Momordica grosvenori mooncake process optimization method and device

Abstract

The invention belongs to the technical field of food processing, and discloses a momordica grosvenori mooncake process optimization method and device, and the method comprises the steps: capturing user evaluation text data of momordica grosvenori mooncakes from an e-commerce platform, and taking the user evaluation text data as evaluation data; a natural language processing technology is utilized to extract taste features related to taste from the evaluation data, and a multi-dimensional taste feature vector is constructed; introducing a time window, and calculating a moving average value of the taste feature vector in a past period of time to obtain a dynamic average taste feature vector; calculating the deviation between the dynamic average taste feature vector and the target taste feature vector to obtain a taste deviation vector; calculating to obtain a taste feature vector adjustment amount according to the taste deviation vector by adopting a proportional-integral control algorithm; determining a production process parameter adjustment amount according to the taste feature vector adjustment amount so as to adjust production process parameters; therefore, the dynamic adjustment of production process parameters based on user feedback can be realized, and the market competitiveness of the mouth feel of the momordica grosvenori mooncake is improved.

Description

Technical Field

[0001] This application relates to the field of food processing technology, and more specifically, to a method and apparatus for optimizing the processing of monk fruit mooncakes. Background Technology

[0002] Currently, quality inspection systems are widely used in intelligent food production lines to monitor product physical parameters and ensure that products meet factory standards. To further enhance market competitiveness, more and more manufacturers are actively seeking to integrate consumer feedback, especially user reviews from e-commerce platforms, to optimize production processes and achieve a more flexible production model that precisely meets market demands. In the niche market of monk fruit mooncakes, intelligent production lines are also gradually exploring how to effectively utilize user taste reviews on e-commerce platforms to guide production adjustments, striving to quickly respond to subtle changes in consumer tastes and thus enhance product competitiveness.

[0003] However, directly adopting and utilizing online user reviews to optimize production processes still faces numerous technical bottlenecks in practice. First, user review data is essentially unstructured natural language text, with highly subjective and diverse content, making it extremely difficult to accurately extract precise instructions for adjusting process parameters from massive amounts of reviews. Second, the taste preferences reflected in user reviews exhibit significant real-time and fluctuating characteristics. Consumer taste preferences may change rapidly over time. If the production line frequently adjusts process parameters based on short-term review fluctuations, it can easily lead to production instability and consequently, product quality deviations. Furthermore, how to achieve a stable and reliable dynamic adjustment mechanism solely through software upgrades based on existing production line hardware is also a critical issue that urgently needs to be addressed. Especially for foods like monk fruit mooncakes, which have refined taste requirements, effectively ensuring the stability of large-scale production and a high degree of product quality consistency while rapidly responding to changes in user taste has become the core of enhancing product market competitiveness.

[0004] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0005] The purpose of this application is to provide a method and apparatus for optimizing the production process of monk fruit mooncakes. This method can utilize user review data from e-commerce platforms to extract taste features from unstructured text and consider dynamic changes in user preferences, thereby enabling dynamic adjustment of production process parameters based on user feedback and enhancing the market competitiveness of monk fruit mooncakes in terms of taste.

[0006] Firstly, this application provides a method for optimizing the production process of monk fruit mooncakes, used to optimize the production process parameters of monk fruit mooncakes to improve their taste. The steps of this method include:

[0007] A1. Extract user review text data of monk fruit mooncakes from e-commerce platforms as review data;

[0008] A2. Using natural language processing technology, extract taste-related features from the evaluation data and construct a multi-dimensional taste feature vector;

[0009] A3. Introduce a time window and calculate the moving average of the taste feature vector over a past period to obtain the dynamic average taste feature vector;

[0010] A4. Calculate the deviation between the dynamic average taste feature vector and the target taste feature vector to obtain the taste deviation vector;

[0011] A5. Using a proportional-integral control algorithm, the adjustment amount of the taste feature vector is calculated based on the taste deviation vector;

[0012] A6. Determine the adjustment amount of the production process parameters based on the adjustment amount of the aforementioned taste feature vector;

[0013] A7. Adjust the production process parameters of the monk fruit mooncake according to the aforementioned production process parameter adjustment amount.

[0014] This method can utilize user review data from e-commerce platforms to extract taste features from unstructured text and take into account dynamic changes in user preferences, thereby enabling dynamic adjustment of production process parameters based on user feedback and enhancing the market competitiveness of monk fruit mooncakes in terms of taste.

[0015] Preferably, step A2 includes:

[0016] A201. The evaluation data is preprocessed to remove interference information, resulting in preprocessed text data; the interference information includes HTML tags, special characters, and stop words;

[0017] A202. Encode the preprocessed text data using a pre-trained BERT model to obtain the corresponding BERT vector;

[0018] A203. Based on a taste dictionary containing multiple taste keywords, calculate the cosine similarity between the BERT vector of each preprocessed text data and the word vector of each keyword in the taste dictionary to obtain a taste similarity vector;

[0019] A204. Based on the taste similarity vector, use a pre-trained taste feature extraction model to extract taste features and construct a multi-dimensional taste feature vector.

[0020] The taste feature vector extracted using the above method can more accurately and comprehensively reflect the taste characteristics of monk fruit mooncakes.

[0021] Preferably, after step A204, the method further includes the following step:

[0022] A205. Based on the knowledge graph of monk fruit mooncake taste, and combined with monk fruit feature information, the multidimensional taste feature vector is modified to highlight the influence of monk fruit's characteristic taste on the overall taste, thus obtaining the final taste feature vector; the knowledge graph contains the correlation between monk fruit feature information and taste features, and the monk fruit feature information includes at least one of monk fruit variety, origin, and harvesting time.

[0023] After correction, the final taste feature vector is obtained, which can more accurately reflect the unique taste of monk fruit mooncakes.

[0024] Preferably, after step A1 and before step A2, the following step is also included:

[0025] A1a. Identify evaluation data containing keywords that indicate abnormal evaluations and calculate the sentiment polarity value of the corresponding evaluation data;

[0026] A1b. If the emotional polarity value is less than the preset negative emotional threshold, the corresponding evaluation data is determined to be malicious evaluation data and is removed.

[0027] A1c. Identify review data containing advertising links or contact information, determine the corresponding review data as advertising review data, and remove it.

[0028] Preferably, step A3 includes:

[0029] A301. Obtain the taste feature vector within the time window, and calculate the weight value of each taste feature vector within the time window according to the time decay function. The closer the time is to the current moment, the greater the weight value.

[0030] A302. Based on the taste feature vector and its weight value within the time window, a weighted average algorithm is used to calculate the dynamic average taste feature vector.

[0031] Preferably, before step A301, the method further includes the following step:

[0032] A303. Monitor the user review data stream of the e-commerce platform in real time, calculate the number of new reviews per unit time, and obtain the review density;

[0033] A304. Based on the evaluation density, the size of the time window is dynamically adjusted using an adaptive adjustment algorithm. When the evaluation density is higher than a preset evaluation density threshold, the time window is reduced; when the evaluation density is lower than the preset evaluation density threshold, the time window is expanded.

