Data processing method of intelligent cabin and vehicle

Abstract

The invention relates to a data processing method of an intelligent cabin and a vehicle, and belongs to the technical field of data processing, and the method comprises the steps: obtaining original dialogue data generated in the intelligent cabin; converting the original dialogue data into sentence vectors, and performing duplicate removal on the sentence vectors to obtain a dialogue list; clustering the sentence vectors in the dialogue list, classifying the sentence vectors with similar semantics into the same cluster, and selecting and collecting dialogue samples corresponding to the sentence vector closest to the centroid of the cluster in each cluster to obtain a selected subset; performing privacy protection processing on the selected subset to obtain a distillation data set; and the distillation data set is uploaded to the cloud server, so that direct transmission of original user data is avoided, the privacy disclosure risk is reduced, fragmentation computing power of the vehicle-mounted terminal can be fully utilized, and distributed collaboration of data processing is realized.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and more particularly to a data processing method for an intelligent cockpit and a vehicle. Background Technology

[0002] With the widespread adoption of AI-powered dialogue applications in smart cockpits, massive amounts of in-vehicle human-machine interaction data are constantly being generated. However, this raw interaction data is riddled with redundancy, noise, and sensitive information. Directly using it for model training would lead to low efficiency and a high risk of privacy breaches. Therefore, cleaning, deduplication, and anonymization of the raw data—interaction data distillation—has become a necessary preprocessing step for extracting high-value training resources from massive amounts of low-value data.

[0003] Currently, existing methods typically upload all raw dialogue data generated during in-vehicle interaction directly to the cloud for processing. The cloud then utilizes its powerful computing capabilities to uniformly perform data distillation operations, ultimately obtaining a refined dataset that can be used for training.

[0004] The aforementioned existing technologies have the following drawbacks: the upload of raw data poses a risk of leakage of user privacy information; the transmission of massive amounts of raw data consumes a large amount of communication bandwidth and generates high storage costs; and cloud servers need to process redundant data uploaded by all terminals, facing enormous computing pressure and resource consumption. Summary of the Invention

[0005] This application addresses, to at least some extent, one of the technical problems in the related art.

[0006] Therefore, this application aims to provide a data processing method and vehicle for a smart cockpit. By acquiring data generated within the smart cockpit, performing sentence vector transformation, deduplication, and clustering to obtain a selected subset, and finally performing privacy protection processing to obtain a distilled dataset, the data is then uploaded to a cloud server.

[0007] To achieve the above objectives, in a first aspect, this application provides a data processing method for an intelligent cockpit, comprising: Acquire raw dialogue data generated within the smart cockpit; The original dialogue data is converted into sentence vectors, and duplicate sentence vectors are removed to obtain a dialogue list; Cluster the sentence vectors in the dialogue list, group the sentence vectors with similar semantics into the same cluster, select and collect the dialogue samples corresponding to the sentence vectors that are closest to the centroid of the cluster in each cluster, and obtain a selected subset; The selected subset is then subjected to privacy protection processing to obtain the distillation dataset; The distillation dataset was uploaded to a cloud server.

[0008] In existing technologies, the interaction data of smart cockpits is usually uploaded directly and without processing to the cloud for processing. This approach leads to a huge risk of privacy leakage for the user's sensitive raw dialogue data during transmission and cloud storage. At the same time, the massive amount of raw data also consumes a lot of network bandwidth and puts a heavy storage and computing burden on cloud servers.

[0009] This application, through the aforementioned solution, completes the entire process of original dialogue data processing locally on the vehicle terminal, from vectorization, deduplication, clustering and selection to privacy protection. Ultimately, only the distilled essence data is uploaded. This fundamental change avoids the direct transmission of original user data from the data source, thereby fundamentally and significantly reducing the risk of privacy leakage and significantly reducing network bandwidth consumption and cloud server pressure.

[0010] Furthermore, this application constructs a complete and collaborative edge data processing pipeline. Compared with the existing architecture that treats the terminal as a simple data collection node and centralizes all intelligent processing in the cloud, this application can make full use of the fragmented computing power of the vehicle terminal to achieve distributed collaboration in data processing. Thus, in the resource-constrained vehicle environment, it can systematically achieve an optimized balance between privacy and security, communication efficiency and computing load.

[0011] In some embodiments of this application, the method for converting the original dialogue data into sentence vectors and deduplicating the sentence vectors to obtain a dialogue list includes: The original dialogue data is converted into sentence vectors using a pre-trained semantic similarity model; The signature of the sentence vector is calculated using a text similarity algorithm. The Hamming distance of the signatures is compared, and signatures whose Hamming distance is within a threshold are removed to obtain the dialogue list.

[0012] In the technical solution, a pre-trained semantic similarity model is used for vectorization transformation, and a signature-based text similarity algorithm is used for deduplication. This ensures that dialogues with similar semantics but different expressions can be accurately identified, thereby efficiently removing redundant content while preserving semantic diversity.

[0013] By employing the computationally efficient signature comparison method described above, a high-quality list of dialogues with no duplicates and comprehensive semantic coverage was quickly obtained while ensuring deduplication, laying a solid data foundation for subsequent fine-grained filtering.

[0014] In some embodiments of this application, the method for clustering sentence vectors in the dialogue list includes: Based on the dialogue list, a clustering algorithm is used to cluster the sentence vectors in the dialogue list, and the sentence vectors with similar semantics are grouped into the same cluster; Calculate the distance from each sentence vector in each cluster to the centroid of the cluster, and select the dialogue sample corresponding to the sentence vector whose distance is closest to the centroid of the cluster; The dialogue sample is used as the representative dialogue of this cluster. The representative dialogues of all clusters are then collected to obtain the selected subset.

[0015] In the technical solution, by adopting a clustering algorithm and implementing a strategy of selecting the sample closest to the centroid, the most representative samples are automatically selected from massive dialogues. The clustering process ensures that all major semantic patterns can be identified and covered, while the mechanism of selecting representative samples from the center of each cluster ensures that each selected sample is the most typical and core instance in its semantic category.

[0016] The above method avoids the subjectivity and high cost of manual screening, thereby automatically generating a high-quality selected subset that significantly compresses the data volume while preserving the semantic diversity and representativeness of the original data to the greatest extent.

[0017] In some embodiments of this application, the method for performing privacy protection processing on the selected subset to obtain the distillation dataset includes: Determine whether the number of dialogue samples in the selected subset meets the preset dynamic sample quantity; If the number of dialogue samples in the selected subset meets the preset dynamic sample number, then the dialogue samples in the selected subset will be subjected to privacy protection processing to obtain the distillation dataset.

[0018] In the technical solution, by introducing a mechanism to determine the number of selected subset samples, dynamic triggering conditions are added to the privacy protection processing stage, which can intelligently determine the subsequent operation path based on the actual amount of data obtained from distillation.

