A big data-based carbon tank failure mode mining system

Abstract

The application discloses a kind of carbon tank failure mode mining systems based on big data, the present application relates to automobile emission control technical field, including: three-domain fusion perception layer, for obtaining the multi-source data of carbon tank in the running process, and the multi-source data is uniformly encoded, constructs the three-domain fusion feature space of the physical structure parameter of carbon tank, chemical adsorption characteristic and real-time operation condition.The carbon tank failure mode mining system based on big data, through multistage collaborative screening and evolution type feedback cycle verification framework, without increasing hardware cost, significantly improve the accuracy and early warning capability of carbon tank fault diagnosis, can effectively identify early weak fault characteristics that traditional method is difficult to capture, and accurately predict fault evolution path and remaining useful life.Overall, it realizes the upgrade of carbon tank fault from passive response maintenance to active predictive maintenance, reduces the risk of emission exceeding standard and vehicle operation failure rate caused by carbon tank failure.

Description

Technical Field

[0001] This invention relates to the field of automotive emission control technology, specifically to a carbon canister failure mode mining system based on big data. Background Technology

[0002] With the rapid development of the automotive industry, the carbon canister, as a key component of the automotive fuel evaporation control system, is widely used to reduce the environmental pollution caused by fuel evaporation. The carbon canister adsorbs fuel vapors from the fuel tank using activated carbon and desorbs them during engine operation, allowing them to enter the intake manifold for combustion, thus effectively reducing emissions. However, because the carbon canister operates under complex conditions for extended periods, it is susceptible to factors such as fuel vapor overload, activated carbon aging, blockage, or poor desorption, leading to system malfunctions and subsequently causing problems such as difficulty starting the vehicle, increased fuel consumption, and excessive emissions. Therefore, timely and accurate identification of carbon canister failure modes is crucial for ensuring the normal operation of automotive emission control systems and improving overall vehicle reliability and environmental performance.

[0003] Currently, fault diagnosis technology for carbon canisters mainly relies on limited sensor data collected by on-board diagnostic systems, such as fuel tank pressure and desorption flow rate, combined with preset threshold rules for judgment. Traditional diagnostic methods typically employ a single data source and fixed logical rules, such as monitoring the operating status of the desorption solenoid valve or changes in fuel tank pressure to determine whether the carbon canister is blocked or leaking. However, these methods have significant limitations. First, carbon canister fault modes are diverse and often concealed, making it difficult for a single data source to comprehensively reflect the deterioration process of its internal condition, resulting in early faults being difficult to detect. Second, traditional methods lack the ability to effectively fuse and analyze multi-source data, failing to uncover potential fault evolution patterns in carbon canisters during long-term operation. Finally, due to significant differences in operating conditions across different vehicle models and driving environments, diagnostic models based on fixed rules have poor generalization ability, easily leading to false alarms or missed alarms.

[0004] To address the above issues, the patent publication "CN120105295A" describes "A Deep Learning-Based Method for Equipment Fault Mode Recognition and Diagnosis." This method collects vibration, temperature, acoustic, current, and image data of equipment using multimodal sensors, employs a multi-head attention mechanism for multimodal feature fusion, and combines a bidirectional long short-term memory network for temporal modeling, effectively improving the accuracy and generalization ability of equipment fault diagnosis. However, this method is mainly aimed at fault diagnosis of general mechanical equipment. For the carbon canister, a specific component of an automotive evaporation control system, its fault modes exhibit obvious chemical adsorption characteristics and dynamic response features of the gas path. Existing deep learning models have not fully incorporated the physicochemical mechanisms of the carbon canister for targeted modeling, making it difficult to directly apply to early fault detection and pattern recognition of the carbon canister.

[0005] In summary, existing technologies lack a fault mode identification system that can fully integrate the working mechanism of the carbon canister and utilize multi-source big data for in-depth analysis. This invention aims to address these issues by proposing a big data-based carbon canister fault mode mining system. By collecting multi-source data throughout the entire lifecycle of the carbon canister and combining deep learning and data mining techniques, it achieves automatic identification and early warning of carbon canister fault modes, thereby improving the reliability and maintenance efficiency of automotive emission control systems. Summary of the Invention

[0006] The purpose of this invention is to provide a carbon canister failure mode mining system based on big data to solve the problems mentioned in the background art.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a carbon canister failure mode mining system based on big data, comprising: The three-domain fusion perception layer is used to acquire multi-source data during the operation of the carbon canister and uniformly encode the multi-source data to construct a three-domain fusion feature space that characterizes the physical structure parameters, chemical adsorption characteristics and real-time operating conditions of the carbon canister. A multi-level collaborative screening layer, connected to the three-domain fusion perception layer, is used to receive the three-domain fusion feature space and perform step-by-step screening of the three-domain fusion feature space based on screening modules with at least three different mechanisms to generate candidate fault features that have undergone multi-level verification. An evolutionary decision layer, connected to the multi-level collaborative screening layer, is used to receive the candidate fault features, extract and predict the fault evolution trajectory of the candidate fault features based on the temporal evolution network, and output preliminary fault mode and evolution trend information. A self-optimizing feedback loop, connected to both the evolutionary decision-making layer and the three-domain fusion perception layer, feeds back the preliminary fault modes and evolutionary trend information as update instructions to the three-domain fusion perception layer. This updates the boundary conditions of the three-domain fusion feature space and triggers at least two iterative verifications between the multi-level collaborative screening layer and the evolutionary decision-making layer until the output fault evolution trajectory converges, resulting in the final fault mode mining result. Real-time data interaction is achieved between modules at each level through a structured data interface. During the iterative verification process, the input parameters of each module are dynamically adjusted according to the feedback instructions, ensuring that the fault mode mining result matches the actual operating state of the carbon canister.

[0008] Preferably, the three-domain fusion perception layer includes: The physical parameter acquisition unit is used to acquire the physical structural parameters of the carbon canister, including activated carbon porosity, carbon canister volume, and desorption solenoid valve opening degree. A chemical property acquisition unit is used to acquire chemical adsorption characteristic parameters of the carbon canister, including desorption efficiency, heat of adsorption, and activated carbon aging index. The operating condition parameter acquisition unit is used to acquire the real-time operating condition parameters of the carbon canister, including ambient temperature, engine load, fuel tank pressure, and desorption flow rate. The feature encoding fusion unit is connected to the physical parameter acquisition unit, the chemical property acquisition unit, and the operating condition parameter acquisition unit, respectively. It is used to uniformly quantize the physical structure parameters, the chemical adsorption characteristic parameters, and the real-time operating condition parameters according to a preset encoding rule, generating the three-domain fusion feature space. The acquisition frequency of each type of parameter is synchronized with the sampling frequency of the vehicle-mounted sensors. The feature encoding process employs an appropriate quantization method for different types of parameters, ensuring the dimensional consistency and data validity of the three-domain fusion feature space.

[0009] Preferably, the multi-level collaborative screening layer includes: The first-level screening module is used to construct a digital twin of the carbon canister based on the physical adsorption model and chemical kinetic model of the carbon canister. The digital twin of the carbon canister is used to perform physical consistency verification on the multi-source data collected in real time and generate a residual feature spectrum as the first-level output feature. The second-level screening module is connected to the first-level screening module. It is used to receive the output features of the first-level module and, based on the working condition decoupled attention network, remove the influence weights of different driving conditions on the carbon canister state features to generate a working condition normalized fault feature vector. The third-level screening module, connected to the second-level screening module, receives the normalized fault feature vector under the operating conditions and performs a stress test on the stability of the normalized fault feature vector under simulated extreme operating conditions based on an adversarial verification network. Feature vectors that pass the stress test are output as candidate fault features to the evolutionary decision layer. The three-level screening module processes data layer by layer in the order of physical verification, operating condition decoupling, and anti-interference verification. The output data of each module provides a standardized input format for the next level module.

[0010] Preferably, the first-level screening module includes: A digital twin construction unit is used to construct a digital twin of the carbon canister based on the physical structure parameters and chemical adsorption characteristic parameters of the carbon canister. The digital twin of the carbon canister includes an adsorption model based on the Langmuir adsorption isotherm and a gas path dynamic response model based on the Ergun pressure drop equation. The residual feature spectrum generation unit, connected to the digital twin construction unit, is used to input real-time acquired multi-source data into the carbon tank digital twin, calculate the evolution spectrum of the deviation between actual sensor data and theoretical simulation values ​​over time, and output the evolution spectrum as the residual feature spectrum. The model parameters of the digital twin correspond one-to-one with the actual physicochemical parameters of the carbon tank, and the generation cycle of the residual feature spectrum is consistent with the data acquisition cycle, intuitively reflecting the deviation between the carbon tank's operating state and the theoretical state.

