A customized furniture demand analysis method and system based on uncertainty modeling

Abstract

The application discloses a customized furniture wardrobe demand analysis method based on uncertainty modeling and belongs to the field of business information and communication technology. The method obtains user characteristic data and determines the prior probability distribution of corresponding customized furniture wardrobe design elements based on a pre-constructed characteristic-demand mapping library, converts the fuzzy demand description input by the user into an observation probability distribution, adopts Bayesian fusion to calculate the posterior probability distribution, starts the interactive process to obtain user feedback and iteratively update the posterior distribution when the interactive starting condition is met, and finally outputs a fusion demand vector containing a probability characteristic to generate a wardrobe design scheme. The application quantifies the demand uncertainty through probability modeling, dynamically clarifies the fuzzy demand through the interactive closed loop, and significantly improves the accuracy of demand analysis and design efficiency.

Description

Technical Field

[0001] This invention relates to the field of business information and communication technology, specifically to a method and system for analyzing customized furniture requirements based on uncertainty modeling. Background Technology

[0002] With the upgrading of consumption and the growth of personalized demands, customized wardrobe furniture is becoming increasingly popular because it can better meet users' spatial adaptation needs and aesthetic preferences. Currently, the customized furniture industry is developing from standardized finished products to whole-house customization and intelligent design, with users placing increasingly refined demands on the functional layout, size adaptation, and material selection of wardrobes. However, a common problem in the customized wardrobe design process is that designers mainly rely on personal experience to interpret user needs, and the quality of the design scheme is highly dependent on the designer's professional level, resulting in unstable design output and severe homogenization.

[0003] Existing technologies face significant technical bottlenecks in addressing user needs: On the one hand, users often express their needs using vague terms like "convenient" or "more advanced," which inherently contain uncertainty, and traditional methods lack quantification tools for this uncertainty; on the other hand, when design solutions deviate from user expectations, existing systems cannot proactively identify conflicting needs and provide targeted clarification, but can only passively wait for user feedback or make repeated modifications, resulting in long design cycles and low user satisfaction. Therefore, how to scientifically quantify the uncertainty in vague user needs and dynamically clarify needs when misunderstandings arise has become a pressing technical problem to be solved in the field of customized wardrobe design. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, this invention provides a demand analysis solution for customized furniture.

[0005] According to a first aspect of the present invention, a method for analyzing customized furniture demand based on uncertainty modeling is proposed, comprising the following steps: S1: Obtain user feature data and, based on a pre-built feature-demand mapping library, determine the prior probability distribution of at least one design element corresponding to the user feature data; the feature-demand mapping library is constructed based on correlation analysis of historical data of customized furniture design. S2: Obtain the fuzzy requirement description input by the user and transform the fuzzy requirement description into the observation probability distribution of at least one design element; S3: For each design element, the prior probability distribution and the observed probability distribution are fused to calculate the posterior probability distribution of the design element; S4: For custom furniture wardrobe design elements that meet the conditions for interactive initiation, initiate the interactive process to obtain user feedback; S5: Transform user feedback into observation data and update the posterior probability distribution of design elements based on the observation data; S6: Determine whether the iteration termination condition is met. If not, return to step S4 and re-evaluate the interaction initiation condition based on the updated posterior probability distribution. If the condition is met, proceed to the next step. S7: Outputs a fused demand vector containing the posterior probability distribution features of design elements. The fused demand vector is used to generate a custom furniture wardrobe design scheme.

[0006] According to some embodiments, in the method of the first aspect of the present invention, the interactive initiation conditions include uncertainty conditions based on posterior probability distribution evaluation and demand conflict conditions based on prior probability distribution and observed probability distribution detection.

[0007] According to some embodiments, in the method of the first aspect of the present invention, the process of determining the uncertainty condition specifically includes: calculating an uncertainty index based on the degree of dispersion of the posterior probability distribution; when the uncertainty index is greater than a first threshold, it is determined that the uncertainty condition is met. The process of determining the condition of demand conflict includes calculating the conflict index, which is determined based on the degree of difference between the expected value of the observed probability distribution and the expected value of the prior probability distribution, as well as the degree of dispersion of the prior probability distribution; when the conflict index exceeds the second threshold, it is determined that there is a demand conflict.

[0008] According to some embodiments, in the method of the first aspect of the present invention, step S4, initiating the interaction process to obtain user feedback includes executing a corresponding interaction processing flow according to the type of interaction initiation condition, specifically including: When the interaction initiation condition is a requirement conflict condition, the interaction processing flow SA is executed: SA1: Diagnose the cause of conflict by matching the most likely conflict type from a pre-built conflict cause knowledge base; SA2: Select the corresponding follow-up question template from the preset follow-up question template library based on the matched conflict type; SA3: Generate follow-up questions based on the follow-up question template and output them to the user; SA4: Receive user feedback on follow-up questions; When the interaction initiation condition is uncertain, the interaction processing flow SB is executed: SB1: Identify at least one design element that satisfies the uncertainty condition; SB2: Generate clarification questions for design elements and output them to the user; SB3: Receive user feedback on clarification issues.

[0009] According to some embodiments, in the method of the first aspect of the present invention, step SA3 further includes: The follow-up question template is dynamically populated with current user data and design element data to generate personalized natural language follow-up questions; The follow-up questions are in multiple-choice format, with options pre-set based on the type of conflict. In cases where multiple conflicting dimensions exist simultaneously, the order of follow-up questions is determined according to a preset priority rule.

[0010] According to some embodiments, in the method of the first aspect of the present invention, step S2, converting the fuzzy demand description into an observation probability distribution of at least one design element, further includes: For natural language text input, a semantic parsing model is used to identify core requirement words and their modification strength, and then convert them into corresponding probability distribution parameters according to the preset semantic-distribution mapping rules; For image input, an image recognition model is used to extract style features, material features, and layout features from the image, and the features are matched with the design element library to generate the corresponding observation probability distribution.

[0011] According to some embodiments, in the method of the first aspect of the present invention, the semantic-distribution mapping rule includes a fuzzy quantifier mapping table and an emotion intensity mapping function; The fuzzy quantifier mapping table is used to map fuzzy quantifiers to the expected value offsets of probability distributions; The emotion intensity mapping function is used to map the intensity of tone to a discrete parameter of a probability distribution.

[0012] According to some embodiments, in the method of the first aspect of the present invention, step S3, fusing the prior probability distribution and the observed probability distribution to calculate the posterior probability distribution, further includes: Select the corresponding conjugate prior distribution family based on the data type of the design element; By utilizing the property of conjugate priors, a closed-form expression for the posterior probability distribution is obtained through analytical calculation; For cases where conjugate priors cannot be used, numerical sampling methods are employed to approximate the posterior probability distribution.

[0013] According to some embodiments, in the method of the first aspect of the present invention, the feature-demand mapping library is constructed by performing correlation analysis on historical design data in advance, and the correlation analysis includes grey correlation analysis and cross analysis; based on the results of the correlation analysis, the statistical correlation law between user features and design elements is fitted to the parameters of the prior probability distribution.

[0014] According to a second aspect of the present invention, a customized furniture demand analysis system based on uncertainty modeling is proposed, comprising: The prior loading module is used to acquire user feature data and, based on a pre-built feature-requirement mapping library, determine the prior probability distribution of at least one design element corresponding to the user feature data. The input parsing module is used to obtain the fuzzy requirement description input by the user and transform the fuzzy requirement description into the observation probability distribution of at least one design element; The fusion analysis module is used to fuse the prior probability distribution and the observed probability distribution for each design element to calculate the posterior probability distribution of the design element. The user interaction module is used to initiate the interaction process and obtain user feedback for custom furniture wardrobe design elements that meet the interaction initiation conditions. The feedback update module converts user feedback into observation data and updates the posterior probability distribution of design elements based on the observation data. The iteration judgment module determines whether the iteration termination condition is met. If not, it returns to the user interaction module, which re-evaluates the interaction initiation condition based on the updated posterior probability distribution. If the condition is met, it enters the output module. The output module is used to output a fused demand vector containing the posterior probability distribution features of design elements. The fused demand vector is used to generate a custom furniture wardrobe design scheme.