[0019] When the amount of data is sufficient, protective processing is performed directly, thereby optimizing the processing flow while satisfying the data utility, avoiding unnecessary waste of computing resources, enhancing the system's adaptability and processing efficiency in different interaction scenarios, and ensuring the flexibility and robustness of the data distillation process.

[0020] In some embodiments of this application, the method for obtaining a distillation dataset by performing privacy protection processing on the selected subset further includes: Determine whether the number of dialogue samples in the selected subset meets the preset dynamic sample quantity; If the number of dialogue samples in the selected subset does not meet the preset dynamic sample quantity, then pseudo samples are generated using the preset local model. The pseudo-samples and the selected subset are combined to obtain a merged dataset. The dialogue samples in the merged dataset are then subjected to privacy protection processing to obtain the distillation dataset.

[0021] In the technical solution, when the number of samples is insufficient, a supplementary mechanism that generates pseudo-samples using the local model is activated, which effectively solves the problem of insufficient or undiversified cloud model training data that may be caused by an initially selected subset being too small.

[0022] By generating pseudo-samples that are semantically similar to the original data locally, the scale and diversity of the dataset can be safely expanded while protecting privacy. This ensures that the data uploaded to the cloud always has sufficient scale and information to support effective model training, thereby improving the reliability and output quality of the entire distillation system.

[0023] In some embodiments of this application, the method for generating pseudo-samples using a preset local model further includes: Obtain user question data and intelligent response data from the dialogue samples in the selected subset; The user question data is segmented into words to identify keywords in the user question data; Synonyms are selected from a preset thesaurus to replace the keywords, resulting in replaced user question data; By combining the replaced user question data with the intelligent response data, enhanced data is obtained; The augmented data is added to the selected subset to obtain the augmented dataset, and the local model is used to generate pseudo-samples based on the dialogue samples in the augmented dataset.

[0024] In this technical solution, by performing synonym substitution on keywords in user questions, new questions with similar sentence structures but different wording can be created while basically maintaining the semantics of the original sentences. This approach has low computational cost and stable results, effectively increasing the language diversity and vocabulary coverage of the dataset. It provides a low-cost and high-quality data source for the model to learn more robust language patterns.

[0025] In some embodiments of this application, the method for generating pseudo-samples using a preset local model includes: The dialogue samples from the selected subset or the augmented dataset are input into the local model, and the local model is used to perform semantic encoding on the dialogue samples to generate new dialogues that are semantically similar to the dialogue samples as pseudo-samples.

[0026] In this technical solution, new dialogues are created by utilizing local models for semantic encoding. This allows for a deeper understanding of the underlying semantics of the original dialogue, and the generation of new dialogues that differ not only in vocabulary and sentence structure but also in core meaning. Figure 1 A new dialogue.

[0027] The data generated by the above methods is richer and more natural in terms of language variation, and can better simulate the various expressions of real users, thereby providing the cloud server with high-quality training samples that are closer to the real data distribution and significantly improving its generalization ability.

[0028] In some embodiments of this application, the method for obtaining a merged dataset by combining the pseudo-samples and the selected subset further includes: Input the selected subset or the dialogue samples in the enhanced dataset into the preset teacher model; The teacher model is used to infer user question data in the dialogue samples and generate a predicted probability distribution as a soft label. The soft labels are attached to the corresponding pseudo-samples, and the merged dataset is obtained by combining the selected subset.

[0029] In the technical solution, by introducing a teacher model and attaching soft labels to the data, the simple data augmentation process is transformed into a knowledge injection process. The soft labels generated by the teacher model contain a richer probabilistic understanding of the dialogue intent, which contains far more information than a single hard label. By attaching these soft labels to pseudo-samples, the uploaded merged dataset not only contains the data itself but also carries the knowledge of the teacher model. This enables more efficient guidance for the training of the model in the cloud, accelerates its convergence, and improves its performance, thus realizing knowledge transfer from the vehicle to the cloud.

[0030] In some embodiments of this application, the method for obtaining a distillation dataset by performing privacy protection processing on the selected subset further includes: The dialogue samples in the selected subset or the merged dataset are anonymized using regular expressions, and zero-mean Gaussian noise is added to the sentence vectors in the dialogue samples to obtain the distillation dataset.

[0031] In the technical solution, a dual protection mechanism is constructed by combining two complementary privacy protection technologies: text desensitization and vector noise addition. Regular expression desensitization directly removes sensitive entity information from the text, thus solving the risk of direct identification.

[0032] Adding Gaussian noise to sentence vectors disrupts the original feature distribution of the data, increases the difficulty of inferring the original content through reverse engineering, solves the risk of indirect inference, and provides comprehensive and reliable security for user privacy while preserving the semantic utility of the data for training to the maximum extent.

[0033] Secondly, this application provides a vehicle, including: The in-vehicle terminal is used to acquire raw dialogue data generated within the smart cockpit. A processor, which is communicatively connected to the vehicle terminal, is equipped with a pre-trained semantic similarity model. The processor is further configured as follows: Receive the original dialogue data, use the pre-trained semantic similarity model to convert the original dialogue data into sentence vectors, and deduplicate the sentence vectors to obtain a dialogue list; Cluster the sentence vectors in the dialogue list, group the sentence vectors with similar semantics into the same cluster, select and collect the dialogue samples corresponding to the sentence vectors that are closest to the centroid of the cluster in each cluster, and obtain a selected subset; The data in the selected subset is processed for privacy protection to obtain the uploaded data; The uploaded data is uploaded to the cloud server.

[0034] In the technical solution, by limiting the communication connection between the processor and the vehicle terminal and configuring a specific model, the real-time performance and stability of the data processing process are guaranteed. This can meet the stringent requirements of the vehicle for system reliability and response speed in the actual driving environment, and ensure that the entire data distillation process can operate independently, efficiently and safely in the vehicle.

[0035] As can be seen from the above technical solutions, additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0036] Figure 1 A schematic flowchart of a data processing method for an intelligent cockpit provided in an embodiment of this application; Figure 2 A schematic diagram of another data processing method for a smart cockpit provided in this application embodiment; Figure 3 A schematic diagram of another data processing method for a smart cockpit provided in this application embodiment; Figure 4 This is a structural schematic diagram of a vehicle provided in an embodiment of this application. Detailed Implementation

[0037] In this application, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0038] The present application will now be described in detail through exemplary embodiments. However, it should be understood that, without further description, elements, structures, and features in one embodiment may be advantageously incorporated into other embodiments. It should be noted that in the automotive field, with the rapid development of smart cockpits, artificial intelligence big data technology has been gradually integrated into in-vehicle human-machine interaction.

