Big data-based automobile industry customer demand analysis and market trend prediction method

Abstract

The application relates to the technical field of natural language processing, and discloses a method for analyzing customer demand and predicting market trends in the automobile industry based on big data, which comprises the following steps: collecting user expression, vehicle configuration, vehicle door handle hardware data, regional time, scene and platform source data to construct an analysis sample; extracting the configuration parameters of the vehicle door handle hardware to construct a hardware benchmark, and aligning the user expression to the hardware benchmark; decoupling the user expression to extract appearance demand degree, safety demand degree and scene mutation degree; combining node information entropy sharing amount, time distribution difference and scene mutation degree to generate a propagation weight and construct a propagation correlation graph; generating a time decay kernel according to a mutation sequence to extract a historical influence degree; inputting a comprehensive state containing the scene mutation degree and the historical influence degree into a time sequence model to obtain a demand trend value, and updating the time sequence model according to a prediction error feedback.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and more specifically, to a method for analyzing customer needs and predicting market trends in the automotive industry based on big data. Background Technology

[0002] Automotive demand forecasting is gradually shifting from simple macro-level time-series forecasting to multi-dimensional data mining based on online behavior, external variables, and user-generated content. In current automotive sentiment mining and analysis, conventional natural language processing models typically categorize user feedback into independent categories such as appearance preferences, intelligent configuration preferences, or after-sales maintenance based on word distribution or sentiment polarity.

[0003] However, with the widespread application of new exterior body components (such as concealed or electrically retractable door handles), user behavior data shows a strong context-dependent nature. In static showroom experiences and routine daily use, users tend to express positive appreciation for the technological feel and aesthetic continuity; but when vehicles encounter specific external events such as low-temperature icing, rain and snow intrusion, water wading, collision power loss, or nighttime rescue, the originally positive demand tends to quickly shift to negative demands for mechanical redundancy release, operability, and safety availability.

[0004] Existing text mining solutions often process customer complaints in isolation due to the lack of a mechanism to ternarily bind natural language with specific vehicle body hardware configurations, external climate, and accident conditions. Furthermore, the regional seasonality, sporadic nature, and time lag in cross-platform diffusion of such usability anxieties lead conventional analysis models to filter them out as ordinary appearance complaints or occasional noise. This lack of analytical dimension prevents existing systems from capturing potential customer complaint crises caused by specific vehicle body hardware, resulting in judgments that deviate from actual engineering safety requirements in the R&D and selection predictions for next-generation vehicle configurations. Summary of the Invention

[0005] This invention provides a method for analyzing customer needs and predicting market trends in the automotive industry based on big data, which solves the technical problems mentioned in the background.

[0006] This invention provides a method for analyzing customer needs and predicting market trends in the automotive industry based on big data, applicable to automotive configurations that include hidden or electrically retractable door handles, including: Collect user expression data, vehicle configuration data, door handle structure data, regional and temporal data, scenario data, and platform-sourced data to construct an analytical sample representing the state of multi-source heterogeneous inputs; The physical configuration parameters of the door handle structure data are extracted to construct a structural benchmark, and the user expression data is aligned to the structural benchmark to eliminate semantic discrepancies in non-standard expressions; The user expression data is decoupled to extract the appearance demand degree representing the visual attribute direction, the safety demand degree representing the reliability of mechanical operation, and the scene change degree representing the environmental trigger state. Construct a propagation relationship graph containing demand nodes, and generate propagation weights that characterize the scale of cross-domain node diffusion by combining the shared information entropy of nodes, the difference in time distribution, and the abruptness of the scenario. Based on the mutation sequence composed of the scene mutation degree, a time decay kernel is generated to extract the historical influence degree across cycles; The comprehensive state, including the scenario mutation rate and the historical influence rate, is input into the time series model to obtain the demand trend value in the future time. The structural influence degree decomposition is performed on the demand trend value, and the time series model is updated based on the prediction error feedback determined by actual observation.

[0007] The beneficial effects of this invention are as follows: This invention integrates and spatially aligns natural language expressions, vehicle hardware benchmarks, and external triggering scenarios to identify semantic reversal phenomena in customer complaints caused by hidden or electrically retractable door handles under specific natural environments and unexpected events. By combining node information entropy and temporal distribution differences to calculate cross-domain propagation weights and introducing a time decay kernel to extract long-tail historical influence, this solution quantifies user safety and usability demands induced by low-frequency, delayed events. The predicted trend values ​​output by the system are directly correlated with specific hardware configuration parameters, avoiding the bias of conventional sentiment analysis models that misclassify them as ordinary after-sales complaints, thus providing data support for automotive engineering research and development and hardware iteration. Attached Figure Description

[0008] Figure 1 This is a flowchart of the big data-based customer demand analysis and market trend prediction method for the automotive industry, as presented in this invention. Detailed Implementation

[0009] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0010] like Figure 1 As shown, the big data-based customer demand analysis and market trend forecasting method for the automotive industry is applied to automotive configurations that include hidden or electrically retractable door handles, including: Collect user expression data, vehicle configuration data, door handle structure data, regional and temporal data, scenario data, and platform-sourced data to construct an analytical sample representing the state of multi-source heterogeneous inputs; The physical configuration parameters of the door handle structure data are extracted to construct a structural benchmark, and the user expression data is aligned to the structural benchmark to eliminate semantic discrepancies in non-standard expressions; The user expression data is decoupled to extract the appearance demand degree representing the visual attribute direction, the safety demand degree representing the reliability of mechanical operation, and the scene change degree representing the environmental trigger state. Construct a propagation relationship graph containing demand nodes, and generate propagation weights that characterize the scale of cross-domain node diffusion by combining the shared information entropy of nodes, the difference in time distribution, and the abruptness of the scenario. Based on the mutation sequence composed of the scene mutation degree, a time decay kernel is generated to extract the historical influence degree across cycles; The comprehensive state, including the scenario mutation rate and the historical influence rate, is input into the time series model to obtain the demand trend value in the future time. The structural influence degree decomposition is performed on the demand trend value, and the time series model is updated based on the prediction error feedback determined by actual observation.

[0011] This embodiment applies to automotive configurations that include concealed or electrically retractable door handles. The method achieves accurate analysis of automotive industry customer needs and reliable prediction of market trends through steps such as multi-source heterogeneous data fusion, natural language and hardware structure alignment, demand decoupling and propagation analysis, and time-series prediction and feedback updates.

[0012] S201: Based on the door handle structure data and vehicle configuration data, extract the outer surface flatness, motion pair morphology, electronic triggering logic, mechanical redundancy accessibility, and operating space scale to construct structural features that reflect physical interference properties.