[0015] The present invention has the following beneficial effects: 1. In response to the problem that traditional customized design often treats vague user needs as deterministic keywords or simple weights, which cannot quantify the inherent uncertainty of the needs themselves, this invention maps user features to a prior probability distribution, transforms vague need descriptions into observed probability distributions, and expresses the complete statistical characteristics of user needs in the form of probability distributions. This achieves the scientific quantification of the uncertainty of vague user needs and provides a computable mathematical basis for subsequent integration and interaction.

[0016] 2. To address the problem that existing technologies separate user features from direct user input and fail to organically integrate the two, this invention uses Bayesian inference to fuse prior probability distributions with observed probability distributions, calculating a posterior probability distribution that simultaneously contains group patterns and individual information. This achieves effective synergy between group statistical patterns and individual fuzzy expressions, making the demand analysis results both statistically reliable and individually targeted.

[0017] 3. To address the problem that traditional systems cannot determine when to ask users for clarification, resulting in either blindly interacting and disturbing users or missing crucial clarification opportunities, this invention sets dual interaction initiation conditions—an uncertainty threshold based on the degree of dispersion of the posterior distribution and a conflict detection threshold based on the difference between prior and observation—to achieve precise triggering of interaction timing. Interaction is initiated only when clarification is truly needed, balancing user experience with the accuracy of demand analysis. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an embodiment 1000 of the custom furniture demand analysis method based on uncertainty modeling of the present invention. Figure 2 for Figure 1 A flowchart illustrating step S2 of Example 1000; Figure 3 for Figure 1 A flowchart illustrating step S3 of Example 1000; Figure 4 for Figure 1 A flowchart illustrating the interactive processing flow SA and the interactive interaction flow SB in step S4 of Example 1000; Figure 5 for Figure 4 A flowchart illustrating sub-step SA3 of the interactive processing flow SA; Figure 6 A schematic diagram of an embodiment 2000 of the custom furniture demand analysis system based on uncertainty modeling of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Figure 1 This is a flowchart illustrating an embodiment 1000 of the custom furniture demand analysis method based on uncertainty modeling according to the present invention. Figure 1 As shown, Example 1000 includes steps S1-S6.

[0021] In step S1, a customized furniture demand analysis system based on uncertainty modeling (hereinafter referred to as the system) acquires user characteristic data and determines the prior probability distribution of at least one design element corresponding to the user characteristic data based on a pre-built feature-demand mapping library.

[0022] The core function of step S1 is to transform statistical patterns at the user group level into personalized prior knowledge for the current user. In the Bayesian framework, the prior distribution represents a reasonable guess made by the system based on user characteristics regarding the user's potential needs before seeing the user's specific expressed needs. This step solves the cold start problem, allowing the system to provide an evidence-based initial judgment based on the user's background characteristics even when the user has just entered the system and has not yet expressed specific needs.

[0023] In some specific embodiments, step S1 is based on the following core assumption: user groups with similar characteristics exhibit statistically predictable patterns in their wardrobe design needs. For example, statistical analysis might show that "users with master's degrees need an average of 3-4 drawers" and "families with children have a significantly higher demand for environmentally friendly materials." These statistical patterns are extracted from historical design data through grey relational analysis and cross-analysis, and then transformed into parameters of probability distributions, forming calculable prior knowledge.

[0024] Specifically, determining the prior probability distribution in step S1 is essentially a parameter estimation problem. Different distribution forms are used for different types of design elements: for count-type design elements, such as the number of drawers and the number of short clothing areas, a Poisson distribution is used, and its mean λ is determined by user characteristics through a mapping function; for continuous design elements, such as hanging rod height and cabinet depth, a normal distribution is used, and its mean and standard deviation are determined by user characteristics; for categorical design elements, such as material selection and style preference, a multinomial distribution is used, and the probability of each category is determined by user characteristics.

[0025] Optionally, the feature-demand mapping library is constructed based on correlation analysis of historical data on customized furniture design. The correlation analysis includes grey relational analysis and cross-analysis; based on the results of the correlation analysis, the statistical correlation between user characteristics and design elements is fitted to the parameters of a prior probability distribution.

[0026] Specifically, the feature-requirement mapping library is pre-built by offline analysis of historical design data, and its construction process includes: (1) Data preparation: Export the completed design order data from the enterprise database. Each record contains user characteristics and the final confirmed design element values.

[0027] (2) Grey relational analysis: Calculate the correlation degree between each user feature and each design element, and filter out highly correlated features with a correlation degree greater than the correlation threshold. Optionally, the correlation threshold is set to 0.8.

[0028] (3) Cross-tabulation analysis: Perform cross-tabulation analysis on highly correlated feature-feature pairs to obtain the distribution of feature values ​​under different feature values.

[0029] (4) Parameter fitting: Fit the distribution obtained from cross-analysis to probability distribution parameters and store them as a mapping table in JSON format.

[0030] Optionally, after constructing the feature-requirement mapping library, the system dynamically generates the prior distribution. After the user feature data is input, the following operations are performed: traverse all design elements; for each design element, retrieve the prior parameters that match the current user features from the feature-requirement mapping library; if multiple features correspond to the same design element, use a weighted fusion strategy to determine the final parameters; and generate a specific probability distribution object according to the distribution type.

[0031] For example, in some specific embodiments, the specific implementation of step S1 includes a prior distribution of the number of drawers based on age and education level. A user's characteristics are: age 35, education level: master's degree, age categorized as 30-39 years old. The system retrieves the following from the mapping database: 30-39 year old group → Poisson distribution of drawer number λ=3.0 Master's Degree Group → Number of Drawers: Poisson Distribution λ = 3.5 A weighted average fusion method is used, with weights determined based on the grey relational degree. For example, the correlation degree between age and education level is 0.75, and the correlation degree between education level and education level is 0.85. = (0.75×3.0 + 0.85×3.5) / (0.75+0.85) = 3.28; The final generated prior distribution: number of drawers ~ Poisson (λ=3.28).

[0032] For example, in some specific embodiments, the specific implementation of step S1 includes, based on the prior distribution of material preferences for the residential structure, and given that the user characteristic is that there are children in the household, the system retrieves the following from the mapping database: Children's material preferences varied: solid wood 65%, particleboard 25%, other 10%; The final prior distribution is: P(material = log) = 0.65, P(material = solid wood particles) = 0.25, P(material = other) = 0.10.

[0033] In step S2, the system acquires the fuzzy requirement description input by the user and transforms it into an observation probability distribution of at least one design element. The core function of step S2 is to convert the fuzzy requirements directly expressed by the user into a computable observation probability distribution. For example, a user might say, "I want a convenient wardrobe," or upload a picture of a favorite piece of furniture. These expressions vary and are semantically ambiguous, making direct mathematical calculation impossible. Step S2 transforms natural human expressions into probabilistic machine language, providing observational data for subsequent Bayesian fusion.

[0034] In step S2, the system acquires the fuzzy requirement description input by the user and transforms it into an observation probability distribution of at least one design element. The observation probability distribution represents quantitative information about the user's wardrobe design preferences, expressed directly by the user through natural language or images, and is one of the important inputs for subsequent Bayesian fusion. Due to the diverse forms of user expression, the system designs differentiated processing paths, separately for natural language text input and image input, ultimately outputting a unified probability distribution.

