Automobile bus test AI diagnosis system and closed loop feedback and self-learning method thereof

Abstract

The application relates to the case management and artificial intelligence application field in automobile testing, and provides an automobile bus test AI diagnosis system and a closed-loop feedback and self-learning method thereof.The method comprises the following steps: S1, a system generates a fault diagnosis suggestion; S2, maintenance feedback data is collected; S3, causal characteristics are extracted, and two causal characteristic vectors are formed; S4, differential coding of the causal characteristics is generated; S5, system model incremental training and fine-tuning are carried out; and S6, the system model after fine-tuning is used for fault diagnosis.The system comprises a fault diagnosis module, a maintenance feedback module, a causal characteristic extraction module, a differential coding module and an incremental training module.The application realizes the improvement of diagnosis accuracy with the increase of use times through the closed-loop feedback and self-learning mechanism; according to evaluation, the diagnosis accuracy can be improved by 35% within 3 months; and the continuous evolution capability enables the system to continuously adapt to the specific test environment and fault modes of enterprises, and to provide more and more accurate diagnosis suggestions.

Description

Technical Field

[0001] This invention relates to the fields of case management and artificial intelligence applications in automotive testing, and particularly to an AI diagnostic system for automotive bus testing and its closed-loop feedback and self-learning method. Background Technology

[0002] Existing AI-assisted diagnostic solutions for automotive testing typically employ a "one-off" interaction model, meaning the interaction ends once the system provides a suggestion. The model cannot know whether its suggestion has been adopted. This one-way interaction model prevents the AI ​​system from self-optimizing based on real-world application scenarios, keeping its diagnostic capabilities at an initial level and unable to adapt to specific enterprise testing environments and fault modes.

[0003] Existing AI diagnostic systems lack effective feedback channels to collect engineers' evaluations and corrections of diagnostic suggestions. This prevents models from iterating and optimizing based on real-world application scenarios. Even if engineers discover errors in the model's suggestions, there is no convenient way to feed this valuable experience back to the system, causing the same errors to recur in future diagnoses, thus reducing engineers' trust in and willingness to use the AI ​​system.

[0004] Existing AI diagnostic systems lack a complete closed loop from "diagnostic suggestions" to "engineer feedback" and then to "model optimization," thus failing to enable self-learning and continuous evolution. This closed loop needs to address three key aspects: (1) how to design an intuitive and convenient user feedback interface to lower the barrier for engineers to provide feedback; (2) how to transform unstructured user feedback into structured training data; and (3) how to utilize this feedback data for efficient incremental training of the local model without affecting the normal operation of the system.

[0005] Existing AI diagnostic systems have not addressed the technical challenges of model fine-tuning, including how to perform efficient incremental training with limited local computing resources, how to avoid catastrophic forgetting (i.e., forgetting old knowledge when learning new knowledge), and how to evaluate and verify the effectiveness of model optimization to ensure that each iteration brings a substantial improvement in diagnostic accuracy. Summary of the Invention

[0006] The purpose of this invention is to overcome at least one of the shortcomings of the prior art and provide an AI diagnostic system for automotive bus testing and its closed-loop feedback and self-learning method, so as to realize the self-learning and continuous evolution of the diagnostic system, enabling the system to continuously improve its diagnostic accuracy with the increase of usage, and ultimately forming a "diagnostic brain" that is exclusive to the enterprise and becomes smarter with use.

[0007] The present invention adopts the following technical solution: On the one hand, this invention provides a closed-loop feedback and self-learning method for an automotive bus testing AI diagnostic system, including: S1. System generates fault diagnosis suggestions: For a certain test fault, the AI ​​diagnostic system provides fault diagnosis suggestions, which are presented in structured data. S2. Collect maintenance feedback data: For the test fault in step S1, after the processing is completed, collect maintenance feedback data, which is also presented as structured data. S3. Extract causal features: When a fault diagnosis suggestion is incorrect, define this fault diagnosis as a negative sample; extract the causal features of the negative sample, and at the same time extract the causal features of the corresponding maintenance feedback data to form two causal feature vectors; S4. Calculate the difference between the two causal feature vectors obtained in step S3, and generate causal feature differential encoding; S5. Incremental training and fine-tuning of the system model: When the differential encoding of causal features obtained in step S4 accumulates to a set quantity, the model training condition is triggered; the differential encoding of causal features is used as training samples to perform incremental training and fine-tuning of the system model. S6. Use the fine-tuned system model for fault diagnosis.

[0008] In addition to any of the possible implementations described above, another implementation is provided in which the process of extracting causal features in step S3 is as follows: S31. Obtain all raw data of bus messages within a period of time before and after the diagnostic time point, the version of the currently loaded DBC (Data Base CAN) file, the firmware version information of the relevant ECU, and the temperature and voltage of the test environment; S32. Extract preset causal features from the data obtained in step S31. The causal features include global features, multi-scale local features, and signal correlation features. The global features are used to capture the macroscopic state of the entire bus system. The multi-scale local features are features related to specific nodes or signals. The signal correlation features are used to determine whether two or more signals exhibit abnormal synchronous changes when a fault occurs. S33. Extract maintenance-related features from the negative samples and the maintenance feedback data respectively. The maintenance-related features are used to reflect the characteristics of interaction with the system during the maintenance process. S34. Combine the causal features preset in step S32 with the maintenance-related features in step S33 to form a high-dimensional causal feature vector; the negative sample and the maintenance feedback data each correspond to one of the causal feature vectors.

