Garment package weight prediction system and method based on mlp fine-tuning

Abstract

The application provides a clothing package weight prediction system and method based on MLP fine-tuning, relates to the technical field of product data management, and comprises an MLP full-amount training module, a system offset detection and correction module, a clothing package incremental fine-tuning module and a weight prediction module. The application can realize reliable guarantee of data quality, elimination of systematic offset and overall distribution standardization based on MLP fine-tuning, improve the accuracy of initial parameters of the model, realize quick adaptation of the model to new commodities and data distribution on the basis of reservation of historical knowledge through an incremental fine-tuning mechanism, avoid overfitting, improve the robustness and generalization ability of the system, and thus ensure the stability, precision and sustainable optimization ability of overall prediction performance.

Description

Technical Field

[0001] This invention relates to the field of product data management technology, and in particular to a garment package weight prediction system and method based on MLP fine-tuning. Background Technology

[0002] With the rapid development of e-commerce and logistics networks, the types of clothing parcels are becoming increasingly diverse, and packaging methods are becoming more varied. Traditional static weight estimation methods are no longer sufficient to meet the needs of accurate prediction and efficient management. To meet the urgent needs of modern logistics for intelligent and refined management, establishing models for predicting parcel weight or reasonable weight ranges to provide data support for parcel detection, anomaly monitoring, and intelligent weighing in logistics operations has gradually become an industry trend. Existing prediction technologies are linear regression models, which have the following shortcomings in data processing for clothing weight prediction scenarios: for example, using the IQR quantile method to remove outliers is not applicable to the multimodal, long-tailed distribution of clothing weights and is prone to accidental deletion of positive values. The model has several drawbacks. The Z-score standardization assumes a Gaussian distribution, but clothing weights actually exhibit a mixed distribution with significant differences between categories, leading to information loss after standardization. The model architecture also has the following shortcomings: the least squares method suffers from parameter estimation issues with large-scale features and samples; calculating the inverse covariance matrix is ​​computationally too complex when the sample size and number of features are extremely large, exceeding the load capacity of commercial servers; matrix inversion operations are numerically unstable in high dimensions, resulting in inaccurate regression coefficient estimations; furthermore, a full model update requires retraining with all historical data, each iteration is time-consuming, and it cannot quickly adapt to the launch of new clothing products, failing to meet the business's need for efficient prediction, and incurring high costs.

[0003] For example, Chinese invention patent CN120875708A discloses a method and electronic device for identifying anomalies in logistics package weight information, which includes: establishing a logistics package weight database based on multiple positive logistics data points at the SKU level (smallest stock unit) dimension for clothing products; training a clothing product weight prediction model using data from the database; receiving weight information of returned packages uploaded by the logistics service provider during the return pickup service process, determining the associated return order information, and the SKU information of the returned clothing products associated with the return order; predicting the reasonable weight range of returned clothing product SKUs in logistics packages using the clothing product weight prediction model; and determining anomalies by judging whether the weight information of the returned packages returned by the logistics service provider is within the reasonable weight range.

[0004] For example, the Chinese invention patent with announcement number CN114330829B discloses a method, system, device, and storage medium for predicting the weight of logistics orders. This includes: obtaining information on the type of goods and the estimated number of pieces with non-zero values ​​in the current logistics order; calculating the root mean square error of the estimated number of pieces, the standard deviation of the weight per piece, and the mean weight per piece for historical logistics orders of the same goods type; combining the root mean square error of the estimated number of pieces and the standard deviation of the weight per piece, determining whether the total weight error of the current order is not greater than a preset weight threshold; if so, the predicted weight of the current order is the product of the estimated number of pieces and the mean weight per piece.

[0005] The method for predicting the weight of clothing parcels primarily involves constructing a logistics parcel weight database based on the SKU dimension. A regression prediction model is trained using historical positive logistics data to predict reasonable weight ranges for goods or parcels. Anomaly detection is performed by combining actual returned weight data, thereby improving the accuracy of parcel weight prediction and anomaly identification capabilities. Simultaneously, weighted or mean-product calculations are performed based on historical statistical features. By acquiring information on each type of goods and its non-zero estimated number of pieces in the current order, and combining this with statistical features of the same type of goods in historical logistics orders—including the root mean square error of the estimated number of pieces, the standard deviation of the weight per piece, and the mean—the overall weight error of the current order is determined. When the overall order prediction error does not exceed a preset threshold, the predicted weight of the current order is calculated as the product of the estimated number of pieces and the mean weight per piece, thus estimating the order weight.

[0006] The above-mentioned technology has at least the following technical problems: In the task of predicting the weight of clothing parcels, new products, packaging methods, seasonality, and category combinations are constantly emerging in the logistics system. The original model only learned the old distribution of product weight parameters based on historical data. When faced with new products or quantity patterns that have not been seen before, the model is prone to prediction bias. If the fine-tuning strategy is not reasonable, it will also cause problems: if the amount of new data is small but the number of training rounds is too large or the learning rate is too high, the model will overfit these small amount of new data; if the initial weights of existing products do not use calibrated historical parameters, the forgetting effect may occur during the fine-tuning process, which will reduce the prediction accuracy of historical products. Summary of the Invention

[0007] In view of this, embodiments of the present invention provide a clothing package weight prediction system and method based on MLP fine-tuning, which can quickly adapt to new products and data distribution while retaining historical knowledge, so that the weight characteristics of clothing packages maintain their original business meaning, the data quality is reliably guaranteed, and the accuracy of the MLP model is improved.

[0008] The technical solution of this invention is implemented as follows: This invention provides a garment package weight prediction system based on MLP fine-tuning, the system comprising: The MLP full training module is used to extract clothing package data from the historical logistics database, perform outlier handling, initialize the weight of clothing goods after outlier handling, and then perform full training of the MLP model.

[0009] The system offset detection and correction module is used to detect the offset of historical clothing wrapping data at a preset offset monitoring period after the MLP model is fully trained, and to correct the data when the detection result is a systemic offset.

[0010] The clothing package incremental fine-tuning module is used to determine the trigger conditions for incremental fine-tuning after data correction, and to determine the initial weight parameter selection strategy when incremental fine-tuning is triggered. It determines the initial weight parameters of the clothing product, sets the training strategy, and then executes incremental fine-tuning.

[0011] The weight prediction module is used to take the sequence and quantity of clothing items in the clothing package as input to the MLP model and output the predicted weight of the clothing package data.

[0012] This application also provides a method for predicting the weight of clothing packages based on MLP fine-tuning. This method is applied to a clothing package weight prediction system based on MLP fine-tuning, and includes: Clothing package data is extracted from historical logistics databases, outlier handling is performed, and clothing weight is initialized after outlier handling, thereby enabling full training of the MLP model.

[0013] After the MLP model is fully trained, historical clothing wrapping data offset detection is performed at a preset offset monitoring period, and data correction is performed when the detection result shows a systematic offset.

[0014] After data correction, the incremental fine-tuning trigger condition is determined, and the initial weight parameter selection strategy is determined when incremental fine-tuning is triggered. The initial weight parameters of the clothing product are determined, the training strategy is set, and incremental fine-tuning is performed accordingly.

