Automobile carbon deposition driving fault diagnosis method based on deep learning

Abstract

The invention discloses an automobile carbon deposition driving fault diagnosis method based on deep learning, and relates to the technical field of intelligent diagnosis. Deploying a multi-target dynamic reward function to synchronously calculate a power loss rate reward value, a carbon deposition coverage rate penalty value, a combustion efficiency gradient value and a carbon deposition growth rate gradient value, and outputting a power attenuation diagnosis state through a value iteration optimization strategy network; and receiving the power attenuation diagnosis state and the carbon deposition type classification code through the student model, loading the pre-trained teacher model, adding a confrontation discriminator in the distillation loss function, comparing the characteristic graph distribution difference of the teacher model and the student model, and outputting a grading treatment instruction matched with the vehicle state. According to the invention, by constructing the reinforcement learning agent and deploying the multi-target dynamic reward function, the comprehensive dynamic quantitative evaluation of the carbon deposition driving fault is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnostic technology, and in particular to a method for diagnosing automotive carbon buildup-related faults based on deep learning. Background Technology

[0002] In recent years, deep learning methods have shown significant potential in engine fault diagnosis, particularly in image recognition-based carbon deposit classification and performance degradation assessment based on sensor time-series data analysis. Convolutional neural networks (CNNs) are widely used to extract visual features of carbon deposits from engine endoscope images, while recurrent neural networks (RNNs) or Transformer architectures are used to process time-series sensor data such as cylinder pressure waveforms, torque, and exhaust gas concentration to capture dynamic patterns of power performance changes. These technologies provide a foundation for building intelligent automotive fault diagnosis systems.

[0003] However, existing technical solutions for diagnosing carbon deposit faults generally suffer from modal fragmentation and insufficient dynamic correlation modeling. Relying solely on image analysis makes it difficult to accurately quantify the actual impact of carbon deposits on instantaneous combustion efficiency and torque output; while relying solely on sensor time-series data cannot intuitively characterize the spatial distribution and evolution of carbon deposits. This fragmentation makes it difficult for diagnostic models to accurately assess the real-time, multi-dimensional comprehensive impact of carbon deposits on the engine power system, particularly failing to dynamically quantify the complex coupling relationship between carbon deposit growth and power attenuation, thus hindering the generation of accurate diagnostic and remediation decisions. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a deep learning-based method for diagnosing automotive carbon buildup-driven faults, solving the problem that existing technologies cannot dynamically quantify the multi-dimensional impact of carbon buildup on the powertrain.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a deep learning-based method for diagnosing automotive carbon buildup-driven faults, comprising: real-time acquisition of carbon buildup image sequences and power performance sensor data of the vehicle; inputting the carbon buildup image sequences into a CNN convolutional layer, extracting local spatial features of carbon buildup morphology through convolutional and pooling layers, and outputting a carbon buildup type classification code through a fully connected layer and a Softmax classifier; inputting the power performance sensor data into an independent Transformer encoder to generate a temporal feature vector, fusing it with the local spatial features extracted by the CNN through a cross-attention mechanism to construct a spatiotemporal fusion feature map; constructing a reinforcement learning agent, using the spatiotemporal fusion feature map as state input, deploying a multi-objective dynamic reward function to synchronously calculate the power loss rate reward value, carbon buildup coverage penalty value, combustion efficiency gradient value, and carbon buildup growth rate gradient value, and outputting a power attenuation diagnostic state through a value iteration optimization strategy network; receiving the power attenuation diagnostic state and carbon buildup type classification code through a student model, loading a pre-trained teacher model, adding an adversarial discriminator in the distillation loss function, comparing the feature map distribution differences between the teacher model and the student model, and outputting a graded treatment instruction matching the vehicle state.

[0007] As a preferred embodiment of the deep learning-based automotive carbon buildup-driven fault diagnosis method of the present invention, the carbon buildup image sequence includes an engine intake valve image sequence, a fuel injector image sequence, and a combustion chamber surface deposit image sequence; the power performance sensor data includes real-time waveforms from the cylinder pressure sensor, timing data from the crankshaft torque sensor, and O2, CO2, and NOx concentration change curves.

[0008] As a preferred embodiment of the deep learning-based automotive carbon buildup-driven fault diagnosis method of the present invention, the specific steps are as follows: Local spatial features of carbon buildup morphology are extracted through convolutional and pooling layers, and a carbon buildup type classification code is output via a fully connected layer and a Softmax classifier. The carbon deposit image sequence is input into a three-dimensional convolutional layer to extract multi-scale spatial features. The multi-scale spatial features are then reduced in dimensionality by a max pooling layer to generate a carbon deposit feature vector. The carbon deposit feature vector is input into the fully connected layer to calculate the nonlinear combination weights; The Softmax classifier is used to output the probability distribution of dry carbon, wet carbon and annular deposition types based on nonlinear combined weights. The type with the highest probability is marked as the carbon deposition type classification code.

[0009] As a preferred embodiment of the deep learning-based automotive carbon buildup-driven fault diagnosis method of the present invention, the specific steps for constructing the spatiotemporal fusion feature map are as follows: The dynamic performance sensor data is divided into fixed-length time windows and input into the Transformer encoder; The cross-cycle dependence of cylinder pressure fluctuation, torque decay and exhaust gas composition is modeled by a multi-head self-attention mechanism, and a time-series feature vector with timestamp alignment is output. Using temporal feature vectors as query vectors, the local spatial features extracted by CNN are used to generate key and value vectors through independent linear layers; The similarity weights of the Query vector and Key vector are calculated through a cross-attention layer, and the Value vector is weighted and summed based on the similarity weights to construct a spatiotemporal fusion feature map.