[0009] In addition to any of the possible implementations described above, another implementation is provided in which, in step S32, the global features include: average bus load rate, peak load rate, frequency of error frames, and bus utilization; the multi-scale local features include: periodicity of specific ID messages, calculation of the standard deviation and jitter of their transmission interval to determine whether there is a timing anomaly; statistical features and rate of change of the signal, and detection of abnormal jumps or drifts, wherein the statistical features include mean, variance, kurtosis, and skewness; the signal correlation features include: determining whether two seemingly independent signals exhibit abnormal synchronous changes when a fault occurs.

[0010] In addition to any of the possible implementations described above, another implementation is provided in which, in step S33, the maintenance-related features include maintenance response signal features, fault code and maintenance action association features, and maintenance text features.

[0011] In addition to any of the possible implementations described above, another implementation is provided, wherein the specific steps of step S4 are as follows: S41. Calculate the original difference vector: the causal feature vector of the negative sample is V_error, the causal feature vector of the maintenance feedback data is V_correct, and the original difference vector V_diff_raw = V_correct - V_error; the negative sample represents the incorrect diagnosis, and the maintenance feedback data represents the correct diagnosis. S42. Normalize the original difference vector V_diff_raw; S43. Use a feature selection algorithm to filter out the key feature dimensions that distinguish between correct and incorrect diagnoses, and obtain a sparsed difference vector V_diff_sparse. S44. Associate the sparsed difference vector V_diff_sparse with the error type label to generate the causal feature differential encoding; the causal feature differential encoding is a structured data unit containing the sparsed difference vector V_diff_sparse, the error type label, and the original context information.

[0012] In addition to any of the possible implementations described above, another implementation is provided, in step S5, The differential encoding of the causal features, the corresponding fault diagnosis suggestions, and the corresponding maintenance feedback data are combined into an incremental data training set. The system model training uses a composite loss function, which is expressed as: L_total = L_ce + λ * L_correct; Where L_ce is the cross-entropy loss, used to optimize the model's classification accuracy on positive samples; L_correct is the error correction loss, used to encourage the model to learn the feature differences indicated by the sparsed difference vector V_diff_sparse; λ is a hyperparameter used to balance the weights of the two loss terms. Taking a given fault scenario as the anchor point, fault diagnosis suggestions as negative samples, and maintenance feedback data as positive samples, the training objective of the system model is to minimize the distance between the anchor point and the positive samples in the feature space, while maximizing the distance between the anchor point and the negative samples in the feature space. During the system model training process, the system model parameters are continuously fine-tuned to obtain the optimized system model parameters.

[0013] In addition to any of the possible implementations described above, a further implementation is provided in which the system model has a dynamically expanding model structure. When it is found that the system model's diagnostic accuracy for a newly emerging type of fault mode is consistently low, or the feature complexity of the new data exceeds the expressive power of the current system model, a structural expansion instruction for the system model is triggered. The system dynamically adds new "enhanced node groups" (which can be one or more neurons, or a more complex sub-network module, such as an attention head or a feedforward network layer) to the network. These new nodes are randomly initialized and connected to existing feature nodes and output nodes. Then, the system model freezes the parameters of the original network and trains only the newly added nodes, enabling them to learn how to fit residual information that the current system model cannot handle.

[0014] In addition to any of the possible implementations described above, another implementation is provided in which the conditions for triggering training in step S5 include: Quantity-driven: When the number of newly collected negative samples reaches a set threshold; or, Performance-driven: When the diagnostic accuracy of the system model for a specific type of fault continuously decreases beyond a set threshold; or, Time-driven: Set a fixed time period for regular system model training and updates.

[0015] In addition to any of the possible implementations described above, a further implementation is provided in which a new version is formed each time the system model is trained and updated; A / B testing is automatically performed to compare the new version with the old version on a validation set; the new version is automatically deployed as the actual use version only when it outperforms the old version in all key performance indicators; if the new version performs poorly, the system model will automatically revert to the old version and generate a detailed analysis report; the key performance indicators include diagnostic accuracy, performance retention rate, inference latency, and resource consumption.

[0016] On the other hand, the present invention also provides an AI diagnostic system for automotive bus testing, the system being used to implement the aforementioned closed-loop feedback and self-learning method, the system comprising: The fault diagnosis module is used to provide fault diagnosis suggestions based on specific fault scenarios; The maintenance feedback module is used to collect maintenance feedback data after the fault repair is completed. The maintenance feedback data is the correct fault diagnosis. The causal feature extraction module is used to extract the causal features from the negative sample when the fault diagnosis suggestion is incorrect, and at the same time extract the causal features from the corresponding maintenance feedback data to form two causal feature vectors. A differential encoding module is used to calculate the difference between the two causal feature vectors and generate causal feature differential encoding. The incremental training module uses the causal feature differential encoding as training samples to perform incremental training and fine-tuning on the system model.

[0017] In addition to any of the possible implementations described above, another implementation is provided in which the system further includes: The training trigger module is used to set training trigger conditions. When the trigger conditions are met, the system model training is triggered. The dynamic expansion module adds new "enhanced node groups" to the network and connects them to existing nodes when the system model's diagnostic accuracy for a newly emerging fault mode remains low, or when the feature complexity of newly collected data exceeds the expressive power of the current system model. The A / B testing module automatically performs A / B testing when the system model training is updated to form a new version, comparing the new version with the old version on the validation set. Only when the new version outperforms the old version in all key performance indicators will it be automatically deployed as the new version. If the new version performs poorly, the system model will automatically revert to the old version.

[0018] The beneficial effects of this invention are as follows: 1. Through closed-loop feedback and self-learning mechanisms, this invention enables diagnostic accuracy to improve with repeated use. Preliminary assessments indicate that diagnostic accuracy can improve by 35% within three months. As the system accumulates more feedback data, the rate of improvement in diagnostic accuracy will gradually slow, but it will continue to converge towards higher levels. This ability to continuously evolve allows the system to adapt to specific enterprise testing environments and failure modes, providing increasingly accurate diagnostic recommendations.