[0015] The MLP model takes the clothing item sequence and quantity in the clothing package as input and outputs the predicted weight of the clothing package data.

[0016] The beneficial effects of the technical solutions provided by the embodiments of the present invention include at least the following: 1. This invention provides a clothing package weight prediction system and method based on MLP fine-tuning. It achieves systematic cleaning, missing value handling, and multi-level outlier filtering of historical clothing package data, ensuring reliable data quality. By calibrating the mean and distribution width by grouping by product, it eliminates systematic bias and standardizes the overall distribution, improving the accuracy of the model's initial parameters. Through forward propagation and gradient optimization based on MLP, it achieves high-precision prediction of package weight and supports weighted modeling of complex product combinations. Through an incremental fine-tuning mechanism, dynamically adjusting the training rounds and learning rate based on the amount of new data, and employing warm-up, decay, and early stopping strategies, the model can quickly adapt to new products and data distribution while retaining historical knowledge, avoiding overfitting and improving the system's robustness and generalization ability. This ensures the stability, accuracy, and sustainable optimization capability of the overall prediction performance.

[0017] 2. This invention effectively eliminates systematic bias and improves data consistency and comparability by calculating the mean and standard deviation by grouping products and performing scaling operations on each group of data. This process not only calibrates the mean but also adjusts the distribution width, so that the overall fluctuation range of each group is consistent with the historical reference, avoiding model deviation caused by single group anomalies or historical changes, and improving the adaptability and generalization ability of the overall system.

[0018] 3. This invention achieves adaptive control of model updates by dynamically monitoring the number of new packages within each preset period and triggering incremental fine-tuning based on the cumulative number of times and a threshold: when the amount of new data is large enough, fine-tuning is triggered in a timely manner, enabling the model to quickly learn the new distribution; when the amount of new data is insufficient, the cumulative mechanism avoids frequent fine-tuning, reducing training overhead and system fluctuations; at the same time, it takes into account both data accumulation and periodic scheduling, ensuring that the model can maintain timely updates when new products are added and data distribution changes, while also improving training efficiency and system stability, thereby achieving efficient, reliable, and intelligent management of incremental fine-tuning.

[0019] 4. This invention dynamically adjusts the training rounds and initial learning rate based on the cumulative number of newly added packages, enabling the model to adaptively learn from new data of different scales. This ensures that the model fully learns the new distribution when the data volume is large, and avoids overfitting when the data volume is small. The warm-up strategy smoothly increases the learning rate from zero to the initial value, ensuring stable weight updates and reducing unstable fluctuations in the early stages of training. Only the model parameters with the best performance on the validation set are saved, allowing the system to quickly adapt to new products and new data distributions while retaining historical knowledge. This enhances the model's continuous adaptability and robustness, ensuring that the model can update stably and efficiently when processing dynamic logistics data, while maintaining prediction accuracy and system reliability. This provides a solid foundation for long-term operation and continuous optimization. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the structure of the clothing package weight prediction system based on MLP fine-tuning provided in an embodiment of the present invention; Figure 2 This is a flowchart of the clothing package weight prediction method based on MLP fine-tuning provided in the embodiments of the present invention; Figure 3 This is a strategy diagram of the clothing package weight prediction method based on MLP fine-tuning provided in the embodiments of the present invention; Figure 4 This is a diagram illustrating the incremental fine-tuning strategy of the clothing package weight prediction method based on MLP fine-tuning provided in this embodiment of the invention. Figure 5 This is a flowchart of the clothing package weight prediction method based on MLP fine-tuning provided in the embodiments of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0023] like Figure 1 The schematic diagram of the garment package weight prediction system based on MLP fine-tuning shown includes: an MLP full training module, a system offset detection and correction module, a garment package incremental fine-tuning module, and a weight prediction module.

[0024] The MLP full training module is used to extract clothing package data from the historical logistics database, handle outliers, initialize the weight of clothing goods after outlier handling, and then conduct full training of the MLP model.

[0025] like Figure 3 This invention provides a full training strategy diagram for a clothing package weight prediction method based on MLP fine-tuning. The method includes obtaining package data from a historical logistics database, cleaning, handling missing values, and removing outliers to make a preliminary estimate of the product weight; then predicting the package weight using an MLP model. Based on the forward propagation calculation of the predicted value, the product weight parameters are continuously optimized using error feedback and gradient updates to achieve accurate modeling of the product weight.

[0026] Furthermore, outlier handling is performed, and the specific analysis methods are as follows: Data on clothing parcels was extracted from historical logistics databases and then cleaned.

[0027] It should be noted that clothing package data is extracted from historical logistics databases, including product SKUs (Stock Keeping Units, codes used to uniquely identify specific products in logistics databases), quantities, categories, packaging methods, seasonal information, and actual package weights. The data is then cleaned, specifically by checking and removing duplicate records to ensure the uniqueness of each package entry. Missing values ​​are filled using methods based on field type. The units and formats of all fields are standardized, with weights uniformly converted to grams and dates to standard date formats. Finally, the classification fields are standardized, unifying category names and seasonal labels. This ensures consistency and usability of the data in subsequent analysis and model training, resulting in complete, standardized, non-redundant clothing package data suitable for modeling.

[0028] The clothing package data is grouped according to the category and season of the clothing products to obtain each true weight outlier detection group. For the package weight data of each true weight outlier detection group, the first quartile, the third quartile, and the interquartile range are calculated respectively. The lower limit of the outlier is obtained based on the first quartile and the interquartile range, and the upper limit of the outlier is obtained based on the third quartile and the interquartile range.

[0029] It should be noted that historical package data is grouped according to the category of clothing goods, and each group is attached with a seasonal tag to form multiple groups for detecting outliers in actual weight. Data within each group is considered to have a similar weight distribution. The lower limit of the outlier is calculated by subtracting a multiple of the interquartile range from the first quartile and multiplying it by the interquartile range. This is used to identify outliers that are too light. The upper limit of the outlier is calculated by adding a multiple of the interquartile range to the third quartile and multiplying it by the interquartile range. This is used to identify outliers that are too heavy. This achieves effective identification and filtering of package weight outliers within each group.

[0030] In this embodiment, the interquartile range multiple is typically set based on the statistical characteristics and empirical rules of historical parcel data to balance the accuracy and stability of outlier detection. The interquartile range multiple is used to expand the first and third quartiles of each group into the lower and upper limits of outliers, thereby determining whether the clothing parcel data is abnormal. Specifically, for groups with relatively concentrated distribution, the multiple can be set to a smaller value to strictly identify outliers, while for groups with large distribution fluctuations, the multiple can be set to a larger value to avoid misjudgment. This ensures that the threshold for outlier detection conforms to statistical laws and adapts to actual business scenarios under different product categories and seasonal conditions.

[0031] For a given outlier value in real weight, a detection group is established. Real weight data that is greater than the upper limit of the outlier value or less than the lower limit of the outlier value is marked as an outlier value, while real weight data that is less than or equal to the upper limit of the outlier value and greater than or equal to the lower limit of the outlier value is marked as a normal value.