[0034] Preferably, step A5 includes:

[0035] A501. Obtain historical adjustment data for each production process parameter, analyze the correlation between each production process parameter, and construct a process parameter coupling relationship matrix;

[0036] A502. Based on the process parameter coupling relationship matrix, a matrix decoupling algorithm is used to decouple the proportional-integral control algorithm to obtain the decoupled proportional-integral control algorithm.

[0037] A503. Using the decoupled proportional-integral control algorithm, the adjustment amount of the taste feature vector is calculated based on the taste deviation vector and the coupling relationship matrix of the process parameters.

[0038] Preferably, step A6 includes:

[0039] A601. Input the adjustment amount of the taste feature vector into the pre-trained mapping model to obtain the initial adjustment amount of each production process parameter;

[0040] A602. Determine whether the adjustment amount of each initial production process parameter exceeds the corresponding preset constraint range;

[0041] A603. If there are initial production process parameter adjustments that exceed the constraint range, calculate the distance between each initial production process parameter adjustment and the corresponding constraint range boundary to obtain the out-of-bounds distance vector.

[0042] A604. Based on the aforementioned cross-boundary distance vector, the adjustment amounts of each initial production process parameter are corrected to minimize the weighted norm of the taste deviation vector, so that the dynamic average taste feature vector approaches the target taste feature vector.

[0043] Preferably, step A604 includes:

[0044] The initial production process parameter adjustment amount that exceeds the constraint range is corrected to the corresponding constraint boundary value to obtain the corrected production process parameter adjustment amount.

[0045] For the initial production process parameter adjustment amount that does not exceed the constraint range, with the goal of minimizing the weighted norm of the taste deviation vector, a quadratic programming algorithm is adopted to adjust the initial production process parameter adjustment amount that does not exceed the constraint range according to the importance of taste features, so as to obtain the optimized production process parameter adjustment amount.

[0046] The corrected production process parameter adjustment amount and the optimized production process parameter adjustment amount are combined to obtain the final production process parameter adjustment amount.

[0047] Secondly, this application provides a process optimization device for monk fruit mooncakes, used to optimize the production process parameters of monk fruit mooncakes to improve their taste. The device includes:

[0048] The evaluation data collection module is used to extract user review text data of monk fruit mooncakes from e-commerce platforms as evaluation data.

[0049] The taste feature extraction module is used to extract taste-related features from the evaluation data using natural language processing technology, and to construct a multi-dimensional taste feature vector.

[0050] The dynamic average calculation module is used to introduce a time window and calculate the moving average of the taste feature vector over a period of time to obtain the dynamic average taste feature vector.

[0051] The deviation calculation module is used to calculate the deviation between the dynamic average taste feature vector and the target taste feature vector to obtain the taste deviation vector.

[0052] The feature adjustment calculation module is used to calculate the adjustment amount of the taste feature vector based on the taste deviation vector using a proportional-integral control algorithm.

[0053] The parameter adjustment calculation module is used to determine the adjustment amount of the production process parameters based on the adjustment amount of the taste feature vector.

[0054] The adjustment execution module is used to adjust the production process parameters of the monk fruit mooncake according to the adjustment amount of the production process parameters.

[0055] Beneficial effects: The method and apparatus for optimizing the production process of monk fruit mooncakes provided in this application utilize user review data from e-commerce platforms to extract taste features from unstructured text and consider the dynamic changes in user preferences, thereby achieving dynamic adjustment of production process parameters based on user feedback and enhancing the market competitiveness of monk fruit mooncakes in terms of taste. Attached Figure Description

[0056] Figure 1 A flowchart illustrating the process optimization method for monk fruit mooncakes provided in this application embodiment.

[0057] Figure 2 This is a schematic diagram of the process optimization device for monk fruit mooncakes provided in an embodiment of this application.

[0058] Labeling Explanation: 1. Evaluation Data Acquisition Module; 2. Taste Feature Extraction Module; 3. Dynamic Average Calculation Module; 4. Deviation Calculation Module; 5. Feature Adjustment Calculation Module; 6. Parameter Adjustment Calculation Module; 7. Adjustment Execution Module. Detailed Implementation

[0059] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0060] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0061] refer to Figure 1 This application proposes a method for optimizing the production process of monk fruit mooncakes, which aims to improve the taste of monk fruit mooncakes by optimizing the production process parameters. The method includes the following steps:

[0062] A1. Extract user review text data of monk fruit mooncakes from e-commerce platforms as review data;

[0063] A2. Using natural language processing technology, extract taste-related features from the evaluation data and construct a multi-dimensional taste feature vector;

[0064] A3. Introduce a time window and calculate the moving average of the taste feature vector over a period of time to obtain the dynamic average taste feature vector;

[0065] A4. Calculate the deviation between the dynamic average taste feature vector and the target taste feature vector to obtain the taste deviation vector;

[0066] A5. A proportional-integral control algorithm is used to calculate the adjustment amount of the taste feature vector based on the taste deviation vector;

[0067] A6. Determine the adjustment amount of production process parameters based on the adjustment amount of the taste feature vector;

[0068] A7. Adjust the production process parameters of Luo Han Guo mooncakes according to the adjustment amount of the production process parameters.

[0069] In step A1, user reviews of monk fruit mooncakes are obtained from e-commerce platforms. This can be achieved using web scraping techniques, such as the Python Scrapy framework or BeautifulSoup library, to periodically collect user review data for monk fruit mooncakes from mainstream e-commerce platforms like Taobao, JD.com, and Pinduoduo. The scraped data can include review text, user ratings, and review dates. This data forms the foundation for subsequent taste analysis and process optimization, ensuring that the optimization process is based on genuine user feedback.

[0070] In step A2, natural language processing (NLP) techniques are used to analyze user reviews, extract and quantify taste features. This can be achieved through the following process: First, the captured review text is preprocessed, such as removing HTML tags, special characters, and stop words. Then, a pre-trained word vector model, such as BERT, Word2Vec, or GloVe, is used to convert the text data into vector representations. Next, a taste dictionary is constructed, containing keywords describing the taste of mooncakes, such as "sweetness," "soft and glutinous," "delicate," and "refreshing." By calculating the similarity between the review text vector and the keyword vectors in the taste dictionary, taste-related features can be extracted from the review text. Finally, the extracted taste features are integrated to construct a multi-dimensional taste feature vector; for example, a five-dimensional vector can be used to represent the five taste dimensions: "sweetness," "glutinousness," "delicateness," "refreshing," and "monk fruit flavor." Through NLP, users' subjective taste descriptions can be transformed into objective and quantifiable data, providing a basis for subsequent adjustments to process parameters.

[0071] In step A3, a time window is introduced to smooth out fluctuations in taste characteristics. Specifically, a time window can be set, such as the most recent week or the most recent month. Then, all taste feature vectors within the time window are collected, and the moving average of these vectors is calculated. The moving average can be calculated using a simple averaging method or a weighted averaging method; for example, evaluation data more recent to the current time can be given higher weights. By calculating the moving average, a dynamic average taste feature vector can be obtained, which reflects consumers' average taste preferences over a period of time, effectively filtering out short-term evaluation fluctuations and avoiding frequent adjustments to the production process.

