A cementing material production energy consumption evaluation method based on internet of things collection

Abstract

The application discloses a cementing material production energy consumption evaluation method based on Internet of Things collection, comprising a meta-learning and transfer learning fusion framework, and is characterized in that the method further comprises the following steps: S1: the meta-learning and transfer learning fusion framework first divides source domain cementing material production data into batches and classification tasks as basic units of training cementing material production data, and the cementing material production data divided in the task unit mode is used as a basic input unit of meta-learning. Through the multi-task learning paradigm of meta-learning, the division of the support set and the query set is utilized to realize rapid adaptive learning of the model under the condition of a small amount of cementing material production data samples, solve the small sample learning problem, in addition, the knowledge transfer from the source domain cementing material production data to the target domain data is realized through the transfer learning, the cross-domain learning problem is solved, and the adaptability of the model to different production environments is enhanced.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for assessing energy consumption in cementitious material production based on Internet of Things (IoT) data acquisition. Background Technology

[0002] The production process of cementitious materials is a complex industrial process involving multiple physical and chemical reactions and the combined effects of various variables, such as temperature, pressure, humidity, material ratio, and energy consumption. Changes in these variables directly affect production energy consumption and product quality, making accurate energy consumption assessment crucial. However, traditional energy consumption assessment methods rely on large numbers of labeled samples and linear assumption models, which are difficult to apply effectively when data is limited and the production environment is complex and variable. In actual industrial production, cementitious material production data has the following characteristics: first, the data scale is relatively small and unevenly distributed; second, the data characteristics are nonlinear, high-dimensional, and complex; and third, there are inter-domain differences in data across different production equipment or environments.

[0003] Chinese invention patent CN118799113A discloses a method for predicting energy consumption in tin smelting processes based on virtual sample generation and a multi-output neural network model, belonging to the field of energy consumption prediction technology for production processes. Specifically, this invention includes: preprocessing data from the tin smelting process; using a mutual information algorithm to test the correlation of input variables; constructing a multi-output neural network model; using a multi-distribution overall trend diffusion technique to generate virtual samples of the input variables; using a particle swarm optimization algorithm to filter virtual samples; and finally reconstructing the multi-output neural network energy consumption prediction model based on the mixed samples. This invention solves the problem of low energy consumption prediction accuracy caused by insufficient data quantity and incomplete data information in the tin smelting process; it improves the predictive ability of the model by integrating virtual and real samples; and it effectively expands the data for small samples, improving the prediction accuracy of the prediction model on small sample datasets.

[0004] Chinese invention patent CN118940030A discloses a method and system for assessing power consumption in factory workshops, relating to the field of energy consumption assessment and diagnosis technology. The method includes: acquiring basic workshop data and preprocessing it; extracting and combining features to generate a multivariate feature sample set; constructing an equipment energy consumption prediction model based on the multivariate feature sample set to obtain predicted energy efficiency values; calculating real-time predicted energy efficiency values ​​and performing deviation calculations to obtain assessment results. The method described in this invention forms a comprehensive multivariate feature sample set through deep feature mining and combination, thereby constructing an equipment energy consumption prediction model. This process not only improves prediction accuracy but also enables real-time monitoring of equipment energy efficiency and rapid deviation assessment, effectively optimizing energy management in the production process, achieving energy conservation and emission reduction goals, improving enterprise economic benefits, and promoting environmental sustainability.

[0005] Chinese invention patent CN119049589A discloses a data-driven method, system, and medium for analyzing, predicting, and screening the oxygen release performance of materials. The method includes: text mining of published literature data on oxygen-releasing materials to obtain structural information and oxygen release performance-related data, constructing an integrated dataset of "material structure-oxygen release performance"; building a machine learning model based on the dataset, inputting material structural features to obtain regression predictions for oxygen release performance; and obtaining structural data of potential oxygen-releasing materials from an open-source crystal database, inputting this data into the trained and optimized model to achieve high-throughput screening of oxygen-releasing materials. Compared with existing technologies, this invention mines data from published literature and constructs a mapping relationship between material structural data and oxygen release performance through a machine learning model. This approach is low-cost, highly reliable, significantly reduces the screening range of oxygen-releasing materials, and avoids the time-consuming and labor-intensive problems of traditional material experimental testing.

[0006] The above-mentioned technical solutions and existing technologies still have the following problems that need to be further addressed: 1. In the task of production energy consumption assessment, existing small-sample learning methods cannot make full use of a small number of samples to achieve effective energy consumption assessment when there is insufficient production data of cementitious materials, resulting in unstable classification performance. In addition, traditional transfer learning methods cannot fully adapt to the cross-domain characteristics of cementitious material production data, and it is difficult to effectively transfer the model trained on the source domain data to the target domain data, resulting in insufficient generalization ability; 2. In the task of production energy consumption assessment, the neural network optimization method based on gradient descent is prone to getting trapped in local optima and cannot efficiently extract complex nonlinear features in cementitious material production data, resulting in limited feature extraction ability; 3. In the task of production energy consumption assessment, traditional Euclidean space classification algorithms cannot fully capture the complex relationships of key features in cementitious material production data, reducing the accuracy and reliability of energy consumption assessment.

[0007] In summary, existing energy consumption assessment techniques struggle to achieve efficient learning under small sample conditions, fail to fully utilize cross-domain data for generalization, and struggle to capture complex nonlinear characteristics, thus affecting assessment accuracy and the practical applicability of the model. With the widespread application of IoT technology, it has become possible to collect real-time data from the production process through sensors, providing richer feature information for energy consumption assessment. This application proposes an energy consumption assessment method for cementitious material production based on IoT-based data acquisition. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a method for assessing energy consumption in cementitious material production based on Internet of Things (IoT) data collection, thus solving the problems mentioned in the background section.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] An energy consumption assessment method for cementitious material production based on IoT data acquisition includes a framework integrating meta-learning and transfer learning, and further includes the following steps:

[0011] S1: The meta-learning and transfer learning fusion framework first divides the source domain cementitious material production data into batches and classification tasks as the basic units for training cementitious material production data. The cementitious material production data is divided into tasks as units. The divided cementitious material production data is used as the basic input unit for meta-learning. Each task is further divided into a support set and a query set. The support set and query set contain various categories of cementitious material production data. Each category contains various cementitious material production data samples. The ratio of the support set to the query set is 8:2.

