Intelligent furniture design system based on AI big data

Abstract

The invention discloses an intelligent furniture design system based on AI big data, and relates to the technical field of artificial intelligence, and the system comprises a data collection module which is used for collecting user behavior data, environment data and market data; the data processing module is used for carrying out cleaning, feature extraction and fusion on the collected data to generate structured design input parameters, and the structured design input parameters comprise user preference vectors and spatial constraint parameters; according to the intelligent furniture design system, accurate analysis of design requirements is achieved by comprehensively collecting user behaviors, environment and market data, the problem that design requirement analysis is incomplete due to the fact that a traditional data collection means is single is solved, the system adopts an advanced data processing technology, and the design requirement analysis accuracy is improved. Valuable design input parameters are effectively extracted and fused, a solid foundation is provided for personalized design, and a design generation and optimization module based on a deep learning model can automatically generate and optimize a furniture design scheme.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an intelligent furniture design system based on AI big data. Background Art

[0002] Against the backdrop of the rapid development of artificial intelligence and big data technologies, smart furniture design systems provide a new path for intelligent and personalized furniture design by integrating user behavior data, environmental parameters, and market trends. Traditional furniture design mainly relies on designer experience and static rules, making it difficult to dynamically adapt to changes in user needs and complex spatial constraints. With the popularization of the concept of smart homes, the market has placed higher demands on the functionality, space utilization, and user experience of furniture, and there is an urgent need to achieve digital transformation of the design process through data-driven methods. Although existing technologies can achieve basic parametric modeling, they have significant shortcomings in multi-source data fusion, in-depth mining of user preferences, and dynamic optimization of design solutions. Especially when responding to large-scale customization needs, the efficiency and accuracy of traditional methods can no longer meet the needs of industry development.

[0003] The core challenges currently facing the smart furniture design field include: First, traditional data collection methods are limited and cannot fully capture real-time data on user behavior, environmental characteristics, and market dynamics, resulting in incomplete design requirements analysis; second, the data processing process lacks effective feature extraction and fusion technologies, making it difficult to extract valuable design input parameters from massive amounts of data; third, the design generation process relies on preset templates and lacks the creative generation capabilities based on deep learning, resulting in severe homogeneity of solutions; fourth, the design optimization process fails to fully consider the dynamic balance between space utilization, cost constraints, and user preferences, making multi-objective optimization difficult; and fifth, the user feedback mechanism lags behind, preventing a closed-loop optimization process for design iterations, leading to deviations between products and actual user needs. Furthermore, existing systems lack collaboration in the manufacturing process, resulting in inefficient and error-prone conversion of design solutions into production drawings. These issues hinder the innovative development and large-scale application of the smart furniture industry. To address this, we propose a smart furniture design system based on AI big data. Summary of the Invention

[0004] In order to solve the above technical problems, an intelligent furniture design system based on AI big data is provided. This technical solution solves the above problems.

[0005] In order to achieve the above objects, the technical solution adopted by the present invention is:

[0006] Smart furniture design system based on AI big data, including:

[0007] The data collection module is used to collect user behavior data, environmental data and market data;

[0008] The data processing module is used to clean, extract and fuse the collected data to generate structured design input parameters, which include: user preference vector and space constraint parameters;

[0009] The AI ​​design module is used to generate and optimize furniture design solutions based on deep learning models;

[0010] The output module is used to convert the optimized design into a 3D model, manufacturing drawings, and bill of materials, and to achieve visual interactive verification through AR / VR devices;

[0011] The system returns user feedback and production data to the data processing module through a data closed loop to achieve iterative updates of the model.

[0012] Preferably, the data acquisition module collects user behavior data including: historical purchase records, browsing time, click hotspots, voice interaction logs and somatosensory data obtained through wearable devices; the data acquisition module collects environmental data including: 3D scanning data of room size, light intensity, temperature and humidity, and space layout obtained through sensors or measuring tools; the data acquisition module collects market data including: furniture sales trends, material price fluctuation data and user review sentiment analysis results.