[0013] Door handle structural data refers to the 3D geometric model data, engineering drawings, and physical parameter data of concealed or electrically retractable door handles, covering information such as the handle's external dimensions, movement trajectory, internal mechanical structure, and electronic control unit parameters. Vehicle configuration data refers to the official configuration list data for different vehicle models, including whether each model is equipped with concealed or electrically retractable door handles, as well as the specific model and version information of the handles.

[0014] Outer surface flatness refers to the height difference between the outer surface of the handle's center position and the outer surface of the door panel's center position when the door handle is closed, at a normal temperature of 25 degrees Celsius, and with the doors fully closed and locked. The unit is millimeters. This parameter directly affects the vehicle's overall appearance and aerodynamic performance. Kinematic pair morphology refers to the connection method and relative motion form between the components during the door handle's movement. It is represented numerically using a discretized coding method, where a rotary joint corresponds to value 1, a sliding joint to value 2, and a helical joint to value 3. Electronic triggering logic refers to the door handle's triggering method and control flow. It is represented numerically using a discretized coding method, where capacitive touch triggering corresponds to value 1, mechanical button triggering to value 2, and inductive automatic triggering to value 3. Mechanical redundancy reachability refers to the ease and reliability of manually opening the door handle mechanically when the electronic control system fails. The value ranges from 0 to 1, where 0 indicates that it is completely impossible to open manually, and 1 indicates that it can be opened within 1 second without any tools. The intermediate value is obtained by weighting the opening time and opening force. ,in The time required to manually start, in seconds. The force required to manually open the door handle is measured in Newtons (N). The operating space dimension refers to the minimum space required for the user to operate the door handle, measured in millimeters. It includes three sub-parameters: the extended length of the handle after it pops out, the width of the operating area, and the height of the operating area. The measurement starts at the outer surface of the door sheet metal and ends at the furthest position of the handle after it pops out.

[0015] The calculation formula for constructing structural features that reflect physical interference properties is as follows: ; in, As a structural feature, it is a 5-dimensional vector; For the flatness of the outer surface; It is a kinetic submorph; For electronic triggering logic; For mechanical redundancy reachability; The scale of the operation space is denoted as , and is a 3D vector obtained by concatenating the three sub-parameters. Here, for simplification, it is represented as a single element. Vehicle configuration data; For the structural data of the car door handle; The feature extraction function can be implemented using a pre-trained convolutional neural network. This network consists of three convolutional layers with a kernel size of 3×3 and a stride of 1. The activation function is ReLU, and the final layer is a fully connected layer, which maps the input three-dimensional geometric model data into a 5-dimensional structural feature vector. Alternatively, it can be implemented using engineering rules. The flatness of the outer surface is calculated by the difference in z-coordinates between the center of the handle and the center of the door sheet metal. The kinematic pair shape and electronic triggering logic are directly encoded by the configuration data. The mechanical redundancy accessibility is calculated by the weighted formula mentioned above, and the operating space scale is calculated by the bounding box size of the three-dimensional model. For sample index, For vehicle model identification.

[0016] The selection criteria for the feature extraction function are as follows: when the three-dimensional geometric model data of the door handle is available, a convolutional neural network is used; when only engineering drawings and configuration data are available, engineering rules are used. The specific measurement methods for the three sub-parameters of the operating space scale are as follows: the extension length is the vertical distance from the outermost point of the handle after it is fully extended to the outer surface of the door sheet metal; the width is the maximum horizontal distance between the left and right edges of the handle operating area; and the height is the maximum vertical distance between the upper and lower edges of the handle operating area. The opening time and opening force measurement conditions in the mechanical redundancy accessibility calculation are as follows: under normal temperature of 25 degrees Celsius and normal pressure of 101 kPa, three measurements are taken by adult males with a height of 170 cm to 180 cm using the standard grip method, and the average value is taken.

[0017] S202: Extract corresponding low temperature, ice and snow, collision and rescue scene factors using regional time data and scene data. Perform distribution unification and dimensionality uncorrelation processing on each scene factor to eliminate the dimensional differences under different climatic conditions and construct scene features.

[0018] Geographic time data refers to the geographic location and timestamp information corresponding to the time the user feedback was generated. The geographic location information is accurate to the city level, and the timestamp information is accurate to the hour level. Scenario data refers to the vehicle usage environment and event information described in the user feedback, including but not limited to weather conditions, road conditions, and vehicle status.

[0019] The low-temperature scenario factor refers to the absolute temperature value when the ambient temperature is below 0 degrees Celsius, expressed in degrees Celsius. The minimum temperature threshold is -40 degrees Celsius; temperatures below this threshold are uniformly calculated as 40 degrees Celsius. Temperature data is derived from hourly observations from local weather stations. The snow and ice scenario factor refers to the comprehensive quantitative value when there is snow or ice on the road surface. Snow thickness refers to the snow thickness at the door handle location, expressed in millimeters. The snow compaction coefficient is 0.8. Icing intensity is divided into three levels: Level 1 corresponds to thin ice less than 5 millimeters thick, Level 2 corresponds to medium ice 5 to 10 millimeters thick, and Level 3 corresponds to thick ice greater than 10 millimeters thick. The formula for calculating the snow and ice scenario factor is... ,in For snow depth, This refers to the icing intensity level. The collision scenario factor indicates the severity of a collision, ranging from 0 to 1, and is calculated using the following formula: ,in The speed of the vehicles at the time of the collision is expressed in kilometers per hour. The collision location is weighted as follows: frontal collision corresponds to 1, side collision to 0.8, and rear collision to 0.6. The rescue scenario factor refers to the urgency of the vehicle's emergency situation requiring rescue, with a value ranging from 0 to 1. Specifically, vehicle power failure corresponds to 1, vehicle rollover corresponds to 0.8, and injuries to occupants correspond to 0.9. The maximum value among these three states is taken as the rescue scenario factor value. When data for a scenario factor is missing, it is filled using the historical average value from the same region and time period.

[0020] Distribution unification processing refers to converting each scene factor into a standard normal distribution with a mean of 0 and a variance of 1. Dimensional uncorrelation processing refers to eliminating the linear correlation between scene factors and reducing the redundancy between feature dimensions through principal component analysis. The historical data used to calculate the uncorrelation covariance matrix is ​​nationwide scene factor data from the past 3 years, with 4 principal components, retaining more than 95% of the data variance.

[0021] The formula for calculating scene features is as follows: ; in, The scene feature is a 4-dimensional vector; In order to be related to the region and time The corresponding scene factor vector includes four scene factors: low temperature, ice and snow, collision and rescue. This is the distribution mean vector of the scene factors, calculated using historical data; To solve the relevant covariance matrix, principal component analysis was performed on historical scene factor data. This is the inverse square root matrix of the correlation covariance matrix, used to implement dimensional uncorrelation processing.

