Weighing equipment control system based on active learning

Abstract

The invention provides a weighing equipment control system based on active learning, and relates to the technical field of data processing, and the system comprises a sample evaluation module which is used for carrying out the sample confidence evaluation of analysis data, obtaining a sample confidence score, and when the sample confidence score is lower than a preset confidence threshold value, outputting an evaluation result; marking the analysis data as to-be-learned data; the deviation calculation module is used for estimating the feature vector according to the feature vector of the to-be-learned data and based on a current weight estimation model, outputting an estimation result and carrying out difference calculation on the estimation result and an actual result; the parameter generation module is used for calculating an error contribution value on each feature dimension in combination with feature dimension weight distribution in the feature vector, and performing nonlinear mapping to generate an updated parameter vector; the model updating module is used for performing incremental training on the current weight estimation model according to the updated parameter vector; according to the invention, the autonomy and accuracy of the weighing equipment control system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and particularly to a weighing device control system based on active learning. Background Art

[0002] In the prior art, the control systems of weighing devices generally use preset parameters and fixed models for weight detection and control. Such systems usually rely on manual calibration to complete the initial setting of the device. By collecting sensor data (such as pressure sensors, resistance strain gauges, etc.) and inputting it into a preset algorithm, the weight calculation of the target object is realized. In scenarios such as industrial production lines and logistics sorting, the weighing system may also be equipped with a PLC or an embedded control unit to achieve automatic weighing and process control. Although some systems have a data feedback mechanism and can adjust the weighing strategy according to set rules, the overall still mainly uses a static learning mode and lacks the ability to actively identify and retrain new samples or edge data, resulting in poor adaptability.

[0003] During the packaging process of agricultural products, due to the large sample differences, such as the different sizes, densities, and surface attachments of natural crops like potatoes and onions, it is difficult for the weighing system based on a fixed model to accurately estimate the actual weight. For example, when the surface of a potato is attached with wet mud or there are multiple connected individuals, the traditional weighing system may misidentify it as a single heavy object, resulting in classification errors, which in turn affect the subsequent packaging process. Due to the lack of the ability to actively learn abnormal samples, the system cannot automatically adjust the model parameters or collect new samples for training. Over time, error accumulation will occur, and frequent manual intervention and recalibration are required, reducing the adaptive level and practicality of the system. Summary of the Invention

[0004] The purpose of the present invention is to provide a weighing device control system based on active learning, aiming to solve the problems mentioned in the background art.

[0005] To solve the above technical problems, the technical solution of the present invention is as follows:

[0006] A weighing device control system based on active learning, the system includes:

[0007] A sample evaluation module, used to evaluate the sample confidence of the analysis data, specifically including: calculating the feature distance between the historical label sample and the analysis data, and combining the timestamp and the weighing frequency to obtain the sample confidence score. When the sample confidence score is lower than the preset confidence threshold, the analysis data is marked as data to be learned;

[0008] A deviation calculation module, used to estimate the feature vector according to the feature vector of the data to be learned and based on the current weight estimation model, output the estimation result, and calculate the difference between it and the actual result to obtain the estimation deviation value;

[0009] A parameter generation module, configured to calculate an error contribution value for each feature dimension according to an estimated deviation value in combination with the feature dimension weight distribution in the feature vector, and perform a non-linear mapping to generate an update parameter vector for adjusting the weight estimation model;

[0010] A model update module, configured to perform incremental training on the current weight estimation model according to the update parameter vector to obtain an updated weight estimation model.

[0011] Preferably, the sample evaluation module includes:

[0012] A feature distance calculation sub-module, configured to calculate a weighted Euclidean distance between the feature vector of the analysis data and multiple feature vectors in the historical label samples to obtain a sample feature distance;

[0013] A time correlation processing sub-module, configured to calculate a time density value according to the timestamp of the analysis data and the number of weighings per unit time, and perform weighted fusion according to the time density value and the sample feature distance to obtain a composite evaluation factor;

[0014] A confidence level evaluation sub-module, configured to obtain a sample confidence level score according to the composite evaluation factor and a preset confidence level scoring model, and mark the current sample as data to be learned when the sample confidence level score is lower than a preset confidence level threshold.

[0015] Preferably, the parameter generation module includes:

[0016] An error contribution analysis sub-module, configured to calculate an error contribution value for each feature dimension according to the correlation between the estimated deviation value and the feature vector in the data to be learned;

[0017] A parameter mapping sub-module, configured to perform interval grading on the error contribution value and perform non-linear mapping according to different mapping response methods for different levels to obtain a mapped error contribution value;

[0018] An update parameter generation sub-module, configured to combine the mapped error contribution values in the order of feature dimensions to generate an update parameter vector.

[0019] Preferably, the model update module includes:

[0020] An orientation training sub-module, configured to input the update parameter vector into the current weight estimation model, and perform incremental training on the current weight estimation model based on parameter replacement, weighted update or learning rate adjustment without resetting the network structure and historical parameters of the weight estimation model to obtain an updated weight estimation model;

[0021] An accuracy evaluation sub-module for performing estimation calculations on the updated weight estimation model on both the new and old sample sets respectively to obtain the corresponding mean error and standard deviation;

[0022] A stability feedback sub-module for constructing an error change curve based on the mean error and standard deviation, thereby determining whether the updated weight estimation model meets the preset stability criteria. When the judgment result is not satisfied, a training adjustment signal is output for correcting the mapping response mode of the subsequent updated parameter vector or adjusting the training frequency.

[0023] Preferably, the feature distance calculation sub-module includes:

[0024] A feature vector extraction unit for extracting the feature vectors constituting the sample features from the analysis data;

[0025] A weight factor determination unit for performing statistical analysis on each feature dimension in the historical label samples. By evaluating the numerical fluctuation degree of each feature dimension within a preset first time window, it is determined that the weight value corresponding to the relatively stable features is higher, while the weight value of the features with large fluctuations is lower, thereby generating a set of weighting factors for distance calculation.

[0026] A distance calculation unit for performing dimension-by-dimension difference calculations on the feature vector of the current analysis data and the multiple feature vectors of the historical label samples, and accumulating the sum after multiplying each difference by the corresponding weighting factor to obtain a feature distance value representing the sample similarity degree.

[0027] Preferably, the time correlation processing sub-module includes:

[0028] A time information extraction unit for extracting the timestamp information of the samples from the analysis data and counting the number of weighing records within the current preset second time window to obtain the weighing frequency per unit time;

[0029] A time density calculation unit for calculating the time interval length of the samples by comparing the timestamp information of the samples with the current system time, and performing weighted combination after normalizing the reciprocal of the time interval length and the weighing frequency per unit time respectively to calculate the time density value;

[0030] A weighted fusion unit for jointly evaluating the feature distance value and the time density value. By setting the adjustment parameter that changes with the time density value, the influence degree of the samples collected during the high-density period on the feature distance value is enhanced, thereby forming a composite evaluation factor for representing the comprehensive evaluation level of the samples.

[0031] Preferably, the confidence evaluation sub-module includes:

[0032] A scoring parameter loading unit for loading a preset confidence scoring model, which is constructed based on multiple influencing factors, including a composite evaluation factor, a time density value corresponding to a sample, and a local fluctuation amplitude of an original signal sequence, and is used to comprehensively evaluate the confidence level of a current sample;

[0033] A scoring calculation unit for inputting multiple influencing factors into the confidence scoring model and calculating a confidence scoring value of the current sample according to the parameter weights corresponding to each factor;

[0034] A threshold comparison unit for comparing the confidence scoring value of the current sample with a preset confidence threshold. When the score is lower than the threshold, it outputs information of a sample to be learned and classifies the sample into a data set to be learned; if the score is higher than the threshold, the current sample is ignored for model training.

[0035] Preferably, the error contribution analysis sub-module includes:

[0036] An eigenvalue analysis unit for calculating the proportion information of each feature in the entire feature vector according to the feature vector of the data to be learned and its corresponding estimated deviation value and based on the relative magnitudes of the eigenvalues of each feature dimension;

[0037] A correlation adjustment unit for determining the relative contribution degree of each feature dimension to the total deviation according to the proportion information and the estimated deviation value, and generating an error contribution value. Among them, a reduction coefficient is set for the feature dimension with a large eigenvalue change range but low stability in training to reduce its dominant role in parameter generation.

[0038] Preferably, the parameter mapping sub-module includes:

[0039] A contribution level discrimination unit for classifying the error contribution value into multiple levels according to the set contribution classification standard based on the error contribution value of each feature dimension, and generating a contribution level classification result. Among them, the error contribution value of a higher level will obtain a larger parameter adjustment range;

[0040] A non-linear mapping strategy unit for performing a non-linear mapping response process on the error contribution value through a mapping function group according to the contribution level classification result, and generating a mapped error contribution value. Among them, a compression mapping is adopted for the error contribution value of a lower level to suppress the model change caused by weak perturbations, and an enhancement mapping is adopted for the error contribution value of a higher level to accelerate model convergence.

[0041] Preferably, the stability feedback sub-module includes:

[0042] An error trend extraction unit for constructing an error change curve according to the change conditions of the error mean and standard deviation in the current multiple update cycles;

[0043] A stability judgment unit, configured to judge whether the current model is in a stable state after being updated according to the slope and fluctuation frequency of the error change curve. If the fluctuation amplitude continues to rise or the slope exceeds a set critical value, it is judged as not meeting the stability standard; otherwise, it meets the standard.

[0044] A feedback generation unit, configured to generate a training adjustment signal when it is judged that the current model is in an unstable state after being updated. The training adjustment signal is used to trigger the parameter mapping sub-module to re-adjust the mapping range or reduce the model update frequency.