[0035] In some specific embodiments, in step S2, the system processes the user input to determine the corresponding observation probability distribution, specifically including: (1) For natural language text input: The system uses a semantic parsing model to identify core requirement words and their modification intensity, and converts them into corresponding probability distribution parameters according to the preset semantic-distribution mapping rules. Optionally, the semantic-distribution mapping rules include a fuzzy quantifier mapping table and an emotion intensity mapping function: the fuzzy quantifier mapping table is used to map fuzzy quantifiers to the expected value offset of the probability distribution; the emotion intensity mapping function is used to map the intensity of tone to the discreteness parameter of the probability distribution.

[0036] (2) For image input: The system uses an image recognition model to extract style features, material features and layout features from the image, and matches the features with the design element library to generate the corresponding observation probability distribution.

[0037] In step S3, for each design element, the system fuses the prior probability distribution with the observed probability distribution to calculate the posterior probability distribution of the design element.

[0038] Optionally, step S3 specifically includes: the system selects the corresponding conjugate prior distribution family according to the data type of the design element; the system uses the properties of the conjugate prior to obtain the closed expression of the posterior probability distribution through analytical calculation; for cases where the conjugate prior cannot be used, the system uses a numerical sampling method to approximate the posterior probability distribution.

[0039] In step S4, if the prior probability distribution, posterior probability distribution, and observed probability distribution meet the interaction initiation conditions, the system initiates the interaction process to obtain user feedback. In step S4, the system determines whether the preset interaction initiation conditions are met. If they are met, the interaction process is initiated to obtain user feedback, clarifying uncertainties in the requirements or resolving conflicts between the prior and observed information. The interaction process includes conflict cause diagnosis, follow-up question template matching, follow-up question generation and output, and a user feedback receiving sub-step.

[0040] Optionally, the interaction initiation conditions mainly include two categories: first, uncertainty conditions based on posterior probability distribution assessment, reflecting a low level of confidence in the system's understanding of the current demand; and second, demand conflict conditions detected based on prior probability distribution and observed probability distribution, reflecting a significant contradiction between the user's expression and expectations based on their background characteristics. Both are triggered independently, and interaction can be initiated if either condition is met.

[0041] Optionally, the process of determining uncertainty conditions specifically includes: calculating an uncertainty index based on the dispersion of the posterior probability distribution; if the uncertainty index exceeds a first threshold, the uncertainty condition is determined to be met. The dispersion of the posterior distribution reflects the system's confidence in the demand estimate. If the posterior distribution is too dispersed, it indicates that the prior and observations have failed to form a consistent inference, the credibility of the current understanding of the demand is low, and further confirmation from the user is required.

[0042] In some specific embodiments, the basic implementation of the uncertainty condition includes: calculating the dispersion index of the posterior distribution of each design element. Different indices are used for different distribution types: for Poisson or gamma distributions corresponding to count elements, variance or coefficient of variation, such as standard deviation / mean, can be used; for normal distributions corresponding to continuous elements, standard deviation can be used directly; for multinomial distributions corresponding to categorical elements, entropy or the reciprocal of the maximum probability can be used. In some specific embodiments, a first threshold is set. When the dispersion index is greater than the first threshold, the uncertainty condition is determined to be met.

[0043] Optionally, the process for determining conflict conditions specifically includes calculating a conflict index, which is determined based on the degree of difference between the expected value of the observed probability distribution and the expected value of the prior probability distribution, as well as the dispersion of the prior probability distribution. When the conflict index exceeds a second threshold, a conflict is determined to exist. A significant difference between the prior distribution and the observed distribution may indicate that the user has special preferences, or that their expression does not conform to conventional expectations, requiring confirmation through interaction.

[0044] In some specific embodiments, in the concrete implementation of the demand conflict condition, a conflict index is defined to measure the degree to which the observed mean deviates from the prior mean, and is normalized using the prior standard deviation: ,in The expected value of the observed distribution. Let the expected value of the prior distribution be . This represents the standard deviation of the prior distribution. When the conflict index exceeds the second threshold... At that time, it was determined that there was a conflict of demand.

[0045] Optionally, in step S4, initiating the interaction process to obtain user feedback includes executing the corresponding interaction processing flow according to the type of interaction initiation condition, specifically including: when the interaction initiation condition is a demand conflict condition, executing the interaction processing flow SA; when the interaction initiation condition is an uncertain condition, executing the interaction processing flow SB.

[0046] Optionally, the interactive processing flow SA specifically includes: diagnosing the cause of the conflict and matching the most likely conflict type from a pre-set conflict cause knowledge base; selecting the corresponding follow-up question template from a pre-set follow-up question template library based on the matched conflict type; generating follow-up questions based on the follow-up question template and outputting them to the user; and receiving user feedback on the follow-up questions. Optionally, the interactive processing flow SB specifically includes: identifying at least one design element that satisfies the uncertainty condition; generating clarification questions for the design element and outputting them to the user; and receiving user feedback on the clarification questions.

[0047] In step S5, the system converts user feedback into observation data and updates the posterior probability distribution of the design elements based on the observation data. After the interactive processing flow ends, the system obtains user feedback on follow-up or clarification questions. Step S5 converts this feedback information into computable observation data and updates the posterior probability distribution of the corresponding design elements based on the new observation data, making the understanding of requirements more accurate.

[0048] User feedback may take the form of multiple-choice options, numerical input, or natural language text. In step S5, the system transforms the feedback into an observation probability distribution similar to that in step S2, based on the feedback type. Then, using the same Bayesian fusion method as in step S3, the new observation is fused with the current posterior (as the new prior) to obtain an updated posterior distribution.

[0049] The transformation rules are consistent with those in step S2, which transforms fuzzy demand descriptions into observed distributions, but may be simpler because the feedback is usually more explicit. For example: Multiple choice options: directly mapped to the values ​​or categories of design elements, and can be assigned an equivalent sample size based on the degree of certainty of the options.

[0050] Numerical input: directly used as the observed value, and can be combined with the input method to estimate uncertainty (e.g., inputting precise numbers results in small variance, while inputting ranges results in large variance).

[0051] Natural language text: The semantic parsing model is invoked again to transform it into an observation distribution.

[0052] The update process is the same as step S3. Based on the data type, select conjugate prior or numerical sampling, take the current posterior as the prior, take the new observation as the observation, and calculate the new posterior.

[0053] In some specific embodiments, in step S5, the feedback conversion process includes: (1) Multiple choice feedback: Each option corresponds to a value or category of a design element. For example, the follow-up question on the number of drawers has the following options: A. Yes, I'm sure → The observation is still 1, high confidence level (n=10) B. I plan to use a storage box instead → The observation is still 1, but with the added information of "using a storage box", the confidence level is moderate (n=5). C. I need more drawers → Observations adjusted to 3, medium confidence (n=5) The system maps options to specific observations and assigns an equivalent sample size based on the degree of certainty of the option. The equivalent sample size can be determined by predefined rules or based on the distribution of the option in historical data.

[0054] (2) Numerical Input: The user inputs "I want the hanging rod height to be around 1500mm". The system parses the value 1500 and determines the standard deviation based on the input method. Example rule: If you enter an exact number such as "1500", the standard deviation will be a smaller value (e.g., 10 mm).

[0055] If the input range is "1500-1600", the mean value is 1550, and the standard deviation is calculated based on the length of the range (e.g., (1600-1500) / (2×1.96) ≈ 25.5mm).

[0056] If the standard deviation contains vague terms such as "around" or "approximately", it will be magnified proportionally (e.g., by 1.2 times).

[0057] (3) Natural Language Feedback: Call the same semantic parsing model as in step S2 to convert the text into an observation distribution. For example, if a user says, "Actually, I prefer the natural wood color, but my budget is limited," the system will parse it as a preference for natural wood material, but with additional cost constraints, and output a weighted observation distribution.

