Concrete mixing quality prediction method based on mixing condition self-adaption and AI model

Abstract

The invention discloses a concrete mixing quality prediction method based on mixing condition self-adaption and an AI model, and belongs to the technical field of concrete mixing quality prediction, and the method comprises the following steps: collecting batch working condition information related to a concrete mixing technology and a multi-source process signal in a mixing process, judging the working condition type of the current batch, and determining the working condition type of the current batch; determining feature normalization parameters of the current batch, performing feature extraction on the collected multi-source process signals according to process stages to obtain normalized feature expression, constructing the normalized feature expression and batch working condition information as feature input of the current batch, and inputting the feature input into an artificial intelligence model to obtain an artificial intelligence model; and obtaining a prediction result including the concrete quality index set. According to the method, the problems that the cross-working-condition generalization ability is insufficient, the dependence on calibration data is relatively strong, and the output stability and the engineering suitability need to be improved in an engineering scene with complex working condition fluctuation, process event disturbance, equipment difference and long-term drift are solved.

Description

Technical Field

[0001] This invention relates to the field of concrete mixing quality prediction technology, and in particular to a method for predicting concrete mixing quality based on adaptive mixing conditions and AI models. Background Technology

[0002] The quality of concrete mixing directly affects its workability and forming effect, especially in the production and on-site mixing of commercial concrete. Workability indicators such as slump are often used as important bases for quality control before the concrete is discharged from the mixer. Traditional quality control relies heavily on sampling and testing after mixing or manual judgment, which suffers from problems such as strong lag, lack of process visibility, and difficulty in timely correction. Therefore, how to use available process information during mixing to estimate and predict the workability of concrete online, thereby supporting quality control and process management before the concrete is discharged from the mixer, has become one of the directions of current technology.

[0003] To achieve the above objectives, existing technologies have developed online prediction schemes that utilize observable signals during the mixing process as proxy quantities for the mixing state. For example, signals such as current, voltage, and sound that can reflect equipment load or mixing rheological changes are collected. After data cleaning and preprocessing, characteristic features are extracted or data-driven models are directly established to estimate and predict indicators such as slump. The prediction results are then used for process monitoring or to provide a reference for process adjustment.

[0004] However, although the above-mentioned online prediction and adjustment schemes can estimate and intervene in slump or related properties under specific sites, specific equipment and relatively stable material systems, they may still face problems of insufficient adaptability and stability when operating in engineering sites for a long time and being promoted across sites, equipment and mix proportions. On the one hand, the mapping relationship between observable signals and working performance is not only related to the process such as mix proportion, volume, feeding sequence and mixing time, but also closely related to the mixer model, transmission system status, liner and blade wear, sensor installation position and sampling link conditions. When the equipment or site changes, the baseline level, dynamic range and noise structure of the current / acoustic signal often change accordingly. This makes it necessary to recalibrate and unify the caliber when migrating prediction methods that rely on fixed feature caliber, fixed threshold range or site-specific model library. Otherwise, problems such as prediction offset, threshold mismatch and inconsistent adjustment actions corresponding to the same judgment result on different equipment may occur.

[0005] On the other hand, engineering quality control is often not solely characterized by slump; on-site considerations also require a comprehensive approach, addressing multi-dimensional constraints such as homogeneity, segregation risk, bleeding risk, and workability stability. Furthermore, the online observable proxies for these dimensions, along with cost-effective calibration data, make it difficult to establish a unified and stable evaluation standard. This leads to situations where existing online prediction methods achieve slump predictions but multi-dimensional quality risks are not simultaneously constrained, thus affecting their usability and scalability in complex operating conditions and under varying process disturbances. In summary, while existing online prediction methods can estimate and predict performance parameters such as slump under certain conditions, they still suffer from insufficient cross-condition generalization ability, strong reliance on calibration data, and room for improvement in output stability and engineering adaptability in engineering scenarios characterized by complex operating condition fluctuations, process event disturbances, equipment differences, and long-term drift. Summary of the Invention

[0006] This invention provides a method for predicting concrete mixing quality based on adaptive mixing conditions and an AI model. The method includes: Collect batch operating condition information related to concrete mixing process and multi-source process signals during the mixing process.

[0007] Based on batch operating condition information, each candidate historical operating condition set is filtered to determine the operating condition type of the current batch. Based on the operating condition type, the feature normalization parameters of the current batch are determined from the corresponding historical operating condition normalization parameter set or the preset new operating condition normalization rules, so as to control the consistency of feature statistical caliber under different operating conditions and avoid cross-operating condition prediction deviation.

[0008] The features of the acquired multi-source process signals are extracted according to the process stage, and the features of each stage are dynamically normalized using feature normalization parameters to obtain normalized feature representations.

[0009] Normalized feature representation and batch working condition information are constructed as feature inputs for the current batch, which are then input into an artificial intelligence model to obtain prediction results including a set of concrete quality indicators.

[0010] Simultaneously, using preset update rules, the process feature vectors and detection feedback data of each batch of working conditions are obtained, and the feature normalization parameters corresponding to each batch of working conditions are recursively updated. During the recursive update process, constraints are imposed on the participation of newly added batches in the update, so as to maintain the predictive effectiveness of the method under the condition evolution.

[0011] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. The concrete mixing quality prediction method based on adaptive mixing conditions and AI model provided by this invention, through the joint acquisition of batch condition information and multi-source process signals during the concrete mixing production process, combined with dynamically screened historical condition normalized parameters, and relying on artificial intelligence model to perform multi-index quality prediction on normalized features, achieves adaptive standardization of production data and accurate prediction of mixing quality under different mix proportions, equipment models, feeding sequences, raw material batches and other conditions. Furthermore, it achieves stable prediction and synchronous risk constraint of multi-dimensional quality indicators such as concrete workability, homogeneity, segregation and bleeding risk under complex on-site conditions, equipment changes and frequent material batch switching. It solves the problems of rigid feature normalization caliber, insufficient model transfer adaptability, and difficulty in balancing single index optimization with multi-dimensional on-site quality in existing technologies, and improves the generalization ability and data availability of intelligent prediction methods in complex engineering scenarios.

