Car product market data analysis system based on AI agent and big data

Abstract

The application discloses an automobile product market data analysis system based on an AI intelligent agent and big data, and relates to the technical field of automobile market data analysis.The system comprises a data management module, a distributed computing module and an AI decision module.The data management module is used for receiving and storing automobile market multi-modal original data from multiple heterogeneous data sources, and performing the operations of de-duplication, outlier elimination and missing value filling on the multi-modal original data to generate a standardized basic data set.The distributed computing module is used for performing distributed iterative calculation of the standardized basic data set in memory to generate a high-dimensional analysis data set containing product correlation and time series features.The AI decision module is used for receiving a user-input natural language analysis intent, and performing dependency syntax analysis and named entity joint identification on the natural language analysis intent according to the high-dimensional analysis data set to generate a structured analysis task queue containing multiple atomic analysis operations with dependency relationships.

Description

Technical Field

[0001] This invention relates to the field of automotive market data analysis technology, specifically to an automotive product market data analysis system based on AI agents and big data. Background Technology

[0002] The automotive market data analysis field has long faced challenges in integrating and processing multi-source heterogeneous data and incurring high barriers to interactive analysis. Existing automotive market data analysis systems typically rely on centralized data warehouses and pre-defined fixed report templates. Multimodal data collected from various channels, such as text reviews, configuration parameters, and sales records, requires significant manual intervention for cleaning and format alignment, which is time-consuming and prone to introducing subjective bias. In the data processing stage, traditional solutions often use single-machine batch processing scripts or simple rules for outlier detection, failing to automatically perceive the contextual relationships between fields within massive amounts of records. This leads to the direct use of global mean or fixed default values ​​when filling missing values, ignoring the distribution characteristics within different data domains. Consequently, the constructed analysis dataset suffers from significant information loss in reflecting implicit relationships between products and dynamic market evolution patterns. In the task construction stage, market analysts must transform business problems into precise query statements or operation sequences. This process requires users to possess both domain knowledge and technical operational capabilities, making it difficult for non-technical users to directly initiate complex multidimensional cross-analysis. Furthermore, during task execution, scheduling mechanisms typically rely solely on first-in-first-out queues or round-robin allocation of node resources, neglecting the data flow dependencies between atomic operations. This leads to frequent cross-node transmission of intermediate results, generating additional network overhead and computational latency. The final result presentation depends on manually stitching together charts, failing to automatically organize the visualization layout based on the hierarchical semantics of the data dimensions. Therefore, how to construct a high-quality analytical foundation from raw multimodal data capable of capturing fine-grained product relationships and dynamic temporal characteristics, and how to automatically decompose dependent atomic analysis processes from unstructured natural language expressions, have become pressing issues in this technical field. Summary of the Invention

[0003] This invention provides an automotive product market data analysis system based on AI intelligent agents and big data, which solves the problems of existing technologies that make it difficult to automatically construct an analysis dataset containing product associations and time-series dynamic features from multimodal raw data, and that it is impossible to automatically deconstruct natural language analysis intentions into adaptively schedulable atomic analysis tasks.

[0004] To achieve the above objectives, the present invention provides the following technical solution: The present invention relates to an automotive product market data analysis system based on AI intelligent agents and big data, the system comprising a data management module, a distributed computing module, an AI decision-making module, a task scheduling module, and a report generation module.

[0005] The data management module receives and stores multimodal raw data of the automotive market from multiple heterogeneous data sources, and performs deduplication, outlier removal, and missing value imputation operations on the multimodal raw data to generate a standardized basic dataset. Specifically, the data management module calculates a data integrity score for each record in the multimodal raw data according to a preset data quality rule base, marks records with data integrity scores below a preset threshold as records to be cleaned, performs field-level anomaly detection on the records to be cleaned, identifies outlier records in numerical fields and illegal value records in categorical fields, deletes the outlier records and writes the deletion operation log to a data traceability table. For missing fields in the remaining records after the deletion operation, the data management module retrieves context records related to the missing field from the multimodal raw data according to the data domain to which the field belongs, and imputes the missing field based on the mode or median value of the same field in the context records to generate the standardized basic dataset. Through the above processing, the present invention can automatically identify and correct various quality problems in the raw data from different channels, ensuring the consistency and reliability of the data on which subsequent analysis depends.

[0006] The distributed computing module performs in-memory distributed iterative computation on the standardized basic dataset to generate a high-dimensional analysis dataset containing product relationships and time-series features. In a preferred embodiment, the distributed computing module partitions the standardized basic dataset according to product identifiers and loads each partition into the memory space of the corresponding computing node, constructing an in-memory data partition. Records in each in-memory data partition are sorted by timestamps to generate time-series data blocks. Sliding window aggregation computation is then performed on these time-series data blocks to generate sales change rate features and market share change rate features for each product within multiple preset time windows. Simultaneously, the distributed computing module calculates the configuration similarity between any two products based on the product configuration parameters in the standardized basic dataset and constructs a product similarity matrix. The sales change rate feature, market share change rate feature, and product similarity matrix are then concatenated to generate the high-dimensional analysis dataset. This processing method organically integrates the dynamic time-series features of the products themselves with the static relationship features between products, providing a data foundation with extremely high information density for upper-level intelligent analysis.

[0007] The AI ​​decision module receives the natural language analysis intent input by the user and, based on the high-dimensional analysis dataset, performs dependency parsing and named entity recognition on the natural language analysis intent to generate a structured analysis task queue. This queue contains multiple atomic analysis operations with dependencies. In a specific implementation, the AI ​​decision module performs word segmentation and part-of-speech tagging on the natural language analysis intent to generate a lexical analysis sequence. It then performs dependency parsing on the lexical analysis sequence to identify the core predicate verb in the natural language analysis intent and the corresponding subject and object dependencies, generating a syntactic dependency tree. Subsequently, named entity recognition is performed on the nominal nodes in the syntactic dependency tree to extract time entities, location entities, product model entities, and indicator name entities from the nominal nodes. Based on the semantic category of the core predicate verb, the entity type in the subject subordination, and the entity type in the object subordination, the AI ​​decision module maps the natural language analysis intent into a structured query description containing analysis operation type, analysis dimension, and analysis constraints. It then performs operation decomposition on the structured query description, breaking it down into multiple atomic analysis operations with data flow dependencies. Based on the data flow dependencies, it sets execution priorities for the atomic analysis operations, generating the structured analysis task queue. Preferably, when performing dependency parsing on the lexical analysis sequence, a dependency classifier based on a graph neural network is used. Each word in the lexical analysis sequence is treated as a node, and the adjacency relationships between words are treated as edges. The module iteratively aggregates contextual semantic features and outputs the probability value of a dependency relationship between any two words. The dependency relationship with the highest probability value is taken as the result of the dependency parsing. This mechanism enables the system to accurately understand the complex analytical requests expressed by the user in natural language and automatically convert them into a sequence of machine-executable analysis instructions.

