A weighing equipment control system based on active learning

Abstract

The present invention provides a weighing equipment control system based on active learning, which relates to the field of data processing technology. The system comprises: a sample evaluation module, which is used to perform sample confidence evaluation on analysis data to obtain a sample confidence score. When the sample confidence score is lower than a preset confidence threshold, the analysis data is marked as data to be learned; a deviation calculation module, which is used to estimate the feature vector of the data to be learned based on the feature vector of the data to be learned and based on the current weight estimation model, output the estimation result, and perform difference calculation between the estimation result and the actual result; a parameter generation module, which is used to combine the feature dimension weight distribution in the feature vector, calculate the error contribution value on each feature dimension, perform nonlinear mapping, and generate an updated parameter vector; and a model update module, which is used to perform incremental training on the current weight estimation model based on the updated parameter vector. The present invention improves the autonomy and accuracy of the weighing equipment control system.

Description

Technical Field

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

[0002] In existing technologies, the control systems of weighing equipment generally use preset parameters and fixed models for weight detection and control. Such systems typically rely on manual calibration to complete the initial equipment settings. By collecting sensor data (such as pressure sensors, resistance strain gauges, etc.) and inputting it into a preset algorithm, the weight of the target object is calculated. In scenarios such as industrial production lines and logistics sorting, weighing systems may also be equipped with PLCs or embedded control units to achieve automated weighing and process control. Although some systems have data feedback mechanisms that can adjust weighing strategies according to set rules, the overall system is still mainly based on static learning mode and lacks the ability to actively identify and retrain new samples or edge data, resulting in poor adaptability.

[0003] During the agricultural product packaging process, large sample variability, such as the varying size, density, and surface attachments of natural crops like potatoes and onions, often makes it difficult for fixed-model weighing systems to accurately estimate actual weight. For example, when potatoes are covered in mud or consist of multiple connected individuals, traditional weighing systems may mistake them for a single heavy object, causing classification errors and impacting subsequent packaging processes. Due to the system's lack of active learning capabilities for anomalous samples, it cannot automatically adjust model parameters or collect new samples for training. Over time, errors accumulate, requiring frequent manual intervention and recalibration, reducing the system's adaptability and practicality. Summary of the Invention

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

[0005] In order to solve the above technical problems, the technical solutions of the present invention are as follows:

[0006] A weighing equipment control system based on active learning, the system comprising:

[0007] The sample evaluation module is used to evaluate the sample confidence of the analysis data. Specifically, it calculates the feature distance between the historical label sample and the analysis data, and combines the timestamp and 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] The deviation calculation module is used to estimate the feature vector of the data to be learned 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;

[0009] A parameter generation module is used to calculate the error contribution value of each feature dimension based on the estimated deviation value and the feature dimension weight distribution in the feature vector, and perform nonlinear mapping to generate an updated parameter vector for adjusting the weight estimation model;

[0010] The model update module is used to perform incremental training on the current weight estimation model according to the updated parameter vector to obtain an updated weight estimation model.

[0011] Preferably, the sample evaluation module includes:

[0012] The feature distance calculation submodule is used to calculate the weighted Euclidean distance between the feature vector of the analysis data and multiple feature vectors in the historical label sample to obtain the sample feature distance;

[0013] The time correlation processing submodule is used to calculate the time density value based on the timestamp of the analysis data and the number of weighings per unit time, and to perform weighted fusion based on the time density value and the sample feature distance to obtain a composite evaluation factor;

[0014] The confidence assessment submodule is used to obtain a sample confidence score based on 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.

[0015] Preferably, the parameter generation module includes:

[0016] The error contribution analysis submodule is used to calculate the error contribution value of each feature dimension based on the correlation between the estimated deviation value and the feature vector in the data to be learned;

[0017] The parameter mapping submodule is used to classify the error contribution values ​​into intervals and perform nonlinear mapping according to different mapping response modes according to different levels to obtain the mapped error contribution values;

[0018] The update parameter generation submodule is used to combine the mapped error contribution values ​​in the order of feature dimensions to generate an update parameter vector.

[0019] Preferably, the model updating module includes:

[0020] A directed training submodule is used to input the updated 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] The accuracy evaluation submodule is used to estimate and calculate the updated weight estimation model on the new and old sample sets respectively to obtain the corresponding error mean and standard deviation;

[0022] The stability feedback submodule is used to construct an error change curve based on the error mean and standard deviation, so as to determine whether the updated weight estimation model meets the preset stability standard. If the judgment result is not satisfied, it outputs a training adjustment signal to correct the mapping response method of the subsequent update parameter vector or adjust the training frequency.

[0023] Preferably, the feature distance calculation submodule includes:

[0024] A feature vector extraction unit, used to extract feature vectors constituting sample features from the analysis data;

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

[0026] The distance calculation unit is used to calculate the dimension-by-dimensional difference between the feature vector of the current analysis data and multiple feature vectors of the historical label samples, and multiply each difference by the corresponding weighting factor and then add them up to obtain the feature distance value representing the degree of sample similarity.

[0027] Preferably, the time correlation processing submodule includes:

[0028] A time information extraction unit is used to extract the timestamp information of the sample from the analysis data and count the number of weighing records within the current preset second time window to obtain the weighing frequency per unit time;

[0029] The time density calculation unit is used to compare the timestamp information of the sample with the current system time, calculate the time interval length of the sample, and calculate the time density value by normalizing the reciprocal of the time interval length and the weighing frequency per unit time and performing weighted combination;

[0030] 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 samples collected during high-density periods on the feature distance value is enhanced, thereby forming a composite evaluation factor for representing the comprehensive evaluation level of the sample.

[0031] Preferably, the confidence assessment submodule includes:

[0032] A scoring parameter loading unit is used to load a preset confidence scoring model. The confidence scoring model is constructed based on multiple influencing factors, including a 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;

[0033] The scoring calculation unit is used to input multiple influencing factors into the confidence scoring model and calculate the confidence score value of the current sample according to the parameter weight corresponding to each factor;

[0034] The threshold comparison unit is used to compare the confidence score value of the current sample with the preset confidence threshold. When the score is lower than the threshold, the to-be-learned label information is output and the sample is included in the to-be-learned data set; if the score is higher than the threshold, the current sample is ignored for model training.

[0035] Preferably, the error contribution analysis submodule includes:

[0036] The eigenvalue parsing unit is used to calculate the proportion of each feature in the entire feature vector based on the eigenvector of the data to be learned and its corresponding estimated deviation value, and based on the relative size of the eigenvalues ​​of each feature dimension;

[0037] The correlation adjustment unit is used to determine the relative contribution of each feature dimension to the total deviation based on the proportion information and the estimated deviation value, and generate an error contribution value. Among them, a reduction coefficient is set for feature dimensions with a large range of feature value changes but low stability during training to reduce their dominant role in parameter generation.

[0038] Preferably, the parameter mapping submodule includes:

[0039] The contribution level discrimination unit is used to classify the error contribution value of each feature dimension into multiple levels according to the set contribution classification standard, and generate a contribution level classification result, wherein the error contribution value of a higher level will obtain a larger parameter adjustment range;

[0040] The nonlinear mapping strategy unit is used to divide the results according to the contribution level, perform nonlinear mapping response processing on the error contribution value through the mapping function group, and generate the mapped error contribution value. Among them, compression mapping is used for low-level error contribution values ​​to suppress the model changes caused by weak disturbances, and enhancement mapping is used for high-level error contribution values ​​to accelerate model convergence.

[0041] Preferably, the stability feedback submodule includes:

[0042] The error trend extraction unit is used to construct an error change curve based on the changes in the error mean and standard deviation in the current multiple update cycles;

[0043] The stability judgment unit is used to judge whether the current model is in a stable state after the update based on the slope and fluctuation frequency of the error change curve. If the fluctuation amplitude continues to increase or the slope exceeds the set critical value, it is judged that the stability standard is not met, otherwise it is met;

[0044] 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 being updated. The training adjustment signal is used to trigger the parameter mapping submodule to readjust the mapping range or reduce the model update frequency.