[0039] Current mainstream smart cockpit systems are generally equipped with voice assistants and multimodal dialogue functions, which can generate interactive feedback such as information cards and recommended content based on user commands.

[0040] However, due to the limitations of computing power and storage resources of the intelligent cockpit terminal itself, such generative artificial intelligence functions usually need to rely on cloud services, which affects the real-time nature of interaction and brings potential risks of user data privacy leakage.

[0041] Although the industry has begun to explore edge-cloud collaborative computing models and attempted to bring some commonly used models down to the vehicle end to alleviate cloud pressure, overall, the existing AI dialogue applications in smart cockpits still have significant limitations in terms of effective use of terminal computing power, acquisition of high-quality scenario-based training data, and protection of user privacy and security.

[0042] In existing technologies, the vast majority of raw interactive data generated by smart cockpits needs to be uploaded to the cloud for unified processing and distillation.

[0043] However, the original dialogue data contains a large amount of sensitive user information, and the entire process of uplink transmission and centralized cloud storage poses a risk of privacy leakage.

[0044] Meanwhile, the massive amount of raw data is full of redundant content. Directly uploading it not only consumes high communication bandwidth, but also puts enormous computing and storage pressure on cloud servers.

[0045] In addition, the cost and efficiency of cloud-based unified data cleaning, deduplication and annotation are high, making it difficult to generate high-quality and diverse labeled data for vehicle scenarios, which restricts the further improvement of model optimization results.

[0046] Based on this, this application proposes a data processing method and vehicle for an intelligent cockpit. By integrating semantic vectorization, deduplication, clustering of dialogue data, and privacy protection processing of output results into the vehicle terminal, the method achieves the effect of distillation and sublimation of data before it leaves the vehicle. This solves the core problems of high user privacy and security risks, large consumption of communication and computing resources, and insufficient quality of scenario-based training data brought about by the traditional cloud-based centralized processing mode.

[0047] In the following, embodiments of this application will be described in detail with reference to the accompanying drawings.

[0048] As attached Figures 1 to 2 As shown in an illustrative embodiment of the data processing method for the intelligent cockpit in this application, the data processing method includes: S1: Acquire raw dialogue data generated within the smart cockpit.

[0049] In some embodiments, the raw dialogue data includes user input data and card results.

[0050] The user input data refers not only to the original speech waveform file, but also to data processed by the front end. User input data includes speech-recognized text, semantic intent, and triggering scenarios.

[0051] Speech-to-text (STR) is the text that is recognized from what a user says. For example, if a user says, "Hello voice assistant, I'm a little hot," the corresponding text is "I'm a little hot."

[0052] Semantic intent is the core user instruction derived from natural language understanding and parsing. For example, the intent parsed from "I'm a little hot" might be "adjust the air conditioner".

[0053] A trigger scenario is the vehicle's state or scene information at the time of the interaction. For example, trigger scenarios could include driving, driver's side, nighttime, etc.

[0054] Card-based results are a method of information presentation in smart cockpits, serving as a structured user interface template. Compared to simple text-based or voice-based responses, card-based results contain richer, more machine-readable information.

[0055] S2: Convert the original dialogue data into sentence vectors, and remove duplicates from the sentence vectors to obtain the dialogue list.

[0056] In some embodiments, a method for converting raw dialogue data into sentence vectors and deduplicating the sentence vectors to obtain a dialogue list includes: The original dialogue data is converted into sentence vectors using a pre-trained semantic similarity model; The signature of the sentence vector is calculated using a text similarity algorithm. The Hamming distance of the signatures is compared, and signatures with a Hamming distance within a threshold are removed to obtain the dialogue list.

[0057] By using a pre-trained semantic similarity model for vectorization and employing a signature-based text similarity algorithm for deduplication, it is ensured that dialogues with similar semantics but different expressions can be accurately identified, thereby efficiently eliminating redundant content while preserving semantic diversity.

[0058] Furthermore, every interaction between the user and the smart cockpit is converted into a sentence vector in real time through the SBERT model. The generated sentence vector is not immediately compared with the previous vector, but is first stored in a local temporary cache or database. The system will preset a trigger condition, and only when the condition is met will the deduplication operation be performed on all sentence vectors in the cache. The trigger condition can typically be time-driven, quantity-driven, or event-driven. When the triggering condition is met, the system will read all sentence vectors accumulated in the past period from the temporary cache.

[0059] Subsequently, a text similarity algorithm is used to generate a signature for each sentence vector, and pairwise comparisons are performed within this batch of signatures to remove duplicates with excessively close Hamming distances.

[0060] After deduplication is complete, the system will output a list of deduplicated conversations; Subsequently, the temporary cache will be cleared or updated to prepare for the next accumulation cycle.

[0061] By employing the computationally efficient signature comparison method described above, a high-quality list of dialogues with no duplicates and comprehensive semantic coverage was quickly obtained while ensuring deduplication, laying a solid data foundation for subsequent fine-grained filtering.

[0062] Furthermore, the original dialogue data undergoes sentence vector transformation and deduplication as a coarse distillation stage. The coarse distillation stage is mainly responsible for transforming the original dialogue data into a machine-understandable semantic vector representation and performing efficient deduplication based on semantic similarity.

[0063] By using a pre-trained semantic similarity model, each original dialogue data is converted into a high-dimensional sentence vector, embedding the semantic information of the text into the vector space. This makes dialogues with similar semantics closer in space, laying the foundation for subsequent similarity calculations.

[0064] The semantic similarity model can be the SBERT model. The SBERT (Sentence-Bidirectional Encoder Representation from Transformers) model is a sentence embedding generation model based on BERT. It fine-tunes BERT through the structure of Siamese network and triplet network to generate sentence embeddings with semantic information.

[0065] For example, in the process of converting raw dialogue data into sentence vectors, it is first necessary to obtain an instance of the ONNXRuntime runtime environment; ONNX Runtime is an open engine specifically designed for efficiently running AI models conforming to the ONNX standard format on various hardware and software platforms. Next, the system will create a session configuration options object. This object is primarily used to set various parameters for model runtime, including configuration options such as the execution provider and the number of threads. Then, the program will open the pre-trained SBERT model file from the Assets folder. The SBERT model file uses INT8 quantization format and reads its entire contents into a byte array. Using this byte array and the previously created session configuration, the system establishes a new model session in the ONNX Runtime environment, which serves as the core instance for subsequent inference computation. To perform the text-to-vector conversion, the system defines a dedicated encoding function. This function accepts text input in string form, processes it, and returns a floating-point array representing semantic features, i.e., a 768-dimensional semantic vector.

[0066] During the function execution, the word segmenter is first called to process the input text, converting it into a sequence of numeric IDs that the model can understand, thus realizing the conversion from natural language to machine-readable numeric representation.

