Product key production parameter mining method and system based on user demand classification

Abstract

The invention discloses a product key production parameter mining method and system based on user demand classification, and relates to the technical field of artificial intelligence. The method comprises the following steps: obtaining user preference data containing paired preference samples and a candidate parameter space formed by mass-producible parameter combinations; based on the data and the space, an improved self-adaptive direct preference optimization algorithm is adopted to train a strategy model, a dynamically calculated self-adaptive reward margin is introduced in the training process so as to adjust a model updating gradient according to a sample differentiation degree, and a Coubeck-Leibler divergence constraint is applied to limit parameter search in a candidate parameter space; and after the strategy model is converged, determining a group of product parameter combinations with the maximum posterior probability from the candidate parameter space according to the optimized model, and outputting the product parameter combinations as key production parameters.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a product key production parameter mining method and system based on user demand classification. BACKGROUND

[0002] In modern manufacturing, especially in the fields closely related to user experience such as consumer electronics, home appliances and automobiles, the market competitiveness of a product largely depends on whether its design can accurately match the individual needs of users. Therefore, how to scientifically mine the key production parameter combination that can maximize user satisfaction from the complex user feedback has become a core issue in product research and development and intelligent manufacturing fields.

[0003] At present, enterprises often rely on traditional user experience evaluation methods when determining product parameters. For example, user preference data is collected through market research, questionnaires (such as Likert scale), etc. However, the data obtained by such methods often has strong subjectivity, and the feedback of different users in different situations has large dispersion and poor consistency, making it difficult to directly and accurately map these qualitative and fuzzy preferences into executable and quantitative engineering parameters on the production line. In order to solve the above problems, the industry has also tried to use some traditional optimization algorithms such as genetic algorithm (GA) or particle swarm optimization (PSO) to search for the optimal solution in the preset parameter space. However, modern products have multiple parameters and complex coupling relationships, forming a huge, non-convex search space. When dealing with such problems, these traditional optimization algorithms are prone to premature convergence and fall into local optimal solutions, making it difficult to guarantee finding a global optimal parameter combination that can truly maximize the preferences of all target users.

[0004] In summary, the existing technology has obvious shortcomings in connecting subjective user preferences with objective production constraints. On the one hand, it is difficult to effectively process and utilize relative and pairwise preference information among users; on the other hand, the efficiency and effectiveness are both lacking when performing global optimization in a limited, discrete, and high-dimensional production feasible parameter space. Therefore, there is an urgent need in the field for a new technical solution to solve the technical problem of accurately and efficiently mining key production parameters that meet both user real preferences and industrial production conditions. SUMMARY

[0005] The purpose of the present application is to solve the problems existing in the prior art, and a product key production parameter mining method and system based on user demand classification are provided. In order to achieve the above purpose, the technical solutions adopted by the embodiments of the present application are as follows: In the first aspect, the embodiment of the present application proposes a product key production parameter mining method based on user demand classification, which comprises the following steps: obtaining user preference data containing paired preference samples, and a candidate parameter space composed of multiple groups of product parameter combinations that can be mass-produced under existing production process constraints; based on the user preference data and the candidate parameter space, training a strategy model using an improved adaptive direct preference optimization algorithm, the steps of training the strategy model including: for each group of paired preference samples, dynamically calculating an adaptive reward margin to adjust the update gradient of the model parameters of the strategy model according to the differentiation degree of the samples; and imposing a Kullback-Leibler divergence constraint on the optimization of the strategy model to limit the parameter search space of the strategy model within the candidate parameter space; After the strategy model converges, a group of product parameter combinations with the maximum posterior probability is determined from the candidate parameter space according to the optimized strategy model, and is output as the key production parameter.

[0006] Preferably, the step of obtaining user preference data containing paired preference samples comprises the following steps: showing paired product parameter combinations to the target user group for the users to make a forced choice between two; recording the selection results of each user in each pair of product parameter combinations, and defining the parameter combination selected by the user as a preferred parameter combination vector , defining the parameter combination abandoned by the user as an abandoned parameter combination vector ; and associating the preferred parameter combination vector and the abandoned parameter combination vector with the corresponding user context information to form multiple groups of triplets in the format of , which constitute a user preference data set .

[0007] Preferably, before the step of training the strategy model using the improved adaptive direct preference optimization algorithm, a pre-processing step is further included for the user preference data, which comprises the following steps: performing data cleaning on the multi-dimensional feature values in the preferred parameter combination vector and the abandoned parameter combination vector to remove invalid or abnormal data; and performing normalization processing on the cleaned multi-dimensional feature values to scale the parameter features of different dimensions and value ranges to a unified numerical interval, so as to eliminate the influence of feature scale difference on model training, and to construct the normalized data into paired tensors for input of the strategy model.