[0058] In some specific embodiments, in step S5, the implementation process of the posterior update, taking the number of drawers as an example, includes: Current posterior: Based on the gamma distribution Gamma obtained from the aforementioned fusion ( ) New observation: Poisson distribution from feedback, expected value Equivalent sample size

[0059] Updated formula: ,

[0060] The updated posterior Gamma is obtained. ).

[0061] In some specific embodiments, in step S5, the posterior update for normally distributed elements is: Current posterior: Normal( ) New observations were obtained based on feedback: Normal ( ) Updated formula: Posterior precision = 1 / +1 /

[0062] Posterior mean = ( ) / (1 / +1 / ) Posterior variance = 1 / posterior precision In some specific embodiments, in step S5, the posterior update is performed for multiple distribution elements: Current and posterior: Dirichlet distribution ) New observation: Multinomial distribution, virtual counting vector (Originated from feedback) renew:

[0063] The updated posterior distribution will be used for the iteration termination judgment in step S6.

[0064] In step S6, the system determines whether the iteration termination condition is met. If not, it returns to step S4 and re-evaluates the interaction initiation condition based on the updated posterior probability distribution. If the condition is met, it proceeds to the next step. After updating the posterior distribution, the system needs to assess whether the current understanding of the requirement is clear enough, or whether the upper limit of the number of interactions has been reached, to decide whether to continue iterating (i.e., return to step S4 to re-evaluate the interaction initiation condition) or terminate the interaction and enter the output stage. Step S6 is the core of iteration control, preventing infinite loops or premature termination.

[0065] Optionally, the iteration termination condition typically includes one or more of the following: The posterior uncertainty of all design elements is below the first threshold: that is, the dispersion index of each element is less than the uncertainty threshold set in step S4. This indicates that the system has sufficient confidence in its understanding of all elements.

[0066] The number of iterations reaches a preset limit: To prevent infinite loops, a maximum number of interaction rounds is set, and the loop will be forcibly terminated once this limit is reached.

[0067] User-initiated termination: During the interaction, the user may choose "no longer answer" or "generate a solution directly", and the system should be able to receive such instructions and terminate the interaction.

[0068] All conflicts have been resolved: For conflict-triggered interactions, it may be set that a conflict is considered resolved when the conflict index of all conflicting elements is below a second threshold.

[0069] After each posterior update, the system recalculates the uncertainty index and conflict index of each element and checks whether the termination condition is met. If any termination condition is met, the iteration stops and proceeds to step S7; otherwise, it returns to step S4 and re-evaluates the interaction initiation condition based on the updated posterior distribution.

[0070] In step S7, the system outputs a fused requirement vector containing the posterior probability distribution features of the design elements. After one or more rounds of interactive iteration, the system finally obtains the posterior probability distribution of each design element. Step S7 packages the features of these posterior distributions into a structured fused requirement vector for use by downstream design systems. Optionally, the posterior probability distribution includes the expected value, confidence interval, and probability quality function.

[0071] In some specific embodiments, in step S7, the demand vector is fused into a data structure containing key information about the posterior distribution of each design element. The output differs for different types of design elements. Counting elements: expected output value, variance or standard deviation, probability mass function.

[0072] Continuous elements: expected output value, standard deviation, confidence interval.

[0073] Categorical elements: Output the posterior probability, maximum probability category, and confidence level for each category.

[0074] In addition, the fusion demand vector may also include an overall confidence score. The output format should be machine-readable structured data (such as JSON or XML) for seamless integration with downstream systems.

[0075] According to such Figure 1 The implementation method shown in this invention maps user features to a prior probability distribution, transforms fuzzy demand descriptions into observed probability distributions, and uses Bayesian fusion to obtain a posterior probability distribution, thus achieving scientific quantification of the uncertainty in users' fuzzy demands. By setting dual interaction initiation conditions based on uncertainty and demand conflict, a targeted interaction process is initiated when necessary to obtain user feedback and iteratively update the posterior distribution. The final output fused demand vector fully represents the user's true demand in probabilistic form, providing a highly confident quantitative basis for downstream wardrobe design, significantly improving the accuracy of demand analysis and design efficiency, effectively reducing design rework and increasing user satisfaction.

[0076] Figure 2 for Figure 1 A flowchart illustrating step S2 of Example 1000. (See attached diagram.) Figure 2 As shown, step S2 includes steps S21 and S22.

[0077] In step S21, for natural language text input, the system uses a semantic parsing model to identify core requirement words and their modification strengths, and converts them into corresponding probability distribution parameters according to preset semantic-distribution mapping rules. In some specific embodiments, step S21 is based on a technical approach combining natural language processing and probabilistic modeling, and its core process is as follows: (1) Semantic parsing: A deep learning model is used to understand the user's text, identify the core requirement words mentioned in the text, and the modifier strength of each core word. At the same time, the model will also evaluate the tone strength of the whole sentence.

[0078] (2) Semantic-Distribution Mapping: This transforms the qualitative information obtained from semantic parsing into quantitative parameters. This transformation relies on two predefined mapping components: Fuzzy quantifier mapping table: An offset that maps fuzzy quantifiers to the expected value of a probability distribution. For example, "many drawers" would increase the expected value of the number of drawers by a specific number.

[0079] The emotion intensity mapping function is an adjustment coefficient that maps the intensity of tone to the dispersion of a probability distribution. The stronger the tone and the smaller the variance, the more certain the user is about the need.

[0080] (3) Generation of observation distribution: Based on the data type of the design elements, select an appropriate distribution form and generate a specific probability distribution using the parameters obtained from the above mapping. For example, for the number of countable drawers, a Poisson distribution is used, and its parameter λ is determined by the baseline value and the offset; for the preference of different types of materials, a multinomial distribution is used, and its probability vector is determined by the user's direct or indirect selection.

[0081] Optionally, the semantic-distribution mapping rule includes a fuzzy quantifier mapping table and an emotion intensity mapping function: the fuzzy quantifier mapping table is used to map fuzzy quantifiers to expected value offsets of probability distributions; the emotion intensity mapping function is used to map the intensity of tone to discreteness parameters of probability distributions.

[0082] The fuzzy quantifier mapping table is a predefined rule base that converts common fuzzy quantifiers into mathematical offsets. Optionally, the specific implementation of the fuzzy quantifier mapping table is shown in Table 1 below: Table 1. Partial examples of the fuzzy quantifier mapping table:

[0083] In some specific embodiments, sentiment intensity is used to adjust the variance of the probability distribution, reflecting the certainty of a user's need. Optionally, the sentiment intensity mapping function is defined as: variance adjustment factor = 1 - α * its intensity coefficient. Wherein, the sentiment intensity coefficient is calculated based on text features, and α is an adjustable parameter. The sentiment intensity coefficient can be accumulated according to the rules shown in Table 2 below: Table 2. Examples of rules for tone intensity coefficients in the emotion intensity mapping function:

[0084] For example, in some embodiments, if the user input includes an exclamation mark and "certain", then the intensity coefficient = 1 + 2 = 3, and the variance adjustment factor = 1 - 0.2 × 3 = 0.4, that is, the variance is reduced to 40% of the original.

[0085] In some specific embodiments, the observation distribution generator in step S21 generates a specific probability distribution based on the data type of the design elements and the parameters mentioned above: For count-type design elements: a Poisson distribution is used. First, determine the baseline expected value. Then, based on the modification strength of the core requirement words, the offset Δλ is obtained from the fuzzy quantifier mapping table to obtain... Optionally, the baseline expected value This can be set based on user profiles or system defaults, such as the expected default number of drawers. =2.5. Final observation distribution: number of drawers ~ Poisson( ).