[0012] S2: Data on the production of source-domain cementitious materials is collected through the Internet of Things. Data collection is achieved by installing sensors on key equipment in the production line. These sensors can monitor key parameters related to the energy efficiency of the production process in real time, such as temperature, pressure, and power consumption.

[0013] S3: Data collected via IoT devices is transmitted to the central data processing center in encrypted form. In the data center, the data is first preprocessed, including data cleaning and format standardization, and then stored in a highly available data warehouse. Data storage adopts a structured query language database management system, optimized for efficient querying and analysis, and the storage format is key-value pairs to ensure fast data retrieval and efficient processing.

[0014] S4: After completing the task learning for all batches, the optimal initialization parameters are obtained. When a new task is encountered, the optimal initialization parameters are fine-tuned using the support set of the new task to obtain the parameters most suitable for the new task.

[0015] S5: A dynamic pulse optimization strategy is adopted to optimize the parameters of the neural network used for feature extraction of cementitious material production data. The dynamic pulse optimization strategy simulates the sound wave propagation and echo localization mechanism in nature. It adopts a parameter adjustment method based on pulse wave propagation and echo feedback, which is different from the traditional gradient descent method during neural network training. By simulating the propagation of pulse waves in a high-dimensional parameter space and its echo signal, the weights and biases of the neural network are dynamically adjusted to solve the local optimum problem in the gradient descent algorithm. At the same time, it balances the relationship between parameter exploration and utilization, thereby improving the robustness and efficiency of the neural network used for feature extraction of cementitious material production data.

[0016] Preferably, the meta-learning and transfer learning fusion framework includes a meta-learning module, a transfer learning module, and a classification decision fusion module. Both the meta-learning module and the transfer learning module have their own feature extraction module and feature classification module, and the two parts complete their training processes independently before the final classification decision fusion. The meta-learning module learns meta-knowledge of different tasks through a multi-task learning paradigm and uses this meta-knowledge to learn quickly in new tasks, enabling it to achieve good learning results even with a small number of cementitious material production data samples. The transfer learning module solves the small-sample and cross-domain learning problems of cementitious material production data by relying on knowledge transfer from source domain cementitious material production data to target domain cementitious material production data, requiring pre-training of the energy consumption assessment model with a large amount of labeled cementitious material production data. The pre-training of the transfer learning module and the meta-learning module uses the same source domain cementitious material production dataset and is performed simultaneously.

[0017] Preferably, the meta-learning and transfer learning fusion framework includes a pre-training stage and a fine-tuning stage. In the pre-training stage, for the feature extraction modules of transfer learning and meta-learning, batch-based cementitious material production data samples are input into the transfer learning feature extraction module, and task-based cementitious material production data samples are input into the meta-learning feature extraction module. In the fine-tuning stage, the training set of the target domain cementitious material production data is divided using batches and tasks as basic units. The training set of the target domain cementitious material production data is a set of labeled cementitious material production data within the target domain cementitious material production data. Pre-trained energy consumption assessment models of meta-learning and transfer learning are used to classify the target domain cementitious material production data, and the scores of the two energy consumption assessment model classification results are fused to obtain the final classification result of the fusion framework.

[0018] Preferably, in the meta-learning and transfer learning fusion framework, the meta-learning module adopts a weak-association meta-learning strategy for energy consumption assessment models. The learning objective of this strategy is to learn the optimal parameters that can quickly adapt to new tasks through a specific optimization algorithm. Specifically, it consists of two loop modes: an inner loop and an outer loop. The inner loop uses a base learner to extract features specific to the task, and the base learner is a neural network model. The outer loop uses a meta-learning strategy to update the initialization parameters of the base learner through gradient descent. In the meta-learning module, each category is treated as a task. The goal of meta-learning is to learn how to quickly distinguish different categories from these category tasks. The energy consumption assessment model not only needs to learn to complete the classification task itself, but also needs to learn to extract knowledge from a small number of examples to solve other unseen category tasks more efficiently. The meta-learning module maps each sample in the support set and each sample in the query set to a high-dimensional feature space as input to the neural network model. During training, several categories are randomly selected for training the energy consumption assessment model.

[0019] Preferably, the training process of the neural network based on dynamic impulse tuning in step S5 includes:

[0020] S501: Initialize a pulse wave transmitter for each neural network parameter used for feature extraction of cementitious material production data. These transmitters are able to send simulated sound waves in the parameter space. The initial state of the neural network parameter space is captured by initializing the frequency and amplitude. Cementitious material data usually has a highly nonlinear feature distribution. Initializing the weights in the form of simulated sound waves can capture data features from multiple dimensions.

[0021] S502: Perform parameter encoding for the neural network, encoding the weights and biases of the current network into the frequency and amplitude of the sound wave, so as to adjust the parameters through pulse wave propagation;

[0022] S503: Simulates the propagation process of sound waves in the parameter space of a neural network. The propagation of sound waves is affected by obstacles in the parameter space, which change the propagation path of the sound waves.

[0023] S504: The nature of the obstacles encountered by the sound wave is determined by the echo of the sound wave, so that the echo information of each neural network parameter is used to evaluate the validity of the current parameter value. There may be some key variables in the cementitious material data that have a significant impact on the production results, while other variables may be insensitive. The simulation of sound wave propagation, through the obstacle assessment method, can distinguish between key variables and minor variables, focusing on the features that are important to the production results.