[0012] 2. This invention performs stratified screening of historical production data based on batch operating condition information to construct a candidate historical operating condition set comparable to the current batch. Based on this set, the operating condition type of the current batch is determined, enabling subsequent feature processing and model prediction to be built upon data with consistent or similar operating conditions. This reduces the probability of feature scale distortion and prediction bias caused by directly mixing historical data with incomparable operating conditions, and improves the adaptability and stability of the artificial intelligence model used for mixing quality prediction across equipment, mix proportions, and long-term operating conditions.

[0013] 3. This invention determines the corresponding feature normalization parameters for the current batch working condition type, so that the process features collected under different batches, different equipment and different material conditions can be expressed and compared under a unified statistical scale. This reduces the impact of equipment differences, raw material fluctuations and sampling link changes on feature distribution, reduces the probability of cross-working condition mismatch introduced by using fixed normalization caliber or single historical parameters, improves the consistency and stability of feature input, and enhances the generalization ability and prediction reliability of artificial intelligence models under complex working conditions and long-term operation scenarios, ensuring the engineering usability of concrete mixing quality prediction results. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 A flowchart of a concrete mixing quality prediction method based on adaptive mixing conditions and an AI model provided in this application embodiment; Figure 2 This is a distribution diagram of the collapse index in the historical training samples involved in the embodiments of the present invention; Figure 3 This is a distribution diagram of the cohesion index in the historical training samples involved in the embodiments of the present invention; Figure 4 This is a distribution diagram of homogeneity index in historical training samples involved in the embodiments of the present invention; Figure 5 This is a schematic diagram comparing the prediction performance of the artificial intelligence model involved in the embodiments of the present invention on historical samples; Figure 6 This is a statistical diagram showing the percentage of prediction error ranges for the three quality indicators—slump, cohesion, and homogeneity—by the artificial intelligence model involved in this embodiment of the invention. Detailed Implementation

[0016] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0017] Embodiment 1 of the present invention: Figure 1 The diagram shows a flowchart of the concrete mixing quality prediction method based on adaptive mixing conditions and an AI model provided in this application. First, batch operating condition information related to the concrete mixing process and multi-source process signals during mixing are collected. Based on the batch operating condition information, candidate historical operating condition sets are filtered to determine the operating condition type of the current batch. Then, according to the operating condition type, feature normalization parameters for the current batch are determined from the corresponding historical operating condition normalization parameter set or a preset new operating condition normalization rule to control the consistency of feature statistical caliber under different operating conditions and avoid prediction deviations across operating conditions. Subsequently, features are extracted from the collected multi-source process signals according to the process stage, and the features of each stage are dynamically normalized using feature normalization parameters to obtain normalized feature expressions. The normalized feature expressions and batch operating condition information are used to construct the feature input for the current batch, which is then input into the artificial intelligence model to obtain prediction results including a set of concrete quality indicators. Furthermore, by using preset update rules, the method obtains the process feature vectors and detection feedback data of each batch of working conditions, recursively updates the feature normalization parameters corresponding to each batch of working conditions, and constrains the participation of newly added batches in the update process, so that the method can maintain the effectiveness and adaptability of prediction under the condition evolution.

[0018] Concrete mixing generally refers to the process of thoroughly mixing various raw materials (including cement, sand, gravel, water, admixtures, etc.) according to design requirements and process parameters, under a specific process sequence, mix proportion, and mixing time, to form a wet concrete mixture with uniform composition, good workability, and construction performance. The core purpose of mixing is to ensure that the raw materials are evenly distributed, that the paste fully coats the aggregates, and that the fluidity, cohesiveness, water retention, and other properties of the concrete paste meet the quality requirements of subsequent construction, pouring, and molding. The mixing process is usually completed using specialized concrete mixing equipment (such as a mixer).

[0019] For each batch of concrete mixing, batch operating condition information related to the batch mixing process is obtained. This batch operating condition information includes: mix proportion identification information (used to uniquely identify the raw material mix proportions used in this batch), mixing equipment identification information (used to characterize the specific equipment type and number performing the mixing operation), material feeding sequence information (used to record the actual order in which various raw materials are added), mixing volume information (i.e., the total amount of concrete fed into this batch), mixing time information (i.e., the total actual mixing time performed in this batch), raw material batch identification information (including the batch numbers of each major raw material such as cement, aggregates, and admixtures), water addition information (the actual amount of water added), admixture dosage information (the actual dosage of various admixtures), and ambient temperature information (real-time temperature data of the mixing site or the environment where the equipment is located).

[0020] Simultaneously, during the mixing process, multi-source process signals are acquired in real-time or near real-time according to the preset data acquisition cycle and synchronization mechanism in the database. These multi-source process signals specifically include: current signals (the curve of the mixer's main motor current changing over time) reflecting the operating status of the mixing equipment; voltage signals (the changes in the operating voltage of the mixing equipment); sound signals (acoustic waveform data collected by microphones installed at key locations on the equipment); and vibration signals (multi-axis vibration data collected by accelerometers installed on the equipment structure). Each of these signals should be accompanied by batch identifiers, acquisition timestamps, and other auxiliary information during acquisition to facilitate subsequent data alignment, segmentation, and process stage mapping.

[0021] To accurately identify the current concrete mixing batch's operating condition type, the first step is to extract pre-determined key operating condition fields based on the batch operating condition information already collected for the current batch. These key operating condition fields characterize the basic attributes and stability features of the mixing condition; in this example, mix proportion identification information, mixing equipment identification information, and material feeding sequence information are preferred. In specific embodiments, these key operating condition fields need to be preset based on engineering experience and historical data analysis results before implementation, and the meaning and order of the fields must remain consistent across batches. After extraction, the key operating condition fields are combined according to the preset field order to construct the key field combination corresponding to the current batch, which is used for subsequent operating condition screening and comparison.

[0022] The key operating condition fields corresponding to each historical batch are sequentially read from the historical operating condition database. Each historical operating condition in the database stores its corresponding key operating condition field information in a structured format. For each historical operating condition, a corresponding historical key field combination is constructed according to the same field order as the current batch's key field combination. This ensures that the meaning and position of fields between different batches correspond one-to-one during subsequent comparisons, avoiding comparison errors introduced due to inconsistent field order. After constructing the current batch's key field combination and each historical key field combination, a field-by-field consistency comparison is performed. Specifically, for each field in the key field combination, the current batch's field value is compared with the corresponding historical operating condition's field value to see if they are the same. In this embodiment, this consistency comparison is based on precise matching of field values, used to quickly filter out historical operating conditions that are clearly incomparable, thereby reducing the computational scale of subsequent similarity calculations.