[0086] For continuous design elements: a normal distribution is used. Baseline mean. and the benchmark standard deviation Provided by system presets or user profiles. The mean is adjusted based on the intensity of the modification. Where Δμ is obtained from the mapping table. The standard deviation is adjusted based on tone intensity: = Variance adjustment factor. Final observed distribution: Pole height ~ Normal( , ).

[0087] For categorized design elements: a multivariate distribution is used. If a user directly mentions a specific category, the probability of that category is set to a higher value, and the remaining probabilities are evenly distributed among other categories. If the user expresses themselves indirectly through vague descriptions, the probability distribution of abstract words onto specific categories is mapped according to a semantic association model. The intensity of the tone can be used to adjust the concentration of probabilities: the stronger the tone, the higher the probability of the target category, and the lower the probability of other categories.

[0088] In step S22, for image input, the system uses an image recognition model to extract style features, material features, and layout features from the image, and matches these features with a design element library to generate corresponding observation probability distributions. Step S22 transforms visual information into quantifiable observation probability distributions, thus complementing text input and more comprehensively capturing user needs. In some specific embodiments, step S22 is based on a technical approach combining computer vision and domain knowledge, and its core process is as follows: (1) Image recognition: Deep convolutional neural networks are used to analyze images and extract features in three dimensions: Style characteristics: Determine the overall style of the image, such as modern minimalist, Nordic style, industrial style, Chinese classical style, etc., and output the probability distribution of each style.

[0089] Material characteristics: Identify the main material of the wardrobe or furniture in the image, such as wood, white paint, glass, metal, etc., and output the probability distribution of the material category.

[0090] Layout features: Analyze the spatial layout information in the image, such as the proportion of open hanging areas, the location and number of drawers, the density of shelves, the height of hanging rods, etc., and output quantitative feature values.

[0091] (2) Feature matching: The features obtained from image recognition are matched with the design element library. The design element library is a pre-built knowledge base that stores typical design element parameters corresponding to each style, material, and layout.

[0092] (3) Generation of observation distribution: Based on the matching results and the recognition confidence, the corresponding probability distribution is generated. For classification features such as style and material, the probability distribution output by the recognition model is used directly; for layout features, a continuous probability distribution is generated based on the recognition value and its uncertainty.

[0093] Specifically, in step S22, the system deploys a pre-trained convolutional neural network and fine-tunes it on a large furniture image dataset to accurately identify visual features related to wardrobe design. The model is designed for multi-task output: The style classification head outputs a softmax probability vector, representing the probability that an image belongs to each style category. Style categories may include: modern minimalist, Nordic, industrial, new Chinese, light luxury, and American country.

[0094] The material classification header outputs a probability vector for each material category. Material categories include: raw wood, particleboard, white paint, glass, metal, and rattan.

[0095] The layout regression model outputs continuous values ​​for the number of drawers, hanging rod height, and percentage of open drawers. Additionally, the model can output the confidence level or uncertainty for each predicted value.

[0096] In some specific embodiments, the design element library is a structured knowledge base that stores typical relationships between style, material, and design elements, constructed based on historical design data statistics, expert experience, or industry standards. Optionally, specific embodiments of some contents of the design element library are shown in Table 3 below: Table 3. Specific implementation examples of some contents of the design element library:

[0097] In some specific embodiments, step S22, the feature matching process specifically includes: weighted fusion of typical parameters corresponding to each style based on the style probability vector obtained from image recognition. For example, if the probability of Nordic style is 0.85 and that of modern minimalist style is 0.10, then the expected value of the number of drawers = 0.85 × (typical mean of Nordic style) + 0.10 × (typical mean of modern minimalist style) + …, while also considering the variance generated by the distribution range.

[0098] In some specific embodiments, in step S22, the observation distribution generator generates a specific probability distribution based on the identification result and confidence level, specifically including: For classification elements such as style and material: the probability vector output by the recognition model is directly used as the parameter of the multinomial distribution. For example, style preference ~ Multinomial(p=[0.85, 0.10, 0.05]).

[0099] For count data in layout elements: if the identification model outputs a continuous value, it can be used as the expected value of a Poisson distribution, and the variance can be adjusted according to the confidence level. For example, the number of drawers ~ Poisson (λ=2.3). If the confidence level is low, the distribution of λ can be expanded, or the uncertainty of the model output can be used directly.

[0100] For continuous data in layout elements: if the recognition model outputs mean μ and standard deviation σ, then a normal distribution is directly adopted: pole height ~ Normal(μ,σ).

[0101] According to such Figure 2 The implementation method shown in this invention transforms ambiguous user requests expressed in different forms into a computable observation probability distribution through two paths: natural language processing and image processing. Whether the user says "I want many drawers" or uploads a reference image, the system ultimately outputs a structured probability representation, providing standardized observation data for subsequent Bayesian fusion. This multimodal processing capability enables the system to more comprehensively understand user intent and also enhances the flexibility of the user experience.

[0102] Figure 3 for Figure 1 A flowchart illustrating step S3 of Example 1000. (See attached diagram.) Figure 3 As shown, step S3 includes steps S31-S33.

[0103] In step S31, the system selects the corresponding conjugate prior distribution family based on the data type of the design element. In Bayesian statistics, if the prior distribution and the likelihood function belong to the same distribution family, then the posterior distribution also belongs to that family, and its parameters can be updated using simple analytical formulas. This property is called conjugateness, and using conjugate priors can greatly simplify the calculation process. The purpose of step S31 is to select a conjugate prior distribution family based on the data type of the design element, laying the foundation for subsequent analytical calculations.

[0104] The correspondence between data types and conjugate priors is based on probability and statistics theory. Common correspondences are as follows: For count data: the generation process of observed data is usually described using the Poisson distribution. The conjugate prior of the Poisson distribution is the gamma distribution. The gamma distribution has two parameters: a shape parameter α and a rate parameter β, and its probability density function is Gamma(α,β). When the observed data follows a Poisson distribution, the posterior distribution is still a gamma distribution, and the parameter update rule is simple.

[0105] Continuous data: Observed data are usually described using a normal distribution. If the variance of the observed data is known, the conjugate prior of the normal mean is a normal distribution; if both the mean and variance are unknown, the conjugate prior is a normal-inverse gamma distribution or a normal-inverse chi-square distribution. In practical applications, it is often assumed that the variance is known or simplified methods are used, employing normal-normal conjugate pairs.

[0106] Categorized data: The observed data follow a multinomial distribution. The conjugate prior of the multinomial distribution is the Dirichlet distribution. The Dirichlet distribution is a multidimensional generalization of the Beta distribution, with parameters denoted by a vector α = (α1, α2, …, α…). k For posterior updates, simply add the observation count to the corresponding component.

[0107] In a specific embodiment, for each design element, step S31 first reads its data type and then retrieves the corresponding conjugate prior family from a predefined mapping table. This step requires no complex calculations and only involves configuration querying and allocation.

[0108] In step S32, the system utilizes the conjugate prior property to analytically calculate the closed-form expression of the posterior probability distribution. After determining the family of conjugate prior distributions, step S32 uses the conjugate property of Bayes' theorem to directly calculate the parameters of the posterior distribution through analytical formulas. This step combines the prior distribution parameters with the observed data to output a complete parameterized representation of the posterior distribution, providing a foundation for subsequent uncertainty assessment and interactive decision-making.

[0109] Optionally, in step S32, the formula for calculating the posterior parameter for different conjugate pairs is as follows: (1) For count data, Poisson-gamma conjugate pairs are used: Prior: Gamma distribution (Gamma) ) Observations: Poisson distribution, observation value x. In this scheme, the observation distribution itself is also a probability distribution, therefore it is necessary to process the observation distribution rather than individual observation values. This involves transforming the observation probability distribution into equivalent observation data. In practice, the expected value of the observation distribution is used as the virtual observation count, or the observation distribution is regarded as a weighted sum of multiple virtual observations. For simplification, the system can extract the expected value of the observation distribution as a representative value and assign it a certain weight.