[0067] Subsequently, the system encapsulates the obtained numeric ID sequence into a tensor format required by the ONNX Runtime. By reorganizing the one-dimensional array into a two-dimensional array, the desired input shape requirement of the model is met. The batch size is set to 1, indicating that only a single sentence is processed at a time.

[0068] After tensor preparation is complete, the system performs model inference through the session object. During this process, a key-value mapping needs to be passed in, explicitly specifying the name of the input node and its corresponding input tensor. By creating a single mapping, it ensures that input data is correctly delivered to the designated input of the model.

[0069] After the model inference is complete, it returns a mapping set containing the results of all output nodes. This is achieved by obtaining the numerical content of the first output node and converting it into a corresponding two-dimensional floating-point array. Since the batch size is 1, the output array has a shape of 1 row and 768 columns, which corresponds precisely to the 768-dimensional semantic vector representation of a single sentence.

[0070] Finally, the system extracts the first row element from the two-dimensional array result and returns it as a one-dimensional 768-dimensional floating-point array, completing the entire conversion process from the original text to the semantic vector.

[0071] Furthermore, based on the obtained sentence vectors, a text similarity algorithm is used to achieve efficient redundancy filtering.

[0072] First, the sentence vector is compressed into a 64-bit binary signature. Then, by calculating and comparing the Hamming distance between different signatures, it is quickly determined whether the corresponding text is approximately repeated.

[0073] For newly arrived data, if the Hamming distance between its signature and any existing signature in the preset set of recognized signatures is less than a set threshold, it is considered redundant content and is removed.

[0074] A text similarity algorithm can be the SimHash algorithm. The SimHash algorithm is a similarity hashing algorithm that can quickly determine the similarity of original text by calculating the Hamming distance between two hash values.

[0075] For example, in the implementation of sentence vector deduplication, two core data structures are first initialized: one is used to store the fingerprint database of the SimHash signatures corresponding to the confirmed unique dialogues, and the other is used to store the list of unique dialogue objects after deduplication. At the start of the processing flow, the input raw dialogue data set is traversed. For each dialogue unit in the set, the user's question and the AI's response are concatenated to form a complete dialogue text. This allows the entire question-and-answer pair to be processed as a complete semantic unit, effectively avoiding the situation where valid dialogues are incorrectly removed because the user's questions are the same but the system's responses are different. After text concatenation, the SimHash algorithm is used to extract features from the concatenated complete dialogue text, generating a corresponding 64-bit hash signature. This signature serves as the unique digital fingerprint of the dialogue segment and will form the basis for subsequent similarity comparisons. Next, a streaming processing method is used to compare the currently generated signature with all existing signatures in the fingerprint database one by one. During the comparison process, bitwise XOR operations are used to calculate the differences between signatures, and the Hamming distance between two signatures is accurately calculated by counting the number of binary bits with 1 in the result. Based on an empirical threshold, when the Hamming distance is less than or equal to 3, the two signatures are considered to be highly similar, and it is inferred that the corresponding dialogue content is semantically similar and repetitive. When it is confirmed that the Hamming distance between the current dialogue signature and all existing signatures in the fingerprint database exceeds a preset threshold, the dialogue content is determined to be a new unique dialogue. At this time, the signature of the dialogue is added to the fingerprint database for subsequent comparison reference, and the dialogue object is added to the unique dialogue result list. After the above processing steps, the final output is a clean set of dialogue data that has undergone rigorous deduplication. This result will serve as input for subsequent data processing, laying the foundation for subsequent cluster analysis and data selection. The entire deduplication process, through an efficient signature comparison mechanism, achieves efficient deduplication of massive amounts of dialogue data while ensuring the accuracy of semantic recognition.

[0076] To further improve processing efficiency for massive amounts of data, the MinHash-LSH library can be introduced to achieve large-scale similar dialogue detection and deduplication within a near-linear time complexity.

[0077] For example, by using MinHash to generate signatures for a set of texts and then using LSH to cluster similar signatures, most duplicate pairs can be found in O(n) time.

[0078] MinHash-LSH is an approximate similarity retrieval technique that combines minimum hashing and locality-sensitive hashing. Among them, MinHash (Miniature Hashing) is a minimal hashing algorithm used to estimate the similarity between two sets; LSH (Locality-Sensitive Hashing) is a locality-sensitive hashing algorithm that uses MinHash signatures to bucket similar items into the same hash bucket, thereby greatly reducing the range that needs to be accurately compared and achieving highly efficient similar item retrieval.

[0079] Through the two-stage processing of vectorization and hash deduplication described above, semantically repetitive or highly similar dialogue entries can be significantly eliminated, greatly reducing the data scale required for subsequent modules and providing a preliminary cleaned data foundation for the entire distillation process.

[0080] S3: Cluster the sentence vectors in the dialogue list, group the sentence vectors with similar semantics into the same cluster, select and collect the dialogue samples corresponding to the sentence vectors that are closest to the centroid of the cluster in each cluster, and obtain a selected subset.

[0081] In some embodiments, a method for clustering sentence vectors in a dialogue list includes: Based on the dialogue list, a clustering algorithm is used to cluster the sentence vectors in the dialogue list, and sentence vectors with similar semantics are grouped into the same cluster; Calculate the distance from each sentence vector in each cluster to the centroid of the cluster, and select the dialogue sample corresponding to the sentence vector that is closest to the centroid of the cluster; The dialogue sample is used as the representative dialogue of this cluster. The representative dialogues of all clusters are collected to obtain a selected subset.

[0082] By employing a clustering algorithm and implementing a strategy of selecting the sample closest to the centroid, the most representative samples are automatically filtered from massive dialogues. The clustering process ensures that all major semantic patterns can be identified and covered, while the mechanism of selecting representative samples from the center of each cluster guarantees that each selected sample is the most typical and core instance in its semantic category.

[0083] The above method avoids the subjectivity and high cost of manual screening, thereby automatically generating a high-quality selected subset that significantly compresses the data volume while preserving the semantic diversity and representativeness of the original data to the greatest extent.

[0084] Furthermore, by selecting the data subset with the largest amount of information and the strongest representativeness from the deduplicated corpus as the core data uploaded to the cloud server, the essence can be extracted from the massive amount of data, and the data scale can be further compressed while retaining key semantic information.

[0085] First, a lightweight and efficient k-means++ clustering algorithm is used to cluster the input dialogue list. The initial cluster centers are selected through a probabilistic long-range initialization strategy, which effectively avoids the defect of the traditional k-means algorithm being sensitive to the initial value and ensures that high-quality initial centers are obtained.

[0086] The k-means++ clustering algorithm is an improved version of the classic k-means clustering algorithm, with the main optimization being the selection of initial cluster centers.