[0010] As a preferred embodiment of the deep learning-based automotive carbon buildup-driven fault diagnosis method of the present invention, the specific steps for constructing the reinforcement learning agent are as follows: Initialize the policy network parameters and set the greedy exploration coefficients, using the spatiotemporal fusion feature map as the environmental state input to the policy network; The strategy network outputs action commands containing the cleaning intensity level and duration. After executing the action commands, new dynamic performance sensor data is acquired via the vehicle bus. The next-moment environmental state is generated based on the new dynamic performance sensor data, and the policy network parameters are updated by calculating the total reward value through a multi-objective dynamic reward function.

[0011] As a preferred embodiment of the deep learning-based automotive carbon buildup-driven fault diagnosis method of the present invention, the construction of the multi-objective dynamic reward function includes the following specific steps: The power loss rate bonus value is calculated based on the difference between the real-time waveform of the cylinder pressure sensor and the standard waveform. Carbon deposit coverage penalty values ​​are obtained by pixel-level segmentation of carbon deposit image sequences; Combustion efficiency gradient values ​​are generated based on the rate of change of the sliding window ratio of the O2 and CO2 concentration change curves. The difference in carbon deposition coverage area between adjacent time windows is calculated using the sliding window difference algorithm to generate the gradient value of carbon deposition growth rate. The total reward value is calculated by weighting and integrating the power loss rate reward value, carbon deposit coverage rate penalty value, combustion efficiency gradient value, and carbon deposit growth rate gradient value.

[0012] As a preferred embodiment of the deep learning-based automotive carbon buildup-driven fault diagnosis method of the present invention, the specific steps of outputting the power attenuation diagnostic status through a value iteration optimization strategy network are as follows: The action value function of the current state is calculated using the Bellman equation, and the action value function is updated through temporal difference learning and stored in the experience replay buffer. The advantage function value is calculated by sampling action value values ​​from the experience replay buffer, and the policy gradient correction is generated based on the linear combination of the advantage function value and the action value value. The gradient ascent algorithm is used to optimize the policy network parameters based on the policy gradient correction amount. The iteration is repeated until the state value function value converges, and the output action command probability distribution is marked as the dynamic decay diagnostic state.

[0013] As a preferred embodiment of the deep learning-based automotive carbon buildup-driven fault diagnosis method of the present invention, the pre-trained teacher model refers to a two-layer Transformer network trained based on carbon buildup treatment cases in a standard vehicle repair database. The specific steps for outputting hierarchical management instructions that match the vehicle status are as follows: The diagnostic status of power attenuation and the type of carbon deposit are classified, encoded, and concatenated into a diagnostic fusion feature vector, which is then input into the student model to output the student feature vector. The pre-trained teacher model is invoked to perform forward computation on the diagnostic fusion feature vector, outputting the teacher feature vector. An adversarial discriminant is constructed in the feature space of the distillation loss function, and the cosine similarity of the feature maps of the teacher feature vector and the student feature vector is calculated. An adversarial loss function is constructed based on the cosine similarity of feature maps, and the parameters of the student model are optimized using the backpropagation algorithm. The student model output layer generates the probability distribution of governance instructions from level P1 to level P6, and sends the instruction with the highest probability as the hierarchical governance instruction to the vehicle control unit.

[0014] The beneficial effects of this invention are as follows: By constructing a reinforcement learning agent and deploying a multi-objective dynamic reward function, a comprehensive dynamic quantitative assessment of carbon deposit-driven faults is achieved. The reinforcement learning agent receives a spatiotemporal fusion feature map that integrates the spatial morphological characteristics of carbon deposits and the temporal dynamic characteristics of power performance as environmental state input. After executing the cleaning action command output by the policy network, the multi-objective dynamic reward function simultaneously calculates four key indicators: a power loss rate reward value based on cylinder pressure waveform differences to generate power recovery parameters; a carbon deposit coverage penalty value based on image segmentation to quantify the deposition removal effect; a combustion efficiency gradient value based on the rate of change of exhaust gas components to generate combustion optimization trend parameters; and a carbon deposit growth rate gradient value based on the difference in coverage area to quantify the deposition suppression effect. The weighted fusion of the four dimensions of reward values ​​is calculated simultaneously to construct a real-time quantitative assessment framework that comprehensively reflects the impact of carbon deposit status and power performance. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of a deep learning-based method for diagnosing automotive carbon buildup-driven faults.

[0017] Figure 2 This is a flowchart of spatiotemporal feature fusion.

[0018] Figure 3 Workflow diagram for reinforcement learning intelligent agents.

[0019] Figure 4 Generate a flowchart for hierarchical governance instructions. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0023] Reference Figure 1 This is one embodiment of the present invention, which provides a deep learning-based method for diagnosing automotive carbon buildup drive faults, comprising the following steps: S1: Real-time acquisition of carbon buildup image sequences and power performance sensor data from vehicles.

[0024] The carbon deposit image sequence includes an engine intake valve image sequence, a fuel injector image sequence, and a combustion chamber surface deposit image sequence; the power performance sensor data includes real-time waveforms from the cylinder pressure sensor, timing data from the crankshaft torque sensor, and O2, CO2, and NOx concentration change curves.

[0025] S1.1: Capture carbon deposit image sequences using endoscopic equipment; the carbon deposit image sequences include engine intake valve image sequences, fuel injector image sequences, and combustion chamber surface deposit image sequences.