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

[0046] First, this invention utilizes an active learning mechanism to dynamically identify and label low-confidence analysis data as samples to be learned, overcoming the limitations of existing weighing control systems, which rely on fixed models and static operation. The sample evaluation module compares the features of the analysis data with historically labeled samples and, combining timestamps with weighing frequency, constructs a confidence score in real time, thus implementing a learning mechanism that simultaneously identifies and samples samples. This feature enables the system to automatically detect new samples, edge cases, or samples in unusual conditions (such as attachments or structural connections on natural crop surfaces), improving the model's ability to identify abnormal conditions and reducing the need for manual intervention.

[0047] Secondly, the present invention constructs an error-driven model update mechanism through a parameter generation module. After the deviation calculation module estimates the learning data, the system decomposes the estimated deviation into each feature dimension through error contribution analysis, and then combines it with a nonlinear mapping strategy to generate a structured and controllable update parameter vector. This solution realizes the fine-tuning of local parameters of the weight estimation model, avoiding the computational burden brought by the overall reconstruction or full training in the existing technology, and realizes incremental training through the model update module, so that the model has the ability to quickly adapt to new data patterns while maintaining stability.

[0048] Thirdly, the accuracy assessment and stability feedback mechanism provided by this invention enables the system to automatically detect the changing trend of prediction errors after model updates, dynamically determine whether the model is in a convergence state, and output training adjustment signals if stability criteria are not met. By limiting the update frequency or adjusting the parameter mapping response range, model oscillation and error amplification can be effectively suppressed, thereby improving the robustness and security of the model's long-term operation.

[0049] Finally, this invention is suitable for weighing applications involving highly variable natural objects, such as the automated weighing of potatoes, onions, and other items in agricultural product packaging lines. The system automatically adapts to changes in sample volume, density, and surface condition, significantly improving weighing and sorting accuracy, reducing false alarm rates and manual calibration frequency, and enhancing the system's practicality and maintainability in complex environments. Overall, this invention possesses the triple capabilities of self-learning, self-adjustment, and self-stabilization, significantly outperforming existing static model control systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This 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] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although 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. Rather, 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 provides a weighing equipment control system based on active learning, the system comprising:

[0053] A data acquisition module is used to obtain raw sensor data of the weighing target, wherein the raw sensor data includes multiple voltage signals collected by the weighing device within a unit time;

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

[0055] The sample evaluation module is used to evaluate the sample confidence of the analysis data. Specifically, it calculates the feature distance between the historical label sample and the analysis data, and combines the timestamp and 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 of the data to be learned 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 is used to calculate the error contribution value of each feature dimension based on the estimated deviation value and the feature dimension weight distribution in the feature vector, and perform nonlinear mapping to generate an updated parameter vector for adjusting the weight estimation model;

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

[0059] The weight control module is used to perform real-time weight estimation on the subsequently input analysis data according to the updated weight estimation model, and output the estimation result as the weighing control output.

[0060] In an embodiment of the present invention, a weighing equipment control system based on an active learning mechanism is constructed, achieving dynamic updating and precision enhancement of traditional weighing control models. This system employs a multi-level modular architecture. First, a data acquisition module acquires raw sensor data from the weighing target. This data is derived from multiple voltage signals output by the pressure sensor per unit time. These signals represent the actual force exerted by the object under test on the weighing platform. High-speed sampling ensures both the temporal resolution of data acquisition and the continuity of the physical response, providing a high-quality data foundation for subsequent processing.

[0061] Next, the data preprocessing module removes outliers, performs smoothing filtering, and normalizes the raw sensor data. Extreme value removal is based on a predefined empirical threshold, smoothing can utilize time series methods such as sliding averages, and normalization is adjusted based on sensor placement. After processing, the resulting analytical data can be used to further assess the value and confidence level of the sample.

[0062] The system incorporates a sample evaluation module that compares the current analysis data with historically labeled samples. This module then combines timestamps and the frequency of weighting per unit time to generate a dynamic sample confidence score. When the score falls below a system-defined threshold, the system marks the sample as "data to be learned." Compared to traditional static systems that rely on manually defined boundaries, this evaluation mechanism proactively identifies model blind spots, thereby maximizing the value of samples.

[0063] After obtaining the feature vectors of the data to be learned, the subsequent deviation calculation module generates an estimated value based on the currently used weight estimation model. This value is then compared with 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 a starting point for subsequent parameter optimization.

[0064] Based on the deviation value and the importance of each dimension in the feature vector (i.e., the dimension weight distribution), the parameter generation module assesses the specific influence of each feature dimension on the deviation and performs nonlinear mapping on it to generate an updated parameter vector for model updating. This process avoids over-response to weak perturbations and strengthens the adaptive weight adjustment of key feature dimensions.

[0065] The model update module does not reset the overall model structure. Instead, it performs incremental training based on the original parameters and only makes targeted corrections to the weights. This update strategy preserves the steady-state performance of the original model while enabling rapid adaptation to new data features.

[0066] Finally, the system performs real-time estimation operations on the subsequently connected analysis data through the weight control module, and outputs the results as weighing control signals to drive downstream mechanical execution equipment such as packaging and sorting to achieve refined control.

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

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

[0069] The feature distance calculation submodule is used to calculate the weighted Euclidean distance between the feature vector of the analysis data and multiple feature vectors in the historical label sample to obtain the sample feature distance;

[0070] The time correlation processing submodule is used to calculate the time density value based on the timestamp of the analysis data and the number of weighings per unit time, and to perform weighted fusion based on the time density value and the sample feature distance to obtain a composite evaluation factor;

[0071] The confidence assessment submodule is used to obtain a sample confidence score based on 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.

[0072] To improve the accuracy and proactive screening capabilities of weighing data confidence assessment, this embodiment of the present invention designed a sample evaluation module consisting of a feature distance calculation submodule, a temporal correlation processing submodule, and a confidence assessment submodule. This module utilizes a dual temporal and spatial feature fusion mechanism to efficiently determine sample confidence levels.

[0073] First, the feature distance calculation submodule receives feature vectors extracted from the analyzed data and performs a weighted Euclidean distance calculation with each feature vector in the historical label sample library. During this process, the weight factor for each feature dimension is set based on its fluctuation within a preset time window, with more stable features receiving higher weights. The resulting weighted feature distance reflects the overall deviation between the current sample and the labeled sample, serving as the primary basis for evaluating the sample's "knownness."

[0074] Next, the time correlation processing submodule extracts sample timestamps from the analyzed data and counts the current system's weighing frequency within the second time window. Based on the interval between sample collection time and the current system time, combined with the frequency level, a time density value reflecting the sample's "freshness" is calculated. To prioritize sample freshness, the system incorporates an adjustable density response factor, which combines the time density value with the characteristic distance to generate a composite evaluation factor.

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

[0076] Finally, the confidence assessment submodule uses a composite evaluation factor, combined with the confidence scoring model currently loaded by the system, to quantify the sample's confidence level. This scoring model considers the sample's characteristic deviation, temporal density, and the volatility of the original signal sequence. After obtaining the score, the system compares it with a preset threshold. If the score falls below the threshold, a flag 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 guided by sample quality, which avoids a large number of invalid samples from participating in training, improves the value of training data, and thus enhances model iteration efficiency and training convergence speed.

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

[0079] The error contribution analysis submodule is used to calculate the error contribution value of each feature dimension based on the correlation between the estimated deviation value and the feature vector in the data to be learned;

[0080] The parameter mapping submodule is used to classify the error contribution values ​​into intervals and perform nonlinear mapping according to different mapping response modes according to different levels to obtain the mapped error contribution values;

[0081] The update parameter generation submodule is used to combine the mapped error contribution values ​​in the order of feature dimensions to generate an update parameter vector.