[0087] The k-means++ clustering algorithm selects initial centroids using a random strategy where the farther away the centroid is, the higher the probability of it being selected. This makes the initial centroids as dispersed as possible throughout the data space, effectively avoiding the problem that the standard k-means algorithm may converge to a local optimum, thus achieving a more stable and better clustering effect.

[0088] Through iterative optimization, sentence vectors with similar semantics are grouped into the same cluster. After the clustering is completed, the most representative dialogues are selected from each cluster. The selection criterion is the dialogue sample point closest to the cluster center. These dialogue samples can best represent the semantic features of the cluster, ensuring that each semantic pattern has a corresponding typical representative selected into the selected subset.

[0089] Representative dialogues selected from each cluster are aggregated to form the final selected subset. The initial center spacing is optimized using the k-means++ principle to ensure that the cluster distribution is balanced and the coverage is extensive. This ensures that the final selected subset retains the semantic diversity and typical pattern features of the original data to the greatest extent possible, while the amount of data is much smaller than the original data size.

[0090] The selected subset obtained through the above clustering not only eliminates semantic redundancy but also ensures the complete preservation of key information, providing the most valuable data foundation for subsequent model training.

[0091] For example, in the process of clustering the dialogue list, the unique dialogue list after deduplication is first traversed, and the sentence vector representation corresponding to each dialogue unit is extracted. Floating-point vector data is converted into double-precision numerical format and encapsulated into a specific data structure required by the clustering algorithm, so that each dialogue is mathematically represented as a coordinate point in a high-dimensional space; During the clustering initialization phase, the optimal number of clusters is automatically determined based on the total number of data points. An empirical formula is used to calculate the number of clusters, and by rounding up the square root of the data volume, effective data compression is achieved while maintaining representativeness, thus achieving the best balance between data diversity and processing efficiency. Next, the improved k-means++ clustering algorithm is executed. Through an intelligent initial centroid selection strategy, it effectively avoids the problem that traditional clustering methods are prone to getting trapped in local optima. Within the set maximum number of iterations, the k-means++ clustering algorithm automatically groups semantically similar sentence vectors into the same cluster, forming multiple dialogue groups with similar semantic features. Each cluster contains a centroid representing the center position of the group, as well as all dialogue vector points belonging to that group.

[0092] In the representative sample selection stage, each generated cluster is processed in detail. By calculating the Euclidean distance from each data point in the cluster to the centroid of the cluster, the best sample that best represents the semantic features of the cluster can be accurately identified. Distance calculations employ the standard Euclidean distance formula to ensure the accuracy and consistency of the evaluation. The system continuously updates the current best representative sample while traversing all points in each cluster by maintaining a minimum distance tracking variable. After traversing all clusters, the selected best vector points are back-located to the corresponding original dialogue data through the established index mapping relationship.

[0093] The carefully selected dialogue samples were compiled into a final, high-quality dataset. This dataset significantly reduced the size of the original data while fully preserving the core semantic patterns and diverse features of the original dialogue data. This optimized dataset provided high-quality input material for subsequent data uploads and model training, ensuring a significant improvement in cloud processing efficiency.

[0094] S4: Perform privacy protection processing on the selected subset to obtain the distillation dataset.

[0095] In some embodiments, such as Figure 3 As shown, methods for obtaining the distillation dataset by performing privacy protection processing on a selected subset include: Determine whether the number of dialogue samples in the selected subset meets the preset dynamic sample quantity; If the number of dialogue samples in the selected subset meets the preset dynamic sample count, then the dialogue samples in the selected subset will undergo privacy protection processing to obtain the distilled dataset.

[0096] By introducing a mechanism to determine the number of selected subset samples, dynamic triggering conditions are added to the privacy protection process, which can intelligently determine the subsequent operation path based on the actual amount of data obtained from distillation.

[0097] When the amount of data is sufficient, protective processing is performed directly, thereby optimizing the processing flow while satisfying the data utility, avoiding unnecessary waste of computing resources, enhancing the system's adaptability and processing efficiency in different interaction scenarios, and ensuring the flexibility and robustness of the data distillation process.

[0098] In some embodiments, the method for obtaining a distillation dataset by performing privacy protection processing on a selected subset further includes: Determine whether the number of dialogue samples in the selected subset meets the preset dynamic sample quantity; If the number of dialogue samples in the selected subset does not meet the preset dynamic sample quantity, then pseudo samples are generated using the preset local model. The pseudo-samples are combined with the selected subset to obtain a merged dataset. The dialogue samples in the merged dataset are then subjected to privacy protection processing to obtain a distilled dataset.

[0099] Furthermore, pseudo-samples refer to synthetic samples generated by algorithms rather than being directly collected from the original data. In knowledge distillation or data augmentation, pre-trained teacher model generators are often used to produce pseudo-samples that are similar to the distribution of real data for student model training.

[0100] By initiating a supplementary mechanism that generates pseudo-samples using the local model when the number of samples is insufficient, the problem of insufficient or undiversified cloud-based model training data, which may be caused by an initially small selected subset, is effectively solved.

[0101] By generating pseudo-samples that are semantically similar to the original data locally, the scale and diversity of the dataset can be safely expanded while protecting privacy. This ensures that the data uploaded to the cloud always has sufficient scale and information to support effective model training, thereby improving the reliability and output quality of the entire distillation system.

[0102] In some embodiments, the method for generating pseudo-samples using a preset local model further includes: Acquire user question data and intelligent response data from a selected subset of dialogue samples; The user question data is segmented into words to identify keywords in the user question data; Select synonyms from a pre-defined thesaurus to replace keywords, and obtain the replaced user question data; By combining the replaced user question data with the intelligent response data, enhanced data is obtained; Add augmented data to a selected subset to obtain an augmented dataset, and use a local model to generate pseudo-samples based on dialogue samples in the augmented dataset.

[0103] By performing synonym substitution on keywords in user questions, new questions with similar sentence structures but different wording can be created while basically maintaining the semantics of the original sentences. This method has low computational cost and stable results, effectively increasing the language diversity and vocabulary coverage of the dataset, and providing a low-cost and high-quality data source for the model to learn more robust language patterns.

[0104] In some embodiments, the method for generating pseudo-samples using a preset local model includes: Dialogue samples from a selected subset or augmented dataset are input into the local model. The local model performs semantic encoding on the dialogue samples and generates new dialogues that are semantically similar to the dialogue samples as pseudo-samples.

[0105] By leveraging local models for semantic encoding to create new dialogues, it is possible to gain a deeper understanding of the underlying semantics of the original dialogue and generate new dialogues that differ not only in vocabulary and sentence structure but also in core meaning. Figure 1 A new dialogue.