[0024] S505: Demodulates echo signals to extract local environmental information of parameters, such as local gradients and extrema of neural network parameters. The demodulation process is equivalent to identifying significant patterns from complex data of cementitious material production.

[0025] S506: Based on the demodulated information, adjust the weights and biases of the neural network. The update of each parameter depends not only on its own gradient, but also on the state of other parameters, thereby realizing dynamic parameter space adjustment and optimizing the overall search strategy.

[0026] S507: Repeat the above process until the stopping condition is met, such as reaching the preset maximum number of iterations or the parameter adjustment change being lower than the preset threshold. After the neural network training is completed, use the preset Softmax function to classify the data after the neural network feature extraction to obtain the energy consumption assessment category.

[0027] Preferably, the meta-learning and transfer learning fusion framework differs from meta-learning based on weak association strategies of energy consumption assessment models. Its fused transfer learning module can uncover subtle features beneath the superficial features of cementitious material production data. The advantage of transfer learning is that it can fully utilize source domain cementitious material production data samples for pre-training the energy consumption assessment model. Specifically, a Riemann neural network energy consumption assessment model is used for transfer learning. The Riemann neural network energy consumption assessment model employs a Riemann neural network classification algorithm based on fractional derivatives. In cementitious material production, the nonlinearity, high dimensionality, and complex structure of the data make it difficult for traditional Euclidean space modeling methods to comprehensively capture the data. Based on the complex relationships between them, the Riemann neural network algorithm based on fractional derivatives utilizes non-Euclidean space to project and extract features from the cementitious material production data, enabling it to capture these complex relationships more accurately. By projecting the cementitious material production data onto points in non-Euclidean space, it further processes the cementitious material production data embedded on the Riemann manifold. Building upon the traditional Riemann neural network algorithm, fractional calculus is used to process the cementitious material production data on the Riemann manifold, allowing the network to more finely adjust weight updates during the learning process, thus improving its adaptability to complex cementitious material production data structures and its classification accuracy.

[0028] Preferably, the training process of the Riemann neural network algorithm based on fractional derivatives includes:

[0029] S5501: Initialize the parameters of the Riemann neural network;

[0030] S5502: In the forward propagation stage, the feature vector of the input cementitious material data is transformed into points on the Riemann manifold through the network layers. Each layer performs corresponding manifold operations to capture the nonlinear characteristics of the data, and the Sigmoid activation function is used to enable the model to simulate the nonlinear changes in cementitious material production.

[0031] S5503: Based on the cementitious material production data points on the Riemann manifold, the network loss is calculated using the cross-entropy loss function based on the regularization term to ensure that the model can accurately classify the cementitious material characteristics.

[0032] S5504: Fractional derivatives are used to calculate the gradient of network parameters to adapt to the distribution characteristics of cementitious material production data on the manifold. The distribution of cementitious material data on Riemannian manifolds often has complex geometric characteristics. Fractional derivatives capture these geometric characteristics more delicately in weight updates, making the learning process more sensitive to complex data structures.

[0033] Preferably, the features in the cementitious material data are often highly correlated, and dynamically updating the weights can gradually optimize the representation of these correlated features, and update the network parameters based on the gradient calculated by backpropagation.

[0034] Preferably, after the pre-training is completed, the parameters of the two feature extraction networks are fixed and transferred to the target domain cementitious material production data for fine-tuning. The cementitious material production data of the target domain is divided into a support set and a query set according to the meta-learning strategy. The energy consumption evaluation model of transfer learning is fine-tuned using the cementitious material production data of the support set. The fine-tuning strategy can be the gradient descent method.

[0035] After fine-tuning the energy consumption assessment model, the query set of cementitious material production data in the target domain is input into the neural network model for feature extraction of cementitious material production data and the Riemann neural network energy consumption assessment model, respectively, to obtain their corresponding prediction scores. Then, the Softmax function is used to normalize the prediction scores. Finally, the two prediction scores are merged and output as the final prediction result. The classification category is the category corresponding to the maximum prediction score.

[0036] Preferably, the collected data attributes may be as follows: Ta represents temperature value, recording the temperature of the equipment operating environment; Pa represents pressure value, recording the working pressure during the production process; Va represents speed value, recording the equipment operating speed; Ca represents power consumption, recording the power consumed per unit time; Ha represents humidity value, recording the ambient humidity; Wa represents weight value, recording the weight of the material; Ea represents energy efficiency ratio, recording the energy consumed per unit of product; Qa represents quality score, recording the product quality control score; Sa represents safety index, recording production safety-related indicators; Da represents duration value, recording the time required to produce one batch. In practical applications, the data attributes are usually more than 10, and the number of data attributes may reach dozens or even hundreds. Furthermore, the collected data is manually labeled, and the labeling categories include: low energy consumption, medium energy consumption, and high energy consumption. The preset stopping iteration condition is reaching a preset maximum number of iterations; preferably, the preset maximum number of iterations is set to 1000.

[0037] Compared with the prior art, the beneficial effects of the present invention are:

[0038] 1. This method for assessing energy consumption in cementitious material production based on IoT data collection utilizes a meta-learning multi-task learning paradigm. By dividing the support set and query set, it enables rapid adaptive learning of the model with a small number of cementitious material production data samples, solving the problem of small sample learning. Furthermore, it addresses cross-domain learning by transferring knowledge from source domain cementitious material production data to target domain data through transfer learning, thereby enhancing the model's adaptability to different production environments.