[0045] The above solution of the present invention has at least the following beneficial effects:

[0046] First of all, by setting up an active learning mechanism, the present invention can dynamically identify and mark low-confidence analysis data as samples to be learned, breaking through the limitation of "relying on a fixed model and static operation" in the existing weighing control system. The sample evaluation module compares the features of the analysis data with the historical label samples, and combines the timestamp and weighing frequency to construct a confidence score in real time, thus realizing a learning mechanism of identifying, sampling while running. This feature enables the system to automatically perceive new samples, edge samples or special state samples (such as natural crops with attachments or structural connectors on the surface), improving the model's ability to identify abnormal states and reducing the frequency of manual intervention.

[0047] Secondly, the present invention constructs an error-driven model update mechanism through the parameter generation module. After the deviation calculation module estimates the data to be learned, the system decomposes the estimated deviation into each feature dimension through error contribution analysis, and then combines the non-linear mapping strategy to generate a structured and controllable update parameter vector. This solution realizes the fine-tuning of the local parameters of the weight estimation model, avoids the computational burden brought by global reconstruction or full-scale training in the prior art, and realizes incremental training through the model update module, enabling the model to have the ability to quickly adapt to new data patterns while maintaining stability.

[0048] Thirdly, the accuracy evaluation and stability feedback mechanism provided by the present invention enables the system to automatically detect the change trend of the prediction error after the model is updated, dynamically judge whether the model is in a convergent state, and output a training adjustment signal when the stability standard is not met. By restricting the update frequency or adjusting the parameter mapping response range, the model oscillation and error amplification phenomena are effectively suppressed, thereby improving the robustness and safety of the model during long-term operation.

[0049] Finally, the present invention is suitable for weighing application scenarios with strong differences in natural objects, such as the automatic weighing of potatoes, onions and other objects in agricultural product packaging lines. The system can automatically adapt to changes in sample volume, density, and surface state, significantly improve weighing accuracy and sorting accuracy, reduce false alarm rate and manual calibration frequency, and enhance the practicality and maintainability of the system in complex environments. Overall, the present invention has the triple capabilities of self-learning, self-adjustment, and self-stabilization, which is significantly better than the static model control system in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is an architecture diagram of a weighing equipment control system based on active learning provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0051] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0052] like Figure 1 As shown, an embodiment of the present invention proposes a weighing equipment control system based on active learning, the system comprising:

[0053] A data acquisition module, used to acquire original sensor data of a weighing target, wherein the original sensor data includes a plurality of voltage signals collected by the weighing device within a unit time;

[0054] The data preprocessing module is used to remove extreme value signals beyond the empirical range based on the original sensor data, smooth the remaining signal sequences, and normalize the data based on the sensor point locations to obtain analysis data;

[0055] The sample evaluation module is used to evaluate the sample confidence of the analysis data, including: calculating the feature distance between the historical label sample and the analysis data, and combining the timestamp and the weighing frequency to obtain the sample confidence score. When the sample confidence score is lower than the preset confidence threshold, the analysis data is marked as data to be learned;

[0056] The deviation calculation module is used to estimate the feature vector according to the feature vector of the data to be learned and based on the current weight estimation model, output the estimated result, and perform difference calculation between the estimated result and the actual result to obtain the estimated deviation value;

[0057] A parameter generation module, configured to calculate an error contribution value for each feature dimension according to the estimated deviation value and in combination with the feature dimension weight distribution in the feature vector, and perform a non-linear mapping to generate an update parameter vector for adjusting the weight estimation model;

[0058] A model update module, configured to perform incremental training on the current weight estimation model according to the update parameter vector to obtain an updated weight estimation model;

[0059] A weight control module, configured to perform real-time weight estimation on the subsequent input analysis data according to the updated weight estimation model, and use the estimation result as the weighing control output.

[0060] In an embodiment of the present invention, by constructing a weighing device control system based on an active learning mechanism, dynamic update and accuracy enhancement of a traditional weighing control model are realized. The system adopts a multi-level module structure. First, a data acquisition module acquires original sensing data of a weighing target. The data is derived from multiple voltage signals output by a pressure sensor within a unit time, and these signals are used to represent the actual force exerted on the weighing platform by the object to be measured. Through a high-speed sampling method, the time resolution of data acquisition and the continuity of physical response are ensured, providing a high-quality data basis for subsequent processing.

[0061] Next, a data preprocessing module performs operations such as outlier removal, smoothing filtering, and normalization on the original sensing data. The removal of extreme values is based on a set empirical threshold, the smoothing process can use time series methods such as moving average, and the normalization operation is adjusted differently according to the sensor layout position. After the processing is completed, the generated analysis data can be used to further evaluate the value and confidence level of the sample.

[0062] The system introduces a sample evaluation module to compare the similarity between the current analysis data and historical labeled samples, and combines the timestamp and the weighing frequency per unit time to form a dynamic sample confidence score. When the score value is lower than the threshold set by the system, the system marks the sample as "data to be learned". Compared with a traditional static system that relies on manual setting of boundaries, this evaluation mechanism can actively identify model blind spots, thereby stimulating the value of samples.

[0063] Subsequently, after obtaining the feature vector of the data to be learned, a deviation calculation module performs a valuation output based on the currently used weight estimation model, and calculates the difference from the actual physical mass to form a deviation value reflecting the model error. This deviation value is used to measure the generalization ability of the current model on new samples and serves as the starting point for subsequent parameter optimization.

[0064] The parameter generation module evaluates the specific influence intensity of each feature dimension in the deviation according to the deviation value and the importance of each dimension in the feature vector (i.e., the dimension weight distribution), and performs non-linear mapping processing on it to generate an update parameter vector for model update. This process avoids over-responding to weak perturbations and strengthens the adaptive weight adjustment for key feature dimensions.

[0065] The model update module does not reset the overall model structure, but performs incremental training on the basis of the original parameters and only makes directional corrections to the weights. This update strategy not only retains the steady-state performance of the original model but also realizes rapid adaptation to the characteristics of new data.

[0066] Finally, the system performs real-time estimation operations on the subsequent analyzed data through the weight control module, and outputs the results as weighing control signals to drive downstream mechanical execution devices such as packaging sorting to complete fine control.

[0067] The above structure strengthens the system's autonomous learning ability and response sensitivity, can significantly reduce the model aging speed in continuous operation scenarios, improve data utilization rate, and finally achieve low-intervention and high-precision intelligent weighing control.

[0068] In a preferred embodiment of the present invention, the sample evaluation module includes:

[0069] A feature distance calculation sub-module for calculating the weighted Euclidean distance between the feature vector of the analyzed data and multiple feature vectors in the historical label samples to obtain the sample feature distance;

[0070] A time correlation processing sub-module for calculating the time density value according to the timestamp of the analyzed data and the number of weighings per unit time, and performing weighted fusion according to the time density value and the sample feature distance to obtain a composite evaluation factor;

[0071] A confidence evaluation sub-module for obtaining the sample confidence score according to the composite evaluation factor and a preset confidence scoring model, and marking the current sample as data to be learned when the sample confidence score is lower than the preset confidence threshold.

[0072] In the embodiment of the present invention, to improve the accuracy and active screening ability of weighing data confidence evaluation, a sample evaluation module composed of a feature distance calculation sub-module, a time correlation processing sub-module and a confidence evaluation sub-module is designed. This module realizes the efficient determination of the sample confidence level through a spatio-temporal dual feature fusion mechanism.

[0073] First, the feature distance calculation sub-module receives the feature vectors extracted from the analysis data and calculates the weighted Euclidean distance with each feature vector in the historical label sample library. In this process, the weight factor of each feature dimension is set according to its fluctuation amplitude within a preset time window. The higher the stability of the feature, the higher its weight. The sample feature distance calculated after weighting reflects the overall deviation degree between the current sample and the label sample, serving as the first basis for evaluating the "knownness" of the sample.

[0074] Secondly, the time correlation processing sub-module extracts the sample timestamp from the analysis data and counts the weighing frequency of the current system within the second time window. According to the interval between the sample collection time and the current system time, combined with the frequency level, a time density value reflecting the "freshness" of the sample is calculated. To reflect the priority of sample timeliness, the system sets an adjustable density response factor to jointly fuse the time density value and the feature distance to generate a composite evaluation factor.

[0075] This factor not only describes the feature difference degree between the sample and the historical label but also weighs the real-time nature of its time distribution, effectively identifying new data types not fully covered by the model.

[0076] Finally, the confidence evaluation sub-module uses the composite evaluation factor and combines the confidence scoring model loaded by the current system to quantitatively calculate the sample confidence level. This scoring model considers the feature deviation degree, time density, and volatility of the original signal sequence of the sample. After obtaining the scoring value, the system compares it with a preset threshold. When the score is lower than the threshold, a marking signal is immediately output, and the sample is included in the set of samples to be learned for subsequent optimization training.

[0077] This module implements a self-learning data screening mechanism oriented by sample quality, avoiding a large number of invalid samples from participating in training, improving the gold content of training data, and thus enhancing the model iteration efficiency and training convergence speed.

[0078] In a preferred embodiment of the present invention, the parameter generation module includes:

[0079] An error contribution analysis sub-module for calculating the error contribution value of each feature dimension according to the correlation between the estimated deviation value and the feature vector in the data to be learned;

[0080] A parameter mapping sub-module for classifying the error contribution value into intervals and performing non-linear mapping according to different levels in different mapping response manners to obtain the mapped error contribution value;

[0081] An updated parameter generation sub-module for combining the mapped error contribution values in the order of feature dimensions to generate an updated parameter vector.

