Method and device for evaluating energy efficiency of coal pulverizing system

Abstract

The invention provides a coal pulverizing system energy efficiency evaluation method and device, and the method comprises the steps: obtaining the multi-dimensional feature data of a coal pulverizing system, and the multi-dimensional feature data comprises an acoustic signal, a vibration signal, an operation parameter, and a coal quality characteristic; constructing a three-dimensional energy efficiency evaluation system including energy consumption, reliability and environmental influence based on the multi-dimensional feature data, wherein the three-dimensional energy efficiency evaluation system allocates weights of an energy consumption dimension, a reliability dimension and an environmental influence dimension through an analytic hierarchy process; establishing a mechanism-data hybrid dynamic model, and embedding an energy conservation equation as a hard constraint into an improved Transform network to realize collaborative modeling of a physical mechanism and data characteristics; and carrying out dynamic optimization on the hybrid dynamic model by adopting a transfer learning framework, minimizing feature distribution differences under different coal working conditions through adversarial training, and implementing consistency regularization in combination with unmarked samples to improve the generalization ability of the model.

Description

Technical Field

[0001] This invention belongs to the field of power system technology, specifically relating to a method and apparatus for evaluating the energy efficiency of a pulverizing system. Background Technology

[0002] As a core process equipment in industries such as thermal power, cement, and chemicals, the energy efficiency of pulverizing systems directly affects production costs, equipment lifespan, and carbon emission intensity. Existing energy efficiency assessment and optimization technologies for pulverizing systems mainly focus on three areas: fault diagnosis, parameter optimization, and energy efficiency evaluation. 1. In the field of fault diagnosis, deep learning-based solutions dominate. For example, the stacked autoencoder (SAE) achieves fault feature recognition of coal mills through fault data training, with an accuracy of 98.97%. The CNN-BiLSTM-Attention model integrates spatiotemporal features to improve early warning performance. 2. In terms of parameter optimization, the NSGA-II algorithm combined with the PSO-BPNN model forms a multi-objective optimization framework, which optimizes parameters in feed grinding processes with productivity and power consumption per ton of feed as objectives. Dynamic neural network optimization (such as LSTM) combined with particle swarm optimization improves thermal efficiency by 4.58% compared to steady-state optimization. 3. Energy efficiency evaluation methods include two categories: single indicators (grinding unit consumption) and comprehensive indicators (weighted evaluation of coal powder fineness, output, power consumption, etc.). Fuzzy comprehensive evaluation rules further incorporate factors such as metal consumption and material costs.

[0003] Existing technologies have multiple shortcomings in energy efficiency assessment and fault diagnosis of pulverizing systems. In terms of feature extraction, they rely solely on traditional data such as vibration signals, neglecting acoustic features, resulting in insufficient early warning of faults and failure to identify early-stage faults. The energy efficiency evaluation system is incomplete, often using single indicators such as power consumption per ton of pulverized coal, ignoring dimensions such as system health and carbon emissions, thus failing to reflect the overall system performance. The model's generalization ability is insufficient, with errors exceeding 10% when coal type changes, lagging dynamic control under complex operating conditions, temperature control accuracy of only ±3.5℃, and load change response time reaching 32 seconds. Data fusion presents challenges, with inadequate preprocessing of multi-source heterogeneous data and a lack of intelligent optimization algorithms, leading to persistently high overall energy consumption. Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems in the related art.

[0005] Therefore, the first objective of this invention is to provide a method for evaluating the energy efficiency of a powder-making system.

[0006] The second objective of this invention is to provide an energy efficiency evaluation device for a powder-making system.

[0007] The third objective of this invention is to provide a computer device.

[0008] A fourth objective of this invention is to provide a non-transitory computer-readable storage medium.

[0009] To achieve the above objectives, a first aspect of the present invention provides a method for evaluating the energy efficiency of a pulverizing system, comprising: S1, acquire multi-dimensional feature data of the pulverizing system, the multi-dimensional feature data including acoustic signals, vibration signals, operating parameters and coal quality characteristics; S2, Based on the multi-dimensional feature data, a three-dimensional energy efficiency evaluation system including energy consumption, reliability and environmental impact is constructed. The three-dimensional energy efficiency evaluation system allocates the weights of the energy consumption dimension, reliability dimension and environmental impact dimension through the analytic hierarchy process. S3. Establish a mechanism-data hybrid dynamic model, embed the energy conservation equation as a hard constraint into the improved Transformer network, and realize the collaborative modeling of physical mechanisms and data characteristics. S4. The hybrid dynamic model is dynamically optimized using a transfer learning framework. The feature distribution differences under different coal types are minimized through adversarial training, and consistency regularization is implemented in combination with unlabeled samples to improve the model's generalization ability.