[0008] Preferably, the training process adopts a reference model , the reference model is an initial policy model copy whose parameters are frozen before the training starts, the policy distribution of the reference model represents the prior engineering knowledge or feasible distribution of initial production parameters before the user preference alignment is performed; The Kullback-Leibler divergence constraint penalizes the policy model for deviating too much from the reference model , the output probability mass of the policy model is concentrated to discrete combinations within the candidate parameter space, thus ensuring that the output key production parameter combination is one of the predefined mass-producible solutions.

[0009] Preferably, the calculation of the adaptive reward margin is realized by constructing an implicit, dynamically adjusted reference model , the implicit reference model fuses the policy distribution of the current policy model and the policy distribution of the reference model by interpolation, and its mathematical expression is associated with the probability ratio of the policy model and the reference model on a specific sample , so that the reward margin can be adaptively adjusted according to the confidence difference of the model on different samples.

[0010] Preferably, the step of training the policy model is performed by minimizing the loss function defined as follows: wherein represents the expectation on the user preference dataset , the user preference dataset is composed of multiple groups of triplets, each group of triplets includes user context information , a preferred parameter combination vector selected by the user , and an abandoned parameter combination vector abandoned by the user ; is a Sigmoid activation function; is a temperature hyperparameter that controls the scale of the reward function; is the probability of the parameter combination generated by the policy model to be optimized when given the user context ; is the probability of the parameter combination generated by the implicit reference model; the loss function drives the model optimization by maximizing the difference in the logarithmic probability of the preferred response and the abandoned response.

[0011] Preferably, the method further comprises: outputting the complete policy distribution vector of the converged policy model together with the output key production parameters; the policy distribution vector is a list containing all product parameter combinations in the candidate parameter space and their corresponding preference probabilities. and storing the strategy distribution vector in a database for supporting downstream personalized product recommendation applications.

[0012] Preferably, the step of supporting downstream personalized product recommendation applications comprises: When a user visits a product customization or recommendation page, the system retrieves the corresponding strategy distribution vector from the database according to the user's context information and ranks the top several groups of product parameter combinations in descending order of preference probability in the strategy distribution vector, and displays them to the user as personalized recommendation options.

[0013] Preferably, the output key production parameters are formatted as data conforming to the interface specifications of the production line programmable logic controller (PLC) or enterprise resource planning (ERP) system, and are directly written into the corresponding device parameter area or generated production order.

[0014] In a second aspect, the embodiments of the present application propose a product key production parameter mining system based on user demand classification, which comprises: a data acquisition module for acquiring user preference data containing pairwise preference samples and a candidate parameter space composed of multiple groups of product parameter combinations that can be mass-produced under existing production process constraints; a parameter mining module for training a strategy model based on user preference data and candidate parameter space using an improved adaptive direct preference optimization algorithm, wherein the parameter mining module is configured to: during the training process, dynamically calculating an adaptive reward margin for each group of pairwise preference samples to adjust the update gradient of the model parameters of the strategy model according to the differentiation degree of the samples; and during the training process, imposing a Kullback-Leibler divergence constraint on the optimization of the strategy model to limit the parameter search space of the strategy model within the candidate parameter space; and a result generation module for determining a group of product parameter combinations with the maximum posterior probability from the candidate parameter space according to the optimized strategy model after the strategy model converges, and outputting them as key production parameters.

[0015] Advantages: ​The product key production parameter mining method based on user demand classification provided by the embodiment of the present application effectively overcomes the technical problems of strong subjectivity of user preference data, difficulty in quantification, and traditional optimization algorithm easily falling into local optimal solution by using an improved self-adaptive direct preference optimization algorithm and combining production constraints for model training. Compared with the prior art, the embodiment of the present application directly uses pair preference samples closer to the human decision-making mode for training, significantly improving the accuracy of capturing the real preferences of users. More importantly, the self-adaptive reward margin calculated dynamically is introduced in the training process, which can intelligently adjust the optimization gradient according to the differentiation degree of the preference samples, avoiding the pattern collapse problem in the model training process, so as to more stably and efficiently converge to the global optimal solution. In addition, by applying the Kullback-Leibler divergence constraint, it is ensured that the key production parameters finally output are always located in the preset producible candidate space, perfectly solving the alignment problem between user preferences and industrial production feasibility, and guaranteeing the direct industrial application value of the mining result. BRIEF DESCRIPTION OF DRAWINGS