[0008] Furthermore, during the operation decomposition process, the AI ​​decision module retrieves an abstract analysis template from a preset analysis operation template library that matches the analysis operation type in the structured query description. This abstract analysis template includes input data slots, output data slots, and data processing logic placeholders. Based on the analysis dimensions and constraints, the input data slots are filled with parameters to generate an analysis unit to be executed. The data reference relationships between these units are identified, and the output data slot of the first analysis unit is matched with the input data slot of the second analysis unit. Based on these data reference relationships, the analysis units are organized into a directed acyclic graph (DAG) structure. Each node in the DAG is considered an atomic analysis operation, and the directed edges in the DAG are considered data flow dependencies. Through templated task decomposition and dependency construction, the controllable decomposition and orderly execution of complex analysis tasks are ensured.

[0009] The task scheduling module receives the structured analysis task queue and, based on the current cluster resource load status, allocates computing nodes to the atomic analysis operations, driving the distributed computing module to execute the atomic analysis operations and obtain an intermediate analysis result set. Specifically, the task scheduling module obtains the current CPU utilization and current memory utilization of each computing node in the distributed computing module, calculates the comprehensive load score of each computing node accordingly, and sorts the computing nodes in ascending order of the comprehensive load score to generate a sequence of available computing nodes. Based on the estimated computing resources required for the atomic analysis operations and the data flow dependencies in the structured analysis task queue, the module selects a target computing node from the sequence of available computing nodes for each atomic analysis operation, sends the data processing logic placeholder and the specific value of the input data slot corresponding to the atomic analysis operation to the target computing node, triggers the execution process on the target computing node, receives the intermediate calculation results returned by the execution process, caches all the intermediate calculation results according to the execution order of the atomic analysis operations, and generates the intermediate analysis result set.

[0010] In a preferred embodiment of the present invention, when selecting a target computing node, the task scheduling module calculates the estimated value of the required memory resources based on the input data size of the atomic analysis operation, calculates the estimated value of the required CPU resources based on the complexity of the data processing logic placeholders of the atomic analysis operation, and traverses the available computing node sequence to select computing nodes whose current remaining memory resources are greater than the estimated value of memory resources and whose current remaining CPU resources are greater than the estimated value of CPU resources as a set of candidate nodes. Simultaneously, based on the data flow dependencies in the structured analysis task queue, a first upstream atomic analysis operation with an upstream dependency relationship to the current atomic analysis operation is identified, and the first computing node containing the first upstream atomic analysis operation is designated as a priority node. If the priority node exists in the set of candidate nodes, then the priority node is determined as the target computing node for the current atomic analysis operation. This resource-aware scheduling strategy, which also considers data locality, effectively reduces data transmission overhead and improves the overall throughput of distributed computing.

[0011] The report generation module performs multi-source result association and aggregation on the intermediate analysis result set and calls the visualization component library to generate a user-oriented automotive product market data analysis dashboard. Specifically, the report generation module obtains each intermediate calculation result and its corresponding atomic analysis operation identifier in the intermediate analysis result set, and extracts the corresponding analysis dimension label and analysis constraint label from the structured analysis task queue according to the atomic analysis operation identifier. Based on the analysis dimension label, multiple intermediate calculation results with the same analysis dimension are merged into the same data group. A time-series alignment operation is performed on multiple intermediate calculation results within the same data group, resampling the intermediate calculation results according to the same time granularity to generate a dataset to be displayed. Based on the data type of the dataset to be displayed, a matching chart template is selected from the visualization component library, and the fields in the dataset to be displayed are mapped to the data channels of the chart template.

[0012] Preferably, when generating the dashboard, the report generation module extracts all the analysis dimension labels corresponding to the dataset to be displayed, constructs a dimension hierarchy tree based on the inclusion relationship between the analysis dimension labels, where the root node of the dimension hierarchy tree represents the entire data range and the child nodes represent progressively subdivided analysis dimensions. A depth-first traversal is performed on the dimension hierarchy tree, attaching the chart templates with completed field mappings to the corresponding nodes of the dimension hierarchy tree. Different display area sizes are assigned to the chart templates attached to nodes at different levels based on the depth of the dimension hierarchy tree, with deeper levels corresponding to smaller display area sizes. Following the traversal order of the dimension hierarchy tree, all the chart templates attached to the dimension hierarchy tree are filled into the corresponding display areas, generating the automotive product market data analysis dashboard. Through the above processing, the originally discrete intermediate analysis results are integrated into a unified data view with inherent logical connections, and presented intuitively in a way that conforms to human cognitive levels. This allows market analysts to penetrate layer by layer from macro to micro, quickly gaining insight into the competitive landscape and changing trends of the automotive product market.