[0010] In one embodiment of the present invention, S1 includes: S11, Acoustic signals are acquired through a 4-channel microphone array. The microphone array has 2 microphones at each end of the coal mill, with a sampling radius of 0.5m and a sampling frequency of 44.1kHz. S12, the acoustic signal is sequentially subjected to anti-aliasing filtering and spectral subtraction noise reduction processing to eliminate environmental noise and retain the characteristic sound waves of the equipment operation.

[0011] In one embodiment of the present invention, S2 includes: S21, the weights of the three-dimensional energy efficiency evaluation system are allocated using the Analytic Hierarchy Process (AHP), with the energy consumption dimension having a weight of 0.5, the reliability dimension having a weight of 0.3, and the environmental impact dimension having a weight of 0.2, and this is done using the formula... Calculate the overall energy efficiency index.

[0012] In one embodiment of the present invention, S3 further includes: S31, the energy conservation equation As an improved Transformer network with hard-constrained embedding, in which For electrical energy input, To carry away heat from the pulverized coal, For heat dissipation loss, This is the grinding work.

[0013] In one embodiment of the present invention, S4 includes: S41 uses adversarial training to bring the feature space distance between the source and target domains to 0.03, and introduces 40% unlabeled samples to implement consistency regularization to improve the model's generalization ability.

[0014] In one embodiment of the present invention, the method further includes: S5 dynamically adjusts the multimodal feature weights based on operating condition characteristics, specifically including: S51, when the moisture content of coal is detected to be higher than the benchmark value, the acoustic feature weight is automatically increased to 0.62 through the working condition attention weight; S52 employs a temporal attention mechanism to rank the feature importance of vibration and acoustic signals, prioritizing the extraction of parameters strongly correlated with energy efficiency.

[0015] To achieve the above objectives, a second aspect of the present invention provides an energy efficiency evaluation device for a pulverizing system, comprising: A multimodal data acquisition module is used to acquire multi-dimensional feature data of the pulverizing system, including acoustic signals, vibration signals, operating parameters, and coal quality characteristics. The three-dimensional evaluation system construction module is used to construct a three-dimensional energy efficiency evaluation system that includes energy consumption, reliability and environmental impact based on the multi-dimensional feature data. The three-dimensional energy efficiency evaluation system allocates the weights of the energy consumption dimension, reliability dimension and environmental impact dimension through the analytic hierarchy process. The physical co-modeling module is used to establish a mechanism-data hybrid dynamic model. It embeds the energy conservation equation as a hard constraint into an improved Transformer network to achieve co-modeling of physical mechanisms and data characteristics. The transfer optimization module is used to dynamically optimize the hybrid dynamic model using a transfer learning framework. It minimizes the differences in feature distribution under different coal types and working conditions through adversarial training, and implements consistency regularization in combination with unlabeled samples to improve the model's generalization ability.

[0016] The present invention provides a method and apparatus for evaluating the energy efficiency of a pulverizing system, which can achieve a comprehensive energy efficiency evaluation of the pulverizing system in three dimensions: energy consumption, reliability, and environmental impact, improve the accuracy and dynamic adaptability of the evaluation model, and significantly enhance the timeliness of fault early warning and carbon emission control capabilities.

[0017] To achieve the above objectives, a third aspect of this application provides a computer device, including a processor and a memory; wherein the processor runs a program corresponding to the executable program code by reading executable program code stored in the memory, for implementing a pulverizing system energy efficiency evaluation method as described in the first aspect embodiment.

[0018] To achieve the above objectives, a fourth aspect of this application provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for evaluating the energy efficiency of a pulverizing system as described in the first aspect embodiment.