[0082] To improve the precision and targeted nature of model updates, this embodiment of the present invention proposes an error-driven parameter generation module. This module generates a structured update parameter vector from the deviation value to guide subsequent incremental model training. This module includes an error contribution analysis submodule, a parameter mapping submodule, and an update parameter generation submodule. Its core concept is to leverage the error representation of feature dimensions to drive the intensity of weight updates, achieving a dynamic balance between control accuracy and convergence speed.

[0083] First, the Error Contribution Analysis submodule receives the feature vector of the data to be learned and its estimated deviation value. It analyzes the contribution of each feature dimension to the feature vector, assessing the relative importance of each dimension's value in the overall vector. The system further calculates the contribution of each feature dimension to the overall deviation based on the combined relationship between feature contribution and deviation value. For feature dimensions that have demonstrated significant volatility or instability during historical training, the system automatically sets a reduction factor to prevent them from dominating the model's training direction.

[0084] The parameter mapping submodule then classifies the error contribution of each feature dimension into intervals, forming a contribution level mapping structure. Each level corresponds to a response mode: dimensions with smaller contribution values ​​adopt a compressed response strategy to suppress excessive parameter amplification caused by subtle perturbations; dimensions with larger contribution values ​​activate an enhanced mapping mechanism, amplifying their weight adjustment proportion in model updates. This mapping process is accomplished based on a nonlinear mapping function, enabling an adaptive curvilinear response to input errors.

[0085] Finally, the update parameter generation submodule recombines the mapped error contributions in the order of the feature dimensions to construct a complete update parameter vector. This vector not only reflects the explanatory power of each dimension in the current deviation but also controls its degree of participation through mapping, thereby achieving differentiated training updates.

[0086] This module enhances the controllability and interpretability of updated parameters, while maintaining the model's convergence stability, achieving rapid adaptation to key features, and ultimately improving the model's generalization ability and response accuracy in actual weighing tasks.

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

[0088] A directed training submodule is used to input the updated 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 submodule is used to estimate and calculate the updated weight estimation model on the new and old sample sets respectively to obtain the corresponding error mean and standard deviation;

[0090] The stability feedback submodule is used to construct an error change curve based on the error mean and standard deviation, so as to determine whether the updated weight estimation model meets the preset stability standard. If the judgment result is not satisfied, it outputs a training adjustment signal to correct the mapping response method of the subsequent update parameter vector or adjust the training frequency.

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

[0092] The directed training submodule receives the generated updated parameter vector and inputs it into the weight estimation model currently in use. This model contains 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 submodule does not adopt a model reinitialization strategy, but instead performs fine-grained parameter adjustments through a preset update mechanism. Specifically, if the amplitude of the updated parameter vector is small, it is directly applied by parameter replacement; if the parameter changes are concentrated and of moderate magnitude, a weighted update method is used; during 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 of new samples while maintaining its original structural capabilities, effectively extending the model life cycle and improving learning efficiency.

[0093] The accuracy assessment submodule objectively measures model performance before and after an update. This submodule incorporates two sample sets: the historical sample set before the update and the updated set of to-be-learned samples. The system estimates the weight of the updated weight estimation model in both sample sets and calculates the mean and standard deviation of the estimation error to determine the generalization performance and accuracy trends after the update. This method quantifies the model's adaptability to both new and old samples, helping to determine the effectiveness of the update.

[0094] The stability feedback submodule constructs an error trend curve based on the aforementioned error mean and standard deviation data. This trend curve determines whether the current model has entered an oscillating state or is in stable convergence by observing the direction and magnitude of the estimated error changes over multiple consecutive training cycles. If the slope of the error curve exceeds a positive threshold, or the error standard deviation continues to increase, the system determines that the model is in an unstable range and immediately triggers a training adjustment signal. This signal is fed back to the parameter generation module, which adjusts the response range of the parameter mapping function or reduces the training frequency until the error returns to a stable state.

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

[0096] The directed training submodule is used to input the updated 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. Specifically, it includes:

[0097] This submodule is the core component of the online model update mechanism, enabling localized, targeted adjustments to model parameters without disrupting the original model structure. Compared to traditional model systems that require complete retraining, this module effectively reduces update latency and computing resource consumption, making it particularly suitable for industrial control scenarios that require learning while running.

[0098] In practice, the system first receives the update parameter vector output by the parameter generation module. This vector consists of multiple parameter increments corresponding to the model's feature dimensions, processed through nonlinear mapping, and contains information about directionality and adjustment strength. After matching this vector with the current model's weight parameters, the system selects one or more update methods based on the set strategy:

[0099] First, parameter replacement strategy: If the change in the updated parameter of 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, a weighted update strategy: For medium-intensity updates, the system performs a weighted fusion of the original weights and the updated parameters. The recommended approach is to set a primary weight coefficient (e.g., 0.7) to assign historical weights to maintain model continuity, and then multiply the updated parameters by a secondary weight coefficient (e.g., 0.3) to fine-tune the direction. This approach is particularly suitable for adjustments after a model has entered a steady-state phase after long-term operation.

[0101] Third, learning rate adjustment strategy: If the update parameters change dramatically in multiple feature dimensions or the system determines that the current training is in an unstable state, the overall learning rate can be adjusted downward, or a dynamic step function can be set to make the model weight adjustment smoother and avoid 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, its force distribution on the sensor may differ from historical samples. This module allows the system to quickly replace or slightly correct the model weights corresponding to the relevant sensor channels without requiring full model regression training. This allows the system to complete model updates within 5 seconds, ensuring continuous weighing accuracy.

[0103] Through the above method, the targeted training submodule achieves the model update goal of "low intrusion, low computation, and high adaptability". It is suitable for actual industrial environments with frequent fluctuations in sensor signals and high object diversity, and improves the overall responsiveness of the system and the stability of long-term operation.

[0104] The accuracy assessment submodule is used to estimate and calculate the updated weight estimation model on the new and old sample sets respectively to obtain the corresponding error mean and standard deviation, specifically including:

[0105] This submodule objectively and quantitatively analyzes the effectiveness of model updates, providing a basis for determining whether to trigger feedback control. In traditional model training, accuracy assessments often rely solely on static testing of the training or validation sets, failing to accurately reflect the model's adaptability to new data. This module dynamically evaluates model performance trends by introducing a dual-set error comparison mechanism.

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

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

[0108] The second set is the set of samples to be learned, which contains new samples that have been recently marked as to be learned and have undergone parameter generation and training updates. It is used to evaluate the model's ability to cover the new feature space.

[0109] The updated weight estimation model predicts each piece of data in both sample sets, generating an estimated output value for each sample. The system then compares the estimated value with the actual weight value, calculates the prediction error for each sample, and then calculates the mean and standard deviation of the error for both sets.

[0110] in:

[0111] The mean error is used to assess the degree of prediction deviation. If the value is large, it means that the model has systematic deviations.

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

[0113] To enhance the ability to identify changing trends, the system can employ a sliding window mechanism to sort and compare the mean error values ​​over successive update cycles. For example, the direction of error change over the three most recent updates can be calculated to determine whether the error is gradually decreasing (converging) or fluctuating repeatedly (instability). Furthermore, a dynamic baseline can be set to assess whether degradation is occurring based on a set historical best accuracy value.

[0114] For example, if after a model update, the mean error in the historical sample set increases by more than 10%, and the standard deviation of the error in the set to be learned also increases significantly, the system may determine that the current model update effect is poor and the stability feedback mechanism should be triggered to limit further updates.