[0023] When a field value in a given historical key field combination matches a corresponding field value in the current batch of key field combinations, the historical work condition is considered comparable to the current batch in at least some key work condition attributes, and is marked as meeting the candidate criteria. By performing the above field-by-field consistency comparison and marking operations on all historical work conditions, the historical work conditions marked as meeting the candidate criteria are collected to construct a candidate historical work condition set. This candidate historical work condition set is used for subsequent more refined work condition similarity calculations and work condition type discrimination, thereby reducing the overall computational complexity while ensuring the effectiveness of work condition screening, and improving the feasibility and stability of the method under long-term operation conditions in engineering sites.

[0024] The batch condition information of the current concrete mixing batch is read, and all batch condition fields are arranged sequentially according to a pre-defined field order to construct the condition vector for the current batch. In this batch condition information, discrete condition fields (such as mix proportion identifier, equipment identifier, and material feeding sequence) and continuous condition fields (such as volume, duration, and water addition) maintain a unified data type and format. Specifically, all discrete fields uniformly use encoded numeric types (such as integer or string types), while all continuous fields uniformly use numeric types (such as float types), ensuring that each type of field has a consistent representation format and unit within the same batch condition, thus guaranteeing comparability between different batches. The condition vectors corresponding to each candidate historical batch are retrieved from the historical condition database. The condition vectors of each historical batch are also arranged according to a pre-defined field order, forming a set of historical condition vectors with a consistent structure.

[0025] A detailed field-level comparison is performed between the current batch of operating condition vectors and each candidate historical operating condition vector. Specifically, all discrete operating condition fields are extracted to form a discrete field sequence for the current batch and a discrete field sequence for each candidate historical batch, and all continuous operating condition fields are extracted to form a continuous field sequence for the current batch and a continuous field sequence for each candidate historical batch.

[0026] The comparison of discrete operating condition fields employs a consistency matching method. Specifically, for each discrete field being compared, if the field value in the current batch is the same as the field value in the corresponding historical batch, the consistency matching result for that field is recorded as consistent; otherwise, it is recorded as inconsistent. All consistency matching results are mapped to numerical values ​​(preferably 1 for consistency and 0 for inconsistency in this example). The average of the consistency matching results for all discrete fields in the same historical batch is then taken to obtain the discrete consistency score between the current batch and the corresponding historical batch. By traversing all candidate historical operating conditions, the discrete consistency scores between the current batch and each historical batch can be obtained.

[0027] The comparison of continuous operating condition fields uses numerical distance calculation. Specifically, the absolute value of the numerical difference between each continuous field in the current batch and the corresponding continuous field in the historical batch is calculated to obtain a sequence of absolute values ​​of continuous field differences. To eliminate the influence of different fields' units and dimensions, the absolute value of each field difference is normalized. This example preferably uses a standardized normalization method. Specifically, for each continuous operating condition field, the mean and standard deviation of the field are first calculated in the historical data. The absolute value of the difference between the current batch and the corresponding field in the historical batch is subtracted from the historical mean of the field to obtain the mean difference of the field. The absolute value of the mean difference of the field is then divided by the standard deviation of the field to obtain the normalized dimensionless difference.

[0028] In other specific embodiments, for fields with extremely skewed value distributions or outliers, methods such as quantile normalization and minimum-maximum normalization can be selected to scale the absolute values ​​of all differences to the 0-1 range to adapt to different engineering data scenarios.

[0029] The average of the absolute values ​​of the normalized field differences is used as the comprehensive distance between the current batch and the corresponding historical batch. This process is repeated for all candidate historical batches to obtain the comprehensive distance between the continuous fields.

[0030] Based on the discrete consistency score and the comprehensive distance value of the continuous field obtained above, a weighted average is performed to obtain the comprehensive similarity between the current batch of working conditions and each historical working condition. Specifically, the weights for the weighted average are preset by relevant personnel based on experience and stored in the database.

[0031] For the overall similarity of all candidate historical working conditions, if the overall similarity of any historical working condition exceeds a preset similarity threshold in the database, the current batch of working conditions is determined to be an existing working condition. Based on this, the normalized parameters of that historical working condition and subsequent model processing procedures are adopted. If the overall similarity of all candidate historical working conditions does not reach the similarity threshold, the current batch of working conditions is determined to be a new working condition. This discrimination process can take into account both the categorical and numerical differences of working conditions, providing a data foundation and logical guarantee for achieving working condition adaptability and model transfer.

[0032] To achieve comparability and normalization of the characteristics of different batches of concrete mixing processes, it is necessary to adopt a differentiated normalization parameter determination strategy for each batch of working conditions, ensuring that the process of obtaining normalization parameters is feasible, reproducible, and adaptive.

[0033] After the current batch is classified into existing operating conditions based on the operating condition type, if a historical operating condition type exists in the historical operating condition database that completely matches the key operating condition fields of the current batch or meets the preset similarity conditions, the corresponding feature statistical parameters are directly retrieved from the historical operating condition database. These feature statistical parameters include, but are not limited to, historical mean, standard deviation, quantiles, and other statistical information for each stage of the features. Based on these historical statistical parameters, the process features extracted from each stage of the current batch are normalized to achieve consistency and comparability with existing historical operating condition data.

[0034] Secondly, if the current batch is identified as a new operating condition, but there are operating conditions with usable feature normalization parameters in its corresponding candidate historical operating condition set, then it is necessary to obtain the comprehensive similarity between the current batch and each candidate historical operating condition calculated in the aforementioned steps. Taking the candidate historical operating condition with the highest comprehensive similarity (i.e., the relatively highly similar operating condition) as a reference, the feature statistical parameters (such as mean and standard deviation) maintained in the historical database for that operating condition are used as the normalization parameters for the current batch. This ensures that even if the new operating condition has no direct historical records, its features can still be normalized through the statistical information of similar operating conditions, reducing initial normalization bias.

[0035] If the current batch is determined to be a new operating condition, and there are no candidate historical operating conditions with available feature normalization parameters (e.g., the first batch of extremely new operating conditions, insufficient historical data coverage, etc.), then the preset new operating condition normalization rule is adopted. The new operating condition normalization rule is typically based on engineering experience, raw material parameter ranges, and process specifications, pre-defining the allowable statistical scale and upper and lower limits of each key process feature. For example, default mean, standard deviation, maximum and minimum values ​​can be set for each feature, or industry-recommended values ​​can be used to ensure that feature values ​​can still be basically normalized even without historical data support, preventing model input imbalance due to abnormal values. The new operating condition normalization parameters can be automatically updated recursively after sufficient batch data has been accumulated, gradually evolving adaptively towards the actual statistical state.