[0019] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0020] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of an energy efficiency evaluation method for a powder-making system according to an embodiment of the present invention; Figure 2 This is an architecture diagram of an energy efficiency evaluation method for a powder-making system according to an embodiment of the present invention; Figure 3 This is a structural diagram of an energy efficiency evaluation device for a powder-making system according to an embodiment of the present invention; Figure 4 It is a computer device according to an embodiment of the present invention. Detailed Implementation

[0021] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0023] The following description, with reference to the accompanying drawings, describes a method and apparatus for evaluating the energy efficiency of a powder-making system according to an embodiment of the present invention.

[0024] Example 1 Figure 1 This is a flowchart of an energy efficiency evaluation method for a pulverizing system according to an embodiment of the present invention, such as... Figure 1 As shown, it includes: S1, acquire multi-dimensional feature data of the pulverizing system, including acoustic signals, vibration signals, operating parameters and coal quality characteristics.

[0025] In some implementations, the acoustic signals are acquired by a 4-channel microphone array deployed at the front and rear ends of the coal mill, with a sampling frequency of [missing information]. This covers key acoustic radiation areas to ensure signal integrity. Vibration signals are then transmitted via a triaxial accelerometer. Frequency acquisition extracts time-frequency domain features such as kurtosis, peak factor, and 1 / 3 octave band energy. Operating parameters are acquired in real-time by the DCS system, including 18 indicators such as current, temperature, and pressure, with a sampling frequency of [frequency value missing]. This is used to reflect the system's operating status. Coal quality characteristics are then measured using an online coal quality analyzer. The data is collected periodically, covering key coal quality parameters such as moisture, ash, and volatile matter, to establish a correlation model between coal quality and energy efficiency.

[0026] Specifically, acoustic signal preprocessing includes anti-aliasing filtering (cutoff frequency). ) and spectral subtraction noise reduction to preserve the characteristic acoustic waves of the device operation. MFCC features are obtained through a 2048-point FFT and Point frame shift extraction, forming The acoustic eigenvectors are generated. Vibration signals are denoised using wavelet filtering to improve the signal-to-noise ratio. Missing values ​​in operating parameters are imputed and normalized to ensure data consistency. Coal quality characteristics are obtained through a standardized coal quality analysis process, with data updates synchronized with the system's operating cycle.

[0027] Furthermore, S1 includes: S11 Acoustic signals are acquired through a 4-channel microphone array. The microphone array has two microphones at each end of the coal mill, with a sampling radius of 0.5m and a sampling frequency of 44.1kHz.

[0028] In some implementations, the microphone array supports The sampling frequency conforms to international standards for audio signal processing (such as ISO 11201:2010), enabling it to fully capture the mid-to-high frequency acoustic characteristics generated during coal mill operation, such as metallic friction sounds, coal particle crushing sounds, and abnormal bearing vibration sounds. During signal acquisition, data from each channel is synchronized using a unified clock source and fused with vibration signals (2kHz) and operating parameters (1Hz) using timestamp alignment to ensure data consistency in the temporal dimension.

[0029] Specifically, the radius of the microphone array This is based on a comprehensive consideration of the sound wave propagation characteristics and the equipment's structural dimensions, ensuring that the sound source localization error is controlled within a certain range. Within range. Sampling frequency Satisfying the Nyquist sampling theorem effectively avoids signal aliasing and supports subsequent extraction of higher-order acoustic features such as Mel-frequency cepstral coefficients (MFCC). An anti-aliasing filter (cutoff frequency) is used in the preprocessing stage. ) and spectral subtraction noise reduction to eliminate environmental noise interference and retain the characteristic sound wave components of equipment operation.

[0030] S12, the acoustic signal is sequentially subjected to anti-aliasing filtering (cutoff frequency 8kHz) and spectral subtraction noise reduction processing to eliminate environmental noise and retain the characteristic sound waves of the equipment operation.

[0031] In some implementations, anti-aliasing filtering uses a low-pass filter (LPF) to preprocess the original acoustic signal, with a cutoff frequency set to 8kHz to prevent spectral aliasing during subsequent sampling or transformation. This filter is typically an IIR or FIR structure; in specific implementations, a Butterworth filter can be used, with its order adjusted based on the difference between the system noise bandwidth and the signal bandwidth, generally set to 4-6 to control computational complexity while ensuring filtering effectiveness. The filtered signal then enters the spectral subtraction denoising stage. This method uses short-time Fourier transform (STFT) to perform time-frequency analysis of the signal, estimating the noise spectrum and subtracting it from the mixed signal to achieve denoising. Specifically, the system first collects background noise samples during quiet periods when no equipment is running and calculates their average spectral power. Then, in the device's operating signal, for each frequency band power After correction, the enhanced signal spectrum is obtained. Finally, the time-domain signal is reconstructed through inverse STFT.