[0032] It should be noted that if the actual weight of a package within a detection group is greater than the upper limit of the outlier value or less than the lower limit of the outlier value, it means that the weight of the package deviates significantly from the normal distribution range of that group, and there may be a weighing error or data anomaly. Therefore, it is marked as an outlier. On the other hand, if the actual weight of a package within a detection group is less than or equal to the upper limit of the outlier value and greater than or equal to the lower limit of the outlier value, it means that the weight of the package is within the reasonable range of that group and conforms to statistical regularity. Therefore, it is marked as a normal value.

[0033] The outlier value of the actual weight is removed from the detection group, and the initial weight signal of the clothing product is generated after the removal.

[0034] It should be noted that weight records that deviate significantly from the normal range are removed from the historical logistics database and are not included in subsequent statistics and model training. After removing outliers, the initial weight signal for clothing products is generated.

[0035] If the actual weight outlier exists, but the detection group does not have an outlier, then the garment weight initialization signal is directly generated.

[0036] It should be noted that if the actual weight outlier exists, the absence of outliers in the detection group indicates that the weight of all packages in that group is within a reasonable range, and no data needs to be removed. The initial weight signal for clothing products can be generated directly.

[0037] Furthermore, the weight of the clothing items is initialized, and the specific analysis method is as follows: Upon receiving the garment weight initialization signal, the garment weight is initialized.

[0038] For individual garment items, the system checks the historical logistics database for individual garment items.

[0039] If a single garment item exists in the historical logistics database, the historical single garment items are statistically analyzed, the average weight of the single garment item is calculated, and this average weight is used as the initial weight of the garment item.

[0040] It should be noted that if a single garment item exists in the historical logistics database, it means that the actual weight data of the single item in past shipments can be obtained, thus providing a reliable basis for the initial weight of the garment item. First, all single package records of the garment item are filtered out from the database, and statistical analysis is performed on the historical single garment items of the garment item to calculate the arithmetic mean of the weight of the historical single garment items. The average weight of the historical single garment items is used as the initial weight parameter of the single garment item.

[0041] If no single garment item exists in the historical logistics database, the average weight of all garment items in the same category will be used as the initial weight of the garment item.

[0042] It should be noted that the absence of a single garment item in the historical logistics database means that historical single-item weight data for this item cannot be directly obtained. Therefore, it is impossible to provide an initial weight based on its own history. Instead, the average single-item weight of all existing garment items in the same category is used as the initial weight parameter for this item to provide a reasonable initial reference value. Even without direct historical data for this item, usable initial weight parameters can still be provided for model training, thus ensuring that the package weight prediction system can quickly adapt to and maintain the stability and accuracy of predictions when faced with new items.

[0043] Furthermore, we perform full training of the MLP model. The specific analysis method is as follows: The sequence and quantity of clothing items in each package are used as input, and the package weight is calculated according to the model's prediction formula for forward propagation.

[0044] It should be noted that the forward propagation process involves taking the sequence and quantity of clothing items in each package as input, and constructing a prediction formula based on the current weight parameters of the clothing items to obtain the predicted weight of the package. This process maps the input sequence and quantity of items to the output of the predicted package weight, thus providing basic data for subsequent error calculation and gradient update.

[0045] The mean squared error is used as the loss function for error calculation.

[0046] It should be noted that the predicted weight of each package is compared with its actual weight, and the mean squared error is used as the loss function for calculation. Specifically, the difference between the predicted value and the actual value for each sample is squared. y u Let y represent the actual weight of the u-th package. u 0 Let W represent the predicted weight of the u-th package, N represent the number of product types in the package, and i represent the number of products in the package, i = 1, 2, 3...N.i Let X represent the weight parameter of the i-th item. i This represents the number of the i-th item in the package.

[0047] The gradient of the weight parameter of each garment is obtained by taking the derivative of the loss function, and the gradient is then used to solve the problem.

[0048] It's important to note that the loss function measures the error between the model's predicted package weight and the actual weight. Taking the derivative of the loss function involves calculating the partial derivative of each garment's weight parameter with respect to the loss function, which is the gradient. This gradient reflects the direction and magnitude by which the loss function will increase or decrease if the weight parameter of that garment is changed. By solving for the gradient, we can determine in which direction the weight parameter of each garment should be adjusted to reduce the overall error. Subsequently, the gradient descent algorithm is used to update these parameters, thereby gradually optimizing the model and making the package weight prediction more accurate.

[0049] The weight parameters of the clothing items are updated using the gradient descent algorithm to obtain new weight parameters for the clothing items.

[0050] It should be noted that, according to the gradient descent algorithm, the weight update value is obtained by multiplying the gradient of each product by the learning rate. The new clothing product weight parameter is obtained by subtracting the weight update value from the product weight parameter.

[0051] It's worth noting that gradient descent is an iterative method for optimizing model parameters. Its core idea is to update the parameters along the negative gradient of the loss function in the parameter space to gradually reduce prediction error. In each iteration, the partial derivatives (i.e., gradients) of the loss function with respect to each parameter under the current model parameters are first calculated. Then, the gradients are applied proportionally to the parameter updates according to the set learning rate, causing the parameters to move along the direction of the fastest descent of the loss function. By repeating this process, the model parameters gradually converge to optimal or near-optimal values ​​that minimize the overall loss, thereby improving the model's prediction accuracy and stability.

[0052] The updated clothing weight parameters will generate new package weight predictions in the next iteration's forward propagation.

[0053] It should be noted that the new weight parameter is multiplied by the quantity of the corresponding goods in each package and summed to generate a new predicted weight for that package. The MLP model uses the latest parameters to recalculate the predicted value, so that each iteration is based on more accurate goods weight information, thereby gradually reducing the prediction error and optimizing the accuracy of the overall package weight prediction.

[0054] In each iteration, the forward propagation, error calculation, gradient solution, and weight update processes are repeated continuously, so that the weight parameters of the clothing products gradually converge to the optimal solution of the overall prediction error.

[0055] It should be noted that in each iteration, the model uses the current clothing weight parameters to calculate the predicted weight of the package. The predicted weight calculated by the model is subtracted from the actual weight, and the difference is squared to obtain the prediction error. The parameters are adjusted according to the gradient. As the number of iterations increases, the weight parameters are continuously optimized, and the gap between the prediction result and the actual weight gradually narrows. Finally, the clothing weight parameters converge to the optimal solution that minimizes the overall prediction error, thus achieving high accuracy and stability in package weight prediction.

[0056] Furthermore, the package weight is calculated based on the model's prediction formula. The specific analysis method is as follows: The input consists of a list of clothing items, and each sample contains a sequence of clothing item IDs and their corresponding quantities.

[0057] It should be noted that each package is represented as a sample input, and each sample includes two core pieces of information: a sequence of clothing item IDs and a corresponding quantity sequence. The item ID sequence is used to uniquely identify each item in the package, and the quantity sequence records the quantity of each item in the package. By using these two sequences as input, the model can encode the composition and quantity information of each package, providing basic data for subsequent forward propagation to calculate the predicted weight of the package.

[0058] The initial weight of each garment item is obtained based on the garment item weight initialization, and is recorded as the garment item weight parameter.