[0082] In the embodiments of the present invention, in order to improve the fineness and pertinence of model updates, an error-driven parameter generation module is proposed, which is used to generate a structured update parameter vector starting from the deviation value to guide subsequent model incremental training. This module includes an error contribution analysis sub-module, a parameter mapping sub-module, and an update parameter generation sub-module. The core idea is to use the error performance of the feature dimension to drive the weight update intensity to achieve a dynamic balance between control accuracy and convergence speed.

[0083] First, the error contribution analysis sub-module receives the feature vector of the data to be learned and its estimated deviation value, and analyzes the proportion of each feature dimension in the feature vector, that is, evaluates the relative importance of the value of each dimension in the overall vector. The system further calculates the contribution degree of each feature dimension to the overall deviation according to the combination relationship between the feature proportion and the deviation value. For those feature dimensions that show strong volatility or instability in historical training, the system will automatically set a reduction factor to avoid them dominating the training direction of the model.

[0084] Subsequently, the parameter mapping sub-module performs interval grading based on the error contribution value of each feature dimension to form a contribution level mapping structure. Each level corresponds to a response mode: dimensions with smaller contribution values will adopt a compression response strategy to suppress the excessive amplification of parameters by minor perturbations; dimensions with larger contribution values will activate an enhanced mapping mechanism to amplify the weight adjustment ratio in model updates. This mapping process is completed based on a non-linear mapping function and can achieve an adaptive curve response to the input error.

[0085] Finally, the update parameter generation sub-module recombines the above-mapped error contribution values in the order of feature dimensions to construct a complete update parameter vector. This vector not only reflects the explanatory ability of each dimension for the current deviation but also controls its participation degree through mapping, thereby realizing differential training updates.

[0086] This module strengthens the controllability and interpretability of the update parameters, realizes the rapid adaptation to key features while maintaining the convergence stability of the model, and finally improves the generalization ability and response accuracy of the model in actual weighing tasks.

[0087] In a preferred embodiment of the present invention, the model update module includes:

[0088] An orientation training sub-module, which is used to input the update parameter vector into the current weight estimation model and perform incremental training on the current weight estimation model based on parameter replacement, weighted update, or learning rate adjustment without resetting the network structure and historical parameters of the weight estimation model to obtain an updated weight estimation model;

[0089] The accuracy evaluation sub-module is used to perform estimation calculations on the updated weight estimation model for the old and new sample sets respectively, and obtain the corresponding mean error and standard deviation.

[0090] The stability feedback sub-module is used to construct an error change curve based on the mean error and standard deviation, so as to judge whether the updated weight estimation model meets the preset stability standard. When the judgment result is not satisfied, a training adjustment signal is output, which is used to correct the mapping response mode of the subsequent updated parameter vector or adjust the training frequency.

[0091] In the embodiment of the present invention, to achieve the continuous optimization and stable update of the weight estimation model, the system sets up a model update module, which is composed of a directional training sub-module, an accuracy evaluation sub-module and a stability feedback sub-module, and is used to gradually inject the updated parameter vector from the parameter generation module into the existing model, so as to form an incremental improvement path without resetting the structure.

[0092] The directional training sub-module receives the generated updated parameter vector and inputs it into the currently used weight estimation model. This model contains the existing structural parameters and historical training weights. To avoid the computational overhead and loss of stability caused by the overall reconstruction of the model, this sub-module does not adopt the model re-initialization strategy, but performs fine-grained parameter adjustment through a preset update mechanism. Specifically, if the amplitude of the updated parameter vector is small, it is directly applied in the form of parameter replacement; if the parameter changes are concentrated and the magnitude is moderate, a weighted update method is adopted; in the unstable training stage, the learning rate is dynamically adjusted to achieve adaptive control of the parameter adjustment speed. Through the above strategy, the model gradually introduces adaptive learning for new samples while maintaining the original structural ability, effectively extending the model life cycle and improving the learning efficiency.

[0093] The accuracy evaluation sub-module is used to objectively measure the performance of the model before and after the update. This sub-module introduces two types of sample sets, namely the historical sample set before the update and the to-be-learned sample set after the update. The system performs weight estimation on the updated weight estimation model in the two sample sets respectively, and calculates the mean error and standard deviation of its estimation error, so as to judge the generalization performance and accuracy change trend of the model after the update. In this way, the adaptation degree of the model to the old and new samples can be quantified, and the effectiveness of the update behavior can be assisted in judgment.

[0094] The stability feedback sub-module constructs an error change trend curve based on the aforementioned error mean and standard deviation data. In this trend curve, by observing the change direction and amplitude of the estimated error over multiple consecutive training cycles, it can be determined whether the current model has entered an oscillating state or is in a stable convergence state. If the slope of the error curve is greater than the positive threshold, or the error standard deviation continuously increases, the system determines that the model is in an unstable interval and immediately triggers a training adjustment signal. This signal will be fed back to the parameter generation module to adjust the response range of the parameter mapping function or reduce the training frequency until the error returns to a stable state.

[0095] This module constructs a highly stable dynamic training mechanism through "strongly controllable directional update + dual-sample accuracy evaluation + closed-loop stability feedback", which not only maintains the continuity of the model structure but also enhances the system's adaptive update ability when dealing with new object or scene data.

[0096] Among them, the directional training sub-module is used to input the updated parameter vector into the current weight estimation model, and without resetting the network structure and historical parameters of the weight estimation model, perform incremental training on the current weight estimation model based on parameter replacement, weighted update, or learning rate adjustment to obtain an updated weight estimation model, specifically including:

[0097] This sub-module is the core component of the model online update mechanism, which is used to achieve local directional adjustment of model parameters without destroying the original model structure. Compared with the traditional model system that requires complete retraining, this module can effectively reduce update latency and computational resource consumption, and is particularly suitable for industrial control scenarios where learning occurs while running.

[0098] In specific implementation, the system first receives the updated parameter vector output by the parameter generation module. This vector consists of multiple parameter increment terms corresponding to the model feature dimensions, has undergone non-linear mapping processing, and contains directional and adjustment intensity information. After the system matches this vector with the weight parameters of the current model, it selects one or more update methods according to the set strategy:

[0099] First, the parameter replacement strategy: If the change amplitude of the updated parameter in a certain feature dimension is within the set range (such as less than 5% of the original weight value), the system can directly use this value to replace the original parameter to achieve rapid iteration.

[0100] Second, the weighted update strategy: For medium-strength updates, the system performs weighted fusion of the original weight and the updated parameter. The recommended method is to set the main weight coefficient (such as 0.7) to be assigned to the historical weight to maintain model continuity, while the updated parameter is multiplied by the secondary weight coefficient (such as 0.3) for fine-tuning the direction. This method is particularly suitable for adjustments when the model enters the steady state after long-term operation.

[0101] Third, learning rate adjustment strategy: If the updated parameters change drastically in multiple feature dimensions or the system determines that the current training is in an unstable state, the overall learning rate can be decreased, or a dynamic step size function can be set to make the adjustment of the model weights tend to be smooth and avoid the oscillations caused by too fast updates.

[0102] For example, in a dynamic weighing system on a packaging line, when a new batch of lightweight materials (such as foam items) appears, the force characteristics distribution on the sensor is different from the historical samples. Through this module, the system can quickly replace or slightly correct only the model weights corresponding to the relevant sensor channels without full model regression training, thus completing the model update within 5 seconds and ensuring continuous weighing accuracy.

[0103] Through the above method, the targeted training sub-module achieves the model update goal of "low intrusion, low calculation, and high adaptability", which is suitable for the actual industrial environment with frequent fluctuations of sensing signals and strong object diversity, and improves the overall response ability and long-term operation stability of the system.

[0104] Among them, the accuracy evaluation sub-module is used to perform estimation calculations on the updated weight estimation model on the old and new sample sets respectively to obtain the corresponding error means and standard deviations, specifically including:

[0105] This sub-module is used to objectively quantify and analyze the model update effect and provide a judgment basis for whether to trigger feedback control in the future. In traditional model training, accuracy evaluation is often only statically tested on the training set or validation set and cannot reflect the adaptability of the model to new data in a timely manner. This module introduces a dual-set error comparison mechanism to dynamically evaluate the change trend of the model performance.

[0106] In the implementation process, the system first constructs two sample sets:

[0107] The first set is the historical label sample set, which contains the analysis data known to the model and obtained through manual calibration or high-confidence learning and their corresponding true weight labels;

[0108] The second set is the sample set to be learned, which contains new samples recently marked as to be learned and updated through parameter generation and training, and is used to evaluate the coverage ability of the model for the new feature space.

[0109] The updated weight estimation model predicts each piece of data in the two sample sets in turn to obtain the estimated output value of each sample. Subsequently, the system compares the estimated value with the actual mass value, calculates the prediction error of each sample, and then statistically calculates the error mean and error standard deviation in the two sets respectively.

[0110] Among them:

[0111] The mean error is used to evaluate the degree of prediction deviation. If this value is large, it indicates that the model has a systematic bias;

[0112] The standard deviation of the error reflects the degree of fluctuation of the model output within the set and is used to evaluate the prediction stability.

[0113] To enhance the ability to judge trend changes, the system can adopt a sliding window mechanism to sort and compare the mean errors in consecutive update cycles. For example, calculate the direction of error change in the last three updates to determine whether the error is gradually decreasing (converging) or fluctuating repeatedly (unstable). In addition, a dynamic baseline can be set to evaluate whether there is a degradation phenomenon based on the set historical best accuracy value.

[0114] For example, if after a certain model update, the increase in the mean error in the historical sample set exceeds 10%, and the standard deviation of the error in the set to be learned also increases significantly, the system can judge that the current model update effect is poor and should trigger the stability feedback mechanism to limit further updates.

[0115] Through this module, the system can monitor the change trajectory of the model accuracy in real time without external manual intervention and provide a basis for training adjustment for subsequent modules, which is an important basic module to ensure the safe self-update of the model.