[0019] 2. This invention enables the creation of a company-specific diagnostic model that becomes increasingly accurate with use, transforming the collective wisdom of engineers into the company's knowledge asset. This proprietary model not only includes general automotive bus fault knowledge but also incorporates the company's specific vehicle platform, ECU supplier, and testing procedures, making its diagnostic recommendations more tailored to the company's actual needs. Over time, this model will become a valuable intangible asset for the company, difficult for competitors to replicate.

[0020] 3. The local fine-tuning mechanism of this invention does not rely on cloud training; all data and models are processed locally, effectively protecting the enterprise's data sovereignty. Sensitive enterprise data (such as DBC files, ECU firmware, test logs, etc.) never leaves the local environment, completely eliminating the risk of data leakage. Simultaneously, local fine-tuning avoids dependence on cloud services, ensuring the system can still operate and evolve normally even when the network is down or cloud services are unavailable. Attached Figure Description

[0021] Figure 1 The diagram shown is a schematic flowchart of a closed-loop feedback and self-learning method for an automotive bus testing AI diagnostic system according to an embodiment of the present invention.

[0022] Figure 2 The diagram shown is a closed-loop feedback flowchart in the embodiment.

[0023] Figure 3 The diagram shown is a fine-tuning architecture diagram of the system model in the embodiment. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0025] The accompanying drawings illustrate a layer structure according to an embodiment of the present invention. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0026] Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0027] In the description of this invention, it should be noted that the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0028] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0029] like Figure 1 As shown in the figure, an embodiment of the present invention provides a closed-loop feedback and self-learning method for an automotive bus testing AI diagnostic system, comprising: S1. System generates fault diagnosis suggestions: For a certain test fault, the AI ​​diagnostic system provides fault diagnosis suggestions, which are presented in structured data. S2. Collect maintenance feedback data: For the test fault in step S1, after the processing is completed, collect maintenance feedback data, which is also presented as structured data. S3. Extract causal features: When a fault diagnosis suggestion is incorrect, define this fault diagnosis as a negative sample; extract the causal features of the negative sample, and at the same time extract the causal features of the corresponding maintenance feedback data to form two causal feature vectors; S4. Calculate the difference between the two causal feature vectors obtained in step S3, and generate causal feature differential encoding; S5. Incremental training and fine-tuning of the system model: When the differential encoding of causal features obtained in step S4 accumulates to a set quantity, the model training condition is triggered; the differential encoding of causal features is used as training samples to perform incremental training and fine-tuning of the system model. S6. Use the fine-tuned system model for fault diagnosis.

[0030] The closed-loop feedback process of the automotive bus test AI diagnostic system of this invention is as follows: Figure 2 As shown.

[0031] In one specific embodiment, the collection of maintenance feedback data in step S2 provides "Accept," "Disaccept," and "Correct" buttons on the diagnostic suggestion interface such as ET Studio, facilitating engineers to provide feedback. Specifically, after the system generates diagnostic suggestions, the interface displays a concise feedback panel containing the following options: Adoption: Once the engineer confirms that the suggestion is correct and useful, the system will record this successful diagnostic case.

[0032] Ignore: The engineer deems the suggestion irrelevant or useless, but provides no further information.

[0033] Correction: The engineer believes the suggestion is incorrect and can enter the correct cause of the failure and solution.

[0034] To reduce the cognitive burden on engineers providing feedback, the system also offers intelligent assistance features. For example, when an engineer selects "Fix," the system automatically pops up a form pre-filled with possible fault categories (such as hardware failure, software bug, configuration error, etc.) and provides a drop-down menu to select common solution templates. Engineers only need to make minor modifications to complete the feedback.

[0035] Feedback vectorization: Engineer feedback (such as "Incorrect suggestion, actual cause is power interference") is converted into structured labels to facilitate model learning. Specifically, the system employs the following strategy: 1. Positive Sample Processing: For "adoption" feedback, the system will match the complete context information (including bus messages, DBC version, ECU firmware, etc.) with the diagnostic suggestions to form a high-quality positive sample.

[0036] 2. Negative Sample Handling: For feedback indicating "correction," the system marks the original incorrect suggestion as a negative sample, while treating the correct solution provided by the engineer as a new positive sample. Furthermore, the system analyzes the semantic differences between the incorrect suggestion and the correct solution, generating a comparative learning signal.

[0037] 3. Feature Engineering: The system extracts key features from the feedback, such as fault type, affected ECU, related signals, and solution complexity, and encodes these features into vectors for subsequent model training.

[0038] Through this structured feedback processing mechanism, the system can make engineers' tacit knowledge explicit and transform it into a format that the model can understand and learn.

[0039] To deeply mine and utilize the valuable knowledge contained in engineer feedback, especially the corrective information for model erroneous suggestions, this invention proposes an innovative negative sample processing mechanism. The core idea of ​​this mechanism is to go beyond simple "error-correct" label comparisons, delving into the deep logical relationship between erroneous suggestions and correct causes, and quantifying this into feature differences that the model can learn. This mechanism integrates cutting-edge technologies such as causal inference, negative sample reinforcement learning (NSR), and contrastive learning, aiming to enable the model not only to "know what," but also to "know why," thereby achieving precise and efficient evolution of diagnostic capabilities. The mechanism mainly comprises three core steps: First, through a causal feature extraction module, the key causal features behind the original erroneous suggestions and the correct causes provided by engineers are analyzed; second, a difference encoder is used to calculate the difference between the two sets of causal features, generating a high-dimensional "causal difference vector"; finally, this difference vector is used as a strong supervisory signal, and the model is fine-tuned through a combination of negative sample reinforcement learning (NSR) and contrastive learning, enabling it to proactively avoid reasoning paths that lead to errors in future diagnoses and strengthen the logical chain leading to correct conclusions.