[0026] Specifically, an endoscope probe is inserted into the engine intake valve area to continuously record video while the vehicle is idling or running stably, and the video frame rate is fixed to generate an engine intake valve image sequence.

[0027] It should be noted that the capture process continues to cover at least one complete engine cycle to ensure that the engine intake valve image sequence contains deposit states at multiple time points.

[0028] The endoscope probe is moved to the fuel injector area, and continuous video recording is performed under the same operating conditions to generate a sequence of fuel injector images. The fuel injector surface is focused during the continuous video recording to capture any blockages or deposits.

[0029] An endoscope probe is inserted into the combustion chamber, and video is recorded when the vehicle is accelerated to a moderate load. The captured video is converted into an image frame sequence, and the image sequence of the deposits on the surface of the combustion chamber is output.

[0030] S1.2: Power performance sensor data acquisition uses standard sensors (cylinder pressure sensor, crankshaft torque sensor and exhaust gas sensor) to record real-time waveforms of cylinder pressure sensor, timing data of crankshaft torque sensor, and O2, CO2 and NOx concentration change curves.

[0031] Specifically, connect the cylinder pressure sensor to the engine cylinder, the crankshaft torque sensor to the crankshaft position, and the exhaust gas sensors (O2 sensor, CO2 sensor, and NOx sensor) to the exhaust pipe, and calibrate all sensors to ensure the accuracy of sensor data.

[0032] When the vehicle starts running and transitions from idle to full load, the cylinder pressure sensor monitors cylinder pressure changes in real time and records the real-time waveform at a high sampling rate, covering at least three engine cycles. Under the same operating conditions, the crankshaft torque sensor measures torque changes in real time, records time-series data, and stores it as a time-torque value sequence, with the time-series data timestamped synchronized with the real-time waveform of the cylinder pressure sensor. The exhaust gas sensor monitors the concentrations of O2, CO2, and NOx in the exhaust pipe in real time, recording concentration change curves. The data point sampling interval is fixed, and the capture process covers stable vehicle operating conditions as well as transient conditions such as acceleration and deceleration.

[0033] All sensor data is transmitted via the vehicle bus (such as CAN bus), and timestamp alignment and standardization of multi-source data are performed to generate a unified dataset containing real-time waveforms from cylinder pressure sensors, timing data from crankshaft torque sensors, and O2, CO2, and NOx concentration change curves.

[0034] S2: Input the carbon deposit image sequence into the CNN convolutional layer, extract the local spatial features of the carbon deposit morphology through the convolutional and pooling layers, and output the carbon deposit type classification code through the fully connected layer and the Softmax classifier.

[0035] S2.1: Input the carbon deposit image sequence into a three-dimensional convolutional layer to extract multi-scale spatial features, and use a max pooling layer to perform dimensionality reduction processing on the multi-scale spatial features to generate a carbon deposit feature vector.

[0036] Specifically, the engine intake valve image sequence, fuel injector image sequence, and combustion chamber surface deposit image sequence from the carbon deposit image sequence are aligned by time frame and the pixel values ​​are normalized to... scope.

[0037] By continuously scanning the normalized carbon deposit image sequence using a 3D convolution kernel, and by covering the temporal and spatial dimensions through the kernel size, the spatiotemporal correlation of carbon deposit morphology is captured. The expression is as follows: ; In the formula, This represents the feature map output by the first convolutional layer. This indicates a modified linear unit activation function. The input is a sequence of carbon deposit images. for Convolution kernel size, This represents a 3D convolution operation.

[0038] Max pooling is performed on the multi-scale spatial features extracted by the convolution kernel in the 3D convolutional layer. The pooling window size is 2×2×1, which compresses the feature dimension while preserving significant texture information. The expression is as follows: ; In the formula, This represents the carbon deposition feature vector output by the max-pooling layer. This represents the max pooling operation, which selects the maximum value within a local region. This represents the feature map input to the pooling layer.

[0039] S2.2: Input the carbon deposit feature vector into the fully connected layer to calculate the nonlinear combination weights.

[0040] Specifically, the carbon deposit feature vector is flattened into a one-dimensional vector and used as the input to the fully connected layer; Two fully connected layers (first layer and second layer) are used to calculate nonlinear combinations through a weight matrix; The first layer has an output dimension of 512 and uses the ReLU activation function; the second layer has an output dimension of 128 and generates a weight vector. .

[0041] S2.3: The Softmax classifier is used to output the probability distribution of dry carbon, wet carbon and annular deposition types based on nonlinear combined weights, and the type corresponding to the highest probability is marked as the carbon deposition type classification code.

[0042] Specifically, the weight vector output by the fully connected layer Input the Softmax classifier to calculate the probability of three types of carbon deposition (dry carbon, wet carbon, and annular deposition), expressed as follows: ; In the formula, This represents the probability value indicating the type of carbon deposit. This indicates dry carbon, wet carbon, and ring deposition. The weight vector at the output of the fully connected layer represents the weight of the first element. Dimension weights, This indicates that the weighted score is subjected to an exponential operation. The sum of the exponents representing the scores of the three types of carbon deposits is used for probability normalization. This represents the loop index variable in the summation operation, used to iterate through the weighted scores of the three carbon deposition types.

[0043] By comparing the probability values ​​of dry carbon, wet carbon, and annular deposition, the type corresponding to the highest probability is marked as the carbon deposition type classification code.