[0022] The method for filling missing values ​​of scene factors is to take the arithmetic mean of the scene factors corresponding to the same city and the same hour over the past 3 years; the steps of principal component analysis are as follows: first, calculate the covariance matrix of all historical scene factor data, then solve the eigenvalues ​​and eigenvectors of the covariance matrix, sort the eigenvalues ​​from largest to smallest, and take the first 4 eigenvectors to form the projection matrix; the calculation rules for ice and snow scene factors are as follows: when there is only snow and no ice, the ice intensity is taken as 0; when there is only ice and no snow, the snow thickness is taken as 0.

[0023] S203: Multimodal concatenation of user expression data, structural features, and scene features is performed to obtain analysis samples.

[0024] User-expressed data refers to textual comments, reviews, complaints, and suggestions about car door handles posted by users on various online platforms, car forums, social media, customer service systems, and other channels. The preprocessing workflow for user-expressed data includes three steps: word segmentation, stop word removal, and spell correction. Word segmentation uses the jieba tool, stop word removal uses the Harbin Institute of Technology stop word list, and spell correction uses a pre-trained Chinese text correction model. The preprocessed user-expressed data is converted into 768-dimensional text vectors, implemented using BERT word embedding. Texts shorter than 512 characters are padded with zeros, while texts longer than 512 characters are truncated at the end.

[0025] Multimodal concatenation refers to connecting the feature vectors of different modalities end to end in the order of structural features, scene features, and user expression data to form a unified high-dimensional feature vector. The structural features are 5-dimensional, the scene features are 4-dimensional, and the user expression data is 768-dimensional, with a total dimension of 777 after concatenation.

[0026] The formula for calculating the analysis sample is as follows: ; in, For the analysis of samples; Structural features; For scene features; The text vector that represents the data expressed to the user.

[0027] In the user expression data preprocessing, jieba word segmentation adopts the precise mode; the stop word removal process involves precisely matching the word segmentation results with the Harbin Institute of Technology stop word list, and directly deleting words that match successfully; the misspelling correction adopts a Chinese error correction model based on bert4keras; the text vector generation method is to take the output of the first token of the last layer of the BERT model as the vector representation of the corresponding text.

[0028] S301: Perform natural language encoding on user expression data to obtain text features.

[0029] Natural language encoding refers to converting user-expressed text data into numerical vectors that can be processed by computers. This is achieved using a BERT-based-cased pre-trained language model. The maximum length of the input text is 512, and long texts are truncated at the end. Emojis and special characters in the user-expressed data are directly filtered out; internet slang is converted to standard Chinese expressions using synonym replacement; and for mixed Chinese and English text, the English portion is also encoded using the same BERT model. This pre-trained model was fine-tuned using a review dataset from the automotive field, with three training epochs, a batch size of 32, and a learning rate of 2e-5.

[0030] The formula for calculating text features is: ; in, The text feature is a 768-dimensional vector; This is the encoding function of the fine-tuned BERT pre-trained language model. The input is the preprocessed text sequence, and the output is the corresponding text feature vector.

[0031] When fine-tuning the pre-trained language model, the ratio of training set to validation set is 8:2. An early stopping strategy is adopted, and training is terminated when the validation set loss does not decrease for three consecutive rounds. The emoji filtering rule is to filter all emojis with Unicode encoding in the range of U+1F600 to U+1F64F. The special character filtering rule is to filter all characters except Chinese characters, uppercase and lowercase English letters, and Arabic numerals. The synonym replacement table for internet slang is derived from the Baidu Encyclopedia entry on internet slang and a compilation of frequently used terms from car forums.

[0032] S302: Combine the structural mapping relationship learned during training to perform spatial transformation on the structural features, and integrate the common features of the associated professional structural lexicon to generate a structural benchmark.

[0033] The structural mapping relationship refers to the linear mapping relationship from the structural feature space to the text feature space obtained through supervised learning training; the structural mapping matrix is ​​the key component of this relationship. The matrix has a dimension of 768×5, with the number of rows matching the dimension of the text features and the number of columns matching the dimension of the structural features. The dataset used to train this matrix consists of 100,000 labeled user reviews of car models. The objective function is the mean squared error loss function, and the optimizer uses the Adam optimization algorithm. The training run consists of 10 epochs.

[0034] The professional terminology database refers to a collection of technical terms related to automotive door handles, specifically including 23 terms such as concealed handle, electric retractable handle, outer surface flushness, kinematic pair, electronic trigger, mechanical redundancy, operating space, and pop-out distance. Different vehicle models correspond to different subsets of the database, filtered according to the vehicle's configuration. For example, the subset for vehicles equipped with concealed handles includes terms related to concealed handles. The database is updated quarterly to supplement and adjust it based on the latest industry terminology.

[0035] The formula for calculating the structural datum is as follows: ; in, The structural baseline is a 768-dimensional vector; The structure mapping matrix; Structural features; In order to match the model The corresponding professional terminology for car door handles; The number of words in the professional structure lexicon; For a single word in a specialized lexicon; Professional terminology The feature vector is obtained by encoding through a pre-trained language model.

[0036] During the training of the structure mapping matrix, the ratio of the training set to the validation set was 9:1, and the optimizer parameters were set to beta1=0.9, beta2=0.999, and epsilon=1e-8. The professional structure lexicon was selected based on the following criteria: it includes all terms related to car door handles in the national standards of the automotive industry, as well as related words that appear more than 1,000 times in automotive forums. The lexicon was updated by collecting the latest industry terms and high-frequency online words every quarter, which were then reviewed and approved by more than 3 experts in the field of automotive engineering before being added to the lexicon.

[0037] S303: Evaluate the distribution similarity between text features and structural benchmarks in the same feature space, and eliminate long-tail text interference through normalization processing to obtain text relevance characterizing the strength of the corresponding entity association, thus completing the soft alignment to specific vehicle body physical hardware.

[0038] Distributional similarity refers to the cosine similarity between two vectors in the feature space, ranging from -1 to 1. A higher value indicates greater similarity between the two vectors. Normalization involves applying softmax normalization to the similarity values ​​of all candidate structures, converting them into a probability distribution form to eliminate numerical bias caused by long-tail text. The candidate vehicle set refers to the set of currently available vehicles of the same brand, class, and price range (within a 20% fluctuation range) as the current vehicle model; this set is updated quarterly. The text relevance threshold is 0.1; user-expressed data below this threshold is considered irrelevant to the door handle and is filtered out.

[0039] The formula for calculating text relevance is: ; in, This represents text relevance, with a value ranging from 0 to 1. A higher value indicates a stronger correlation between the user's expressed data and the corresponding car model's door handle. Text features; As a structural reference; The dot product of text features and structural baselines is used to calculate cosine similarity. In time With car model A collection of all relevant candidate models; Candidate models The corresponding structural reference.

