterion. The RBF network takes coil current, real-time voltage, and historical confidence levels as inputs, and its hidden layers contain 20 Gaussian kernel functions. It outputs the weight adjustment and real-time approximation error of the neural network online, thereby identifying and compensating for the nonlinear disturbances of the system.

[0043] The weight adjustment output of the RBF neural network is coupled and fused with the real-time acquired differential coil current signal and the flux change estimated by voltage-current integration to generate an adaptive control signal with nonlinear compensation characteristics. The adaptive control signal not only includes basic error feedback, but also integrates feedforward compensation for unmodeled dynamics of the system and external disturbances.

[0044] Then, the phase and amplitude of a reference pulse-width modulation (PWM) carrier (e.g., a triangular wave with a frequency of 1 kHz) are dynamically reconstructed using an adaptive control signal. Specifically, a voltage-controlled oscillator (VCO) and a digital amplifier are used to adjust the rising slope (to change the equivalent voltage) and falling edge delay (to adjust the energy injection duration) of the carrier signal according to the amplitude of the adaptive control signal, thereby generating a drive pulse sequence that is co-modulated in terms of frequency and duty cycle. This co-modulation strategy can control the energy injected into the coil more precisely than traditional fixed-frequency PWM, avoiding overshoot and oscillation.

[0045] Finally, the drive pulse sequence is applied to both ends of the solenoid valve coil via an H-bridge power amplifier circuit. This forms a closed-loop control loop based on Lyapunov stability theory. This loop can sense the changing trend of the valve body state in real time and adjust the drive characteristics in real time, actively and proactively suppressing response deviations, ensuring that the solenoid valve can operate quickly, smoothly, and reliably under various disturbances.

[0046] In this embodiment, it also includes: When modeling the long-range dependency of multi-channel signals in the Transformer module, the transient magnetic field spatial distribution generated in the valve core of the adjacent soft magnetic material is analyzed by the Biot-Savart law based on the coil drive current waveform in the multi-channel synchronous timing data, and the transient magnetic field intensity tensor is generated. The transient magnetic field strength tensor is analyzed based on the Stoner-Wolfas model to obtain its perturbation effect on the displacement energy barrier of the magnetic domain wall inside the valve body material, and the probability distribution of micro-magnetic domain flipping is quantified to obtain the macro-magnetic-strictive strain perturbation vector. Using the magnetostrictive strain perturbation vector, an equivalent electromagnetic valve core mechanical stress spectrum induced by the magnetoelastic effect is constructed, and the equivalent electromagnetic valve core mechanical stress spectrum is encoded into a frequency domain physical constraint mask using Fourier descriptors. During the multi-head attention computation process, the frequency domain physical constraint mask and the attention score matrix are subjected to Hadamard product operation, thereby embedding the long-range physical dependencies induced by the magnetic field into the attention weight distribution in a hard constraint manner, thus optimizing the representation quality of context features.

[0047] In practice, firstly, based on the coil drive current waveform in multi-channel synchronous timing data, the spatial distribution of the transient magnetic field generated inside the valve core in the adjacent soft magnetic material is analyzed using the Biot-Savart law, generating a transient magnetic field strength tensor. This tensor quantifies the magnetic field strength and direction at each point in space. Next, based on the Stoner-Wolface physics model, the transient magnetic field strength tensor is analyzed to obtain the perturbation effect of the external magnetic field on the displacement energy barrier of the magnetic domain walls inside the valve body material, thereby quantifying the magnetic domain flipping probability distribution at the microscale. Based on this, a macroscopically observable magnetostrictive strain perturbation vector is derived to reflect the deformation effect caused by changes in the magnetization state of the material. Using the magnetostrictive strain perturbation vector, an equivalent electromagnetic valve core mechanical stress spectrum induced by magnetoelastic effects is constructed to characterize how magnetostriction is converted into mechanical stress and propagates within the valve body. The equivalent mechanical stress spectrum is then encoded using Fourier descriptors and converted into a frequency-domain physical constraint mask. This mask identifies which frequency components are dominated by the aforementioned physical processes in the frequency domain. Finally, in the key step of multi-head attention calculation, the frequency domain physical constraint mask and the attention score matrix obtained by model self-learning are subjected to Hadamard product operation. This embeds the prior knowledge of the long-range physical dependence caused by the magnetic field directly into the distribution of attention weights in a hard constraint manner, forcing the model to focus in a direction that conforms to physical laws. Ultimately, this optimizes the representation quality of context features, making them more realistically reflect the internal state of the solenoid valve.