[0036] To standardize the input of multi-source signal characteristics in the concrete mixing process, the first step is to time-align all collected multi-source process signals. Specifically, for each batch, based on the timestamp information of signal acquisition, current, voltage, sound, and vibration signals are organized in chronological order to form a complete and synchronized process signal sequence. Subsequently, based on pre-defined process stage division rules in the database, the start and end time intervals corresponding to each stage are determined, and the process signal data is segmented and mapped according to each stage.

[0037] In this example, the process stages are divided based on typical operational nodes in the entire concrete mixing process, comprising the initial feeding stage, the main mixing stage, the final mixing stage, and the discharge stage. Specifically, the initial feeding stage refers to the period from when raw materials begin to be added to the mixing equipment until all major raw materials are added. This stage begins when the mixing equipment starts and detects the first raw material signal and ends when all major materials have been added. The main mixing stage refers to the process where, after the raw materials are fully mixed, the mixing equipment runs continuously at a set speed until the peak mixing time. The start point of the main mixing is the end of the feeding process, and the end point can be determined by the mixing power, equipment current, or sound signal reaching a steady state, or by a pre-set mixing duration range. The final mixing stage refers to the period after the main mixing is completed, during which low-speed or variable-speed mixing continues according to the formula or process requirements. This stage begins at the end of the main mixing process and ends at the start of the discharge stage or the set duration of the final mixing. The discharge stage refers to the time interval from when the concrete material is discharged from the mixer to the transport equipment. The starting point of this stage is the time when the mixer starts the discharge mechanism, and the ending point is when the material is discharged and no residual signal is detected.

[0038] To achieve accurate division of process stages, this embodiment preferably uses the following signal thresholds and criteria as the basis for automated stage switching. The termination point of the initial feeding stage is based on the timestamp of the completion of all raw material input. This node can be automatically determined by monitoring the feeding sensor signals (such as the cumulative value of the weighing sensor reaching a set target). The termination point of the main mixing stage is determined by the criterion that the sliding average value of the mixing motor current signal changes less than a preset current change threshold (such as less than 1% of the rated current) within n consecutive seconds (n is preset by relevant personnel based on experience and historical data and stored in the database), or by the auxiliary criterion that the amplitude of the main frequency fluctuation of the sound signal is lower than a preset main frequency fluctuation range (such as less than 50Hz). The start and end points of the final mixing stage and the discharge stage are automatically recorded by the mixing mode switching and discharge execution signals issued by the control system, respectively.

[0039] In practical applications, the above-mentioned stage divisions can be flexibly adjusted according to different equipment models, engineering requirements, or monitoring signal characteristics. For example, for special mixing processes, it can be further subdivided into stages such as premixing, intermittent stirring, and auxiliary feeding, and the switching nodes of each stage can be automatically identified by combining real-time signal threshold changes and control command records. The process stage division rules are usually stored in a database in the form of parameter tables or rule sets, supporting batch-level adaptive calling and dynamic adjustment to ensure that the stage divisions of different batches and different process flows are consistent and reproducible.

[0040] Within each process stage, typical process features, such as peak current, steady current, dominant sound frequency, and voltage fluctuation, are calculated for the aforementioned multi-source process signals according to feature extraction rules. The resulting stage features provide a quantitative description of the signal state within that process stage. Specific calculation methods include signal window statistics, frequency domain analysis, and time-series feature extraction. After feature extraction, normalization is performed on each stage feature based on the feature normalization parameters (such as mean, standard deviation, extreme values, or quantiles) determined by the current batch's identified operating condition type. Normalization often employs standardization (Z-score) or interval scaling to transform the stage features into a dimensionless, comparable standard format, thereby eliminating the influence of differences in equipment model, material batch, and sampling link on the feature scale. The normalized stage feature data from all process stages are integrated and summarized in stage order to form a complete normalized feature expression representing the current batch's concrete mixing process state. This normalized feature expression can be directly used as the basis for subsequent model input or operating condition comparison, achieving unified and reproducible processing of feature data across different batches and operating conditions.

[0041] Specifically, this example focuses on feature extraction from multi-source process signals. The specific calculation methods are as follows: Peak current is obtained by calculating the maximum value of the current curve within each stage; steady current is obtained by taking the average value of the last N seconds of the main stirring stage (N is preset by relevant personnel based on experience and historical data and stored in the database) as a representative value; the dominant frequency of sound is obtained by performing a Fast Fourier Transform (FFT) on the sound wave signal of this stage and taking the frequency component with the largest amplitude as the dominant frequency; voltage fluctuation is obtained by calculating the absolute value of the difference between the maximum and minimum values ​​of the voltage curve within each stage; vibration amplitude is based on the root mean square of the acceleration signal within the stage.

[0042] For the current batch, acquire the characteristic data of each process stage that has undergone normalization. Following a preset feature organization order, sequentially concatenate the normalized features within each process stage to form a process feature vector reflecting the dynamic evolution of the concrete mixing process in this batch. The process feature vector may include, but is not limited to, normalized representations of current signal characteristics (such as peak current, steady current, average slope, minimum slope, etc.), voltage signal characteristics (such as voltage fluctuation, stability, etc.), sound signal characteristics (such as dominant frequency, frequency fluctuation, etc.), and vibration signal characteristics (such as vibration amplitude, vibration frequency, etc.) for each process stage.

[0043] The system synchronously reads the batch operating condition information of the current batch (such as mix ratio, equipment model, feeding sequence, mixing volume, mixing time, raw material batch, water addition, admixture dosage, ambient temperature, etc.), and performs structured and numerical processing on the above operating condition information according to the field order and encoding method consistent with the input requirements of the artificial intelligence model to obtain the operating condition feature vector.

[0044] The process feature vector and the operating condition feature vector are concatenated to construct a structured feature input vector. This feature input vector reflects both the temporal dynamic state of the mixing process in this batch and contains static constraint information of the production operating conditions, ensuring the comprehensiveness and consistency of the model input.