[0013] Preferably, the method for the data collection module to collect market data is:

[0014] Among them, the method for obtaining furniture sales trends is:

[0015] Based on the sales database, historical sales data {S1, S2, ..., S t}, the future sales trend is predicted by the Holt-Winters three-exponential smoothing model, and the calculation formula is:

[0016] L t =αS t +(1-α)(L t-1 +T t-1 )

[0017] T t =β(L t -L t-1 )+(1-β)T t-1

[0018]

[0019] Where S t is the actual sales data of period t, L t is the horizontal component of period t, representing the basic sales volume after removing seasonality and trend, T tis the trend component of period t, indicating the slope of sales volume changing over time, I t is the seasonal component of the tth period, which represents the influencing factor of periodic fluctuations. α, β, γ∈[0,1] are smoothing coefficients, which control the update speed of the level, trend and seasonal components respectively. m is the length of the seasonal cycle, and k is the number of future periods to be predicted. is the predicted sales volume for period t+k;

[0020] The method for obtaining material price fluctuation data is as follows:

[0021] Collect historical price series {P1,P2,…,P n}, calculate the logarithmic rate of return:

[0022] r t =ln(P t / P t-1 )

[0023] Where r t is the logarithmic rate of return, P t is the material price in period t;

[0024] Model volatility using the GARCH(1,1) model based on the calculated log returns:

[0025]

[0026] Where, is the price volatility of period t, ω is a constant term, representing the long-term average volatility, which must satisfy ω>0, α is the coefficient of the squared residual term, measuring the impact of previous price fluctuations on the current one, and satisfying α≥0, β is the lagged volatility coefficient, measuring the persistence of previous volatility, and satisfying β≥0 and α+β<1;

[0027] Among them, the method for obtaining the user comment sentiment analysis results is:

[0028] Segment the comment text and remove stop words to construct the word vector {w1,w2,…,w m};

[0029] Calculate sentiment scores based on pre-trained sentiment lexicon:

[0030]

[0031] Where w i is the i-th word in the comment text, Polarity(w i ) is the word w i The sentiment polarity is {-1, 0, 1}, which represent negative, neutral, and positive respectively. Intensity(w i ) is the word wi The emotional intensity weight of , range [0,1];

[0032] The context dependency is modeled by the LSTM model, and the final sentiment classification probability p positive Output from the Softmax layer:

[0033]

[0034] Where h T is the hidden state vector of the last layer of the LSTM model, W c and b c is the classifier parameter matrix and bias term, corresponding to the emotion category c, p positive is the probability that the review is positive.

[0035] Preferably, in the data processing module, the user preference vector is calculated by analyzing the correlation between user behavior data and similar user groups through a collaborative filtering algorithm, and the spatial constraint parameters include: room length, width and height, door and window position coordinates and load-bearing wall distribution matrix.

[0036] Preferably, the data processing module performs cleaning, feature extraction and fusion on the collected data, specifically including:

[0037] Based on the collected data, the missing values ​​in the user behavior data are supplemented by linear interpolation. The interpolation formula is:

[0038]

[0039] Where x prev and x next is the value of the adjacent data points before and after the missing value, and t is the timestamp;

[0040] For outliers in environmental data, the box plot method is used to identify and eliminate them, specifically:

[0041] The environmental data set D={d1,d2,…,d n} Arrange in ascending order to form an ordered sequence D sorted =[d (1) ,d (2) ,…,d (n) ];

[0042] Quartile calculation is performed, where the first quartile Q1 is located at the value of 25% of the data position, and its calculation formula is:

[0043]

[0044] If n×0.25 is an integer, then k=n×0.25. If n×0.25 is not an integer, then

[0045] Where the third quartile Q3 is the value at the 75% position of the data, and its calculation formula is:

[0046]

[0047] If n×0.75 is an integer, then m = n×0.75; if n×0.75 is not an integer, then

[0048] Based on the calculated Q1 and Q3, calculate the interquartile range:

[0049] IQR = Q3 - Q1

[0050] In the formula, IQR represents the interquartile range;

[0051] Perform outlier determination. The lower bound is LowerBound = Q1 - 1.5×IQR, and the upper bound is UpperBound = Q3 + 1.5×IQR. If the data point d i satisfies d i < LowerBound or d i > UpperBound, then it is determined as an outlier. For the target determined as an outlier, delete all data points d that meet the above conditions from the data set D i , and obtain the cleaned data set D cleaned = {d j |d j ∈D and LowerBound ≤ d j ≤ UpperBound};

[0052] For duplicate records in the market data, use the hash deduplication algorithm for processing. The hash function is:

[0053] H(data) = SHA - 256(data)

[0054] For the said feature extraction, calculate the user preference vector through the collaborative filtering algorithm, and extract the spatial constraint parameters by quantifying the room size into a numerical combination of length L, width W, and height H;

[0055] Among them, the method for extracting the user preference vector feature is:

[0056] Construct a user - design element rating matrix R m×n , where m is the number of users and n is the number of design element categories;

[0057] Calculate the cosine similarity between the target user and other users, screen the top k similar user groups with the highest similarity, and generate the user preference vector P = [p1, p2,..., pn ] , where p iThe calculation formula is:

[0058]

[0059] Where p i is the user’s preference weight for the i-th type of design element, S(U,U j ) is the target user U and similar users U j The similarity of For similar users U j The preference value for the i-th type of design element, k is the number of similar users selected, For similar users U j The score of the i-th category design element is weighted by similarity, is the sum of the similarities between all similar users and the target user;

[0060] Among them, the method for extracting spatial constraint parameter features is:

[0061] The image recognition algorithm is used to locate the center coordinates of doors and windows (x d ,y d );

[0062] Construct the load-bearing wall distribution matrix M wall , the matrix element M wall (i, j) = 1 means there is a load-bearing wall at the coordinate (i, j), otherwise it is 0;

[0063] The feature fusion is achieved by concatenating the user preference vector and the spatial constraint parameter into an input vector, and its expression is:

[0064] X=[P;L;W;H;vec(M wall )]

[0065] Where X represents the fused input vector, which is used to input the deep learning model of the AI ​​design module, P represents the user preference vector, L; W; H represent the length, width, and height of the room, and M wall Represents the load-bearing wall distribution matrix, which is a two-dimensional binary matrix. The matrix element value 1 represents the wall position and 0 represents the passable area. vec(M wall ) represents the matrix M wall Flattened into a one-dimensional vector for input into the neural network.

[0066] Preferably, the AI ​​design module includes: a design generation submodule and an optimization submodule. The design generation submodule is used to adopt a generative adversarial network to receive user preference vectors and space constraint parameters to generate an initial furniture design sketch. The optimization submodule is used to dynamically adjust the design scheme through a reinforcement learning algorithm combined with space utilization, user preference matching and cost constraints.

[0067] Preferably, the design generation submodule generates an initial furniture design sketch specifically including:

[0068] The design generation submodule includes a generator and a discriminator. The network structure of the generator includes: an input layer, a hidden layer and an output layer, wherein the input layer receives the concatenated vector [P, L, W, H, M wall ], with a dimension of n+4+256, where 256 is the dimension of the flattened load-bearing wall matrix; the hidden layer contains 5 fully connected layers, with 512, 1024, 2048, 1024, and 512 neurons in each layer, and the activation function is LeakyReLU (α=0.2); the output layer generates a pixel matrix of 256×256×3 of the design drawing, and the activation function is Tanh;

[0069] The network structure of the discriminator includes: an input layer, a convolutional layer, and an output layer, wherein the input layer receives the generated design image; the convolutional layer is a 4-layer convolution, the number of filters is 64, 128, 256, 512, the step size is 2, the padding is 1, and the activation function is LeakyReLU (α=0.2); the fully connected layer in the output layer outputs the discrimination probability p∈[0,1], and the activation function is Sigmoid;

[0070] The input of the generator is (P, L, W, H), the output is the furniture design drawing, and the loss function of the discriminator is:

[0071]

[0072] Where, L GAN is the overall loss value of the generated adversarial network, D(D real ) is the discriminator's response to the real design graph D real The discriminant probability of G(z) is the furniture design generated by the generator based on the noise vector z, and z is the noise vector input to the generator.