[0016] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The accompanying drawings are included to provide a description of the preferred embodiments, and are not intended to limit the scope of the present application. Moreover, the same reference numerals are used throughout the various drawings to designate similar or equivalent parts. In the drawings: Figure 1 The flowchart of the product key production parameter mining method based on user demand classification provided by the embodiment of the present application is shown in Figure 1. Figure 2 The structure diagram of the product key production parameter mining system based on user demand classification provided by the embodiment of the present application is shown in Figure 2. Figure 3 The structure diagram of the computer device provided by the embodiment of the present application is shown in Figure 3. DETAILED DESCRIPTION

[0017] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be understood as limiting the present application. It should be noted that similar reference numerals and letters represent similar items in the following drawings, so that once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0018] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be described below in detail with reference to the drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0019] In the first aspect, see Figure 1 A product key production parameter mining method based on user demand classification provided by the embodiment of the present application can be executed on a computer device. The computer device can be a server, a personal computer (PC), or an embedded product key production parameter mining system based on user demand classification with sufficient computing capability. The device usually includes but is not limited to a processor, a memory, a communication interface and an input / output device. The memory stores computer program instructions, and the processor executes the instructions to realize the method described in the present application. In the embodiment, taking the research and development of a new intelligent hair dryer as an example, the steps of a product key production parameter mining method based on user demand classification are described in detail. The hair dryer aims to optimize its core performance parameters (such as wind speed, temperature, power, weight, noise, negative ion concentration, etc.) to the greatest extent to meet the individual preferences of different user groups. The method described in the embodiment of the present application includes the following steps: S101: Obtain user preference data containing paired preference samples, and a candidate parameter space composed of multiple groups of product parameter combinations that can be mass-produced under existing production process constraints.

[0020] In this step, the core task is to prepare two kinds of key input information for subsequent algorithm models; one is the data reflecting the real preferences of users, and the other is the constraint data indicating the boundaries of industrial production. The following sub-steps are included: S1011: Obtain user preference data.

[0021] In order to obtain high-quality user preference data that can be directly used for model training, the embodiment adopts a forced two-choice paired comparison method. Compared with the traditional Likert scale scoring method, this method can reveal the real trade-offs between different options of users, the signal is stronger, and the subjective bias is smaller. Preferably, this step specifically includes: S10111: Show the target user group paired product parameter combinations for the user to make a forced two-choice selection.

[0022] Specifically, in a constant temperature and humidity professional user experience laboratory, 500 target users with different hair types (e.g. long hair, short hair, curly hair, fine and soft hair), different ages and different gender backgrounds are recruited. Each user is presented with a series of (e.g. 20 sets) carefully designed pairs of product parameter combination comparison items. Each comparison item is displayed in the form of "A scheme" and "B scheme", clearly listing the respective core parameters. For example, one comparison item can be as follows: A scheme, wind speed 21 m / s, working noise 71 dB, whole machine weight 550 g. B scheme, wind speed 18 m / s, working noise 68 dB, whole machine weight 520 g. The user is required to choose his preferred one from the two explicit and perceptible schemes. In this way, the user is forced to make a trade-off between "faster dry hair speed" and "lower noise and lighter weight", thereby capturing his inner preference tendency.

[0023] S10112: Record the selection results of each user in each pair of product parameter combinations, and define the parameter combination selected by the user as the preferred parameter combination vector , and define the parameter combination abandoned by the user as the abandoned parameter combination vector .

[0024] The system automatically and accurately records each selection of each user. In the above example, if the user selects the A scheme, the parameter combination [21, 71, 550,...] of the A scheme is marked as the "preferred" or "winning" vector of the user, denoted as (winner); while the parameter combination [18, 68, 520,...] of the B scheme is marked as the "abandoned" or "losing" vector, denoted as (loser).

[0025] S10113: Associate the preferred parameter combination vector and the abandoned parameter combination vector with the corresponding user context information to form a set of triplets in the format of , which constitutes the user preference data set .

[0026] Before the user participates in the experiment, relevant context information of the user is collected, such as recording his hair type (long hair / short hair), hair volume (more / less), sensitivity to noise (high / medium / low), demand for styling (high / low), etc. through a questionnaire. These context information are bound with each selection behavior of the user. Finally, each raw data record forms a triplet The structure of a record is, for example, ({hair: long, noise sensitivity: high}, [21, 71,...], [18, 68,...]). All the choices made by all 500 users (500 x 20 = 10,000 records) are collected to form the final user preference dataset for model training .

[0027] S1012: Preprocess the user preference data.