[0044] S3: Input the dynamic performance sensor data into an independent Transformer encoder to generate a temporal feature vector. This vector is then fused with the local spatial features extracted by the CNN through a cross-attention mechanism to construct a spatiotemporal fusion feature map. For details, please refer to [link to relevant documentation]. Figure 2 .

[0045] S3.1: Divide the dynamic performance sensor data into fixed-length time windows and input them into the Transformer encoder.

[0046] Specifically, standardization processing is performed on the data from each power performance sensor. The values ​​of cylinder pressure waveform, torque time series data, and exhaust gas concentration curve are scaled to the same dimension range through zero-mean and unit variance transformation to ensure the consistency of subsequent processing of multimodal data.

[0047] The standardized dynamic performance sensor data is divided into sequence windows with a fixed time length.

[0048] It should be noted that the time window length is fixed to ensure that all data points are aligned with the timestamps. The number of windows covers the entire data sequence to avoid omissions.

[0049] The split time window data is input into an independent Transformer encoder, which contains encoder stack layers and a fixed number of layers.

[0050] S3.2: Model the cross-cycle dependence of cylinder pressure fluctuation, torque decay and exhaust gas composition through a multi-head self-attention mechanism, and output a time-series feature vector with timestamp alignment characteristics.

[0051] Inside the Transformer encoder, a multi-head self-attention mechanism is used to calculate the cross-cycle dependence of cylinder pressure fluctuations, torque decay, and exhaust gas composition.

[0052] In the multi-head self-attention mechanism, the number of self-attention heads is fixed, and the query vector... (Query), key vector (Key) and value vector (Value) is generated through a linear transformation.

[0053] Multi-head self-attention mechanism The expression is: ; In the formula, This represents the independent computational process of each head in a multi-head self-attention mechanism. express The feature dimension is used to scale the dot product result. Represents the key vector (Key) The dimension of the vector. express and Matrix multiplication is used to calculate similarity scores. Represents the normalized exponential function, This represents the matrix transpose operation.

[0054] The feature vector output by the multi-head self-attention layer retains the timestamp alignment property through position encoding and residual connections, generating a temporal feature vector. The temporal feature vector has a fixed dimension and represents the dynamic temporal pattern of the sensor data.

[0055] It should be noted that in step S3.2, after the dynamic performance sensor data is input into the independent Transformer encoder, the temporal dependencies are modeled using a multi-head self-attention mechanism. At this point, the Query vector, Key vector, and Value vector are all generated from the same input data (sensor temporal data) through three independent linear transformation layers, used to calculate the self-attention weights. These three linear transformation layers have different weights, projecting the input data into the Query space, Key space, and Value space respectively, thus representing the same temporal data in different feature spaces.

[0056] S3.3: Use the temporal feature vector as the query vector, and generate key and value vectors from the local spatial features extracted by the CNN through independent linear layers.

[0057] The carbon deposit feature vector is input into two independent linear layers (fully connected layers). The first linear layer generates the key vector, and the second linear layer generates the value vector. The weights of the linear layers are initialized independently, without using an activation function.

[0058] The time-series feature vector is used as the query vector, requiring no additional processing and maintaining its original dimensions.

[0059] S3.4: Calculate the similarity weights of the Query vector and Key vector through a cross-attention layer, and construct a spatiotemporal fusion feature map by weighted summation of the Value vector based on the similarity weights.

[0060] In the cross-attention layer, the dot product similarity between the Query vector (the temporal feature vector in S3.3) and the Key vector (the Key vector generated in S3.3) is calculated, and the softmax function is applied to generate the similarity weight distribution.

[0061] The Value vector (generated in S3.3) is weighted and summed based on similarity weights to fuse spatiotemporal information. The weight distribution adjusts the contribution of the Value vector to generate fused features.

[0062] The fused features are expanded into a two-dimensional feature matrix along the channel-time dimension to construct a spatiotemporal fusion feature map. The row vectors of the two-dimensional feature matrix encode the spatial morphological distribution of carbon deposits, while the column vectors map the temporal patterns of cylinder pressure fluctuations and torque decay, providing a joint feature representation for fault diagnosis.

[0063] It should be noted that in steps S3.3-S3.4, the carbon deposition feature vector (derived from CNN spatial features) is processed through two independent linear layers to generate Key and Value vectors, respectively, while the Query vector directly uses the temporal feature vector output by the Transformer encoder. Here, the Key and Value vectors differ in origin from the Query vector: the Key and Value vectors originate from the carbon deposition feature vector, while the Query vector originates from the temporal feature vector. Cross-attention calculates the similarity between the Query (temporal) and Key (spatial) vectors, and then weightedly fuses the Value (spatial) vector to achieve spatiotemporal feature alignment.

[0064] S4: Construct a reinforcement learning agent, using a spatiotemporal fusion feature map as state input. Deploy a multi-objective dynamic reward function to simultaneously calculate the power loss rate reward, carbon deposit coverage penalty, combustion efficiency gradient, and carbon deposit growth rate gradient. Output the power attenuation diagnostic state through a value-iterative optimization strategy network. (See details below.) Figure 3 .

[0065] S4.1: Initialize the policy network parameters and set the greedy exploration coefficients, using the spatiotemporal fusion feature map as the environmental state input to the policy network.