[0039] 2. This method for assessing energy consumption in cementitious material production based on Internet of Things (IoT) data acquisition avoids the local optima problem of traditional gradient descent by simulating sound wave propagation and echo mechanisms and dynamically adjusting the weights and biases of the neural network, thereby improving the feature extraction effect of cementitious material production data. Furthermore, it adopts a parameter optimization mechanism based on sound wave propagation to capture the global characteristics of the data and improve the training efficiency of the model when processing complex nonlinear and high-dimensional production data.

[0040] 3. This method for assessing energy consumption in cementitious material production based on Internet of Things (IoT) data projection onto a non-Euclidean space, utilizing Riemannian manifolds to capture complex nonlinear relationships between data, enhances the model's ability to capture key features in the production process, and employs fractional calculus to optimize weight updates, thereby improving the model's adaptability to high-dimensional, nonlinear cementitious material production data and enhancing classification accuracy. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the method of the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] Example 1: Refer to Figure 1 A method for assessing energy consumption in cementitious material production based on IoT data acquisition, comprising a framework integrating meta-learning and transfer learning, characterized by further including the following steps:

[0044] S1: The meta-learning and transfer learning fusion framework first divides the source domain cementitious material production data into batches and classification tasks as the basic units for training cementitious material production data. The cementitious material production data is divided into tasks as units. The divided cementitious material production data is used as the basic input unit for meta-learning. Each task is further divided into a support set and a query set. The support set and query set contain various categories of cementitious material production data. Each category contains various cementitious material production data samples, and the ratio of the support set to the query set is 8:2.

[0045] S2: Data on the production of cementitious materials in the source region is collected through the Internet of Things. Data collection is achieved by installing sensors on key equipment in the production line. The sensors can monitor key parameters related to the energy efficiency of the production process in real time, such as temperature, pressure, and power consumption.

[0046] S3: Data collected via IoT devices is transmitted to the central data processing center in encrypted form. In the data center, the data is first preprocessed, including data cleaning and format standardization, and then stored in a highly available data warehouse. Data storage adopts a structured query language database management system, optimized for efficient querying and analysis, and the storage format is key-value pairs to ensure fast data retrieval and efficient processing.

[0047] S4: After completing the task learning for all batches, the optimal initialization parameters are obtained. When a new task is encountered, the optimal initialization parameters are fine-tuned using the support set of the new task to obtain the parameters most suitable for the new task.

[0048] S5: A dynamic pulse optimization strategy is adopted to optimize the parameters of the neural network used for feature extraction of cementitious material production data. The dynamic pulse optimization strategy simulates the sound wave propagation and echo localization mechanism in nature. It adopts a parameter adjustment method based on pulse wave propagation and echo feedback, which is different from the traditional gradient descent method during neural network training. By simulating the propagation of pulse waves in high-dimensional parameter space and its echo signal, the weights and biases of the neural network are dynamically adjusted to solve the local optimum problem in the gradient descent algorithm. At the same time, it balances the relationship between parameter exploration and utilization, thereby improving the robustness and efficiency of the neural network used for feature extraction of cementitious material production data.

[0049] In this invention, the meta-learning and transfer learning fusion framework includes a meta-learning module, a transfer learning module, and a classification decision fusion module. Both the meta-learning and transfer learning modules have their own feature extraction and feature classification modules, which are trained independently before the final classification decision fusion. The meta-learning module learns meta-knowledge of different tasks through a multi-task learning paradigm and uses this meta-knowledge to learn quickly in new tasks, enabling it to achieve good learning results even with a small number of cementitious material production data samples. The transfer learning module solves the small-sample and cross-domain learning problems of cementitious material production data by relying on knowledge transfer from source domain cementitious material production data to target domain cementitious material production data. It requires pre-training the energy consumption assessment model with a large amount of labeled cementitious material production data. The transfer learning module and the meta-learning module use the same source domain cementitious material production dataset for pre-training and are performed simultaneously.

[0050] In this invention, the meta-learning and transfer learning fusion framework includes a pre-training stage and a fine-tuning stage. In the pre-training stage, the feature extraction modules for transfer learning and meta-learning are used. Data samples of cementitious material production, trained on a batch basis, are input into the transfer learning feature extraction module, while data samples of cementitious material production, trained on a task basis, are input into the meta-learning feature extraction module. In the fine-tuning stage, the training set of the target domain cementitious material production data is divided using batches and tasks as basic units. The training set of the target domain cementitious material production data is a set of labeled cementitious material production data within the target domain cementitious material production data. Pre-trained energy consumption assessment models based on meta-learning and transfer learning are used to classify the target domain cementitious material production data. The scores of the two energy consumption assessment models are then fused to obtain the final classification result of the fusion framework.

[0051] In this invention, within the framework of meta-learning and transfer learning, the meta-learning module employs a weakly correlated meta-learning strategy based on an energy consumption assessment model. The learning objective of this strategy is to learn optimal parameters that can quickly adapt to new tasks using a specific optimization algorithm. Specifically, it consists of two loops: an inner loop and an outer loop. The inner loop uses a base learner to extract features specific to the task; the base learner is a neural network model. The outer loop uses a meta-learning strategy to update the initialization parameters of the base learner through gradient descent. In the meta-learning module, each category is considered a task. The goal of meta-learning is to learn how to quickly distinguish different categories from these category tasks. The energy consumption assessment model must not only learn to complete the classification task itself but also learn to extract knowledge from a small number of examples to more efficiently solve other unseen category tasks. The meta-learning module maps each sample in the support set and each sample in the query set to a high-dimensional feature space, which serves as the input to the neural network model. During training, several categories are randomly selected for training the energy consumption assessment model.