[0110] Posterior parameters: ,

[0111] Optionally, the observation distribution is a Poisson distribution, with parameters... The system can The expected value of an observation is considered as the expected value of a single observation, but to reflect the uncertainty of the observation, the observation distribution needs to be treated as multiple virtual samples. A practical approach is to use the expected value of the observation distribution as the observed value and set an equivalent sample size. Then the posterior parameter: , .

[0112] (2) For continuous data with known variance, normal-normal conjugate pairs are used: Prior: Normal distribution ) Observation: Normal distribution ),in Known.

[0113] Posterior mean:

[0114] Posterior variance:

[0115] That is, posterior precision is the sum of prior precision and observational precision, and posterior mean is the weighted average of the two precisions.

[0116] (3) Categorical data are classified using multinomial-Dirichlet conjugate pairs: Prior: Dirichlet distribution ),in Observations: Multinomial distribution, observation count vector is (That is, the number of times each category appears).

[0117] Posterior parameters:

[0118] In this scheme, the observation distribution itself is a multinomial distribution, and its probability vector is... This can be converted into a virtual count. For example, an equivalent sample size can be set. Then virtual counting . It can be determined based on the concentration of the observed distribution (e.g., maximum probability); the more concentrated the probability, the higher the probability. The larger.

[0119] In step S33, for cases where conjugate priors cannot be used, the system employs a numerical sampling method to approximate the posterior probability distribution. Although conjugate priors cover most common data types, suitable conjugate priors may not be found in some complex scenarios. For example, analytical methods are no longer applicable when the prior distribution is not in a standard form, the observation distribution is very complex, or there are correlations between design elements that require joint modeling.

[0120] Step S33, as a backup plan, employs a numerical sampling method to approximate the posterior distribution, ensuring the system can still perform fusion under complex conditions. Optionally, the numerical sampling method used includes Markov Chain Monte Carlo (MCMC) or importance sampling. In some specific embodiments, when the system detects that the combination of current design elements cannot use conjugate priors, it switches to numerical sampling mode. The specific process is as follows: (1) Define the target posterior distribution: According to Bayes' theorem, the posterior distribution ∝ the prior distribution × the likelihood function of the observed distribution. The system needs to be able to calculate the prior probability density and likelihood function value at any parameter point.

[0121] (2) Selection of sampling algorithm: For low-dimensional problems, the Metropolis-Hastings algorithm can be used. Algorithm steps: Initialize parameter values

[0122] For t=1 to T: Generate candidates from the proposal distribution.

[0123] Calculate the probability of acceptance Accept candidates with probability α, otherwise keep the current value. Collect converged samples as posterior samples (3) Posterior approximation: After collecting a sufficient number of samples (e.g., 10,000), calculate the sample mean and variance as an estimate of the posterior distribution. Histograms or kernel density estimates can also be plotted. These statistics will be passed to subsequent steps.

[0124] Because numerical sampling involves significant computational costs, systems typically treat it as offline preprocessing or a low-frequency task, or only trigger it when user interaction is infrequent. In online systems, approximate analytical forms of the posterior can be pre-calculated for complex cases, or faster variational inference can be used.

[0125] According to such Figure 3 The embodiment shown in this invention achieves a scientific integration of user needs, specifically, it achieves the following beneficial effects: Through conjugate prior analytical computation, millisecond-level posterior updates are achieved in most scenarios, meeting the needs of real-time interaction; numerical sampling methods are used as alternatives to ensure that the system can handle probabilistic models of arbitrary complexity; all probabilistic features of the posterior distribution are output, providing rich information for subsequent uncertainty assessment and interactive decision-making; the data type-driven design makes it easy to add new design elements or adjust the modeling method to adapt to business development.

[0126] Figure 4 for Figure 1 A flowchart illustrating the interactive processing flow SA and the interactive interaction flow SB in step S4 of Example 1000. Figure 4 As shown, the interactive processing flow SA includes steps SA1-SA4, and the interactive processing flow SB includes steps SB1-SB3.

[0127] In some specific embodiments, when the interaction is triggered by a conflict of needs, the system executes process SA, which clarifies the real needs behind the conflict by diagnosing the cause of the conflict, selecting a follow-up question template, generating follow-up questions, and receiving feedback.

[0128] In some specific embodiments, in step SA1, the system diagnoses the cause of the conflict and matches the most likely conflict type from a pre-built conflict cause knowledge base. The possible causes of the conflict are analyzed to allow for targeted follow-up questions. Different causes require different question designs. In some specific embodiments, the system has a built-in conflict cause knowledge base that stores common conflict types and their diagnostic rules. Optionally, each conflict type corresponds to a set of conditions, for example: User cognitive bias: Users may underestimate or overestimate their actual needs. For example, when customizing a wardrobe for the first time.

[0129] Special usage habits: Users have unique storage methods. For example, they prefer storage boxes to drawers.

[0130] Future changes in demand: Users plan for changes in their future family structure, such as having children.

[0131] Feature presets are not applicable: The user's characteristics do not conform to general patterns and belong to niche preferences.

[0132] Specifically, in step SA1, during diagnosis, the system outputs the most likely conflict type based on information such as the design elements that triggered the conflict, the size of the conflict index, user characteristics, and historical interaction data, through rule matching or a simple decision tree.

[0133] For example, in some specific embodiments, assume a drawer number conflict triggers, with a conflict index of 0.58. The system checks user characteristics: age 35, master's degree, and living alone. Knowledge base rules: If the conflict index < 1.0 and the user lives alone → possible reason: people living alone usually have lower needs, but higher education may raise expectations, falling under "feature presets and individual differences". If the conflict index > 1.5 and there are no special records → possible reason: special usage habits. This example matches "feature presets and individual differences".

[0134] In step SA2, the system selects a corresponding follow-up question template from a pre-set follow-up question template library based on the matched conflict type. Optionally, in step SA2, the system selects a corresponding template from the pre-set follow-up question template library as a framework for generating questions based on the diagnosed conflict type.

[0135] Optionally, the question template library is categorized by conflict type. Each template includes: template ID, applicable conflict type, question text framework (including placeholders such as [feature name], [recommended value], [observation value]), option list (if it is a multiple-choice question), and priority (used for multi-dimensional conflict sorting). The system retrieves templates from the library based on the conflict type output in step SA1. If multiple templates are available, the default can be selected or further filtered based on user characteristics.

[0136] In step SA3, the system generates follow-up questions based on the follow-up question template and outputs them to the user. In some specific embodiments, in step SA3: the system dynamically populates the follow-up question template with the current user data and design element data to generate personalized natural language follow-up questions; in cases where multiple conflicting dimensions exist simultaneously, the system determines the follow-up question order according to a preset priority rule. Optionally, the follow-up questions are in the form of multiple-choice questions, with options preset according to the conflict type.

[0137] In step SA4, the system receives user feedback on follow-up questions. Optionally, in step SA4, the system receives user feedback on follow-up questions and records the feedback data for subsequent steps to transform into new observation data, used to update the posterior probability distribution. Specifically, depending on the question format, the feedback can be multiple-choice options, numerical input, or natural language text. The system needs to parse the feedback content, extract valid information, and transform it into structured data.

[0138] In some specific embodiments, when the interaction is triggered by uncertain conditions, the system executes process SB, which directly generates clarification questions for design elements with high uncertainty, without the need for complex diagnosis.

[0139] In step SB1, the system identifies at least one design element that satisfies the uncertainty condition. From all design elements, a list of elements whose posterior distribution dispersion exceeds a threshold and requires clarification is selected. The system iterates through all design elements and, based on the uncertainty index calculated in step S4, records elements whose index is greater than the first threshold. Multiple elements may satisfy the condition simultaneously.