[0073] Preferably, the optimization submodule is used to dynamically adjust the design scheme by combining space utilization, user preference matching and cost constraints through a reinforcement learning algorithm, specifically including:

[0074] The optimization criteria of the reinforcement learning algorithm are determined by the reward function, which is expressed as:

[0075] R=α·U space +β·S p -γ·C

[0076] Where R is the reward function, U space is the space utilization rate, S p is the user preference matching degree, C is the cost;

[0077] The space utilization calculation formula is:

[0078]

[0079] Where V i is the volume of furniture;

[0080] The calculation formula for user preference matching is:

[0081]

[0082] Where P is the user preference vector, and D is the design scheme feature vector;

[0083] The cost calculation formula is:

[0084]

[0085] Where c k is the unit price of the kth material, q k is the dosage, t cut is the processing time, r machine is the equipment rate;

[0086] The reinforcement learning algorithm of the optimization submodule adopts proximal strategy optimization, and its objective function is:

[0087]

[0088] Where, L CLUP (θ) is the clipping loss function optimized by the proximal strategy, r t (θ) is the probability ratio of the new and old strategies, is the advantage function, representing action a t The advantage relative to the average level is ∈, which is the clipping range hyperparameter.

[0089] Preferably, the 3D modeling process of the output module includes: parametric modeling, physical simulation verification and manufacturing drawing generation;

[0090] The parametric modeling method is:

[0091] The design drawing D is parsed into a B-rep (boundary representation) model. The key parameters include: the coordinates of the curve control points (x i ,y i ,z i ); Material map index T ID ;Connector type and degrees of freedom DOF;

[0092] The physical simulation verification method is:

[0093] The stress distribution σ of the furniture structure is calculated by finite element analysis (FEA) to meet the following requirements:

[0094] σ max ≤0.8·σ yield

[0095] where σ yield is the yield strength of the material;

[0096] The method for generating manufacturing drawings is as follows:

[0097] Convert 3D models into CNC machining code, including:

[0098] Cutting path planning, optimizing tool path based on the Traveling Salesman Problem (TSP) to minimize processing time T cut

[0099]

[0100] where d i is the length of the i-th path, v is the tool speed, t pause It is the tool change pause time.

[0101] Preferably, the system further includes a user feedback closed-loop mechanism:

[0102] Real-time scoring interface: users score the key indicators of the generated solution S user =[s1,s2,s3], including aesthetics s1, functionality s2, and cost-effectiveness s3;

[0103] Feedback reinforcement learning, S user As a reward signal to update the PPO policy network, the weight adjustment formula is:

[0104] β new =β old +η·(s2-5)

[0105] γ new =γ old +η·(8-s3)

[0106] Where η = 0.05 is the learning rate;

[0107] Data is returned, and the data of the solution finally selected by the user is stored in the database and used to update the user similarity matrix of the collaborative filtering model.

[0108] Compared with the prior art, the present invention has the following beneficial effects:

[0109] The intelligent furniture design system proposed in the present invention realizes accurate analysis of design requirements by comprehensively collecting user behavior, environment and market data, and solves the problem of incomplete design requirement analysis caused by the single traditional data collection method. The system adopts advanced data processing technology to effectively extract and integrate valuable design input parameters, providing a solid foundation for personalized design. The design generation and optimization module based on the deep learning model can automatically generate and optimize furniture design schemes, significantly improving the creativity and practicality of the design, and avoiding the homogenization of schemes caused by reliance on preset templates. The system fully considers the dynamic balance of space utilization, cost constraints and user preferences, realizes multi-objective optimization, and improves the overall quality of the design. Through parametric modeling and physical simulation verification, the manufacturability and practicality of the design scheme are ensured. The introduction of the user feedback closed-loop mechanism enables the system to receive user feedback in real time, continuously optimize the design strategy, realize closed-loop optimization of design iteration, and ensure the fit between the product and the actual needs of users. BRIEF DESCRIPTION OF THE DRAWINGS