[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows: When performing deduplication, outlier removal, and missing value imputation on multimodal raw data, the data management module calculates the integrity score of each record based on a preset data quality rule base. Records that fail to meet the score undergo field-level anomaly detection, distinguishing between numerical outliers and categorical illegal values, and deleting them. The entire deletion process is logged in the data traceability table. For missing fields in the remaining records, relevant context records are retrieved from the raw data based on their data domain, and the missing values ​​are imputed using the mode or median value of the same field within that context. This approach abandons the practice of uniform imputation across the entire table, ensuring that the imputed values ​​conform to the true distribution of the corresponding data domain. The generated standardized basic dataset retains the local statistical characteristics of the original data while eliminating the distribution bias introduced by cross-domain imputation. The distributed computing module loads the standardized dataset into the memory of each computing node, partitioned by product identifier. It then performs sliding window aggregation on the time-series data blocks generated by timestamp sorting within each partition, obtaining sales change rate and market share change rate features across multiple time windows. Based on product configuration parameters, it calculates the configuration similarity between any two products, constructing a product similarity matrix. Finally, it concatenates the time-series features with the similarity matrix to create a high-dimensional analysis dataset. This dataset simultaneously encodes horizontal competitor relationships and vertical market evolution trends, enabling subsequent analysis to simultaneously perceive the configuration substitution relationships between products and the periodic fluctuations of market rhythm. The AI ​​decision-making module performs word segmentation, part-of-speech tagging, and dependency parsing on the user's input natural language analysis intent. It identifies the core predicate verb and its subject-object subordination relationships, then extracts named entities such as time, region, product model, and indicator name from noun nodes. Based on the predicate semantic category and argument entity type, it maps the analysis intent into a structured query description containing operation types, dimensions, and constraints. Next, by retrieving abstract analysis templates from a pre-defined template library, it decomposes the query description into atomic analysis operations with data flow dependencies, organizing them into a structured analysis task queue in the form of a directed acyclic graph. This solution allows users to express analysis needs at any level without needing to master query language or operation scripts. The automatically decomposed atomic operation granularity provides a precise task orchestration basis for subsequent fine-grained resource scheduling. When allocating computing nodes for atomic analysis operations, the task scheduling module estimates the required memory and CPU resources based on the size of the input data and the complexity of the data processing logic. It then iterates through the available computing nodes sorted by comprehensive load score to select candidate nodes that meet the resource requirements. Simultaneously, it identifies upstream dependencies of the current operation and prioritizes assigning tasks to computing nodes that already possess upstream intermediate results. Data processing logic placeholders and the specific values ​​of the input data slots are directly sent to the target node to trigger execution. This data-local scheduling strategy reduces unnecessary data migration across nodes and compresses the end-to-end latency of the analysis pipeline.After acquiring intermediate analysis results, the report generation module merges results within the same dimension according to the dimension labels corresponding to the atomic analysis operations and performs time-series resampling alignment. It automatically selects visualization component templates based on data type and maps data channels. Then, it constructs a dimension hierarchy tree based on the inclusion relationships between dimension labels, attaches chart templates to corresponding nodes, and allocates display area sizes according to hierarchy depth, ultimately generating a dashboard. This process automates the organization of multi-source analysis results into multi-granularity visualization dashboards, ensuring consistency in the logical hierarchy and spatial layout of the displayed content. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0015] Figure 1 This is a schematic diagram of the structure of an automotive product market data analysis system based on AI intelligent agents and big data; Figure 2 This is a flowchart of the multimodal automotive market data integrity assessment and cleaning process; Figure 3 This is a flowchart of the distributed computing module generating high-dimensional analysis datasets; Figure 4 This is a flowchart of the AI ​​decision-making module. Figure 5 This is a flowchart for generating a dashboard for automotive product market data analysis. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] See Figure 1This invention provides an automotive product market data analysis system based on AI agents and big data, comprising: a data management module, a distributed computing module, an AI decision-making module, a task scheduling module, and a report generation module. The data management module receives and stores multimodal raw automotive market data from multiple heterogeneous data sources, and performs deduplication, outlier removal, and missing value imputation operations on the multimodal raw data to generate a standardized basic dataset. The distributed computing module performs in-memory distributed iterative computation on the standardized basic dataset to generate a high-dimensional analysis dataset containing product relationships and time-series features. The AI ​​decision-making module receives natural language analysis intent input by the user and, based on the high-dimensional analysis dataset, performs dependency parsing and named entity joint recognition on the natural language analysis intent to generate a structured analysis task queue, which contains multiple atomic analysis operations with dependencies. The task scheduling module receives the structured analysis task queue and, based on the current cluster resource load status, allocates computing nodes to the atomic analysis operations, driving the distributed computing module to execute the atomic analysis operations to obtain a set of intermediate analysis results. The report generation module performs multi-source result association and aggregation on the intermediate analysis result set, and calls the visualization component library to generate a user-oriented automotive product market data analysis dashboard.

[0018] Example 1: In specific implementation, please refer to Figure 2 The data management module receives multimodal raw data from multiple heterogeneous data sources in the automotive market. This multimodal raw data includes structured sales records, semi-structured vehicle configuration documents, and unstructured user review text. The data management module calculates a data integrity score for each record in the multimodal raw data based on a pre-defined data quality rule base. This data quality rule base includes the integrity weight, set of valid value ranges, upper limit for missing value tolerance, and inter-field correlation constraints for each data field. The data integrity score is calculated using the following formula:

[0019] in, This indicates the total number of fields contained in a single record; Indicates the first The integrity weights corresponding to each field are read from the data quality rule base, and the key fields... The value is 0.5, which is important for this field. The value is 0.3, which is typical for most fields. The value is 0.2; Indicates the first The integrity indicator value of each field, when the field contains non-empty and valid data. The value is 1, otherwise The value is 0.

[0020] A data integrity score is calculated for each record. Next, the data integrity score is compared with a preset threshold. The preset threshold, obtained from the data quality rule base, has a value of 0.6, determined based on the minimum tolerance for data integrity in business analysis. Records with data integrity scores below the preset threshold are marked as records to be cleaned. Field-level anomaly detection is performed on these records to identify outlier records in numeric fields and illegal values ​​in categorical fields. For numeric fields, the first quartile is calculated. and the third and fourth quartiles And set the outlier detection interval as ,in For interquartile range (ICM) fields, records whose numerical field values ​​fall outside the outlier detection range are identified as outlier records. For categorical fields, a set of valid value ranges corresponding to the categorical field is extracted from the data quality rule base, and records whose categorical field values ​​are not included in the set of valid value ranges are identified as invalid value records.

[0021] Deletion operations are performed on the identified outlier records and illegal value records. During the deletion process, the identifier, deletion reason, and a snapshot of the fields before deletion for each deleted record are written to the data traceability table. The data traceability table is stored in the persistent storage medium of the data management module for data auditing. For missing fields remaining in the records after the deletion operation, context records are retrieved from the multimodal raw data according to the data domain to which the missing field belongs. The context records refer to other undeleted records that are related to the current record in terms of timestamp, product identifier, or data source. For numerical missing fields, all non-missing values ​​of the same field are extracted from the retrieved context records, the median value of these non-missing values ​​is calculated, and the median value is filled into the missing field. For categorical missing fields, all non-missing values ​​of the same field are extracted from the retrieved context records, the frequency of each value is counted, and the mode value with the highest frequency is filled into the missing field. After all missing fields are filled, a standardized basic dataset is generated, which no longer contains null values, outliers, or illegal values.

[0022] Example 2: In specific implementation, please refer to Figure 3The distributed computing module performs in-memory distributed iterative computation on the standardized basic dataset. The module partitions the standardized basic dataset according to product identifiers, where each product identifier is a field uniquely identifying a product instance in each record. The partitioning process employs a hash partitioning strategy, taking the hash value of the product identifier modulo the total number of computing nodes, and grouping records with the same modulo result into the same partition. After partitioning, each partition is fully loaded into the memory space of the corresponding computing node, constructing an in-memory data partition. The in-memory data partition is organized in a columnar storage format, where the timestamp column, sales volume column, market share column, and product configuration parameter column are stored in contiguous memory address segments.

[0023] The records in each in-memory data partition are sorted by timestamp. The distributed computing module uses an in-memory merge sort algorithm to perform the sorting operation, organizing the sorted records into time-series data blocks in chronological order. One in-memory data partition corresponds to one product's time-series data block. Each record in the time-series data block is appended with the timestamp index of the previous record and the timestamp index of the next record to accelerate subsequent sliding window operations.