[0040] One of the core innovations of this invention lies in the introduction of causal feature extraction and differential encoding, aiming to fundamentally understand and quantify the logical gap between incorrect model suggestions and correct engineer corrections. Traditional negative sample processing methods often remain superficial, such as simply marking incorrect answers as negative samples, without delving into the root cause of the error. This mechanism, however, draws on causal inference theory, aiming to identify the true cause leading to a specific diagnostic result, rather than merely correlation. In specific implementation, the system first constructs a structural causal model (SCM) for automotive bus fault diagnosis. This model explicitly defines the causal relationship between fault phenomena (such as abnormal messages), potential causes (such as ECU software defects, signal interference, power fluctuations), and diagnostic conclusions. When the engineer corrects the model's diagnostic suggestions, the system activates the causal feature extractor. This extractor takes the original incorrect suggestion, the correct cause provided by the engineer, and relevant bus data (such as message sequences, signal values, and timestamps) as input, and performs inference through the SCM to generate two sets of high-dimensional causal feature vectors. One set of vectors represents the features of the "pseudo-causal path" leading to incorrect suggestions, while the other set represents the features of the "true causal path" leading to the correct cause. The difference encoder then compares these two sets of feature vectors, calculating their differences across various dimensions, ultimately generating a compact and highly informative "causal difference encoding" (CDE). This CDE vector accurately captures the logical direction of correction from error to correctness, providing invaluable guidance for subsequent model training.

[0041] In one specific embodiment, the process of extracting causal features in step S3 is as follows: S31. Obtain all raw data of bus messages within a period of time before and after the diagnostic time point, the version of the currently loaded DBC (Data Base CAN) file, the firmware version information of the relevant ECU, and the temperature and voltage of the test environment; S32. Extract preset causal features from the data obtained in step S31. The causal features include global features, multi-scale local features, and signal correlation features. The global features are used to capture the macroscopic state of the entire bus system. The multi-scale local features are features related to specific nodes or signals. The signal correlation features are used to determine whether two or more signals exhibit abnormal synchronous changes when a fault occurs. S33. Extract maintenance-related features from the negative samples and the maintenance feedback data respectively. The maintenance-related features are used to reflect the characteristics of interaction with the system during the maintenance process. S34. Combine the causal features preset in step S32 with the maintenance-related features in step S33 to form a high-dimensional causal feature vector; the negative sample and the maintenance feedback data each correspond to one of the causal feature vectors.

[0042] In one specific embodiment, in step S32, the global features include: average bus load rate, peak load rate, frequency of error frames, and bus utilization; the multi-scale local features include: periodicity of specific ID messages, calculation of the standard deviation and jitter of their transmission interval to determine whether there is a timing anomaly; statistical features and rate of change of the signal, and detection of abnormal jumps or drifts, wherein the statistical features include mean, variance, kurtosis, and skewness; the signal correlation features include: determining whether two seemingly independent signals exhibit abnormal synchronous changes when a fault occurs.

[0043] In one specific embodiment, in step S33, the maintenance-related features include maintenance response signal features, fault code and maintenance action association features, and maintenance text features.

[0044] Maintenance response signal characteristics: This mainly focuses on the system's feedback when receiving maintenance instructions or detecting maintenance actions. For example, after a technician uses diagnostic tools to clear fault codes, does the system immediately regenerate the same fault code, or generate a new, related fault code? This information contains important diagnostic clues. Furthermore, the response behavior of certain ECUs when receiving specific diagnostic service requests (such as 2F service, input / output control) (e.g., whether the request was successfully executed, response time, etc.) can also be used as characteristics. These characteristics reflect the system's "reaction" to maintenance operations, helping the model determine whether the fault is temporary, intermittent, or persistent, thus providing a deeper basis for diagnosis.

[0045] Fault Code and Repair Action Association Features: This feature aims to establish the correlation between fault codes (DTCs) and specific repair actions. By analyzing a large number of historical repair records, it's possible to statistically determine the most common repair actions (e.g., replacing a sensor, flashing the software of an ECU, checking a circuit) that occur after a specific fault code. This correlation can be quantified as a feature, where each dimension represents a repair action, and its value indicates the strength of the association between that action and the current fault code. When the system generates a diagnostic suggestion, it simultaneously predicts the most likely repair action. If the engineer's feedback (positive samples) does not match the system's prediction, this discrepancy can be used to train the model, enabling it to learn a more accurate fault-repair mapping.

[0046] Repair Text Features: When providing repair feedback, engineers typically input text descriptions such as "Fault resolved after replacing the left front wheel speed sensor" or "Inspection revealed a short circuit to ground in the CAN-H line." This textual information contains rich, structured knowledge. Through Natural Language Processing (NLP) techniques, such as keyword extraction, Named Entity Recognition (NER), and sentiment analysis, this unstructured text can be transformed into structured features. For example, component names mentioned in the text (e.g., "left front wheel speed sensor"), repair actions (e.g., "replacement," "inspection"), and fault symptoms (e.g., "short circuit to ground") can be extracted and encoded into feature vectors. These textual features directly reflect the engineer's diagnostic approach and repair process, serving as powerful signals for distinguishing between positive and negative samples.