[0110] Figure 1 It is a system module diagram of the present invention. DETAILED DESCRIPTION

[0111] The following description is intended to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are merely examples, and those skilled in the art may conceive of other obvious variations.

[0112] Reference Figure 1 As shown in the figure, the intelligent furniture design system based on AI big data includes: data acquisition module, data processing module, AI design module and output module.

[0113] The specific implementation of the present invention is as follows: During system operation, the system first comprehensively acquires design input parameters through a multi-source data acquisition system. Data acquisition modules deployed on user terminals and servers utilize IoT sensors to capture user behavior data in real time, including posture data recorded by wearable devices, interaction logs of smart home systems, and historical behavior traces on shopping platforms. Environmental data is collected using a spatial scanning device consisting of a laser rangefinder, a temperature and humidity sensor array, and an RGB-D camera, generating three-dimensional point cloud data with centimeter-level accuracy. A market trend analysis module integrates with the e-commerce platform API to dynamically capture sales data and user reviews, and uses time series analysis models to predict future market demand fluctuations.

[0114] After the collected raw data enters the data processing module, it undergoes multi-dimensional data cleaning. Missing time series values ​​in user behavior data are repaired using a timestamp-based linear interpolation algorithm to ensure data continuity. Outliers in environmental data are automatically identified and removed using the interquartile range detection method. Spatial scanning data is coordinate-normalized to establish a standardized three-dimensional coordinate system. Market data is deduplicated using hash fingerprinting technology, and semantic parsing is performed on unstructured review text to extract sentiment polarity features.

[0115] During the feature extraction phase, a hybrid collaborative filtering algorithm was used to construct a user preference model. Matrix decomposition techniques were used to explore users' potential interests, decomposing design elements into feature subspaces such as material, shape, and function. The weighted similarity between the target user and similar groups was then calculated. Spatial constraint parameters were extracted from 3D point cloud data using a convolutional neural network. The building's structural features were automatically identified using door and window locations and the distribution of load-bearing walls, generating a binary space occupancy matrix. The feature fusion process involved tensor concatenation of the user preference vector and the spatial constraint parameters to form a composite feature vector with semantic associations.

[0116] The AI ​​design module uses a deep generative model to achieve creative design. The generator of the generative adversarial network uses a residual connection structure to improve feature propagation efficiency. A five-layer fully connected network maps the input vector into a high-resolution design sketch. The discriminator uses a progressive convolutional architecture to gradually extract image features. Through an adversarial training mechanism, the generator is guided to output design solutions that meet aesthetic standards. The reinforcement learning optimizer deploys a proximal policy optimization algorithm and establishes a multi-objective reward function to evaluate space utilization, cost-effectiveness, and style matching. It simulates furniture placement in a virtual environment and dynamically adjusts the design solution until the overall optimization goals are met.

[0117] The design is converted into a boundary representation model using a parametric modeling engine. Finite element analysis verifies structural strength and automatically annotates material properties and connection node information. The manufacturing drawing generation module integrates a computer-aided manufacturing system, discretizing the 3D model into a set of CNC machining instructions. An ant colony algorithm optimizes tool motion trajectories and generates a highly efficient machining path plan. An augmented reality verification system projects a virtual furniture model through a head-mounted device, allowing users to interactively navigate the space and conduct functional testing, collecting operational feedback data in real time.