[0066] Specifically, a fully connected neural network is created as the policy network. The input layer dimension matches the spatiotemporal fusion feature map, and the output layer dimension corresponds to the cleaned action space. The weights are randomly initialized, and the bias is set to zero. The greedy exploration coefficient ε is set to a fixed value to control the action exploration probability (ε is the probability of a random action, and 1-ε is the probability of the network output action). It should be noted that setting the greedy exploration coefficient ε is a key hyperparameter in the training process of a reinforcement learning agent, serving to balance the trade-off between exploration (trying new actions) and exploitation (relying on existing experience). The greedy exploration coefficient ε is specifically set using a linear annealing strategy. For example, initially setting ε = 0.5 (high exploration rate), and then linearly decreasing it to ε = 0.1 (low exploration rate) as the training rounds increase, thus balancing early exploration with later exploitation.

[0067] It should also be noted that the cleaning action space refers to a set of discrete-continuous mixed actions consisting of two dimensions: cleaning intensity level (P1-P6 level governance instructions) and duration (minutes).

[0068] Specifically, the cleaning intensity levels (P1-P6 treatment instructions) are designed to provide progressive treatment solutions based on the severity of carbon buildup and vehicle symptoms. Level P1 addresses minor carbon buildup accompanied by slightly higher fuel consumption or reduced power, employing basic cleaning measures. Level P2 is suitable for moderate carbon buildup causing noticeable noise or decreased combustion efficiency, requiring enhanced cleaning. Level P3 addresses severe carbon buildup causing abnormal exhaust emissions or increased oil consumption, implementing deep cleaning. Level P4 treats stubborn localized deposits accompanied by significant power loss, employing targeted intensive cleaning. Level P5 addresses complex carbon buildup combined with high temperatures or mechanical wear risks, requiring multi-stage combined treatment. Level P6 addresses extreme carbon buildup leading to system critical failure, employing high-intensity comprehensive treatment in conjunction with component maintenance.

[0069] The spatiotemporal fusion feature map is processed into a one-dimensional vector and used as the input to the policy network for the environment state.

[0070] S4.2: The strategy network output includes the cleaning intensity level (discrete values ​​of P1-P6) and duration of the action command. After the action command is executed, new power performance sensor data is obtained through the vehicle bus.

[0071] Specifically, the strategy network outputs the cleaning intensity level and duration (continuous value such as minutes), which are combined into an action command. The action command is sent to the vehicle control unit via the vehicle bus (CAN bus) to execute the carbon deposit cleaning operation.

[0072] After the carbon deposit cleaning is completed, real-time waveforms from the new cylinder pressure sensor, timing data from the crankshaft torque sensor, O2, CO2 and NOx concentration change curves, and carbon deposit image sequences are acquired.

[0073] S4.3: Generate the next moment's environmental state based on the new dynamic performance sensor data, and update the policy network parameters by calculating the total reward value through a multi-objective dynamic reward function.

[0074] The specific steps for constructing the multi-objective dynamic reward function are as follows: The power loss rate bonus value is calculated based on the difference between the real-time waveform of the cylinder pressure sensor and the standard waveform.

[0075] Specifically, the real-time waveforms of the new and old cylinder pressure sensors are compared, and the mean square error (MSE) between the standard waveform (new vehicle state) and the real-time waveform is calculated. The difference is normalized into a reward value, and the smaller the difference, the higher the reward.

[0076] The carbon deposition coverage penalty value is obtained by pixel-level segmentation of the carbon deposition image sequence.

[0077] Specifically, the new carbon deposit image sequence is segmented at the pixel level, and the carbon deposit pixel region is identified using the threshold segmentation method. The proportion of carbon deposit pixels is calculated as the coverage rate, and the higher the coverage rate, the greater the penalty value.

[0078] Combustion efficiency gradient values ​​are generated based on the rate of change of the sliding window ratio of the O2 and CO2 concentration change curves.

[0079] Specifically, based on the new O2 concentration change curve and CO2 concentration change curve, the O2 to CO2 ratio is calculated using a fixed-size sliding window, and the combustion efficiency gradient value is generated by the rate of change of the ratio between adjacent windows. When the rate of change increases positively, the reward value is increased.

[0080] The difference in carbon deposition coverage area between adjacent time windows is calculated using a sliding window difference algorithm to generate a gradient value for carbon deposition growth rate.

[0081] Specifically, the sliding window difference algorithm is used to compare new and old carbon deposition image sequences, calculate the difference in carbon deposition coverage area between adjacent time windows, and normalize the difference in carbon deposition coverage area to generate gradient values. The lower the growth rate, the higher the reward corresponding to the gradient value.

[0082] The total reward value is calculated by weighting and integrating the power loss rate reward value, carbon deposit coverage rate penalty value, combustion efficiency gradient value, and carbon deposit growth rate gradient value.

[0083] It should be noted that in the process of weighted fusion of the total reward value, the four types of indicators—power loss rate reward value, carbon deposit coverage penalty value, combustion efficiency gradient value, and carbon deposit growth rate gradient value—are comprehensively calculated according to preset weight ratios. The power loss rate reward value reflects the degree of cylinder pressure waveform recovery; a higher value indicates a more significant improvement in power performance. The carbon deposit coverage penalty value quantifies the deposit removal effect based on image segmentation results; a lower coverage rate results in a smaller penalty value. The combustion efficiency gradient value assesses the combustion optimization trend through the rate of change in exhaust gas components; a positive change increases the contribution. The carbon deposit growth rate gradient value measures the deposit inhibition effect; a decrease in the growth rate increases the gradient value. The four types of indicators are weighted to reflect the priority of treatment, and finally, a weighted sum is generated to produce the total reward value. The total reward value directly reflects the comprehensive effectiveness of the current cleaning action and is used to guide the updating of strategy network parameters.