[0048] In summary, this embodiment deeply integrates first-principles physics with a data-driven model, improving the model's ability to represent complex electromagnetic-mechanical coupled systems and its state recognition accuracy. It also enhances the model's interpretability and generalization performance under unseen operating conditions.

[0049] like Figure 2 As shown, the second aspect of the present invention discloses an intelligent vehicle-mounted LNG and CNG multi-channel solenoid valve monitoring system. The intelligent vehicle-mounted LNG and CNG multi-channel solenoid valve system includes a memory and a processor. The memory stores an intelligent vehicle-mounted LNG and CNG multi-channel solenoid valve method program. When the intelligent vehicle-mounted LNG and CNG multi-channel solenoid valve method program is executed by the processor, the steps of the intelligent vehicle-mounted LNG and CNG multi-channel solenoid valve method described in any one of the claims are implemented.

[0050] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0051] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0052] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0053] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0054] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0055] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for monitoring intelligent on-board LNG and CNG multi-channel solenoid valves, characterized in that, Includes the following steps: The coil voltage, current and temperature signals of each parallel solenoid valve are collected synchronously, and the signals are cleaned and aligned to generate multi-channel synchronous timing data. From the multi-channel synchronous time-series data, electrical and spectral characteristics characterizing the response of the solenoid valve are extracted, and a high-dimensional feature vector is formed by fusing temperature parameters. The high-dimensional feature vector is input into a lightweight Transformer-CNN hybrid network model. The CNN module captures the local details of the electrical characteristics, and the Transformer module models the long-range dependencies between multi-channel signals to obtain the valve body state confidence. Fault warning decisions are made based on the confidence level of the valve body state, and the vehicle-side operating data and decision results are uploaded to the cloud platform via a federated learning framework to trigger a global incremental update of the hybrid network model. The updated model parameters are then sent back to the vehicle to complete self-evolution.

2. The intelligent on-board LNG and CNG multi-channel solenoid valve monitoring method according to claim 1, characterized in that, The coil voltage, current, and temperature signals of each parallel solenoid valve are synchronously acquired, and the signals are cleaned and aligned to generate multi-channel synchronous timing data, specifically: Synchronously acquire signals from various sensors, including the voltage and current signals of the solenoid valve coil and the temperature signal acquired by the temperature sensor, to obtain the original synchronization signal sequence; The original synchronization signal sequence is filtered and denoised to generate a preprocessed signal sequence; The preprocessed signal sequence is subjected to outlier detection and removal, and missing data is filled by interpolation to form a clean signal sequence; A unified time reference is assigned to each of the clean signal sequences and time alignment is performed to generate a time-aligned multi-channel signal sequence. The time-aligned multi-channel signal sequence is reassembled according to the sampling time to construct multi-channel synchronous time-series data.

3. The intelligent on-board LNG and CNG multi-channel solenoid valve monitoring method according to claim 1, characterized in that, From the multi-channel synchronous time-series data, the electrical and spectral characteristics characterizing the solenoid valve response are extracted, and a high-dimensional feature vector is formed by fusing temperature parameters, specifically: Empirical mode decomposition is performed on the voltage and current signals in the synchronous time-series data to obtain a set of intrinsic mode functions. The Hilbert marginal spectral energy of each mode is then extracted as the nonlinear spectral energy distribution feature. Calculate the coherence coefficient of the transfer function of the current signal relative to the voltage signal, and combine the coherence coefficient of the transfer function with the differential inductance parameter to form the frequency domain impedance characteristic set of the solenoid valve; The temperature signal and voltage signal are coupled and analyzed to extract temperature-voltage joint modulation characteristic parameters; The high-dimensional feature vector is generated by fusing the nonlinear spectral energy distribution characteristics, frequency domain impedance feature set, and temperature-voltage joint modulation feature parameters through multi-domain feature fusion.

4. The intelligent on-board LNG and CNG multi-channel solenoid valve monitoring method according to claim 3, characterized in that, The coherence coefficient of the transfer function of the current signal relative to the voltage signal is calculated, and the coherence coefficient of the transfer function, together with the differential inductance parameter, constitutes the frequency domain impedance characteristic set of the solenoid valve, specifically: Complex analytic envelope extraction is performed on the voltage and current signals to obtain voltage envelope signals and current envelope signals; Based on the voltage envelope signal and the current envelope signal, the instantaneous phase difference is calculated by Hilbert transform to generate a phase coherence index; The current signal is subjected to high-order differential operation to extract the current change acceleration feature, and the current change acceleration feature is nonlinearly mapped to the instantaneous rate of change of the voltage signal to generate differential inductance parameters; Based on the phase coherence index and differential inductance parameters, the time-series characteristics are converted into a two-dimensional feature image through Gram angle field transformation to generate an impedance feature spectrum. The impedance feature map is input into a lightweight convolutional network for deep feature extraction to obtain a frequency domain impedance feature set.