[0118] User feedback data enters the model update pipeline via a distributed message queue, building an online learning mechanism to continuously optimize design strategies. Rating data serves as a reinforcement signal and is injected into the strategy network, where network weight parameters are adjusted using a gradient descent algorithm. Success case data flows back into the collaborative filtering model, updating the user similarity matrix to improve recommendation accuracy. The system establishes a version control mechanism to manage model iterations and ensure traceability and stability of the update process. The entire implementation process forms a complete closed loop from data collection to design generation and production verification, enabling intelligent and adaptive optimization of the furniture design process.

[0119] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions merely illustrate the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. The intelligent furniture design system based on AI big data is characterized by: include: The data collection module is used to collect user behavior data, environmental data and market data; The data processing module is used to clean, extract and fuse the collected data to generate structured design input parameters, which include: user preference vector and space constraint parameters; The AI ​​design module is used to generate and optimize furniture design solutions based on deep learning models; The output module is used to convert the optimized design into a 3D model, manufacturing drawings, and bill of materials, and to achieve visual interactive verification through AR / VR devices; The system returns user feedback and production data to the data processing module through a data closed loop to achieve iterative updates of the model.

2. The intelligent furniture design system based on AI big data according to claim 1 is characterized by: The data acquisition module collects user behavior data including: historical purchase records, browsing time, click hotspots, voice interaction logs and somatosensory data obtained through wearable devices; the data acquisition module collects environmental data including: 3D scanning data of room size, light intensity, temperature and humidity, and space layout obtained through sensors or measuring tools; the data acquisition module collects market data including: furniture sales trends, material price fluctuation data and user review sentiment analysis results.

3. The intelligent furniture design system based on AI big data according to claim 2 is characterized in that: The method for the data acquisition module to collect market data is: Among them, the method for obtaining furniture sales trends is: Based on the sales database, historical sales data {S1, S2, ..., S t }, the future sales trend is predicted by the Holt-Winters three-exponential smoothing model, and the calculation formula is: L t =αS t +(1-β)(L t-1 +T t-1 ) T t =β(L t -L t-1 )+(1-β)T t-1 Where S t is the actual sales data of period t, L t is the horizontal component of period t, representing the basic sales volume after removing seasonality and trend, T t is the trend component of period t, indicating the slope of sales volume changing over time, I t is the seasonal component of the tth period, which represents the influencing factor of periodic fluctuations. α, β, γ∈[0,1] are smoothing coefficients, which control the update speed of the level, trend and seasonal components respectively. m is the length of the seasonal cycle, and k is the number of future periods to be predicted. is the predicted sales volume for period t+k; The method for obtaining material price fluctuation data is as follows: Collect historical price series {P1,P2,…,P n }, calculate the logarithmic rate of return: r t =ln(P t / P t-1 ) Where r t is the logarithmic rate of return, P t is the material price in period t; Model volatility using the GARCH(1,1) model based on the calculated log returns: Where, is the price volatility of period t, ω is a constant term, representing the long-term average volatility, which must satisfy ω>0, α is the coefficient of the squared residual term, measuring the impact of previous price fluctuations on the current one, and satisfying α≥0, β is the lagged volatility coefficient, measuring the persistence of previous volatility, and satisfying β≥0 and α+β<1; Among them, the method for obtaining the user comment sentiment analysis results is: Segment the comment text and remove stop words to construct the word vector {w1,w2,…,w m }; Calculate sentiment scores based on pre-trained sentiment lexicon: Where w i is the i-th word in the comment text, Polarity(w i ) is the word w i The sentiment polarity is {-1, 0, 1}, which represent negative, neutral, and positive respectively. Intensity(w i ) is the word w i The emotional intensity weight of , range [0,1]; The context dependency is modeled by the LSTM model, and the final sentiment classification probability p positive Output from the Softmax layer: Where h T is the hidden state vector of the last layer of the LSTM model, W c and b c is the classifier parameter matrix and bias term, corresponding to the emotion category c, p positive is the probability that the review is positive.

4. The intelligent furniture design system based on AI big data according to claim 1 is characterized in that: In the data processing module, the user preference vector is calculated by analyzing the correlation between user behavior data and similar user groups through a collaborative filtering algorithm. The spatial constraint parameters include: room length, width and height, door and window position coordinates and load-bearing wall distribution matrix.