[0059] The model prediction formula is defined as follows: Multiply the weight parameter of each clothing item by its corresponding quantity, sum the results over all clothing items, forming a linear weighted superposition process, thus outputting the predicted weight of the corresponding clothing package. It should be noted that by multiplying the weight parameter of each item in the package by its corresponding quantity, the contribution of each item to the total weight of the package is obtained. Then, the contributions of all items are summed to form a linear weighted superposition, and the predicted weight of the package is finally output. This not only reflects the composition of the items in the package, but also provides a basis for subsequent error calculation and model parameter optimization.

[0060] The system offset detection and correction module is used to detect the offset of historical clothing wrapping data at a preset offset monitoring period after the MLP model is fully trained, and to correct the data when the detection result is a systemic offset.

[0061] Furthermore, historical clothing package data offset detection was performed, and data correction was conducted when the detection results showed a systematic offset. The specific analysis method is as follows: Historical clothing package data is grouped and statistically analyzed by clothing item to obtain each drift presence detection group. The mean weight and standard deviation of clothing items in each drift presence detection group are calculated, and the baseline mean weight and standard deviation of clothing items in each drift presence detection group are extracted from the database.

[0062] It should be noted that grouping statistics by clothing product is based on the clothing product itself. All package data is categorized according to the product ID of each clothing product. In other words, all package records of the same product are assigned to the same group, so the data in each group corresponds to the product information of the same product.

[0063] After obtaining the drift detection groups, the weight of the packages in each group is statistically analyzed, and the average weight and standard deviation of the product in historical packages are calculated. These are then compared with the preset baseline average weight and standard deviation of the product in the database to detect whether there is a systematic shift or drift in the product, thus providing a basis for subsequent calibration and model training.

[0064] In this embodiment, the preset logic for the baseline mean weight and baseline standard deviation of clothing items within each drift detection group in the database is based on standard reference values ​​obtained from statistical analysis of a large amount of historical package data. These values ​​are used to measure the typical weight level and fluctuation range of each item under normal circumstances. Specifically, the logic is as follows: statistical analysis is performed on the weight records of each clothing item in historical packages, its average value is calculated as the baseline mean weight, and its standard deviation is calculated as the baseline standard deviation. These values ​​are then stored in the database. In subsequent drift detection, the weight of new packages will be compared with these preset benchmarks to identify whether there are systematic shifts or abnormal fluctuations, thereby providing a reliable basis for data calibration and model updates.

[0065] Extract the preset drift threshold from the database.

[0066] In this embodiment, the preset logic of the drift threshold in the database is based on historical statistical characteristics and business tolerance. It is used to determine whether the weight distribution of clothing products in a certain drift detection group has undergone a systematic shift. Referring to the baseline weight mean and baseline weight standard deviation of the group, upper and lower drift thresholds are set according to the maximum allowable deviation range of the business, which are set as the baseline mean plus or minus a certain multiple of the baseline standard deviation. When the mean of a new data group exceeds this threshold, it is determined that the group has a systematic drift. Through this logic, the drift threshold can reflect the normal fluctuation range of the historical weight of the products, and also take into account the sensitivity of actual business to abnormal changes, providing a basis for judgment for data calibration and model fine-tuning.

[0067] The deviations of the mean weight and the standard deviation of the weight are obtained based on the mean weight and the standard deviation of the weight of the clothing items in each drift presence detection group and the mean weight and the standard deviation of the standard weight of each drift presence detection group.

[0068] It should be noted that the mean weight and standard deviation of each drift detection group are obtained by subtracting the mean weight and standard deviation of each drift detection group from the mean weight and standard deviation of each drift detection group.

[0069] If the mean weight deviation of a group is greater than or equal to the drift threshold, then the group is determined to have a systematic bias.

[0070] It should be noted that if the mean weight deviation of a certain group is greater than or equal to the drift threshold, it means that the overall weight distribution of clothing products in that group has changed significantly compared with the historical benchmark and has exceeded the normal fluctuation range. Based on this, it is determined that there is a systematic bias in the group, indicating that the weight data of the group needs to be calibrated to ensure the accuracy of subsequent package weight prediction and model stability.

[0071] The group weight sequence is obtained by taking the weight data of each data point and the mean weight deviation within the group, and the mean weight of the group is obtained by taking the mean of the group weight sequence.

[0072] It should be noted that by subtracting the mean deviation of the group from the weight data of each package within the group, a group weight sequence is obtained, which is the new value of each data point relative to the deviation correction. Statistical processing is performed on this group weight sequence to calculate its arithmetic mean, thereby obtaining the group's baseline mean. This baseline mean represents the central level of the package weight in the group after eliminating the mean deviation, providing a reference for subsequent distribution width scaling and overall data calibration, so that the calibrated data is aligned with the historical baseline while maintaining the relative fluctuation characteristics within the group.

[0073] The scaling factor for the distribution width is obtained based on the weight standard deviation.

[0074] It should be noted that the scaling factor for the distribution width is obtained by dividing the standard deviation of the baseline weight by the standard deviation of the weight of the drift-existing detection group.

[0075] Scaling is performed on each weight data point within a group based on the scaling factor and the mean of the group baseline to obtain the weight data for each drift presence detection group. After processing, the updated weight data for each group is obtained.

[0076] It should be noted that each weight data point within a group is first subtracted from the group's baseline mean, then multiplied by a scaling factor, and finally added back to the group's baseline mean. This ensures that while aligning the mean, the overall range and fluctuation of the weight distribution within the group conforms to the historical baseline, thus achieving data distribution calibration.

[0077] After calibration, the grouped updated weight data is used as the distribution standard for incremental fine-tuning.

[0078] It should be noted that after drift calibration, the weight data of each group has undergone mean deviation correction and distribution width scaling to ensure that its mean and fluctuation range are consistent with the historical benchmark. Using these calibrated group weight data as the distribution standard for incremental fine-tuning means that when the model is fine-tuned in the future, the new package data will be learned and updated with reference to this calibration distribution. This ensures that the model retains historical knowledge and can adapt to the changes in new data during the fine-tuning process, thereby achieving the overall stability and accuracy of package weight prediction.

[0079] If the mean weight deviation of a certain group is less than the drift threshold, it is determined that there is no systematic deviation in that group, and the weight data of that group is directly used as the distribution standard for incremental fine-tuning.

[0080] It should be noted that if the mean weight deviation of a certain group is less than the drift threshold, it means that the weight distribution of the group has not changed much from the historical benchmark and is still within the normal fluctuation range. There has been no significant systematic drift. It is determined that there is no systematic deviation in the group. There is no need to perform mean or distribution width calibration. The original weight data of the group can be directly used as the distribution standard for incremental fine-tuning and used for subsequent model fine-tuning, so that the model can smoothly adapt to the new data while maintaining historical knowledge.

[0081] The clothing package incremental fine-tuning module is used to determine the trigger conditions for incremental fine-tuning after data correction, and to determine the initial weight parameter selection strategy when incremental fine-tuning is triggered. It determines the initial weight parameters of the clothing product, sets the training strategy, and then executes incremental fine-tuning.