[0106] The data generated by the above methods is richer and more natural in terms of language variation, and can better simulate the various expressions of real users, thereby providing the cloud server with high-quality training samples that are closer to the real data distribution and significantly improving its generalization ability.

[0107] In some embodiments, the method for obtaining a merged dataset by combining pseudo-samples and a selected subset further includes: Input selected subsets or augmented datasets of dialogue samples into the pre-defined teacher model; The teacher model is used to infer user question data in the dialogue samples and generate a predicted probability distribution as a soft label. Soft labels are attached to the corresponding pseudo-samples, and a selected subset is combined to obtain the merged dataset.

[0108] Furthermore, compared to hard labels, soft labels are a form of label that includes a category probability distribution; for example, the probability of a teacher model predicting each category of an input is a soft label.

[0109] The teacher model is a high-performance, lightweight classification model that is pre-trained and deployed on the vehicle. As a carrier of knowledge transfer, the teacher model is used to endow data samples with deep semantic understanding information.

[0110] The teacher model generates a predictive probability distribution containing rich information, namely soft labels, by performing deep semantic analysis on user questions. The soft labels contain the most likely category judgments and fully preserve the teacher model's confidence assessment of other related categories, forming a continuous and fine-grained supervision signal.

[0111] The user question data from a selected subset or augmented dataset is input into the teacher model in batches. After the teacher model performs inference calculations on each question, it outputs the corresponding probability distribution array as a soft label. This can accurately reflect the probability weight of the question in different intent categories, fully record the teacher model's multi-dimensional understanding of the question semantics, and form a supervision signal with more information than traditional hard labels.

[0112] In subsequent processing, the generated soft labels are associated and bound with the corresponding pseudo samples and selected subsets to construct a merged dataset rich in knowledge information. This knowledge-enhanced data will provide high-quality supervision information for the training of the cloud-based student model, significantly improving the efficiency of model training and the final performance.

[0113] By introducing a teacher model and attaching soft labels to the data, simple data augmentation is transformed into a knowledge injection process. The soft labels generated by the teacher model contain a richer probabilistic understanding of the dialogue intent, which is far more informative than a single hard label. Attaching these soft labels to pseudo-samples ensures that the uploaded merged dataset not only contains the data itself but also carries the knowledge of the teacher model. This enables more efficient guidance for the training of models in the cloud, accelerating their convergence and improving their performance, thus achieving knowledge transfer from the vehicle to the cloud.

[0114] Furthermore, by determining whether the number of samples is sufficient, and by expanding the data and adding knowledge annotations based on a selected subset, the quality and scale of the dataset can be further improved.

[0115] When terminal computing resources permit, data augmentation can be performed on selected subsets using pre-built local models.

[0116] We employ a small GPT model to perform semantic learning on the original dialogue corpus, generating new dialogue samples that retain the original meaning but have diverse expressions. We use rule templates to perform data augmentation operations on the existing text, such as synonym rewriting, word replacement, or spelling perturbation, thereby effectively expanding the data scale and enhancing data diversity.

[0117] Small GPT (Generative Pre-trained Transformer) models are generative pre-trained transformers. "Generative" means the model can generate new, coherent text, rather than simply classifying or analyzing it. "Pre-trained" means the model is first trained on massive amounts of unlabeled text data to learn general language knowledge and rules. "Transformer" refers to the model's core architecture, which is based on the Transformer model and uses a self-attention mechanism to understand contextual relationships, making it suitable for processing sequential data. Small GPT models refer to GPT models with fewer parameters that have been compressed or pruned to run on devices with limited computing power.

[0118] The teacher model deployed on the vehicle is used to intelligently label data samples, and the teacher model is used to perform reasoning analysis on selected data and generated pseudo samples. Generate the corresponding soft label probability distribution, which includes the model's confidence in judging various intentions, and attach these soft labels rich in model knowledge as supervision signals to the corresponding samples; By leveraging the synergy of data augmentation and knowledge annotation, the scale and diversity of training data are increased, injecting rich semantic supervision information into the samples and forming a high-quality augmented dataset, providing a reliable data foundation for the accurate training of models in the cloud.

[0119] For example, in the process of data augmentation, an empty list is first created to store the augmented dialogue data. Then, each dialogue unit in the selected subset is traversed, and the user questions contained therein are semantically augmented. By calling a specially designed synonym replacement tool, keywords in the questions are automatically identified, and suitable synonyms are randomly selected from a pre-set thesaurus for replacement, thereby generating new questions with the same semantics but different expressions. For the intelligent response section in each dialogue unit, its original content is usually kept unchanged, but in some implementations it can be rewritten in a lightweight manner as needed. Based on the processed user questions and intelligent responses, new dialogue unit instances are created, which fully preserve the semantic structure of the original dialogue while achieving diversification in language expression. All newly generated augmented dialogue units are added to a dedicated augmented data list. After the entire selected subset has been traversed and processed, the original selected subset is merged with the augmented dataset to form a larger-scale augmented dataset. Augmented datasets can maintain the semantic integrity of the original data, effectively improve the diversity and richness of the data, and provide more sufficient and diverse learning samples for subsequent model training; Data augmentation mechanisms can significantly improve the scale and quality of data resources available for training without introducing additional annotation costs.

[0120] In some embodiments, the method for obtaining a distillation dataset by performing privacy protection processing on a selected subset further includes: Regular expressions are used to de-identify dialogue samples in a selected subset or merged dataset, and zero-mean Gaussian noise is added to the sentence vectors in the dialogue samples to obtain the distilled dataset.

[0121] By combining two complementary privacy protection techniques, text desensitization and vector noise enhancement, a dual protection mechanism is constructed. Regular expression desensitization directly removes sensitive entity information from the text, thus solving the risk of direct identification.

[0122] Adding Gaussian noise to sentence vectors disrupts the original feature distribution of the data, increases the difficulty of inferring the original content through reverse engineering, solves the risk of indirect inference, and provides comprehensive and reliable security for user privacy while preserving the semantic utility of the data for training to the maximum extent.

[0123] Furthermore, by implementing privacy protection processing on the data to be uploaded, we ensure that the data complies with privacy and security regulations.

[0124] First, directly personal identification information in the selected subset or merged dataset is identified and removed, including but not limited to sensitive information such as user names, vehicle identification numbers, and contact information; by performing pattern recognition and content replacement on the dialogue text, identifiers that can be directly associated with individuals or devices are eliminated, thereby achieving data anonymization. After completing the basic desensitization, the processed dialogue text undergoes an irreversible transformation to ensure that the original text content cannot be reversed and restored through uploaded data, thus eliminating the risk of privacy leakage at the source.

[0125] It also employs a temperature noise injection mechanism, which effectively obscures individual data features by adding random noise that meets differential privacy requirements to the data features, thereby enhancing the strength of privacy protection while ensuring the overall statistical utility of the data. Through the above multi-layered protection mechanism, the uploaded data is ensured to meet the compliance requirements for privacy protection while retaining its training value, thus providing reliable security for the entire data distillation process.