[0115] Through this module, the system can monitor the trajectory of model accuracy changes in real time without external human intervention, and provide a basis for training and adjustment for subsequent modules. It 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 submodule includes:

[0117] A feature vector extraction unit is used to extract a feature vector constituting a sample feature from the analysis data, wherein the feature vector is composed of multiple sensor signal values ​​after normalization processing and is used to reflect the distribution characteristics of the sample in multiple feature dimensions;

[0118] A weight factor determination unit is used to perform statistical analysis on each feature dimension in the historical label samples. By evaluating the degree of value fluctuation of each feature dimension within a preset first time window, it is determined that relatively stable features have higher weight values, while features with large fluctuations have lower weight values, thereby generating a set of weighting factors for distance calculation;

[0119] The distance calculation unit is used to calculate the dimension-by-dimensional difference between the feature vector of the current analysis data and multiple feature vectors of the historical label samples, and multiply each difference by the corresponding weighting factor and then add them up to obtain the feature distance value representing the degree of sample similarity.

[0120] In an embodiment of the present invention, in order to realize the feature similarity evaluation between the analysis data and the historical label samples, the feature distance calculation submodule is divided into a feature vector extraction unit, a weight factor determination unit and a distance calculation unit, which is 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 multidimensional feature vectors that characterize the samples from the analyzed 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 across various physical measurement dimensions. These vectors represent the relative distribution of the samples in space and serve as the basic data structure for subsequent similarity calculations.

[0122] The weight factor determination unit assesses the importance of each feature dimension and assigns a weighting coefficient. Specifically, this unit analyzes the fluctuation amplitude of each feature dimension within a preset first time window. Dimensions exhibiting a small variance or standard deviation within this window are considered stable and given a higher calculation weight. Conversely, dimensions with high volatility have their influence in the calculation weakened. This approach ensures that when calculating feature distances, the system prioritizes feature components that have substantive discriminatory power.

[0123] The distance calculation unit uses the current sample's feature vector as a benchmark and compares it to multiple feature vectors in the historical label sample library. After performing a difference calculation on each corresponding dimension, it is multiplied by a weighting factor and the weighted differences across all dimensions are summed to obtain a weighted distance value that measures the similarity between the current sample and historical samples. The resulting feature distance not only reflects the overall proximity between samples but also takes into account the differences in discriminative power between feature dimensions, facilitating accurate confidence scoring.

[0124] This submodule can effectively avoid the interference of low-value features on the similarity judgment results in a mixed scenario of multiple types of samples, and enhance the robustness and generalization of sample evaluation.

[0125] The feature vector extraction unit is used to extract feature vectors that constitute sample features from the analysis data. The feature vectors are composed of multiple sensor signal values ​​after normalization and are used to reflect the distribution characteristics of the sample in multiple feature dimensions. Specifically, they include:

[0126] This unit converts preprocessed time series signals into structured multidimensional data representations, enabling the samples to participate in feature similarity analysis in subsequent processes. The analysis data is output by the data preprocessing module, which has already undergone basic processing such as extreme value removal, signal smoothing, and normalization. The sensor signal values ​​include voltage data collected by pressure sensors installed at different locations on the weighing platform. After normalization, these data are uniformly mapped to the [0, 1] interval.

[0127] During implementation, the system treats each piece of analytical data as a sample segment and extracts its normalized response value at that time slice, measured by the sensor channel dimension. Assuming 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 left of the platform's center, the output voltages of the two sensors on the left are significantly higher than those of the two on the right. After normalization, the feature vector of this sample might appear as "0.82, 0.77, 0.35, 0.29," representing the relative proportions of the pressure distribution on the platform. This structured vector can then be used as input for subsequent feature distance calculations.

[0129] This unit extracts the response ratio of samples on different sensors to form a unified feature representation. 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 samples. By evaluating the degree of numerical fluctuation of each feature dimension within a preset first time window, it is determined that relatively stable features have higher weight values, while features with large fluctuations have lower weight values. In this way, a set of weighting factors for distance calculation is generated, specifically including:

[0131] The core function of this unit is to identify which feature dimensions are more representative and discriminative in historical samples and assign them higher weights, thereby improving the discriminative power of feature distance calculations. In existing technologies, all feature dimensions are often treated as equivalent, causing certain volatile but low-information features to cause unnecessary interference in distance calculations. This unit statistically evaluates the fluctuations of each dimension and dynamically assigns its importance in the overall calculation.

[0132] The system sets a fixed-length time window as the analysis interval, for example, the set of the most recent 200 labeled samples. Within this window, the system calculates the fluctuation range of the values ​​for each feature dimension. Volatility can be characterized by methods such as the difference between the maximum and minimum values ​​or the standard deviation. To simplify implementation, we recommend using 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] The system reversely generates weighting factors based on the aforementioned volatility results: that is, smaller fluctuations result in higher weights, while larger fluctuations result in lower weights. To avoid overweighting or underweighting certain dimensions, which could affect overall calculation stability, it is recommended to normalize all raw weights so that their sum is 1.

[0135] For example, in a four-channel system, analysis reveals that channels 1 and 2 have relatively low standard deviations, 0.04 and 0.06, respectively, while channels 3 and 4 have relatively high standard deviations, 0.11 and 0.15, respectively. The system can then assign weights of 0.35, 0.30, 0.20, and 0.15 to the four channels, respectively. This weighting method emphasizes the importance of low-variability channels for sample classification, helping the system focus on key features and suppressing noise interference.

[0136] This unit effectively constructs a mapping mechanism from "statistical characteristics" to "weight factors", providing a basis for discriminant priority sorting for subsequent distance calculations, and improving the accuracy of the system's similarity judgment in actual weighing applications.

[0137] The distance calculation unit is used to calculate the dimension-by-dimension difference between the feature vector of the current analysis data and multiple feature vectors of the historical label samples, and multiply each difference by the corresponding weighting factor and then add them up to obtain the feature distance value indicating the degree of sample similarity, which specifically includes:

[0138] This unit quantifies the differences between two feature vectors across multiple feature dimensions. Traditional Euclidean distance calculations simply use the sum of the squared differences in each dimension as the basis for similarity, ignoring the importance differences between feature dimensions and making the distance results overly sensitive to local perturbations. To address this issue, the present invention introduces a weighted distance mechanism. This calculation incorporates the weighting factors generated above to adjust the contribution of the differences, resulting in a distance result that more closely reflects the similarity between real samples.

[0139] In practice, the system retrieves the feature vector (denoted as A) from the current sample and sequentially reads the comparison vectors (denoted as B) from the historical labeled sample set. For each corresponding feature dimension in A and B, the absolute difference is calculated and multiplied by the weighting factor corresponding to that dimension. All weighted differences are then summed 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" and the sample to be compared is "0.80, 0.75, 0.42, 0.34", the corresponding dimension weights are "0.35, 0.30, 0.20, 0.15" respectively. The system then calculates the difference in each dimension as: 0.02, 0.02, 0.07, 0.05, multiplies each by the weight to obtain the weighted result, and finally sums them to obtain the final distance value.

[0141] It should be pointed out that in order to improve the computational efficiency of the system, 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 between samples in the multi-dimensional sensor feature space, avoiding the deviation of distance results caused by irrelevant features, thereby more accurately judging the similarity between the current analysis data and historical label samples, and ultimately improving the accuracy of selecting samples to be learned in active learning and the intelligence of system response.

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

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

[0145] The time density calculation unit 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 based on the reciprocal of the time interval length and the weighing frequency per unit time, and calculate the time density value. The shorter the time interval length and the higher the unit frequency, the greater the time density value obtained;

[0146] 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 samples collected during high-density periods on the feature distance value is enhanced, thereby forming a composite evaluation factor for representing the comprehensive evaluation level of the sample.

[0147] In an embodiment of the present invention, in order to strengthen the time sensitivity evaluation mechanism of samples and improve the responsiveness of sample confidence scores to temporal characteristics, the system sets up a time correlation processing submodule, which consists of a time information extraction unit, a time density calculation unit and a weighted fusion unit, and is used to dynamically weight the samples in combination with time information.