[0140] In step SB2, the system generates clarification questions for the design elements and outputs them to the user. In some specific embodiments, in step SB2, the system generates simple and direct clarification questions for each uncertain design element, inquiring about the user's specific preferences. Optionally, the system uses a preset general clarification template, such as: "What are your specific requirements for [element name]?" or "What would you like for [element name]?" The template may vary slightly for different types of elements.

[0141] In step SB3, the system receives feedback from the user regarding the clarification issue. Optionally, the processing method of step SB3 is similar to that of step SA4, and will not be described again here.

[0142] According to such Figure 4 The implementation method shown in this invention organically combines Bayesian fusion with active interaction, upgrading the system from passive reception to active understanding, significantly improving the accuracy and efficiency of customized wardrobe demand analysis. Specifically, it achieves the following: Intelligent decision-making: Based on quantitative indicators, accurately determine when and why interaction is needed.

[0143] Differentiated processing: Design separate interaction processes for conflicts and uncertainties, with in-depth diagnosis for the former and direct clarification for the latter, balancing efficiency and relevance.

[0144] Closed-loop optimization: User feedback becomes a new observation in the iteration, enabling the understanding of requirements to continuously evolve.

[0145] User experience: The timing of interactions is reasonable, the questions are personalized, and the format is user-friendly, which enhances user engagement and satisfaction.

[0146] Figure 5 for Figure 4 A flowchart illustrating sub-step SA3 of the interactive processing flow SA. (See diagram below.) Figure 5 As shown, step SA3 further includes steps SA31-SA32.

[0147] In step SA31, the system dynamically populates the follow-up question template with the current user data and design element data to generate personalized natural language follow-up questions. Step SA31 also involves replacing placeholders in the selected template with actual data to generate natural language questions tailored to the current user and design elements.

[0148] Optionally, in step SA31, the system extracts typical recommended values ​​from the prior distribution, extracts the user's current expression value from the observed distribution, and extracts feature information from the user data, filling these values ​​into the template. If it is a multiple-choice question, the options are also preset and filled according to the conflict type. Optionally, the typical recommended value is the mean or a commonly used range. Optionally, follow-up questions are in the form of multiple-choice questions, with options preset according to the conflict type.

[0149] In some specific embodiments, step SA31 is implemented in the following ways: Template: "Based on your background, most users would choose [recommended value] [feature name], are you sure you only need [observations]?" Recommended value: Calculated from the prior distribution. For example, if the prior mean of the number of drawers is 3.43, it can be rounded up to "3-4".

[0150] Element name: "Drawer".

[0151] Observed values: Observational distribution expectation 1.

[0152] After filling: "Based on your background, most users would choose 3-4 drawers. Are you sure you only need 1?" If it is a multiple-choice question, options such as "A. Yes, I'm sure; B. I plan to use storage boxes instead; C. I need more drawers" will be generated.

[0153] In step SA32, for situations where multiple conflicting dimensions exist simultaneously, the system determines the order of follow-up questions according to a preset priority rule. When multiple design elements trigger conflict conditions simultaneously, the system needs to determine the order of follow-up questions to avoid confusing the user by outputting too many questions at once. The system sorts the elements to be followed up by priority and processes them sequentially, completing one before moving on to the next. Optionally, the preset priority rule can be based on: Conflict index: The more severe the conflict, the higher the priority.

[0154] Design element importance weights: Predefine the weights of each element. Optionally, the predefined weights follow the order of Material > Size > Layout.

[0155] Dependencies: Clarifying certain elements may affect other elements; for example, determining the material may influence subsequent color selection.

[0156] Figure 6 A schematic diagram of an embodiment 2000 of the customized furniture demand analysis system based on uncertainty modeling of the present invention. (See diagram below.) Figure 6As shown, embodiment 2000 includes a priori loading module 201, an input parsing module 202, a fusion analysis module 203, a user interaction module 204, a feedback update module 205, an iterative judgment module 206, and an output module 207.

[0157] The prior loading module 201 is used to acquire user feature data and, based on a pre-built feature-requirement mapping library, determine the prior probability distribution of at least one design element corresponding to the user feature data.

[0158] In some specific embodiments, the prior loading module 201 transforms the user's objective characteristics into computable prior knowledge. Internally, the prior loading module 201 maintains a feature-demand mapping library pre-built through grey relational analysis and cross-analysis, storing the correspondence between different user feature values ​​and the probability distribution parameters of each design element. Upon receiving user feature data, the prior loading module retrieves matching prior parameters from the mapping library and generates corresponding prior probability distributions based on the data type of the design elements, providing initial judgments at the group statistical level for subsequent Bayesian fusion.

[0159] The input parsing module 202 is used to acquire the fuzzy requirement description input by the user and transform the fuzzy requirement description into an observation probability distribution of at least one design element. The input parsing module 202 transforms the fuzzy requirement description input by the user in natural language or image form into a standardized observation probability distribution.

[0160] For natural language input, a semantic parsing model finely tuned for the furniture customization field is employed to identify core requirement words and their modification strength in the text. Based on a pre-defined fuzzy quantifier mapping table and sentiment intensity mapping function, qualitative descriptions are transformed into expected offsets and dispersion parameters of probability distributions. For image input, a multi-task image recognition model is used to simultaneously extract style features, material features, and layout features from the image. The recognition results are then matched with a design element database to generate corresponding observation probability distributions. The output of the input parsing module 202 has the same probability distribution format as the output of the prior loading module 201, laying the foundation for subsequent fusion.

[0161] The fusion analysis module 203 is used to fuse the prior probability distribution and the observed probability distribution for each design element to calculate the posterior probability distribution of the design element. In some specific embodiments, the fusion analysis module 203 scientifically fuses prior knowledge with individual expressions.

[0162] For each design element, the fusion analysis module 203 selects the corresponding conjugate prior distribution family based on its data type and performs analytical calculations using Bayes' theorem to obtain a closed-form expression for the posterior probability distribution. For special cases where conjugate priors cannot be used, the fusion analysis module 203 switches to numerical sampling mode to approximate the posterior distribution. The posterior distribution output by the fusion analysis module 203 simultaneously incorporates information from both group statistical regularities and individual user expressions, and represents the confidence level of the demand estimate in a complete probabilistic form.

[0163] User interaction module 204 is used to initiate an interaction process and obtain user feedback for customized furniture wardrobe design elements that meet the interaction initiation conditions. In some specific embodiments, user interaction module 204 initiates a dialogue with the user when the understanding of the requirements is uncertain or contradictory. Its specific functions include: First, the uncertainty index is calculated based on the posterior distribution output by the fusion analysis module. Simultaneously, a conflict index is calculated based on the prior and observed distributions, and compared with a preset threshold to determine if the interaction initiation conditions are met. When the interaction is triggered by uncertainty, a clarification process is executed, directly generating general questions for design elements with high uncertainty. When the interaction is triggered by requirement conflict, a deep diagnostic process is executed, calling the built-in conflict cause knowledge base to analyze possible causes of the conflict, selecting a matching template from the follow-up question template library, and dynamically generating targeted multiple-choice or open-ended questions. For situations where multiple elements trigger the interaction simultaneously, the follow-up question order is determined according to preset priority rules to avoid information overload.

[0164] The feedback update module 205 is used to convert user feedback into observation data and update the posterior probability distribution of design elements based on the observation data. Optionally, the feedback update module 205 converts the user's answers to follow-up questions into new, computable observation data and updates the posterior probability distribution based on this.