[0040] The specific screening process for the candidate model set is as follows: first, models from the same brand as the current model are screened; then, models with the same body type are screened; and finally, models with an official guide price between 80% and 120% of the current model's official guide price are screened. The method for determining the text relevance threshold is to statistically analyze the text relevance distribution of 100,000 labeled data points and take the highest F1 value of 0.1 as the final threshold. Soft alignment refers to establishing the association between user expression and hardware structure through probability distribution, rather than a one-to-one hard match.

[0041] S401: Perform cross-domain projection and feature fusion on text features, scene features and structural features using independent mapping matrices, and obtain the demand distribution value through nonlinear activation.

[0042] Independent mapping matrices refer to three different linear mapping matrices designed separately for text features, scene features, and structural features. The mapping matrix for text features is... The dimension is 64×768, which is a mapping matrix for scene features. The dimension is 64×4, which is the mapping matrix for structural features. The dimension is 64×5, used to project features from different domains onto the same 64-dimensional demand feature space. Feature fusion refers to the element-wise addition of the three projected feature vectors. The temperature coefficient of the softmax function is 0.8, used to adjust the smoothness of the demand distribution. The demand category system contains 64 different user demand categories, covering multiple dimensions such as appearance, security, usability, and reliability. The dataset used to train these three mapping matrices is labeled user review data, the objective function is the cross-entropy loss function, the optimizer uses the Adam optimization algorithm, and the training epochs are 10.

[0043] The formula for calculating the demand distribution value is: ; in, The demand distribution value is a 64-dimensional probability distribution vector, representing the probability distribution of user-expressed data across various demand categories. This is an independent mapping matrix for text features; Text features; This is an independent mapping matrix for scene features; For scene features; This is an independent mapping matrix for structural features; is the structural feature; softmax is the normalization exponential function used to convert the fused features into a probability distribution form.

[0044] The requirement category system is specifically divided into 12 categories of appearance, 18 categories of security, 15 categories of usability, and 19 categories of reliability. The independent mapping matrix training set contains 500,000 labeled user comments, and each requirement category contains at least 5,000 samples. The softmax temperature coefficient of 0.8 is the optimal value in the experiment. The smaller the temperature coefficient, the sharper the requirement distribution, and the larger the coefficient, the smoother the distribution.

[0045] S402: Extract features from the demand distribution values ​​along the preset principal components of visual attributes and safety and availability, and parse out the corresponding appearance demand and safety demand.

[0046] The visual attribute principal components refer to the principal component vectors obtained by performing principal component analysis on a large amount of user evaluation data on car appearance over the past three years, representing the main focus of users on the visual attributes of car appearance. The safety and usability principal components refer to the principal component vectors obtained by performing principal component analysis on a large amount of user evaluation data on car safety and usability over the past three years, representing the main focus of users on car safety and usability. Both principal component vectors have a dimension of 64 and have undergone L2 normalization. The obtained appearance demand and safety demand are mapped to a value range of 0 to 1 through min-max normalization. When outliers occur, values ​​less than 0 are corrected to 0, and values ​​greater than 1 are corrected to 1.

[0047] The formulas for calculating appearance requirements and safety requirements are as follows: ; ; in, The value represents the degree of demand for the appearance of the door handles, ranging from 0 to 1. The higher the value, the higher the degree of demand for the appearance attributes of the door handles. The principal component vector of visual attributes; This represents the demand distribution value; The safety requirement level ranges from 0 to 1, with a higher value indicating a higher level of user demand for the safety and usability of the door handles. This is the principal component vector for security and availability.

[0048] The visual attribute principal component dataset contains 2 million user reviews related to car appearance, and the safety and usability principal component dataset contains 1.5 million user reviews related to car safety. The number of principal components is determined by counting those with a cumulative variance contribution rate of 95%. The specific formula for min-max normalization is as follows: ,in and These are the minimum and maximum values ​​of appearance requirement in the training set, respectively. The same method is used to normalize the security requirement.

[0049] S403: Extract the attribute offsets of safety requirements and appearance requirements, and perform boundary constraint processing by integrating the coupling interference terms of scene features and structural features, and then combine them with text relevance to determine the scene mutation degree.

[0050] The attribute offset refers to the difference between safety requirements and appearance requirements, representing the degree to which user needs shift from appearance to safety. This difference is first normalized to a range of -1 to 1 using min-max normalization. The coupling interference term refers to the interaction between scene features and structural features, representing the differences in the impact of different door handle structures on user needs under different scenarios. The coupling interference term matrix... The matrix has a dimension of 4×5, and the training dataset consists of labeled scene structure samples. The objective function is the mean squared error loss function, and the optimizer uses the Adam optimization algorithm with 5 training epochs. Boundary constraint processing involves using the tanh nonlinear function to limit the calculation results to a range of -1 to 1. The input value of the tanh function is limited to a range of -5 to 5; input values ​​exceeding this range are truncated to prevent gradient vanishing. In the scene abruptness value, a positive value indicates a sudden shift in user demand towards safety, a negative value indicates a sudden shift in user demand towards appearance, and a value of 0 indicates a balance between the two demands.

[0051] The formula for calculating the abrupt change in scene is: ; in, This represents the scene abruptness, with a value ranging from -1 to 1. For text relevance; For safety requirements; For appearance requirements; The attribute offset between safety requirements and appearance requirements; For scene features; This is the coupling interference term matrix between the scene and the structure; Structural features; is the coupling interference term between scene features and structural features; tanh is the hyperbolic tangent function, used to implement boundary constraint processing.

[0052] The attribute offset min-max normalization parameter is -1 for the minimum and 1 for the maximum. The coupling interference term matrix training set contains 100,000 labeled scene structure samples, and each scene structure combination contains at least 1,000 samples. The tanh function input truncation is implemented such that the input value is 5 when it is greater than 5 and -5 when it is less than -5. A scene abruptness of 0 indicates that the user's needs for appearance and security are in a balanced state, with no obvious tendency for abrupt changes.

[0053] S501: Construct a propagation relationship graph with vehicle type, structure, region, scenario, platform, and demand node as the topological core.

[0054] A propagation relationship graph is a directed graph structure used to represent the information propagation relationships between different nodes. The nodes in the graph include six types: vehicle model nodes represent different car models currently on sale, containing 1200 nodes; structure nodes represent different types of door handle structures, containing 3 nodes: hidden, electric retractable, and traditional; region nodes represent different prefecture-level city geographical regions, containing 337 nodes; scenario nodes represent different vehicle usage scenarios, containing 4 nodes: low temperature, snow and ice, collision, and roadside assistance; platform nodes represent different user expression and publishing platforms, containing 7 nodes: Autohome, Dongchedi, Douyin, Weibo, Taobao, JD.com, and customer service systems; and demand nodes represent different user demand categories, containing 64 nodes, corresponding one-to-one with the 64 categories of demand distribution values. Each node's unique identifier is generated by adding a type prefix to the node ID; for example, the identifier for the region node Beijing is city_110000, and the identifier for the platform node Autohome is platform_autohome.