[0072] In step A4, the difference between the current product's taste and the target taste is calculated. Specifically, a target taste feature vector is pre-defined, representing the desired product taste standard. This vector can be determined based on market research, expert experience, or historical best-in-class data. Then, the dynamic average taste feature vector obtained in step A3 is compared with the target taste feature vector, and the deviation between them is calculated. The deviation can be calculated using vector subtraction or by calculating the distance between vectors, such as Euclidean distance or cosine distance. The calculated taste deviation vector clearly indicates the gap between the current product's taste and the target taste, providing a clear direction and quantitative indicator for subsequent process adjustments.

[0073] In step A5, a proportional-integral (PI) control algorithm is used to calculate the adjustment amount of the taste feature vector. PI control is a commonly used control algorithm that outputs a corresponding control quantity based on the magnitude and trend of the deviation. In this application, the deviation is the taste deviation vector obtained in step A4, and the control quantity is the adjustment amount of the taste feature vector. Specifically, the proportional control part outputs an adjustment amount proportional to the current magnitude of the taste deviation vector; the integral control part eliminates steady-state error based on the accumulated deviation over a period of time. Through the PI control algorithm, precise and stable adjustment of the taste feature vector can be achieved.

[0074] In step A6, the adjustment amount of the taste feature vector is converted into specific production process parameter adjustments. This requires establishing a mapping relationship between taste features and production process parameters. For example, the "sweetness" dimension in the taste feature vector may be related to the "sugar addition amount" and "monk fruit extract addition amount" in the production process parameters; the "glutinousness" dimension may be related to the "glutinous rice flour ratio" and "cooking time". The mapping relationship can be established through experimental analysis, expert experience, or machine learning methods. Based on the established mapping relationship, the taste feature vector adjustment amount calculated in step A5 can be converted into specific production process parameter adjustments, such as increasing the sugar addition by 2 grams, increasing the glutinous rice flour ratio by 5%, and reducing the cooking time by 5 minutes.

[0075] In step A7, the production process parameters for the monk fruit mooncakes are adjusted based on the adjustments calculated in step A6. Specific production process parameters may include, but are not limited to, some or all of the following: monk fruit extract addition, sugar addition, glutinous rice flour ratio, wheat flour ratio, moisture content, steaming time, baking temperature, and baking time. Adjusting these production process parameters can alter the texture of the monk fruit mooncakes, making them closer to the target texture. The execution of step A7 completes the closed-loop control of the entire process optimization method, enabling the production process to be dynamically adjusted based on user feedback, continuously improving the texture of the monk fruit mooncakes.

[0076] Specifically, the method provided in this application first collects user review data from e-commerce platforms in step A1, providing a data foundation for subsequent optimization. Step A2 uses natural language processing technology to extract taste features from unstructured text, achieving objective quantification of users' subjective feelings. Step A3 introduces a time window to smooth short-term fluctuations, reflect long-term trends, and ensure the stability of process adjustments. Step A4 calculates taste deviations, clarifying the optimization direction. Step A5 employs a proportional-integral control algorithm to achieve automated process adjustments. Steps A6 and A7 establish the relationship between taste features and process parameters, and implement the adjustments into the production process, ultimately achieving dynamic optimization of the monk fruit mooncake process based on user feedback. This method can quickly respond to changes in consumer tastes and enhance product competitiveness.

[0077] Through the above technical solution, this application realizes the dynamic optimization of the production process parameters of monk fruit mooncakes based on user evaluation data from e-commerce platforms. This enables the production process of monk fruit mooncakes to be adjusted in real time according to user taste feedback, continuously improving the taste of the mooncakes, more accurately meeting consumer taste needs, and enhancing the market competitiveness of the product.

[0078] Preferably, step A2 includes:

[0079] A201. The evaluation data is preprocessed to remove interference information, resulting in preprocessed text data; interference information includes HTML tags, special characters, and stop words;

[0080] A202. Encode the pre-trained BERT (Bidirectional Encoder Representations from Transformers) model to obtain the corresponding BERT vectors;

[0081] A203. Based on a taste dictionary containing multiple taste keywords, calculate the cosine similarity between the BERT vector of each preprocessed text data and the word vector of each keyword in the taste dictionary to obtain the taste similarity vector;

[0082] A204. Based on the taste similarity vector, use a pre-trained taste feature extraction model to extract taste features and construct a multi-dimensional taste feature vector.

[0083] Among them, in step A201, preliminary data cleaning operations are performed on the collected evaluation data, aiming to improve the quality and efficiency of subsequent data processing. Specifically, methods such as regular expression matching, string replacement, and stop word list filtering can be used to remove HTML tags, special characters, and stop words. For example, HTML tags can be matched and removed by the pattern "<.*?>", special characters can be CJK punctuation marks, full-width characters, or Unicode control characters, and stop words can be high-frequency words without practical meaning such as "的", "地", and "得". Through preprocessing, the noise information in the evaluation data is effectively reduced, and the purity and normality of the text data are ensured.

[0084] Among them, in step A202, the preprocessed text data is vectorized and characterized using the BERT model, aiming to convert text information into a numerical form that can be understood and calculated by machines. Specifically, a pre-trained Chinese BERT-base model, BERT-wwm model, or RoBERTa model can be used to encode the text data. The BERT model can deeply capture the semantic information and context relationships of the text. Compared with traditional bag-of-words models or recurrent neural networks, the BERT model can represent the semantics of the text more comprehensively and accurately. Thus, each text data is converted into a fixed-length BERT vector, providing high-quality vectorized input for subsequent similarity calculation and feature extraction.

[0085] Among them, in step A203, the cosine similarity between the BERT vector of each text data and the keyword vector in the taste dictionary is calculated, aiming to quantify the degree of correlation between the evaluation data and the taste characteristics. Specifically, a pre-constructed dictionary containing taste keywords such as sweet, greasy, soft and glutinous, crispy, etc. can be used, and the cosine similarity between the BERT vector of each text and the BERT vector of each keyword in the dictionary is calculated. Cosine similarity measures the similarity of two vectors in direction. The closer its value is to 1, the more similar the vectors are; the closer it is to -1, the less similar the vectors are; and the closer it is to 0, the smaller the correlation between the vectors. By calculating the cosine similarity, the similarity score of each text data with respect to different taste keywords in the taste dictionary can be obtained, forming a taste similarity vector (the similarity scores corresponding to each keyword for the same text data form a taste similarity vector), which is used to characterize the tendency of the evaluation data in different taste dimensions.