[0024] A sliding window aggregation calculation is performed on the time series data blocks. The sliding window aggregation calculation uses a preset time window length and sliding step size. The preset time window length includes three values: 7 days, 30 days, and 90 days, corresponding to weekly, monthly, and quarterly analysis granularities, respectively. The sliding step size is uniformly set to 1 day. Within each preset time window, the time series data block is truncated, and the sales change rate and market share change rate characteristics for each product within that window are calculated. The sales change rate characteristic is calculated using the following formula:

[0025] in, Indicates the characteristics of sales change rate; This represents the product sales volume at the last point in time within the current sliding window; This represents the product sales volume at the first point in time within the current sliding window; The smoothing constant takes a value of This is used to avoid cases where the denominator is zero.

[0026] The market share change rate feature is calculated by first calculating the difference between the market share value at the last time point within the current sliding window and the market share value at the first time point, and then dividing the difference by the market share value at the first time point. This sliding window aggregation calculation process is repeated for each preset time window and each product to generate multiple sets of sales change rate features and market share change rate features for each product.

[0027] The distributed computing module calculates the configuration similarity between any two products based on the product configuration parameters in the standardized basic dataset. These parameters include numerical configuration items such as engine displacement, maximum power, vehicle length, wheelbase, number of seats, and fuel type. For each product, all numerical configuration items are organized into a configuration vector. Cosine similarity is used to calculate the similarity value between two configuration vectors. This calculation is repeated for any two product combinations, and the results are populated into the product similarity matrix. The product similarity matrix is ​​a symmetric square matrix, where the row and column indices correspond to product identifiers, and the matrix elements represent the configuration similarity values ​​between products in the corresponding row and column.

[0028] The distributed computing module concatenates the sales change rate feature, market share change rate feature, and product similarity matrix for each product. For each product, the sales change rate feature and market share change rate feature calculated under each preset time window are expanded into a one-dimensional feature vector in the order of the time window. Then, the row vector corresponding to the product in the product similarity matrix is ​​concatenated with the one-dimensional feature vector to generate a high-dimensional feature vector for that product. The high-dimensional feature vectors of all products are combined into a high-dimensional analysis dataset. The high-dimensional analysis dataset uses the product identifier as the primary key and includes columns for sales change rate feature, market share change rate feature, and configuration similarity feature.

[0029] Example 3: In specific implementation, please refer to Figure 4 The AI ​​decision-making module receives the user's natural language processing (NLP) intent, which is input as a text string. The module performs word segmentation and part-of-speech tagging on the NLP intent, generating a lexical analysis sequence. The word segmentation operation uses a conditional random field-based model to divide the input text into a sequence of words. The part-of-speech tagging operation assigns a part-of-speech tag to each word based on the segmentation results. The part-of-speech tag set includes categories such as nouns, verbs, adjectives, adverbs, prepositions, conjunctions, numerals, quantifiers, and locative words. The lexical analysis sequence consists of multiple lexical units, each containing the text form of the word, its part-of-speech tag, and its position index in the original sentence.

[0030] The AI ​​decision module performs dependency parsing on the lexical analysis sequence, employing a dependency relation classifier based on a graph neural network. This graph neural network-based dependency relation classifier consists of three graph convolutional network layers, a multi-head attention aggregation layer, and a bilinear affine classification layer. In each of the three graph convolutional network layers, neighborhood features are aggregated on the input node features. The neighborhood is defined as the adjacent edges between words in the lexical analysis sequence. The initial node features are composed of 768-dimensional word vectors extracted from each word by a pre-trained BERT language model. The first graph convolutional network layer receives the initial node features and outputs a 256-dimensional node hidden state; the second graph convolutional network layer receives the 256-dimensional node hidden state output by the first layer and outputs a 128-dimensional node hidden state; the third graph convolutional network layer receives the 128-dimensional node hidden state output by the second layer and outputs a 64-dimensional node hidden state. The multi-head attention aggregation layer concatenates the hidden states of nodes at different granularities output by the three-layer graph convolutional network, calculates the attention weights between any two nodes through an 8-head attention mechanism, and then sums the attention weights with the node hidden states to generate a 64-dimensional fused node representation.

[0031] The bilinear affine classification layer receives the fused node representations of any two nodes and calculates the probability that a dependency relationship exists between the two nodes. The probability value is calculated using the following formula:

[0032] in, Lexical units and lexical units The probability value of an existence of a dependency relationship; Lexical units The fusion node representation vector has a dimension of 64; Lexical units The fusion node representation vector has a dimension of 64; It is a bilinear weight matrix with dimension 1. It is obtained through model training; The weight vector, with a dimension of 128, is obtained through model training. Indicates will and A 128-dimensional vector obtained by concatenating along the dimensional direction; It is a bias scalar, obtained through model training; This represents the Sigmoid activation function.

[0033] The training process of the dependency classifier based on graph neural networks uses a pre-labeled dependency syntax tree library as training data. During training, dependency pairs in the labeled syntax trees are used as positive samples, and word pairs without dependency relationships are randomly selected as negative samples. The binary cross-entropy loss function is used to optimize the model parameters. The initial learning rate is set to 0.001, the batch size is set to 32, and the training epochs are set to 50. In the 30th epoch, the learning rate is decayed to one-tenth of the initial value.

[0034] After calculating the dependency relation probability values ​​for all word pairs in the lexical analysis sequence, the AI ​​decision module uses the maximum spanning tree algorithm to select the tree structure with the largest sum of probability values ​​from all word pairs as the dependency parsing result. Based on the dependency parsing result, the core predicate verb in the natural language analysis intent is identified; the core predicate verb is the root node word of the dependency tree. Subject and object dependencies corresponding to the core predicate verb are extracted from the dependency tree. Subject dependencies are determined by finding child nodes with dependency relation labels of "subject" to the core predicate verb, and object dependencies are determined by finding child nodes with dependency relation labels of "object" to the core predicate verb. The identified relation structures are then organized into a dependency tree.

[0035] The AI ​​decision-making module performs named entity recognition on nominal nodes in the syntactic dependency tree. Named entity recognition employs a sequence labeling model based on a bidirectional long short-term memory (BSSM) network and a conditional random field (CRF) to extract time entities, location entities, product model entities, and indicator name entities from nominal nodes. The BSSM network consists of forward BSSM layers and backward BSSM layers, each with 128 hidden units. The CRF layer uses the label scores output by the BSSM network as emission scores and introduces a label transition score matrix, outputting the optimal entity label sequence through Viterbi decoding. The word text of nominal nodes is input into the sequence labeling model, which outputs corresponding entity type labels. Words labeled as time are classified as time entities, words labeled as location entities as location entities, words labeled as product model entities as product model entities, and words labeled as indicator entities as indicator name entities.