[0047] After supplementing with maintenance-related features, the new causal feature vector will be a high-dimensional fusion containing multiple types of information. It includes not only global, local, and correlated features describing the physical phenomena of the fault, but also maintenance-related features describing maintenance behavior and decision-making processes. This fusion allows the feature vector to more comprehensively characterize a fault scenario, thus providing the model with richer input information. During model training, these different types of features will be adaptively fused through the weights of the neural network, and the model will automatically learn which features are more important in distinguishing different fault types.

[0048] By introducing maintenance-related features, the model's discriminative ability is significantly enhanced. When faced with a complex failure scenario, the model no longer relies solely on physical similarity but can incorporate maintenance-related evidence to make judgments. For example, for two failures with similar physical characteristics, if their maintenance histories or engineers' textual feedback differ significantly, the model can distinguish them using these maintenance-related features, labeling one as a negative sample and the other as a positive sample. This allows the model to learn deeper diagnostic logic from engineers' feedback, rather than simply relying on surface-level pattern matching. Ultimately, this leads to a substantial improvement in the model's diagnostic accuracy and better adaptability to enterprise-specific testing environments and failure modes.

[0049] In one specific embodiment, step S4 consists of the following steps: S41. Calculate the original difference vector: the causal feature vector of the negative sample is V_error, the causal feature vector of the maintenance feedback data is V_correct, and the original difference vector V_diff_raw = V_correct - V_error; the negative sample represents the incorrect diagnosis, and the maintenance feedback data represents the correct diagnosis. S42. Normalize the original difference vector V_diff_raw; S43. Use a feature selection algorithm (e.g., L1 regularization or mutual information-based methods) to filter out the key feature dimensions that distinguish between correct and incorrect diagnoses, and obtain a sparsed difference vector V_diff_sparse; the non-zero elements in it correspond to the key causal features that lead to the model's misdiagnosis. S44. Associate the sparsed difference vector V_diff_sparse with the error type label to generate the causal feature differential encoding; the causal feature differential encoding is a structured data unit containing the sparsed difference vector V_diff_sparse, the error type label, and the original context information.

[0050] After obtaining the Causal Differential Encoding (CDE), this invention employs an efficient Negative Sample Reinforcement (NSR) strategy to optimize the model. Recent research has shown that simply penalizing the model's erroneous outputs without explicitly rewarding correct outputs can significantly improve the model's reasoning ability. The core mechanism of NSR is that it doesn't simply suppress incorrect answers, but rather redistributes the model's probabilistic quality to more reasonable candidate paths by suppressing erroneous reasoning paths. In this invention, when the system receives a negative sample corrected by an engineer, it not only treats the original, erroneous diagnostic suggestion as a negative sample, but more importantly, it pairs this erroneous suggestion with the corresponding Causal Differential Encoding (CDE). During training, the model's goal is to minimize the probability of generating this erroneous suggestion. Simultaneously, the CDE vector acts as a strong supervisory signal, guiding the model's attention to understand "why" the suggestion is wrong. Specifically, the loss function is designed to penalize not only the final erroneous conclusion but also the intermediate reasoning steps leading to that conclusion (if the model is a chained reasoning model). In this way, NSR training enables the model to learn a more robust decision boundary, allowing it to proactively avoid reasoning patterns that have been proven to lead to erroneous conclusions when faced with similar scenarios, thereby significantly improving diagnostic accuracy without sacrificing generative diversity.

[0051] To further enhance the model's ability to understand the subtle differences between "correct" and "incorrect," this invention introduces a contrastive learning mechanism based on NSR. The core idea of ​​contrastive learning is to "bring similar things closer together and push dissimilar things further apart." In the scenario of this invention, for a given fault scenario (anchor point), the correct correction solution provided by the engineer is its "positive sample," while the model's original, incorrect suggestion is its "negative sample." The system constructs a triplet containing (anchor point, positive sample, negative sample). During training, the model aims to minimize the distance between the anchor point and the positive sample in the feature space, while maximizing the distance between the anchor point and the negative sample. This mechanism forces the model to learn more discriminative feature representations, enabling it to clearly distinguish between correct and incorrect diagnostic logic. Furthermore, this invention innovatively combines Causal Differential Encoding (CDE) with contrastive learning. The CDE vector is used as an additional feature, fused with the model's original output features. This fusion can be achieved in various ways, such as feature concatenation, attention mechanisms, or weighted summation. By injecting the deep logical correction information contained in CDE into the feature representation of the model, the effect of contrastive learning is significantly enhanced. The model no longer merely learns to distinguish between "right" and "wrong," but learns to distinguish between "right" and "why it is wrong," thereby achieving a deeper understanding and mastery of diagnostic knowledge.

[0052] In one specific embodiment, in step S5, The differential encoding of the causal features, the corresponding fault diagnosis suggestions, and the corresponding maintenance feedback data are combined into an incremental data training set. The system model training uses a composite loss function, which is expressed as: L_total = L_ce + λ * L_correct; Where L_ce is the cross-entropy loss, used to optimize the model's classification accuracy on positive samples; L_correct is the error correction loss, used to encourage the model to learn the feature differences indicated by the sparsed difference vector V_diff_sparse; λ is a hyperparameter used to balance the weights of the two loss terms. Taking a given fault scenario as the anchor point, fault diagnosis suggestions as negative samples, and maintenance feedback data as positive samples, the training objective of the system model is to minimize the distance between the anchor point and the positive samples in the feature space, while maximizing the distance between the anchor point and the negative samples in the feature space. During the system model training process, the system model parameters are continuously fine-tuned to obtain the optimized system model parameters.

[0053] In this way, the model is explicitly "taught" to identify and avoid previous mistakes while learning new knowledge. This reinforcement training based on differential encoding ensures that every correction made by an engineer is efficiently absorbed by the model, thereby significantly improving the model's diagnostic accuracy and robustness in specific enterprise environments.