[0084] The preset weight ratio refers to the contribution allocation of each indicator to the power attenuation diagnostic state determined by training with historical vehicle maintenance case data based on the Pareto front analysis method in multi-objective optimization theory. The historical vehicle maintenance case data is a structured dataset collected from actual vehicle maintenance records, specifically including recorded engine operating parameters (such as collected carbon deposit image sequences and cylinder pressure waveforms), maintenance measures (such as P1-P6 level treatment instructions), and post-maintenance performance improvement indicators (such as changes in power loss rate).

[0085] S4.4: The action value function value of the current state is calculated using the Bellman equation, and the action value function value is updated through temporal difference learning and stored in the experience playback buffer.

[0086] Specifically, regarding the current state and actions The action value function is calculated using the following expression: ; In the formula, Representing state Next action Action value function value, Indicates the reward value. This represents the discount factor (a fixed value). Indicates the next state Next, for all possible next actions The maximum value of the action.

[0087] It should be noted that during the training process of a reinforcement learning agent, the current state... This comprehensive carbon deposit status of a vehicle is derived from the fusion of carbon deposit image features and power performance data. Specifically, it includes the visual morphological characteristics of carbon deposits inside the engine, as well as dynamic parameters such as real-time cylinder pressure fluctuations, torque changes, and exhaust gas composition. After processing by a neural network, a high-dimensional feature vector is formed. This high-dimensional feature vector fully depicts the spatial distribution characteristics and temporal evolution trend of the vehicle's carbon deposit problem at the current moment.

[0088] action Specifically, this involves cleaning operations of different intensity levels and their durations. The action space includes six discrete cleaning intensity levels and continuous duration parameters, which together constitute a complete governance scheme. When the agent selects a specific action, the vehicle control system executes the corresponding cleaning program, and then collects new operational data through the sensor network. This operational data is processed to form the state for the next moment. This is used to evaluate the effectiveness of actions and update strategies. The immediate reward value generated during state transitions comprehensively considers multiple target indicators such as power performance recovery and carbon deposit removal effectiveness.

[0089] Furthermore, the timing difference error is calculated using the following expression: ; In the formula, This represents the time-series difference error, reflecting the deviation between the current action value estimate and the previous estimate. This represents the old action value function estimate, stored in the experience replay buffer.

[0090] The action value function is updated using the following expression: ; In the formula, The learning rate is fixed at 0.001. quadruple Stored in the experience replay buffer (fixed capacity queue).

[0091] S4.5: Sample action value from the experience replay buffer to calculate the advantage function value, and generate policy gradient correction based on the linear combination of the advantage function value and the action value value.

[0092] Specifically, batches of data are randomly drawn from the experience replay buffer, with each batch containing... Quadruple ; This indicates the sampling batch size, which is the number of samples randomly selected from the experience playback buffer.

[0093] The advantage function is calculated as follows: ; ; In the formula, The dominance function value is used to measure the action. Relative to state The degree of advantage in average value The state value function value represents the state. The average expected cumulative reward, Indicates all possible actions Summation operation, Indicates policy network parameters In state The next action after the output The probability, Indicates the state Execute the next action The expected cumulative reward afterward.

[0094] It should be noted that expected cumulative reward is the target for calculating maximum action value. The two are essentially the same, but they have different emphases (expected cumulative reward emphasizes the result of the benefit, while maximum action value emphasizes the decision-making process).

[0095] Based on a linear combination of the advantage function value and the action value function value, the policy gradient correction is generated, expressed as: ; In the formula, The policy gradient value represents the policy network parameters. The policy gradient adjustment amount for the expected cumulative reward. This represents the normalization coefficient for the sampling batch. This represents the logarithmic gradient of the probability of the policy network output action. This represents the difference between the current action value function and the state value function.

[0096] S4.6: Optimize the policy network parameters using the gradient ascent algorithm based on the policy gradient correction amount, repeat the iteration until the state value function value converges, and mark the output action command probability distribution as the dynamic decay diagnostic state.

[0097] Specifically, the policy network weights are updated along the gradient direction based on the policy gradient correction amount, and a fixed learning rate is used in the update process to avoid parameter oscillations.

[0098] The system monitors the changes in the probability distribution of actions output by the policy network in real time. If the change in the probability of each action (P1-P6 level governance instructions) is less than the preset action probability convergence threshold in multiple consecutive iterations, it is considered stable.

[0099] It should be noted that the convergence threshold of action probabilities is determined by the rate of change of the KL divergence of the probability distribution of the policy network output layer, and the theoretical boundary is determined by the convergence condition of the policy improvement theorem.

[0100] The stabilized policy network generates an action probability distribution vector from the input spatiotemporal fusion feature map, and outputs the probability distribution vector directly as the power attenuation diagnostic state according to (P1-P6 level governance instructions).

[0101] S5: The student model receives the power attenuation diagnostic status and carbon deposit type classification codes, loads the pre-trained teacher model, adds an adversarial discriminator to the distillation loss function, compares the feature map distribution differences between the teacher and student models, and outputs a graded treatment instruction matching the vehicle state. For details, please refer to [link to relevant documentation]. Figure 4 .

[0102] The pre-trained teacher model refers to a two-layer Transformer network trained on carbon deposit treatment cases from a standard vehicle repair database. The standard vehicle repair database is a structured collection of engine repair cases, containing a linked dataset of standardized fields such as engine operating parameters (e.g., carbon deposit images and cylinder pressure waveforms), repair measures (e.g., cleaning intensity levels), and post-repair performance indicators (e.g., power recovery rate).