[0148] The time information extraction unit extracts the timestamp information for each sample from the analyzed data and counts the number of weighing records within the second time window currently set by the system. This window setting ensures that the system can adaptively determine the frequency of data generation per unit time, i.e., the weighing frequency per unit time, even when sampling density varies. This frequency value reflects the operation density within the sample's time period and is a key basis for calculating sample timeliness.

[0149] The time density calculation unit compares the sample timestamp with the current system time, calculates the time interval length, and combines this interval length with the unit time frequency to calculate the time density value. The calculation result reflects the "freshness" of the sample: the newer and more frequently occurring samples generate a larger time density value. This density value is used to measure the importance of the time period in which the sample occurs.

[0150] The weighted fusion unit fuses the aforementioned temporal density values ​​with the sample's characteristic distance values. During this fusion process, the system sets a tuning parameter that changes with temporal density. This parameter increases the weight of the current sample in the comprehensive evaluation factor as temporal density increases, ensuring that samples with high frequency in the dataset and close to the current moment receive more model attention. The resulting composite evaluation factor integrates the spatial similarity and temporal timeliness of the samples and serves as the core input for subsequent confidence calculations.

[0151] This module improves the system's responsiveness to fresh data and effectively avoids outdated data dominating the training process, which is conducive to building a weighing identification system with real-time update capabilities.

[0152] The time density calculation unit is used to compare the timestamp information of the sample with the current system time to calculate the time interval length of the sample; and based on the reciprocal of the time interval length and the weighing frequency per unit time, normalize and perform weighted combination to obtain a time density value. The shorter the time interval and the higher the frequency per unit time, the larger the time density value obtained. Specifically, it includes:

[0153] This unit aims to measure the "currentness" and "activity" of samples in the time dimension, which is used to assist in determining whether the sample has the time value to participate in model updates first. 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 lagging prediction capabilities. By constructing a "time density" indicator, this invention comprehensively evaluates the timeliness of samples and the current sampling activity of the system, making the system more sensitive to data evolution trends.

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

[0155] In the first step, the system extracts the timestamp of the current sample and compares it with the current system time to determine the time interval. This time interval can be measured in seconds or minutes, reflecting the time since the sample was collected. The shorter the interval, the fresher the sample, and the more prioritized it should be.

[0156] In the second step, the system uses the time information extraction unit to obtain the weighing frequency per unit time. This frequency is defined as the number of weighing records actually completed by the system within a set time window (for example, the past 10 minutes) divided by the window length. A higher frequency indicates that the system has recently generated more data, indicating that the current model is actively updating.

[0157] In the third step, the system normalizes the two indicators mentioned above. The normalization range is recommended to be fixed between 0 and 1, where:

[0158] The longer the time interval, the lower the normalized value. 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. The historical maximum frequency or the empirical maximum value can be set as the upper limit standard.

[0160] In the fourth step, the system weights the reciprocal of the normalized time interval and the unit time frequency value proportionally to generate the final time density value. The weighting ratio can be set by the system based on experience. For example, weighting factors of 0.4 and 0.6, respectively, indicate that real-time performance is slightly lower than activity performance in time density judgment.

[0161] For example, assuming the timestamp of an analysis sample is 120 seconds from the current time and the maximum reference interval is 300 seconds, the normalized reciprocal score is approximately 0.6. The unit time frequency is 15 times / minute, the upper reference limit is 20 times / minute, and the normalized value is 0.75. The final time density value is 0.6 × 0.4 + 0.75 × 0.6 = 0.69 (for illustration only).

[0162] The resulting temporal density value serves as a weighting factor for subsequent "feature distance fusion" and "confidence scoring" steps. The system can dynamically prioritize a sample based on this value. For example, when the temporal density value exceeds a set threshold (e.g., 0.7), the system can prioritize the sample in active learning screening, accelerating its adaptation to fresh data patterns.

[0163] This unit combines the two dimensions of time interval and data frequency to construct a timeliness judgment indicator associated with the task status, avoiding the traditional system's excessive reliance on "old samples" and significantly improving the model's training efficiency and responsiveness in a dynamic data environment.

[0164] The weighted fusion unit is used to jointly evaluate the feature distance value and the time density value. By setting the adjustment parameter that changes with the time density value, the influence of samples collected during high-density periods 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 metric, making it a key node in active learning sample selection strategies. Existing methods focus solely on spatial distance (such as Euclidean distance and Mahalanobis distance), overlooking the crucial influence of temporal density on data timeliness. This can easily lead to the model collecting structurally "unfamiliar" but outdated data. This unit dynamically adjusts the component weights of the sample evaluation criteria by setting an adjustable fusion strategy.

[0166] During implementation, the system receives the feature distance value (reflecting the difference between the current sample and historically labeled samples) output by the feature distance calculation submodule, and the time density value output by the time density calculation unit. The system integrates these two data sets using a fusion factor, which controls the degree to which time density influences the overall evaluation factor. The value range of the fusion factor can be set adaptively by the system or adjusted based on fluctuations in historical system data.

[0167] Specifically, the system can set the fusion factor to a function that increases linearly or nonlinearly with the time density value. When the time density value is in the 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. Conversely, when the time density value is low, the fusion factor tends to be conservative, and the system relies more strictly on 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 the weight so that its comprehensive evaluation factor is greater than a certain set threshold, and thus it will be marked as a high-value learning sample.

[0169] Through the dynamic fusion mechanism of this unit, the system has achieved a transition from "spatial difference" to "time-driven", which can more accurately identify potential samples that are structurally different 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 is used to load a preset confidence scoring model. The confidence scoring model is constructed based on multiple influencing factors, including a 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;

[0172] The scoring calculation unit is used to input multiple influencing factors into the confidence scoring model and calculate the confidence score value of the current sample according to the parameter weight corresponding to each factor;

[0173] The threshold comparison unit is used to compare the confidence score value of the current sample with the preset confidence threshold. When the score is lower than the threshold, the to-be-learned label information is output and the sample is included in the to-be-learned data set; if the score is higher than the threshold, the current sample is ignored for model training.

[0174] In this embodiment of the present invention, to implement sample confidence scoring based on multi-source information and to determine whether to introduce active learning processes, the system has designed a confidence assessment submodule. This module consists of a scoring parameter loading unit, a scoring calculation unit, and a threshold comparison unit. This module jointly models three factors: features, time, and signal stability, to form a dynamically adjusted sample confidence evaluation system.

[0175] The scoring parameter loading unit is responsible for loading the system's pre-trained confidence scoring model. This scoring model is constructed using a multi-factor regression or fitting approach. 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, derived from pre-processing, represents the degree of deviation between the sample and the labeled data in both feature space and time dimensions; the time density value represents the timeliness of sample collection; and the signal fluctuation amplitude reflects the stability of the current sample's sensor signal. Applying these three factors as input to the scoring model significantly improves the representativeness of the scoring results.

[0176] The scoring calculation unit is used to input the above-mentioned multiple influencing factors into the confidence scoring model. The model weights each factor according to the weight coefficients learned during training and outputs a numerical confidence score result. The score fluctuates between 0 and 1, with lower values ​​indicating more valuable learning samples.

[0177] The threshold comparison unit receives the scoring result and compares it with the confidence threshold set in the system. If the score is below the threshold, the current sample is deemed unreliable and the system marks it as "data to be learned" and writes this marking to the training sample management list. If the score is above the threshold, the sample's features have been overwritten by the model and the current sample is not included in the update.