[0045] The aforementioned feature input vectors are fed into a pre-trained artificial intelligence model; in this example, Automated Machine Learning (AutoML) is chosen. Based on the feature-index mapping relationship learned from internal parameters and historical training samples, the AI ​​model outputs the predicted quality indicators for the current batch of concrete. The set of concrete quality indicators includes, but is not limited to, slump (characterizing workability), cohesion (characterizing bonding ability), homogeneity (reflecting mixing uniformity), segregation risk (reflecting stratification trend), and bleeding risk (reflecting the tendency for free water to precipitate). Each quality indicator is presented as a numerical or graded result, providing a data foundation for subsequent concrete quality assessment, production adjustments, and process optimization. This process is highly reproducible and engineering-applicable, and can be flexibly extended to joint prediction scenarios for other related performance indicators.

[0046] like Figure 2 The figure shows the distribution of slump index in the historical training samples involved in the embodiments of the present invention. As can be seen from the figure, the slump values ​​of the historical samples are mainly concentrated in the higher range, and the overall distribution pattern is mainly peaked at about 170mm-190mm, with a small number of low slump samples forming a long tail. The figure gives the average and median of the slump as a reference for the statistical center location, which is used to characterize the overall level of the samples and their skewness characteristics, thus indicating that the training samples have a certain coverage range in the slump dimension and are mainly in the common working condition range.

[0047] like Figure 3 The figure shows the distribution of cohesion index in the historical training samples involved in this embodiment of the invention. As can be seen from the figure, the cohesion samples are mainly distributed and continuously cover the range of approximately 1.5 MPa to 4.5 MPa, with a small number of samples in the low-value region (approximately 0.5 MPa to 1.5 MPa). Overall, the distribution exhibits a multi-peak or broad-peak characteristic, indicating that the historical data simultaneously contains samples of multiple working conditions corresponding to different cohesion states. The figure also marks the positions of the mean and median, which can be used to reflect the concentration trend and distribution skewness of the cohesion samples, supporting subsequent model training in learning and distinguishing samples at different cohesion levels.

[0048] like Figure 4The figure shows the distribution of homogeneity index in the historical training samples involved in this embodiment of the invention. As can be seen from the figure, the homogeneity index is mainly concentrated in the range of approximately 0.08-0.16, exhibiting a single-peak shape centered around approximately 0.11-0.13. Different degrees of tail samples exist on both sides of high / low homogeneity, reflecting the differences in homogeneity under different mixing stability or material states. The average and median values ​​given in the figure are used to characterize the typical level of homogeneity samples, providing a statistical basis for subsequent prediction modeling and threshold determination based on homogeneity labels.

[0049] Specifically, the steps for automated modeling and optimal model selection using the AutoML platform are as follows: Prepare a structured dataset containing multiple historical concrete production batches. Each batch of data includes batch working condition information (such as mix proportion, equipment identification, feeding sequence, mixing volume, etc.), normalized feature representation of multi-source process signals, and concrete quality inspection labels for the corresponding batch (such as slump, cohesion, homogeneity, segregation risk, bleeding risk, etc.).

[0050] After data preparation, all historical batch datasets are input into the AutoML modeling platform. The platform first automatically performs preprocessing operations on the input data, including missing value completion, outlier detection, feature encoding, and data normalization. Next, the AutoML platform automatically identifies the input feature types and, based on its internal modeling strategy, automatically generates various candidate machine learning models, including but not limited to gradient boosting trees, random forests, support vector machines, neural networks, and ensemble learning models. Simultaneously, it automatically searches and adjusts the hyperparameter space for each model. During model training, the AutoML platform employs mechanisms such as cross-validation and automatic training / validation set partitioning to evaluate the performance of each candidate model on target quality metrics (such as regression error, classification accuracy, and F1 score). Combining feature importance ranking and model interpretability metrics, it automatically selects the final model with the best predictive performance and strongest generalization ability.

[0051] After model training is complete, the AutoML platform outputs the optimal model along with its entire structure and parameters. In the final production deployment phase, the pre-trained AutoML model is directly loaded, and subsequent batches of feature input vectors can be directly fed into the model to obtain the corresponding concrete quality index prediction results. If new batches of labeled data are continuously accumulated during model operation, new data can be periodically added to the AutoML platform to trigger automatic model retraining, parameter fine-tuning, and optimization updates, ensuring that the model's prediction performance adapts to long-term changes in actual working conditions and materials.

[0052] For each completed concrete mixing batch, a unique batch-level association is established between the batch's process feature vector (i.e., a multi-stage feature splicing expression representing the entire mixing process) and the corresponding quality inspection feedback data (such as measured slump values, measured bleeding results, etc.). Using the batch's unique identifier as an index, the process feature vector and inspection feedback data are assigned to the current batch's determined working condition type and written into the working condition data subset for that type.

[0053] like Figure 5 The figure shows a comparison of the prediction performance of the artificial intelligence model involved in this invention on historical samples. The red curves represent the measured value sequence of the corresponding quality indicators, the blue scatter points represent the prediction results of the training set samples, and the green scatter points represent the prediction results of the test set samples. The horizontal axis represents the sample number, and the vertical axes represent the slump index (mm), cohesion index (MPa), and homogeneity index, respectively. As can be seen from the figure, the predicted points of each indicator in the training set and the test set closely follow the distribution of the measured value curves, and maintain good consistency across different value ranges. This indicates that the artificial intelligence model can effectively learn the mapping relationship between the characteristics of the mixing process and the quality indicators, and still has good generalization ability on test samples that were not trained. Therefore, it can be used to stably predict concrete quality indicators such as slump, cohesion, and homogeneity.

[0054] like Figure 6 The figure shows a statistical diagram illustrating the proportion of prediction error intervals for three quality indicators—slump, cohesion, and homogeneity—by the artificial intelligence model involved in this embodiment of the invention. The figure divides the prediction errors into several intervals using slump error (mm), cohesion error (MPa), and homogeneity error (coefficient) as the horizontal axis, and statistically analyzes the sample proportion of each interval (the vertical axis represents the error proportion). As can be seen from the figure, the prediction errors for each indicator are mainly concentrated in the intervals close to zero error, and the proportion of samples falling into the larger deviation intervals is low. This indicates that the artificial intelligence model can achieve relatively stable prediction accuracy across different quality indicator dimensions. Furthermore, statistically analyzing the proportion of error intervals can provide a quantitative basis for subsequent threshold setting, confidence assessment, and abnormal working condition identification.