[0086] In step A204, a pre-trained taste feature extraction model is used to extract taste features based on the taste similarity vector, constructing a multi-dimensional taste feature vector. The aim is to extract high-level, abstract taste features from the quantified taste similarity vector. Specifically, a pre-trained neural network model, support vector machine model, or gradient boosting tree model can be used to extract these taste features. The taste feature extraction model takes the taste similarity vector as input and learns the mapping relationship between the taste similarity vector and the taste features, thus automatically extracting multi-dimensional taste features from the taste similarity vector, such as sweetness, richness, and firmness. The constructed multi-dimensional taste feature vector can comprehensively represent the user's evaluation of the taste of the monk fruit mooncake, providing quantitative indicators for subsequent process optimization. For example, the pre-trained taste feature extraction model uses a deep learning model containing a 3-layer fully connected neural network, with 100 nodes in the input layer, 256 and 128 nodes in the hidden layers, and 64 nodes in the output layer. The ReLU activation function is used. The training dataset for the texture feature extraction model consists of texture similarity vectors corresponding to manually labeled monk fruit mooncake evaluation data. Each evaluation data point is labeled with a 64-dimensional texture feature vector. The texture feature extraction model is trained using the Adam optimizer and mean squared error loss function. After training, the texture similarity vector obtained in step A203 is input into the trained texture feature extraction model to obtain a 64-dimensional texture feature vector.

[0087] Specifically, this solution defines the implementation of step A2. First, step A201 performs preliminary cleaning of the e-commerce platform's review data, removing non-critical information such as HTML tags, special characters, and stop words, providing a clean and standardized data foundation for subsequent text analysis. Then, step A202 uses a pre-trained BERT model to perform deep semantic encoding on the pre-processed text data. BERT's powerful semantic understanding ensures that the review data can be accurately vectorized, overcoming the limitations of traditional text representation methods in effectively capturing contextual information. Next, in step A203, the solution combines a taste dictionary, calculating the cosine similarity between the BERT vector and the keyword vectors in the taste dictionary to achieve the correlation calculation between the review data and the preset taste dimension, quantifying the taste-level bias of the review data. Finally, step A204 uses a pre-trained taste feature extraction model, taking the taste similarity vector obtained in step A203 as input, to further extract multi-dimensional taste features from the review data and construct these features into a taste feature vector. Through the above steps, this solution provides a technical approach to effectively and accurately extract taste features from user reviews. The extracted taste feature vectors can provide data support for optimizing the production process of monk fruit mooncakes.

[0088] In some preferred embodiments, after step A204, the following step is further included:

[0089] A205. Based on the knowledge graph of monk fruit mooncake taste, and combined with the characteristic information of monk fruit, the multi-dimensional taste feature vector is modified to highlight the influence of the characteristic taste of monk fruit on the overall taste, and the final taste feature vector is obtained. The knowledge graph contains the relationship between the characteristic information of monk fruit and the taste features. The characteristic information of monk fruit includes at least one of the following: monk fruit variety, place of origin, and harvesting time.

[0090] The taste feature vector correction step adjusts the multidimensional taste feature vector constructed in step A204. A pre-constructed knowledge graph of monk fruit mooncake taste is used to store the relationships between monk fruit characteristics and taste features. The knowledge graph contains information such as monk fruit variety, origin, and harvest time, as well as descriptions of how these characteristics affect the mooncake's taste. For example, the knowledge graph might indicate that a specific variety of monk fruit brings a richer sweetness, or that monk fruit from a specific origin has a more refreshing taste. During the correction process, the monk fruit characteristic information used in the current batch of mooncakes is first obtained, such as the variety and origin. Then, based on the obtained monk fruit characteristic information, the corresponding taste feature relationships are searched in the monk fruit mooncake taste knowledge graph. Therefore, based on the relationships provided by the knowledge graph, the multidimensional taste feature vector obtained in step A204 is adjusted. The adjustment process aims to enhance the dimensions of the taste feature vector related to the monk fruit's own characteristics, thereby highlighting the impact of the monk fruit's characteristic taste on the overall taste. Finally, after correction processing, the final taste feature vector is obtained, which can more accurately reflect the unique taste of monk fruit mooncakes.

[0091] Specifically, after obtaining the multi-dimensional taste feature vector output in step A204, the system accesses a pre-constructed knowledge graph of monk fruit mooncake taste. This knowledge graph stores data on the correlation strength between different monk fruit varieties, origins, harvest times, and various taste characteristics. For example, when the monk fruit used in production is of the "Guilin Premium" variety, the system searches the knowledge graph and finds a high correlation between "Guilin Premium" monk fruit and the taste characteristics of "sweet" and "throat-soothing." At this point, for the taste feature vector output in step A204, the system increases the weights of the dimensions corresponding to "sweet" and "throat-soothing" while maintaining or decreasing the weights of other non-monk fruit characteristic taste dimensions. This weight adjustment is achieved through a weighted algorithm, with the weighting coefficients derived from the correlation strength data in the knowledge graph. Through this correction process, the initial taste feature vector is adjusted, and the adjusted vector better highlights the unique taste brought by the monk fruit variety, thus providing more targeted input for subsequent process parameter adjustments.

[0092] In some specific implementations, the knowledge graph of monk fruit mooncake taste is constructed in the form of triplets, such as (monk fruit variety - Yongfu monk fruit, taste feature - sweetness, association strength - 0.8), indicating that the association strength between the Yongfu monk fruit variety and the sweetness taste feature is 0.8. When Yongfu monk fruit is used in production, and the taste feature vector obtained in step A204 is [sweetness: 0.7, aroma: 0.6, texture: 0.8], a correction step is performed. The correction step first queries the knowledge graph to obtain the association strength of 0.8 between Yongfu monk fruit and sweetness taste. Then, based on this association strength, the taste feature vector is corrected. The correction method is to amplify the value of the "sweetness" dimension in the taste feature vector according to the association strength. For example, the value of the "sweetness" dimension can be corrected from 0.7 to 0.7*(1+0.8) = 1.26. The revised taste feature vector becomes [sweetness: 1.26, aroma: 0.6, texture: 0.8]. Thus, the varietal characteristics of monk fruit are incorporated into the taste feature vector, allowing it to more accurately express the taste characteristics of monk fruit mooncakes.

[0093] Preferably, after step A1 and before step A2, the following step may be included:

[0094] A1a. Identify evaluation data containing keywords that indicate abnormal evaluations and calculate the sentiment polarity value of the corresponding evaluation data;

[0095] A1b. If the emotional polarity value is less than the preset negative emotional threshold, the corresponding evaluation data is determined to be malicious evaluation data and is removed.

[0096] A1c. Identify review data containing advertising links or contact information, determine the corresponding review data as advertising review data, and remove it.

[0097] In step A1a, a list of abnormal review keywords can be pre-constructed, containing preset abnormal review keywords. These keywords may include, but are not limited to, "bad review," "unpalatable," and "quality problem." When the review data contains abnormal review keywords, a sentiment polarity value is calculated. The sentiment polarity value can be calculated using a pre-trained sentiment analysis model or a sentiment dictionary.

[0098] In step A1b, the negative sentiment threshold can be set according to actual needs, for example, the negative sentiment threshold can be set to -0.5. When the sentiment polarity value of the evaluation data is less than the preset negative sentiment threshold, the evaluation data is judged as malicious evaluation data and is removed from the evaluation dataset.

[0099] In step A1c, ad review data is determined by identifying whether it contains ad links or contact information. Ad links can be URLs, and contact information can be phone numbers, WeChat IDs, or email addresses. When review data contains ad links or contact information, it is identified as ad review data and removed from the review dataset.