[0028] Preferably, before formally training the model, the original user preference dataset needs to be preprocessed in order to improve the training efficiency and performance of the model. The preprocessing step includes: S10121: Data cleaning of the multi-dimensional feature values in the preference parameter combination vector and the discarded parameter combination vector .

[0029] Data cleaning aims to remove invalid or abnormal data. For example, check if there are parameter values outside the normal range (such as wind speed of 0 or negative, data missing, etc.) in the dataset due to user misoperation or system recording errors. These abnormal data will be corrected or removed to ensure the data quality of the input model.

[0030] S10122: Normalization of the cleaned multi-dimensional feature values.

[0031] Due to the huge differences in the dimensions and value ranges of product parameters (for example, wind speed unit is m / s, value range may be 15-25; while negative ion concentration unit is ions / cm³, value range may be 1.0 x 10 6 -2.5 x 10 6 ), such scale differences will seriously affect the stability and convergence speed of model training. Therefore, normalization processing is needed for all parameter features, for example, using the Min-Max Scaling method to linearly map all feature values to a unified interval. This can eliminate the negative impact of feature scale differences on model training.

[0032] S10123: Construct the normalized data into a paired tensor for strategy model input.

[0033] After cleaning and normalization, the entire dataset is converted into a numerical multi-dimensional array or tensor, usually with dimensions [sample number, 3, feature dimension]. In this example, the dimensions are [10,000, 3, N], where N is the parameter dimension of the hair dryer (such as 10 dimensions), and 3 corresponds to the context information , preference vector and abandoned vectors This structured tensor format is the standard input form for batch training of deep learning models.

[0034] S1013: Obtain candidate parameter space.

[0035] The candidate parameter space defines the "solution space" boundary of the technical solution of the present application, ensuring that all optimization results can be manufactured in the real world. This space is derived from the enterprise's internal production system.

[0036] Specifically, the candidate parameter space is a pre-defined set of all directly manufacturable product parameter combinations according to the enterprise's existing bill of materials, production process, supply chain capabilities, and cost control requirements. It is not a continuous, infinite space, but a discrete, finite list. Taking the hair dryer in this embodiment as an example, the research and development team and the production department jointly determine the following feasible parameter options: Air volume level: 15 m / s, 18 m / s, 21 m / s, 24 m / s (4 types); Rated power: 1100W, 1300W, 1600W (3 types); Diameter of air duct: 44mm, 46mm (2 types); Negative ion concentration: 1.0×106 ions / cm³, 2.5×106 ions / cm³ (2 types).

[0037] Different combinations of these parameters constitute all theoretically producible models. The total number of combinations is 4x3x2x2=48. This list of 48 specific parameter combinations constitutes the candidate parameter space of the present application This list is usually stored in the enterprise resource planning (ERP) system, and each row corresponds to a bill of materials (BOM) scheme that can be directly called by the production line.

[0038] S102: Based on user preference data and candidate parameter space, use an improved adaptive direct preference optimization algorithm to train the strategy model.

[0039] This step is the core of the present application, which aims to learn a strategy model that can accurately predict user preferences under the premise of meeting production constraints by using an advanced algorithm. The embodiment uses an improved adaptive direct preference optimization algorithm (Adaptive Direct Preference Optimization, ). Direct Preference Optimization (DPO) is an algorithm that can directly utilize preference data for optimization, which avoids the complex two-stage process of training a reward model first and then optimizing the policy in traditional reinforcement learning methods, and is more stable and efficient. The method used in the present application , which is based on the standard DPO, makes key improvements to solve the reward hijacking and mode collapse problems that may be encountered in practical applications. The step of training the policy model specifically includes: S1021: Adopting a reference model .

[0040] Preferably, before training begins, a reference model .

[0041] Specifically, the reference model is usually a copy of the initial policy model , and its model parameters are "frozen" at the beginning of training and remain unchanged throughout the training process. The policy distribution of the reference model represents the prior engineering knowledge or feasible distribution of initial production parameters before aligning with user preferences. For example, it can be a uniform distribution, i.e., it is believed that the probability of selecting 48 BOM schemes is equal without user preference data; or it can also be an initial distribution based on historical sales or expert experience. The role of the reference model is to provide a stable benchmark and anchor point for the optimization of the policy model.

[0042] S1022: Apply Kullback-Leibler divergence constraint to the optimization of the policy model.

[0043] In order to ensure that the output of the policy model always falls within the industrial feasible region defined by the candidate parameter space , the present application imposes a Kullback-Leibler (KL) divergence constraint on the optimization process of the model.

[0044] KL divergence is an index that measures the difference between two probability distributions. Here, the KL divergence between the distribution of the policy model and the distribution of the reference model cannot be too large, i.e. , where is a pre-set smaller threshold (e.g., 0.08).