[0116] In a preferred embodiment of the present invention, the feature distance calculation sub-module includes:

[0117] A feature vector extraction unit for extracting the feature vectors that make up the sample features from the analysis data, where the feature vectors are composed of multiple sensor signal values after normalization processing and are used to reflect the distribution characteristics of the sample in multiple feature dimensions;

[0118] A weight factor determination unit for statistically analyzing each feature dimension in the historical label samples, and determining that the weight value corresponding to the relatively stable feature is higher by evaluating the degree of numerical fluctuation of each feature dimension within a preset first time window, while the weight value of the feature with large volatility is lower, so as to generate a set of weighting factors for distance calculation;

[0119] A distance calculation unit for calculating the difference between the feature vectors of the current analysis data and the multiple feature vectors of the historical label samples dimension by dimension, and accumulating and summing the products of each difference and the corresponding weighting factor to obtain a feature distance value representing the similarity degree of the samples.

[0120] In the embodiment of the present invention, to realize the evaluation of the feature similarity between the analysis data and the historical label samples, the feature distance calculation sub-module is divided into a feature vector extraction unit, a weight factor determination unit and a distance calculation unit, which are used to form a weighted similarity calculation process that conforms to the multi-dimensional signal feature distribution.

[0121] The feature vector extraction unit is responsible for extracting the multi-dimensional feature vectors that constitute the sample features from the analysis data output by the data preprocessing module. Each feature dimension corresponds to a normalized sensor signal channel value, and the overall feature vector reflects the comprehensive state of the sample in each physical measurement dimension. These vectors can be used to represent the relative distribution position of the sample in space and are the basic data structure for subsequent similarity calculation.

[0122] The weight factor determination unit is used to evaluate the importance of each feature dimension and assign a weighting coefficient. Specifically, within a preset first time window, this unit performs a fluctuation amplitude analysis on each feature dimension. If a certain dimension shows a small variance or standard deviation within this window, it is regarded as a stable dimension and is assigned a higher calculation weight; conversely, dimensions with large fluctuations will have their influence in the calculation weakened. This way ensures that when calculating the feature distance, the system pays more attention to the feature components that have a substantial discriminatory ability for the determination result.

[0123] The distance calculation unit takes the feature vector of the current sample as a reference and calculates it one by one with the feature vectors in the historical label sample library. After performing a difference process on each corresponding dimension, it multiplies by the weight factor, and sums up the weighted differences of all dimensions to obtain a weighted distance value used to measure the similarity degree between the current sample and the historical sample. The obtained feature distance not only reflects the overall proximity between samples but also takes into account the differences in the discriminatory ability between feature dimensions, which helps the subsequent accurate judgment of the confidence score.

[0124] This sub-module can effectively avoid the interference caused by low-value features to the similarity judgment result in the multi-type sample mixing scenario, and enhance the robustness and generalization of sample evaluation.

[0125] Among them, the feature vector extraction unit is used to extract the feature vectors that constitute the sample features from the analysis data, where the feature vectors are composed of multiple normalized sensor signal values and are used to reflect the distribution characteristics of the sample in multiple feature dimensions, specifically including:

[0126] The function of this unit is to convert the preprocessed time series signal into a structured multi-dimensional data representation, enabling the sample to participate in the feature similarity analysis in the subsequent process. The analysis data is output by the data preprocessing module and has completed basic processing such as extreme value elimination, signal smoothing, and normalization. The sensor signal values include voltage data collected by pressure sensors installed at different positions on the weighing platform, and these data can be uniformly mapped to the [0,1] interval after normalization.

[0127] During the implementation, the system treats each piece of analysis data as a sampling segment and extracts its normalized response value at that time slice according to the sensor channel dimension. Assuming that the system is equipped with N pressure sensors, the feature vector corresponding to each sample will contain N components, representing the signal response strength of each sensor channel at the current sample moment.

[0128] For example, in a logistics weighing platform, suppose the system is equipped with four pressure sensors. When an object is placed slightly to the left of the center of the platform, the output voltage values ​​of the two sensors on the left are significantly higher than those of the two on the right. After normalization, the feature vector of the sample may appear in the form of "0.82, 0.77, 0.35, 0.29", representing the relative proportional characteristics of the pressure distribution on the platform. This structured vector can be used as the input for subsequent feature distance calculation.

[0129] This unit extracts the response ratio of samples on different sensors to form a feature representation in a unified format. It has the advantages of low dimension, dimensionlessness, and clear structure. It is the basic condition for performing feature similarity judgment and pattern recognition.

[0130] The weight factor determination unit is used to perform statistical analysis on each feature dimension in the historical label sample, and to determine that the weight value corresponding to the relatively stable feature is higher, and the weight value of the feature with large volatility is lower by evaluating the value fluctuation degree of each feature dimension within the preset first time window, so as to generate a set of weighting factors for distance calculation, specifically including:

[0131] The core function of this unit is to identify which feature dimensions are more representative and discriminative in historical samples, and give them higher participation weights, thereby improving the discriminative ability of feature distance calculation. In the prior art, all feature dimensions are often treated as equivalent, resulting in unnecessary interference in distance calculation for some features that are prone to fluctuation but have low information value. This unit dynamically allocates its importance in the overall calculation by statistically evaluating the fluctuation of each dimension.

[0132] The system sets a fixed-length time window as the analysis interval, such as a set of the most recent 200 labeled samples. Within this window, the system calculates the numerical fluctuation amplitude of each feature dimension. The volatility can be characterized by "the difference between the maximum and minimum values" or "standard deviation". To simplify the implementation, it is recommended to use the sliding standard deviation as the volatility criterion.

[0133] For feature dimensions with smaller standard deviations, it means that their values ​​in different samples are relatively concentrated and the change trend is stable, which may reflect basic characteristics such as physical structure or stable posture, and have stronger consistency and discrimination ability; on the contrary, feature dimensions with larger standard deviations often indicate that they are affected by occasional changes in samples or system noise, and their participation intensity should be appropriately reduced.

[0134] Based on the above volatility results, the system generates a weighted factor in reverse: that is, the smaller the volatility, the higher the weight, and the larger the volatility, the lower the weight. To avoid the overall calculation stability being affected by too large or too small weights in some dimensions, it is recommended to normalize all the original weight values so that their sum is 1.

[0135] For example, in a four-channel system, after analysis, it is found that the standard deviations of channels 1 and 2 are relatively small, 0.04 and 0.06 respectively, while the standard deviations of channels 3 and 4 are relatively large, 0.11 and 0.15 respectively. Then the system can assign weights to the four channels as: 0.35, 0.30, 0.20, 0.15. This way of weight allocation emphasizes the importance of low-volatility channels for sample classification, helps the system focus on the main features, and suppresses noise interference.

[0136] This unit effectively constructs a mapping mechanism from "statistical characteristics" to "weight factors", provides a basis for discriminative priority sorting for subsequent distance calculation, and improves the accuracy of similarity judgment in the actual weighing application of the system.

[0137] Among them, the distance calculation unit is used to perform a dimension-by-dimension difference calculation on the feature vector of the current analysis data and multiple feature vectors of historical label samples, and accumulate the sum after multiplying each difference by the corresponding weighted factor to obtain a feature distance value representing the similarity degree of the samples, specifically including:

[0138] The function of this unit is to quantitatively express the differences between two feature vectors in multiple feature dimensions. The traditional Euclidean distance calculation method directly adds the squares of the differences in each dimension as the similarity basis, ignoring the importance differences of each feature dimension, resulting in the distance result being too sensitive to local perturbations. To solve this problem, the present invention introduces a weighted distance mechanism, combines the previously generated weight factors in the calculation, and adjusts the contribution of the differences, making the distance result closer to the similarity degree between real samples.

[0139] In specific implementation, the system obtains its feature vector (denoted as A) from the current sample, and sequentially reads the comparison vectors (denoted as B) from the historical label sample set. For each corresponding feature dimension in A and B, calculate their absolute differences, and multiply them by the weighted factor corresponding to this dimension. Then sum up all the weighted differences item by item to obtain the final weighted distance value.

[0140] For example, if the feature vector of the current sample is "0.82, 0.77, 0.35, 0.29", the sample to be compared is "0.80, 0.75, 0.42, 0.34", and the corresponding dimensional weights are "0.35, 0.30, 0.20, 0.15". Then the system calculates the difference for each dimension as: 0.02, 0.02, 0.07, 0.05, multiplies each by the weight to get the weighted result, and finally sums them up to obtain the final distance value.

[0141] It should be noted that to improve the system calculation efficiency, all difference calculations and weight operations can be completed within the range of low-precision floating-point numbers, which is suitable for fast matching in real-time scenarios.

[0142] This unit reasonably weights the differences of samples in the multi-dimensional sensing feature space, avoiding deviations in the distance results caused by irrelevant features, thus more accurately judging the similarity between the current analysis data and the historical labeled samples, and ultimately improving the accuracy of selecting samples to be learned in active learning and the intelligence of the system response.

[0143] In a preferred embodiment of the present invention, the time correlation processing sub-module includes:

[0144] A time information extraction unit, which is used to extract the timestamp information of the sample from the analysis data and count the number of weighing records within the currently preset second time window, so as to obtain the weighing frequency per unit time;

[0145] A time density calculation unit, which is used to compare the timestamp information of the sample with the current system time, calculate the time interval length of the sample, and perform weighted combination after normalizing the reciprocal of the time interval length and the weighing frequency per unit time respectively to calculate the time density value. The shorter the time interval length and the higher the unit frequency, the larger the obtained time density value;

[0146] A weighted fusion unit, which is used to jointly evaluate the feature distance value and the time density value, and by setting an adjustment parameter that changes with the time density value, the influence degree of the sample collected during the high-density period on the feature distance value is enhanced, so as to form a composite evaluation factor representing the comprehensive evaluation level of the sample.