[0052] In this invention, the training process of the neural network based on dynamic impulse tuning in step S5 includes:

[0053] S501: Initialize a pulse wave transmitter for each neural network parameter used for feature extraction from cementitious material production data. These transmitters are capable of sending simulated sound waves in the parameter space. The initial state of the neural network parameter space is captured by initializing the frequency and amplitude. Cementitious material data typically has a highly nonlinear feature distribution. Initializing the weights to the form of simulated sound waves (frequency and amplitude) can capture data features from multiple dimensions. The initial characteristics of the sound waves are defined as follows:

[0054]

[0055] In the formula, f p A is the frequency of the sound wave; p W represents the amplitude of the sound wave. pThese are the weight parameters of the neural network; σ p The standard deviation for initializing the weight parameters of the neural network is preferably set to 0.1.

[0056] S502: Perform parameter encoding for the neural network, encoding the current network's weights and biases as the frequency and amplitude of a sound wave, so that the parameters can be adjusted through pulse wave propagation. The encoding process is represented as follows:

[0057]

[0058] In the formula, α p First encoding learning rate, β p The second encoding learning rate is preferably set to α. p β is 0.05. p It is 0.03; Δf p It is the frequency change adjusted based on network feedback; ΔA p The amplitude change is adjusted based on network feedback;

[0059] It is an adaptive amplitude modulation factor; The frequency of the encoded sound wave;

[0060] This represents the amplitude of the encoded sound wave.

[0061] Furthermore, feature extraction in cementitious material production may involve a large amount of historical data (such as production process history and raw material quality changes). The adaptive amplitude modulation factor depends on the mean and variance, dynamically adjusting the exploration range to enable the model to focus on the long-term stability of the production process. The adaptive amplitude modulation factor uses variance and mean to capture the trend and stability of amplitude changes, relying not only on current gradient information but also on exploring and utilizing the relationship of historical performance balance parameters. The calculation method is expressed as follows:

[0062]

[0063] In the formula, Ac p It is the set of amplitudes from previous iterations; ρ p is the amplitude adjustment factor, preferably set to 0.1; Var() is the variance function; Mean() is the mean function.

[0064] Furthermore, the calculation method for the frequency change based on network feedback adjustment is expressed as follows:

[0065]

[0066] In the formula, L p It is the loss function of the neural network used for feature extraction from cementitious material production data; λp To adjust the frequency sensitivity, in the feature extraction of cementitious materials, the dynamic changes in production data (such as real-time temperature and humidity fluctuations) may cause the objective function (such as prediction error) to change rapidly. The frequency adjustment sensitivity provides a fine-tuning mechanism, enabling the algorithm to quickly adapt to the dynamic changes in feature distribution. Preferably, it is set to 0.01.

[0067] S503: Simulates the propagation of sound waves in the parameter space of a neural network. The propagation of sound waves is affected by obstacles in the parameter space (such as local optima in the neural network parameter space), which alter the propagation path of the sound waves. The simulation of sound wave propagation in the parameter space can be represented as:

[0068]

[0069] In the formula, int(t) represents the current iteration number; S p () represents the propagation function, which simulates the dynamic changes of parameters in different directions through sound wave propagation; φ p This is the initial phase, assumed to be 0;

[0070] S504: The nature of obstacles encountered by sound waves is determined based on their echoes (i.e., reflected sound waves). This allows the echo information of each neural network parameter to be used to evaluate the validity of the current parameter value. Cementitious material data may contain some key variables that significantly impact production results (e.g., the water-cement ratio's effect on concrete strength), while other variables may be insensitive (e.g., minor temperature fluctuations). The simulation of sound wave propagation, through obstacle assessment methods, can distinguish between key and secondary variables, focusing on features important to production results. Echo reception processing can be simulated using the following model:

[0071]

[0072] In the formula, γ p For the echo reception intensity, preferably, it is set to 0.9; E p ( ) represents the echo reception function; This is the standard deviation of the time delay, preferably set to 1;

[0073] τ p For the echo reception time delay, dτ p This represents the derivative with respect to the echo reception time delay.

[0074] Furthermore, based on the propagation effect of sound waves in complex environments, the distribution of echo reception time delay is affected not only by the standard deviation but also by the time delay caused by environmental noise. Data from the cementitious material production process is often accompanied by noise (such as limitations in equipment sensor accuracy). By filtering out the noise and extracting reliable parameter information, the calculation method is expressed as follows:

[0075] τ p =ω p ·t+κ p randn(0,0.01)

[0076] In the formula, ω p The propagation speed factor is preferably set to 0.95; κ p The noise level factor is preferably set to 0.05; randn(0, 0.01) represents a normal distribution with a mean of 0 and a variance of 0.01.

[0077] S505: Demodulate the echo signal to extract local environmental information of parameters, such as local gradients and extrema of neural network parameters. The demodulation process is equivalent to identifying significant patterns (such as the nonlinear relationship between material proportions and compressive strength) from the complex data of cementitious material production, expressed as:

[0078]

[0079] In the formula, η p The learning rate for parameter updates is preferably set to 0.01; ΔW p Adjust the increments for the neural network parameters;

[0080] S506: Based on the demodulated information, adjust the weights and biases of the neural network. The update of each parameter depends not only on its own gradient but also on the state of other parameters, thereby achieving dynamic parameter space adjustment and optimizing the overall search strategy. The adjustment method is expressed as follows:

[0081]

[0082] In the formula, The adjusted neural network weight parameters will be used as the neural network weight parameters for the next iteration.

[0083] S507: Repeat the above process until the stopping condition is met, such as reaching the preset maximum number of iterations or the parameter adjustment change being lower than the preset threshold. After the neural network training is completed, use the preset Softmax function to classify the data after the neural network feature extraction to obtain the energy consumption assessment category.