[0045] Preferably, the Kullback-Leibler divergence constraint is implemented by penalizing the policy model for the reference model overfitting, the output probability mass of the policy model is concentratedly assigned to discrete combinations within the candidate parameter space. Since the reference model is itself defined on the 48 feasible BOM schemes, this hard constraint acts like a “gravitational field”, ensuring that no matter how the policy model is updated, its probability distribution is always “locked” on these 48 points, and cannot “drift” to any infeasible parameter combination. This fundamentally guarantees the industrial manufacturability of the final output.

[0046] S1023: Dynamically compute the adaptive reward margin.

[0047] This is the core innovation of the algorithm compared to the standard DPO, aiming to solve the training instability problem that may be caused by the fixed reward margin in the standard DPO.

[0048] Preferably, the computation of the adaptive reward margin is achieved by constructing an implicit, dynamically adjusted reference model .

[0049] Specifically, this implicit reference model is not an independent model that needs to be trained, but dynamically integrates the policy distribution of the policy model that is currently being updated and the policy distribution of the fixed reference model through interpolation. Its mathematical expression is related to the probability ratio of the policy model and the reference model on a specific sample .

[0050] The design is that the reward difference between the preference vector and the abandoned vector is no longer a fixed value, but can be adaptively adjusted according to the “confidence” difference of the model on different samples. For example, if for a certain type of user (such as a user with long hair and sensitive to noise), the preference data almost overwhelmingly chooses the combination of high wind speed and low noise, it indicates that the preference signal is very strong. At this time, the implicit reference model will automatically adjust so that the reward margin of this sample increases, thereby giving stronger “push” during gradient update, forcing the policy model to optimize faster and more firmly in the direction of strong preference. Conversely, if the difference between a certain preference pair is small and the user's choice is ambiguous, the reward margin will be reduced accordingly to avoid over-optimization of the model on ambiguous signals.

[0051] S1024: Train by minimizing the loss function.

[0052] The driving force of the entire training process is to minimize a carefully designed loss function. Preferably, the step of training the strategy model is performed by minimizing a loss function defined as follows: Wherein, the detailed description of each parameter is as follows: : represents the entire : represents the entire : represents the entire

[0053] : represents the entire : represents the entire : represents the entire : represents the entire

[0054] : represents the entire : represents the entire : represents the entire : represents the entire : represents the entire : represents the entire : represents the entire : represents the entire : represents the entire : represents the entire : represents the entire : represents the entire : represents the entire : represents the entire : represents the entire : represents the entire : represents the entire

[0055] The core of this loss function is that for each preference pair , it maximizes the difference in the logarithmic probability of the preference response relative to the abandoned response , and this difference is adjusted by the adaptive reward margin (reflected by the term). By using optimization algorithms such as gradient descent (e.g., Adam optimizer) to continuously iterate and update the parameters of the strategy model , the value of the loss function is continuously reduced, and the ability of the strategy model is also improved, until convergence.

[0056] S103: After the policy model converges, a set of product parameter combinations with the maximum posterior probability is determined from the candidate parameter space according to the optimized policy model, and is output as the key production parameter.

[0057] When the model training is completed and converges, an optimized policy model that can accurately reflect user preferences and whose output strictly complies with production constraints is obtained . The final step is to use this model to derive the final conclusion.

[0058] S1031: Determine a set of product parameter combinations with the maximum posterior probability (Maximum A Posteriori, MAP).

[0059] For a given user context (e.g., a vector representing a "long hair, noise sensitive" user), input this into the optimized policy model . The model will output the probability of each solution in the candidate parameter space (i.e., 48 BOM solutions).

[0060] The process of determining the key production parameter is to find the solution with the highest probability from the 48 probability values. This process is called maximum posterior probability decoding. The mathematical expression is: where is the key production parameter combination for a specific user group that is finally mined.

[0061] S1032: Output the key production parameter and support downstream applications.

[0062] Preferably, the present application not only outputs a single optimal solution, but also provides more comprehensive information to support different business scenarios.

[0063] S10321: Output the key production parameter table and realize automatic production docking. The found in the above step will be formatted into a clear "key production parameter table". As shown in Table 1: Preferably, the output key production parameter is formatted into data that complies with the interface specifications of the programmable logic controller (PLC) or ERP system of the production line, and is directly written into the corresponding device parameter area or generates a production order. This realizes seamless docking from user demand mining to automatic production instruction issuance, greatly improving the conversion efficiency from research and development to production.

[0064] S10322: output the complete strategy distribution vector and support personalized recommendation.