[0036] The AI ​​decision-making module maps natural language analysis intent into a structured query description containing analysis operation type, analysis dimension, and analysis constraints based on the semantic category of the core predicate verb, the entity type in the subject-subject subordination, and the entity type in the object-subordination. The semantic category of the core predicate verb is determined by the mapping rules from verb to analysis operation type: the verb "compare" maps to the comparison analysis operation type, the verb "trend" maps to the trend analysis operation type, the verb "rank" maps to the ranking analysis operation type, and the verb "percentage" maps to the percentage analysis operation type. The entity type in the subject-subject subordination determines the analysis dimension. When the subject-subject subordination contains a time entity, the time value corresponding to the time entity is set as the time dimension in the analysis dimension; when the subject-subject subordination contains a geographic entity, the geographic value corresponding to the geographic entity is set as the geographic dimension in the analysis dimension; and when the subject-subject subordination contains a product model entity, the model value corresponding to the product model entity is set as the product dimension in the analysis dimension. The entity type in the object subordination determines the analysis constraints. Set the indicator value in the indicator name entity as the analysis indicator constraint, set the time range in the time entity as the time range constraint, set the region list in the region entity as the region filter constraint, and set the product list in the product model entity as the product filter constraint.

[0037] The AI ​​decision-making module performs operation decomposition on structured query descriptions, breaking them down into multiple atomic analysis operations with data flow dependencies. This decomposition process follows a pre-defined rule base for analysis operation decomposition, which defines atomic operation sequence templates for each type of analysis operation. For example, comparative analysis operations are decomposed into data filtering atomic operations, grouping and aggregation atomic operations, and difference calculation atomic operations; trend analysis operations are decomposed into time series construction atomic operations and moving average calculation atomic operations; ranking analysis operations are decomposed into data filtering atomic operations, numerical sorting atomic operations, and top N item extraction atomic operations; and proportion analysis operations are decomposed into data filtering atomic operations, sum calculation atomic operations, and ratio calculation atomic operations. The decomposed atomic operations are instantiated based on the analysis dimension entity and constraint entity, embedding the required input fields, filtering conditions, and grouping keys into the atomic operation instances. Data flow dependencies between atomic operation instances are identified, determined by the matching relationship between the input data source of one atomic operation instance and the output data target of another atomic operation instance.

[0038] Execution priorities are assigned to atomic analysis operations based on data flow dependencies. Execution priorities are assigned in a hierarchical, increasing manner: atomic analysis operations without upstream dependencies are assigned a priority of 1; subsequent atomic analysis operations directly dependent on an atomic analysis operation with priority 1 are assigned a priority of 2, and so on, until all atomic analysis operations are assigned an execution priority. All atomic analysis operations and their corresponding execution priorities are organized into a structured analysis task queue, stored as an ordered list. Each element in the list contains an atomic analysis operation identifier, operation type, parameter configuration, execution priority, and a list of upstream dependent operation identifiers.

[0039] Example 4: In practical implementation, the AI ​​decision-making module decomposes the structured query description into operations. The AI ​​decision-making module retrieves an abstract analysis template from a pre-set analysis operation template library that matches the analysis operation type in the structured query description. The pre-set analysis operation template library is stored in key-value pairs, where the key is the analysis operation type identifier and the value is the corresponding abstract analysis template. The abstract analysis template includes input data slots, output data slots, and data processing logic placeholders. The input data slots define the data source description required when the abstract analysis template is executed, including the data source type, filter condition expression placeholders, and field name placeholders. The output data slots define the resulting data structure after the abstract analysis template is executed, including output field names and output data types. The data processing logic placeholders define the core computational logic description of the abstract analysis template, recorded in the form of function signatures and combinations of operation operators.

[0040] The AI ​​decision-making module populates input data slots with parameters based on the analysis dimensions and constraints, generating executable analysis units. The parameter population process replaces the placeholders in the filter expression of the input data slots with the specific values ​​from the analysis dimension entities, and fills the field name placeholders with the indicator names from the analysis constraints. For example, when the analysis dimension includes a product dimension and the analysis constraint includes a sales volume indicator, the list of product identifiers corresponding to the product dimension is filled into the product identifier field of the filter expression placeholders, and the field name corresponding to the sales volume indicator is filled into the field name placeholders, generating executable analysis units containing specific parameter values. Each executable analysis unit contains a complete description of the input data specifications, an output data specifications description, and data processing logic.

[0041] The AI ​​decision-making module identifies data reference relationships between the analysis units to be executed. The identification method involves traversing all analysis units to be executed, extracting the output field name and output data type from the output data slot of each analysis unit, and comparing each analysis unit: if the output field name in the output data slot of the first analysis unit is semantically equivalent to the field name placeholder in the input data slot of the second analysis unit, and the output data type is compatible with the input data type required by the input data slot, then the output data slot of the first analysis unit is determined to match the input data slot of the second analysis unit, establishing a data reference relationship from the first analysis unit to the second analysis unit.

[0042] Based on data reference relationships, the analysis units to be executed are organized into a directed acyclic graph (DAG) structure. The DAG construction process is as follows: a vertex set is created with the analysis units to be executed as nodes; an edge set is created with the data reference relationships as directed edges, with the direction of the directed edges pointing from the analysis unit producing the data to the analysis unit consuming the data. After construction, cycle detection is performed on the directed graph. A depth-first search algorithm is used to traverse all vertices in the graph. If a backward edge is detected, a cycle is determined to exist, triggering a graph structure correction process. The correction process includes inserting data cache operation nodes to cut off the cycle path. After confirming no cycles, each node in the graph is treated as an atomic analysis operation, and the directed edges in the directed graph are treated as data flow dependencies.

[0043] In practice, the task scheduling module obtains the current CPU utilization and current memory utilization of each computing node in the distributed computing module. The current CPU utilization is obtained by reading the process scheduler statistics from the computing node's operating system, and is represented as the percentage of CPU time slices used out of the total CPU time slices. The current memory utilization is obtained by reading the memory management statistics from the computing node's operating system, and is represented as the percentage of allocated physical memory out of the total physical memory. The task scheduling module calculates a comprehensive load score for each computing node based on the current CPU utilization and current memory utilization. The comprehensive load score is calculated using the following formula:

[0044] in, Indicates the first The overall load score of each computing node. The range of values ​​is A higher value indicates a heavier node load. Indicates the first The current CPU utilization of each computing node. The range of values ​​is ; Indicates the first The current memory utilization of each compute node. The range of values ​​is ; This is the CPU load weighting coefficient, with a value of 0.6. This value is set based on the experience that CPU resources have a greater impact on the execution speed of distributed computing tasks than memory resources. This is the memory load weighting coefficient, with a value of 0.4.

[0045] The task scheduling module sorts the computing nodes according to their comprehensive load scores from lowest to highest, generating a sequence of available computing nodes. The sorting operation uses the quicksort algorithm, placing the computing node with the lowest comprehensive load score at the beginning of the available computing node sequence and the computing node with the highest comprehensive load score at the end. For computing nodes with the same comprehensive load score, their order is determined by the lexicographical order of their identifiers.