[0011] Preferably, the second-level screening module includes: The driving condition identification unit is used to perform cluster analysis on real-time operating condition parameters and identify the current driving condition type, which includes urban congestion condition, highway cruising condition and idling condition. An attention weight calculation unit, connected to the working condition identification unit, is used to dynamically calculate the contribution of each sensor feature to the fault mode under the current working condition based on the working condition type output by the working condition identification unit through a self-attention mechanism, and generate working condition adaptive attention weights. A normalized feature generation unit, connected to the attention weight calculation unit, is used to perform weighted normalization processing on the first-level output features according to the adaptive attention weight of the operating conditions, eliminating the influence of operating condition differences on fault features, and generating the operating condition normalized fault feature vector. The feature dimension of the operating condition clustering analysis is consistent with the dimension of the real-time operating condition parameters, and the calculation process of the attention weight is dynamically adjusted according to the operating condition type, allowing the normalized fault features to be compared and analyzed across operating conditions.

[0012] Preferably, the third-level screening module includes: The adversarial example generation unit is used to simulate various extreme working conditions and noisy environments to generate adversarial test samples; The stability verification unit is connected to the adversarial sample generation unit and the second-level screening module, respectively, and is used to compare and verify the normalized fault feature vector of the working condition with the adversarial test sample to determine the representation stability of the normalized fault feature vector of the working condition in the adversarial environment. The output filtering unit, connected to the stability verification unit, outputs feature vectors that pass stability verification as candidate fault features and discards feature vectors that fail stability verification. The generation dimension of the adversarial test samples is consistent with the dimension of the normalized fault feature vectors under operating conditions. The determination index for stability verification is based on the statistical analysis of historical operating data of the carbon canister, ensuring that the selected feature vectors can accurately characterize the actual fault state of the carbon canister.

[0013] Preferably, the evolutionary decision-making layer includes: An evolutionary trend memory network is used to perform time-series modeling on the candidate fault features and extract the full life cycle evolution law of carbon canister faults from the budding stage to the deterioration stage. The evolutionary trend memory network includes memory decay and reinforcement units, which are used to dynamically adjust the contribution weight of historical fault features to the current fault mode recognition. A fault evolution trajectory generation unit, connected to the evolution trend memory network, is used to generate a fault evolution trajectory based on the output of the evolution trend memory network. The fault evolution trajectory includes a fault type probability distribution and a remaining effective lifetime prediction range. The preliminary result output unit, connected to the fault evolution trajectory generation unit, is used to output the fault evolution trajectory as the preliminary fault mode and evolution trend information to the self-optimizing feedback loop. The training data of the evolution trend memory network comes from the operation and fault data of the carbon canister throughout its entire life cycle. The fault evolution trajectory continuously fits the candidate fault features in time series, fully presenting the degradation process of the carbon canister's condition.

[0014] Preferably, the self-optimizing feedback closed loop includes: The convergence judgment unit, connected to the preliminary result output unit, is used to determine whether the difference between the fault evolution trajectory output in the current cycle and the fault evolution trajectory output in the previous cycle is less than a preset convergence threshold. The feedback triggering unit, connected to the convergence judgment unit, is used to send the currently output preliminary fault mode and evolution trend information as an update instruction to the three-domain fusion perception layer when the convergence judgment unit determines that the difference is greater than or equal to the preset convergence threshold, so as to update the boundary conditions of the chemical adsorption characteristic parameters and real-time operating condition parameters in the three-domain fusion feature space. The loop counting unit is connected to the feedback triggering unit and the convergence judgment unit respectively. It is used to record the number of times the loop verification is executed, and when the number of executions reaches a preset upper limit, it forcibly outputs the current fault evolution trajectory as the final result. The final result output unit, connected to the convergence judgment unit, is used to output the current fault evolution trajectory as the final fault mode mining result when the difference determined by the convergence judgment unit is less than a preset convergence threshold. The convergence threshold is determined based on the historical data fluctuation range of the carbon canister fault evolution, and the upper limit of the loop count is set comprehensively based on the computational efficiency and result accuracy of data processing, taking into account both the accuracy and timeliness of fault mining.

[0015] Preferably, the update instruction triggered by the feedback triggering unit specifically includes: The probability distribution of fault types contained in the preliminary fault mode and evolution trend information is used as a prior probability and input into the feature encoding fusion unit of the three-domain fusion perception layer to adjust the weight allocation of the chemical adsorption characteristic parameters in the three-domain fusion feature space. The remaining effective life prediction range contained in the preliminary failure mode and evolution trend information is used as a dynamic constraint and input into the digital twin construction unit of the first-level screening module to update the decay rate parameter of the activated carbon aging index in the digital twin of the carbon canister. The weight adjustment magnitude of the chemisorption characteristic parameter is positively correlated with the value of the failure type probability distribution, and the update of the activated carbon aging index decay rate parameter is combined with the deviation ratio of the remaining effective life to calculate the correction factor.

[0016] Preferred options also include: A visualization interaction layer, connected to the self-optimizing feedback closed loop, is used to receive the final failure mode mining results and generate a carbon canister failure mode evolution map. The carbon canister failure mode evolution map displays the evolution path and key feature change nodes of the carbon canister from normal state to failure state in the form of a time axis. The operation and maintenance decision support unit, connected to the visualization interaction layer, is used to generate maintenance suggestion information based on the carbon canister failure mode evolution map. This maintenance suggestion information includes faulty component location, failure cause analysis, and maintenance priority ranking. The time axis scale of the failure mode evolution map matches the time granularity of data acquisition, and the generation of maintenance suggestion information achieves precise matching between the failure type and the actual state of failure evolution.

[0017] This invention provides a carbon canister failure mode mining system based on big data. It has the following beneficial effects: This big data-based carbon canister failure mode mining system achieves a technological leap from single data triggering to multi-dimensional evolution trend mining of carbon canister failure modes by constructing a deeply coupled architecture consisting of a three-domain fusion perception layer, a multi-level collaborative screening layer, an evolutionary decision-making layer, and a self-optimizing feedback closed loop.

[0018] This big data-based carbon canister failure mode mining system employs a multi-level collaborative screening and evolutionary feedback cyclical verification architecture. Without increasing hardware costs, it significantly improves the accuracy of carbon canister failure diagnosis and early warning capabilities. It effectively identifies early, subtle fault characteristics that are difficult to capture using traditional methods and accurately predicts fault evolution paths and remaining service life. Overall, it upgrades carbon canister failure maintenance from reactive response to proactive predictive maintenance, reducing the risk of emissions exceeding standards and vehicle malfunctions caused by carbon canister failures. It has broad industrial application value and promising prospects for widespread adoption. Attached Figure Description

[0019] Figure 1 This is a detailed flowchart of a carbon canister failure mode mining system based on big data according to the present invention. Figure 2This is a flowchart of the self-optimizing feedback closed-loop internal process of a carbon canister failure mode mining system based on big data according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1 and Figure 2 This invention provides a technical solution: a carbon canister failure mode mining system based on big data, comprising: The three-domain fusion perception layer is used to acquire multi-source data during the operation of the carbon canister and to uniformly encode the multi-source data to construct a three-domain fusion feature space that characterizes the physical structure parameters, chemical adsorption characteristics and real-time operating conditions of the carbon canister. A multi-level collaborative screening layer, connected to the three-domain fusion perception layer, is used to receive the three-domain fusion feature space and perform step-by-step screening of the three-domain fusion feature space based on screening modules with at least three different mechanisms to generate candidate fault features that have undergone multi-level verification. The evolutionary decision layer, connected to the multi-level collaborative screening layer, is used to receive candidate fault features and extract and predict fault evolution trajectories based on the temporal evolution network, outputting preliminary fault mode and evolution trend information. The self-optimizing feedback loop is connected to the evolutionary decision-making layer and the three-domain fusion perception layer, respectively. It is used to feed back the preliminary fault mode and evolution trend information as update instructions to the three-domain fusion perception layer to update the boundary conditions of the three-domain fusion feature space, and trigger the multi-level collaborative screening layer and the evolutionary decision-making layer to perform no less than two cyclic verifications until the output fault evolution trajectory converges and the final fault mode mining result is output.

[0022] It should be further explained that a big data-based carbon canister failure mode mining system achieves in-depth mining and dynamic tracking of carbon canister failure modes by constructing a linkage architecture consisting of a three-domain fusion perception layer, a multi-level collaborative screening layer, an evolutionary decision-making layer, and a self-optimizing feedback closed loop.