[0065] Preferably, while outputting the key production parameters, the method also outputs the complete strategy distribution vector after the strategy model converges. This vector is a list containing all 48 product parameter combinations in the candidate parameter space and their corresponding preference probabilities.

[0066] The strategy distribution vector is stored in the database to support downstream personalized product recommendation applications. For example, on the product customization page of an e-commerce platform: When a user identified as "long hair, noise sensitive" accesses the page, the system retrieves the corresponding strategy distribution vector from the database based on the user's context information. Then, according to the preference probabilities in the vector in descending order, the user is preferentially shown the top several (e.g., 3) product parameter combinations as personalized recommendation options. This achieves "thousand faces for a thousand people" precise recommendation for different user preferences, significantly improving user experience and conversion rate.

[0067] So far, this embodiment has fully demonstrated how to use the method proposed in this invention to collect user preference data, through algorithm training and optimization, finally mine the key production parameters that can maximize user satisfaction and meet industrial production constraints, and apply them to automated production and personalized recommendation.

[0068] Please refer to Figure 2 , the second aspect, the embodiments of the present invention also propose a product key production parameter mining system based on user demand classification, which is suitable for the product key production parameter mining method based on user demand classification described in the foregoing embodiments. This system can be deployed on a single server or a distributed server cluster, and can realize full-process automation from user preference data collection to key production parameter output, providing efficient and accurate decision support for product research and development and intelligent manufacturing. The following will be described in detail in combination with the intelligent hair dryer research and development scene in the foregoing embodiments. The system includes a data acquisition module, a parameter mining module, and a result generation module, wherein each module works cooperatively to form an integrated data-driven decision framework.

[0069] The data acquisition module is used to acquire user preference data containing paired preference samples, and a candidate parameter space composed of multiple groups of product parameter combinations that can be mass-produced under existing production process constraints.

[0070] Specifically, this module is the data input hub of the entire system, which can be further divided into multiple sub-units that work collaboratively. Among them, the user interaction unit is responsible for generating and displaying the front-end interface for data collection, such as the questionnaire interface mentioned in the previous embodiment, which is used for forced choice selection by users in the professional user experience laboratory. This interface clearly shows the user the paired product parameter combinations (such as A and B schemes), and receives the user's selection input. The data recording and association unit is tightly coupled with the user interaction unit, and is used to record the user's every choice in real time and accurately. This unit marks the user's selected parameter combination as the preferred parameter combination vector , marks the abandoned parameter combination as the abandoned parameter combination vector , and automatically binds these selection results with the pre-collected user context information x (such as hair quality, noise sensitivity, etc.) to form a structured triple , and finally gathers into a user preference dataset . The production constraint interface unit is responsible for communicating with the production management system (such as the ERP system) within the enterprise, and is used to obtain and maintain an up-to-date candidate parameter space composed of multiple groups of directly producible product parameter combinations. In the hair dryer scene of the previous embodiment, this unit obtains a discrete table composed of 48 different BOM schemes. The data preprocessing unit, as the last checkpoint before the data flows to the next module, performs cleaning, normalization, and format conversion on the collected user preference data, ensuring that the data input into the parameter mining module is high-quality, standardized paired tensor, laying a solid foundation for subsequent model training.

[0071] The parameter mining module is used to train the strategy model based on user preference data and candidate parameter space, using an improved adaptive direct preference optimization algorithm.

[0072] This module is the core computing engine of the system, responsible for performing complex model training tasks. Its internal integration is based on a deep learning framework (such as PyTorch or TensorFlow) to implement algorithm core. This module is specially configured to perform the following key functions: first, at the beginning of training, the model initialization unit will build an initial strategy model , and create a copy of its parameters as a reference model to provide a stable benchmark for the entire optimization process. Secondly, during the training iteration process, the training controller will send the preprocessed data received from the data acquisition module to the algorithm core in batches. This core is configured to: for each paired preference sample, by constructing an implicit, dynamically adjusted reference model , to dynamically calculate the adaptive reward margin. This mechanism enables the model to intelligently adjust the update gradient of its parameters according to the degree of differentiation of the samples, effectively avoiding the problems of reward hijacking and mode collapse. At the same time, the core is also configured to strictly impose the Kullback-Leibler divergence constraint during optimization, by punishing the excessive deviation of the policy model from the reference model, to ensure that its parameter search space is strictly limited within the set of producible solutions defined by the candidate parameter space. Finally, the module continuously iteratively updates the parameters of the policy model through optimization algorithms such as gradient descent, until the model performance converges.

[0073] A result generation module is used to determine a set of product parameter combinations with the maximum posterior probability from the candidate parameter space according to the optimized policy model after the policy model converges, and output as the key production parameters.