5. The intelligent furniture design system based on AI big data according to claim 4 is characterized in that: The data processing module performs cleaning, feature extraction and fusion of the collected data, specifically including: Based on the collected data, the missing values ​​in the user behavior data are supplemented by linear interpolation. The interpolation formula is: Where x prev and x next is the value of the adjacent data points before and after the missing value, and t is the timestamp; For outliers in environmental data, the box plot method is used to identify and eliminate them, specifically: The environmental data set D={d1,d2,…,d n } Arrange in ascending order to form an ordered sequence D sorted =[d (1) ,d (2) ,…,d (n) ]; Quartile calculation is performed, where the first quartile Q1 is located at the value of 25% of the data position, and its calculation formula is: If n×0.25 is an integer, then k=n×0.

25. If n×0.25 is not an integer, then The third quartile Q3 is located at the 75% position of the data, and its calculation formula is: If n×0.75 is an integer, then m=n×0.

75. If n×0.75 is not an integer, then Calculate the interquartile range based on the calculated Q1 and Q3: IQR=Q3-Q1 In the formula, IQR represents interquartile range; Perform outlier determination. The lower bound is LowerBound = Q1 - 1.5 × IQR, and the upper bound is UpperBound = Q3 + 1.5 × IQR. If the data point d i satisfies d i < LowerBound or d i > UpperBound, it is determined as an outlier. For the target determined as an outlier, all data points d i satisfying the above conditions are deleted from the data set D cleaned to obtain the cleaned data set D j ={d j |d j ∈ D and LowerBound ≤ d m×n ≤ UpperBound}; The duplicate records in the market data are processed using a hash deduplication algorithm. The hash function is: H(data)=SHA-256(data) The feature extraction calculates the user preference vector through collaborative filtering algorithm and extracts the space constraint parameters by quantifying the room size into a numerical combination of length L, width W and height H; Among them, the method of extracting user preference vector features is: Constructing the user-design element rating matrix R m×n , where m is the number of users and n is the number of design element categories; Calculate the cosine similarity between the target user and other users, filter the top k similar user groups, and generate the user preference quantity P = [p1, p2, ..., p n ], where p i The calculation formula is: Where p i is the user’s preference weight for the i-th type of design element, S(U,U j ) is the target user U and similar users U j The similarity of For similar users U j The preference value for the i-th type of design element, k is the number of similar users selected, For similar users U j The score of the i-th category design element is weighted by similarity, is the sum of the similarities between all similar users and the target user; Among them, the method for extracting spatial constraint parameter features is: The center coordinates of doors and windows (x d ,y d ); Construct the load-bearing wall distribution matrix M wall , the matrix element M wall (i, j) = 1 means there is a load-bearing wall at the coordinate (i, j), otherwise it is 0; The feature fusion is achieved by concatenating the user preference vector and the spatial constraint parameter into an input vector, and its expression is: X=[P;L;W;H;vec(M wall )] Where X represents the fused input vector, which is used to input the deep learning model of the AI ​​design module, P represents the user preference vector, L; W; H represent the length, width, and height of the room, and M wall Represents the load-bearing wall distribution matrix, which is a two-dimensional binary matrix. The matrix element value 1 represents the wall position and 0 represents the passable area. vec(M wall ) represents the matrix M wall Flattened into a one-dimensional vector for input into the neural network.

6. The intelligent furniture design system based on AI big data according to claim 1 is characterized in that: The AI ​​design module includes: a design generation submodule and an optimization submodule. The design generation submodule is used to use a generative adversarial network to receive user preference vectors and space constraint parameters to generate an initial furniture design sketch. The optimization submodule is used to dynamically adjust the design scheme through a reinforcement learning algorithm combined with space utilization, user preference matching and cost constraints.