[0023] The three-domain fusion sensing layer is used to acquire multi-source data during the operation of the carbon canister and uniformly encode the multi-source data to construct a three-domain fusion feature space characterizing the physical structural parameters, chemical adsorption characteristics, and real-time operating conditions of the carbon canister. Specifically, this layer acquires the physical structural parameters of the carbon canister through the physical parameter acquisition unit, including activated carbon porosity, carbon canister volume, and desorption solenoid valve opening; acquires the chemical adsorption characteristic parameters of the carbon canister through the chemical characteristic acquisition unit, including desorption efficiency, adsorption heat, and activated carbon aging index; and acquires the real-time operating condition parameters of the carbon canister through the operating condition parameter acquisition unit, including ambient temperature, engine load, fuel tank pressure, and desorption flow rate.

[0024] The three types of data mentioned above are uniformly quantized by the feature coding fusion unit according to the preset coding rules to generate a three-domain fusion feature space with the unique physicochemical properties of the carbon canister.

[0025] A multi-level collaborative screening layer is connected to a three-domain fusion perception layer to receive the three-domain fusion feature space. Based on at least three different screening modules, the three-domain fusion feature space is progressively screened to generate candidate fault features that have undergone multi-level verification. The first-level screening module constructs a digital twin of the carbon canister based on its physical adsorption and chemical kinetic models. This digital twin is used to verify the physical consistency of real-time multi-source data, generating a residual feature spectrum as the first-level output feature.

[0026] The second-level screening module is connected to the first-level screening module. It is used to receive the output features of the first-level module and, based on the working condition decoupled attention network, remove the influence weights of different driving conditions on the carbon canister state features to generate a working condition normalized fault feature vector.

[0027] The third-level screening module is connected to the second-level screening module. It is used to receive the normalized fault feature vector under the working condition and perform stress testing on the stability of the normalized fault feature vector under simulated extreme working conditions based on the adversarial verification network. The feature vector that passes the stress test is output as the candidate fault feature.

[0028] An evolutionary decision-making layer connects to a multi-level collaborative screening layer to receive candidate fault features. Based on a temporal evolution network, it extracts and predicts the fault evolution trajectory of these features, outputting preliminary fault modes and evolutionary trend information. Specifically, an evolutionary trend memory network is used to perform temporal modeling of candidate fault features. This network includes memory decay and reinforcement units, which can dynamically adjust the contribution weight of historical fault features to the current fault mode recognition. This allows for the extraction of the entire lifecycle evolution law of carbon canister faults from the nascent stage to the deterioration stage, generating a fault evolution trajectory that includes the probability distribution of fault types and the remaining effective lifespan prediction range.

[0029] The self-optimizing feedback loop is connected to the evolutionary decision-making layer and the three-domain fusion perception layer, respectively. It is used to feed back the preliminary fault mode and evolution trend information as update instructions to the three-domain fusion perception layer to update the boundary conditions of the three-domain fusion feature space, and trigger the multi-level collaborative screening layer and the evolutionary decision-making layer to perform no less than two cyclic verifications until the output fault evolution trajectory converges and the final fault mode mining result is output.

[0030] Specifically, the convergence judgment unit in this closed loop is used to determine whether the difference between the fault evolution trajectory output by the current loop and the fault evolution trajectory output by the previous loop is less than a preset convergence threshold. The feedback triggering unit is used to send the preliminary fault mode and evolution trend information output by the current loop as an update instruction to the three-domain fusion perception layer when the convergence judgment unit determines that the difference is greater than or equal to the preset convergence threshold, so as to update the boundary conditions of the chemical adsorption characteristic parameters and real-time operating condition parameters in the three-domain fusion feature space.

[0031] The loop counting unit records the number of times the loop verification is executed and forces the current fault evolution trajectory as the final result when the number of executions reaches a preset upper limit. The final result output unit outputs the current fault evolution trajectory as the final fault mode mining result when the difference determined by the convergence judgment unit is less than a preset convergence threshold.

[0032] Through the deep coupling of the above three-domain fusion perception, multi-level collaborative screening, evolutionary decision-making and self-optimizing feedback loop, a technological leap has been achieved from single threshold triggering of carbon canister failures to multi-dimensional evolution trend mining.

[0033] The three-domain fusion perception layer includes: The physical parameter acquisition unit is used to acquire the physical structural parameters of the carbon canister, including activated carbon porosity, carbon canister volume, and desorption solenoid valve opening. The chemical property acquisition unit is used to acquire the chemical adsorption characteristic parameters of the carbon canister, including desorption efficiency, heat of adsorption, and activated carbon aging index. The operating condition parameter acquisition unit is used to acquire the real-time operating condition parameters of the carbon canister, including ambient temperature, engine load, fuel tank pressure, and desorption flow rate. The feature encoding fusion unit is connected to the physical parameter acquisition unit, the chemical property acquisition unit, and the operating condition parameter acquisition unit, respectively. It is used to uniformly quantize the physical structure parameters, chemical adsorption characteristic parameters, and real-time operating condition parameters according to the preset encoding rules to generate a three-domain fusion feature space.

[0034] It should be further explained that the three-domain fusion perception layer achieves unified representation and fusion encoding of multi-source heterogeneous data from the carbon tank through the collaborative work of the physical parameter acquisition unit, chemical characteristic acquisition unit, operating condition parameter acquisition unit, and feature encoding fusion unit.

[0035] The physical parameter acquisition unit is used to acquire the physical structural parameters of the carbon canister, including activated carbon porosity, carbon canister volume, and desorption solenoid valve opening. Activated carbon porosity is determined by mercury intrusion porosimetry or nitrogen adsorption and characterized by pore diameter distribution data; the carbon canister volume is calculated from the geometric dimensions of the internal cavity of the carbon canister shell; and the desorption solenoid valve opening is acquired in real time by a valve core position sensor and recorded as a percentage or angle value.

[0036] The chemical property acquisition unit is used to acquire the chemical adsorption characteristic parameters of the carbon canister, including desorption efficiency, heat of adsorption, and activated carbon aging index. Desorption efficiency is calculated by comparing the change in the mass of fuel vapor in the carbon canister before and after desorption; heat of adsorption is measured by a calorimeter, and the amount of heat released by the activated carbon during adsorption is recorded in joules per mole; the activated carbon aging index is calculated based on a weighted function of the cumulative usage time of the carbon canister and the cumulative number of desorptions, and is used to characterize the degree of decay of the activated carbon's adsorption capacity.

[0037] The operating parameter acquisition unit is used to acquire real-time operating parameters of the carbon canister, including ambient temperature, engine load, fuel tank pressure, and desorption flow rate. Ambient temperature is acquired by a thermocouple sensor located near the carbon canister; engine load is read by the engine control unit and characterized by throttle opening or injection pulse width; fuel tank pressure is monitored in real time by a pressure sensor installed on top of the fuel tank; and desorption flow rate is measured by a flow sensor installed in the desorption line and recorded in standard liters per minute.

[0038] The feature encoding fusion unit is connected to the physical parameter acquisition unit, chemical property acquisition unit, and operating condition parameter acquisition unit, respectively, and is used to uniformly quantify the above three types of parameters according to preset encoding rules. Specifically, for continuous numerical parameters such as desorption flow rate and ambient temperature, the maximum and minimum value normalization method is used to map them to a preset numerical range; for discrete parameters such as desorption solenoid valve opening, one-heat encoding or embedded encoding is used to convert them into vector form; for time series parameters such as fuel tank pressure fluctuation, the sliding window statistical feature extraction method is used to calculate its mean, variance, and rate of change.

[0039] The transformed feature vectors are then concatenated along three dimensions: the physical structure of the carbon canister, its chemical adsorption characteristics, and its real-time operating conditions, forming a three-domain fusion feature space with carbon canister-specific attributes. This feature space is stored as a multidimensional tensor and serves as input data for subsequent multi-level collaborative screening layers.

[0040] The multi-level collaborative screening layers include: The first-level screening module is used to construct a digital twin of the carbon canister based on the physical adsorption model and chemical kinetic model of the carbon canister. The digital twin of the carbon canister is used to perform physical consistency verification on the multi-source data collected in real time and generate residual feature spectrum as the first-level output feature. The second-level screening module is connected to the first-level screening module. It is used to receive the output features of the first-level module and, based on the working condition decoupled attention network, remove the influence weights of different driving conditions on the carbon canister state features to generate a working condition normalized fault feature vector. The third-level screening module, connected to the second-level screening module, is used to receive the normalized fault feature vector under operating conditions and to perform stress tests on the stability of the normalized fault feature vector under simulated extreme operating conditions based on the adversarial verification network. The feature vector that passes the stress test is output as a candidate fault feature to the evolutionary decision layer.

[0041] It should be further explained that the multi-level collaborative screening layer specifically achieves multi-dimensional purification and verification of the three-domain integrated feature space through a progressive screening mechanism consisting of a first-level screening module, a second-level screening module, and a third-level screening module.