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

[0179] The confidence scoring model is constructed based on multiple influencing factors, including a composite evaluation factor, the time density value corresponding to the sample, and the local fluctuation amplitude of the original signal sequence. It is used to comprehensively evaluate the confidence level of the current sample, including:

[0180] This scoring model aims to provide a quantitative evaluation method for sample credibility based on a multi-factor fusion approach, assisting in determining whether a sample should be included in the active learning process. Existing techniques often use simple thresholds or fixed rules to screen samples, lacking comprehensive consideration of multi-dimensional indicators and prone to screening errors and overfitting risks. This scoring model improves the comprehensiveness and robustness of sample evaluation by incorporating multiple evaluation dimensions, including spatial, temporal, and signal fluctuations.

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

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

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

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

[0185] Scoring models can be constructed using empirical regression models, rule-based models, or simple weighted linear combination strategies. A weighted superposition model is recommended, with weight coefficients for each factor obtained through training with historically annotated samples. For example, during the research phase, 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 of the scoring model is a real value between 0 and 1, which represents the credibility of the sample.

[0186] In actual operation, the system can set multiple threshold ranges. For example, 0–0.4 indicates low confidence and requires active learning; 0.4–0.7 indicates that the observation is retained; and above 0.7 indicates high confidence and no training is required. For example, sample B has a composite evaluation factor of 0.65, a time density value of 0.91, and a local fluctuation of 0.10 (indicating relatively stable). After processing by the scoring model, it scores 0.38, and the system marks it as a sample for learning.

[0187] By introducing this scoring model, the system can make accurate judgments on sample value under a dynamic, multi-source indicator system, thereby improving the accuracy and reliability of active learning strategies and reducing training costs caused by model overfitting and sample redundancy.

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

[0189] The eigenvalue parsing unit is used to calculate the proportion of each feature in the entire feature vector based on the eigenvector of the data to be learned and its corresponding estimated deviation value, and based on the relative size of the eigenvalues ​​of each feature dimension;

[0190] The correlation adjustment unit is used to determine the relative contribution of each feature dimension to the total deviation based on the proportion information and the estimated deviation value, and generate an error contribution value. Among them, a reduction coefficient is set for feature dimensions with a large range of feature value changes but low stability during training to reduce their dominant role in parameter generation.

[0191] In this embodiment of the present invention, to address the issues of unclear parameter adjustment granularity and dimensional contributions during model updates, the system incorporates an error contribution analysis submodule. This module analyzes the structural relationship between the feature vectors of the data to be learned and its estimated deviations, extracting feature contributions dimension by dimension to form the fundamental data used to drive model optimization.

[0192] The eigenvalue parsing unit first receives the feature vectors marked as data to be learned, and simultaneously receives the estimated deviation values ​​output by the deviation calculation module. This feature vector is a multidimensional structure, with each dimension representing a normalized sensor channel value, reflecting the state of the sample across different measurement dimensions. To clarify the relative weight of each feature dimension, the unit normalizes each eigenvalue into a percentage, forming feature proportion information that indicates the contribution of that dimension to the overall feature structure.

[0193] The correlation adjustment unit calculates the relative contribution of each feature dimension to the overall deviation based on the interaction between the aforementioned feature proportion information and the estimated deviation value. The system sets a rule that if the feature value proportion of a dimension is high and consistent with the direction of deviation change, it is considered the deviation-dominant feature; otherwise, it is considered a non-dominant dimension. For dimensions with poor stability in historical training, that is, feature dimensions with high error volatility during training, the system will set a reduction coefficient to weaken their influence on the direction of parameter adjustment and prevent them from causing instability in the update path.

[0194] The final output error contribution value forms a one-dimensional vector sequence corresponding to the dimension of the feature vector, providing a clear structure, reasonable proportion and strong robust error information reference for the subsequent parameter mapping module.

[0195] Through the above structure, the system achieves a quantitative explanation of the role of each feature dimension in the deviation formation mechanism, avoids blindly allocating errors equally to all feature dimensions, and significantly improves the accuracy control capability of model updates.

[0196] The correlation adjustment unit is used to determine the relative contribution of each feature dimension to the total deviation based on the proportion information and the estimated deviation value, and generate an error contribution value. A reduction coefficient is set for feature dimensions with a large range of feature value variations but low stability during training to reduce their dominant role in parameter generation. Specifically, the following are included:

[0197] This unit is used to distribute error deviations appropriately across feature dimensions, preventing a single feature from receiving excessive weight in model updates due to occasional fluctuations. Traditionally, errors are simply averaged or distributed proportionally to the number of features. This approach ignores the actual influence of each dimension on error formation and can lead to model parameter misadjustment or overfitting risks. This unit combines the contribution of each dimension in the feature vector with the estimated deviation of the current sample to generate a more targeted error contribution structure.

[0198] In practice, the system first determines the relative weight of each dimension in the feature structure based on the percentage information output by the eigenvalue parsing unit. For example, in a four-dimensional feature vector, if the first dimension has a normalized value of 0.82 and the total vector modulus is 2.0, its percentage is approximately 41%. This percentage indicates the representativeness of that dimension in the overall sample state.

[0199] The system then combines the contribution of that dimension with the current estimated deviation. To achieve linear correspondence or nonlinear enhancement, a scaling factor can be set to multiply the contribution by the deviation to obtain a preliminary error contribution value. To prevent high-volatility 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 feature dimension has exhibited high volatility or abnormal repetition in past rounds of training, a smaller reduction coefficient (e.g., below 0.6) is added to it to weaken its dominant ability to explain the current error.

[0200] For example, the fourth dimension of sample A accounts for 18% and has a current estimated deviation of 0.3. Historical data shows that the standard deviation of this dimension has been higher than that of other dimensions over the past five rounds of training. Therefore, the system can set a reduction factor of 0.5 for this dimension. The final error contribution of this dimension is 0.18 × 0.3 × 0.5 = 0.027 (for illustrative purposes only), which is significantly lower than that of other dimensions. This reduces its involvement in the subsequent parameter mapping process.

[0201] Through the error distribution and reduction mechanism of this unit, the system can ensure that representative features obtain the dominant update status while limiting the disturbance 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 submodule includes:

[0203] The contribution level discrimination unit is used to classify the error contribution value of each feature dimension into multiple levels according to the set contribution classification standard, and generate a contribution level classification result, wherein the error contribution value of a higher level will obtain a larger parameter adjustment range;

[0204] The nonlinear mapping strategy unit is used to divide the results according to the contribution level, perform nonlinear mapping response processing on the error contribution value through the mapping function group, and generate the mapped error contribution value. Among them, compression mapping is used for low-level error contribution values ​​to suppress the model changes caused by weak disturbances, and enhancement mapping is used for high-level error contribution values ​​to accelerate model convergence.

[0205] To address the issue of error contribution values ​​being unable to be directly used for model updates, this embodiment of the present invention incorporates a parameter mapping submodule. This module performs a nonlinear transformation on the original error contribution values, enhancing the flexibility and stratification capabilities of the model update response. This module comprises a contribution level determination unit and a nonlinear mapping strategy unit, completing the logical mapping process from error to update strength.

[0206] The contribution level identification unit is responsible for classifying the error contribution values ​​for each feature dimension. Based on the defined contribution grading criteria, the system divides the contribution value range into multiple levels, such as low, medium, and high. The specific grading is determined by setting boundaries based on the mean and standard deviation of the contribution value distribution in the current training round, dynamically adjusting the corresponding threshold for each level. The grading results are output as contribution level labels to guide downstream response strategy formulation.

[0207] The nonlinear mapping strategy unit divides the results based on the contribution level and calls the corresponding mapping function to perform nonlinear response processing on the error contribution values ​​of each dimension. For feature dimensions with low contribution values, a compression mapping strategy is adopted, that is, the output value after mapping changes slowly and approaches zero to suppress the impact 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 increase the mapping output rapidly to amplify the dominant 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 goal is the same: to make the updated parameters discriminative, responsive, and stable.