[0055] Under the same operating condition, as new batches of data accumulate, when the number of newly added batches reaches a preset update trigger threshold in the database (e.g., 10 batches), the feature statistical status information maintained under the current operating condition is read. This feature statistical status information includes historical statistics such as mean, variance, and extreme values ​​for each feature dimension. The process feature vector of the newly added batch is used as the incremental input, and according to preset update rules (such as moving average recursion, exponentially weighted moving average, Welford algorithm, etc.), the new data is combined with the existing statistical status to obtain the updated feature statistical status. The entire update process only needs to operate on the current statistical status and the newly added data, without having to backtrack all the original features of all historical batches, thus improving computational efficiency and data management flexibility.

[0056] Based on the latest feature statistics, statistics (such as mean, standard deviation, quantiles, etc.) for normalization are re-extracted for each feature dimension, and numerical shift parameters and scaling parameters required for normalization are generated accordingly. The newly generated parameter combination is used as the current effective feature normalization parameter for this working condition type. Subsequent batches of the same working condition type will use this parameter for feature normalization processing, thereby achieving long-term adaptive adjustment of feature statistical caliber.

[0057] Specifically, the numerical shift parameter, i.e., the centralization parameter for each feature, is preferably the mean (or median) of that feature in the latest subset of operating condition data. It is calculated as follows: under the current operating condition type, accumulate all values ​​of that feature participating in the recursive update batches, calculate the arithmetic mean, and record this as the shift parameter for that feature. Alternatively, after weighting and constraining abnormal batches, only the mean of the representative sample set can be used as the parameter. For cases where some features are susceptible to extreme values, the median or quantile can be used as the central benchmark to improve robustness.

[0058] The scaling parameter, also known as the scaling factor, is used to unify the units and fluctuation range of feature values. The standard deviation, or the more robust interquartile range, is preferably used as the scaling standard. It is calculated as follows: based on all values ​​of the feature in a representative batch under the current operating condition, the standard deviation is calculated, and this standard deviation is used as the scaling parameter for the feature. If the feature distribution does not follow a normal distribution, the difference between the 75th and 25th percentiles can be used as a substitute to accommodate heavy-tailed or non-Gaussian characteristics.

[0059] To prevent abnormal or non-representative batches from adversely affecting the normalized parameters, a batch participation weight constraint mechanism is introduced during recursive updates. Specifically, the deviation of the process feature vector of a newly added batch from the maintained statistical status of the current operating condition (such as Mahalanobis distance or standardized distance) is first calculated. This deviation is compared to a preset update tolerance range. When the deviation is not within the update tolerance range, the critical value of the update tolerance range closest to the deviation is taken as the update tolerance critical value. The larger of the update tolerance critical value and the deviation value is subtracted from the smaller value to obtain the update tolerance deviation value. Using this deviation value as a lookup key, the corresponding weight reduction value is extracted from a weight reduction mapping table pre-established and stored in the database based on experience and historical data. This reduces the weight of the batch in the parameter recursion, allowing only representative or stable batches to have a major impact on the statistical parameters. This effectively improves the robustness of the normalized parameters and prevents a few abnormal samples from causing normalization scale drift.

[0060] It should be noted that the critical values ​​of the update tolerance range mentioned above refer to the upper and lower bounds of the update tolerance range.

[0061] Through the aforementioned batch-level recursive update and anomaly constraint mechanism, efficient adaptation of feature normalization parameters is achieved under complex conditions such as long-term evolution of working conditions, equipment changes, and material batch fluctuations in the concrete production process. This ensures the consistency of input features and the effectiveness of prediction for subsequent artificial intelligence models, demonstrating high engineering practicality and reproducibility.

[0062] To ensure efficient management and traceability among batches, operating conditions, and parameters, this embodiment employs a three-tiered structured database design: The first tier is the operating condition type master table, recording operating condition IDs, key field groups, normalized parameter versions, etc.; the second tier is the batch data table, where each batch data entry is bound to a unique operating condition ID and batch identifier, containing all batch operating condition fields and multi-source signal characteristics; the third tier is the normalized parameter table, where each operating condition ID is associated with normalized parameters (mean, standard deviation, quantiles, etc.), along with a parameter update timestamp and valid status marker. All tables are indexed through operating condition IDs and batch IDs, ensuring full-process traceability and version control of feature parameters, model inputs, and quality labels, supporting long-term dynamic updates and multi-model deployment requirements.

[0063] Example 2: While keeping the other steps of Example 1 unchanged, a heuristic discrimination strategy different from the weighted average is introduced for the calculation of comprehensive similarity, so as to further enhance the stability and interpretability of working condition type discrimination in complex engineering scenarios.

[0064] In one specific implementation, the calculation of comprehensive similarity employs a multi-level discrimination method. This method hierarchically sets the order of action for discrete consistency scores and continuous field comprehensive distance values, rather than directly fusing numerical values. Specifically, firstly, candidate historical work conditions are initially screened based on discrete consistency scores to determine whether they meet a preset consistency threshold. When the discrete consistency score of a candidate historical work condition is higher than the preset consistency threshold in the database, it is considered that the historical work condition has high consistency with the current batch in terms of discrete work condition attributes, and only historical work conditions that meet this condition continue to be compared with the continuous field comprehensive distance value; for historical work conditions whose discrete consistency scores do not reach the consistency threshold, their comprehensive similarity with the current batch is directly determined to be low, and they are not included in the subsequent continuous field analysis stage. Subsequently, among the historical work conditions that pass the initial screening, they are sorted according to the magnitude of the continuous field comprehensive distance value. The smaller the distance value, the closer the continuous work condition parameters are, and the higher the comprehensive similarity. Through this multi-level discrimination method, discrete fields have a priority in work condition identification, while avoiding interference from continuous field differences in the judgment of key work condition categories.

[0065] In another specific implementation, the calculation of comprehensive similarity employs a rule-based approach. This method pre-constructs a set of discrimination rule tables based on engineering experience and historical statistical results. These rule tables describe the correspondence between discrete consistency score intervals and continuous field comprehensive distance intervals. Specifically, the rule table can define various combinations of rules. For example, when the discrete consistency score reaches a preset high consistency interval and the continuous field comprehensive distance value is below a preset distance threshold, the corresponding historical work condition is classified as a highly similar work condition; when the discrete consistency score is in a medium interval and the continuous field comprehensive distance value is low, the corresponding historical work condition is classified as a medium similar work condition; and when the discrete consistency score is low or the continuous field comprehensive distance value is too large, it is classified as a low similarity work condition. By traversing the rule table, the discrete consistency scores and continuous field comprehensive distance values ​​corresponding to the current batch and each candidate historical work condition are mapped to the corresponding similarity levels, thereby completing the determination of comprehensive similarity.