[0126] For example, in the process of implementing privacy protection processing for dialogue samples, a cryptographically secure random number generator is first initialized to ensure that the random noise generated subsequently has a high degree of unpredictability, providing a solid security foundation for data protection.

[0127] After the processing flow begins, the merged dataset to be processed is traversed, and dual privacy protection operations are performed on each dialogue unit in it; The first layer of protection involves direct desensitization of text content. Through predefined regular expression patterns, it automatically identifies and replaces mobile phone numbers in the form of 11 consecutive digits in user questions, and identifies and replaces vehicle identification numbers consisting of 17 uppercase letters and numbers in smart replies. All identified sensitive information is replaced by a unified specific label, thereby effectively eliminating direct personal identification information.

[0128] The second layer of protection is noise perturbation processing for semantic vectors. By traversing all dimensions of the sentence vector corresponding to each dialogue unit, random noise conforming to a Gaussian distribution is generated for each dimension. By controlling the standard deviation of the noise within a finely set range, it is ensured that while effectively perturbing the data features, the integrity of the semantic information of the vector itself will not be significantly affected. The generated random noise is superimposed on each dimension of the original vector to achieve micro-perturbation of the semantic features.

[0129] The data, after being processed by a dual protection mechanism, eliminates directly identifiable sensitive information. Noise injection reduces the risk of inferring the original content from vector features. The final output dataset maintains sufficient semantic utility while fully protecting user privacy, providing safe and reliable data support for subsequent cloud model training and ensuring the security and compliance of the data throughout its lifecycle.

[0130] The above method adds tiny random noise to each dimension of the sentence vector, so that each sentence vector undergoes a slight and unpredictable perturbation. Even if the same sentence is processed at different times, the resulting vector will be slightly different. This effectively prevents malicious parties from accurately reconstructing the original text content by comparing vector libraries or performing model reverse engineering attacks, providing a second line of defense for privacy and security.

[0131] S5: Upload the distillation dataset to the cloud server.

[0132] In existing technologies, the interaction data of smart cockpits is usually uploaded directly and without processing to the cloud for processing. This approach leads to a huge risk of privacy leakage for the user's sensitive raw dialogue data during transmission and cloud storage. At the same time, the massive amount of raw data also consumes a lot of network bandwidth and puts a heavy storage and computing burden on cloud servers.

[0133] This application, through the aforementioned solution, completes the entire process of original dialogue data processing locally on the vehicle terminal, from vectorization, deduplication, clustering and selection to privacy protection. Ultimately, only the distilled essence data is uploaded. This fundamental change avoids the direct transmission of original user data from the data source, thereby fundamentally and significantly reducing the risk of privacy leakage and significantly reducing network bandwidth consumption and cloud server pressure.

[0134] Furthermore, this application constructs a complete and collaborative edge data processing pipeline. Compared with the existing architecture that treats the terminal as a simple data collection node and centralizes all intelligent processing in the cloud, this application can make full use of the fragmented computing power of the vehicle terminal to achieve distributed collaboration in data processing. Thus, in the resource-constrained vehicle environment, it can systematically achieve an optimized balance between privacy and security, communication efficiency and computing load.

[0135] In some embodiments, the distillation dataset is encapsulated in a standardized JSON format and prepared for upload; JSON (JavaScript Object Notation) is a lightweight data interchange format that organizes data using key-value pairs. It has a clear structure, distinct hierarchy, and is easy to read and write, as well as easy for machines to parse and generate.

[0136] For example, the code for encapsulating a distillation dataset is represented as follows: {"dialog_id": "veh1234_1609459200_001", "context": ["Navigate to Sanlitun, Beijing", "Okay, route planned for you"], "card": {"type": "nav_route", "destination": "a landmark in Beijing", "eta":15}, "source": "onboard_distilled", "weight": 1.0}; Among them, "dialog_id" is a unique dialog identifier, which consists of vehicle number, timestamp and serial number to ensure data traceability; “context” is an array of dialogue contexts, which stores text pairs of user queries and system responses in order; “card” is structured response data, which includes a type identifier and a set of parameters; the set of parameters includes the destination, estimated arrival time, etc. “source” is the data source identifier, clearly marked as end-side distillation onboard_distilled, which facilitates cloud-based identification of the processing pipeline; "weight" is the sample weight parameter, which provides a reference for the importance of samples during cloud training.

[0137] Ultimately, the destination field has been obfuscated and, through standardized encapsulation, the complete semantic context of the dialogue and the system response logic have been preserved, ensuring the proper handling of sensitive information and ultimately forming a standardized data package suitable for cloud model training that combines data value and privacy security.

[0138] This structured data organization facilitates cloud-based parsing and batch processing, providing high-quality data input for subsequent model training.

[0139] like Figure 4 As shown, this application also provides a vehicle. The vehicle includes: The in-vehicle terminal is used to acquire raw dialogue data generated within the smart cockpit. The processor, which communicates with the vehicle-mounted terminal; The processor is configured to execute the aforementioned data processing method for the smart cockpit.

[0140] In some embodiments, the processor is further configured to: receive raw dialogue data, convert the raw dialogue data into sentence vectors, and deduplicate the sentence vectors to obtain a dialogue list; Cluster the sentence vectors in the dialogue list, group the sentence vectors with similar semantics into the same cluster, select and collect the dialogue samples corresponding to the sentence vectors that are closest to the centroid of the cluster in each cluster, and obtain a selected subset; A selected subset is processed for privacy protection to obtain the distillation dataset; Upload the distillation dataset to the cloud server.

[0141] By limiting the communication connection between the processor and the vehicle terminal and configuring a specific model, the real-time performance and stability of the data processing process are guaranteed. This meets the stringent requirements of the vehicle for system reliability and response speed in the actual driving environment, ensuring that the entire data distillation process can operate independently, efficiently and safely in the vehicle.

[0142] In some embodiments, the processor is further configured to: The original dialogue data is converted into sentence vectors using a pre-trained semantic similarity model; The signatures of sentence vectors are calculated using a text similarity algorithm. The Hamming distances of the signatures are compared, and signatures with Hamming distances within a threshold are removed to obtain a list of dialogues.

[0143] In some embodiments, the processor is further configured to: Based on the dialogue list, a clustering algorithm is used to cluster the sentence vectors in the dialogue list, and sentence vectors with similar semantics are grouped into the same cluster; Calculate the distance from each sentence vector in each cluster to the centroid of the cluster, and select the dialogue sample corresponding to the sentence vector that is closest to the centroid of the cluster; The dialogue sample is used as the representative dialogue of this cluster. The representative dialogues of all clusters are collected to obtain a selected subset.