5. The intelligent on-board LNG and CNG multi-channel solenoid valve monitoring method according to claim 1, characterized in that, The high-dimensional feature vector is input into a lightweight Transformer-CNN hybrid network model. The CNN module captures the local details of the electrical characteristics, and the Transformer module models the long-range dependencies between multi-channel signals to obtain the valve body state confidence. Specifically: The high-dimensional feature vector is adaptively sampled through a deformable one-dimensional convolutional layer to generate a geometrically adaptive local feature map. The local feature map is input into the multi-head causal convolution module to model temporal causal relationships and outputs a spatiotemporal correlation feature tensor. After encoding the spatiotemporal correlation feature tensor with relative position, it is input into the gated attention unit to calculate the causal attention weights between channels and generate attention-weighted context features. After residual connection of the context features and the original local features, feature filtering is performed through gated linear units to generate refined feature representations. The refined feature representation is input into the time series focusing module to calculate the feature importance weights at different time steps, and the feature importance weights at different time steps are used to perform probability divergence calculation based on the reference distribution to generate focusing coefficients. When the focusing coefficient exceeds a preset coefficient threshold, key timing features are enhanced, and finally, the valve body state confidence is generated through the sigmoid activation function.

6. The intelligent on-board LNG and CNG multi-channel solenoid valve monitoring method according to claim 5, characterized in that, The refined feature representation is input into the temporal focusing module to calculate the feature importance weights at different time steps. Then, using these feature importance weights, a probability divergence calculation based on a reference distribution is performed to generate focusing coefficients. Specifically: Temporal autocorrelation analysis is performed on the refined feature representation to extract the feature activation saliency at each time step and generate a temporal saliency mapping. Based on the aforementioned temporal saliency mapping, the feature importance weights for each time step are obtained through information entropy weighted calculation. The JS divergence of the feature importance weights and the preset feature importance prior distribution is used to obtain the time-series distribution difference measure. The time-series distribution difference metric is adaptively scaled and converted into a focusing coefficient.

7. The intelligent on-board LNG and CNG multi-channel solenoid valve monitoring method according to claim 1, characterized in that, The fault warning decision is made based on the confidence level of the valve body state, specifically as follows: An energy integral operation is performed on the confidence level of the valve body state using a sliding time window to generate a confidence level energy sequence; the Mahalanobis distance between the confidence level energy sequence and the normal operation reference template is calculated to generate a confidence level deviation index. The confidence deviation index is processed by nonlinear mapping and combined with real-time operating parameters for adaptive sensitivity adjustment, and the fault probability estimate is output. Based on the estimated fault probability, a multimodal decision fusion mechanism is used to generate the final hierarchical early warning instruction. The graded warning instructions are encoded into a standard CAN message format and sent to the main control ECU via the CAN vehicle bus. The main control ECU then triggers the corresponding audible and visual alarm response based on the message content.

8. The intelligent on-board LNG and CNG multi-channel solenoid valve monitoring method according to claim 1, characterized in that, Vehicle-side operational data and decision results are uploaded to the cloud platform via a federated learning framework to trigger a global incremental update of the hybrid network model. The updated model parameters are then distributed back to the vehicle to complete its self-evolution. Specifically: Local differential privacy processing is performed on the multi-channel synchronous time-series data and corresponding valve body state decision results in the vehicle-side operation data to generate a desensitized feature dataset. The local update gradient of the lightweight hybrid network model is calculated based on the desensitized feature dataset, and the local update gradient is pruned and quantized to generate an encrypted model update fragment. The encrypted model update fragments from multiple vehicle terminals are uploaded to the cloud platform and fused using a secure aggregation algorithm to generate a global model update gradient. The global model update gradient is used to incrementally update the cloud-based baseline model, generating parameters for a new generation of lightweight hybrid network models. After differential compression and signature verification of the new generation model parameters, they are distributed to each vehicle via OTA through the vehicle communication module to complete the self-evolution and update of the vehicle model.

9. An intelligent on-board LNG and CNG multi-channel solenoid valve monitoring system, characterized in that, The intelligent on-board LNG and CNG multi-way solenoid valve system includes a memory and a processor. The memory stores a method program for the intelligent on-board LNG and CNG multi-way solenoid valve. When the method program for the intelligent on-board LNG and CNG multi-way solenoid valve is executed by the processor, the steps of the intelligent on-board LNG and CNG multi-way solenoid valve method as described in any one of claims 1 to 8 are implemented.

Images