[0054] The anchor point is the initial fault scenario that triggers the entire diagnostic process. All comparison operations, whether bringing positive samples closer or pushing negative samples further away, are performed relative to this anchor point. The anchor point specifically includes: the fault code that triggers the diagnosis, related bus message data, and environmental and status information.

[0055] Fault codes triggering diagnosis: Diagnostic Trouble Codes (DTCs) are the core component of the "anchor point." They are generated by the vehicle's self-diagnostic system based on preset fault logic and provide a preliminary qualitative description of the fault. For example, a fault code like "P0128 - Coolant Thermostat Performance" directly indicates that the fault may be related to the cooling system's thermostat. This fault code not only initiates the diagnostic process but also provides important clues and directions for subsequent diagnosis.

[0056] Relevant bus message data: A fault code is merely a result; to deeply understand the cause of the fault, it is necessary to analyze the process data that led to that result. Therefore, the "anchor point" also includes all message data transmitted on the vehicle bus within a certain period before and after the fault code was set. This data typically exists in the form of CAN or LIN messages, containing detailed information about the communication between various ECUs. By analyzing these messages, it is possible to understand which ECUs were communicating when the fault occurred, what signals they transmitted, and whether the values ​​of these signals were normal. For example, by analyzing the messages between the engine control unit (ECU) and the transmission control unit (TCU), it can be determined whether shift shock is related to engine torque control.

[0057] Environmental and Status Information: In addition to bus data, the vehicle's environment and operating status at the time are also crucial for fault diagnosis. This information includes, but is not limited to, vehicle speed, engine speed, coolant temperature, battery voltage, and ambient temperature. This information helps determine whether the fault occurs only under specific operating conditions or exists under all conditions. For example, a fault related to cold starts may only occur when the engine temperature is below a certain threshold, making ambient temperature information particularly important in this case.

[0058] To address the emerging new failure modes and enterprise-specific testing scenarios in automotive bus testing, this invention proposes an innovative incremental training model: a causal generalized incremental learning model. Traditional incremental learning methods, such as those based on experience replay or knowledge distillation, while mitigating catastrophic forgetting to some extent, often struggle with scenarios involving significant changes in data distribution, and their model structures are typically fixed, lacking dynamic scalability. This invention aims to overcome these limitations. It not only continuously learns new knowledge but also dynamically adjusts and expands its structure based on the characteristics of new knowledge, achieving true continuous evolution. The core idea of ​​this model is to combine causal inference with generalized incremental learning. By identifying causal features in new samples, it determines whether to perform parameter fine-tuning or structural expansion. This mechanism enables the model to absorb new knowledge more intelligently and efficiently while maximizing the retention of old knowledge, ensuring that the system's diagnostic capabilities steadily and continuously improve during the learning process.

[0059] The incremental learning phase of this invention forms the basis for its continuous evolution. When the system collects new feedback data (whether positive or negative), it first initiates a rapid evaluation module. This module utilizes a pre-trained causal feature extractor to analyze the similarity between the causal features of the new sample and existing knowledge in the existing model knowledge base. If the causal features of the new sample are highly similar to existing knowledge, i.e., they belong to a variant of a known failure mode, the system triggers a "parameter fine-tuning" mode. In this mode, the system employs efficient parameter fine-tuning techniques, such as LoRA (Low-Rank Adaptation) or Adapter, to quickly update the local parameters of the model. This method has low computational overhead, is fast, and can quickly adapt to subtle changes in known failure modes. However, if the causal features of the new sample show significant differences from any knowledge in the existing knowledge base, the system determines that this may be a completely new and unknown failure mode. At this point, the system triggers a "structure adaptation" mode. In this mode, the system does not immediately update the parameters but instead marks the new sample as a "sample to be structured and expanded" and stores it in a dedicated cache. This adaptive mechanism ensures that the model can adopt the most appropriate response strategy when faced with different types of incremental data, avoiding ineffective parameter adjustments on irrelevant new knowledge, thereby improving the efficiency and stability of incremental learning.

[0060] When the number of "samples awaiting structural expansion" in the cache reaches a certain level, or when their importance (e.g., marked as "critical faults" by engineers) reaches a threshold, the system model of this invention will initiate a structural incremental learning mechanism. This is the core innovation that distinguishes it from traditional incremental learning methods. Traditional incremental learning models have static structures and cannot effectively handle entirely new knowledge domains. This invention, however, dynamically expands its network structure to accommodate new knowledge. Specifically, the system analyzes the causal characteristics of these new samples and attempts to construct a new sub-network module specifically designed to handle these new fault modes. This sub-network module can be a new attention head, a new feedforward network layer, or a complete Transformer block, its specific structure determined by the complexity and characteristics of the new knowledge. After constructing the new module, the system uses a "teacher-student" framework for training. The original, unexpanded model acts as the "teacher," providing stable guidance on old knowledge; while the newly expanded model (containing the new module) acts as the "student," learning how to handle new fault modes under the supervision of the teacher model. This knowledge distillation training method ensures that while learning new knowledge, catastrophic forgetting of old knowledge is minimized. In this way, the model structure of the present invention can dynamically evolve with the growth of knowledge, giving it a powerful ability to handle open-world problems.

[0061] In one specific embodiment, the system model has a dynamically expandable model structure. When it is found that the system model's diagnostic accuracy for a newly emerging type of fault mode is consistently low, or the feature complexity of the new data exceeds the expressive power of the current system model, a structural expansion instruction for the system model is triggered. The system dynamically adds new "enhanced node groups" to the network. These new nodes are randomly initialized and connected to existing nodes. Then, the system model freezes the parameters of the original network and trains only the newly added nodes, enabling them to learn how to fit residual information that the current system model cannot handle.