[0103] Specifically, the training process of the two-layer Transformer network for the pre-trained teacher model is as follows: Based on historical carbon deposit treatment case data in the standard vehicle maintenance database, engine operating parameters, maintenance measure records, and performance improvement indicators are standardized to construct a supervised learning dataset containing input feature vectors and corresponding treatment labels; the network architecture adopts two stacked Transformer encoding layers. The first layer processes the embedding representation of raw sensor data and image features, and models cross-modal feature association through a multi-head self-attention mechanism. The second layer refines the attention weights and outputs the latent space representation; during the training phase, a labeled smooth cross-entropy loss function is used, and the Adam optimizer is used for end-to-end optimization with a fixed learning rate. The training is terminated when the validation set loss no longer decreases through an early stopping mechanism, and the network parameters are saved as the teacher model.

[0104] S5.1: The diagnostic status of power attenuation and the type of carbon deposit are classified and encoded and concatenated into a diagnostic fusion feature vector, which is then input into the student model and the student feature vector is output.

[0105] Specifically, the probability distribution vector of the power attenuation diagnostic status and the unique thermal encoding vector of the carbon deposit type classification code are connected end to end along the dimension to form a diagnostic fusion feature vector.

[0106] It should be noted that the unique thermal coding vector refers to the standard basis vector that maps the discrete categories (dry carbon, wet carbon, and annular deposition) in the carbon deposition type classification coding to a three-dimensional unit vector space.

[0107] S5.2: Call the pre-trained teacher model to perform forward computation on the diagnostic fusion feature vector and output the teacher feature vector.

[0108] Specifically, the diagnostic fusion feature vector is input into the first layer of the Transformer encoder of the pre-trained teacher model and processed through a self-attention mechanism.

[0109] The output is refined by the second-layer Transformer encoder to generate teacher feature vectors.

[0110] S5.3: Construct an adversarial discriminator in the feature space of the distillation loss function and calculate the cosine similarity of the feature maps of the teacher feature vector and the student feature vector.

[0111] Specifically, the diagnostic fusion feature vector is input into the student model (fully connected network), and the student feature vector is output. An adversarial discriminator is constructed in the feature space, and the cosine similarity between the teacher feature vector and the student feature vector is calculated. The teacher feature vector and the student feature vector are then normalized respectively. Calculate the dot product of the normalized vectors; the dot product value is the cosine similarity score.

[0112] S5.4: Construct an adversarial loss function based on the cosine similarity of the feature maps and optimize the student model parameters using the backpropagation algorithm.

[0113] Specifically, the cosine similarity score is converted into an adversarial loss value, and the gradient of the adversarial loss with respect to the student model parameters is calculated using the gradient descent algorithm. The student model weights are updated based on the gradient, thus optimizing the similarity between the student feature vector and the teacher feature vector.

[0114] S5.5: The student model output layer generates the probability distribution of governance instructions from level P1 to level P6, and sends the instruction with the highest probability as the hierarchical governance instruction to the vehicle control unit.

[0115] Specifically, the optimized student model output layer performs calculations: the last fully connected layer of the student model outputs a six-dimensional vector, which is then transformed into a probability distribution through a normalized exponential function.

[0116] Select the governance command level corresponding to the highest probability value in the probability distribution, and send the command to the vehicle control unit for execution via the vehicle bus.

[0117] This embodiment also provides a computer device, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the deep learning-based automotive carbon buildup drive fault diagnosis method proposed in the above embodiment.

[0118] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0119] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the deep learning-based automotive carbon buildup drive fault diagnosis method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0120] In summary, this invention achieves a comprehensive dynamic quantitative assessment of carbon deposit-driven faults by constructing a reinforcement learning agent and deploying a multi-objective dynamic reward function. The reinforcement learning agent receives a spatiotemporal fusion feature map, which integrates the spatial morphological characteristics of carbon deposits with the temporal dynamic characteristics of power performance, as the environmental state input. After executing the cleaning action command output by the policy network, the multi-objective dynamic reward function simultaneously calculates four key indicators: a power loss rate reward value based on cylinder pressure waveform differences to generate power recovery parameters; a carbon deposit coverage penalty value based on image segmentation to quantify the deposition removal effect; a combustion efficiency gradient value based on the rate of change of exhaust gas components to generate combustion optimization trend parameters; and a carbon deposit growth rate gradient value based on the difference in coverage area to quantify the deposition suppression effect. The four-dimensional reward values ​​are simultaneously calculated and weighted to construct a real-time quantitative assessment framework that comprehensively reflects the impact of carbon deposit status and power performance.

[0121] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A deep learning-based method for diagnosing automotive carbon buildup-driven faults, characterized in that: include: Real-time acquisition of carbon buildup image sequences and power performance sensor data from vehicles; The carbon deposit image sequence is input into the CNN convolutional layer. The local spatial features of the carbon deposit morphology are extracted through the convolutional and pooling layers, and the carbon deposit type classification code is output through the fully connected layer and the Softmax classifier. The dynamic performance sensor data is input into an independent Transformer encoder to generate a temporal feature vector, which is then fused with the local spatial features extracted by the CNN through a cross-attention mechanism to construct a spatiotemporal fusion feature map. A reinforcement learning agent is constructed, using a spatiotemporal fusion feature map as the state input. A multi-objective dynamic reward function is deployed to simultaneously calculate the power loss rate reward value, carbon deposit coverage penalty value, combustion efficiency gradient value, and carbon deposit growth rate gradient value. The power attenuation diagnostic state is output through a value iterative optimization strategy network. The student model receives the power attenuation diagnostic status and carbon deposit type classification codes, and loads the pre-trained teacher model. An adversarial discriminator is added to the distillation loss function to compare the feature map distribution differences between the teacher model and the student model, and outputs a graded treatment instruction that matches the vehicle status.