[0032] Specifically, the acoustic signal sampling frequency is 44.1 kHz, the frame length is 2048 points, and the frame shift is 512 points to ensure the continuity and resolution of the signal in the time and frequency domains. The noise estimation window length in the spectral subtraction is typically 10–30 seconds to accommodate noise variations under different operating conditions. The noise reduction effect of this step can be quantitatively evaluated by the improvement in signal-to-noise ratio (SNR). Experiments show that in typical industrial noise environments, the SNR can be improved by approximately 6–10 dB.

[0033] S2. Based on the multi-dimensional feature data, a three-dimensional energy efficiency evaluation system including energy consumption, reliability, and environmental impact is constructed. The three-dimensional energy efficiency evaluation system allocates the weights of the energy consumption dimension, reliability dimension, and environmental impact dimension through the analytic hierarchy process.

[0034] In some implementations, the three-dimensional energy efficiency evaluation system adopts a hierarchical structure, divided into a basic parameter layer, a core indicator layer, and a comprehensive energy efficiency index (EEI) layer. The basic parameter layer includes key operational data such as power consumption, failure frequency, and CO2 concentration; the core indicator layer consists of the net energy efficiency ratio (NER), health index (HI), and carbon emission factor (CF), representing the system's performance in terms of energy consumption, equipment health status, and environmental impact, respectively. The top-level comprehensive energy efficiency index (EEI) is calculated using a weighted summation method, and its mathematical expression is:

[0035] In this framework, NER is the ratio of effective energy output to total input energy; HI is calculated based on a weighted average of relative entropy and wear coefficient to determine equipment health status; and CF represents the CO2 emissions per unit of pulverized coal production. The weight allocation employs the Analytic Hierarchy Process (AHP), ensuring the rationality and scientific validity of each dimension's weights by constructing a judgment matrix and conducting consistency checks. In practical applications, the AHP judgment matrix is ​​generated by experts who score the relative importance of each dimension to the system's energy efficiency, and the final weight values ​​are determined using the eigenvector method.

[0036] Specifically, the benchmark value for NER refers to industry-leading data for platform-type wheat milling equipment at 70%-90% design capacity utilization, with a unit power consumption of [missing data]. The powder yield is 76% - 78%; the HI benchmark value is set at... It is necessary to consider the advance warning time of equipment failure (such as the advance warning time of rolling bearing failure). (hours) and number of annual failures ( The value of CF is dynamically adjusted (per unit / time); the baseline value of CF is... And adjustments are made based on differences in coal type.

[0037] Furthermore, S2 includes: S21, the weights of the three-dimensional energy efficiency evaluation system are allocated using the Analytic Hierarchy Process (AHP), with the energy consumption dimension having a weight of 0.5, the reliability dimension having a weight of 0.3, and the environmental impact dimension having a weight of 0.2, and this is done using the formula... Calculate the overall energy efficiency index.

[0038] In some implementations, the specific operational process for weight allocation is as follows: First, based on the energy efficiency assessment objectives of the pulverizing system, a hierarchical model containing three dimensions is established. Energy consumption (NER) serves as the core indicator of system operating efficiency, reliability (HI) reflects equipment health and operational stability, and environmental impact (CF) measures ecological burdens such as carbon emissions. Using expert scoring or historical data-driven methods, pairwise comparison matrices are constructed, and the relative weights of each dimension are calculated. The final determined weights are: energy consumption dimension weight 0.5, reliability dimension weight 0.3, and environmental impact dimension weight 0.2, satisfying the constraint that the sum of weights in the AHP method is 1.

[0039] Furthermore, the formula for calculating the comprehensive energy efficiency index is as follows: The formula derives from the following: NER (Net Energy Efficiency Ratio), defined as the ratio of effective energy output to total input energy; HI (Health Index), calculated by weighting the relative entropy of equipment operating status with wear coefficient; and CF (Carbon Emission Factor), representing the CO2 emissions per unit of pulverized coal production. This formula linearly weights and integrates multiple indicators to form a quantifiable system energy efficiency assessment result.