7. The intelligent furniture design system based on AI big data according to claim 6 is characterized in that: The design generation submodule generates an initial furniture design sketch specifically including: The design generation submodule includes a generator and a discriminator. The network structure of the generator includes: an input layer, a hidden layer and an output layer, wherein the input layer receives the concatenated vector [P, L, W, H, M wall ], with a dimension of n+4+256, where 256 is the dimension of the flattened load-bearing wall matrix; the hidden layer contains 5 fully connected layers, with 512, 1024, 2048, 1024, and 512 neurons in each layer, and the activation function is LeakyReLU (α=0.2); the output layer generates a pixel matrix of 256×256×3 of the design drawing, and the activation function is Tanh; The network structure of the discriminator includes: an input layer, a convolutional layer, and an output layer, wherein the input layer receives the generated design image; the convolutional layer is a 4-layer convolution, the number of filters is 64, 128, 256, 512, the step size is 2, the padding is 1, and the activation function is LeakyReLU (α=0.2); the fully connected layer in the output layer outputs the discrimination probability p∈[0,1], and the activation function is Sigmoid; The input of the generator is (P, L, W, H), the output is the furniture design drawing, and the loss function of the discriminator is: Where, L GAN is the overall loss value of the generated adversarial network, D(D real ) is the discriminator's response to the real design graph D real The discriminant probability of G(z) is the furniture design generated by the generator based on the noise vector z, and z is the noise vector input to the generator.

8. The intelligent furniture design system based on AI big data according to claim 6 is characterized in that: The optimization submodule is used to dynamically adjust the design scheme by combining space utilization, user preference matching and cost constraints through a reinforcement learning algorithm, specifically including: The optimization criteria of the reinforcement learning algorithm are determined by the reward function, which is expressed as: R=α·U space +β·S p -γ·C Where R is the reward function, U space is the space utilization rate, S p is the user preference matching degree, C is the cost; The space utilization calculation formula is: Where V i is the volume of furniture; The calculation formula for user preference matching is: Where P is the user preference vector, and D is the design scheme feature vector; The cost calculation formula is: Where c k is the unit price of the kth material, q k is the dosage, t cut is the processing time, r machine is the equipment rate; The reinforcement learning algorithm of the optimization submodule adopts proximal strategy optimization, and its objective function is: Where, L CLIP (θ) is the clipping loss function optimized by the proximal strategy, r t (θ) is the probability ratio of the new and old strategies, is the advantage function, representing action a t The advantage relative to the average level is ∈, which is the clipping range hyperparameter.

9. The intelligent furniture design system based on AI big data according to claim 1 is characterized in that: The 3D modeling process of the output module includes: parametric modeling, physical simulation verification and manufacturing drawing generation; The parametric modeling method is: The design drawing D is parsed into a B-rep (boundary representation) model. The key parameters include: the coordinates of the curve control points (x i ,y i ,z i ); Material map index T ID ;Connector type and degrees of freedom DOF; The physical simulation verification method is: The stress distribution σ of the furniture structure is calculated by finite element analysis (FEA) to meet the following requirements: s max ≤0.8·s yield where σ yield is the yield strength of the material; The method for generating manufacturing drawings is as follows: Convert 3D models into CNC machining code, including: Cutting path planning, optimizing tool path based on the Traveling Salesman Problem (TSP) to minimize processing time T cut where d i is the length of the i-th path, v is the tool speed, t pause It is the tool change pause time.

10. The intelligent furniture design system based on AI big data according to claim 1 is characterized in that: The system also includes a user feedback closed loop mechanism: Real-time scoring interface: users score the key indicators of the generated solution S user =[s1,s2,s3], including aesthetics s1, functionality s2, and cost-effectiveness s3; Feedback reinforcement learning, S user As the reward signal updates the PPO policy network, the weight adjustment formula is: b new =b old +η·(s2-5) c new =c old +η·(8-s3) Where η = 0.05 is the learning rate; Data reflux: The user's final selected solution data is stored in the database and used to update the user similarity matrix of the collaborative filtering model.

Images