[0055] When two nodes have more than 100 co-occurring samples in the past 30 days, a directed edge is generated between the two nodes, with the edge pointing from the information source node to the information receiving node, such as from the region node to the vehicle model node. The activation status of a node is determined by the presence of at least 10 valid samples in the past 30 days, and active nodes are updated daily.

[0056] The formula for constructing the propagation relationship graph is: ; in, For a moment Propagation relationship diagram; For a moment The topology node set includes all active vehicle models, structures, regions, scenarios, platforms, and demand nodes; For a moment The set of edges contains the edges that allow information propagation between all pairs of nodes.

[0057] The unique identifiers for nodes are generated as follows: vehicle type node identifier is car_vehicle_ID, structure node identifier is structure_structure_ID, region node identifier is city_administrative division code, scene node identifier is scene_scene_scene ID, platform node identifier is platform_platform ID, and demand node identifier is demand_demand_demand ID. The direction of edges is determined based on the causal relationship of information propagation; for example, if region affects vehicle type sales, then the edge points from the region node to the vehicle type node. The node activation status update process is as follows: every day at midnight, the number of valid samples for each node over the past 30 days is counted, and if the number of samples is greater than or equal to 10, it is marked as an active node.

[0058] S502: Within a specified time window, extract the mutation sample set associated with specific node pairs, and obtain the scene gain value after removing local sporadic noise by using the global mutation state ratio and anti-zero division regularization processing.

[0059] The specified time window refers to the continuous time interval used for statistical information dissemination. It is set to 1 day for short-term analysis needs, 7 days for medium-term analysis needs, and 30 days for long-term analysis needs. The mutation sample set refers to the set of all samples within the specified time window whose absolute value of scene mutation degree is greater than a preset threshold of 0.3. The same sample is counted only once to avoid duplicate calculations. The global sample set refers to the set of all valid samples within the specified time window. (Zero-prevention regularization parameter) The value is This is used to avoid calculation errors where the denominator is zero.

[0060] The formula for calculating the scene gain value is: ; in, For a moment Node pairs The scene gain value ranges from 0 to 1. The larger the value, the greater the impact of scene changes on the information propagation between nodes. and This represents a specific pair of nodes in the propagation relationship graph; For a moment Associated node pairs A set of mutation samples; For the sample The absolute value of the scene abruptness; For a moment The global sample set; To prevent division by zero regularization parameters.

[0061] The criteria for selecting the time window length are as follows: a 1-day time window is used to analyze the impact of sudden events, a 7-day time window is used to analyze weekly trends, and a 30-day time window is used to analyze monthly trends. The method for determining the threshold of scene mutation degree is to statistically analyze the scene mutation degree distribution of all historical mutation events and take 0.3 corresponding to the 95th quantile as the threshold. The deduplication rule for mutation sample set is to deduplicate based on user ID and timestamp, and for multiple comments from the same user within the same hour, only the one with the largest absolute value of scene mutation degree is retained.

[0062] S503: Determine the propagation weight by combining the information entropy sharing between node pairs, the temporal decay coefficient determined by the time distribution difference, and the comprehensive adjustment effect of the scene gain value.

[0063] Information entropy sharing refers to the mutual information between two nodes, used to measure the statistical correlation between them. It is calculated based on the co-occurrence frequency of nodes, and the specific calculation formula is as follows: The calculation results are normalized using min-max to map to a value range of 0 to 1. The time distribution feature vector refers to the daily posting frequency vector of the node over the past 30 days, with a dimension of 30. Covariance matrix. This is the covariance matrix of the time distribution characteristics over the past year, with dimensions 30×30. Mahalanobis distance is used to measure the difference between the time distributions of two nodes, and the calculation result is used to generate the time-series decay coefficient.

[0064] The formula for determining the propagation weight is: ; in, For a moment Node pairs The propagation weight, the larger the value, the more information spreads from the node. propagation to nodes The higher the strength; For a moment node With nodes The amount of information entropy shared between them; and They are nodes and nodes At any moment The time distribution feature vector; The Mahalanobis distance between two time distribution eigenvectors; Let be the covariance matrix of the time distribution; This is the timing decay coefficient; This represents the scene gain value. This is the overall adjustment effect term for the scene gain value.

[0065] Information entropy sharing is calculated based on the node co-occurrence frequency over the past 30 days. Co-occurrence frequency refers to the number of times two nodes appear simultaneously in the same sample divided by the total number of samples. The method for generating the time distribution feature vector is to count the number of samples of a node each day over the past 30 days and then divide it by the total number of samples over 30 days to obtain a normalized 30-dimensional vector. The covariance matrix is ​​calculated based on the time distribution feature vector of all active nodes over the past year.

[0066] S601: Based on specific vehicle models, regions, and demand categories, the scene mutation degree of historical events is recombined according to the temporal topology to construct a mutation sequence.

[0067] The specific vehicle model, region, and demand category combination refers to combining the vehicle model identifier, region identifier, and demand category identifier to form a unique identifier triplet. The demand category identifier has a value from 1 to 64, corresponding one-to-one with 64 demand categories. The timestamp precision of historical events is in hours to ensure the accuracy of the chronological order. The process of constructing the mutation sequence involves filtering out all historical events that match the triplet, then arranging them in ascending order of their occurrence time. Multiple events at the same time point are arranged in descending order of the absolute value of the scene mutation degree. Only one event from the same user at the same time is retained to avoid duplication.

[0068] The formula for constructing the mutant sequence is: ; in, For model ,area Demand Categories The corresponding mutation sequence; For historical events The time of occurrence; For historical events The degree of scene abruptness; For historical events Corresponding vehicle model identification; For historical events Corresponding regional identifier; For historical events The corresponding requirement category identifier.

[0069] The complete correspondence of the demand category identifiers is as follows: 1 corresponds to the technological feel of the appearance, 2 corresponds to the overall appearance, 3 corresponds to low temperature reliability, 4 corresponds to ice and snow adaptability, 5 corresponds to collision safety, and 6 corresponds to rescue convenience. The historical event timestamps are synchronized using the NTP network time protocol, with a synchronization error of no more than 1 second. The mutation sequence sorting rule is that for multiple events from the same user at the same time, only the one with the largest absolute value of scene mutation degree is retained.

[0070] S602: Extract the state autocorrelation features of the mutant sequence at the target time interval, perform global distribution smoothing to filter out time discontinuity interference, and dynamically generate an adaptive time decay kernel.