[0042] The first-level screening module is used to construct a digital twin of the carbon canister based on the physical adsorption model and chemical kinetic model of the carbon canister. This digital twin is based on the physical structural parameters and chemical adsorption characteristic parameters of the carbon canister, integrating the Langmuir adsorption isotherm model to describe the adsorption equilibrium relationship of activated carbon on fuel vapor, integrating the Ergun pressure drop equation model to simulate the pressure loss characteristics of gas flowing through the carbon canister's packed bed, and incorporating the Arrhenius equation which describes the temperature dependence of desorption rate, forming a virtual mirror model capable of dynamically responding to changes in operating conditions.

[0043] During real-time operation, the first-level screening module inputs the currently collected multi-source data into the carbon canister digital twin, driving the digital twin to synchronously calculate the expected output values ​​under theoretical conditions, including theoretical desorption flow rate, theoretical fuel tank pressure, and theoretical internal temperature of the carbon canister. Then, the actual sensor-collected values ​​are compared with the theoretical expected values ​​at time points, the residuals between the two are calculated, and the relationship between the residuals and time is plotted as a residual feature spectrum. This residual feature spectrum, recorded in time series form, effectively filters out fluctuations under normal operating conditions and sensor noise, highlighting characteristic information reflecting abnormal physicochemical processes inside the carbon canister, and is transmitted as the first-level output feature to the second-level screening module.

[0044] The second-level screening module is connected to the first-level screening module. It receives the output features from the first-level module and performs deep processing on these features based on a driving condition decoupled attention network. This driving condition decoupled attention network consists of an input layer, a driving condition encoding layer, an attention weight calculation layer, and a feature normalization layer. The driving condition encoding layer performs cluster analysis on the real-time collected driving condition parameters to identify the specific type of the current driving condition and output a driving condition label.

[0045] The attention weight calculation layer takes the first-level output features and operating condition labels as input. It calculates the contribution weight of each sensor feature dimension under the current operating condition using a multi-head self-attention mechanism. This weight reflects the sensitivity of different features to fault modes under different operating conditions. The feature normalization layer performs weighted summation and vector normalization on the first-level output features based on the calculated attention weights, eliminating feature distribution shifts caused by changes in operating conditions and generating a condition-normalized fault feature vector. This vector eliminates the influence of operating condition differences on fault features, ensuring that the same fault mode has a consistent representation under different operating conditions.

[0046] The third-level screening module is connected to the second-level screening module and is used to receive the normalized fault feature vector under operating conditions, and to perform stability stress testing on the vector based on the adversarial verification network. The adversarial verification network includes an adversarial example generation unit and a discriminant verification unit. The adversarial example generation unit simulates extreme operating conditions and sensor fault environments by introducing Gaussian white noise, random spectral cropping, and time series misalignment, generating multiple adversarial test samples.

[0047] The discrimination and verification unit compares the similarity of the normalized fault feature vector under the operating condition with each adversarial test sample, and calculates the degree of shift of the feature vector under adversarial interference. If the feature vector can still maintain a similarity higher than a preset threshold with the original feature vector in a preset number of adversarial samples, it is determined to have passed the stability test, and the feature vector is output as a candidate fault feature to the evolutionary decision layer; if it fails the test, the feature vector is discarded and the first-level screening module is triggered to re-collect data segments for a new round of screening.

[0048] Through the above three-level screening and stress testing mechanism, it is ensured that the candidate fault characteristics that finally enter the evolutionary decision-making layer have physical consistency, operational robustness and anti-interference stability.

[0049] The first-level screening module includes: The digital twin building unit is used to construct a digital twin of the carbon canister based on its physical structure parameters and chemical adsorption characteristics. The digital twin of the carbon canister includes an adsorption model based on the Langmuir adsorption isotherm and a gas path dynamic response model based on the Ergun pressure drop equation. The residual feature spectrum generation unit, connected to the digital twin construction unit, is used to input real-time multi-source data into the carbon canister digital twin, calculate the evolution spectrum of the deviation between the actual sensor data and the theoretical simulation value over time, and output the evolution spectrum as the residual feature spectrum.

[0050] It should be further explained that the first-level screening module specifically achieves the physical consistency verification between the real-time operating data of the carbon tank and the theoretical expected state through the collaborative work of the digital twin construction unit and the residual feature spectrum generation unit.

[0051] The digital twin building block is used to construct a digital twin of the carbon canister based on its physical structure and chemisorption characteristics. The construction process is based on the carbon canister's geometry and bed parameters. First, an adsorption model is built based on the Langmuir adsorption isotherm. This model uses the Langmuir equation to describe the monolayer adsorption behavior on the activated carbon surface, with the equation in the form of: Where q is the equilibrium adsorption amount, q m Let q be the monolayer saturated adsorption capacity, K be the adsorption equilibrium constant, and P be the adsorbate partial pressure. q was obtained by experimentally measuring adsorption isotherm data at different temperatures and fitting the data. m The values ​​of K and K serve as fundamental parameters for the chemisorption characteristics of the carbon canister.

[0052] Secondly, a dynamic response model for the gas path is constructed based on the Ergun pressure drop equation. This model uses the Ergun equation to describe the pressure loss of gas flowing through the carbon canister-filled bed. The equation is in the form of: Where ΔP is the pressure drop, L is the bed height, ε is the porosity, μ is the gas dynamic viscosity, v is the apparent velocity, and d p Let ρ be the equivalent diameter of the activated carbon particles and ρ be the gas density. By substituting the activated carbon porosity and particle size distribution parameters of the carbon canister into this equation, a mapping relationship between the airflow resistance and flow rate inside the carbon canister is established.

[0053] Furthermore, the digital twin building blocks also integrate a desorption rate model based on the Arrhenius equation, which describes the effect of temperature on the desorption rate constant, in the form: Where k is the desorption rate constant, A is the pre-exponential factor, and E a R is the desorption activation energy, T is the gas constant, and T is the absolute temperature. Through the coupling of the above three models, a digital twin of the carbon canister is formed that can dynamically respond to changes in operating conditions. This digital twin is deployed on edge computing devices or cloud servers in the form of computer executable code. It receives real-time data such as ambient temperature, engine load, fuel tank pressure, and desorption flow as input and simultaneously calculates the expected output value under theoretical conditions.

[0054] The residual feature spectrum generation unit is connected to the digital twin construction unit. It is used to input multi-source data collected in real time into the carbon canister digital twin, drive the digital twin to run and output the theoretical expected value. Then, the actual sensor collected value is compared with the theoretical expected value at time points to calculate the deviation between the two. For continuous parameters such as desorption flow rate and fuel tank pressure, the residual is calculated in the form of absolute deviation or relative deviation. For discrete parameters such as desorption solenoid valve opening, the residual is calculated in the form of state matching degree.

[0055] All the calculated residuals are arranged in chronological order to form a multidimensional residual time series. This residual time series is stored in the form of a two-dimensional matrix. The rows of the matrix correspond to different sensor types, and the columns of the matrix correspond to different sampling time points. Each element in the matrix represents the residual value of the sensor at the corresponding time. At the same time, the residual time series is segmented using a sliding window method. The length of each segment is a preset time window, and the step size is half the length of the window.

[0056] The residual data within each time window is plotted as a two-dimensional image or a three-dimensional surface plot to form a residual feature spectrum. This residual feature spectrum is output in a visual image format or a numerical matrix format and passed as the first-level output feature to the second-level screening module to characterize the degree of deviation between the actual operating state of the carbon tank and the ideal physicochemical process.

[0057] The second-level screening module includes: The driving condition identification unit is used to perform cluster analysis on real-time operating condition parameters and identify the current driving condition type, which includes urban congestion condition, highway cruising condition and idling condition. The attention weight calculation unit, connected to the working condition identification unit, is used to dynamically calculate the contribution of each sensor feature to the fault mode under the current working condition based on the working condition type output by the working condition identification unit, and generate working condition adaptive attention weights through a self-attention mechanism. The normalization feature generation unit, connected to the attention weight calculation unit, is used to perform weighted normalization processing on the first-level output features according to the adaptive attention weight based on the working conditions, thereby eliminating the influence of working condition differences on fault features and generating a working condition normalized fault feature vector.

[0058] It should be further explained that the second-level screening module specifically achieves the decoupling and elimination of the working condition coupling components in the first-level output features through the coordinated operation of the working condition identification unit, the attention weight calculation unit, and the normalized feature generation unit.