[0040] S3 establishes a mechanism-data hybrid dynamic model, embedding the energy conservation equation as a hard constraint into an improved Transformer network to achieve collaborative modeling of physical mechanisms and data characteristics.

[0041] In some implementations, this hybrid model is constructed based on the energy balance equation of a medium-speed coal mill: ,in Indicates electrical energy input (unit: kWh). Heat (kWh) is carried out by pulverized coal. System heat loss (kWh). The effective energy consumed in the grinding process (kWh). This equation serves as a physical constraint, and during model training, it corrects the prediction output in real time through an energy conservation verification mechanism, thereby reducing prediction bias caused by data noise or sudden changes in operating conditions.

[0042] The improved Transformer network receives multimodal feature sequences at the input layer, including vibration signals, acoustic signals (such as MFCC, SPL, and spectral centroid), and operating parameters (such as 18 indicators including current, temperature, and pressure). The encoder incorporates a temporal attention mechanism and operating condition attention weights to dynamically adjust the contribution of different features under different operating conditions. For example, under high-moisture coal conditions, the weight of acoustic features is automatically increased to 0.62, significantly enhancing the model's responsiveness to changes in coal quality.

[0043] Furthermore, this hybrid model achieves joint optimization of physical constraints and data fitting by incorporating the energy conservation equation as part of the loss function. During training, the model not only minimizes prediction errors (such as MSE) but also satisfies energy conservation constraints, thereby improving the model's physical consistency and robustness. Experimental data show that this mechanism can reduce the model's prediction error by 2.3%, and under complex coal conditions such as lignite, the energy efficiency prediction deviation is controlled within 3.8%, representing a 27% improvement compared to a purely data-driven model.

[0044] Furthermore, S3 includes: S31, the energy conservation equation As an improved Transformer network with hard-constrained embedding, in which For electrical energy input (kWh), To remove heat (kWh) from pulverized coal. For heat loss (kWh), The grinding work is expressed in kWh.

[0045] In its implementation, this energy conservation equation is introduced into the model by constructing a constraint loss function. Specifically, in the output layer of the Transformer network, the predicted... , and With known Real-time validation is performed, and the residuals are calculated and incorporated into the total loss function. For example, the constraint loss term can be defined as:

[0046] in The constraint weight coefficient is usually set to This balances the priorities of data fitting and physical constraints. This constraint term, together with the traditional mean squared error (MSE) loss term, constitutes a hybrid loss function, driving the model to achieve high-precision prediction of energy efficiency parameters while satisfying physical laws.

[0047] S4. The hybrid dynamic model is dynamically optimized using a transfer learning framework. The feature distribution differences under different coal types are minimized through adversarial training, and consistency regularization is implemented in combination with unlabeled samples to improve the model's generalization ability.

[0048] In some implementations, transfer learning frameworks use Domain-Adaptive Neural Networks (DANNs) as their core, introducing adversarial learning mechanisms to align feature distributions between the source domain (historical coal types) and the target domain (new coal types). Specifically, the model is first supervised pre-trained in the source domain, and the loss function combines mean squared error (MSE) with L2 regularization. ,in This is the regularization coefficient, usually set to... To prevent model overfitting, a gradient reversal layer (GRL) is introduced during the domain alignment stage. This allows the feature extractor to undergo adversarial training between the classification and domain discrimination tasks, thereby minimizing the maximum mean discrepancy (MMD) between the source and target domains and converging the feature space distance to a minimum. Within this range, the model's adaptability to changes in coal type is significantly improved.

[0049] Furthermore, to enhance the model's generalization ability to unknown operating conditions, this invention introduces a consistency regularization strategy, utilizing 40% of unlabeled samples for semi-supervised fine-tuning. Specifically, the model applies consistency constraints to the prediction results of the same unlabeled sample under different noise perturbations. By minimizing the difference in prediction outputs, the robustness of the model under uneven data distribution is improved. The consistency loss function can be expressed as... ,in and This refers to the predicted output for the same sample under different perturbations.

[0050] Furthermore, S4 includes: S41 uses adversarial training to bring the feature space distance between the source domain (historical coal types) and the target domain (new coal types) to converge to 0.03, and introduces 40% unlabeled samples to implement consistency regularization to improve the model's generalization ability.