[0074] The module is the decision output end of the system, responsible for converting the trained model into production instructions with commercial value and business insights. Its internal contains a model inference unit that loads the final policy model trained and converged by the parameter mining module When receiving specific user context information , the unit performs a forward propagation calculation, outputting a preference probability for each product parameter combination in the candidate parameter space (such as the 48 BOM solutions in the blow-drying scene). Subsequently, the MAP decoding unit analyzes these probability values, automatically identifies and extracts the product parameter combination with the maximum posterior probability as the optimal solution for the user group. Finally, the output interface unit is responsible for formatting and distributing the optimal solution. On the one hand, it generates the parameter combination into data that conforms to the interface specifications of the production line PLC or ERP system, realizing the automatic issuance of instructions to the production line. On the other hand, it also stores the complete strategy distribution vector (i.e. all candidate solutions and their corresponding probabilities) to the database and provides it to the downstream personalized product recommendation system through the Application Programming Interface (API) interface, to support advanced applications such as "thousand faces for a thousand people" recommendation.

[0075] In summary, the system proposed by the embodiment of the present application builds an automated and intelligent mining link from raw user preferences to final producible key parameters through the seamless cooperation of the above modules, and can efficiently and accurately solve the problem of aligning product design with user demand in modern manufacturing.

[0076] Please refer to Figure 3In a third aspect, an embodiment of the present application provides a product key production parameter mining device based on user demand classification, comprising a memory and a processor connected in sequence and in communication, wherein the memory is configured to store a computer program, and the processor is configured to read the computer program and execute a product key production parameter mining method based on user demand classification according to the first aspect. Specifically, the memory can include, but is not limited to, a random access memory (RAM), a read-only memory (ROM), a flash memory, a first-in first-out memory (FIFO), a first-in last-out memory (FILO), and the like; the processor can be, but is not limited to, a microprocessor of STM32F105 series, an ARM (Advanced RISC Machines) processor, an X86 architecture processor, or a processor integrated with an NPU (neural-network processing units). The working process, working details, and technical effects of the device provided in the third aspect of the present embodiment can be referred to the first aspect, and will not be repeated here.

[0077] In a fourth aspect, an embodiment of the present application provides a computer readable storage medium storing instructions of a product key production parameter mining method based on user demand classification according to the first aspect, i.e., the computer readable storage medium stores instructions, and when the instructions run on a computer, the computer executes a product key production parameter mining method based on user demand classification according to the first aspect. The computer readable storage medium refers to a carrier storing data, which can include, but is not limited to, a floppy disk, an optical disc, a hard disk, a flash memory, a USB flash disk, a memory stick, and the like. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device systems. The working process, working details, and technical effects of the computer readable storage medium provided in the fourth aspect of the present embodiment can be referred to the first aspect, and will not be repeated here.

[0078] In a fifth aspect, an embodiment of the present application provides a computer program product containing instructions, which, when running on a computer, causes the computer to execute a product key production parameter mining method based on user demand classification according to the first aspect. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device systems.

[0079] The above-described embodiments are merely illustrative for the present application, wherein the units illustrated as separated components can or can not be physically separated, and the components illustrated as units can or can not be physical units, i.e., can be located in one place, or can be distributed to multiple network units. Part or all of the modules can be selected to achieve the purpose of the present embodiment according to actual needs. Those skilled in the art can understand and implement it without creative labor.

[0080] Through the description of the above embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software and necessary universal hardware platforms, and of course can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of software products, and the computer software products can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and include a plurality of instructions to make a computer or programmable device execute the methods of the embodiments or some parts of the embodiments.

[0081] Finally, it should be noted that: although the embodiments of the present application have been shown and described, those skilled in the art can understand that various changes, modifications, replacements and variations can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the claims and their equivalents.

Claims

1. A product key production parameter mining method based on user demand classification, characterized in that, The method comprises: obtaining user preference data containing paired preference samples, and a candidate parameter space composed of multiple groups of product parameter combinations that can be mass-produced under existing production process constraints; based on the user preference data and the candidate parameter space, training a strategy model using an improved adaptive direct preference optimization algorithm, the steps of training the strategy model comprising: for each group of paired preference samples, dynamically calculating an adaptive reward margin to adjust the update gradient of the model parameters of the strategy model according to the differentiation degree of the samples; and imposing a Kullback-Leibler divergence constraint on the optimization of the strategy model to limit the parameter search space of the strategy model within the candidate parameter space; after the strategy model converges, determining a group of product parameter combinations with the maximum posterior probability from the candidate parameter space according to the optimized strategy model, and outputting the group of product parameter combinations as the key production parameters.