[0100] Specifically, for the collected user review data, step A1a is first executed. The identification process is completed using a keyword matching method, with a pre-defined dictionary of abnormal review keywords used for keyword matching. The sentiment polarity value calculation process uses sentiment analysis tools, such as the SnowNLP sentiment analysis toolkit. Then, step A1b is executed. Assuming the preset negative sentiment threshold is set to -0.3, review data with a sentiment polarity value less than -0.3 is identified as malicious review data and removed. Next, step A1c is executed. Advertising links and contact information are identified using regular expressions, such as "https?: / / " for identifying URLs and "\d{11}" for identifying phone numbers. Review data containing the identified advertising links or contact information is judged as advertising review data and removed. Through the above steps, malicious review data and advertising review data contained in the original review data are effectively removed, ensuring the quality of the data used for subsequent taste feature extraction.

[0101] By adding a preprocessing step to the evaluation data, the raw evaluation data scraped by e-commerce platforms is first cleaned. This cleaning process removes malicious and advertising reviews from the raw data, reducing the interference of non-genuine user feedback on subsequent taste feature extraction. This allows the extracted taste features to more accurately reflect users' evaluations of the monk fruit mooncakes' taste. Therefore, adjusting production process parameters based on these more authentic taste features makes it possible to improve the taste and quality of the monk fruit mooncakes.

[0102] In some implementations, step A3 includes:

[0103] A301. Obtain the taste feature vector within the time window, and calculate the weight value of each taste feature vector within the time window according to the time decay function. The closer the time is to the current moment, the greater the weight value.

[0104] A302. Based on the taste feature vector and its weight value within the time window, a weighted average algorithm is used to calculate the dynamic average taste feature vector.

[0105] In step A301, a time decay function is introduced to determine the weight of each taste feature vector within a time window. Specifically, the time decay function is designed to take the temporal attribute of the evaluation data into account; evaluation data closer to the current moment is assigned a higher weight to its corresponding taste feature vector, while earlier evaluation data receives a relatively lower weight. For example, the time decay function can take various forms, such as linear decay or exponential decay. In this way, recent user taste preferences play a more significant role in the calculation of the dynamic average taste feature vector.

[0106] In step A302, a weighted average algorithm is employed to synthesize all taste feature vectors within the time window and combine them with the weight values ​​calculated in step A301 to obtain a dynamic average taste feature vector. This dynamic average taste feature vector can more timely and accurately reflect the user's latest taste preferences.

[0107] Specifically, the calculation method for the dynamic average taste feature vector was optimized. By introducing a time decay function and combining it with a weighted averaging algorithm, the calculation result of the dynamic average taste feature vector better reflects the real-time changes in user taste preferences. Therefore, the dynamic average taste feature vector can more timely and effectively reflect the latest trends in user taste, thus providing a more effective data foundation for subsequent adjustments to process parameters and improving the response speed and optimization effect of the monk fruit mooncake production process optimization method.

[0108] Preferably, before step A301, the following step may be included:

[0109] A303. Monitor the user review data stream of the e-commerce platform in real time, calculate the number of new reviews per unit time, and obtain the review density;

[0110] A304. Based on the evaluation density, the size of the time window is dynamically adjusted using an adaptive adjustment algorithm. When the evaluation density is higher than the preset evaluation density threshold, the time window is reduced; when the evaluation density is lower than the preset evaluation density threshold, the time window is expanded.

[0111] In step A303, a statistical period can be set, such as every minute or every hour, and then the number of new user reviews added to the e-commerce platform within this period can be counted. This number is the review density.

[0112] In step A304, the adaptive adjustment algorithm adjusts the time window size based on the evaluation density. As a preferred implementation, the adaptive adjustment algorithm can take the form of a piecewise function. For example, two evaluation density thresholds can be preset: a high threshold and a low threshold. When the evaluation density is higher than the high threshold, the time window is reduced to a smaller value, such as 1 week; when the evaluation density is lower than the low threshold, the time window is expanded to a larger value, such as 4 weeks; when the evaluation density is between the high and low thresholds, the time window remains unchanged, such as 2 weeks. The adjustment of the time window size can also be continuous, for example, using a linear or non-linear function to continuously adjust the size of the time window according to the evaluation density.

[0113] Specifically, to address the issue of inaccurate dynamic average taste feature vector caused by unstable evaluation numbers, this solution monitors the user evaluation data stream of the e-commerce platform in real time through step A303, calculates the number of new evaluations per unit time, and obtains the evaluation density. Evaluation density reflects the activity level and data volume of user evaluations. Therefore, in step A304, an adaptive adjustment algorithm dynamically adjusts the size of the time window based on the evaluation density. When the evaluation density is higher than a preset evaluation density threshold, it indicates a large number of new evaluations per unit time and active user evaluations. To more quickly reflect the latest user taste preferences, the time window is narrowed, reducing the impact of historical evaluation data on the dynamic average taste feature vector and improving the response speed to changes in current user taste preferences. When the evaluation density is lower than the preset evaluation density threshold, it indicates a small number of new evaluations per unit time and inactive user evaluations. To ensure the accuracy and representativeness of the dynamic average taste feature vector, the time window is widened, increasing the amount of data used to calculate the dynamic average taste feature vector and avoiding distortion of the dynamic average taste feature vector due to insufficient data. Through steps A303 and A304, the size of the time window can be dynamically adjusted according to the evaluation density, so that an accurate and timely dynamic average taste feature vector that reflects user taste preferences can be obtained under different evaluation quantities.

[0114] In some implementations, step A5 includes:

[0115] A501. Obtain historical adjustment data for each production process parameter, analyze the correlation between each production process parameter, and construct a process parameter coupling relationship matrix;

[0116] A502. Based on the coupling relationship matrix of process parameters, a matrix decoupling algorithm is used to decouple the proportional-integral control algorithm to obtain the decoupled proportional-integral control algorithm.

[0117] A503. A decoupled proportional-integral control algorithm is adopted to calculate the adjustment amount of the taste feature vector based on the taste deviation vector and the coupling relationship matrix of process parameters.

[0118] In step A501, historical adjustment records of various production process parameters are obtained. This can be achieved in the following ways: Data can be extracted from the production line's historical database, which records detailed information about each process adjustment, including the type of parameter adjusted, the time of adjustment, and the magnitude of the adjustment; alternatively, sensors deployed on the production line can monitor and record the adjustment data of each production process parameter in real time. These sensors must possess high precision and reliability to ensure data accuracy; or, production line operators can manually record each adjustment of the production process parameters, providing foundational data for subsequent data analysis.

[0119] This involves analyzing the correlations between various production process parameters and constructing a coupling matrix. Specifically, this includes: using the Pearson correlation coefficient method to analyze the linear correlation between historical adjustment data of different production process parameters, thereby quantifying the coupling relationship between parameters; or using the Granger causality test to analyze whether the historical adjustment of one production process parameter affects the future adjustment of another, thus revealing the causal coupling relationship between parameters; or using the mutual information method to assess the degree of information sharing between adjustment data of different production process parameters, with higher mutual information values ​​indicating stronger coupling between parameters. After the coupling matrix is ​​constructed, each element in the matrix represents the coupling strength between two corresponding production process parameters.