[0082] like Figure 4 The diagram shows the incremental fine-tuning strategy of the clothing package weight prediction method based on MLP fine-tuning provided in this embodiment of the invention. It includes: continuously detecting newly added clothing package data, determining the triggering condition based on the cumulative number of newly added packages and the cumulative number of times, determining the initial weight parameter selection strategy, determining the initial weight parameters of the clothing product, and dynamically configuring a complete training strategy including the learning rate, training rounds, decay ratio, and early stopping mechanism. Finally, the incremental fine-tuning process is started to achieve efficient model iteration and performance improvement while retaining historical knowledge.

[0083] Furthermore, the incremental fine-tuning trigger condition is determined, and the specific analysis method is as follows: Continuously monitor newly added clothing package data. When newly added clothing package data is detected, extract, clean and filter outliers from the data, and construct a triplet input sample for each package.

[0084] It should be noted that we continuously monitor newly added clothing parcel data. Once a new parcel record is detected, we extract the necessary field information, such as product ID, quantity, and actual weight. We then perform data cleaning, including deduplication, handling missing values, and unit standardization, to ensure data integrity and consistency. Next, we filter out outliers, removing weight data that deviates significantly from the normal range. After processing, each parcel is organized into a triplet input sample, including the product ID sequence, the corresponding quantity sequence, and the actual weight, providing standardized model input for subsequent incremental fine-tuning training.

[0085] Extract the preset fine-tuning period, package quantity threshold, and cumulative number threshold from the database.

[0086] It should be noted that the settings for the fine-tuning period, package quantity threshold, and cumulative number threshold are usually determined based on the changing patterns of historical package data and business needs. The fine-tuning period can be set according to the frequency of new package additions and business response requirements. A period that is too short will lead to frequent fine-tuning and increased computational overhead, while a period that is too long may delay model updates. The package quantity threshold is usually set with reference to the statistical distribution of new packages in each historical period, and is set to a quantity level that can significantly reflect changes in data distribution. The cumulative number threshold is used to take into account situations where there is insufficient data within a period, but continuous accumulation may lead to significant changes. It is usually set based on historical cumulative data fluctuations and the model's tolerance for data delays. Through these preset logics, it is possible to ensure timely model updates while avoiding excessive fine-tuning, thereby improving the efficiency and stability of incremental learning.

[0087] The cumulative number of new parcels is counted within each fine-tuning cycle.

[0088] When the cumulative number of newly added packages reaches or exceeds the package quantity threshold, an incremental fine-tuning is triggered.

[0089] It should be noted that when the cumulative number of newly added packages reaches or exceeds the package number threshold, it means that the amount of new data added in the current fine-tuning cycle is large enough to significantly reflect the changes in package weight distribution, and it is necessary to update the model. Therefore, an incremental fine-tuning is immediately triggered to improve the prediction accuracy.

[0090] If the cumulative number of newly added packages does not reach the package quantity threshold, the cumulative number of times in the fine-tuning cycle is checked. If the cumulative number is greater than or equal to the cumulative number threshold, incremental fine-tuning is triggered directly. If the cumulative number is less than the cumulative number threshold, the cumulative number of newly added packages in the current cycle is carried over to the next cycle, and the cumulative number is incremented by one.

[0091] It should be noted that if the cumulative number of newly added packages does not reach the package number threshold, the cumulative count is checked. If the cumulative count is greater than or equal to the cumulative count threshold, it means that even if the amount of new data in a single period is insufficient, the accumulated data over multiple periods has reached the point where fine-tuning is necessary. Therefore, the system will directly perform an incremental fine-tuning. If the cumulative count is still less than the cumulative count threshold, the system will not fine-tune immediately. Instead, it will accumulate the cumulative number of newly added packages in the current period to the next period, while incrementing the cumulative count by one, waiting for further data accumulation in subsequent periods to ensure that fine-tuning is performed only when the amount of data is sufficient, thus balancing the timeliness and stability of model updates.

[0092] Furthermore, the initial weight parameter selection strategy is determined, and the initial weight parameters of the clothing products are determined. The specific analysis method is as follows: For newly added clothing package data, extract clothing product information. If the clothing product information is found in the existing clothing product information in the historical logistics database, the initial weight parameter selection strategy is set as the first priority acquisition strategy.

[0093] It should be noted that for newly added clothing package data, the clothing product information is extracted first. If the historical data of the clothing product can be found in the existing clothing product information in the historical logistics database, that is, the product has data in previous packages, the historical weight data of the product can be used directly to extract the initial weight parameters. In this case, the selection strategy of the initial weight parameters is marked as the first priority acquisition strategy, which means that the system prioritizes using the product's own historical data to initialize the weight parameters, thereby ensuring that the initial value is closest to the real weight and providing a reliable foundation for subsequent fine-tuning training.

[0094] Extract the preset sample quantity threshold from the database.

[0095] It should be noted that the preset logic of the sample size threshold is to determine whether the amount of historical package data for a single garment item is reliable enough, and to determine the calculation method for the initial weight parameter. This threshold is usually determined based on the statistical distribution of historical package data and the business requirements for prediction accuracy. The specific setting method usually combines the mean or quantile of the historical package quantity distribution, as well as the model's tolerance for the stability of small sample data, so as to flexibly deal with differences in data volume while ensuring the reliability of the weight parameter.

[0096] The weight of the garment is calculated based on historical single-item garment package data, and all weight records of the garment in historical single-item packages are extracted for outlier filtering.

[0097] When the number of samples is less than or equal to the sample number threshold, outlier removal is performed using the seasonal quantile boundary of the product category to obtain the weight data sequence of the clothing product.

[0098] It should be noted that when the sample size is less than or equal to the sample size threshold, it means that the amount of historical data for the product is small and insufficient to accurately estimate the weight parameters by relying entirely on its own data. In order to ensure the robustness of outlier removal, the quantile boundaries of the product category and season are used as a reference to filter outliers in the weight data of its historical single-item packages, thereby removing abnormal records that deviate significantly from the normal range. After filtering, the remaining weight data forms the weight data sequence of the clothing product, which is used to calculate the initial weight parameters of the product, ensuring that a reasonable and stable weight estimate can be obtained even if the sample size is insufficient.

[0099] It should be explained that, in a specific embodiment, the quantile boundaries of the product category and season are Q1−3×IQR and Q3+6×IQR, where Q1 is the first quartile, indicating that 25% of the data are below this value, Q3 is the third quartile, indicating that 75% of the data are below this value, and IQR is the interquartile range, equal to Q3−Q1, reflecting the middle 50% distribution range of the data. By multiplying IQR by 3 and adding it to Q3, or by subtracting IQR multiplied by 6 from Q1, the upper and lower outlier boundaries can be obtained, which are used to identify data with extreme deviations. Using this method, when the sample size of a single product is insufficient, outliers can be reasonably removed from historical weight data by utilizing the overall distribution characteristics of the product category and season, thereby ensuring the reliability of the remaining data for initial weight calculation.

[0100] When the number of samples exceeds the sample size threshold, multi-layer anomaly filtering is performed by comprehensively using the interquartile range of clothing weight, category boundaries, and the 3σ principle to obtain the weight data sequence of the clothing product.