[0066] By introducing the aforementioned multi-level discrimination method or rule table-driven method, the calculation process of comprehensive similarity no longer relies on a single numerical fusion method, but can make flexible judgments by combining the importance level of the working condition field and engineering experience. This improves the stability and robustness of working condition type discrimination in application scenarios with significant fluctuations in equipment, mix ratios, and working conditions.

[0067] The various features and processes described above can be used independently of each other or combined in various ways. All possible combinations and sub-combinations are intended to fall within the scope of this disclosure. Furthermore, certain methods may be omitted in some embodiments.

[0068] The various operations of the example methods described herein can be performed, at least in part, by one or more processors that are temporarily configured (e.g., by software) or permanently configured to perform the relevant operations. Whether temporarily or permanently configured, such processors can constitute the engine of a processor implementation that operates to perform one or more of the operations or functions described herein.

[0069] In this specification, multiple instances may implement components, operations, or structures described as single instances. Although individual operations of one or more methods are shown and described as separate operations, one or more of the separate operations may be performed simultaneously and do not need to be performed in the order shown. Structures and functions presented as separate components in the example configuration may be implemented as composite structures or components. Similarly, structures and functions presented as single components may be implemented as separate components. These and other variations, modifications, additions, and improvements fall within the scope of this document.

[0070] While an overview of the subject matter has been described with reference to specific example embodiments, various modifications and changes can be made to these embodiments without departing from the broader scope of embodiments of this disclosure. Such embodiments of the subject matter are referred to herein, individually or collectively, by the term "invention," and are used for convenience only and are not intended to limit the scope of this application to any single disclosure or concept, should more than one disclosure or concept be disclosed in fact.

[0071] The embodiments described herein have been described in sufficient detail to enable those skilled in the art to practice the disclosed teachings. Other embodiments may be used and derived therefrom, such that structural and logical substitutions and changes may be made without departing from the scope of this disclosure. Therefore, the detailed description should not be construed as limiting, and the scope of the various embodiments is defined only by the appended claims and the full scope of their equivalents.

Claims

1. A method for predicting concrete mixing quality based on adaptive mixing conditions and AI models, characterized in that, Includes the following steps: Collect batch operating condition information related to concrete mixing process and multi-source process signals during the mixing process; Based on batch operating condition information, each candidate historical operating condition set is filtered to determine the operating condition type of the current batch. Based on the operating condition type, the feature normalization parameters of the current batch are determined from the corresponding historical operating condition normalization parameter set or the preset new operating condition normalization rule. The features of the acquired multi-source process signals are extracted according to the process stage, and the features of each stage are dynamically normalized using the feature normalization parameters to obtain the normalized feature expression. Normalized feature representation and batch working condition information are used to construct the feature input of the current batch, which is then input into the artificial intelligence model to obtain prediction results including a set of concrete quality indicators. Simultaneously, using preset update rules, the process feature vectors and detection feedback data of each batch of working conditions are obtained, and the feature normalization parameters corresponding to each batch of working conditions are updated recursively. During the recursive update process, constraints are imposed on the participation of newly added batches in the update.

2. The method for predicting concrete mixing quality based on adaptive mixing conditions and AI models as described in claim 1, characterized in that: The acquisition includes batch operating condition information related to the concrete mixing process and multi-source process signals during the mixing process, wherein: The batch operating information includes mix proportion identification information, mixing equipment identification information, feeding sequence information, mixing volume information, mixing time information, raw material batch identification information, water addition information, admixture dosage information, and ambient temperature information. The multi-source process signals include current signals, voltage signals, sound signals, and vibration signals that reflect the operating status of the mixing equipment.

3. The method for predicting concrete mixing quality based on adaptive mixing conditions and AI models as described in claim 1, characterized in that: The specific steps for filtering each candidate historical operating condition set are as follows: Extract preset key operating condition fields from the batch operating condition information of the current batch, and construct the key field combination of the current batch; Read the key operating condition fields corresponding to each historical operating condition sequentially from the historical operating condition database, and construct a historical key field combination that is consistent with the field order of the current batch key field combination. Perform a field-by-field consistency comparison between the current batch of key field combinations and each historical key field combination. When the value of an existing field in a certain combination of historical key fields is consistent with the corresponding value of the field in the current batch of key field combinations, the historical working condition is marked as meeting the candidate condition, and the historical working conditions marked as meeting the candidate condition are constructed into a set of candidate historical working conditions.

4. The method for predicting concrete mixing quality based on adaptive mixing conditions and AI model as described in claim 1, characterized in that: The specific steps for determining the operating condition type of the current batch are as follows: Read the batch operating condition information corresponding to the current batch, and construct the batch operating condition information into the current batch operating condition vector according to the preset field order; Retrieve the operating condition vectors corresponding to each candidate historical batch from the historical operating condition database and record them as each candidate historical operating condition vector. The current batch of working condition vectors is compared with each candidate historical working condition vector at the field level to obtain the discrete field sequence of the current batch, the continuous field sequence of the current batch, the discrete field sequence of each candidate historical working condition vector, and the continuous field sequence of each candidate historical working condition vector. The field-level comparison includes performing consistency matching judgment on discrete operating condition fields and performing numerical distance calculation on continuous operating condition fields; Obtain the consistency matching results of discrete working condition fields, and then calculate the discrete consistency score between the current batch and each candidate historical working condition. Obtain the numerical distance results of the continuous working condition fields, and then calculate the comprehensive distance value of the continuous fields between the current batch and each candidate historical working condition. Based on the discrete consistency score and the continuous field comprehensive distance value, calculate the comprehensive similarity between the current batch of working conditions and each candidate historical working condition; The comprehensive similarity is judged, and when there are candidate historical working conditions with a comprehensive similarity greater than the similarity threshold, the working condition type of the current batch is determined to be an existing working condition. If there are no candidate historical working conditions with a comprehensive similarity greater than the aforementioned similarity threshold, the working condition type of the current batch will be determined as a new working condition.