[0051] Specifically, in the source domain pre-training phase, the model is based on historical operating data from three 300MW units (totaling 1.2 × 10⁻⁶). 6Supervised learning is performed on a sample-by-sample model, with a loss function combining mean squared error (MSE) and L2 regularization to improve model stability and generalization ability. In the domain alignment optimization stage, an adversarial learning mechanism is introduced to minimize the distribution difference between the source and target domains while maximizing task prediction performance. This process is achieved through a gradient reversal layer (GRL), whose core idea is to backpropagate the discriminative loss of the feature space to the feature extractor, forcing the feature distribution to converge. In this invention, through adversarial training, the feature space distance between the source domain (historical coal type) and the target domain (new coal type) eventually converges to 0.03, meeting the stability requirements for cross-coal type modeling.

[0052] To further enhance the model's generalization ability under unknown operating conditions, this invention introduces 40% unlabeled samples on top of adversarial training and employs a consistency regularization strategy. This strategy applies prediction consistency constraints to the unlabeled samples, ensuring the model's output remains stable under different input perturbations, thereby improving its adaptability to new coal types. Specifically, unlabeled samples are input into the model after data augmentation (e.g., additive Gaussian noise, time window sliding), and the difference loss in the predicted output is calculated. This loss is jointly optimized with the task loss to form the final hybrid loss function. Experiments show that this method reduces the energy efficiency prediction bias to 3.8% under complex coal types such as lignite, a 27% improvement over traditional models. This step, as a key component of the hybrid dynamic modeling module, effectively solves the performance degradation problem of the model during coal type migration, providing a highly robust prediction foundation for the three-dimensional energy efficiency evaluation system.

[0053] S5 dynamically adjusts the weights of multimodal features based on operating condition characteristics.

[0054] In some implementations, this step is based on the collaborative modeling of multimodal features and operating condition features. The specific operation involves: First, simultaneously acquiring multi-source data from the DCS system, vibration sensors, microphone arrays, and online coal quality analyzers, ensuring data consistency through timestamp alignment; second, performing hierarchical preprocessing on each modal data, such as wavelet filtering for noise reduction of vibration signals, anti-aliasing filtering and spectral subtraction for noise reduction of acoustic signals, imputation of missing values ​​for operating parameters, and obtaining coal quality characteristics through chemical composition analysis; finally, inputting the preprocessed multimodal features into an improved Transformer network, where the encoder incorporates an operating condition attention weight module to dynamically adjust the weight allocation of each feature according to the current operating conditions (such as coal type, equipment wear status, load changes, etc.). For example, under high-moisture coal conditions, the acoustic feature weight is automatically increased to 0.62 to enhance sensitivity to anomalies in the grinding process.

[0055] Key parameters involved in this step include: the dynamic adjustment threshold for attention weights, the temporal window length for feature fusion, the sampling frequency of each modal feature (e.g., 44.1 kHz for acoustic signals, 2 kHz for vibration signals, and 1 Hz for operating parameters), and the coal quality characteristic update cycle (5 minutes / time). Furthermore, the weight allocation is set according to the Analytic Hierarchy Process (AHP), where the weight changes of acoustic features under specific operating conditions must meet certain requirements. The constraints are set to ensure the stability and generalization ability of the model under different operating conditions.

[0056] S51, when the moisture content of coal is detected to be higher than the benchmark value, the acoustic feature weight is automatically increased to 0.62 through the working condition attention weight.

[0057] In some implementations, this adjustment process relies on a multi-channel structure of the feature input layer, where acoustic features are acquired by a 4-channel microphone array at a sampling frequency of 44.1 kHz, and then processed using anti-aliasing filtering (cutoff frequency 8 kHz) and spectral subtraction noise reduction to extract key parameters such as 13-dimensional MFCC features, sound pressure level (SPL), and spectral centroid. In the attention mechanism, the weight adjustment of the acoustic features is based on the feature correlation matrix under the current operating conditions, implemented through a learnable attention weight allocator, with the following output: , where α and β are attention weight parameters, which are obtained through training with historical working condition data.

[0058] S52 employs a temporal attention mechanism to rank the feature importance of vibration and acoustic signals, prioritizing the extraction of parameters strongly correlated with energy efficiency (such as sudden changes in sound pressure level and vibration kurtosis anomalies).