[0120] In step A502, the proportional-integral (PI) control algorithm is decoupled. This can be achieved using the following methods: Singular Value Decomposition (SVD) can be used, which decomposes the coupling matrix into multiple singular values ​​and vectors. The coupling strength is assessed by analyzing the magnitude of the singular values, and a decoupling controller is designed accordingly. Alternatively, eigenvalue decomposition can be used, which calculates the eigenvalues ​​and eigenvectors of the coupling matrix, decomposing the original control system into multiple independent subsystems. A PI controller is designed independently in each subsystem, thus achieving decoupling control. Finally, an iterative decoupling algorithm can be used, which continuously optimizes the decoupling matrix, gradually reducing the coupling influence between parameters until the control system ultimately reaches a decoupled state.

[0121] In step A503, the decoupled proportional-integral (PII) control algorithm is used to calculate the adjustment amount of the taste feature vector. Specifically, when calculating the adjustment amount, the decoupled PPI control algorithm considers the coupling relationship matrix of the production process parameters. For example, when calculating the adjustment amount of a certain taste feature vector, it considers not only the components of the taste deviation vector directly related to that taste feature, but also the indirect influence of other taste feature deviation components on the adjustment amount through the coupling relationship matrix. Through matrix operations, the decoupled PPI control algorithm can more accurately calculate the adjustment amount of each taste feature vector, thereby achieving more refined taste control.

[0122] Specifically, in optimizing the production process of Luo Han Guo mooncakes, this application first analyzes historical adjustment data of production process parameters through step A501 to deeply explore the potential coupling relationships between parameters and construct a process parameter coupling relationship matrix. This matrix can quantitatively describe the degree of mutual influence between parameters. Therefore, in step A502, based on the coupling relationship matrix, a matrix decoupling algorithm is used to modify the traditional proportional-integral control algorithm, resulting in a decoupled control algorithm. This improvement allows the control algorithm to fully consider the coupling effect between parameters when making parameter adjustment decisions, avoiding a decrease in control effect due to mutual constraints between parameters. Finally, in step A503, the decoupled proportional-integral control algorithm, combined with the taste deviation vector and the process parameter coupling relationship matrix, calculates the taste feature vector adjustment amount. This ensures that the calculation of the adjustment amount is no longer a simple proportional-integral operation based on taste deviation, but fully considers the parameter coupling relationship. This allows the calculated adjustment amount to more accurately reflect the actual process adjustment needs, effectively avoiding control imbalances or production oscillations caused by neglecting parameter coupling, and ensuring the stability and effectiveness of the taste adjustment process. Compared with the scheme that directly adopts the proportional-integral control algorithm, the scheme of this application has advantages in both control accuracy and stability.

[0123] Through the above technical solution, this application can solve the problem that in the process of optimizing the production process of monk fruit mooncakes, due to the coupling relationship between various production process parameters, when directly using the proportional-integral control algorithm to adjust the taste, it is easy to ignore this coupling relationship, resulting in poor control effect or even causing oscillation in the production process. It achieves more accurate and stable taste control, ensuring the continuous optimization and improvement of the taste quality of monk fruit mooncakes.

[0124] In some implementations, step A6 includes:

[0125] A601. Input the taste feature vector adjustment amount into the pre-trained mapping model to obtain the initial production process parameter adjustment amount for each production process parameter;

[0126] A602. Determine whether the adjustment amount of each initial production process parameter exceeds the corresponding preset constraint range;

[0127] A603. If there are initial production process parameter adjustments that exceed the constraint range, calculate the distance between each initial production process parameter adjustment and the corresponding constraint range boundary to obtain the out-of-bounds distance vector.

[0128] A604. Based on the out-of-bounds distance vector, the adjustment amount of each initial production process parameter is corrected to minimize the weighted norm of the taste deviation vector, so that the dynamic average taste feature vector approaches the target taste feature vector.

[0129] In step A601, a pre-trained mapping model is used to establish the mapping relationship between the adjustment amount of the taste feature vector and the adjustment amount of the production process parameters. As one implementation, this mapping model can employ a neural network model. The input of this neural network model is the adjusted taste feature vector, and the output is the initial adjustment amount of each corresponding production process parameter. This mapping model is trained in advance using historical production data and taste data to learn the correlation between changes in taste features and adjustments to production process parameters.

[0130] Steps A602 and A603 involve determining the constraint range and calculating the out-of-bounds distance. Specifically, the preset constraint range is determined based on the physical limitations of the production process and product quality requirements. For example, the constraint range for baking temperature can be set to 150℃ to 200℃. If the current baking temperature is set to 180℃, then the constraint range for adjusting the baking temperature is -30℃ to +20℃. Step A602 checks whether each initial production process parameter adjustment obtained in step A601 falls within the corresponding constraint range. If the initial adjustment exceeds the constraint range, step A603 calculates the distance between the adjustment and the nearest constraint boundary, and summarizes these distances to form an out-of-bounds distance vector for subsequent adjustment correction.

[0131] Step A604 is the core step in the adjustment correction. In this step, the initial production process parameter adjustment is corrected based on the out-of-bounds distance vector obtained in step A603. The goal of the correction is to minimize the taste deviation as much as possible while ensuring that the adjusted production process parameters do not exceed the constraint range. As a preferred implementation, step A604 can prioritize directly correcting the initial production process parameter adjustment that exceeds the constraint range to the corresponding constraint boundary value. For the initial production process parameter adjustment that does not exceed the constraint range, a quadratic programming algorithm can be used for optimization to further minimize the weighted norm of the taste deviation vector within the constraint range. The design of the weighted norm of the taste deviation vector considers the different importance of the impact of different taste characteristics on the overall taste.

[0132] Specifically, in actual production, the adjustment of production process parameters is limited by physical and technological constraints. If the adjustment amount is directly determined based on the taste feature vector without considering these constraints, the adjusted parameter values ​​may exceed the practically feasible range, making the adjustment plan unfeasible or leading to production instability and product quality degradation. This solution effectively addresses this problem by adding constraint range judgment and adjustment amount correction steps. First, the initial production process parameter adjustment amount is quickly obtained through a pre-trained mapping model, ensuring rapid adjustment. Then, the constraint range judgment step identifies infeasible adjustment amounts and quantifies the degree of exceeding the limit using an out-of-bounds distance vector. Finally, the adjustment amount correction step optimizes and corrects the initial adjustment amount while considering the constraints and the importance of taste features, yielding the final production process parameter adjustment amount. This ensures that the production process parameter adjustment amount effectively optimizes product taste without exceeding the constraints of actual production, making the user-evaluated production process optimization method more reasonable and feasible, and improving the stability of the production process and the reliability of product quality.