[0147] In the embodiment of the present invention, to strengthen the time sensitivity evaluation mechanism of the sample and improve the response ability of the sample confidence score to the time series characteristics, the system sets a time correlation processing sub-module, which is composed of a time information extraction unit, a time density calculation unit and a weighted fusion unit, and is used to perform dynamic weighting processing on the sample by combining time information.

[0148] The time information extraction unit extracts the timestamp information of each sample from the analysis data and counts the number of weighing records within the second time window set by the current system. The setting of this window ensures that the system can adaptively obtain the data generation frequency per unit time, that is, the weighing frequency per unit time, under the background of changing sampling density. This frequency value can reflect the operation density within the time period to which the sample belongs and is the key basis for calculating the timeliness of the sample.

[0149] The time density calculation unit compares the sample timestamp with the current system time, calculates the length of the time interval, and combines this interval length with the per-unit time frequency to calculate the time density value. The calculation result reflects the "freshness" of the sample from the current time: the newer the sample and the higher its occurrence frequency, the greater the generated time density value. This density value is used to measure the importance of the time period in which the sample is located.

[0150] The weighted fusion unit fuses and processes the aforementioned time density value with the feature distance value of the sample. During the fusion process, the system sets an adjustment parameter that changes with the time density. This parameter increases the weight of the current sample in the comprehensive evaluation factor as the time density increases, ensuring that samples with a high occurrence frequency in the dataset and close to the current moment receive more attention from the model. The finally formed composite evaluation factor integrates the spatial similarity and time timeliness of the sample and is the core input item for subsequent confidence calculation.

[0151] This module improves the system's response ability to fresh data and effectively avoids the dominance of outdated data in the training process, which is beneficial to constructing a weighing discrimination system with real-time update capabilities.

[0152] Among them, the time density calculation unit is used to compare the timestamp information of the sample with the current system time to calculate the length of the sample time interval; and based on the reciprocal of this time interval length and the weighing frequency per unit time, they are respectively normalized and then weighted and combined to obtain the time density value. The shorter the time interval and the higher the per-unit time frequency, the greater the obtained time density value, specifically including:

[0153] This unit aims to measure the "currency" and "activity" of the sample in the time dimension and is used to assist in judging whether the sample has the timeliness value to participate in model update preferentially. In existing active learning technologies, sample confidence is usually judged only based on feature space similarity, ignoring the time factor may cause the model to deviate from the current data distribution, resulting in a lag in prediction ability. The present invention constructs a "time density" index to comprehensively evaluate the timeliness of the sample and the current sampling activity of the system, enabling the system to more sensitively respond to the data evolution trend.

[0154] The processing flow of this unit includes the following steps:

[0155] In the first step, the system extracts the timestamp information of the current sample and compares it with the current system time to obtain the length of the time interval. This time interval can be measured in seconds or minutes, reflecting the time distance since the sample was collected. The shorter the interval, the "fresher" the sample, and it should be considered in terms of priority.

[0156] In the second step, the system obtains the weighing frequency per unit time through the time information extraction unit. This frequency is defined as the number of actual weighing records completed by the system within a set time window (such as the past 10 minutes) divided by the window length. The higher the frequency, the denser the recent data generation of the system, indicating that the current is an active update period for the model.

[0157] In the third step, the system performs normalization processing on the above two indicators respectively. The normalization range is recommended to be fixed from 0 to 1, where:

[0158] The longer the time interval length, the lower the normalized value, and the "maximum set time interval" can be used as the normalization upper limit;

[0159] The higher the frequency per unit time, the higher the normalized value, and the historical maximum frequency or empirical maximum value can be set as the upper limit standard.

[0160] In the fourth step, the system inversely weights and superimposes the normalized time interval reciprocal value and the frequency value per unit time in proportion to generate the final time density value. The weighting ratio can be set by the system according to experience. For example, the weight factors are 0.4 and 0.6 respectively, indicating that in the time density judgment, the real-time nature is slightly lower than the activity.

[0161] For example, assume that the timestamp of a certain analysis sample is 120 seconds from the current time, and the maximum reference interval is 300 seconds, then the reciprocal score after normalization is approximately 0.6; the frequency per unit time is 15 times / minute, the reference upper limit is 20 times / minute, and the normalized value is 0.75. Then the final time density value is 0.6×0.4 + 0.75×0.6 = 0.69 (for illustration only).

[0162] The finally output time density value can be used as the weighting basis for subsequent "feature distance value fusion" and "confidence score" links, and the system can dynamically increase or decrease the participation priority of the sample according to this value. For example, when the time density value is higher than the set threshold (such as 0.7), the system can increase the priority of this sample in the active learning screening process, thus accelerating the adaptation to the fresh data pattern.

[0163] This unit constructs a timeliness judgment index associated with the task status by combining two dimensions of time interval and data frequency, avoiding the over-reliance of traditional systems on "old samples", and significantly improving the training efficiency and response ability of the model in a dynamic data environment.

[0164] Among them, the weighted fusion unit is used to jointly evaluate the feature distance value and the time density value. By setting the adjustment parameters that change with the time density value, the influence of the samples collected in the high-density period on the feature distance value is enhanced, thereby forming a composite evaluation factor for representing the comprehensive evaluation level of the sample, which specifically includes:

[0165] This unit is responsible for integrating "spatial feature similarity" and "temporal importance" into a single evaluation index, and is a key node in the active learning sample selection strategy. In the existing technology, most methods only focus on spatial distance (such as Euclidean distance and Mahalanobis distance), ignoring the important impact of temporal density on data timeliness, which can easily lead to the model collecting structurally "strange" but outdated data. This unit dynamically adjusts the composition weights of the sample evaluation criteria by setting an adjustable fusion strategy.

[0166] During implementation, the system receives the feature distance value (reflecting the degree of difference between the current sample and the historical label sample) output by the feature distance calculation submodule, and receives the time density value output by the time density calculation unit. The system integrates the two data through a fusion factor, which is used to control the degree of intervention of time density on the total evaluation factor. The value range of the fusion factor can be set by the system adaptively, or adjusted based on the fluctuation of the system's historical data.

[0167] Specifically, the system can set the fusion factor to a function that grows linearly or nonlinearly with the time density value. When the time density value is in a high range (such as above 0.8), the corresponding fusion factor is also high. At this time, the system will lower the suppression threshold for feature distance, making it easier for samples to enter the learning judgment area; on the contrary, when the time density value is low, the fusion factor tends to be conservative, and the system relies more strictly on the feature space distance for judgment.

[0168] For example, in actual applications, if the feature distance of sample A is 0.65 (medium to high), but the time density value is 0.92 (indicating that the sample is very new and appears frequently), the system will improve its overall score according to the fusion strategy, such as by increasing weights so that its comprehensive evaluation factor is greater than a set threshold, thereby being marked as a high-value learning sample.

[0169] Through the dynamic fusion mechanism of this unit, the system realizes the transformation from "spatial difference" to "time-driven", which can more accurately identify potential samples that are different in structure and have time advantages, and effectively improve the response speed and effectiveness of model updates.

[0170] In a preferred embodiment of the present invention, the confidence assessment submodule includes:

[0171] A scoring parameter loading unit for loading a preset confidence scoring model, which is constructed based on multiple influencing factors, including a composite evaluation factor, a time density value corresponding to a sample, and the local fluctuation amplitude of an original signal sequence, and is used to comprehensively evaluate the confidence level of the current sample;

[0172] A scoring calculation unit for inputting multiple influencing factors into the confidence scoring model and calculating the confidence score value of the current sample according to the parameter weights corresponding to each factor;

[0173] A threshold comparison unit for comparing the confidence score value of the current sample with a preset confidence threshold. When the score is lower than the threshold, it outputs a to-be-learned marker information and classifies the sample into the to-be-learned data set; if the score is higher than the threshold, the current sample is ignored for model training.

[0174] In the embodiment of the present invention, to implement the sample confidence scoring based on multi-source information and accordingly determine whether to introduce an active learning process, the system designs a confidence evaluation sub-module, which is composed of a scoring parameter loading unit, a scoring calculation unit, and a threshold comparison unit. This module jointly models three types of factors, namely features, time, and signal stability, and is used to form a dynamically adjusted sample credibility evaluation system.

[0175] The scoring parameter loading unit is responsible for loading the confidence scoring model pre-trained by the system. This scoring model is constructed based on multi-factor regression or fitting methods, and the input factors include a composite evaluation factor, a time density value, and the local fluctuation amplitude of the original signal sequence. The composite evaluation factor is obtained from the previous processing and represents the double deviation degree of the sample and the label data in the feature space and the time dimension; the time density value represents the acquisition timeliness of the sample; the signal fluctuation amplitude reflects the stability of the sensor signal of the current sample. Taking these three items as inputs and acting on the scoring model together can significantly improve the representativeness of the scoring results.

[0176] The scoring calculation unit is used to input the above multiple influencing factors into the confidence scoring model. The model performs weighted processing on each factor according to the weight coefficients learned during training and outputs a numerical confidence scoring result. This scoring value fluctuates between 0 and 1, and the lower the value, the more learning value the sample has.

[0177] The threshold comparison unit is used to receive the scoring calculation result and compare it with the confidence threshold set in the system. When the scoring value is lower than the threshold, it is determined that the current sample has insufficient credibility. The system marks it as "data to be learned" and writes the marking result into the training sample management list; if the scoring value is higher than the threshold, it indicates that the sample features have been covered by the model, and the current sample is not involved in the update.