[0101] It should be noted that when the sample size exceeds the sample size threshold, it means that the data for the product itself is rich enough to support more refined outlier detection. Three methods are used for multi-layered outlier filtering: first, outlier removal is based on the interquartile range (IQR) of the product's own weight; second, auxiliary filtering is performed by referring to the quantile boundaries of the category; and third, the 3σ principle is adopted to remove extreme values ​​exceeding three times the standard deviation based on the sample mean and standard deviation. Through this multi-layered filtering, abnormal fluctuations can be removed while maintaining data integrity, resulting in a reliable weight data sequence for the clothing product, providing a high-quality data foundation for the initial weight parameter calculation.

[0102] The arithmetic mean of the weight data series of the clothing item is used as the initial weight parameter of the clothing item.

[0103] It should be noted that the average weight data series of the clothing item is used as the initial weight parameter for the forward propagation and prediction calculation of the model. In this way, the initial weight parameter can comprehensively reflect the typical weight level of the item in historical packages, while eliminating the influence of outliers and improving the accuracy and stability of the prediction.

[0104] If the clothing product information is not found in the existing clothing product information in the historical logistics database, check whether there are other clothing products of the same category and season as the newly added clothing product in the historical logistics database.

[0105] It should be noted that if the clothing product information cannot be found in the existing clothing product information in the historical logistics database, it means that the product is a brand new addition and there is no historical package data for reference. In this case, we further check whether there are other clothing products of the same category and season as the new product in the database, and extract the historical weight data of these products as a reference. In this way, even if the new product lacks its own data, the typical weight characteristics of similar products in the same season can be used to provide a reasonable estimate of the initial weight parameters of the product, ensuring that the model can quickly adapt to the prediction needs of the new product.

[0106] If there are clothing items of the same category and season as the newly added clothing item in the historical logistics database, they are recorded as similar clothing items, the initial weight parameter selection strategy is recorded as the second priority acquisition strategy, and the average weight of similar clothing items is taken as the initial weight parameter of the clothing item.

[0107] It should be noted that if there are clothing items of the same category and season as the newly added clothing item in the historical logistics database, these items are recorded as similar clothing items. The initial weight parameter selection strategy is marked as the second priority acquisition strategy. This means that the new item itself lacks historical data, but the weight can be estimated by using the statistical characteristics of similar items. Specifically, the arithmetic mean of the historical weight data of these similar clothing items is calculated, and the result is used as the initial weight parameter of the new item, providing a reasonable initial reference value for subsequent model forward propagation and fine-tuning training.

[0108] If there is no clothing of the same category and season as the newly added clothing item in the historical logistics database, the initial weight parameter selection strategy will be set as the third priority acquisition strategy, and the overall average weight of the clothing category in all seasons will be used as the initial weight parameter of the clothing item.

[0109] It should be noted that if there is a lack of similar data for clothing of the same category and season as the newly added clothing item in the historical logistics database, the initial weight parameter selection strategy will be designated as the third priority acquisition strategy. That is, the overall average weight of the clothing category under all seasons will be used as the initial weight parameter of the new item. In this way, even if there is no specific historical data, a reasonable initial weight estimate can still be provided for the new item based on the overall characteristics of the category, ensuring that the model can successfully carry out forward prediction and incremental fine-tuning training.

[0110] Furthermore, the training strategy is set up, and the specific analysis method is as follows: The training rounds, initial learning rate, and partitioning ratio are determined based on the cumulative number of newly added packages.

[0111] It should be noted that in this embodiment, the database pre-defines specific rules for determining the learning rate training epochs, initial model learning rate, and partition ratio based on the cumulative number of newly added packages, and stores these rules in a configuration file or parameter table for unified invocation during model training. At runtime, the cumulative number of newly added packages is first obtained, reflecting the changes in dataset size or the growth of business load. Then, according to the preset rules, a correspondence is established between the cumulative number of newly added packages and the learning rate training epochs, initial model learning rate, and partition ratio using interval mapping tables, function fitting, or interpolation methods, thereby automatically determining the matching parameter values. The core mapping logic is that a larger cumulative number of newly added packages indicates a larger data volume or higher business complexity. To ensure the sufficiency and stability of model training, the learning rate training epochs, initial model learning rate, and partition ratio are correspondingly larger. Through this positively correlated mapping method, the system can dynamically optimize training parameters, ensuring efficient learning and reliable performance of the model in a data-driven environment.

[0112] After determining the initial weight parameters, fine-tune the training strategy settings, and determine the number of warm-up rounds and the initial value decay ratio based on the learning rate and training rounds.

[0113] It should be noted that in this embodiment, the database pre-sets specific rules for determining the warm-up epochs and initial value decay ratio based on the learning rate training epochs, and stores these rules in a configuration file or parameter table for unified invocation during model training. At runtime, the currently set learning rate training epochs are first obtained. These epochs reflect the expected total duration and iterative complexity of the model training process. Based on the preset rules, a correspondence between the learning rate training epochs, warm-up epochs, and initial value decay ratio is established using interval mapping tables, function fitting, or interpolation methods, thereby automatically determining the matching parameter values. The core mapping logic is that a larger learning rate training epoch indicates a longer total training cycle and a more sufficient training process. To match this long-term characteristic, the warm-up epochs are set accordingly larger to ensure a longer stable upward phase in the early stages of training, avoiding getting trapped in local optima due to excessively high initial learning rate values. Simultaneously, the initial value decay ratio is set smaller, thus lengthening the learning rate decrease process, allowing it to decrease smoothly and stably over a longer training cycle, ultimately ensuring that the model can fully converge and achieve better performance.

[0114] The detected new clothing package data is divided into training set and validation set according to the division ratio.

[0115] It should be noted that upon detecting newly added clothing package data, the system retrieves the currently set partitioning ratio from the database and initiates the data partitioning process. Based on this ratio, a globally unique sequence identifier is assigned to each newly added clothing package data record. A random number generator is used to apply uniformly distributed random seeds to all data to be partitioned, ensuring the randomness and unbiasedness of each partition. The system strictly calculates the precise data splitting points according to the preset partitioning ratio, dividing the total dataset into two parts. The larger portion is used as the training set for model learning and parameter fitting; the smaller portion is used as the validation set for evaluating model performance and adjusting hyperparameters. This ensures that from the very beginning of model training, there is a data subset with a clear structure and purpose, laying a solid foundation for training a model with strong generalization ability.

[0116] For the training set, the learning rate is linearly increased to the initial value of the model learning rate during the warm-up rounds.

[0117] It should be noted that the system retrieves the pre-set warm-up rounds and initial model learning rate from the database. In the first training round at the start of the warm-up phase, the learning rate does not directly use the initial value but starts from 0. In each subsequent training round, the system dynamically calculates the current learning rate according to the formula: current learning rate = initial learning rate multiplied by the current round divided by the total number of warm-up rounds. This causes the learning rate to increase linearly with the increase of training rounds. This process continues until the current training round reaches the preset upper limit of the warm-up rounds. At this point, the learning rate has just steadily increased to the predetermined initial model learning rate value. Thus, the warm-up phase ends, and the learning rate enters the predetermined decay phase. This mechanism effectively avoids gradient oscillations caused by excessively high learning rates in the early stages of training, ensuring the stability of the early training phase.

[0118] Subsequently, following a multi-step decay strategy, the learning rate is reduced in segments and jumps according to preset nodes after the warm-up rounds, until the learning rate is reduced to the initial decay ratio.