[0046] The task scheduling module selects a target computing node for each atomic analysis operation from the available computing node sequence based on the estimated computational resources required and the data flow dependencies in the structured analysis task queue. The estimated computational resources are calculated using two methods: estimated memory resources and estimated CPU resources. Based on the input data size of the atomic analysis operation, the estimated memory resources are calculated. The input data size is extracted from the input data slots and represented as the number of input records multiplied by the number of fields per record multiplied by the average number of bytes per field. The estimated memory resources are three times the input data size to cover the needs of input data storage, intermediate calculation result storage, and output data storage. Based on the complexity of the data processing logic placeholders in the atomic analysis operation, the estimated CPU resources are calculated. The complexity is determined by the number and type of operators contained in the data processing logic placeholders. The mapping rule is 1 CPU unit for each basic filtering operator, 3 CPU units for each aggregation operator, 5 CPU units for each sorting operator, and 8 CPU units for each connection operator. The estimated CPU resources are the sum of the CPU units for all operators.

[0047] The task scheduling module identifies the first upstream atomic analysis operation that has an upstream dependency relationship with the current atomic analysis operation. Upstream dependencies are queried from the data flow dependencies in the structured analysis task queue. When the endpoint of a directed edge in a data flow dependency relationship is the current atomic analysis operation, the atomic analysis operation corresponding to the starting point of that directed edge is the first upstream atomic analysis operation. The computing node containing the first upstream atomic analysis operation is recorded as a priority selection node. The task scheduling module traverses the sequence of available computing nodes, obtaining the current remaining memory resources and current remaining CPU resources for each node. The current remaining memory resources are the total physical memory of the computing node minus the allocated physical memory; the current remaining CPU resources are the total number of CPU cores of the computing node multiplied by the unutilized rate. Computing nodes with current remaining memory resources greater than the estimated memory resource value and current remaining CPU resources greater than the estimated CPU resource value are selected to form a set of candidate nodes.

[0048] The task scheduling module determines whether the preferred node exists in the candidate node set. This is done by iterating through each compute node identifier in the candidate node set; if any compute node identifier matches the identifier of the preferred node, then the preferred node exists in the candidate node set. If the preferred node exists in the candidate node set, it is designated as the target compute node for the current atomic analysis operation. If the preferred node does not exist in the candidate node set, the compute node with the lowest overall load score is selected as the target compute node from the candidate node set.

[0049] The task scheduling module sends the data processing logic placeholders corresponding to the atomic analysis operation and the specific values ​​of the input data slots to the target computing node. The sending process uses a remote procedure call protocol, serializing the data processing logic placeholders into executable code strings and the specific values ​​of the input data slots into a set of key-value pairs. These are then encapsulated into a task execution request message and sent to the execution process on the target computing node via a message queue. Upon receiving the task execution request message, the execution process on the target computing node parses the data processing logic placeholders, executes the corresponding computation logic, reads data from the data partition specified by the input data slot, and returns the intermediate computation results after completing the calculation.

[0050] The task scheduling module receives intermediate computation results returned by the execution process and caches all intermediate computation results according to the execution order of atomic analysis operations. The caching method involves creating an intermediate analysis result set storage area in the task scheduling module's memory space, constructing a hash mapping table with atomic analysis operation identifiers as keys and intermediate computation results as values. Intermediate computation results are received and stored sequentially according to the execution priority of atomic analysis operations. When the intermediate computation results of all upstream dependent operations of a certain atomic analysis operation have been cached, the intermediate computation result corresponding to that atomic analysis operation is marked as ready. After all atomic analysis operations have been executed, the task scheduling module provides the entire intermediate analysis result set to the report generation module.

[0051] Example 5: In specific implementation, please refer to Figure 5The report generation module retrieves each intermediate calculation result and its corresponding atomic analysis operation identifier from the intermediate analysis result set. Based on the atomic analysis operation identifier, it extracts the corresponding analysis dimension label and analysis constraint label from the structured analysis task queue. The atomic analysis operation identifier is a unique identifier for each atomic analysis operation in the structured analysis task queue. The structured analysis task queue contains the analysis dimension label and analysis constraint label bound to each atomic analysis operation during execution. The analysis dimension label records the analysis dimension type and specific dimension value involved in the atomic analysis operation, while the analysis constraint label records the filtering conditions and parameter values ​​applied by the atomic analysis operation. The report generation module iterates through all entries in the intermediate analysis result set, using the atomic analysis operation identifier as the search key to read the analysis dimension label and analysis constraint label from the structured analysis task queue. It then appends the read analysis dimension label and analysis constraint label to the corresponding intermediate calculation result, forming a labeled intermediate result. Based on the analysis dimension label, the report generation module merges multiple intermediate calculation results with the same analysis dimension into the same data group. The analysis dimension is uniquely determined by the combination of the dimension type and dimension value in the analysis dimension label. The merge operation uses a combination of dimension type and dimension value as the grouping key to perform grouping aggregation on all labeled intermediate results, grouping labeled intermediate results with the same grouping key into the same data group. Each data group stores multiple labeled intermediate results in list form, with each element containing the original value, the corresponding timestamp, and the analysis constraint label. The report generation module performs a time-series alignment operation on multiple intermediate calculation results within the same data group. The time-series alignment operation reads the timestamp field of all labeled intermediate results within the same data group, extracts the minimum and maximum values ​​of all timestamps, uses the minimum value as the alignment start point, and the maximum value as the alignment end point, generating a continuous timestamp sequence. The time granularity of the continuous timestamp sequence is read from the analysis constraint label. If a time granularity is specified in the analysis constraint label, the specified time granularity is used; otherwise, a default time granularity of 1 day is used. Each labeled intermediate result within the same data group is matched with a time point in the continuous timestamp sequence according to its timestamp. The alignment uses a forward padding method. For time points in the continuous timestamp sequence lacking a corresponding intermediate calculation result, the value of the intermediate calculation result from the nearest past time point is used to fill the gap. If no past time point value exists, the value from the nearest future time point is used. The report generation module resamples the aligned intermediate calculation results at the same time granularity to generate the dataset to be displayed. The resampling operation extracts the corresponding value for each time point in the continuous timestamp sequence from the aligned data group and assembles it into a record. The record contains a resampled timestamp field, a value field, an analysis dimension label field, and an analysis constraint label field.All data groups are resampled and merged to form the dataset to be displayed. This dataset is stored in tabular form, with each row corresponding to a combination of a time point and an analysis dimension. Columns include timestamp, numerical, dimension type, dimension value, and constraint columns. The report generation module selects a matching chart template from the visualization component library based on the data type of the dataset. The visualization component library includes line chart templates, bar chart templates, pie chart templates, scatter plot templates, area chart templates, and table templates. The data type is determined by the numerical characteristics of the numerical columns in the dataset and the combination of the data records' dimensions: a line chart template is selected when the data group contains continuous time-series values ​​and the analysis dimension is time; a bar chart template is selected when the data group contains categorical comparison values; a pie chart template is selected when the data group contains values ​​composed of proportions; a scatter plot template is selected when the data group contains relationships between two numerical fields; an area chart template is selected when the data group requires cumulative value display; and a table template is selected when the data needs to be displayed in detail in rows and columns. The report generation module maps fields from the dataset to be displayed to the data channels of the selected chart template. The chart template's data channels include an X-axis data channel, a Y-axis data channel, a categorical data channel, and a numerical label data channel. The mapping rules are as follows: timestamp fields in the dataset are mapped to the X-axis data channel; numerical fields are mapped to the Y-axis data channel; dimension value fields are mapped to the categorical data channel; and the concatenated values ​​of indicator names in the constraint label fields are mapped to the numerical label data channel. After mapping, each data channel of the chart template is bound to a specific field in the dataset to be displayed, and the data binding engine within the chart template automatically draws graphical elements based on the bound fields. The report generation module then arranges the chart template with the mapped fields according to the logical hierarchy of the analysis dimension labels, generating an automotive product market data analysis dashboard. The report generation module extracts all analysis dimension labels corresponding to the dataset to be displayed and constructs a dimension hierarchy tree based on the inclusion relationships between the analysis dimension labels. The inclusion relationship is determined as follows: if the set of dimension values ​​for the first analysis dimension label is a subset of the set of dimension values ​​for the second analysis dimension label, and the dimension type of the first analysis dimension label is a subordinate concept of the dimension type of the second analysis dimension label, then the first analysis dimension label is a child node of the second analysis dimension label. The root node of the dimension hierarchy tree represents the entire data range and does not contain any specific dimension labels. The child nodes of the dimension hierarchy tree represent progressively subdivided analysis dimensions, and each child node is bound to an analysis dimension label and a chart template for that dimension. The report generation module performs a depth-first traversal of the dimension hierarchy tree, attaching the chart templates with completed field mappings to the corresponding nodes in the dimension hierarchy tree. The depth-first traversal starts from the root node, sequentially visiting the first child node of the root node, and continuing downwards along the child nodes of that child node until a leaf node is reached, after which it backtracks.During the traversal, at each node visited, it is checked whether the analysis dimension label bound to that node matches the analysis dimension label of a chart template with a completed field mapping. If they match, the chart template is mounted to the current node. The mounting operation stores a reference to the chart template object in the node data structure. The report generation module allocates different display area sizes to chart templates mounted on nodes at different levels based on the depth of the dimension hierarchy tree. The display area size is calculated according to the following formula:

[0052] in, Indicates the hierarchy depth as The area of ​​the chart template displayed on the node, in square pixels; This represents the baseline area of ​​the display area corresponding to the root node, with a value of 160,000 square pixels, corresponding to a baseline display area of ​​400 pixels × 400 pixels; This is the level scaling factor, with a value of 0.5. This value is set based on the principle that the display area of ​​each level is reduced to half of the level above, so as to ensure that the charts of higher-level dimensions occupy the visual dominance. This represents the level depth of a node in a hierarchical tree, with the root node being the most significant element. A value of 1 represents the depth of the direct child nodes of the root node. The value is 2, and so on. The display area size is obtained by taking the square root of the display area and rounding it down; that is, both the width and height of the display area are rounded down to the nearest integer. The integer part. The deeper the node, the smaller the display area, allowing charts of fine-grained analysis dimensions to be presented at a smaller size in the dashboard.

[0053] The report generation module fills the corresponding display areas with all chart templates mounted on the dimension hierarchy tree according to the traversal order, generating an automotive product market data analysis dashboard. The filling process follows the dashboard grid layout, which starts from the root node and fills the entire dashboard canvas. Each child node is arranged within its parent node's display area according to the traversal order, filling row by row from left to right and top to bottom. When the remaining width in a row is insufficient to accommodate the next child node, it automatically wraps to the next row. Each display area renders the corresponding chart template's graph. The chart template's title is taken from the dimension value in the bound analysis dimension label, and the chart's axis labels are taken from the indicator names in the analysis constraint label. After all display areas are filled, a complete automotive product market data analysis dashboard is generated. The dashboard contains various charts organized hierarchically by dimension, allowing users to directly observe automotive product market data under different analysis dimensions.

[0054] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A car product market data analysis system based on AI intelligent agents and big data, characterized in that: include: The data management module is used to receive and store multimodal raw data of the automotive market from multiple heterogeneous data sources, and to perform deduplication, outlier removal and missing value imputation operations on the multimodal raw data to generate a standardized basic dataset. The distributed computing module is used to perform in-memory distributed iterative computing on the standardized basic dataset to generate a high-dimensional analysis dataset containing product relationships and time series features; The AI ​​decision module is used to receive the natural language analysis intent input by the user, and perform dependency parsing and named entity joint recognition on the natural language analysis intent according to the high-dimensional analysis dataset to generate a structured analysis task queue, which contains multiple atomic analysis operations with dependencies. The task scheduling module is used to receive the structured analysis task queue, allocate computing nodes for the atomic analysis operation according to the current cluster resource load status, drive the distributed computing module to execute the atomic analysis operation, and obtain an intermediate analysis result set. The report generation module is used to perform multi-source result correlation aggregation on the intermediate analysis result set and call the visualization component library to generate a user-oriented automotive product market data analysis dashboard.

2. The automotive product market data analysis system based on AI intelligent agents and big data as described in claim 1, characterized in that, The multimodal raw data is subjected to deduplication, outlier removal, and missing value imputation operations to generate a standardized basic dataset, specifically including: Based on a preset data quality rule base, a data integrity score is calculated for each record in the multimodal raw data; Records with data integrity scores below a preset threshold are marked as records to be cleaned. Perform field-level anomaly detection on the records to be cleaned to identify outlier records in numeric fields and illegal value records in categorical fields; Perform deletion operations on the outlier records and the illegal value records, and write the deletion operation log to the data traceability table; For missing fields in the remaining records after deletion, context records related to the missing field are retrieved from the multimodal raw data according to the data domain to which the field belongs, and the missing fields are filled in according to the mode or median value of the same field in the context records to generate the standardized basic dataset.

3. The automotive product market data analysis system based on AI intelligent agents and big data according to claim 1, characterized in that, Perform in-memory distributed iterative computation on the standardized basic dataset to generate a high-dimensional analysis dataset containing product relationships and time series features, specifically including: The standardized basic dataset is partitioned according to product identifiers, and each partition is loaded into the memory space of the corresponding computing node to construct in-memory data partitions; The records in each memory data partition are sorted according to their timestamps to generate time-series data blocks; Perform sliding window aggregation calculations on the time series data blocks to generate sales change rate characteristics and market share change rate characteristics for each product within multiple preset time windows; Based on the product configuration parameters in the standardized basic dataset, calculate the configuration similarity between any two products and construct a product similarity matrix; The sales change rate feature, the market share change rate feature, and the product similarity matrix are concatenated to generate the high-dimensional analysis dataset.