[0178] As a key component for active learning decision-making, this module can effectively control the input quality of training data, improve the sample acquisition efficiency and the model convergence speed, reduce the risk of overfitting, and provide support for the system to build an intelligent judgment and automatic labeling mechanism.

[0179] Among them, the confidence score model is constructed based on multiple influencing factors, including the composite evaluation factor, the time density value corresponding to the sample, and the local fluctuation amplitude of the original signal sequence, and is used to comprehensively evaluate the confidence level of the current sample. Specifically, it includes:

[0180] This scoring model aims to provide a quantitative evaluation method for sample credibility based on multi-factor fusion, and is used to assist in judging whether the current sample should be included in the active learning process. In the existing technology, simple thresholds or fixed rules are often used to screen samples, lacking comprehensive consideration of multi-dimensional indicators, and prone to screening errors or overfitting risks. By introducing multiple evaluation dimensions such as space, time, and signal fluctuation, this scoring model improves the comprehensiveness and robustness of sample evaluation.

[0181] The three factors input into the model are respectively:

[0182] First, the composite evaluation factor, output by the aforementioned weighted fusion unit, integrates two dimensions of feature distance and time density, representing the overall performance of the sample in terms of structural difference and timeliness;

[0183] Second, the time density value, which reflects the update priority of this sample under the current system state;

[0184] Third, the local fluctuation amplitude, which comes from the original signal sequence and is used to measure the intensity of the voltage signal of this sample per unit time. The sliding window standard deviation or the absolute change rate is commonly used for measurement. The larger the value, the stronger the signal fluctuation, which may be an abnormal state.

[0185] The scoring model can be constructed using an empirical regression model, a rule engine model, or a simple weighted linear combination strategy. It is recommended to use a weight superposition model, and the weight coefficients of each factor are obtained through training with historical labeled samples. For example, during the research stage, the weight of the composite evaluation factor can be set to 0.5, the weight of the time density value to 0.3, and the weight of the local fluctuation amplitude to 0.2. The output result of the scoring model is a real number between 0 and 1, which is used to represent the credibility level of this sample.

[0186] In actual operation, the system can set multiple threshold intervals. For example, 0–0.4 indicates low confidence and requires active learning; 0.4–0.7 indicates observable retention; above 0.7 indicates high credibility and does not need to participate in training. For example, the composite evaluation factor of sample B is 0.65, the time density value is 0.91, and the local fluctuation is 0.10 (indicating relatively stable). After being processed by the scoring model, the score is 0.38, and the system thus marks it as a sample to be learned.

[0187] By introducing this scoring model, the system can accurately judge the sample value under a dynamic and multi-source index system, thereby improving the accuracy and reliability of the active learning strategy and reducing the training costs brought by model overfitting and sample redundancy.

[0188] In a preferred embodiment of the present invention, the error contribution analysis sub-module includes:

[0189] An eigenvalue analysis unit, configured to calculate the proportion information of each feature in the entire feature vector based on the feature vector of the data to be learned and its corresponding estimated deviation value, and based on the relative magnitudes of the eigenvalues of each feature dimension.

[0190] A correlation adjustment unit, configured to determine the relative contribution degree of each feature dimension to the total deviation according to the proportion information and the estimated deviation value, and generate an error contribution value. Among them, a reduction coefficient is set for the feature dimension with a large eigenvalue change range but low stability in training to reduce its dominant role in parameter generation.

[0191] In the embodiment of the present invention, to solve the problems of unclear parameter adjustment granularity and unclear dimension contribution in model update, the system sets an error contribution analysis sub-module. This module analyzes the structural relationship between the feature vector of the data to be learned and its estimated deviation value, extracts the feature contributions dimension by dimension, and forms the basic data for driving model optimization.

[0192] The eigenvalue analysis unit first receives the feature vector marked as the data to be learned, and synchronously receives the estimated deviation value output by the deviation calculation module. This feature vector is a multi-dimensional structure, and each dimension represents a normalized sensor channel value, which is used to reflect the state information of the sample in different measurement dimensions. To clarify the relative weight of each feature dimension, this unit normalizes each eigenvalue into a percentage form to form feature proportion information, which is used to represent the proportion of this dimension in the entire feature structure.

[0193] The correlation adjustment unit calculates the relative contribution degree of each feature dimension to the overall deviation based on the interaction between the above-mentioned feature proportion information and the estimation deviation value. The system is set that if the proportion of the feature value of a certain dimension is high and consistent with the deviation change direction, it is determined as the deviation-dominant feature, otherwise it is a non-dominant dimension. For the dimensions with poor stability in historical training, that is, the feature dimensions with high error volatility during the training process, the system will set a reduction coefficient for them to weaken their influence on the parameter adjustment direction and prevent them from causing unstable update paths.

[0194] The finally output error contribution values form a one-dimensional vector sequence corresponding to the feature vector dimensions, providing a reference for the subsequent parameter mapping module with clear structure, reasonable proportion, and strong robustness of error information.

[0195] Through the above structure, the system realizes the quantitative explanation of each feature dimension in the deviation formation mechanism, avoids blindly equally distributing the error to all feature dimensions, and significantly improves the accuracy control ability of model update.

[0196] Among them, the correlation adjustment unit is used to determine the relative contribution degree of each feature dimension to the total deviation according to the proportion information and the estimation deviation value, and generate the error contribution value. Among them, a reduction coefficient is set for the feature dimension with a large change range of the feature value but low stability during training to reduce its dominant role in parameter generation, specifically including:

[0197] This unit is used to reasonably distribute the error deviation among the feature dimensions, so as to avoid a certain feature obtaining too high a weight in model update due to accidental fluctuations. In traditional technologies, the error is often simply equally averaged or apportioned according to the proportion of the number of features. This processing method ignores the actual influence of each dimension feature in error formation and may cause risks of model parameter misadjustment or overfitting. This unit combines the proportion information of each dimension in the feature vector with the estimation deviation value of the current sample to generate a more targeted error contribution structure.

[0198] During specific implementation, the system first clarifies the relative weight of each dimension in the feature structure based on the proportion information output by the feature value analysis unit. For example, in a four-dimensional feature vector, if the normalized value of the first dimension is 0.82 and the total vector modulus is 2.0, its proportion is about 41%. This proportion represents the representativeness of this dimension to the entire sample state.

[0199] Subsequently, the system combines the proportion of this dimension with the current estimated deviation value. To achieve linear correspondence or non-linear enhancement, a scaling factor can be set to multiply the proportion by the deviation value to obtain a preliminary error contribution value. To prevent high-fluctuation features from misleading the training direction, the system further introduces a "reduction coefficient", which is generated based on the error volatility or feature sensitivity index recorded in historical training. If a certain feature dimension shows high fluctuations or abnormal repetitions in the past few rounds of training, a smaller reduction coefficient (such as below 0.6) is attached to it to weaken its dominant ability to explain the current error.

[0200] For example, the proportion of the fourth dimension of sample A is 18%, the current estimated deviation value is 0.3, and historical data shows that the standard deviation of this dimension in the past 5 rounds of training is higher than other dimensions. Then the system can set the reduction coefficient of this dimension to 0.5. Finally, the error contribution value of this dimension is 0.18×0.3×0.5 = 0.027 (for numerical illustration only), which is significantly lower than other dimensions, thus controlling its participation intensity in the subsequent parameter mapping process.

[0201] Through the error allocation and reduction mechanism of this unit, the system can ensure that representative features obtain the dominant update status while limiting the perturbation risk brought by unstable dimensions, significantly improving the robustness of the model update path.

[0202] In a preferred embodiment of the present invention, the parameter mapping sub-module includes:

[0203] A contribution level discrimination unit, which is used to divide the error contribution value into multiple levels according to the set contribution grading standard based on the error contribution value of each feature dimension, and generate a contribution level division result. Among them, the error contribution value of a higher level will obtain a larger parameter adjustment range;

[0204] A non-linear mapping strategy unit, which is used to perform non-linear mapping response processing on the error contribution value through a set of mapping functions according to the contribution level division result, and generate the mapped error contribution value. Among them, a compression mapping is used for the error contribution value of a lower level to suppress the model changes caused by weak perturbations, and an enhancement mapping is used for the error contribution value of a higher level to accelerate model convergence.

[0205] In the embodiment of the present invention, to solve the problem that the error contribution value cannot be directly used for model update, the system designs a parameter mapping sub-module to perform non-linear transformation on the original error contribution value to improve the flexibility and hierarchical ability of the model update response. This module includes a contribution level discrimination unit and a non-linear mapping strategy unit, which complete the logical mapping process between the error and the update intensity.

[0206] The contribution level discrimination unit is responsible for classifying and identifying the error contribution values on each feature dimension. The system divides the contribution value range into multiple levels according to the set contribution classification criteria, such as low level, medium level, and high level. The specific division can refer to the mean and standard deviation of the contribution value distribution in the current training round to set the boundaries, thereby dynamically adjusting the decision thresholds corresponding to each level. The division results will be output as contribution level labels to guide the formulation of downstream response strategies.

[0207] According to the contribution level division results, the non-linear mapping strategy unit calls the corresponding mapping functions to perform non-linear response processing on the error contribution values of each dimension. For feature dimensions with relatively low contribution values, a compression mapping strategy is adopted, that is, the output value after mapping changes slowly and approaches zero to suppress the influence of weak perturbations on the model weights; for medium-level contribution values, a slightly enhanced but restricted response strategy is adopted; for high-level contribution values, an enhanced response strategy is used to make the mapping output rise rapidly to amplify the leading role of key dimensions in model updates. Each mapping strategy can be constructed based on piecewise functions, smooth curves, or exponential responses, but their core objectives are the same: to make the updated parameters have distinguishability, responsiveness, and stability.