2. The method of claim 1, wherein, The step of obtaining user preference data containing paired preference samples comprises: showing paired product parameter combinations to a target user group for forced two-choice selection by the users; record the selection result of each user in each pair of product parameter combination, define the parameter combination selected by the user as a preferred parameter combination vector define the parameter combination abandoned by the user as an abandoned parameter combination vector ; and combining the preference parameter combination vector and the discarded parameter combination vector with corresponding user context information correlation, forming a plurality of sets of triplets in the form constituting the user preference dataset .

3. The method of claim 2, wherein, Before the step of training the strategy model using the improved adaptive direct preference optimization algorithm, the user preference data is preprocessed, and the preprocessing step comprises: a preference parameter combination vector and a discarded parameter combination vector Data cleaning is performed on the multi-dimensional feature values in the preference parameter combination vector and the discarded parameter combination vector to remove invalid or abnormal data. and normalizing the cleaned multi-dimensional feature values to scale parameters of different dimensions and value ranges to a unified numerical interval to eliminate the influence of feature scale differences on model training, and constructing the normalized data into paired tensors for input of the strategy model.

4. The method of claim 1, wherein, The process of training employs a reference model The reference model is an initial policy model copy whose parameters are frozen before training starts, whose policy distribution represents a priori engineering knowledge or a feasible distribution of initial production parameters before user preference alignment is performed; The Korbek-Leibler divergence constraint is implemented through a penalty policy model. For the reference model Excessive deviation will concentrate the output probability quality of the strategy model on discrete combinations within the candidate parameter space, thereby ensuring that the key production parameter combination of the output is one of the predefined mass-producible solutions.

5. The method of claim 4, wherein, The calculation of the adaptive reward margin is realized by constructing an implicit and dynamically adjusted reference model , which fuses the policy distribution of the current policy model and the policy distribution of the reference model by interpolation, and the mathematical expression thereof is associated with the probability ratio of the policy model and the reference model on a specific sample , so that the reward margin can be adaptively adjusted according to the confidence difference of the model on different samples.

6. The method of claim 5, wherein, The step of training the strategy model is performed by minimizing the loss function defined as follows: wherein, is computed on a user preference dataset representing the expectation, said user preference dataset is composed of multiple sets of triplets, each set of triplets comprising a user context information , a user-selected preference parameter combination vector and a discarded parameter combination vector abandoned by the user; is a Sigmoid activation function; is a temperature hyperparameter that controls the scale of the reward function; is the probability of the parameter combination generated by the policy model to be optimized given the user context ; is the probability of the parameter combination generated by the implicit reference model; this loss function drives the model optimization by maximizing the difference in log-probabilities of the preference response and the discarded response.

7. The method of claim 1, wherein, The method further comprises: outputting the complete strategy distribution vector of the strategy model after convergence together with the key production parameters; the strategy distribution vector is a list containing all product parameter combinations in the candidate parameter space and their corresponding preference probabilities; and storing the strategy distribution vector in a database to support downstream personalized product recommendation applications.

8. The method of claim 7, wherein, The step of supporting downstream personalized product recommendation applications comprises: When a user accesses a product customization or recommendation page, the system retrieves the corresponding strategy distribution vector from the database according to the context information of the user . and ranking the top several groups of product parameter combinations in the strategy distribution vector in descending order of preference probability to show the users as personalized recommendation options.

9. The method of claim 1, wherein, The output key production parameters are formatted to comply with the interface specifications of programmable logic controllers (PLCs) or enterprise resource planning (ERP) systems on production lines, and are directly written into the corresponding device parameter area or production orders are generated.

10. A product key production parameter mining system based on user demand classification, characterized by, The system comprises: a data acquisition module for obtaining user preference data containing paired preference samples, and a candidate parameter space composed of multiple groups of product parameter combinations that can be mass-produced under existing production process constraints; a parameter mining module for training a strategy model based on the user preference data and the candidate parameter space using an improved adaptive direct preference optimization algorithm, wherein the parameter mining module is configured to: In the training process, for each set of paired preference samples, an adaptive reward margin is dynamically calculated to adjust the update gradient of the model parameters of the policy model according to the differentiation degree of the samples; And, in the training process, a Kullback-Leibler divergence constraint is imposed on the optimization of the policy model, and the parameter search space of the policy model is limited within the candidate parameter space; And a result generation module is configured to, after the policy model converges, determine a set of product parameter combinations with the maximum posterior probability from the candidate parameter space according to the optimized policy model, and output the set of product parameter combinations as the key production parameters.

Images