[0208] This module converts the intensity of model parameter updates from "direct error drive" to "strategy adjustment drive", 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] The contribution level discrimination unit is used to classify the error contribution value of each feature dimension into multiple levels according to the set contribution classification standard, and generate a contribution level classification result. The error contribution value of a higher level will obtain a larger parameter adjustment range, specifically including:

[0210] This unit maps the raw error contribution values ​​into graded labels, providing a basis for subsequent mapping function selection and parameter adjustment. Unlike traditional models that directly use error values ​​for gradient updates, this method implements a segmented processing of the error response through a hierarchical strategy, avoiding extreme behaviors such as over-amplification of errors or neglect of weak signals.

[0211] In its implementation, the system first receives a sequence of error contribution values ​​output by the correlation adjustment unit, each corresponding to a feature dimension. The system then sets a contribution grading standard, which can be either a fixed threshold segmentation or a distributed segmentation method generated based on the dynamic distribution of samples.

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

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

[0214] Level 2 (intermediate level): Contribution value is between the average value ± 20%;

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

[0216] Of course, you can also set percentile intervals based on system experience or historical training data, such as setting the top 25% as high, the middle 50% as medium, and the rest as low. After the division, each feature dimension is marked as low, medium, or high, providing a basis for the mapping strategy unit to make decisions.

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

[0218] This hierarchical mechanism introduces a logical hierarchy to the error response, so that the model update is no longer simply responded proportionally, but has the ability to adjust according to influence, which optimizes the error adjustment efficiency and improves the response differentiation between feature dimensions.

[0219] The nonlinear mapping strategy unit is used to divide the results according to the contribution level, and performs nonlinear mapping response processing on the error contribution value through the mapping function group to generate the mapped error contribution value. The compression mapping is used for the low-level error contribution value to suppress the model changes caused by weak disturbances, and the enhancement mapping is used for the high-level error contribution value to accelerate the model convergence. Specifically, it includes:

[0220] The goal of this unit is to map the contribution level discrimination results into parameter adjustment values ​​of varying response strengths, achieving nonlinear amplification or suppression of model parameter changes due to errors. Traditional gradient updates often employ linear strategies, which can lead to unnecessary model responses to small errors or insufficient responses to large errors, impacting training efficiency. This 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 for each dimension. For low-level dimensions, the system uses a compression mapping method, mapping the original error contribution value to a smaller output value. The compression ratio can be set between 0.3 and 0.5. This ensures that these dimensions have only a slight impact on the weight adjustment, thereby shielding against noise disturbances.

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

[0223] For high-level dimensions, the system uses an enhanced mapping function that amplifies the original value by 2 times or even more than the original contribution. The specific amplification factor can be adjusted according to the system operating status. This amplification operation can be completed using table lookup, interpolation, exponential function, or empirical curve fitting. It is recommended to use a smooth curve to avoid training discontinuities.

[0224] For example, if the third dimension's contribution is 0.12, which is high, and the system sets an enhancement factor of 2.0, the output after mapping is 0.24. If the fifth dimension's contribution is 0.04, which is low, and the compression factor is 0.4, the output is 0.016. The above mapping results are directly used to update the parameter generation module and serve as the input for actual model parameter adjustment.

[0225] This unit implements the strategy of "fast response to important dimensions and slow update of secondary dimensions" at the numerical level, enabling the model to focus on high-impact directions more quickly while avoiding low-impact signals interfering with 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] ;

[0228] in, For the The original error contribution value of each feature dimension is output by the error contribution analysis submodule. It indicates the proportion of the feature dimension to the overall estimation error in the current sample;

[0229] For the The error contribution value after mapping of the feature dimension is used as an item in the updated parameter vector for model parameter adjustment.

[0230] For the The contribution level label of each feature dimension is output by the contribution level discrimination unit, with values ​​of 1 (low level), 2 (medium level), and 3 (high level);

[0231] For the The error volatility of a feature dimension over multiple historical training rounds reflects the training stability of that feature. This can be calculated by the mean error fluctuation over the last N training rounds.

[0232] For the A smoothing factor for each feature dimension, used to control the response strength of nonlinear mapping. A larger value indicates a more cautious system for that dimension. This value can be set as a system parameter or dynamically adjusted by a stability feedback mechanism.

[0233] Among them, for (Low level): Use a suppressed denominator structure to indicate that low-importance features need to suppress their error responses. is the response damping term; the response value fluctuates with the history It rises and falls quickly, forming a high volatility and low response strategy.

[0234] for (Medium level): Introduce a ratio response term to achieve neutral amplification control. The numerator is , represents the historical error offset; the denominator is , forming a gain balance structure that is sensitive to system stability; it can achieve "model stability → enhanced response, model instability → natural suppression".

[0235] for (High level): Uses nonlinear enhancement structure to ensure that strong influence features get enough model feedback. Contains linear terms and nonlinear amplification , simulating enhanced response; increasing the participation intensity of important dimensions is conducive to rapid convergence.

[0236] Example description: Taking the second dimension feature as an example, if its original error contribution value ,grade , the corresponding historical error offset factor , smoothing factor , the result after mapping is:

[0237] ; The system uses this value as the second item in the updated parameter vector for weight correction during the training process.

[0238] In a preferred embodiment of the present invention, the stability feedback submodule includes:

[0239] The error trend extraction unit is used to construct an error change curve based on the changes in the error mean and standard deviation in the current multiple update cycles;

[0240] The stability judgment unit is used to judge whether the current model is in a stable state after the update based on the slope and fluctuation frequency of the error change curve. If the fluctuation amplitude continues to increase or the slope exceeds the set critical value, it is judged that the stability standard is not met, otherwise it is met;

[0241] 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 being updated. The training adjustment signal is used to trigger the parameter mapping submodule to readjust the mapping range or reduce the model update frequency.

[0242] To ensure the stability and continuity of the weight estimation model during long-term operation, this embodiment of the present invention incorporates a stability feedback submodule, which dynamically adjusts the model update strategy based on error fluctuation trends. This module comprises an error trend extraction unit, a stability determination unit, and a feedback generation unit, establishing a cross-cycle monitoring and feedback control mechanism.

[0243] The error trend extraction unit receives output data from the accuracy assessment submodule, including the mean and standard deviation of the error after each model update. The system records error data over multiple consecutive update cycles and constructs a time series curve to reflect the model's responsiveness to data changes at different stages. This trend curve is used to detect unstable oscillations, increased overfitting, or parameter drift during model training.

[0244] The stability assessment unit analyzes the slope and frequency of the error trend curve to determine the model's state. If the slope of the curve is positive and has not converged for multiple consecutive cycles, or if the standard deviation fluctuates frequently beyond the system's set limits, the system will determine that the current model update path is unstable and trigger a self-correction mechanism.

[0245] If the feedback generation unit determines that the stability criteria are not met, it immediately outputs a training adjustment signal, which can be passed to the parameter mapping submodule 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 to reduce the impact of parameter updates on the model, thereby restoring the controllability and convergence of the system training state.

[0246] Through this module, the system can realize model status monitoring and update rhythm management without relying on human intervention, effectively improving the stability and reliability of the system's long-term operation. It is particularly suitable for industrial automatic weighing equipment that has dual requirements for accuracy and robustness.

[0247] 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 being updated. The training adjustment signal is used to trigger the parameter mapping submodule to readjust the mapping range or reduce the model update frequency, specifically including:

[0248] The core function of this unit is to take timely adaptive adjustment measures for unstable behaviors detected during model training 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 often rely on static learning rates and fixed update frequencies for parameter updates. They lack real-time monitoring and feedback response to training stability, and are prone to update oscillations or performance degradation when sample distribution changes, input disturbances 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, giving it self-stabilizing adjustment capabilities.

[0249] 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.