[0071] State autocorrelation refers to the Pearson correlation coefficient between the degree of scene mutation at different time points in a mutation sequence, used to measure the variation of the influence of historical mutation events on the current moment over time intervals. Global distribution smoothing employs a sliding window averaging method with a window size of 7 days, and Gaussian smoothing with a standard deviation of 1 to eliminate time discontinuity interference caused by missing data or outliers. Time interval The value range is from 1 to 365 days, with a step size of 1 day. Indicator function This means that a value of 1 is assigned when the difference between two times falls within ±0.5 days of the target time interval; otherwise, a value of 0 is assigned. After the time decay kernel is generated, normalization verification is performed to ensure that the sum of the weights of all time intervals is 1. If the error between the sum and 1 exceeds [a certain threshold], the value is set to 0. At that time, the weights are readjusted.

[0072] The formula for generating the time decay kernel is: ; in, For model ,area Demand Categories The corresponding time decay kernel in the time interval The value at this location ranges from 0 to 1, with larger values ​​indicating longer time intervals. The greater the impact of historical catastrophes on the present moment; The target time interval; For all possible time intervals; For indicator functions; and The degree of abrupt change in the scene of a historical event; and For the time of historical events; To prevent division by zero regularization, the value is set to [value to be filled in]. .

[0073] The specific formula for calculating the Pearson correlation coefficient is as follows: The sliding window averaging method is implemented by taking the arithmetic mean of the correlation coefficients of 7 time intervals from Δt-3 to Δt+3 for each time interval Δt. Gaussian smoothing uses a Gaussian kernel function with a standard deviation of 1 to convolve the correlation coefficient sequence. The time decay kernel normalization method is to divide the weights of all time intervals by the sum of the weights so that the sum equals 1.

[0074] S603: Utilizes the time decay kernel to perform forgetting mechanism regulation and cumulative fusion on historical mutation events before the current moment, and extracts the historical influence degree that reflects the long-tail lag effect.

[0075] The forgetting mechanism modulates the impact by assigning different weights to historical abrupt events at different time intervals based on the value of the time decay kernel, with events from more distant time intervals receiving smaller weights. Cumulative fusion involves summing the weighted impacts of all historical abrupt events to obtain the total historical impact. The time span for historical events is 365 days; events exceeding this span are ignored. When outliers appear in the historical impact, a 3-sigma truncation principle is applied: values ​​exceeding the mean plus three standard deviations are corrected to the mean plus three standard deviations, and values ​​below the mean minus three standard deviations are corrected to the mean minus three standard deviations.

[0076] The formula for calculating historical influence is as follows: ; in, For a moment Model ,area Demand Categories Corresponding historical impact; The current moment; For historical events The time of occurrence; For historical events The degree of scene abruptness; For time decay kernels in time intervals The value at; It is a mutated sequence.

[0077] The forgetting mechanism refers to the fact that the time decay kernel value decreases as the time interval increases, and the more distant the historical event, the lower the weight is assigned. The determination of the historical event time span of 365 days is based on the fact that experiments show that the impact of historical events older than 365 days on current demand is less than 0.01 and can be ignored. The 3 sigma principle is implemented by calculating the mean μ and standard deviation σ of all historical impact values, replacing values ​​greater than μ+3σ with μ+3σ, and replacing values ​​less than μ-3σ with μ-3σ.

[0078] S701: Combine structural features, scene features, scene abruptness, historical influence, and platform-sourced data into a tensor to form a comprehensive state.

[0079] Platform source data refers to the platform information on which user-expressed data is published. It is represented numerically using one-hot encoding, with each of the seven platforms corresponding to a seven-dimensional vector. For example, Autohome corresponds to... Dongchedi (a car-related platform) And so on. Tensor composition refers to concatenating feature vectors of different dimensions into a one-dimensional tensor in the order of structural features, scene features, scene mutability, historical influence, and platform source data. The structural features are 5-dimensional, the scene features are 4-dimensional, the scene mutability is 1-dimensional, the historical influence is 1-dimensional, and the platform source data is 7-dimensional, resulting in a total of 18 dimensions after concatenation. The tensor data type is 32-bit floating-point numbers.

[0080] The formula for calculating the overall state is as follows: ; in, For a moment The comprehensive state tensor; For model Corresponding structural features; For a moment Scene characteristics; For a moment The degree of scene abruptness; For a moment The extent of its historical influence; For a moment The one-hot encoded vector corresponding to the platform source data.

[0081] The complete encoding correspondence of the platform source data is as follows: Autohome corresponds to [1,0,0,0,0,0,0], Dongchedi corresponds to [0,1,0,0,0,0,0], Douyin corresponds to [0,0,1,0,0,0,0], Weibo corresponds to [0,0,0,1,0,0,0], Taobao corresponds to [0,0,0,0,1,0,0], JD.com corresponds to [0,0,0,0,0,1,0], and the customer service system corresponds to [0,0,0,0,0,0,1]. The tensor data type adopts 32-bit single-precision floating-point numbers according to the IEEE 754 standard.

[0082] S702: Based on the adjacency relationship established by each propagation weight using a graph neural network, spatial feature aggregation is performed to extract aggregated features.

[0083] Graph Neural Networks (GNNs) are neural network models specifically designed for processing graph-structured data, with a number of layers. The number of neurons in each layer is 2, with 64 neurons. The ReLU activation function is used, and residual connections are used between layers. The adjacency matrix is ​​constructed from propagation weights, with self-loops added, each with a weight of 1. Then, symmetric normalization is performed. ,in For degree matrix, This is the adjacency matrix. Feature extraction matrix. For the first The parameter matrix of the layers has a dimension of 18×64 for the first layer and 64×64 for the second layer. The optimizer for training this network uses the Adam optimization algorithm and has 10 training rounds.

[0084] The formula for calculating the extracted aggregation features is: ; in, For a moment The aggregated feature is a 64-dimensional vector; It is the ReLU activation function; The number of layers in the graph neural network; For the first The layer is based on a normalized adjacency matrix established by propagation weights; For the comprehensive state tensor; For the first Feature extraction matrix of layered graph neural network.

[0085] The specific type of graph neural network is a graph convolutional neural network (GCN). Residual connections are implemented by adding the output and input of each layer before passing them through an activation function. The symmetric normalization steps for the adjacency matrix are as follows: first, calculate the degree matrix D, where the diagonal elements of D represent the degrees of the corresponding nodes; then, calculate the inverse square root matrix of D; finally, obtain the normalized adjacency matrix. The training set for the graph neural network contains 1 million graph structure samples, and the optimizer parameters are the same as those for S302.

[0086] S703: Utilize the time attention mechanism to evaluate the temporal contribution of aggregated features of each historical period to the current time series prediction.