[0059] The driving condition identification unit performs cluster analysis on real-time acquired driving condition parameters to identify the current driving condition type. This unit first acquires driving condition parameters from the three-domain fusion perception layer, including engine load, vehicle speed, ambient temperature, and fuel tank pressure fluctuation frequency, and constructs a multi-dimensional driving condition feature vector. Then, a Gaussian mixture model is used to estimate the probability density of the driving condition feature vector. The mean, covariance, and weight parameters of the Gaussian components are iteratively calculated using the expectation-maximization algorithm, assigning the driving condition feature vector at each sampling time to the Gaussian component with the highest probability. Each Gaussian component corresponds to a preset driving condition type, including urban congestion, highway cruising, idling, and rapid acceleration. The driving condition identification unit outputs the driving condition label and its confidence level at the current moment in real time.

[0060] The attention weight calculation unit is connected to the operating condition identification unit. It dynamically calculates the contribution of each sensor feature to the fault mode under the current operating condition based on the operating condition labels output by the operating condition identification unit using a self-attention mechanism. This unit takes the first-level output feature, i.e., the residual feature spectrum, as input. It splits the residual feature spectrum into multiple feature channels according to sensor type, with each feature channel corresponding to a residual time series of a specific sensor data type. Simultaneously, it converts the operating condition labels into operating condition embedding vectors and concatenates them with the feature vectors of each feature channel to form a feature sequence labeled with the operating condition.

[0061] Then, a multi-head self-attention mechanism is used to calculate the attention score between each element in the feature sequence. The specific calculation process includes linearly mapping the input feature sequence into a query matrix, a key matrix, and a value matrix. Attention weights are calculated by the dot product of the query matrix and the key matrix. After softmax normalization, these weights are weighted and summed with the value matrix to obtain the attention weight of each feature channel under the current operating condition. This weight reflects the relative importance of different sensor residual features for fault identification. The attention weight calculation unit outputs the attention weight vector for each feature channel.

[0062] The normalization feature generation unit is connected to the attention weight calculation unit. It performs weighted normalization on the first-level output features based on the attention weight vector to generate a normalized fault feature vector for the operating condition. This unit multiplies each feature channel of the first-level output features by its corresponding attention weight to form a weighted feature matrix. Then, layer normalization is performed on the weighted feature matrix, i.e., subtracting the mean and dividing by the standard deviation of the feature vector at each time step, so that the feature vectors have a uniform scale distribution while maintaining the relative relationships between channels.

[0063] Furthermore, the normalization feature generation unit employs a residual connection structure, adding the original features before normalization to the normalized features element-wise, preserving the effective information in the original features. This ultimately generates a normalized fault feature vector containing decoupling information related to the operating conditions. This vector is output as a multidimensional tensor to the third-level screening module as input data for subsequent stability testing.

[0064] Through the above-mentioned operating condition identification and adaptive weighting mechanism, the influence of different driving conditions on the carbon canister state characteristics is effectively removed, so that the same fault mode presents a consistent feature representation under different operating conditions.

[0065] The third-level screening module includes: The adversarial example generation unit is used to simulate various extreme working conditions and noisy environments to generate adversarial test samples; The stability verification unit is connected to the adversarial example generation unit and the second-level screening module, respectively. It is used to compare and verify the normalized fault feature vector under the working condition with the adversarial test sample to determine the stability of the representation of the normalized fault feature vector under the adversarial environment. The output filtering unit, connected to the stability verification unit, is used to output the feature vectors that pass the stability verification as candidate fault features and discard the feature vectors that fail the stability verification.

[0066] It should be further explained that the third-level screening module achieves robust screening of the normalized fault feature vector under extreme disturbances through the coordinated operation of the adversarial sample generation unit, the stability verification unit, and the screening output unit.

[0067] The adversarial example generation unit simulates various extreme operating conditions and noisy environments to generate adversarial test samples. This unit first collects atypical operating condition segments recorded in historical operating data, including scenarios such as drastic fluctuations in fuel tank pressure, sudden drops in desorption flow, and abrupt changes in ambient temperature, using these segments as base samples. Then, various perturbation operations are applied to the base samples to generate adversarial examples.

[0068] The perturbation operations include adding Gaussian white noise to the time-series data, with the noise intensity controlled by the signal-to-noise ratio (SNR), ranging from 5 dB to 20 dB; randomly cropping the spectrum of acoustic or vibration signals, i.e., randomly deleting some frequency components before reconstructing the time-domain signal; and performing time axis misalignment or local time stretching / compression on the time-series data to simulate asynchronous sensor sampling or transmission delays. Furthermore, adversarial attack methods from generative adversarial networks (GANs) are employed to calculate small perturbations in the feature vector along the gradient direction of the loss function, generating adversarial examples that maximally alter the classification results. These adversarial examples are generated in batches, with each example having the same dimension as the original normalized fault feature vector, forming an adversarial test set.

[0069] The stability verification unit is connected to the adversarial example generation unit and the second-level screening module, respectively. It is used to compare and verify the normalized fault feature vector of the working condition with the adversarial test samples to determine the stability of the feature vector in the adversarial environment. The unit first inputs the original normalized fault feature vector of the working condition and each adversarial sample into a preset feature similarity calculation model, calculates the cosine similarity or Euclidean distance between the two, and obtains the similarity score.

[0070] Simultaneously, a pre-trained fault classifier is used to classify and predict the original feature vector and each adversarial example, calculating the consistency of the classification results, i.e., the proportion of predicted categories of the original feature vector and the adversarial example. The stability verification unit weights and fuses the similarity score and classification consistency to obtain a stability index. Verification is considered successful when the stability index exceeds a preset threshold. The preset threshold is set based on the statistical distribution of historical data, typically selecting a quantile value that keeps the false positive rate below 5%.

[0071] The filtering output unit is connected to the stability verification unit and is used to output the feature vectors that pass the stability verification as candidate fault features to the evolutionary decision layer. Feature vectors that fail the verification are discarded, and a resampling command is sent to the first-level screening module, triggering the first-level screening module to reselect data segments of adjacent time windows to regenerate the residual feature spectrum and re-enter the screening process.

[0072] Through the aforementioned adversarial testing and screening mechanisms, it is ensured that the candidate fault characteristics that ultimately enter the evolutionary decision-making layer possess characterization stability under extreme operating conditions and sensor anomalies, thereby avoiding misdiagnosis caused by accidental disturbances.

[0073] Evolutionary decision-making layers include: The evolutionary trend memory network is used to perform temporal modeling of candidate fault features and extract the evolutionary pattern of carbon canister faults from the initiation stage to the deterioration stage. The evolutionary trend memory network includes memory decay and reinforcement units, which are used to dynamically adjust the contribution weight of historical fault features to the current fault mode recognition. The fault evolution trajectory generation unit is connected to the evolution trend memory network and is used to generate fault evolution trajectories based on the output of the evolution trend memory network. The fault evolution trajectory includes the fault type probability distribution and the remaining effective lifetime prediction range. The preliminary result output unit is connected to the fault evolution trajectory generation unit and is used to output the fault evolution trajectory as preliminary fault mode and evolution trend information to the self-optimizing feedback closed loop.

[0074] It should be further explained that the evolutionary decision-making layer achieves the leap from static identification to dynamic evolutionary analysis of candidate fault features through the deep coupling of the evolutionary trend memory network, the fault evolution trajectory generation unit, and the preliminary result output unit.

[0075] An evolutionary trend memory network is used to perform temporal modeling of candidate fault features to extract the full life cycle evolution of carbon canister faults from the nascent stage to the deterioration stage. This network is based on an improved gated recurrent unit architecture, introducing memory decay and reinforcement units on top of the traditional recurrent neural network. The memory decay unit uses an exponential decay function to weight historical hidden states, and the decay rate parameter is correlated with the real-time monitoring value of the activated carbon aging index of the carbon canister. This ensures that the further back in time a historical state is from the current moment, the lower its contribution to the current fault identification, simulating the natural degradation of the physicochemical properties of the carbon canister over time.

[0076] The memory enhancement unit uses an attention mechanism to amplify the gain of key features within a specific time window. When a continuous decrease in desorption flow or abnormal fluctuation in fuel tank pressure is detected, the enhancement unit automatically increases the retention weight of the features in memory during that period, so that key turning points in the fault evolution path are effectively captured.

[0077] The evolutionary trend memory network comprises a forward propagation path and a backward propagation path. The forward path processes candidate fault feature sequences in chronological order, generating hidden states for each time step; the backward path processes the same sequence in reverse chronological order, generating reverse hidden states. The forward and reverse hidden states are concatenated and input into a fully connected layer to obtain the fault type probability distribution for each time step.