2. The deep learning-based automotive carbon buildup drive fault diagnosis method as described in claim 1, characterized in that: The carbon deposit image sequence includes an engine intake valve image sequence, a fuel injector image sequence, and a combustion chamber surface deposit image sequence; the power performance sensor data includes real-time waveforms from the cylinder pressure sensor, timing data from the crankshaft torque sensor, and O2, CO2, and NOx concentration change curves.

3. The deep learning-based automotive carbon buildup drive fault diagnosis method as described in claim 2, characterized in that: The process involves extracting local spatial features of carbon deposit morphology through convolutional and pooling layers, and then outputting a carbon deposit type classification code via a fully connected layer and a Softmax classifier. The specific steps are as follows: The carbon deposit image sequence is input into a three-dimensional convolutional layer to extract multi-scale spatial features. The multi-scale spatial features are then reduced in dimensionality by a max pooling layer to generate a carbon deposit feature vector. The carbon deposit feature vector is input into the fully connected layer to calculate the nonlinear combination weights; The Softmax classifier is used to output the probability distribution of dry carbon, wet carbon and annular deposition types based on nonlinear combined weights. The type with the highest probability is marked as the carbon deposition type classification code.

4. The deep learning-based automotive carbon buildup drive fault diagnosis method as described in claim 3, characterized in that: The specific steps for constructing the spatiotemporal fusion feature map are as follows: The dynamic performance sensor data is divided into fixed-length time windows and input into the Transformer encoder; The cross-cycle dependence of cylinder pressure fluctuation, torque decay and exhaust gas composition is modeled by a multi-head self-attention mechanism, and a time-series feature vector with timestamp alignment is output. Using temporal feature vectors as query vectors, the local spatial features extracted by CNN are used to generate key and value vectors through independent linear layers; The similarity weights of the Query vector and Key vector are calculated by a cross-attention layer. The Value vector is then weighted and summed based on the similarity weights to generate fusion features, and a spatiotemporal fusion feature map is constructed.

5. The deep learning-based automotive carbon buildup drive fault diagnosis method as described in claim 4, characterized in that: The specific steps for constructing the reinforcement learning agent are as follows: Initialize the policy network parameters and set the greedy exploration coefficients, using the spatiotemporal fusion feature map as the environmental state input to the policy network; The strategy network outputs action commands containing the cleaning intensity level and duration. After executing the action commands, new dynamic performance sensor data is acquired via the vehicle bus. The next-moment environmental state is generated based on the new dynamic performance sensor data, and the policy network parameters are updated by calculating the total reward value through a multi-objective dynamic reward function.

6. The deep learning-based automotive carbon buildup drive fault diagnosis method as described in claim 5, characterized in that: The construction of the multi-objective dynamic reward function is carried out through the following steps: The power loss rate bonus value is calculated based on the difference between the real-time waveform of the cylinder pressure sensor and the standard waveform. Carbon deposit coverage penalty values ​​are obtained by pixel-level segmentation of carbon deposit image sequences; Combustion efficiency gradient values ​​are generated based on the rate of change of the sliding window ratio of the O2 and CO2 concentration change curves. The difference in carbon deposition coverage area between adjacent time windows is calculated using the sliding window difference algorithm to generate the gradient value of carbon deposition growth rate. The total reward value is calculated by weighting and integrating the power loss rate reward value, carbon deposit coverage rate penalty value, combustion efficiency gradient value, and carbon deposit growth rate gradient value.

7. The deep learning-based automotive carbon buildup drive fault diagnosis method as described in claim 6, characterized in that: The specific steps for outputting the dynamic decay diagnostic state through the value iteration optimization strategy network are as follows: The action value function of the current state is calculated using the Bellman equation, and the action value function is updated through temporal difference learning and stored in the experience replay buffer. The advantage function value is calculated by sampling action value values ​​from the experience replay buffer, and the policy gradient correction is generated based on the linear combination of the advantage function value and the action value value. The gradient ascent algorithm is used to optimize the policy network parameters based on the policy gradient correction amount. This process is repeated iteratively until the state value function converges. The output action command probability distribution is then labeled as the dynamic decay diagnostic state.

8. The deep learning-based automotive carbon buildup drive fault diagnosis method as described in claim 7, characterized in that: The pre-trained teacher model refers to a two-layer Transformer network trained based on carbon deposit treatment cases in a standard vehicle maintenance database. The specific steps for outputting hierarchical management instructions that match the vehicle status are as follows: The diagnostic status of power attenuation and the type of carbon deposit are classified, encoded, and concatenated into a diagnostic fusion feature vector, which is then input into the student model to output the student feature vector. The pre-trained teacher model is invoked to perform forward computation on the diagnostic fusion feature vector, outputting the teacher feature vector. An adversarial discriminant is constructed in the feature space of the distillation loss function, and the cosine similarity of the feature maps of the teacher feature vector and the student feature vector is calculated. An adversarial loss function is constructed based on the cosine similarity of feature maps, and the parameters of the student model are optimized using the backpropagation algorithm. The student model output layer generates the probability distribution of governance instructions from level P1 to level P6, and sends the instruction with the highest probability as the hierarchical governance instruction to the vehicle control unit.

Images