[0087] Temporal attention is a mechanism that automatically learns the importance of features from different historical moments to the current prediction task, where parameters... It is a 64-dimensional vector. It is a 64×64 matrix. It is a 1×64 matrix. The number of columns matches the dimension of the aggregated feature. The number of columns matches the dimension of historical influence. The maximum historical period is 30 days, meaning it considers historical data from the past 30 days. After the attention weights are generated, normalization validation is performed to ensure that the sum of the time-series contributions of all historical periods is 1. If the error between the sum and 1 exceeds a certain threshold... At that time, the weights are readjusted.

[0088] The formula for calculating the contribution of time series data is as follows: ; in, For a moment Distance from the current time The temporal contribution of the aggregated features of each historical period, with a value range from 0 to 1; and This represents the historical cycle offset. , and Parameters for evaluating time attention mechanisms; For a moment Aggregation characteristics; For a moment The historical influence; tanh is the hyperbolic tangent activation function.

[0089] The specific type of the time attention mechanism is an additive attention mechanism; all parameters are initialized uniformly using Xavier; the historical period refers to the time interval in units of 1 day, with the maximum historical period of 30 days corresponding to the data of the past 30 days; the attention weight normalization method is to divide the attention weights of all historical periods by the sum of the weights, so that the sum equals 1.

[0090] S704: Perform sequence feature reconstruction based on temporal contribution and map it to the future time dimension to output demand trend value.

[0091] Sequence feature reconstruction refers to the weighted summation of aggregated features from different historical periods based on their temporal contribution, resulting in reconstructed temporal features. Mapping to the future time dimension refers to converting the reconstructed temporal features into predicted demand trends for future time steps through a linear mapping, where the feature mapping matrix... Its dimensions are 64×1. The dimension is 1×1. The matrix has a 1×1 dimension, and the optimizer used to train these matrices employs the Adam optimization algorithm for 10 training epochs. The demand trend value represents user attention to the car door handle, ranging from 0 to 1. The future time step is set to 1 day for short-term forecasts, 7 days for medium-term forecasts, and 30 days for long-term forecasts. Multi-step forecasts use a direct forecasting method, simultaneously outputting forecast results for multiple time steps.

[0092] The formula for calculating the output demand trend value is: ; in, For the future Demand trend value; For future time dimension step; , and The feature mapping matrix; Contribution to time series; This represents the aggregation characteristics of historical cycles; The sequence features reconstructed based on temporal contribution. Historical impact; This represents the degree of scene mutability.

[0093] Sequence feature reconstruction refers to the weighted summation of aggregated features from different historical periods based on attention weights to obtain comprehensive time-series features; the demand trend value is the relative attention users pay to the door handle of this car model at future moments, with higher values ​​indicating higher user attention; multi-step prediction adopts a direct prediction method, training a model to simultaneously output prediction results for three time steps: 1 day, 7 days, and 30 days.

[0094] S801: Evaluate the gradient sensitivity of demand trend values ​​to a single physical configuration parameter, and combine the data discrete distribution characteristics of each physical configuration parameter to perform a normalized metric, and analyze the structural influence degree used to characterize the hardware configuration interference weight.

[0095] First, z-score normalization is performed on all physical configuration parameters to convert each parameter into a dimensionless normalized value. Specifically: ,in The mean of the parameters. Let be the standard deviation of the parameters. Gradient calculation uses an automatic differentiation method, implemented using the `autograd` module of the PyTorch framework, to calculate the partial derivatives of the demand trend value with respect to the standardized physical configuration parameters. The indices of the physical configuration parameters correspond one-to-one with the elements of the structural features: index 1 corresponds to outer surface flatness, index 2 to kinematic pair morphology, index 3 to electronic triggering logic, index 4 to mechanical redundancy reachability, and index 5 to the operational space scale. Zero-prevention regularization parameters are used. The value is This is used to avoid calculation errors where the denominator is zero.

[0096] The formula for calculating the structural influence is as follows: ; in, For the first The structural influence degree corresponding to each physical configuration parameter ranges from 0 to 1. The larger the value, the greater the influence of the physical configuration parameter on the demand trend value. This represents the demand trend value. For the standardized first One physical configuration parameter; For the demand trend value to the first Partial derivatives of a standardized physical configuration parameter; For the standardized first The variance of each physical configuration parameter takes a value of 1; For the index of each physical configuration parameter being traversed; To prevent division by zero regularization parameters.

[0097] In z-score standardization Let be the mean of the k-th physical configuration parameter across all training samples. The corresponding standard deviation is given; automatic differentiation is implemented through the autograd.backward() method of the PyTorch framework; the complete correspondence of the physical configuration parameter index is as follows: k=1 corresponds to the outer surface flatness, k=2 corresponds to the kinematic pair morphology, k=3 corresponds to the electronic triggering logic, k=4 corresponds to the mechanical redundancy accessibility, and k=5 corresponds to the operational space scale; the variance of the standardized parameters being 1 is an inherent property of z-score standardization.

[0098] S802: Compare the future demand trend value with the actual observation status collected synchronously, and extract the prediction error that characterizes the degree of prediction offset.

[0099] Actual observed state refers to the state at a future time. The frequency of corresponding demand categories in user expression data over the past 7 days was statistically analyzed, with values ​​ranging from 0 to 1. Mean squared error was used as the metric for prediction error; when the demand trend value was a multi-dimensional vector, the error for each dimension was calculated separately and then averaged.

[0100] The formula for calculating the prediction error is as follows:

[0101] in, For the future The prediction error; For the future The actual observation status acquired synchronously; For the future Demand trend value.

[0102] The actual observation state is calculated by dividing the number of times the corresponding demand category appears in all user expression data regarding the door handle of this vehicle model within 7 days before the future time t+Δ by the total number of user expressions during that time period; the specific formula for calculating the mean squared error is as follows. The method for calculating multidimensional prediction error is to take the arithmetic mean of the mean square error for each dimension.

[0103] S803: Based on prediction error and topological constraints, an objective optimization function is constructed. The network weights of the structural baseline and the time series model are iteratively corrected through a parameter optimization algorithm to complete the update of the time series model.

[0104] Topological constraint matrix The specific formula for calculating the graph regularization norm, which is used to propagate the Laplacian matrix of the association graph, is as follows: ,in Let be the Laplace matrix. The parameter optimization algorithm uses the Adam optimization algorithm, with iterative correction step size. The initial value is The gradient decays by 0.99 after each iteration. The gradient clipping threshold is 1.0 to prevent gradient explosion. Trainable parameter set. Includes structure mapping matrix Independent mapping matrix , , Coupled Interference Term Matrix Graph Neural Network Feature Extraction Matrix Attention mechanism parameters , , eigenmap matrix , , All trainable parameters.

[0105] The calculation formula for completing the time series model update is:

[0106] in, and These are the network weight parameter sets for the updated and unupdated temporal models and structural baselines, respectively. This is the iterative correction step size for the parameter optimization algorithm; Optimize the objective function with respect to the parameter set The partial derivative operator; This is the sum of squares of the prediction errors; These are topological constraint terms; For propagation-based association graph The graph regularization norm.