[0078] The network also incorporates gating mechanisms to control the inflow and forgetting of information, specifically including update gates, reset gates, and output gates. The update gate determines how much information from the previous hidden state is retained in the current time step, the reset gate determines how the current input is combined with historical memory, and the output gate generates an output vector based on the current hidden state. The activation functions for these gating units are the sigmoid and tanh functions. The network weights are trained using a backpropagation algorithm, and the loss function is a weighted sum of cross-entropy and mean squared error to simultaneously optimize classification accuracy and evolutionary trajectory fitting precision.

[0079] The fault evolution trajectory generation unit is connected to the evolution trend memory network to generate fault evolution trajectories based on the network's output. This unit first extracts the hidden states from the network's output at each time step and maps them to a high-dimensional feature space. Then, Gaussian process regression is used to fit the hidden state sequence, generating a continuous function form of the fault evolution trajectory. This trajectory, with time on the horizontal axis and fault probability or severity on the vertical axis, visually demonstrates the gradual degradation of the carbon canister from a normal state to the fault threshold.

[0080] The fault evolution trajectory includes a fault type probability distribution, i.e., the probability of occurrence of each type of fault at each time point; it also includes a remaining effective lifetime prediction interval. The remaining effective lifetime is obtained by calculating the time required for the current state to reach a preset fault threshold along the evolution trajectory. The prediction interval uses the Monte Carlo dropout method to estimate uncertainty and generate a confidence interval range for the remaining effective lifetime.

[0081] The preliminary results output unit is connected to the fault evolution trajectory generation unit, and is used to output the fault evolution trajectory as preliminary fault mode and evolution trend information to the self-optimizing feedback closed loop. This unit encapsulates the fault evolution trajectory in the form of structured data, including trajectory start time, trajectory end time, number of sampling points, fault type probability vector of each sampling point, remaining effective lifetime prediction value and its confidence interval.

[0082] Simultaneously, the attention weight matrix of the intermediate layer of the evolutionary trend memory network is visualized and encoded to generate fault-critical node markers, indicating the time points and corresponding feature channels that lead to a sudden increase in fault probability or a change in evolutionary direction. The data packets output by the above preliminary result output unit serve as input to the self-optimizing feedback loop, triggering subsequent cyclic verification processes.

[0083] Through the collaboration of the aforementioned evolutionary trend memory network and fault evolution trajectory generation unit, a paradigm shift has been achieved in the identification of carbon canister faults from discrete events to the tracking of continuous evolution processes.

[0084] The self-optimizing feedback loop includes: The convergence judgment unit, connected to the preliminary result output unit, is used to determine whether the difference between the fault evolution trajectory output in the current loop and the fault evolution trajectory output in the previous loop is less than the preset convergence threshold. The feedback triggering unit, connected to the convergence judgment unit, is used to send the current output preliminary fault mode and evolution trend information as an update instruction to the three-domain fusion perception layer when the convergence judgment unit determines that the difference is greater than or equal to the preset convergence threshold, so as to update the boundary conditions of the chemical adsorption characteristic parameters and real-time operating condition parameters in the three-domain fusion feature space. The loop counting unit is connected to the feedback triggering unit and the convergence judgment unit respectively. It is used to record the number of times the loop verification is executed, and when the number of executions reaches the preset upper limit, it forces the output of the current fault evolution trajectory as the final result. The final result output unit is connected to the convergence judgment unit. When the difference determined by the convergence judgment unit is less than the preset convergence threshold, the current fault evolution trajectory is output as the final fault mode mining result.

[0085] It should be further explained that the self-optimizing feedback closed loop achieves cyclic verification and dynamic correction of the preliminary fault mode and evolution trend information through the close cooperation of the convergence judgment unit, feedback trigger unit, loop counting unit and final result output unit, until the final fault mode mining result of convergence is output.

[0086] The convergence judgment unit is connected to the preliminary result output unit of the evolutionary decision layer. It is used to receive the preliminary fault mode and evolution trend information output in the current cycle, compare it with the preliminary fault mode and evolution trend information output in the previous cycle, and determine whether the difference between the two is less than the preset convergence threshold.

[0087] This unit first extracts the fault type probability distribution sequence and remaining effective lifetime prediction interval from the fault evolution trajectory output by the current loop, and simultaneously extracts the corresponding data stored in the previous loop. Then, it calculates the dynamic time warping distance between the two trajectories. The dynamic time warping distance is obtained by nonlinearly aligning the two time series and then calculating the cumulative Euclidean distance, which can effectively measure the similarity between evolution trajectories with different lengths or phase shifts.

[0088] Simultaneously, the absolute difference of the median of the remaining effective lifetime prediction interval and the cosine similarity of the failure type probability vectors of the two trajectories at key time nodes are calculated. The above dynamic time-normalized distance, absolute difference of remaining effective lifetime, and cosine similarity are weighted and summed to obtain a comprehensive difference index. When the comprehensive difference index is lower than a preset convergence threshold, the convergence judgment unit determines that the current evolution trajectory has converged. The preset convergence threshold is determined based on the statistical analysis of the fluctuation range of the same carbon canister's multiple cycle outputs in historical data, typically selected as half of the upper limit of the fluctuation range.

[0089] The feedback triggering unit is connected to the convergence judgment unit. When the convergence judgment unit determines that the comprehensive difference index is greater than or equal to the preset convergence threshold, it sends the preliminary fault mode and evolution trend information of the current loop output as an update instruction to the three-domain fusion perception layer.

[0090] Specifically, the feedback triggering unit uses the fault type probability distribution contained in the currently output fault evolution trajectory as a priori probability and transmits it in the form of a data packet to the feature encoding fusion unit of the three-domain fusion perception layer. This feature encoding fusion unit adjusts the weight allocation of chemisorption characteristic parameters in the three-domain fusion feature space based on the received prior probabilities. The adjustment method involves mapping the fault type probability distribution to weighting coefficients and weighting the encoded values ​​of parameters such as activated carbon aging index and desorption efficiency, so that feature dimensions highly correlated with the current fault mode receive greater attention in subsequent feature fusion.

[0091] Simultaneously, the feedback triggering unit transmits the currently output remaining effective lifespan prediction interval as a dynamic constraint to the digital twin construction unit of the first-level screening module. This digital twin construction unit updates the decay rate parameter of the activated carbon aging index in the carbon canister digital twin based on the received remaining effective lifespan prediction interval. The update method involves comparing the median of the remaining effective lifespan prediction interval with the remaining lifespan simulated by the current digital twin, calculating a correction factor, and multiplying the correction factor by the original decay rate parameter to align the aging process of the digital twin with the actual monitored fault evolution trend.

[0092] The loop counting unit is connected to both the feedback triggering unit and the convergence judgment unit, and is used to record the number of loop verification executions triggered for data within the same time period. This unit has an internal counter; each time the feedback triggering unit sends an update command, the counter value increases by one. Simultaneously, the loop counting unit stores a preset upper limit for the number of loops. This upper limit is determined based on the carbon canister failure evolution cycle and computational resource consumption, and ranges from three to five times.

[0093] When the counter value reaches the upper limit of the loop count, even if the convergence judgment unit has not yet determined convergence, the loop counting unit will forcibly terminate the loop verification process and send a forced output command to the final result output unit to avoid infinite loop due to local data fluctuations.

[0094] The final result output unit is connected to the convergence judgment unit and the loop counting unit. When the convergence judgment unit determines that the comprehensive difference index is less than the preset convergence threshold, the current fault evolution trajectory is output as the final fault mode mining result; or when the loop counting unit triggers the forced output command, the fault evolution trajectory of the current loop output is output as the final result.

[0095] The final output unit encapsulates the final failure mode mining results in a structured data format, including the carbon canister identifier, diagnostic timestamp, final failure type label, failure evolution trajectory curve data, remaining effective life prediction value, and confidence interval. It also outputs a record of the number of iterations and a convergence process log for subsequent analysis and traceability.

[0096] Through the aforementioned self-optimizing feedback closed-loop cyclic verification mechanism, the failure mode mining results have undergone a qualitative change from single output to multi-round iterative convergence, ensuring the stability and reliability of the final output results.

[0097] The update instructions triggered by the feedback triggering unit specifically include: The probability distribution of fault types contained in the initial fault mode and evolution trend information is used as a prior probability and input into the feature encoding fusion unit of the three-domain fusion perception layer to adjust the weight distribution of chemical adsorption characteristic parameters in the three-domain fusion feature space. The remaining effective life prediction range contained in the preliminary failure mode and evolution trend information is used as a dynamic constraint and input into the digital twin building unit of the first-level screening module to update the decay rate parameter of the activated carbon aging index in the digital twin of the carbon canister.

[0098] It should be further explained that the update instructions triggered by the feedback triggering unit specifically include weight adjustment operations for the feature encoding fusion unit in the three-domain fusion perception layer and model parameter correction operations for the digital twin construction unit in the first-level screening module.