[0133] In some specific implementations, the adjustment of production process parameters for monk fruit mooncakes can consider three key parameters: baking temperature, baking time, and the amount of monk fruit syrup added. The preset constraint range can be set as follows: baking temperature between 160℃ and 190℃, baking time between 20 minutes and 30 minutes, and monk fruit syrup addition between 5% and 15%. Assuming the initial production process parameter adjustments obtained through the mapping model in step A601 are: baking temperature adjustment of +10℃, baking time adjustment of +5 minutes, and monk fruit syrup addition adjustment of +2%, and the current baking temperature is 185℃, baking time is 25 minutes, and monk fruit syrup addition is 10%, then the adjusted initial production process parameter values ​​are: baking temperature of 195℃, baking time of 30 minutes, and monk fruit syrup addition of 12%. After step A602, the baking temperature of 195℃ exceeded the preset upper limit of 190℃, while the baking time of 30 minutes and the amount of monk fruit syrup added (12%) were within the constraints. In step A603, the deviation distance of the baking temperature was calculated to be 5℃, and the deviation distance vector can be represented as [5℃, 0, 0]. In step A604, the baking temperature was first corrected to the constraint boundary value of 190℃. Then, for the baking time and the amount of monk fruit syrup added, a quadratic programming algorithm was used to optimize and adjust the parameters, aiming to minimize the weighted norm of the taste deviation vector. For example, if the weight of "sweetness" in the taste characteristics is high, and the taste deviation vector shows that the sweetness is low, the amount of monk fruit syrup added can be appropriately increased within the constraints to improve the sweetness of the mooncake, making the dynamic average taste characteristic vector closer to the target taste characteristic vector. Finally, the corrected production process parameter adjustment amounts are obtained and used to guide the actual production process adjustment.

[0134] In some preferred embodiments, step A604 includes:

[0135] The initial production process parameter adjustment amount that exceeds the constraint range is corrected to the corresponding constraint boundary value to obtain the corrected production process parameter adjustment amount.

[0136] For the initial production process parameter adjustment amount that does not exceed the constraint range, with the goal of minimizing the weighted norm of the taste deviation vector, a quadratic programming algorithm is adopted to adjust the initial production process parameter adjustment amount that does not exceed the constraint range according to the importance of taste features, so as to obtain the optimized production process parameter adjustment amount.

[0137] The revised and optimized production process parameter adjustments are combined to obtain the final production process parameter adjustments.

[0138] Specifically, the adjustment amount of the initial production process parameters that exceeds the constraint range is corrected to the corresponding constraint boundary value to obtain the corrected production process parameter adjustment amount. This means that for those initial production process parameter adjustment amounts that are too large or too small, they are directly adjusted to the boundary value of the preset process parameter constraint range. The constraint boundary value can be the maximum or minimum value allowed by the process parameter. Specifically, it can be implemented by using a limiter or clamping function to ensure that the adjusted parameter value does not exceed the physical or process limitations of actual production.

[0139] Specifically, for the initial production process parameter adjustments within the constraints, a quadratic programming algorithm is used to minimize the weighted norm of the taste deviation vector. Based on the importance of taste features, the initial production process parameter adjustments within the constraints are adjusted to obtain optimized production process parameter adjustments. This means that for the initial production process parameter adjustments within the constraints, further optimization calculations are performed using the minimization of the weighted norm of the taste deviation vector as the objective function. Quadratic programming is an effective method for finding the optimal value of a quadratic function under constraints. This method can find a set of optimal parameter adjustments while satisfying process parameter constraints, making the final product taste as close as possible to the target taste. The importance of taste features can be reflected by weight coefficients; the higher the weight coefficient, the greater the impact of the taste feature on the overall taste, and the more emphasis will be placed on adjusting this feature during the optimization process. Specifically, mature quadratic programming solvers in software tools such as MATLAB and Python can be used for implementation.

[0140] The process of merging the revised and optimized production process parameter adjustments to obtain the final production process parameter adjustments involves integrating the two adjustment amounts obtained in the preceding steps to create a complete set containing all production process parameter adjustments. This allows for subsequent adjustments to the production line's process parameters based on these adjustments. The merging process can involve combining the revised and optimized adjustment amounts into a vector or list to ensure that each production process parameter has a corresponding adjustment value.

[0141] Specifically, this solution is an improvement measure proposed to ensure the effectiveness and optimization of the initial production process parameter adjustments, based on the already obtained initial adjustments. First, the solution considers the constraints on process parameters in actual production. If the initially calculated adjustments exceed these ranges, direct application may lead to production instability or product quality degradation. Therefore, for adjustments exceeding the constraints, this solution adopts a "correction" strategy, forcibly pulling these adjustments back to the constraint boundaries, ensuring the rationality and feasibility of the adjustments. For adjustments within the constraints, this solution does not stop there but further seeks to maximize taste optimization. By introducing a quadratic programming algorithm and considering the importance of different taste characteristics in the optimization objective, parameter adjustments within the constraints can be more refined and better suited to the needs of product taste optimization. This differentiated approach—correcting adjustments exceeding the constraints and optimizing those within them—demonstrates the comprehensiveness and effectiveness of the solution. Finally, by merging the two sets of adjustment results, the solution obtains a production process parameter adjustment scheme that satisfies process constraints while optimizing product taste as much as possible, providing more accurate and reliable guidance for subsequent production process adjustments.

[0142] Through the above technical solution, this application can adjust the production process parameters more precisely and effectively, ensuring that the adjusted parameters meet the constraints of actual production, and optimizing the product taste to the maximum extent, so that the dynamic average taste feature vector can more accurately approximate the target taste feature vector.

[0143] refer to Figure 2 This application provides a process optimization device for monk fruit mooncakes, used to optimize the production process parameters of monk fruit mooncakes to improve their taste. The device includes:

[0144] Evaluation data collection module 1 is used to capture user review text data of monk fruit mooncakes from e-commerce platforms as evaluation data (for details, please refer to step A1 above).

[0145] The taste feature extraction module 2 is used to extract taste features related to taste from the evaluation data using natural language processing technology, and construct a multi-dimensional taste feature vector (for details, refer to step A2 above).

[0146] The dynamic average calculation module 3 is used to introduce a time window and calculate the moving average of the taste feature vector over a period of time to obtain the dynamic average taste feature vector (refer to step A3 above for the specific process).

[0147] Deviation calculation module 4 is used to calculate the deviation between the dynamic average taste feature vector and the target taste feature vector to obtain the taste deviation vector (for details, refer to step A4 above).

[0148] The feature adjustment calculation module 5 is used to calculate the adjustment amount of the taste feature vector based on the taste deviation vector using a proportional-integral control algorithm (refer to step A5 above for the specific process).

[0149] The parameter adjustment calculation module 6 is used to determine the adjustment amount of production process parameters based on the adjustment amount of the taste feature vector (for details, refer to step A6 above).

[0150] Adjust execution module 7, which is used to adjust the production process parameters of monk fruit mooncakes according to the adjustment amount of the production process parameters (refer to step A7 above for details).

[0151] In some embodiments, the monk fruit mooncake process optimization device further includes:

[0152] The sentiment polarity calculation module is used to identify evaluation data containing abnormal evaluation keywords and calculate the sentiment polarity value of the corresponding evaluation data (for details, refer to step A1a above).