[0084] In this invention, the meta-learning and transfer learning fusion framework differs from meta-learning based on weak association strategies of energy consumption assessment models. Its fused transfer learning module can uncover subtle features beneath the superficial characteristics of cementitious material production data. The advantage of transfer learning is its ability to fully utilize source domain cementitious material production data samples for pre-training the energy consumption assessment model. Specifically, a Riemann neural network energy consumption assessment model is used for transfer learning. This model employs a Riemann neural network classification algorithm based on fractional derivatives. In cementitious material production, the nonlinearity, high dimensionality, and complex structure of the data (e.g., the influence of raw material ratios on strength and energy consumption) make traditional Euclidean space modeling methods very difficult. It is difficult to fully capture the complex relationships between data. Based on fractional derivatives, the Riemann neural network algorithm uses non-Euclidean space (Riemann manifold) to project and extract features from cementitious material production data, which can capture these complex relationships more accurately. By projecting cementitious material production data onto points in non-Euclidean space, the data embedded in the Riemann manifold is further processed. On the basis of the traditional Riemann neural network algorithm, fractional calculus is used to process cementitious material production data on the Riemann manifold, which allows the network to adjust the weight updates during the learning process more finely, improving the adaptability and classification accuracy of complex cementitious material production data structures.

[0085] In this invention, the training process of the Riemann neural network algorithm based on fractional derivatives includes:

[0086] S5501: Initialize the parameters of the Riemann neural network. The initialization method is expressed as follows:

[0087]

[0088] In the formula, The initial values ​​of the network weights are represented by a standard normal distribution. This represents a normal distribution with a mean of 0 and a variance of 1; n qin It represents the number of nodes in the input layer, reflecting the dimensions of cementitious material production data, such as temperature, humidity, and raw material ratio.

[0089] S5502: In the forward propagation stage, the input cementitious material data feature vector is transformed into points on a Riemannian manifold through network layers. Each layer performs corresponding manifold operations to capture the nonlinear characteristics of the data. The Sigmoid activation function is used to enable the model to simulate nonlinear changes in cementitious material production (such as the nonlinear effect of mix proportion changes on material strength). The forward propagation method is represented as follows:

[0090]

[0091] In the formula, It is the output of the l-th layer of the Riemann neural network. It is the output of the (l+1)th layer of the Riemann neural network. and These are the weights and biases of the l-th layer of the Riemann neural network, respectively, and Sig() is the Sigmoid non-linear activation function.

[0092] S5503: Based on the cementitious material production data points on the Riemannian manifold, the network loss is calculated using a cross-entropy loss function based on the regularization term to ensure that the model can accurately classify the characteristics of cementitious materials (such as energy consumption being classified as low, medium, and high), as follows:

[0093]

[0094] In the formula, L q It is the loss function, y qc It is the one-hot encoding of the sample's true label. These are the probabilities predicted by the Riemann neural network model, where C is the total number of categories, and R is... q (W) is the regularization term.

[0095] Furthermore, The calculation method is expressed as follows:

[0096]

[0097] In the formula, z qc The model's raw output score for class c is transformed into predicted probabilities using the Softmax function.

[0098] Furthermore, the regularization term prevents the model from overfitting to specific gelling data, making the model more generalizable when dealing with complex production data. The calculation method is expressed as follows:

[0099]

[0100] In the formula, λ q (t) is the regularization coefficient that is dynamically adjusted as the training progresses. This represents the sum of squares of the weights in the Riemannian neural network.

[0101] In one embodiment, the regularization coefficient λ q The adjustment of (t) is based on the current overfitting status of the model, and the adjustment method is expressed as follows:

[0102]

[0103] In the formula, β, k, and t0 are hyperparameters that control the height, steepness, and midpoint position of the curve, respectively. Preferably, β, k, and t0 are set to 1, 3, and 0.5, respectively.

[0104] S5504: Fractional derivatives are used to calculate the gradients of network parameters to adapt to the distribution characteristics of cementitious material production data on the manifold. The distribution of cementitious material data on Riemannian manifolds often has complex geometric characteristics. Fractional derivatives capture these geometric characteristics more delicately in weight updates, making the learning process more sensitive to complex data structures. The derivative method is expressed as follows:

[0105]

[0106] In the formula, ΔW q It is the update amount of the Riemann neural network weights, η lm It is the learning rate of the Riemann neural network. This indicates the loss function L for the Riemann neural network. q Perform a partial derivative of order 0.5. Preferably, η lm Set it to 0.01; repeat the above steps until the preset stopping iteration condition is met, which means that the Riemann neural network model training is complete.

[0107] In this invention, the features in cementitious material data are often highly correlated (such as the combined effect of proportion and temperature on material properties). Dynamically updating the weights can progressively optimize the representation of these correlated features. The network parameters are updated based on the gradient calculated by backpropagation, and the update method is expressed as follows:

[0108]

[0109] In the formula, These are the weights updated by the Riemann neural network.

[0110] In this invention, after pre-training, the parameters of the two feature extraction networks are fixed and transferred to the target domain cementitious material production data for fine-tuning. The cementitious material production data of the target domain is divided into a support set and a query set according to the meta-learning strategy. The energy consumption evaluation model of transfer learning is fine-tuned using the cementitious material production data of the support set. The fine-tuning strategy can be the gradient descent method.

[0111] After fine-tuning the energy consumption assessment model, the query set of cementitious material production data in the target domain is input into the neural network model for feature extraction of cementitious material production data and the Riemann neural network energy consumption assessment model, respectively, to obtain their corresponding prediction scores. Then, the Softmax function is used to normalize the prediction scores. Finally, the two prediction scores are merged and output as the final prediction result. The classification category is the category corresponding to the maximum prediction score.