[0144] In some embodiments, the processor is further configured to: Determine whether the number of dialogue samples in the selected subset meets the preset dynamic sample quantity; If the number of dialogue samples in the selected subset meets the preset dynamic sample count, then the dialogue samples in the selected subset will undergo privacy protection processing to obtain the distilled dataset.

[0145] In some embodiments, the processor is further configured to: Determine whether the number of dialogue samples in the selected subset meets the preset dynamic sample quantity; If the number of dialogue samples in the selected subset does not meet the preset dynamic sample quantity, then pseudo samples are generated using the preset local model. The pseudo-samples are combined with the selected subset to obtain a merged dataset. The dialogue samples in the merged dataset are then subjected to privacy protection processing to obtain a distilled dataset.

[0146] In some embodiments, the processor is further configured to: Acquire user question data and intelligent response data from a selected subset of dialogue samples; The user question data is segmented into words to identify keywords in the user question data; Select synonyms from a pre-defined thesaurus to replace keywords, and obtain the replaced user question data; By combining the replaced user question data with the intelligent response data, enhanced data is obtained; Augmented data is added to the selected subset to obtain an augmented dataset, and pseudo-samples are generated using the local model based on the dialogue samples in the augmented dataset.

[0147] In some embodiments, the processor is further configured to: Dialogue samples from a selected subset or augmented dataset are input into the local model. The local model performs semantic encoding on the dialogue samples and generates new dialogues that are semantically similar to the dialogue samples as pseudo-samples.

[0148] In some embodiments, the processor is further configured to: Input the selected subset or dialogue samples from the augmented dataset into the preset teacher model; The teacher model is used to infer user question data in the dialogue samples and generate a predicted probability distribution as a soft label. Soft labels are attached to the corresponding pseudo-samples, and a selected subset is combined to obtain the merged dataset.

[0149] In some embodiments, the processor is further configured to: Regular expressions are used to de-identify dialogue samples in a selected subset or merged dataset, and zero-mean Gaussian noise is added to the sentence vectors in the dialogue samples to obtain the distilled dataset.

[0150] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A data processing method for an intelligent cockpit, characterized in that, Applications to vehicle-mounted terminals include: Acquire raw dialogue data generated within the smart cockpit; The original dialogue data is converted into sentence vectors, and duplicate sentence vectors are removed to obtain a dialogue list; Cluster the sentence vectors in the dialogue list, group the sentence vectors with similar semantics into the same cluster, select and collect the dialogue samples corresponding to the sentence vectors that are closest to the centroid of the cluster in each cluster, and obtain a selected subset; The selected subset is then subjected to privacy protection processing to obtain the distillation dataset; The distillation dataset was uploaded to a cloud server.

2. The data processing method for the intelligent cockpit according to claim 1, characterized in that, The method for converting the original dialogue data into sentence vectors and then deduplicating the sentence vectors to obtain a dialogue list includes: The original dialogue data is converted into sentence vectors using a pre-trained semantic similarity model; The signature of the sentence vector is calculated using a text similarity algorithm. The Hamming distance of the signatures is compared, and signatures whose Hamming distance is within a threshold are removed to obtain the dialogue list.

3. The data processing method for the intelligent cockpit according to claim 1, characterized in that, Methods for clustering sentence vectors in the dialogue list include: Based on the dialogue list, a clustering algorithm is used to cluster the sentence vectors in the dialogue list, and the sentence vectors with similar semantics are grouped into the same cluster; Calculate the distance from each sentence vector in each cluster to the centroid of the cluster, and select the dialogue sample corresponding to the sentence vector whose distance is closest to the centroid of the cluster; The dialogue sample is used as the representative dialogue of this cluster. The representative dialogues of all clusters are then collected to obtain the selected subset.

4. The data processing method for an intelligent cockpit according to claim 1, characterized in that, Methods for performing privacy protection processing on the selected subset to obtain the distillation dataset include: Determine whether the number of dialogue samples in the selected subset meets the preset dynamic sample quantity; If the number of dialogue samples in the selected subset meets the preset dynamic sample quantity, then the dialogue samples in the selected subset will be subjected to privacy protection processing to obtain the distillation dataset.

5. The data processing method for the intelligent cockpit according to claim 1 or 4, characterized in that, The method for obtaining the distillation dataset by performing privacy protection processing on the selected subset also includes: Determine whether the number of dialogue samples in the selected subset meets the preset dynamic sample quantity; If the number of dialogue samples in the selected subset does not meet the preset dynamic sample quantity, then pseudo samples are generated using the preset local model. The pseudo-samples and the selected subset are combined to obtain a merged dataset. The dialogue samples in the merged dataset are then subjected to privacy protection processing to obtain the distillation dataset.

6. The data processing method for an intelligent cockpit according to claim 5, characterized in that, Methods for generating pseudo-samples using a pre-defined local model also include: Obtain user question data and intelligent response data from the dialogue samples in the selected subset; The user question data is segmented into words to identify keywords in the user question data; Synonyms are selected from a preset thesaurus to replace the keywords, resulting in replaced user question data; By combining the replaced user question data with the intelligent response data, enhanced data is obtained; The augmented data is added to the selected subset to obtain the augmented dataset, and the local model is used to generate pseudo-samples based on the dialogue samples in the augmented dataset.

7. The data processing method for an intelligent cockpit according to claim 6, characterized in that, Methods for generating pseudo-samples using a pre-defined local model include: The dialogue samples from the selected subset or the augmented dataset are input into the local model, and the local model is used to perform semantic encoding on the dialogue samples to generate new dialogues that are semantically similar to the dialogue samples as pseudo-samples.

8. The data processing method for an intelligent cockpit according to claim 6, characterized in that, The method for obtaining the merged dataset by combining the pseudo-samples and the selected subset also includes: Input the selected subset or the dialogue samples in the enhanced dataset into the preset teacher model; The teacher model is used to infer user question data in the dialogue samples and generate a predicted probability distribution as a soft label. The soft labels are attached to the corresponding pseudo-samples, and the merged dataset is obtained by combining the selected subset.

9. The data processing method for an intelligent cockpit according to claim 5, characterized in that, The method for obtaining the distillation dataset by performing privacy protection processing on the selected subset also includes: The dialogue samples in the selected subset or the merged dataset are anonymized using regular expressions, and zero-mean Gaussian noise is added to the sentence vectors in the dialogue samples to obtain the distillation dataset.

10. A vehicle, characterized in that, include: The in-vehicle terminal is used to acquire raw dialogue data generated within the smart cockpit. A processor that is communicatively connected to the vehicle-mounted terminal; The processor is configured to perform the data processing method for the smart cockpit as described in any one of claims 1 to 9.

Images