[0119] It should be noted that after the warm-up phase, a multi-step decay strategy is immediately initiated to finely adjust the learning rate. First, a series of key decay nodes and the decay ratio corresponding to each node are read from the database. During training, the current training epoch is continuously monitored and compared in real time with the preset decay node sequence. Once the current epoch reaches a certain preset node, the decay operation is immediately executed, multiplying the current learning rate by the decay ratio corresponding to that node, thereby achieving a segmented and jump-like reduction of the learning rate. This process is repeated throughout the entire training cycle until all preset decay nodes are traversed. By performing multiplicative decay sequentially at multiple key nodes, the learning rate is gradually and segmentally reduced from its initial value to the final value determined by the cumulative decay ratio, thus promoting model convergence to the optimal solution with a more refined step size in the later stages of training.

[0120] After each training epoch with the learning rate, forward inference is performed using the validation set to calculate the validation set loss. Training automatically terminates when the validation set loss no longer decreases in consecutive epochs, triggering an early stopping mechanism. After training, only the model parameter file with the best performance on the validation set is saved, achieving optimal overall prediction performance. It's important to note that after each training epoch with a different learning rate, a comprehensive performance evaluation of the current model is performed using validation set data that wasn't used in training. The validation set loss is calculated through forward inference, and this loss is a key indicator of the model's generalization ability. Based on this indicator, an early stopping mechanism is activated for continuous monitoring. When the validation set loss fails to decrease within a pre-set threshold for consecutive epochs, the model's performance is automatically determined to have stopped improving, and training is immediately terminated. This effectively prevents overfitting and conserves computational resources. Simultaneously, throughout the entire training process, the performance on the validation set is continuously tracked. Instead of simply saving the model from the last epoch, only the model parameters from the epoch with the best performance on the validation set (i.e., the epoch with the smallest validation set loss) are selectively saved as the final file. This ensures that the deployed model is the version with the strongest generalization ability generated during its training, thereby maximizing the overall predictive performance.

[0121] It's important to note that the consecutive round threshold is a predefined key parameter. Its logic lies in determining whether the model's performance has plateaued. This value is explicitly set and stored in the configuration, representing the minimum number of rounds for which the validation set loss can be monitored without decreasing. During runtime, a counter is maintained. Each time the validation set loss fails to surpass the historical best value in the next round, the counter is incremented; once the loss reaches a new low, the counter is immediately reset. The core pre-set logic is that early stopping is only triggered when this counter accumulates to the pre-set consecutive round threshold. This logic requires a trade-off between avoiding premature termination of training with potential and preventing unnecessary resource consumption after performance saturation. A larger threshold allows the system to tolerate performance plateaus for a longer period, resulting in more thorough training but also higher computational costs. Conversely, a smaller threshold allows the system to stop more quickly but may miss opportunities for subsequent model convergence.

[0122] The weight prediction module is used to take the sequence and quantity of clothing items in the clothing package as input to the MLP model and output the predicted weight of the clothing package data.

[0123] like Figure 2 This is a flowchart of the clothing package weight prediction method based on MLP fine-tuning provided in the embodiments of the present invention, including: extracting clothing package data from the historical logistics database, performing outlier processing, initializing the weight of clothing goods after outlier processing, and then performing full training of the MLP model.

[0124] After the MLP model is fully trained, historical clothing wrapping data offset detection is performed at a preset offset monitoring period, and data correction is performed when the detection result shows a systematic offset.

[0125] After data correction, the incremental fine-tuning trigger condition is determined, and the initial weight parameter selection strategy is determined when incremental fine-tuning is triggered. The initial weight parameters of the clothing product are determined, the training strategy is set, and incremental fine-tuning is performed accordingly.

[0126] The MLP model takes the clothing item sequence and quantity in the clothing package as input and outputs the predicted weight of the clothing package data.

[0127] like Figure 5 This is a flowchart of the clothing package weight prediction method based on MLP fine-tuning provided in this embodiment of the invention. It includes collecting historical logistics data and handling outliers, then initializing the weight of each product based on the cleaned data, and then using the initialized product weight parameters to perform full model training to obtain a preliminary package weight prediction output. When new data arrives, the data is processed and the product weight is initialized first, and then the model parameters are updated through incremental fine-tuning training. Finally, the fine-tuned package weight prediction result is output.

[0128] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0129] It should be understood that determining B based on A does not mean determining B solely based on A; it also means determining B based on A and / or other information.

[0130] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0131] The above description is only an optional embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A garment package weight prediction system based on MLP fine-tuning, characterized in that, The system includes: an MLP full training module, a system offset detection and correction module, a clothing wrapping incremental fine-tuning module, and a weight prediction module; The MLP full training module is used to extract clothing package data from the historical logistics database, perform outlier handling, initialize the weight of clothing goods after outlier handling, and thus perform full training of the MLP model. The system offset detection and correction module is used to detect the offset of historical clothing wrapping data at a preset offset monitoring period after the MLP model is fully trained, and to correct the data when the detection result is a system offset. The incremental fine-tuning module for clothing packaging is used to determine the triggering condition for incremental fine-tuning after data correction, and to determine the initial weight parameter selection strategy when incremental fine-tuning is triggered, to determine the initial weight parameters of the clothing product, to set the training strategy and thereby execute incremental fine-tuning. The weight prediction module is used to take the sequence and quantity of clothing items in the clothing package as input to the MLP model and output the predicted weight of the clothing package data. The outlier handling process is described in the following specific analysis method: Extract clothing parcel data from historical logistics databases and perform data cleaning; The clothing package data is grouped according to the category and season of the clothing products to obtain each true weight outlier detection group. The first quartile, third quartile and interquartile range are calculated for the package weight data of each true weight outlier detection group. The lower limit of the outlier is obtained based on the first quartile and the interquartile range, and the upper limit of the outlier is obtained based on the third quartile and the interquartile range. For a certain real weight outlier, there is a detection group. Real weight data that is greater than the upper limit of the outlier or less than the lower limit of the outlier is marked as an outlier, and real weight data that is less than or equal to the upper limit of the outlier and greater than or equal to the lower limit of the outlier is marked as a normal value. The outlier value of the actual weight is removed from the detection group, and an initial weight signal for the clothing product is generated after the removal. If the actual weight outlier exists, but the detection group does not have an outlier, then the garment weight initialization signal is directly generated.

2. The garment package weight prediction system based on MLP fine-tuning as described in claim 1, characterized in that, The specific analysis method for initializing the weight of clothing items is as follows: Upon receiving the garment weight initialization signal, the garment weight is initialized. For individual garment items, check the individual garment items in the historical logistics database; If a single garment item exists in the historical logistics database, the historical single garment items are statistically analyzed, the average weight of the single garment item is calculated, and this average weight is used as the initial weight of the garment item. If no single garment item exists in the historical logistics database, the average weight of all garment items in the same category will be used as the initial weight of the garment item.