[0112] Example 2: Refer to Figure 1The collected data attributes can be as follows: Ta represents temperature value, recording the temperature of the equipment operating environment; Pa represents pressure value, recording the working pressure during the production process; Va represents speed value, recording the equipment operating speed; Ca represents power consumption, recording the power consumed per unit time; Ha represents humidity value, recording the ambient humidity; Wa represents weight value, recording the weight of the material; Ea represents energy efficiency ratio, recording the energy consumed per unit of product; Qa represents quality score, recording the product quality control score; Sa represents safety index, recording production safety-related indicators; Da represents duration value, recording the time required to produce one batch. In practical applications, the data attributes are usually more than 10, and the number of data attributes may reach dozens or even hundreds. The collected data is manually labeled, and the labeling categories include: low energy consumption, medium energy consumption, and high energy consumption. The preset stopping iteration condition is reaching the preset maximum number of iterations, preferably, the preset maximum number of iterations is set to 1000.

[0113] It should be noted that, in this document, terms such as “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0114] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for assessing energy consumption in cementitious material production based on IoT data acquisition, comprising a framework integrating meta-learning and transfer learning, characterized in that, It also includes the following steps: S1: The meta-learning and transfer learning fusion framework first divides the source domain cementitious material production data into batches and classification tasks as the basic units for training cementitious material production data. The cementitious material production data is divided into tasks as units. The divided cementitious material production data is used as the basic input unit for meta-learning. Each task is further divided into a support set and a query set. The support set and query set contain various categories of cementitious material production data. Each category contains various cementitious material production data samples. The ratio of the support set to the query set is 8:

2. S2: Data on the production of source-domain cementitious materials is collected through the Internet of Things, and data collection is achieved by installing sensors on key equipment in the production line; S3: Data collected via IoT devices is transmitted to the central data processing center in encrypted form. In the data center, the data is first preprocessed, including data cleaning and format standardization, and then stored in a highly available data warehouse. Data storage is managed using a structured query language database management system. S4: After completing the task learning for all batches, the optimal initialization parameters are obtained. When a new task is encountered, the optimal initialization parameters are fine-tuned using the support set of the new task to obtain the parameters most suitable for the new task. S5: The parameters of the neural network used for feature extraction of cementitious material production data are optimized by adopting a dynamic pulse tuning strategy. The dynamic pulse tuning strategy simulates the sound wave propagation and echo localization mechanism in nature and adopts a parameter adjustment method based on pulse wave propagation and echo feedback. It is different from the traditional gradient descent method during neural network training. By simulating the propagation of pulse waves in high-dimensional parameter space and its echo signal, the weights and biases of the neural network are dynamically adjusted to solve the local optimum problem in the gradient descent algorithm. The proposed meta-learning and transfer learning fusion framework differs from meta-learning based on weak association strategies in energy consumption assessment models. Its fused transfer learning module mines subtle features beneath the surface characteristics of cementitious material production data. The advantage of transfer learning is its ability to fully utilize source domain cementitious material production data samples for pre-training the energy consumption assessment model. Specifically, a Riemann neural network energy consumption assessment model is used for transfer learning. This model employs a Riemann neural network classification algorithm based on fractional derivatives. In cementitious material production, the nonlinearity, high dimensionality, and complex structure of the data make it difficult for traditional Euclidean space modeling methods to comprehensively capture the data. To address the complex relationships between cementitious materials, the Riemann neural network algorithm based on fractional derivatives utilizes non-Euclidean space for projection and feature extraction of the cementitious material production data, enabling more accurate capture of these complex relationships. Projecting the cementitious material production data onto points in non-Euclidean space further processes the data embedded in the Riemann manifold. Building upon the traditional Riemann neural network algorithm, the use of fractional derivatives to process the cementitious material production data on the Riemann manifold allows for more precise adjustment of weight updates during the learning process, improving adaptability to complex cementitious material production data structures and classification accuracy.

2. The method for assessing energy consumption in cementitious material production based on IoT data collection as described in claim 1, characterized in that, The meta-learning and transfer learning fusion framework includes a meta-learning module, a transfer learning module, and a classification decision fusion module. The meta-learning module and the transfer learning module each have their own feature extraction module and feature classification module. The two parts complete the training process independently before the final classification decision fusion. The meta-learning module learns meta-knowledge of different tasks through a multi-task learning paradigm and uses this meta-knowledge to learn quickly in new tasks, enabling it to achieve good learning results with a small number of cementitious material production data samples. The transfer learning module solves the problem of small sample and cross-domain learning of cementitious material production data by relying on knowledge transfer from source domain cementitious material production data to target domain cementitious material production data. It requires a large amount of labeled cementitious material production data to pre-train the energy consumption assessment model. The transfer learning module and the meta-learning module are pre-trained using the same source domain cementitious material production dataset and are performed simultaneously.

3. The method for assessing energy consumption in cementitious material production based on IoT data collection as described in claim 1, characterized in that, The meta-learning and transfer learning fusion framework includes a pre-training stage and a fine-tuning stage. In the pre-training stage, the feature extraction modules for transfer learning and meta-learning are used. Data samples of cementitious material production with batch as training unit are input into the transfer learning feature extraction module, and data samples of cementitious material production with task as training unit are input into the meta-learning feature extraction module. In the fine-tuning phase, the training set of cementitious material production data in the target domain is divided into batches and tasks as basic units. The training set of cementitious material production data in the target domain is a set of cementitious material production data with labels in the target domain cementitious material production data. Pre-trained energy consumption assessment models of meta-learning and transfer learning are used to classify the cementitious material production data in the target domain. The scores of the classification results of the two energy consumption assessment models are fused and used as the final classification result of the fusion framework.