[0059] In some implementations, the temporal attention mechanism is based on an improved Transformer architecture. Its core lies in the weighted fusion of temporal features of vibration and acoustic signals through self-attention and cross-attention modules. Specifically, the vibration signal is acquired using a triaxial accelerometer at a sampling frequency of 2kHz, extracting temporal features (such as kurtosis and peak factor) and frequency domain features (such as 1 / 3 octave band energy). The acoustic signal is acquired by a 4-channel microphone array at a sampling frequency of 44.1kHz. After anti-aliasing filtering (cutoff frequency 8kHz) and spectral subtraction noise reduction, 13-dimensional MFCC features, sound pressure level (SPL), and spectral centroid are extracted. These features are input into the Transformer encoder, where a multi-head attention mechanism calculates the correlation weights of each feature at different time steps, thereby achieving dynamic focusing on key features.

[0060] Specifically, the weight allocation of the attention mechanism is based on characteristics and energy efficiency loss ( The correlation of the features is optimized. For example, during the operation of a coal mill, when the sound pressure level (SPL) changes abruptly, its corresponding attention weight can be increased to 0.62, which is significantly higher than the weight of other features. In addition, vibration kurtosis anomaly, as an important indicator reflecting the nonlinear vibration behavior of the equipment, has a 40% faster response speed in the attention mechanism than traditional methods, which helps to identify equipment wear or failure at an early stage.

[0061] The present invention provides a method for evaluating the energy efficiency of a pulverizing system, which enables a comprehensive evaluation of the pulverizing system across three dimensions: energy consumption, reliability, and environmental impact. This method improves the accuracy and dynamic adaptability of the evaluation model and significantly enhances the timeliness of fault warning and carbon emission control capabilities.

[0062] Example 2 The following describes in detail, with reference to the accompanying drawings, a method for evaluating the energy efficiency of a powder-making system according to an embodiment of the present invention.

[0063] This invention is divided into three main modules, such as Figure 2 As shown, its specific implementation method and functions are as follows: 1. Multi-dimensional feature extraction module: The multi-dimensional feature extraction module introduces acoustic signals as a core innovative dimension, combining vibration, operating parameters, and coal quality characteristics to construct a multi-modal feature system, effectively overcoming the limitations of existing technologies that rely solely on vibration signals. In existing technologies, coal mill fault detection systems primarily acquire vibration signals (2kHz sampling frequency) using a triaxial accelerometer, extracting time-domain features such as kurtosis and peak factor, and frequency-domain features such as 1 / 3 octave band energy, and acquiring 18 operating parameters (1Hz sampling frequency) based on the DCS system. This module adds an acoustic feature dimension, forming a four-dimensional feature architecture of "vibration-acoustics-operation-coal quality." Through multi-modal fusion, the feature information entropy is increased by 40%, significantly enhancing the comprehensiveness and accuracy of energy efficiency assessment.

[0064] 1.1 Hardware deployment and algorithm implementation for acoustic feature extraction: The acoustic feature extraction system employs a 4-channel microphone array (sampling frequency 44.1kHz). The specific deployment is as follows: two microphones are placed at each end of the coal mill, with a sampling radius of 0.5m, ensuring coverage of the equipment's critical acoustic radiation area. The signal preprocessing stage sequentially performs anti-aliasing filtering (cutoff frequency 8kHz) and spectral subtraction noise reduction to eliminate environmental interference and preserve the characteristic sound waves of the equipment operation. Acoustic feature parameters include Mel-frequency cepstral coefficients (MFCC), sound pressure level (SPL), and spectral centroid. The Mel-frequency cepstral coefficients (MFCC) can be extracted using the librosa library as 13-dimensional MFCC features, employing a 2048-point FFT and a 512-point frame shift to reflect the nonlinear acoustic characteristics of the equipment vibration. The sound pressure level can be calculated based on the formula SPL = 20log(signal peak value / 2e-5) to quantify the sound energy intensity. The spectral centroid characterizes the center frequency of the energy distribution, assisting in the identification of abnormal sound wave frequency bands.

[0065] 1.2 Multi-dimensional Feature System and Integration Advantages: The specific parameters of the feature dimensions for module expansion are shown in Table 1. Among them, the introduction of acoustic features provides key supplementary information for fault early warning and energy efficiency assessment. Table 1

[0066] Multimodal fusion distinguishes effective signals from noise through a dynamic and static feature separation mechanism. For example, in a coal mill fault early warning scenario, the complementarity of acoustic features and vibration signals increases the early warning lead time to 25 minutes, significantly improving the accuracy compared to traditional single vibration signal early warning. Furthermore, drawing on experience with multi-source data fusion in platform-type wheat milling equipment, this module links acoustic features with coal quality parameters to construct a mapping relationship between "coal quality - acoustic features - energy efficiency loss," providing high-resolution feature input for the subsequent three-dimensional energy efficiency evaluation system.