5. The method for predicting concrete mixing quality based on adaptive mixing conditions and AI model as described in claim 4, characterized in that: The discrete consistency score is calculated by performing consistency matching on the discrete field sequence of the current batch and the discrete field sequence corresponding to each candidate historical working condition, and then performing numerical mapping on the consistency matching results. The comprehensive distance value of continuous fields is obtained by performing numerical difference calculation on the continuous field sequence of the current batch and the continuous field sequence corresponding to each candidate historical working condition, and then normalizing the field difference.

6. The method for predicting concrete mixing quality based on adaptive mixing conditions and AI model as described in claim 1, characterized in that: The specific analysis process for determining the feature normalization parameters of the current batch is as follows: When the current batch's operating condition type is an existing operating condition, retrieve the characteristic statistical parameters corresponding to the existing operating condition from the historical operating condition database and use them as the characteristic normalization parameters for the current batch. If the current batch of working conditions is determined to be a new working condition and there is a set of candidate historical working conditions with available feature normalization parameters, the candidate historical working condition with the highest comprehensive similarity between the current batch of working conditions and each candidate historical working condition is taken as the relatively high similarity working condition. Based on the feature normalization parameters corresponding to the relatively high similarity working condition, feature normalization parameters for the current batch are constructed. If the current batch of operating conditions is determined to be a new operating condition and there are no candidate historical operating conditions with available feature normalization parameters, feature normalization parameters for the current batch are determined based on the preset new operating condition normalization rules. The new operating condition normalization rules are used to constrain the statistical scale and value range of each process feature under the condition of lack of historical operating condition parameters.

7. The method for predicting concrete mixing quality based on adaptive mixing conditions and AI model as described in claim 1, characterized in that: The process involves extracting features from the acquired multi-source process signals according to the process stages, and then using the feature normalization parameters to dynamically normalize the features of each stage to obtain a normalized feature representation. The specific execution steps are as follows: Read the multi-source process signal data corresponding to the current batch, and perform time alignment of the multi-source process signals based on the timestamp information to form a process signal sequence arranged in chronological order; Based on the preset process stage division rules, the process signal sequence is divided into stages, the time interval corresponding to each process stage is determined, and the process signal sequence is mapped to the corresponding process stage. Within each process stage, for the corresponding multi-source process signals, the stage feature data corresponding to that process stage is calculated according to the preset feature extraction rules. Based on the feature normalization parameters already determined for the current batch, normalization processing is performed on the stage feature data corresponding to each process stage to eliminate differences in feature dimensions and scales under different operating conditions. The normalized characteristic data of each process stage are collected to form a normalized characteristic expression that represents the state of the current batch mixing process.

8. The method for predicting concrete mixing quality based on adaptive mixing conditions and AI model as described in claim 1, characterized in that: The process of constructing the normalized feature representation and batch working condition information as the feature input for the current batch, and then inputting it into the artificial intelligence model to obtain the prediction results including the set of concrete quality indicators, is as follows: Read the normalized feature expression corresponding to the current batch, and splice the normalized features corresponding to each process stage according to the preset feature organization order to form a process feature vector that represents the dynamic state of the mixing process. Read the batch operating condition information corresponding to the current batch, and encode the batch operating condition information according to the field order consistent with the input of the artificial intelligence model to obtain the operating condition feature vector; The process feature vector and the working condition feature vector are concatenated to construct the feature input vector corresponding to the current batch; The feature input vector is input into the artificial intelligence model, and the artificial intelligence model outputs the prediction result of the set of concrete quality indicators corresponding to the current batch based on its internal model parameters; The set of concrete quality indicators includes at least the slump index, cohesion index, homogeneity index, segregation risk index, and bleeding risk index of concrete.

9. The method for predicting concrete mixing quality based on adaptive mixing conditions and AI model as described in claim 1, characterized in that: The process feature vectors and detection feedback data of each batch of working conditions are obtained according to a preset update rule, and the feature normalization parameters corresponding to each batch of working conditions are updated recursively. The specific execution steps are as follows: After completing the mixing process of the current batch of concrete, the process feature vector corresponding to the current batch is obtained, and the detection feedback data corresponding to the batch is obtained simultaneously. The process feature vector and the detection feedback data are then stored in batch-level association. Based on the batch-level association, the process feature vector and the detection feedback data are mapped to the working condition type corresponding to the current batch, and written into the working condition data subset corresponding to the working condition type; When the number of new batches accumulated under the same working condition type reaches the preset update trigger threshold, the feature statistical status information maintained under the current working condition type is read. The feature statistical status information is used to characterize the historical statistical status of each feature dimension under the corresponding working condition type. The process feature vector corresponding to the new batch is used as an incremental input and introduced into the feature statistical status information. According to the preset update rules, the process feature vector of the new batch and the feature statistical status information are combined to form the updated feature statistical status information. The combination process does not backtrack the original feature data of the entire historical batch during the update process. Based on the updated feature statistical state information, statistics are extracted for each feature dimension to characterize the central location and dispersion of the feature's numerical distribution. Based on the statistics, the corresponding numerical translation parameters and scale adjustment parameters are determined for each feature dimension. The numerical translation parameter and the scale adjustment parameter are combined as the feature normalization parameter corresponding to this working condition type, which is used to perform unified scale conversion processing on the feature values ​​of subsequent batches of the same working condition type. The recursively updated feature normalization parameters are stored as the current valid feature normalization parameters for this working condition type, and used for subsequent feature normalization processing of the same working condition type in batches, so as to realize the adaptive adjustment of feature statistical caliber during long-term operation of the working condition.

10. The method for predicting concrete mixing quality based on adaptive mixing conditions and AI model as described in claim 1, characterized in that: The constraint on the participation of newly added batches in the update during the recursive update process is as follows: Based on the process feature vector corresponding to the new batch and the feature statistical status information maintained under the current working condition, calculate the degree of deviation of the process feature vector of the new batch relative to the feature statistical status information. The degree of deviation is compared with a preset update tolerance range; When the deviation exceeds the update tolerance range, the overall deviation of the feature vector of the new batch in the feature dimension is calculated, and the update weight of the new batch in the feature normalization parameter recursive update is reduced based on the overall deviation, so as to reduce the impact of abnormal or non-representative batches on the feature normalization parameter recursive update process.

Images