[0099] For the weight adjustment operation of the feature encoding fusion unit, the feedback triggering unit analyzes the probability distribution of the initial fault mode and the fault type in the evolution trend information of the current loop output, extracts the probability value corresponding to each type of fault, and maps the probability value to the feature dimension weighting coefficient.

[0100] Specifically, for fault types related to activated carbon aging, their probability values ​​are assigned to the feature dimensions corresponding to chemical characteristic parameters such as activated carbon aging index and desorption efficiency, so that these dimensions obtain higher weight coefficients in the subsequent feature fusion process; for fault types related to gas path blockage, their probability values ​​are assigned to the feature dimensions corresponding to physical structural parameters such as activated carbon porosity and desorption solenoid valve opening, so that the proportion of these dimensions in the feature space is increased.

[0101] The feature encoding fusion unit recalibrates the currently constructed three-domain fusion feature space based on the received weighting coefficients. The calibration method involves multiplying the original encoded value of each feature dimension by the corresponding weighting coefficient, and then performing vector normalization to generate an updated three-domain fusion feature space.

[0102] For the model parameter correction operation of the digital twin building unit, the feedback triggering unit analyzes the preliminary fault mode output by the current loop and the remaining effective lifetime prediction interval in the evolution trend information, extracting the interval median and interval width. It then compares this interval median with the theoretical remaining lifetime output by the current digital twin simulation, calculating the deviation ratio between the two. Simultaneously, it determines the correction step size based on the interval width; a smaller correction step size is used when the interval width is narrow to ensure stability, and a larger correction step size is used when the interval width is wide to accelerate convergence.

[0103] The digital twin building unit updates the degradation rate parameter of the activated carbon aging index in the digital twin of the carbon canister based on the received deviation ratio and correction step size. The update method involves multiplying the original degradation rate parameter by a correction factor, calculated as: Correction factor = 1 + (Deviation ratio × Correction step size). When the actual monitored remaining lifetime is shorter than the theoretical simulation value, the deviation ratio is positive, the degradation rate parameter increases, and the aging process of the digital twin accelerates; when the actual monitored remaining lifetime is longer than the theoretical simulation value, the deviation ratio is negative, the degradation rate parameter decreases, and the aging process of the digital twin slows down.

[0104] Through the aforementioned weight adjustment and parameter correction operations, the feature construction process of the three-domain fusion perception layer and the digital twin simulation process of the first-level screening module can dynamically respond to changes in the fault evolution trend, achieving bidirectional alignment and collaborative optimization between front-end data representation and back-end fault mining results.

[0105] Also includes: The visualization interaction layer, connected to the self-optimizing feedback closed loop, is used to receive the final failure mode mining results and generate a carbon canister failure mode evolution map. The carbon canister failure mode evolution map displays the evolution path and key feature change nodes of the carbon canister from normal state to failure state in the form of a time axis. The operation and maintenance decision support unit, connected to the visualization interaction layer, is used to generate maintenance suggestion information based on the carbon canister failure mode evolution map. The maintenance suggestion information includes faulty component location, failure cause analysis, and maintenance priority ranking.

[0106] It should be further explained that it also includes a visualization interaction layer and an operation and maintenance decision support unit, which are used to present the final failure mode mining results in an intuitive form and provide maintenance personnel with maintenance decision assistance.

[0107] The visualization interaction layer is connected to the final result output unit of the self-optimizing feedback loop to receive the final failure mode mining results. These results include the canister identifier, diagnostic timestamp, final failure type label, failure evolution trajectory curve data, remaining usable life prediction value, and confidence interval. The visualization interaction layer inputs this data into the graphics rendering engine to generate a canister failure mode evolution map.

[0108] This graph plots the failure evolution curve of the carbon canister from the initial monitoring time to the current time, with time as the horizontal axis and failure probability or failure severity as the vertical axis. Different colors are used to distinguish various failure types, including activated carbon aging failure, gas path blockage failure, desorption valve sticking failure, and fuel vapor leakage failure. The curves corresponding to each failure type are displayed in parallel on the time axis.

[0109] The graph simultaneously marks key feature change nodes in the fault evolution trajectory. These key feature change nodes are highlighted as markers, each accompanied by a pop-up information box displaying the corresponding original sensor data segment, residual feature spectrum image, and attention weight distribution of each sensor feature's contribution to the fault at that moment. The evolution graph employs an interactive design, supporting zooming, panning, and click-to-query operations. Maintenance personnel can view detailed diagnostic information at any time point by clicking on a touchscreen or mouse.

[0110] The operation and maintenance decision support unit connects to the visualization interaction layer to automatically generate maintenance recommendations based on the carbon canister failure mode evolution map. This unit has a built-in fault-repair association knowledge base, which is stored in a relational database. Each record contains the fault type, a list of fault causes, recommended maintenance measures, a list of required tools and spare parts, and a maintenance priority coefficient.

[0111] The operation and maintenance decision support unit first reads the final fault type label and fault evolution trajectory slope at the current moment from the evolution graph. The fault evolution trajectory slope reflects the rate of fault deterioration. Then, it queries the knowledge base using the fault type label as an index to match the corresponding fault cause and maintenance measures. At the same time, it adjusts the maintenance priority coefficient according to the fault evolution trajectory slope. When the fault evolution trajectory slope exceeds a preset rate threshold, the maintenance priority is automatically increased by one level.

[0112] Maintenance recommendations are generated in the form of a structured report. The report includes faulty component location information, which is determined based on the physical location of the sensor channel with the highest weight in the attention weight distribution; the fault cause analysis section lists the three most likely causes that match the current fault type and their confidence levels; the maintenance priority ranking section arranges multiple faults to be repaired from high to low urgency, and each fault is accompanied by a suggested maintenance time window, which is calculated by subtracting the safety margin from the remaining effective life prediction value.

[0113] The operation and maintenance decision support unit pushes the generated maintenance suggestion information to the interface of the visualization interaction layer and displays it on the same screen as the evolution map. It also supports exporting the report as a portable document format or sending it to the maintenance work order system via message queue.

[0114] Through the collaboration between the aforementioned visualization interaction layer and the operation and maintenance decision support unit, the complete transformation of carbon canister failure mode mining results from data to decision-making was achieved.

[0115] This system achieves a technological leap from single data triggering to multi-dimensional evolution trend mining of carbon canister failure modes by constructing a deeply coupled architecture consisting of a three-domain fusion perception layer, a multi-level collaborative screening layer, an evolutionary decision-making layer, and a self-optimizing feedback loop.

[0116] The three-domain fusion perception layer uniformly encodes the physical structural parameters, chemical adsorption characteristics, and real-time operating conditions of the carbon canister, constructing a feature space with carbon canister-specific attributes, and providing a multi-dimensional data foundation for subsequent fault identification.

[0117] The multi-level collaborative screening layer employs a three-tiered progressive mechanism: physical consistency verification based on digital twins, operational condition impact stripping based on operational condition decoupling attention networks, and stability stress testing based on adversarial verification networks. This mechanism purifies and verifies the original features step by step, effectively filtering out interference from normal operational condition fluctuations and sensor noise on fault features. This ensures that candidate fault features entering the evolutionary decision-making layer possess physical consistency, operational condition robustness, and anti-interference stability.

[0118] The evolutionary decision layer uses an evolutionary trend memory network containing memory decay and reinforcement units to perform time-series modeling of candidate fault features, extracts the full life cycle evolution law of carbon canister faults from the budding stage to the deterioration stage, generates fault evolution trajectory and remaining effective life prediction range, and realizes a paradigm shift from discrete fault event identification to continuous fault evolution process tracking.

[0119] The self-optimizing feedback loop feeds back the initial fault modes and evolution trend information as update instructions to the three-domain fusion perception layer, triggering at least two cyclic verifications until the evolution trajectory converges. This forms a two-way alignment and collaborative optimization mechanism between the front-end data representation and the back-end fault mining results, resulting in higher stability and reliability of the final output fault mode mining results.

[0120] The overall technical solution of this system, through a multi-level collaborative screening and evolutionary feedback cyclical verification architecture, significantly improves the accuracy of carbon canister fault diagnosis and early warning capability without increasing hardware costs. It can effectively identify early and subtle fault characteristics that are difficult to capture by traditional methods, and accurately predict the fault evolution path and remaining service life.

[0121] The visualization and interaction layer presents the final failure mode mining results intuitively in the form of a carbon canister failure mode evolution map, enabling maintenance personnel to quickly locate the key areas and evolution nodes where the failure occurred. The maintenance decision support unit automatically generates maintenance suggestions based on the failure type and evolution trajectory slope, including fault component location, failure cause analysis, and maintenance priority ranking, providing data support for maintenance decisions.