[0062] This incremental structural expansion is similar to continuously building new modules on an existing knowledge framework, rather than starting from scratch. It effectively solves the problem of "catastrophic forgetting" because existing knowledge is fully preserved in the frozen network, while new knowledge is learned through the newly added structure. In this way, the system can function like a continuously evolving expert, its knowledge base and diagnostic capabilities becoming increasingly complex and powerful with accumulated experience.

[0063] In one specific embodiment, the conditions for triggering training in step S5 include: Quantity-driven: When the number of newly collected negative samples reaches a set threshold; or, Performance-driven: When the diagnostic accuracy of the system model for a specific type of fault continuously decreases beyond a set threshold; or, Time-driven: Set a fixed time period for regular system model training and updates.

[0064] The continuous evolution of the network model in this invention is ensured by a comprehensive performance monitoring and evaluation system. After each parameter fine-tuning or structural expansion, the system generates a new model version and automatically performs a series of rigorous evaluations. These evaluations include, but are not limited to: diagnostic accuracy on new samples, performance retention on existing test sets (to measure the degree of catastrophic forgetting), model inference latency, and resource consumption. The system employs A / B testing to compare the new and old models in a small-scale real-world environment, collecting feedback data from engineers to objectively evaluate the new model's actual performance. If the new model performs well across all evaluation metrics, it will be officially deployed, replacing the old model. If the new model performs poorly in some aspects (e.g., while improving the diagnosis rate of new faults, it significantly reduces the diagnosis rate of old faults), the system will trigger a rollback mechanism, reverting to the previous stable version and recording the failed incremental learning attempt for subsequent analysis and optimization. Furthermore, the system regularly archives and analyzes all model versions, summarizing the patterns and trends of incremental learning to provide data support for future model architecture design and training strategy optimization. This closed-loop monitoring and evaluation system ensures that every "evolution" of the system model is beneficial and controllable, thereby guaranteeing the long-term stability and reliability of the entire AI diagnostic system.

[0065] The system fine-tuning architecture of the present invention is as follows: Figure 3 As shown.

[0066] This invention discloses an AI diagnostic system for automotive bus testing. The system is used to implement the aforementioned closed-loop feedback and self-learning method. The system includes: The fault diagnosis module is used to provide fault diagnosis suggestions based on specific fault scenarios; The maintenance feedback module is used to collect maintenance feedback data after the fault repair is completed. The maintenance feedback data is the correct fault diagnosis. The causal feature extraction module is used to extract the causal features from the negative sample when the fault diagnosis suggestion is incorrect, and at the same time extract the causal features from the corresponding maintenance feedback data to form two causal feature vectors. A differential encoding module is used to calculate the difference between the two causal feature vectors and generate causal feature differential encoding. The incremental training module uses the causal feature differential encoding as training samples to perform incremental training and fine-tuning on the system model.

[0067] In one specific embodiment, the system further includes: The training trigger module is used to set training trigger conditions. When the trigger conditions are met, the system model training is triggered. The dynamic expansion module adds new "enhanced node groups" to the network and connects them to existing nodes when the system model's diagnostic accuracy for a newly emerging fault mode remains low, or when the feature complexity of newly collected data exceeds the expressive power of the current system model. The A / B testing module automatically performs A / B testing when the system model training is updated to form a new version, comparing the new version with the old version on the validation set. Only when the new version outperforms the old version in all key performance indicators will it be automatically deployed as the new version. If the new version performs poorly, the system model will automatically revert to the old version.

[0068] While several embodiments of the present invention have been provided herein, those skilled in the art should understand that modifications can be made to these embodiments without departing from the spirit of the invention. The above embodiments are merely exemplary and should not be construed as limiting the scope of the invention.

Claims

1. A closed-loop feedback and self-learning method for an automotive bus testing AI diagnostic system, characterized in that, The method includes: S1. System generates fault diagnosis suggestions: For a certain test fault, the AI ​​diagnostic system provides fault diagnosis suggestions, which are presented in structured data. S2. Collect maintenance feedback data: For the test fault in step S1, after the processing is completed, collect maintenance feedback data, which is also presented as structured data. S3. Extract causal features: When a fault diagnosis suggestion is incorrect, define this fault diagnosis as a negative sample; extract the causal features of the negative sample, and at the same time extract the causal features of the corresponding maintenance feedback data to form two causal feature vectors; S4. Calculate the difference between the two causal feature vectors obtained in step S3, and generate causal feature differential encoding; S5. Incremental training and fine-tuning of the system model: When the differential encoding of causal features obtained in step S4 accumulates to a set quantity, the model training condition is triggered; the differential encoding of causal features is used as training samples to perform incremental training and fine-tuning of the system model. S6. Use the finely tuned system model for fault diagnosis; In step S3, the process for extracting causal features is as follows: S31. Obtain all raw data of bus messages, the currently loaded DBC file version, the firmware version information of the relevant ECU, and the temperature and voltage of the test environment within a period of time before and after the diagnostic time point. S32. Extract preset causal features from the data obtained in step S31. The causal features include global features, multi-scale local features, and signal correlation features. The global features are used to capture the macroscopic state of the entire bus system. The multi-scale local features are features related to specific nodes or signals. The signal correlation features are used to determine whether two or more signals exhibit abnormal synchronous changes when a fault occurs. S33. Extract maintenance-related features from the negative samples and the maintenance feedback data respectively. The maintenance-related features are used to reflect the characteristics of interaction with the system during the maintenance process. S34. Combine the causal features preset in step S32 with the maintenance-related features in step S33 to form a high-dimensional causal feature vector; the negative sample and the maintenance feedback data each correspond to one of the causal feature vectors.