4. The method for assessing energy consumption in cementitious material production based on IoT data collection according to claim 2, characterized in that, In the proposed framework for fusing meta-learning and transfer learning, the meta-learning module employs a weakly correlated meta-learning strategy based on an energy consumption assessment model. The learning objective of this strategy is to learn optimal parameters that can quickly adapt to new tasks using a specific optimization algorithm. Specifically, it consists of two loops: an inner loop and an outer loop. The inner loop uses a base learner (a neural network model) to extract features specific to the task. The outer loop uses a meta-learning strategy to update the initialization parameters of the base learner through gradient descent. In the meta-learning module, each category is considered a task. The goal of meta-learning is to learn how to quickly distinguish between different categories from these category tasks. The energy consumption assessment model must not only learn to complete the classification task itself but also learn to extract knowledge from a small number of examples to more efficiently solve other unseen category tasks. The meta-learning module maps each sample in the support set and each sample in the query set to a high-dimensional feature space as input to the neural network model. During training, several categories are randomly selected for training the energy consumption assessment model.

5. The method for assessing energy consumption in cementitious material production based on IoT data collection according to claim 1, characterized in that, The training process of the neural network based on dynamic impulse tuning in step S5 includes: S501: Initialize a pulse wave transmitter for each neural network parameter used for feature extraction of cementitious material production data. These transmitters are able to send simulated sound waves in the parameter space. The initial state of the neural network parameter space is captured by initializing the frequency and amplitude. Cementitious material data has a highly nonlinear feature distribution. The weights are initialized in the form of simulated sound waves to capture data features from multiple dimensions. S502: Perform parameter encoding for the neural network, encoding the weights and biases of the current network into the frequency and amplitude of the sound wave, so as to adjust the parameters through pulse wave propagation; S503: Simulates the propagation process of sound waves in the parameter space of a neural network. The propagation of sound waves is affected by obstacles in the parameter space. S504: The nature of the obstacles encountered by the sound wave is determined by the echo of the sound wave, so that the echo information of each neural network parameter is used to evaluate the effectiveness of the current parameter value. There are some key variables in the cementitious material data that have a significant impact on the production results, while other variables are not sensitive. The simulation of sound wave propagation, through the obstacle assessment method, can distinguish between key variables and minor variables, focusing on the features that are important to the production results. S505: Demodulate the echo signal to extract local environmental information of the parameters, including the local gradient and extremum information of the neural network parameters. The demodulation process is equivalent to identifying significant patterns from the complex data of cementitious material production. S506: Based on the demodulated information, adjust the weights and biases of the neural network. The update of each parameter depends not only on its own gradient, but also on the state of other parameters, thereby realizing dynamic parameter space adjustment and optimizing the overall search strategy. S507: Repeat the above process until the stopping conditions are met. The stopping conditions include reaching the preset maximum number of iterations or the parameter adjustment change being lower than the preset threshold. After the neural network training is completed, the preset Softmax function is used to classify the data after the neural network feature extraction to obtain the energy consumption assessment category.

6. The method for assessing energy consumption in cementitious material production based on IoT data collection according to claim 5, characterized in that, The training process of the Riemann neural network algorithm based on fractional derivatives includes: S5501: Initialize the parameters of the Riemann neural network; S5502: In the forward propagation stage, the feature vector of the input cementitious material data is transformed into points on the Riemann manifold through the network layers. Each layer performs corresponding manifold operations to capture the nonlinear characteristics of the data, and the Sigmoid activation function is used to enable the model to simulate the nonlinear changes in cementitious material production. S5503: Based on the cementitious material production data points on the Riemann manifold, the network loss is calculated using the cross-entropy loss function based on the regularization term to ensure that the model can accurately classify the cementitious material characteristics. S5504: Fractional derivatives are used to calculate the gradient of network parameters to adapt to the distribution characteristics of cementitious material production data on the manifold. The distribution of cementitious material data on the Riemannian manifold has complex geometric characteristics, and fractional derivatives capture these geometric characteristics more delicately in weight updates.

7. The method for assessing energy consumption in cementitious material production based on IoT data collection according to claim 6, characterized in that, The features in the cementitious material data are highly correlated, and dynamically updating the weights can gradually optimize the representation of these correlated features, and update the network parameters based on the gradient calculated by backpropagation.

8. The method for assessing energy consumption in cementitious material production based on IoT data collection according to claim 6, characterized in that, After the pre-training is completed, the parameters of the two feature extraction networks are fixed and transferred to the target domain cementitious material production data for fine-tuning. The cementitious material production data of the target domain is divided into support set and query set according to the meta-learning strategy. The energy consumption evaluation model of transfer learning is fine-tuned using the cementitious material production data of the support set. The fine-tuning strategy adopts the gradient descent method. After fine-tuning the energy consumption assessment model, the query set of cementitious material production data in the target domain is input into the neural network model for feature extraction of cementitious material production data and the Riemann neural network energy consumption assessment model, respectively, to obtain their corresponding prediction scores. Then, the Softmax function is used to normalize the prediction scores. Finally, the two prediction scores are merged and output as the final prediction result. The classification category is the category corresponding to the maximum prediction score.

9. The method for assessing energy consumption in cementitious material production based on IoT data collection according to claim 7, characterized in that, The collected data attributes can be as follows: Ta represents temperature value, recording the temperature of the equipment operating environment; Pa represents pressure value, recording the working pressure during the production process; Va represents speed value, recording the equipment operating speed; Ca represents power consumption, recording the power consumed per unit time; Ha represents humidity value, recording the ambient humidity; Wa represents weight value, recording the weight of the material; Ea represents energy efficiency ratio, recording the energy consumed per unit of product produced; Qa represents quality score, recording the product quality control score; Sa represents safety index, recording production safety-related indicators; Da represents duration value, recording the time required to produce one batch. In practical applications, the data has more than 10 attributes, and the number of attributes can reach dozens or even hundreds. Furthermore, the collected data is manually labeled, with the labeling categories including: low energy consumption, medium energy consumption, and high energy consumption. The preset stopping iteration condition is reaching the preset maximum number of iterations, which is set to 1000.

Images