[0208] This module converts the intensity of model parameter updates from "directly driven by errors" to "driven after policy adjustment", which not only enhances the adaptability and robustness of the system but also provides a prerequisite for subsequent stability control and training frequency optimization.

[0209] Among them, the contribution level discrimination unit is used to divide the error contribution values into multiple levels according to the set contribution classification criteria based on the error contribution values of each feature dimension, generating contribution level division results. Among them, higher-level error contribution values will obtain a larger range of parameter adjustment, specifically including:

[0210] This unit is used to map the original error contribution values to level labels, providing a decision basis for subsequent mapping function selection and parameter adjustment amplitude. Different from directly using the error values for gradient updates in traditional models, the present invention realizes the segmented processing of error responses through a grading strategy, avoiding extreme behaviors such as excessive error amplification or ignoring weak signals.

[0211] In a specific implementation, the system first receives the error contribution value sequence output by the correlation adjustment unit, and each value corresponds to a feature dimension. The system sets a contribution classification criterion, which can adopt a fixed threshold segmentation or a distributed division method based on the dynamic distribution of samples.

[0212] The recommended method is: first obtain the set of error contribution values of all dimensions in the current sample and calculate their average value and distribution range. The system can divide the contribution values into three level intervals, for example:

[0213] Level 1 (low level): The contribution value is less than 80% of the overall average;

[0214] Level 2 (medium level): The contribution value is between the average value ±20%;

[0215] Level 3 (high level): The contribution value is higher than 120% of the average value.

[0216] Of course, the percentile interval can also be set according to system experience or historical training data. For example, the top 25% is set as the high level, the middle 50% is set as the medium level, and the rest is set as the low level. After division, each feature dimension is marked as low, medium or high level, providing a decision basis for the mapping strategy unit.

[0217] For example, in a six-dimensional feature, if the contribution value of the fifth dimension is 0.18 and the current average contribution value is 0.10, then this dimension will be marked as the high level, and this label will activate the enhanced mapping function in the next mapping, thereby enhancing its influence on model update.

[0218] This grading mechanism introduces a logical hierarchy into the error response, enabling the model update to no longer simply respond proportionally, but to have the ability to adjust according to the influence, optimizing the error adjustment efficiency and enhancing the response discrimination between feature dimensions.

[0219] Among them, the non-linear mapping strategy unit is used to perform non-linear mapping response processing on the error contribution value through a set of mapping functions according to the result of the contribution level division, generating the mapped error contribution value. Among them, the compression mapping is used for the low-level error contribution value to suppress the model changes caused by weak perturbations, and the enhanced mapping is used for the high-level error contribution value to accelerate the model convergence, specifically including:

[0220] The goal of this unit is to map the contribution level discrimination result into a parameter adjustment value with different response intensities, realizing the "non-linear amplification or suppression" of the error to the change of model parameters. Traditional gradient updates often adopt a linear strategy, resulting in the model still possibly generating unnecessary responses to small errors or insufficient responses to large errors, affecting the training efficiency. The present invention significantly improves the flexibility and stability of model training by establishing a hierarchical response mechanism.

[0221] Specifically, the system first reads the contribution level label of each dimension. For the low-level dimension, the system adopts the compression mapping method, that is, maps the original error contribution value into a smaller output value, and the compression ratio can be set between 0.3 - 0.5, ensuring that these dimensions only have a weak impact when participating in the weight adjustment, thus shielding the noise perturbation.

[0222] For the medium-level dimension, weak enhancement or near-linear mapping can be adopted, that is, the original value is slightly amplified, such as multiplied by 1.2 - 1.5 times, to maintain its basic response function.

[0223] For high-level dimensions, the system enables an enhanced mapping function that can amplify the original value to twice or even higher than the original contribution. The amplification factor can be adjusted according to the system's running state. This amplification operation can be completed through look-up table method, interpolation method, exponential function, or empirical curve fitting. It is recommended to use a smooth curve to avoid discontinuous training.

[0224] For example, if the contribution value of the 3rd dimension is 0.12, which belongs to the high level, and the system sets the enhancement factor to 2.0, then the output after mapping is 0.24; if the contribution value of the 5th dimension is 0.04, which belongs to the low level, and the compression factor is 0.4, then the output is 0.016. The results after the above mapping will be directly used to update the parameter generation module as the input for actual model parameter adjustment.

[0225] This unit implements the strategy of "fast response for important dimensions and slow update for minor dimensions" at the numerical level, enabling the model to focus more quickly on high-impact directions and avoiding interference from low-impact signals on the training path. It is one of the core mechanisms for achieving efficient and robust training updates.

[0226] More specifically, the calculation formula of the mapping function group is: ;

[0227] Among them, is the original error contribution value of the th feature dimension, output by the error contribution analysis sub-module. represents the proportion of the overall estimation error of this feature dimension in the current sample; is the error contribution value after mapping of the th feature dimension, which is an item in the updated parameter vector and is used for model parameter adjustment.

[0228] is the contribution level label of the th feature dimension, output by the contribution level discrimination unit, with values: 1 (low level), 2 (medium level), 3 (high level); is the error volatility of the th feature dimension in multiple historical training rounds, which is used to reflect the training stability of this feature. It can be calculated from the change range of the error mean in the recent N training rounds.

[0229] is the smoothing factor of the th feature dimension, which is used to control the response intensity of the non-linear mapping. The larger the value, the more cautious the system is about this dimension. This value can be set as a system parameter or dynamically adjusted by the stability feedback mechanism.

[0230] Among them, for (Low level): It uses a structure with a suppressed denominator, indicating that low-importance features need to compress their error responses. In the denominator, is the response damping term; the response value rapidly decreases as the historical fluctuation increases, forming a strategy of high fluctuation and low response.

[0231] For (Medium level): It introduces a ratio response term to achieve neutral amplification control. The numerator is , representing the historical error offset; the denominator is , forming a gain balance structure that is sensitive to system stability; it can achieve "when the model is stable → enhance the response, when the model is unstable → natural suppression".

[0232] For (High level): It adopts a non-linear enhancement structure to ensure that strongly influential features obtain sufficient model feedback. It includes a linear term and a non-linear amplification term , simulating an enhanced response; it improves the participation intensity of important dimensions, which is beneficial for rapid convergence.

[0233] Example illustration: Taking the second-dimensional feature as an example, if its original error contribution value , level , the corresponding historical error offset factor , smoothing factor , then its mapped result is: ; The system uses this value as the second item in the updated parameter vector for weight correction in the training process.

[0234] In a preferred embodiment of the present invention, the stability feedback sub-module includes:

[0235] An error trend extraction unit, which is used to construct an error change curve according to the changes of the error mean and standard deviation in the current multiple update cycles;

[0236] A stability judgment unit, which is used to judge whether the current model is in a stable state after update according to the slope and fluctuation frequency of the error change curve. If the fluctuation amplitude continues to rise or the slope exceeds the set critical value, it is judged that the stability standard is not met, otherwise it is met;

[0237] A feedback generation unit, which is used to generate a training adjustment signal when it is judged that the current model is in an unstable state after update. The training adjustment signal is used to trigger the parameter mapping sub-module to re-adjust the mapping range or reduce the model update frequency.

[0238] In the embodiments of the present invention, to ensure the stability and continuity of the weight estimation model during long-term operation, the system designs a stability feedback sub-module for dynamically adjusting the model update strategy based on the error fluctuation trend. This module includes an error trend extraction unit, a stability judgment unit, and a feedback generation unit, and constructs a cross-cycle monitoring and feedback control mechanism.

[0239] The error trend extraction unit receives the output data from the accuracy evaluation sub-module, including the error mean and standard deviation after each model update. The system records the error data within multiple consecutive update cycles and constructs a time series curve to reflect the model's response ability to data changes at different stages. This trend curve is used to capture behaviors such as unstable oscillations, increasing overfitting, or parameter drift during model training.

[0240] The stability judgment unit analyzes the slope and frequency of the error trend curve to determine the model state. When the curve slope is positive and does not converge for multiple consecutive cycles, or the standard deviation fluctuates frequently beyond the system-set limit, the system will determine that the current model update path is not stable and needs to trigger a self-correction mechanism.

[0241] When the feedback generation unit determines that the stability standard is not met, it immediately outputs a training adjustment signal, which can be transmitted to the parameter mapping sub-module or the model update module. The response mechanism includes control actions such as shrinking the parameter mapping response interval, extending the training cycle interval, and reducing the learning rate, which are used to weaken the impact of parameter updates on the model and thus restore the controllability and convergence of the system training state.

[0242] Through this module, the system can monitor the model state and manage the update rhythm without relying on manual intervention, effectively improving the stability and reliability of the system during long-term operation, and is particularly suitable for the dual requirements of accuracy and robustness of industrial automatic weighing equipment.

[0243] Among them, the feedback generation unit is used to generate a training adjustment signal when it is determined that the current model is in an unstable state after update. The training adjustment signal is used to trigger the parameter mapping sub-module to readjust the mapping range or reduce the model update frequency, specifically including:

[0244] The core function of this unit is to take timely adaptive adjustment measures for the unstable behaviors detected during the model training process to prevent the model from continuing to train along an unfavorable path or falling into a divergent state. In the prior art, common model training systems mostly rely on static learning rates and fixed update frequencies for parameter updates, lacking real-time monitoring and feedback response for training stability, and are extremely prone to update oscillations or performance degradation when the sample distribution changes, input perturbations increase, or model weights do not match. To solve this problem, the present invention introduces a dynamic control path into the training process through a feedback generation mechanism, enabling it to have self-stabilizing adjustment capabilities.