3. The garment package weight prediction system based on MLP fine-tuning as described in claim 1, characterized in that, The specific analysis method for performing full training of the MLP model is as follows: The sequence and quantity of clothing items in each package are used as input, and the package weight is calculated according to the model prediction formula for forward propagation. The mean squared error is used as the loss function for error calculation. The gradient of the weight parameter of each garment item is obtained by taking the derivative of the loss function, and the gradient solution is then performed accordingly. The weight parameters of the clothing items are updated using the gradient descent algorithm to obtain new weight parameters for the clothing items. The updated clothing weight parameters will generate new package weight predictions in the next iteration of forward propagation; In each iteration, the forward propagation, error calculation, gradient solution, and weight update processes are repeated continuously, so that the weight parameters of the clothing products gradually converge to the optimal solution of the overall prediction error.

4. The garment package weight prediction system based on MLP fine-tuning as described in claim 3, characterized in that, The specific analysis method for calculating the package weight based on the model prediction formula is as follows: The input consists of a list of clothing items, and each sample contains a sequence of clothing item IDs and their corresponding quantity sequences. The initial weight of each garment item is obtained based on the initial weight of the garment item, and is recorded as the garment item weight parameter; The model prediction formula is defined as follows: multiply the weight parameter of each clothing item by its corresponding quantity, sum the results for all clothing items, and form a linear weighted superposition process to output the predicted weight of the corresponding clothing package.

5. The garment package weight prediction system based on MLP fine-tuning as described in claim 1, characterized in that, The process involves detecting historical clothing package data offsets and correcting data when systematic offsets are detected. The specific analysis method is as follows: Historical clothing package data is grouped and statistically analyzed by clothing item to obtain each drift detection group. The mean weight and standard deviation of clothing items in each drift detection group are calculated. The baseline mean weight and standard deviation of clothing items in each drift detection group are extracted from the database. Extract the preset drift threshold from the database; The mean weight and standard deviation weight of each garment item within each drift detection group are used as the basis for the mean weight and standard deviation weight of each drift detection group to obtain the mean weight and standard deviation weight values. If the mean weight deviation of a certain group is greater than or equal to the drift threshold, then the group is determined to have a systematic bias. The group weight sequence is obtained by taking the weight data of each weight data point and the weight mean deviation within the group, and the mean of the group weight sequence is then processed to obtain the benchmark mean within the group. And the scaling factor of the distribution width is obtained based on the weight standard deviation value; Scaling is performed on each weight data in the group based on the scaling factor and the mean of the group baseline to obtain the weight data of each drift existence detection group. After processing, the updated weight data of the group is obtained. After calibration, the grouped updated weight data is used as the distribution standard for incremental fine-tuning. If the mean weight deviation of a certain group is less than the drift threshold, it is determined that there is no systematic deviation in that group, and the weight data of that group is directly used as the distribution standard for incremental fine-tuning.

6. The garment package weight prediction system based on MLP fine-tuning as described in claim 1, characterized in that, The specific analysis method for determining the trigger condition for incremental fine-tuning is as follows: Continuously monitor newly added clothing package data. When newly added clothing package data is detected, extract, clean and filter outliers from the newly added clothing package data, and construct triplet input samples for each package. Extract the preset fine-tuning period, package quantity threshold, and cumulative number threshold from the database; The cumulative number of newly added packages is counted within each fine-tuning cycle; When the cumulative number of newly added packages reaches or exceeds the package quantity threshold, an incremental fine-tuning is triggered. If the cumulative number of newly added packages does not reach the package quantity threshold, the cumulative number of times in the fine-tuning cycle is checked. If the cumulative number is greater than or equal to the cumulative number threshold, incremental fine-tuning is triggered directly. If the cumulative number is less than the cumulative number threshold, the cumulative number of newly added packages in the current cycle is carried over to the next cycle, and the cumulative number is incremented by one.

7. The garment package weight prediction system based on MLP fine-tuning as described in claim 1, characterized in that, The strategy for determining the initial weight parameters, specifically for determining the initial weight parameters of clothing products, is analyzed using the following method: For newly added clothing package data, extract clothing product information. If the clothing product information is found in the existing clothing product information in the historical logistics database, the initial weight parameter selection strategy is recorded as the first priority acquisition strategy. Extract a preset sample size threshold from the database; The weight of the garment is calculated based on the historical single-item garment package data, and all weight records of the garment in the historical single-item packages are extracted for outlier filtering. When the number of samples is less than or equal to the sample number threshold, outlier removal is performed using the seasonal quantile boundary of the category to obtain the weight data sequence of the clothing product. When the number of samples exceeds the sample size threshold, multi-layer anomaly filtering is performed by comprehensively using the interquartile range of clothing weight, category boundaries, and the 3σ principle to obtain the weight data sequence of the clothing product. The arithmetic mean of the weight data series of the clothing item is used as the initial weight parameter of the clothing item. If the clothing product information is not found in the existing clothing product information in the historical logistics database, check whether there are other clothing products of the same category and season as the newly added clothing product in the historical logistics database; If there are clothing items of the same category and season as the newly added clothing item in the historical logistics database, they are recorded as similar clothing items, the initial weight parameter selection strategy is recorded as the second priority acquisition strategy, and the average weight of similar clothing items is taken as the initial weight parameter of the clothing item. If there is no clothing of the same category and season as the newly added clothing item in the historical logistics database, the initial weight parameter selection strategy will be set as the third priority acquisition strategy, and the overall average weight of the clothing category in all seasons will be used as the initial weight parameter of the clothing item.

8. The garment package weight prediction system based on MLP fine-tuning as described in claim 1, characterized in that, The specific analysis method for setting the training strategy is as follows: The training rounds, initial learning rate, and partitioning ratio are determined based on the cumulative number of newly added packages. After the initial weight parameters are determined, the training strategy is fine-tuned, and the number of warm-up rounds and the initial value decay ratio are determined according to the number of training rounds of the learning rate. The detected new clothing package data is divided into training set and validation set according to the division ratio; For the training set, the learning rate is linearly increased to the initial value of the model learning rate during the warm-up rounds; Then, following a multi-step decay strategy, the learning rate is reduced in segments and jumps according to preset nodes after the warm-up round, until the learning rate is reduced to the initial decay ratio. After each training epoch with a learning rate, forward inference is performed using the validation set to calculate the validation set loss. Training is automatically terminated when the validation set loss no longer decreases in consecutive epochs, thus triggering an early stopping mechanism. After training is completed, only the model parameter file with the best performance on the validation set is saved to achieve optimal overall prediction performance.

9. A method for predicting the weight of clothing packages based on MLP fine-tuning, applied to the clothing package weight prediction system based on MLP fine-tuning according to any one of claims 1-8, characterized in that, The method includes: Extract clothing package data from historical logistics database, perform outlier handling, initialize clothing weight after outlier handling, and then perform full training of MLP model. After the MLP model is fully trained, historical clothing wrapping data offset detection is performed at a preset offset monitoring period, and data correction is performed when the detection result is a systematic offset. After data correction, the incremental fine-tuning trigger condition is determined, and the initial weight parameter selection strategy is determined when incremental fine-tuning is triggered. The initial weight parameters of the clothing product are determined, the training strategy is set, and incremental fine-tuning is performed accordingly. The MLP model takes the clothing item sequence and quantity in the clothing package as input and outputs the predicted weight of the clothing package data.

Images