2. The closed-loop feedback and self-learning method of the automotive bus test AI diagnostic system as described in claim 1, characterized in that, In step S32, the global features include: average bus load rate, peak load rate, frequency of error frames, and bus utilization; the multi-scale local features include: periodicity of specific ID messages, calculation of the standard deviation and jitter of their transmission interval to determine whether there is a timing anomaly; statistical features and rate of change of the signal, and detection of abnormal jumps or drifts, the statistical features including mean, variance, kurtosis, and skewness; the signal correlation features include: determining whether two seemingly independent signals exhibit abnormal synchronous changes when a fault occurs; In step S33, the maintenance-related features include maintenance response signal features, fault code and maintenance action association features, and maintenance text features.

3. The closed-loop feedback and self-learning method of the automotive bus test AI diagnostic system as described in claim 1, characterized in that, The specific steps of step S4 are as follows: S41. Calculate the original difference vector: the causal feature vector of the negative sample is V_error, the causal feature vector of the maintenance feedback data is V_correct, and the original difference vector V_diff_raw = V_correct - V_error; the negative sample represents the incorrect diagnosis, and the maintenance feedback data represents the correct diagnosis. S42. Normalize the original difference vector V_diff_raw; S43. Use a feature selection algorithm to filter out the key feature dimensions that distinguish between correct and incorrect diagnoses, and obtain a sparsed difference vector V_diff_sparse. S44. Associate the sparsed difference vector V_diff_sparse with the error type label to generate the causal feature differential encoding; the causal feature differential encoding is a structured data unit containing the sparsed difference vector V_diff_sparse, the error type label, and the original context information.

4. The closed-loop feedback and self-learning method of the automotive bus test AI diagnostic system as described in claim 3, characterized in that, In step S5, The differential encoding of the causal features, the corresponding fault diagnosis suggestions, and the corresponding maintenance feedback data are combined into an incremental data training set. The system model training uses a composite loss function, which is expressed as: L_total = L_ce + λ * L_correct; Where L_ce is the cross-entropy loss, used to optimize the model's classification accuracy on positive samples; L_correct is the error correction loss, used to encourage the model to learn the feature differences indicated by the sparsed difference vector V_diff_sparse; λ is a hyperparameter used to balance the weights of the two loss terms. Taking a given fault scenario as the anchor point, fault diagnosis suggestions as negative samples, and maintenance feedback data as positive samples, the training objective of the system model is to minimize the distance between the anchor point and the positive samples in the feature space, while maximizing the distance between the anchor point and the negative samples in the feature space. During the system model training process, the system model parameters are continuously fine-tuned to obtain the optimized system model parameters.

5. The closed-loop feedback and self-learning method of the automotive bus test AI diagnostic system as described in claim 1, characterized in that, The system model has a dynamically expandable model structure. When it is found that the system model has a persistently low accuracy in diagnosing a new type of fault mode, or when the feature complexity of the new data exceeds the expressive power of the current system model, the system model's structure expansion instruction is triggered. The system model dynamically adds new "enhanced node groups". These new nodes are randomly initialized and connected to the existing nodes. Then, the system model freezes the parameters of the original system model and only trains the newly added nodes, so that they learn how to fit the residual information that the current system model cannot handle.

6. The closed-loop feedback and self-learning method of the automotive bus test AI diagnostic system as described in claim 1, characterized in that, In step S5, the conditions for triggering training include: Quantity-driven: When the number of newly collected negative samples reaches a set threshold; or, Performance-driven: When the diagnostic accuracy of the system model for a specific type of fault continuously decreases beyond a set threshold; or, Time-driven: Set a fixed time period for regular system model training and updates.

7. The closed-loop feedback and self-learning method of the automotive bus test AI diagnostic system as described in claim 5, characterized in that, Each time the system model is trained and updated, a new version is generated. A / B testing is automatically performed to compare the new version with the old version on a validation set. The new version is automatically deployed as the actual use version only when it outperforms the old version in all key performance indicators. If the new version performs poorly, the system model will automatically revert to the old version and generate a detailed analysis report. The key performance indicators include diagnostic accuracy, performance retention, inference latency, and resource consumption.

8. An AI diagnostic system for automotive bus testing, characterized in that, The system is used to implement the closed-loop feedback and self-learning method as described in any one of claims 1-7, the system comprising: The fault diagnosis module is used to provide fault diagnosis suggestions based on specific fault scenarios; The maintenance feedback module is used to collect maintenance feedback data after the fault repair is completed. The maintenance feedback data is the correct fault diagnosis. The causal feature extraction module is used to extract the causal features from the negative sample when the fault diagnosis suggestion is incorrect, and at the same time extract the causal features from the corresponding maintenance feedback data to form two causal feature vectors. A differential encoding module is used to calculate the difference between the two causal feature vectors and generate causal feature differential encoding. The incremental training module uses the causal feature differential encoding as training samples to perform incremental training and fine-tuning on the system model.

9. The automotive bus testing AI diagnostic system as described in claim 8, characterized in that, The system also includes: The training trigger module is used to set training trigger conditions. When the trigger conditions are met, the system model training is triggered. The dynamic expansion module can dynamically add new "enhanced node groups" and connect them to existing nodes when the system model is found to have a persistently low accuracy in diagnosing a new type of fault mode, or when the feature complexity of newly collected data exceeds the expressive power of the current system model. The A / B testing module automatically performs A / B testing when the system model training is updated to form a new version, comparing the new version with the old version on the validation set. Only when the new version outperforms the old version in all key performance indicators will it be automatically deployed as the new version. If the new version performs poorly, the system model will automatically revert to the old version.

Images