[0107] The formula for calculating the Laplace matrix is: ,in For degree matrix, The adjacency matrix is ​​used; the Adam optimization algorithm parameters are set to beta1=0.9, beta2=0.999, and epsilon=1e-8; the iterative correction step size decay strategy is to multiply by 0.99 after each iteration, with a minimum step size of 1e-6; gradient clipping is implemented by scaling the gradient to an L2 norm of 1.0 when the L2 norm exceeds 1.0; the trainable parameter set is used. It contains the weights and bias parameters of all neural network layers.

[0108] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A method for analyzing customer needs and predicting market trends in the automotive industry based on big data, applied to automotive configurations including hidden or electrically retractable door handles, characterized in that... include: Collect user expression data, vehicle configuration data, door handle structure data, regional and temporal data, scenario data, and platform-sourced data to construct an analytical sample representing the state of multi-source heterogeneous inputs; The physical configuration parameters of the door handle structure data are extracted to construct a structural benchmark, and the user expression data is aligned to the structural benchmark to eliminate semantic discrepancies in non-standard expressions; The user expression data is decoupled to extract the appearance demand degree representing the visual attribute direction, the safety demand degree representing the reliability of mechanical operation, and the scene change degree representing the environmental trigger state. Construct a propagation relationship graph containing demand nodes, and generate propagation weights that characterize the scale of cross-domain node diffusion by combining the shared information entropy of nodes, the difference in time distribution, and the abruptness of the scenario. Based on the mutation sequence composed of the scene mutation degree, a time decay kernel is generated to extract the historical influence degree across cycles; The comprehensive state, including the scenario mutation rate and the historical influence rate, is input into the time series model to obtain the demand trend value in the future time. The structural influence degree decomposition is performed on the demand trend value, and the time series model is updated based on the prediction error feedback determined by actual observation.

2. The method for analyzing customer demand and predicting market trends in the automotive industry based on big data as described in claim 1, characterized in that, Construct analytical samples to characterize the states of multi-source heterogeneous inputs, including: Based on the door handle structure data and vehicle configuration data, the outer surface flatness, kinematic pair shape, electronic triggering logic, mechanical redundancy accessibility and operating space scale are extracted to construct structural features that reflect physical interference properties. Using the regional time data and the scene data, corresponding low temperature, ice and snow, collision and rescue scene factors are extracted. Distribution unification and dimensionality uncorrelation processing are performed on each scene factor to eliminate the dimensional differences under different climatic conditions and construct scene features. The analysis sample is obtained by multimodal concatenation of the user expression data, the structural features, and the scene features.

3. The method for analyzing customer demand and predicting market trends in the automotive industry based on big data, as described in claim 2, is characterized in that... Aligning the user expression data to the structural baseline includes: Natural language encoding is performed on the user expression data to obtain text features; The structural features are spatially transformed by combining the structural mapping relationships learned during training, and the common features of the associated professional structural lexicon are fused to generate the structural benchmark. The similarity of the distribution of the text features and the structural benchmark in the same feature space is evaluated, and long-tail text interference is eliminated through normalization processing to obtain the text relevance characterizing the correlation strength of the corresponding entities, thus completing the soft alignment to the specific vehicle body physical hardware.

4. The method for analyzing customer demand and predicting market trends in the automotive industry based on big data, as described in claim 3, is characterized in that... Calculate the degree of appearance requirements, safety requirements, and the degree of scene abruptness, including: Cross-domain projection and feature fusion are performed on the text features, scene features and structural features using an independent mapping matrix, and the demand distribution value is obtained through nonlinear activation. The demand distribution values ​​are extracted along the preset visual attribute principal components and safety and availability principal components to obtain the appearance demand degree and the safety demand degree respectively. Extract the attribute offsets of the security requirement degree and the appearance requirement degree, and perform boundary constraint processing by fusing the coupling interference terms of the scene features and the structural features, and then combine the text relevance to determine the scene mutation degree.

5. The method for analyzing customer demand and predicting market trends in the automotive industry based on big data, as described in claim 4, is characterized in that... Generate propagation weights that characterize the scale of cross-domain node diffusion, including: Construct the propagation relationship graph with vehicle type, structure, region, scenario, platform, and the aforementioned demand node as the topological core; Within a specified time window, extract the mutation sample set associated with specific node pairs, and obtain the scene gain value after removing local sporadic noise by using the global mutation state ratio and anti-zero division regularization. The propagation weight is determined by combining the information entropy sharing between the node pairs, the temporal decay coefficient determined by the time distribution difference, and the comprehensive adjustment effect of the scene gain value.

6. The method for analyzing customer demand and predicting market trends in the automotive industry based on big data, as described in claim 5, is characterized in that... Generate a time decay kernel to extract historical influence across cycles, including: Based on a specific combination of vehicle type, region, and demand category, the mutation sequence is constructed by reorganizing the scene mutation degree of historical events according to the temporal topology; Extract the state autocorrelation features of the mutation sequence at the target time interval, and perform global distribution smoothing to filter out time discontinuity interference, and dynamically generate the adaptive time decay kernel; By utilizing the time decay kernel, the historical abrupt events prior to the current moment are subjected to a forgetting mechanism for regulation and cumulative fusion, and the historical influence degree reflecting the long-tail lag effect is extracted.

7. The method for analyzing customer demand and predicting market trends in the automotive industry based on big data, as described in claim 6, is characterized in that... The comprehensive state, including the scenario abruptness and the historical influence, is input into the time series model to obtain future demand trend values, including: The comprehensive state is constructed by combining the structural features, scene features, scene abruptness, historical influence, and platform-sourced data using tensors. Based on the adjacency relationships established by the propagation weights, spatial feature aggregation is performed using a graph neural network to extract aggregated features; The temporal contribution of the aggregated features of each historical period to the current time series prediction is evaluated using a time attention mechanism. Based on the time-series contribution, sequence features are reconstructed and mapped to the future time dimension to output the demand trend value.

8. The method for analyzing customer demand and predicting market trends in the automotive industry based on big data, as described in claim 7, is characterized in that... Perform structural impact decomposition on the demand trend value, and update the time series model based on the prediction error feedback determined by actual observations, including: The gradient sensitivity of the demand trend value to a single physical configuration parameter is evaluated, and a normalized measure is performed in combination with the data discrete distribution characteristics of each physical configuration parameter to analyze the structural influence degree used to characterize the hardware configuration interference weight. By comparing the demand trend value at the future time with the actual observation status collected synchronously, the prediction error, which characterizes the degree of prediction offset, is extracted. Based on the prediction error and topological constraints, an objective optimization function is constructed. The network weights of the structural baseline and the time series model are iteratively corrected through a parameter optimization algorithm to complete the update of the time series model.

Images