[0153] The malicious evaluation removal module is used to determine that the corresponding evaluation data is malicious evaluation data when the emotional polarity value is less than the preset negative emotional threshold, and remove it (for details, refer to step A1b above).

[0154] The ad removal module is used to identify review data containing ad links or contact information, determine the corresponding review data as ad review data, and remove it (for details, refer to step A1c above).

[0155] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0156] Furthermore, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0157] Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0158] In this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, without necessarily requiring or implying any such actual relationship or order between these entities or operations.

[0159] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for optimizing the production process of monk fruit mooncakes, used to optimize the production process parameters of monk fruit mooncakes to improve their taste, characterized in that, The steps of this method include: A1. Extract user review text data of monk fruit mooncakes from e-commerce platforms as review data; A2. Using natural language processing technology, extract taste-related features from the evaluation data and construct a multi-dimensional taste feature vector; A3. Introduce a time window and calculate the moving average of the taste feature vector over a past period to obtain the dynamic average taste feature vector; A4. Calculate the deviation between the dynamic average taste feature vector and the target taste feature vector to obtain the taste deviation vector; A5. Using a proportional-integral control algorithm, the adjustment amount of the taste feature vector is calculated based on the taste deviation vector; A6. Determine the adjustment amount of the production process parameters based on the adjustment amount of the aforementioned taste feature vector; A7. Adjust the production process parameters of the monk fruit mooncake according to the aforementioned production process parameter adjustment amount.

2. The method for optimizing the processing of monk fruit mooncakes according to claim 1, characterized in that, Step A2 includes: A201. The evaluation data is preprocessed to remove interference information, resulting in preprocessed text data; the interference information includes HTML tags, special characters, and stop words; A202. Encode the preprocessed text data using a pre-trained BERT model to obtain the corresponding BERT vector; A203. Based on a taste dictionary containing multiple taste keywords, calculate the cosine similarity between the BERT vector of each preprocessed text data and the word vector of each keyword in the taste dictionary to obtain a taste similarity vector; A204. Based on the taste similarity vector, use a pre-trained taste feature extraction model to extract taste features and construct a multi-dimensional taste feature vector.

3. The method for optimizing the processing of monk fruit mooncakes according to claim 2, characterized in that, Following step A204, the following steps are also included: A205. Based on the knowledge graph of monk fruit mooncake taste, and combined with monk fruit feature information, the multidimensional taste feature vector is modified to highlight the influence of monk fruit's characteristic taste on the overall taste, thus obtaining the final taste feature vector; the knowledge graph contains the correlation between monk fruit feature information and taste features, and the monk fruit feature information includes at least one of monk fruit variety, origin, and harvesting time.

4. The method for optimizing the processing of monk fruit mooncakes according to claim 1, characterized in that, The steps following step A1 and before step A2 include: A1a. Identify evaluation data containing keywords that indicate abnormal evaluations and calculate the sentiment polarity value of the corresponding evaluation data; A1b. If the emotional polarity value is less than the preset negative emotional threshold, the corresponding evaluation data is determined to be malicious evaluation data and is removed. A1c. Identify review data containing advertising links or contact information, determine the corresponding review data as advertising review data, and remove it.

5. The method for optimizing the processing of monk fruit mooncakes according to claim 1, characterized in that, Step A3 includes: A301. Obtain the taste feature vector within the time window, and calculate the weight value of each taste feature vector within the time window according to the time decay function. The closer the time is to the current moment, the greater the weight value. A302. Based on the taste feature vector and its weight value within the time window, a weighted average algorithm is used to calculate the dynamic average taste feature vector.

6. The method for optimizing the processing of monk fruit mooncakes according to claim 5, characterized in that, Before step A301, the following steps are also included: A303. Monitor the user review data stream of the e-commerce platform in real time, calculate the number of new reviews per unit time, and obtain the review density; A304. Based on the evaluation density, the size of the time window is dynamically adjusted using an adaptive adjustment algorithm. When the evaluation density is higher than a preset evaluation density threshold, the time window is reduced; when the evaluation density is lower than the preset evaluation density threshold, the time window is expanded.

7. The method for optimizing the processing of monk fruit mooncakes according to claim 1, characterized in that, Step A5 includes: A501. Obtain historical adjustment data for each production process parameter, analyze the correlation between each production process parameter, and construct a process parameter coupling relationship matrix; A502. Based on the process parameter coupling relationship matrix, a matrix decoupling algorithm is used to decouple the proportional-integral control algorithm to obtain the decoupled proportional-integral control algorithm. A503. Using the decoupled proportional-integral control algorithm, the adjustment amount of the taste feature vector is calculated based on the taste deviation vector and the coupling relationship matrix of the process parameters.

8. The method for optimizing the processing of monk fruit mooncakes according to claim 1, characterized in that, Step A6 includes: A601. Input the adjustment amount of the taste feature vector into the pre-trained mapping model to obtain the initial adjustment amount of each production process parameter; A602. Determine whether the adjustment amount of each initial production process parameter exceeds the corresponding preset constraint range; A603. If there are initial production process parameter adjustments that exceed the constraint range, calculate the distance between each initial production process parameter adjustment and the corresponding constraint range boundary to obtain the out-of-bounds distance vector. A604. Based on the aforementioned cross-boundary distance vector, the adjustment amounts of each initial production process parameter are corrected to minimize the weighted norm of the taste deviation vector, so that the dynamic average taste feature vector approaches the target taste feature vector.

9. The method for optimizing the processing of monk fruit mooncakes according to claim 8, characterized in that, Step A604 includes: The initial production process parameter adjustment amount that exceeds the constraint range is corrected to the corresponding constraint boundary value to obtain the corrected production process parameter adjustment amount. For the initial production process parameter adjustment amount that does not exceed the constraint range, with the goal of minimizing the weighted norm of the taste deviation vector, a quadratic programming algorithm is adopted to adjust the initial production process parameter adjustment amount that does not exceed the constraint range according to the importance of taste features, so as to obtain the optimized production process parameter adjustment amount. The corrected production process parameter adjustment amount and the optimized production process parameter adjustment amount are combined to obtain the final production process parameter adjustment amount.

10. A process optimization device for monk fruit mooncakes, used to optimize the production process parameters of monk fruit mooncakes to improve their taste, characterized in that... The device includes: The evaluation data collection module is used to extract user review text data of monk fruit mooncakes from e-commerce platforms as evaluation data. The taste feature extraction module is used to extract taste-related features from the evaluation data using natural language processing technology, and to construct a multi-dimensional taste feature vector. The dynamic average calculation module is used to introduce a time window and calculate the moving average of the taste feature vector over a period of time to obtain the dynamic average taste feature vector. The deviation calculation module is used to calculate the deviation between the dynamic average taste feature vector and the target taste feature vector to obtain the taste deviation vector. The feature adjustment calculation module is used to calculate the adjustment amount of the taste feature vector based on the taste deviation vector using a proportional-integral control algorithm. The parameter adjustment calculation module is used to determine the adjustment amount of the production process parameters based on the adjustment amount of the taste feature vector. The adjustment execution module is used to adjust the production process parameters of the monk fruit mooncake according to the adjustment amount of the production process parameters.

Images