[0245] During the implementation process, the system first receives the judgment result output from the stability judgment unit. If it is determined that the currently updated weight estimation model does not meet the set stability standard, that is, the error change curve shows a continuous upward trend or the fluctuation frequency exceeds the tolerance range, the feedback generation logic is activated.

[0246] The system generates a set of training adjustment signals according to the instability degree. Each set of signals consists of two sub-strategies:

[0247] The first strategy is the mapping range reset instruction, which will act on the non-linear mapping strategy unit in the parameter mapping sub-module. The specific method is as follows: According to the perturbation degree of the high-level error contribution dimension in the model convergence behavior during this training, redefine the upper and lower limits of its corresponding mapping factor. For example, if the model shows parameter oscillation after high-level mapping, the system will reduce the magnification factor of this level of mapping from the original set value (such as 2.0) to within 1.3; if the error fluctuates intensively in the medium and low-level dimensions, the system can further shrink the slope of the overall mapping curve to make its response smoother at all levels. This reset process can be achieved by table lookup adjustment, function scaling or controlling the curve shape.

[0248] The second strategy is the update frequency reduction instruction, which will control the directional training sub-module in the model update module to reduce the rhythm density of weight adjustment. The specific methods include but are not limited to: extending the training period (for example, adjusting from updating once every 10 samples to updating once every 30 samples), freezing the parameter update channels of some feature dimensions, or setting a sample sliding window for sparse sampling in the training samples to ensure that the parameter perturbation source is reduced before the model is stably restored.

[0249] For example, in an automatic warehousing weighing system, if it is detected that the standard deviation of the error has increased from 0.08 to 0.14 in the recent three rounds of model updates, and the mean value continues to increase, the system will determine that the model has deviated from the stable interval. The feedback generation unit immediately issues an adjustment signal, instructing to reduce the high-level magnification ratio of the parameter mapping from 2.0 to 1.2, extend the update interval from 10 times to 25 times, and freeze the update permission of the feature dimension with the largest standard deviation fluctuation in the next round of training until the error trend drops.

[0250] This feedback mechanism has continuous responsiveness and multi-channel adjustment capabilities, can effectively buffer systematic anomalies in the training process, improve the stable adaptation ability of the model in a changing sample environment, and significantly reduce the risk of model performance degradation caused by overfitting or excessive adjustment.

[0251] With the above configuration, the feedback generation unit not only enhances the self-regulation ability of the system, but also endows the model training process with an elastic recovery mechanism, enabling the entire weighing control system to have the ability to operate stably for a long time, especially suitable for industrial weighing application scenarios driven by complex and multi-source signals.

[0252] The above is the preferred embodiment of the present invention. It should be pointed out that for those of ordinary skill in the art of the present technology, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A weighing device control system based on active learning, characterized in that, The system includes: A sample evaluation module for evaluating the confidence of samples in the analysis data, specifically including: calculating the feature distance between the historical labeled samples and the analysis data, and combining the timestamp and the weighing frequency to obtain a sample confidence score. When the sample confidence score is lower than the preset confidence threshold, the analysis data is marked as data to be learned; A deviation calculation module for estimating the feature vector of the data to be learned based on the current weight estimation model, outputting an estimation result, and calculating the difference between the estimation result and the actual result to obtain an estimation deviation value; A parameter generation module for calculating the error contribution value of each feature dimension according to the estimation deviation value and combining the feature dimension weight distribution in the feature vector, and performing a non-linear mapping to generate an updated parameter vector; A model update module for incrementally training the current weight estimation model according to the updated parameter vector to obtain an updated weight estimation model.

2. The control system of a weighing device based on active learning according to claim 1, wherein The sample evaluation module includes: A feature distance calculation sub-module for calculating the weighted Euclidean distance between the feature vector of the analysis data and multiple feature vectors in the historical labeled samples to obtain a sample feature distance; A time correlation processing sub-module for calculating the time density value according to the timestamp of the analysis data and the number of weighings per unit time, and performing weighted fusion based on the time density value and the sample feature distance to obtain a composite evaluation factor; A confidence evaluation sub-module for obtaining a sample confidence score according to the composite evaluation factor and the preset confidence scoring model. When the sample confidence score is lower than the preset confidence threshold, the current sample is marked as data to be learned.

3. The control system of a weighing device based on active learning according to claim 1, characterized in that, The parameter generation module includes: An error contribution analysis sub-module for calculating the error contribution value of each feature dimension according to the correlation between the estimation deviation value and the feature vector in the data to be learned; A parameter mapping sub-module for classifying the error contribution value into intervals and performing a non-linear mapping according to different mapping response methods for different levels to obtain a mapped error contribution value; An updated parameter generation sub-module for combining the mapped error contribution values in the order of feature dimensions to generate an updated parameter vector.

4. The control system of a weighing device based on active learning according to claim 1, wherein, The model update module includes: An orientation training sub-module for inputting the updated parameter vector into the current weight estimation model and performing incremental training on the current weight estimation model based on parameter replacement, weighted update, or learning rate adjustment without resetting the network structure and historical parameters of the weight estimation model to obtain an updated weight estimation model; An accuracy evaluation sub-module for performing estimation calculations on the updated weight estimation model on the old and new sample sets respectively to obtain the corresponding error mean and standard deviation; A stability feedback sub-module for constructing an error change curve according to the error mean and standard deviation to determine whether the updated weight estimation model meets the preset stability standard. When the judgment result is not met, a training adjustment signal is output for correcting the mapping response method of the subsequent updated parameter vector or adjusting the training frequency.

5. The control system of a weighing device based on active learning according to claim 2, characterized in that, The feature distance calculation sub-module includes: A feature vector extraction unit, configured to extract a feature vector constituting a sample feature from analysis data; A weight factor determination unit, configured to perform statistical analysis on each feature dimension in historical label samples, and by evaluating the numerical fluctuation degree of each feature dimension within a preset first time window, determine that relatively stable features correspond to higher weight values, while features with large fluctuations have lower weight values, thereby generating a set of weighting factors for distance calculation; A distance calculation unit, configured to perform a dimension-by-dimension difference calculation on the feature vector of current analysis data and multiple feature vectors of historical label samples, and accumulate the sum after multiplying each difference by the corresponding weighting factor to obtain a feature distance value representing the similarity degree of samples.

6. The control system of a weighing device based on active learning according to claim 5, wherein The time correlation processing sub-module includes: A time information extraction unit, configured to extract the timestamp information of a sample from analysis data, and count the number of weighing records within a current preset second time window, thereby obtaining the weighing frequency per unit time; A time density calculation unit, configured to calculate the time interval length of a sample by comparing the timestamp information of the sample with the current system time, and perform weighted combination after normalizing the reciprocal of the time interval length and the weighing frequency per unit time respectively to calculate a time density value; A weighted fusion unit, configured to jointly evaluate the feature distance value and the time density value, and by setting an adjustment parameter that varies with the time density value, enhance the influence degree of samples collected during high-density periods on the feature distance value, thereby forming a composite evaluation factor for representing the comprehensive evaluation level of samples.

7. The control system of a weighing device based on active learning according to claim 6, wherein, The confidence evaluation sub-module includes: A scoring parameter loading unit, configured to load a preset confidence scoring model, where the confidence scoring model is constructed based on multiple influencing factors, including the composite evaluation factor, the time density value corresponding to the sample, and the local fluctuation amplitude of the original signal sequence; A scoring calculation unit, configured to input multiple influencing factors into the confidence scoring model, and calculate the confidence score value of the current sample according to the parameter weights corresponding to each factor; A threshold comparison unit, configured to compare the confidence score value of the current sample with a preset confidence threshold. When the score is lower than the threshold, output the to-be-learned marker information and classify the sample into the to-be-learned data set; if the score is higher than the threshold, ignore the current sample for model training.

8. The control system of a weighing device based on active learning according to claim 3, characterized in that, The error contribution analysis sub-module includes: A feature value analysis unit, configured to calculate the proportion information of each feature in the entire feature vector according to the feature vector of the to-be-learned data and its corresponding estimated deviation value, and based on the relative magnitudes of the feature values of each feature dimension; A correlation adjustment unit, configured to determine the relative contribution degree of each feature dimension to the total deviation according to the proportion information and the estimated deviation value, and generate an error contribution value. Among them, a reduction coefficient is set for the feature dimension with a large change range of feature values but low stability during training to reduce its dominant role in parameter generation.

9. The control system of a weighing device based on active learning according to claim 8, characterized in that, The parameter mapping sub-module includes: A contribution level discrimination unit, which is used to divide the error contribution values into multiple levels according to the set contribution grading criteria based on the error contribution values of each feature dimension, and generate a contribution level division result. Among them, the error contribution values of higher levels will obtain a larger parameter adjustment range; A non-linear mapping strategy unit, which is used to perform non-linear mapping response processing on the error contribution values through a mapping function group according to the contribution level division result, and generate the mapped error contribution values. Among them, a compression mapping is used for the error contribution values of lower levels to suppress the model changes caused by weak perturbations, and an enhancement mapping is used for the error contribution values of higher levels to accelerate the model convergence.

10. A control system for a weighing device based on active learning according to claim 4, wherein, The stability feedback sub-module includes: An error trend extraction unit, which is used to construct an error change curve according to the changes of the error mean and standard deviation in the current multiple update cycles; A stability judgment unit, which is used to judge whether the current model is in a stable state after update according to the slope and fluctuation frequency of the error change curve. If the fluctuation amplitude continues to rise or the slope exceeds the set critical value, it is judged that the stability standard is not met, otherwise it is met; A feedback generation unit, which is used to generate a training adjustment signal when it is judged that the current model is in an unstable state after update. The training adjustment signal is used to trigger the parameter mapping sub-module to re-adjust the mapping range or reduce the model update frequency.

Images