[0122] Overall, it has achieved an upgrade from passive response maintenance to proactive predictive maintenance for carbon canister failures, reducing the risk of emissions exceeding standards and vehicle malfunctions caused by carbon canister failures, and has broad industrial application value and promotion prospects.

[0123] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus 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, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0124] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A carbon canister failure mode mining system based on big data, characterized in that, include: The three-domain fusion perception layer is used to acquire multi-source data during the operation of the carbon canister and uniformly encode the multi-source data to construct a three-domain fusion feature space that characterizes the physical structure parameters, chemical adsorption characteristics and real-time operating conditions of the carbon canister. A multi-level collaborative screening layer, connected to the three-domain fusion perception layer, is used to receive the three-domain fusion feature space and perform step-by-step screening of the three-domain fusion feature space based on screening modules with at least three different mechanisms to generate candidate fault features that have undergone multi-level verification. An evolutionary decision layer, connected to the multi-level collaborative screening layer, is used to receive the candidate fault features, extract and predict the fault evolution trajectory of the candidate fault features based on the temporal evolution network, and output preliminary fault mode and evolution trend information. The self-optimizing feedback loop is connected to the evolutionary decision-making layer and the three-domain fusion perception layer, respectively. It is used to feed back the preliminary fault mode and evolution trend information as update instructions to the three-domain fusion perception layer to update the boundary conditions of the three-domain fusion feature space, and trigger the multi-level collaborative screening layer and the evolutionary decision-making layer to perform no less than two cyclic verifications until the output fault evolution trajectory converges and the final fault mode mining result is output.

2. The carbon canister failure mode mining system based on big data according to claim 1, characterized in that: The three-domain fusion perception layer includes: The physical parameter acquisition unit is used to acquire the physical structural parameters of the carbon canister, including activated carbon porosity, carbon canister volume, and desorption solenoid valve opening degree. A chemical property acquisition unit is used to acquire chemical adsorption characteristic parameters of the carbon canister, including desorption efficiency, heat of adsorption, and activated carbon aging index. The operating condition parameter acquisition unit is used to acquire the real-time operating condition parameters of the carbon canister, including ambient temperature, engine load, fuel tank pressure, and desorption flow rate. The feature encoding fusion unit is connected to the physical parameter acquisition unit, the chemical property acquisition unit, and the operating condition parameter acquisition unit, respectively, and is used to uniformly quantize the physical structure parameters, the chemical adsorption property parameters, and the real-time operating condition parameters according to a preset encoding rule to generate the three-domain fusion feature space.

3. The carbon canister failure mode mining system based on big data according to claim 1, characterized in that: The multi-level collaborative screening layer includes: The first-level screening module is used to construct a digital twin of the carbon canister based on the physical adsorption model and chemical kinetic model of the carbon canister. The digital twin of the carbon canister is used to perform physical consistency verification on the multi-source data collected in real time and generate a residual feature spectrum as the first-level output feature. The second-level screening module is connected to the first-level screening module. It is used to receive the output features of the first-level module and, based on the working condition decoupled attention network, remove the influence weights of different driving conditions on the carbon canister state features to generate a working condition normalized fault feature vector. The third-level screening module, connected to the second-level screening module, is used to receive the normalized fault feature vector under the operating conditions, and to perform stress tests on the stability of the normalized fault feature vector under simulated extreme operating conditions based on an adversarial verification network. The feature vector that passes the stress test is output as the candidate fault feature to the evolutionary decision layer.

4. The carbon canister failure mode mining system based on big data according to claim 3, characterized in that: The first-level screening module includes: A digital twin construction unit is used to construct a digital twin of the carbon canister based on the physical structure parameters and chemical adsorption characteristic parameters of the carbon canister. The digital twin of the carbon canister includes an adsorption model based on the Langmuir adsorption isotherm and a gas path dynamic response model based on the Ergun pressure drop equation. The residual feature spectrum generation unit, connected to the digital twin construction unit, is used to input real-time acquired multi-source data into the carbon canister digital twin, calculate the evolution spectrum of the deviation between actual sensor data and theoretical simulation values ​​over time, and output the evolution spectrum as the residual feature spectrum.

5. The carbon canister failure mode mining system based on big data according to claim 3, characterized in that: The second-level screening module includes: The driving condition identification unit is used to perform cluster analysis on real-time operating condition parameters and identify the current driving condition type, which includes urban congestion condition, highway cruising condition and idling condition. An attention weight calculation unit, connected to the working condition identification unit, is used to dynamically calculate the contribution of each sensor feature to the fault mode under the current working condition based on the working condition type output by the working condition identification unit through a self-attention mechanism, and generate working condition adaptive attention weights. The normalized feature generation unit, connected to the attention weight calculation unit, is used to perform weighted normalization processing on the first-level output features according to the working condition adaptive attention weight, to eliminate the influence of working condition differences on fault features, and to generate the working condition normalized fault feature vector.

6. The carbon canister failure mode mining system based on big data according to claim 3, characterized in that: The third-level screening module includes: The adversarial example generation unit is used to simulate various extreme working conditions and noisy environments to generate adversarial test samples; The stability verification unit is connected to the adversarial sample generation unit and the second-level screening module, respectively, and is used to compare and verify the normalized fault feature vector of the working condition with the adversarial test sample to determine the representation stability of the normalized fault feature vector of the working condition in the adversarial environment. The filtering output unit, connected to the stability verification unit, is used to output the feature vectors that pass the stability verification as the candidate fault features and discard the feature vectors that fail the stability verification.

7. The carbon canister failure mode mining system based on big data according to claim 6, characterized in that: The evolutionary decision-making layer includes: An evolutionary trend memory network is used to perform time-series modeling on the candidate fault features and extract the full life cycle evolution law of carbon canister faults from the budding stage to the deterioration stage. The evolutionary trend memory network includes memory decay and reinforcement units, which are used to dynamically adjust the contribution weight of historical fault features to the current fault mode recognition. A fault evolution trajectory generation unit, connected to the evolution trend memory network, is used to generate a fault evolution trajectory based on the output of the evolution trend memory network. The fault evolution trajectory includes a fault type probability distribution and a remaining effective lifetime prediction range. The preliminary result output unit is connected to the fault evolution trajectory generation unit and is used to output the fault evolution trajectory as the preliminary fault mode and evolution trend information to the self-optimizing feedback closed loop.

8. The carbon canister failure mode mining system based on big data according to claim 7, characterized in that: The self-optimizing feedback closed loop includes: The convergence judgment unit, connected to the preliminary result output unit, is used to determine whether the difference between the fault evolution trajectory output in the current cycle and the fault evolution trajectory output in the previous cycle is less than a preset convergence threshold. The feedback triggering unit, connected to the convergence judgment unit, is used to send the currently output preliminary fault mode and evolution trend information as an update instruction to the three-domain fusion perception layer when the convergence judgment unit determines that the difference is greater than or equal to the preset convergence threshold, so as to update the boundary conditions of the chemical adsorption characteristic parameters and real-time operating condition parameters in the three-domain fusion feature space. The loop counting unit is connected to the feedback triggering unit and the convergence judgment unit respectively. It is used to record the number of times the loop verification is executed, and when the number of executions reaches a preset upper limit, it forcibly outputs the current fault evolution trajectory as the final result. The final result output unit is connected to the convergence judgment unit and is used to output the current fault evolution trajectory as the final fault mode mining result when the convergence judgment unit determines that the difference is less than the preset convergence threshold.

9. A carbon canister failure mode mining system based on big data according to claim 8, characterized in that: The update instructions triggered by the feedback triggering unit specifically include: The probability distribution of fault types contained in the preliminary fault mode and evolution trend information is used as a prior probability and input into the feature encoding fusion unit of the three-domain fusion perception layer to adjust the weight allocation of the chemical adsorption characteristic parameters in the three-domain fusion feature space. The remaining effective life prediction range contained in the preliminary failure mode and evolution trend information is used as a dynamic constraint and input into the digital twin construction unit of the first-level screening module to update the decay rate parameter of the activated carbon aging index in the digital twin of the carbon canister.

10. A carbon canister failure mode mining system based on big data according to claim 1, characterized in that: Also includes: A visualization interaction layer, connected to the self-optimizing feedback closed loop, is used to receive the final failure mode mining results and generate a carbon canister failure mode evolution map. The carbon canister failure mode evolution map displays the evolution path and key feature change nodes of the carbon canister from normal state to failure state in the form of a time axis. The operation and maintenance decision support unit, connected to the visualization interaction layer, is used to generate maintenance suggestion information based on the carbon canister failure mode evolution map. The maintenance suggestion information includes fault component location, failure cause analysis, and maintenance priority ranking.

Images