4. The automotive product market data analysis system based on AI intelligent agents and big data according to claim 1, characterized in that, The natural language analysis intent is subjected to dependency parsing and named entity joint recognition to generate a structured analysis task queue, specifically including: Perform word segmentation and part-of-speech tagging operations on the stated natural language analysis intent to generate a lexical analysis sequence; Dependency parsing is performed on the lexical analysis sequence to identify the core predicate verb in the natural language analysis intent and the subject-object subordination and object-object subordination corresponding to the core predicate verb, and a syntactic dependency tree is generated. Named entity recognition is performed on the noun nodes in the syntactic dependency tree to extract time entity, region entity, product model entity and indicator name entity from the noun nodes; Based on the semantic category of the core predicate verb, the entity type in the subject subordination, and the entity type in the object subordination, the natural language analysis intent is mapped into a structured query description that includes analysis operation type, analysis dimension, and analysis constraints. The structured query description is decomposed into multiple atomic analysis operations with data flow dependencies, and the execution priority of the atomic analysis operations is set according to the data flow dependencies to generate the structured analysis task queue.

5. The automotive product market data analysis system based on AI intelligent agents and big data according to claim 4, characterized in that, When performing dependency parsing on the lexical analysis sequence, a dependency relation classifier based on a graph neural network is used. Each word in the lexical analysis sequence is used as a node and the adjacency relationship between words is used as an edge. The contextual semantic features are iteratively aggregated, and the probability value of a dependency relationship between any two words is output. The dependency relationship with the highest probability value is taken as the result of the dependency parsing.

6. The automotive product market data analysis system based on AI intelligent agents and big data according to claim 4, characterized in that, The structured query description is decomposed into multiple atomic analysis operations with data flow dependencies, specifically including: Retrieve an abstract analysis template from a preset analysis operation template library that matches the analysis operation type in the structured query description. The abstract analysis template includes input data slots, output data slots, and data processing logic placeholders. Based on the analysis dimensions and analysis constraints, the input data slots are filled with parameters to generate an analysis unit to be executed. Identify the data reference relationships between the analysis units to be executed, and match the output data slot of the first analysis unit to be executed with the input data slot of the second analysis unit to be executed; Based on the data reference relationship, the analysis unit to be executed is organized into a directed acyclic graph structure, each node in the directed acyclic graph is taken as an atomic analysis operation, and the directed edges in the directed acyclic graph are taken as the data flow dependencies.

7. The automotive product market data analysis system based on AI intelligent agents and big data according to claim 6, characterized in that, Based on the current cluster resource load status, computing nodes are allocated to the atomic analysis operation, driving the distributed computing module to execute the atomic analysis operation and obtain a set of intermediate analysis results, specifically including: Obtain the current CPU utilization and current memory utilization of each computing node in the distributed computing module, and calculate the comprehensive load score of each computing node based on the current CPU utilization and current memory utilization. The computing nodes are sorted in ascending order of the comprehensive load score to generate a sequence of available computing nodes; Based on the estimated computational resources required for the atomic analysis operation and the data flow dependencies in the structured analysis task queue, a target computational node is selected from the available computational node sequence for each atomic analysis operation; The data processing logic placeholder corresponding to the atomic analysis operation and the specific value of the input data slot are sent to the target computing node to trigger the execution process on the target computing node. The intermediate calculation results returned by the execution process are received, and all the intermediate calculation results are cached in the execution order of the atomic analysis operations to generate the intermediate analysis result set.

8. The automotive product market data analysis system based on AI intelligent agents and big data according to claim 7, characterized in that, Based on the estimated computational resources required for the atomic analysis operation and the data flow dependencies in the structured analysis task queue, a target computational node is selected from the available computational node sequence for each atomic analysis operation, specifically including: Calculate the estimated memory resource requirements of the atomic analysis operation based on the input data size of the atomic analysis operation; Calculate the estimated CPU resource requirement of the atomic analysis operation based on the complexity of the data processing logic placeholders of the atomic analysis operation. Traverse the available computing node sequence and select computing nodes whose current remaining memory resources are greater than the estimated memory resource value and whose current remaining CPU resources are greater than the estimated CPU resource value as the candidate node set; Based on the data flow dependencies in the structured analysis task queue, identify the first upstream atomic analysis operation that has an upstream dependency relationship with the current atomic analysis operation, and select the first computing node where the first upstream atomic analysis operation is located as the preferred node. Determine whether the preferred node exists in the set of candidate nodes. If it does, then determine the preferred node as the target computing node for the current atomic analysis operation.

9. The automotive product market data analysis system based on AI intelligent agents and big data according to claim 1, characterized in that, The intermediate analysis result set is subjected to multi-source result correlation and aggregation, and a visualization component library is invoked to generate a user-oriented automotive product market data analysis dashboard, specifically including: Obtain each intermediate calculation result and its corresponding atomic analysis operation identifier from the intermediate analysis result set, and extract the corresponding analysis dimension label and analysis constraint label from the structured analysis task queue according to the atomic analysis operation identifier; Based on the analysis dimension label, multiple intermediate calculation results with the same analysis dimension are merged into the same data group; Perform a time-series alignment operation on multiple intermediate calculation results within the same data group, resample the intermediate calculation results according to the same time granularity, and generate a dataset to be displayed. Based on the data type of the dataset to be displayed, a matching chart template is selected from the visualization component library, and the fields in the dataset to be displayed are mapped to the data channels of the chart template; The chart templates with completed field mappings are laid out and arranged according to the logical hierarchy of the analysis dimension labels to generate the automotive product market data analysis dashboard.

10. The automotive product market data analysis system based on AI intelligent agents and big data according to claim 9, characterized in that, The chart templates with completed field mappings are laid out and arranged according to the logical hierarchy of the analysis dimension labels to generate the automotive product market data analysis dashboard, specifically including: Extract all the analysis dimension labels corresponding to the dataset to be displayed, and construct a dimension hierarchy tree based on the inclusion relationship between the analysis dimension labels. The root node of the dimension hierarchy tree represents the entire data range, and the child nodes represent the analysis dimensions that are subdivided level by level. Perform a depth-first traversal on the dimension hierarchy tree and mount the chart template with completed field mapping to the corresponding node of the dimension hierarchy tree; Based on the depth of the dimension hierarchy tree, different display area sizes are assigned to the chart templates mounted on nodes at different levels, wherein the display area size is smaller for nodes at deeper levels. Following the traversal order of the dimension hierarchy tree, all the chart templates mounted on the dimension hierarchy tree are filled into the corresponding display areas to generate the automotive product market data analysis dashboard.

Images