[0165] In some specific embodiments, the feedback update module 205 uses the same method as the input parsing module 202 to convert the feedback into an observation probability distribution according to the feedback form. Then, it takes the current posterior distribution as the new prior and the observation obtained from the feedback conversion as new evidence. It then calls the update mechanism of the fusion analysis module 203 again to obtain the updated posterior distribution, so that each interaction can substantially improve the system's understanding accuracy of user needs.

[0166] The iteration judgment module 206 is used to determine whether the iteration termination condition is met. If not, it returns to the fusion analysis module; if the condition is met, it enters the output module. In some specific embodiments, the iteration judgment module 206 decides whether to continue iteratively optimizing the requirement understanding.

[0167] Optionally, the iteration judgment module 206 re-evaluates, after each feedback update, whether the posterior uncertainty of all design elements has fallen below the termination threshold, whether the number of iterations has reached a preset upper limit, and whether the user has actively issued a termination command. If any termination condition is met, the control flow enters the output module 207; otherwise, the flow returns to the user interaction module 204, re-evaluating the interaction initiation conditions based on the updated posterior distribution, and starting a new round of interaction loop. This judgment mechanism ensures that the system outputs information promptly when the requirements are sufficiently clear, and avoids infinite loops when the requirements remain ambiguous.

[0168] Output module 207 is used to output a fused requirement vector containing the posterior probability distribution features of design elements. The fused requirement vector is used to generate a customized furniture wardrobe design scheme. As the final delivery component of the system, output module 207 is responsible for encapsulating the posterior probability distribution after one or more rounds of interactive optimization into a structured fused requirement vector.

[0169] Optionally, the fused requirement vector outputs the probabilistic characteristics of each design element: for count elements, it outputs the expected value and standard deviation; for continuous elements, it outputs the expected value, standard deviation, and confidence interval; and for categorical elements, it outputs the posterior probability of each category and the category with the highest probability. Simultaneously, the fused requirement vector can also include user preference notes and interaction logs recorded during the interaction process, which can be directly used by downstream design systems. The output module 207 ensures a closed-loop process from fuzzy user input to precise quantification of requirements.

[0170] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0171] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for analyzing customized furniture demand based on uncertainty modeling, characterized in that, Includes the following steps: S1: Obtain user feature data and, based on a pre-built feature-demand mapping library, determine the prior probability distribution of at least one design element corresponding to the user feature data; the feature-demand mapping library is constructed based on correlation analysis of historical data of customized furniture design. S2: Obtain the fuzzy requirement description input by the user, and transform the fuzzy requirement description into the observation probability distribution of at least one of the design elements; S3: For each design element, the prior probability distribution and the observed probability distribution are fused to calculate the posterior probability distribution of the design element; S4: If the prior probability distribution, posterior probability distribution and observed probability distribution meet the interaction initiation conditions, initiate the interaction process to obtain user feedback. S5: Convert the user feedback into observation data, and update the posterior probability distribution of the design element based on the observation data; S6: Determine whether the iteration termination condition is met. If not, return to step S4 and re-evaluate the interaction initiation condition based on the updated posterior probability distribution. If the condition is met, proceed to the next step. S7: Output a fused demand vector containing the posterior probability distribution features of the design elements, which is used to generate a custom furniture wardrobe design scheme.

2. The method according to claim 1, characterized in that, The interactive initiation conditions include uncertainty conditions based on posterior probability distribution evaluation and demand conflict conditions based on prior probability distribution and observed probability distribution detection.

3. The method according to claim 2, characterized in that, The determination process of the uncertainty condition specifically includes: calculating an uncertainty index based on the degree of dispersion of the posterior probability distribution; when the uncertainty index is greater than a first threshold, it is determined that the uncertainty condition is met. The process for determining the demand conflict condition specifically includes calculating a conflict index, which is determined based on the degree of difference between the expected value of the observed probability distribution and the expected value of the prior probability distribution, as well as the degree of dispersion of the prior probability distribution; when the conflict index exceeds a second threshold, a demand conflict is determined to exist.

4. The method according to claim 2, characterized in that, In step S4, obtaining user feedback by initiating the interaction process includes executing a corresponding interaction processing flow based on the type of interaction initiation condition, specifically including: When the interaction initiation condition is the requirement conflict condition, the interaction processing flow SA is executed: SA1: Diagnose the cause of conflict by matching the most likely conflict type from a pre-built conflict cause knowledge base; SA2: Select the corresponding follow-up question template from the preset follow-up question template library based on the matched conflict type; SA3: Generate follow-up questions based on the aforementioned follow-up question template and output them to the user; SA4: Receive user feedback on the follow-up questions; When the interaction initiation condition is the uncertain condition, the interaction processing flow SB is executed: SB1: Determine at least one design element that satisfies the aforementioned uncertainty condition; SB2: Generate clarification questions for the aforementioned design elements and output them to the user; SB3: Receive user feedback on the clarification issue.

5. The method according to claim 4, characterized in that, Step SA3 further includes: The aforementioned follow-up question template is dynamically populated with current user data and design element data to generate personalized natural language follow-up questions; The follow-up questions are in the form of multiple-choice questions, with options pre-set according to the type of conflict. In cases where multiple conflicting dimensions exist simultaneously, the order of follow-up questions is determined according to a preset priority rule.

6. The method according to claim 1, characterized in that, In step S2, transforming the fuzzy demand description into an observation probability distribution of at least one design element further includes: For natural language text input, a semantic parsing model is used to identify core requirement words and their modification strength, and then convert them into corresponding probability distribution parameters according to the preset semantic-distribution mapping rules; For image input, an image recognition model is used to extract style features, material features, and layout features from the image, and these features are matched with a design element library to generate a corresponding observation probability distribution.

7. The method according to claim 6, characterized in that, The semantic-distribution mapping rule includes a fuzzy quantifier mapping table and a sentiment intensity mapping function; The fuzzy quantifier mapping table is used to map fuzzy quantifiers to the expected value offset of a probability distribution; The emotional intensity mapping function is used to map the intensity of tone to a discrete parameter of a probability distribution.

8. The method according to claim 1, characterized in that, In step S3, the prior probability distribution and the observed probability distribution are fused to calculate the posterior probability distribution, which further includes: Select the corresponding conjugate prior distribution family based on the data type of the design element; By utilizing the property of conjugate priors, a closed-form expression for the posterior probability distribution is obtained through analytical calculation; For cases where conjugate priors cannot be used, numerical sampling methods are employed to approximate the posterior probability distribution.

9. The method according to claim 1, characterized in that, The feature-demand mapping library is constructed by performing correlation analysis on historical design data in advance. The correlation analysis includes grey correlation analysis and cross analysis. Based on the results of the correlation analysis, the statistical correlation between user features and design elements is fitted to the parameters of the prior probability distribution.

10. A customized furniture demand analysis system based on uncertainty modeling, characterized in that, include: The prior loading module is used to acquire user feature data and, based on a pre-built feature-requirement mapping library, determine the prior probability distribution of at least one design element corresponding to the user feature data. The input parsing module is used to obtain the fuzzy requirement description input by the user and convert the fuzzy requirement description into the observation probability distribution of at least one of the design elements; The fusion analysis module is used to fuse the prior probability distribution and the observed probability distribution for each design element to calculate the posterior probability distribution of the design element. The user interaction module is used to initiate the interaction process and obtain user feedback for custom furniture wardrobe design elements that meet the interaction initiation conditions. The feedback update module converts the user feedback into observation data and updates the posterior probability distribution of the design elements based on the observation data. The iteration judgment module determines whether the iteration termination condition is met. If not, it returns to the user interaction module and re-evaluates the interaction initiation condition based on the updated posterior probability distribution. If the condition is met, it enters the output module. The output module is used to output a fused demand vector containing the posterior probability distribution features of the design elements, and the fused demand vector is used to generate a customized furniture wardrobe design scheme.

Images