[0250] The system generates a set of training adjustment signals based on the degree of instability. Each set of signals consists of two sub-strategies:

[0251] The first strategy is the mapping range reset instruction, which will act on the nonlinear mapping strategy unit in the parameter mapping submodule. The specific method is: based on the degree of disturbance of the high-level error contribution dimension in the model convergence behavior during this training, the corresponding upper and lower limits of the mapping factor are redefined. For example, if the model experiences parameter oscillation after high-level mapping, the system will reduce the mapping magnification of that level from the original set value (such as 2.0) to within 1.3; if the error fluctuates concentratedly 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 through table adjustment, function scaling, or controlling the curve shape.

[0252] The second strategy involves reducing the update frequency. This directive controls the targeted training submodule within the model update module, reducing the frequency of weight adjustments. Specific methods include, but are not limited to, extending the training cycle (for example, from updating every 10 samples to every 30 samples), freezing the parameter update channels for some feature dimensions, or setting a sliding window to sparsely sample the training samples, ensuring that the source of parameter perturbations is reduced before the model stabilizes.

[0253] For example, in a warehouse automatic weighing system, if the error standard deviation in the last three rounds of model updates increases from 0.08 to 0.14, accompanied by a continued increase in the mean, the system will determine that the model has deviated from the stable range. The feedback generation unit immediately issues an adjustment signal, instructing to reduce the high-level amplification ratio of the parameter mapping from 2.0 to 1.2, extend the update interval from 10 to 25, 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 subsides.

[0254] This feedback mechanism has continuous responsiveness and multi-channel adjustment capabilities, which can effectively buffer systematic anomalies during training, improve the model's stable adaptability in a changing sample environment, and significantly reduce the risk of model performance degradation caused by overfitting or too-rapid adjustment.

[0255] Through the above structure, the feedback generation unit not only improves the self-regulation ability of the system, but also gives the model training process an elastic recovery mechanism, so that the entire weighing control system has the ability to operate stably in the long term, which is especially suitable for complex, multi-source signal-driven industrial weighing application scenarios.

[0256] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A weighing equipment control system based on active learning, characterized in that: The system comprises: The sample evaluation module is used to evaluate the sample confidence of the analysis data. Specifically, it calculates the feature distance between the historical label sample and the analysis data, and combines the timestamp and 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. The deviation calculation module is used to estimate the feature vector of the data to be learned 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; The parameter generation module is used to calculate the error contribution value of each feature dimension based on the estimated deviation value and the feature dimension weight distribution in the feature vector, and perform nonlinear mapping to generate an updated parameter vector; A model update module is used to perform incremental training on the current weight estimation model according to the updated parameter vector to obtain an updated weight estimation model; The sample evaluation module includes: The feature distance calculation submodule is used to calculate the weighted Euclidean distance between the feature vector of the analysis data and multiple feature vectors in the historical label sample to obtain the sample feature distance; The time correlation processing submodule is used to calculate the time density value based on the timestamp of the analysis data and the number of weighings per unit time, and to perform weighted fusion based on the time density value and the sample feature distance to obtain a composite evaluation factor; The confidence assessment submodule is used to obtain a sample confidence score based on 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.

2. A weighing equipment control system based on active learning according to claim 1, characterized in that: The parameter generation module includes: The error contribution analysis submodule is used to calculate the error contribution value of each feature dimension based on the correlation between the estimated deviation value and the feature vector in the data to be learned; The parameter mapping submodule is used to classify the error contribution values ​​into intervals and perform nonlinear mapping according to different mapping response modes according to different levels to obtain the mapped error contribution values; The update parameter generation submodule is used to combine the mapped error contribution values ​​in the order of feature dimensions to generate an update parameter vector.

3. The weighing equipment control system based on active learning according to claim 1, characterized in that: The model updating module includes: A directed training submodule is used to input the updated 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; The accuracy evaluation submodule is used to estimate and calculate the updated weight estimation model on the new and old sample sets respectively to obtain the corresponding error mean and standard deviation; The stability feedback submodule is used to construct an error change curve based on the error mean and standard deviation, so as to determine whether the updated weight estimation model meets the preset stability standard. If the judgment result is not satisfied, it outputs a training adjustment signal to correct the mapping response method of the subsequent update parameter vector or adjust the training frequency.

4. The weighing equipment control system based on active learning according to claim 1, characterized in that: The feature distance calculation submodule includes: A feature vector extraction unit, used to extract feature vectors constituting sample features from the analysis data; A weight factor determination unit is used to perform statistical analysis on each feature dimension in the historical label samples. By evaluating the degree of value fluctuation of each feature dimension within a preset first time window, it is determined that relatively stable features have higher weight values, while features with large fluctuations have lower weight values, thereby generating a set of weighting factors for distance calculation; The distance calculation unit is used to calculate the dimension-by-dimensional difference between the feature vector of the current analysis data and multiple feature vectors of the historical label samples, and multiply each difference by the corresponding weighting factor and then add them up to obtain the feature distance value representing the degree of sample similarity.

5. The weighing equipment control system based on active learning according to claim 4, characterized in that: The time correlation processing submodule includes: A time information extraction unit is used to extract the timestamp information of the sample from the analysis data and count the number of weighing records within the current preset second time window to obtain the weighing frequency per unit time; The time density calculation unit is used to compare the timestamp information of the sample with the current system time, calculate the time interval length of the sample, and calculate the time density value by normalizing the reciprocal of the time interval length and the weighing frequency per unit time and performing weighted combination; 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 samples collected during high-density periods on the feature distance value is enhanced, thereby forming a composite evaluation factor for representing the comprehensive evaluation level of the sample.

6. A weighing equipment control system based on active learning according to claim 5, characterized in that: The confidence assessment submodule includes: A scoring parameter loading unit is used to load a preset confidence scoring model, wherein the confidence scoring model 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 the original signal sequence; The scoring calculation unit is used to input multiple influencing factors into the confidence scoring model and calculate the confidence score value of the current sample according to the parameter weight corresponding to each factor; The threshold comparison unit is used to compare the confidence score value of the current sample with the preset confidence threshold. When the score is lower than the threshold, the to-be-learned label information is output and the sample is included in the to-be-learned data set; if the score is higher than the threshold, the current sample is ignored for model training.

7. The weighing equipment control system based on active learning according to claim 2, characterized in that: The error contribution analysis submodule includes: The eigenvalue parsing unit is used to calculate the proportion of each feature in the entire feature vector based on the eigenvector of the data to be learned and its corresponding estimated deviation value, and based on the relative size of the eigenvalues ​​of each feature dimension; The correlation adjustment unit is used to determine the relative contribution of each feature dimension to the total deviation based on the proportion information and the estimated deviation value, and generate an error contribution value. Among them, a reduction coefficient is set for feature dimensions with a large range of feature value changes but low stability during training to reduce their dominant role in parameter generation.

8. The weighing equipment control system based on active learning according to claim 7, characterized in that: The parameter mapping submodule includes: The contribution level discrimination unit is used to classify the error contribution value of each feature dimension into multiple levels according to the set contribution classification standard, and generate a contribution level classification result, wherein the error contribution value of a higher level will obtain a larger parameter adjustment range; The nonlinear mapping strategy unit is used to divide the results according to the contribution level, perform nonlinear mapping response processing on the error contribution value through the mapping function group, and generate the mapped error contribution value. Among them, compression mapping is used for low-level error contribution values ​​to suppress the model changes caused by weak disturbances, and enhancement mapping is used for high-level error contribution values ​​to accelerate model convergence.

9. The weighing equipment control system based on active learning according to claim 3, characterized in that: The stability feedback submodule includes: The error trend extraction unit is used to construct an error change curve based on the changes in the error mean and standard deviation in the current multiple update cycles; The stability judgment unit is used to judge whether the current model is in a stable state after the update based on the slope and fluctuation frequency of the error change curve. If the fluctuation amplitude continues to increase or the slope exceeds the set critical value, it is judged that the stability standard is not met, otherwise it is met; 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 being updated. The training adjustment signal is used to trigger the parameter mapping submodule to readjust the mapping range or reduce the model update frequency.

Images