[0067] 1.3 Feature Extraction Process Design: The overall process of multi-dimensional feature extraction follows a four-step architecture: signal acquisition, preprocessing, feature calculation, and modality fusion. Synchronous signal acquisition: Vibration (2kHz), acoustic (44.1kHz), and operating parameter (1Hz) data are aligned using timestamps; Layered preprocessing: The vibration signal is denoised by wavelet filtering, the acoustic signal is denoised by anti-aliasing filtering and spectral subtraction, and the running parameters are imputed for missing values; Feature parameter calculation: Extract features of each dimension according to the table above, among which acoustic features are calculated automatically using Python; Multimodal fusion: Employs an attention mechanism to dynamically allocate the weights of each feature, focusing on key parameters that are strongly correlated with energy efficiency (such as sudden changes in sound pressure level and vibration kurtosis anomalies).

[0068] This process, through collaborative design of hardware and software, realizes the transformation from multi-source heterogeneous data to high-value characteristics, providing a solid data foundation for the energy efficiency assessment of pulverizing systems.

[0069] 2. Hybrid Dynamic Modeling Module: The hybrid dynamic modeling module adopts a mechanism-data fusion-driven architecture, achieving a balance between high accuracy and strong interpretability in the energy efficiency assessment of pulverizing systems through the synergistic coupling of physical and data models. This module consists of a mechanism constraint layer, an improved data model layer, and a transfer learning layer, which form a closed-loop optimization system through a dynamic interaction mechanism.

[0070] 2.1 Mechanism - Data Coupling Mechanism: The mechanistic layer uses the energy conservation principle of the pulverizing system as the core constraint, establishing the energy balance equation for a medium-speed coal mill: Qin = Qout + Qloss + Qgrind, where Qin is the electrical energy input (kWh), Qout is the heat carried out by the pulverized coal (kWh), Qloss is the heat loss (kWh), and Qgrind is the grinding work (kWh). This equation is embedded as a hard constraint in the data model training process, and the prediction results are corrected in real time through energy conservation verification. Experimental data show that the mechanistic constraint can reduce the data model error by 2.3%, significantly improving the physical interpretability of the model.

[0071] The data layer adopts a multimodal fusion architecture, with the basic model being an improved Transformer network. The input layer receives multi-channel feature sequences composed of vibration, acoustic, and operational parameters. The encoder introduces a temporal attention mechanism and working condition attention weights to dynamically allocate the importance of features under different working conditions. For example, under high-moisture coal conditions, the acoustic feature weight is automatically increased to 0.62, improving the feature response speed by 40% compared to the traditional model.

[0072] 2.2 Transfer Learning Implementation Framework: To address the issue of data distribution discrepancies under complex operating conditions such as varying coal types and wear levels, Domain Adaptive Networks (DANN) are used as the core of transfer learning. The implementation steps include: Source domain pre-training: using one year of operating data from three 300 MW units (1.2 × 10⁻⁶). 6 Supervised learning is performed on samples, and the loss function uses MSE+L2 regularization; Domain alignment optimization: By minimizing the distribution difference between the source domain (historical coal type) and the target domain (new coal type) through adversarial training, the feature space distance converges to 0.03; Semi-supervised fine-tuning: Introducing 40% unlabeled samples to implement consistency regularization enhances the model's ability to generalize to unknown operating conditions.

[0073] 2.3 Modeling Process and Validation: The model training employs a two-stage strategy: the first stage completes the identification of mechanistic parameters and pre-training of the data model; the second stage achieves dynamic optimization through energy conservation verification and transfer learning. Experimental results show that the module maintains stable accuracy under six typical coal types, including bituminous coal, lean coal, and lignite. Among them, the energy efficiency prediction deviation under lignite conditions is only 3.8%, which is 27% higher than that of a purely data-driven model.

[0074] By deeply integrating physical mechanisms with data intelligence, this module effectively solves the contradiction between the traditional model's "poor interpretability" and "weak generalization ability," providing a systematic solution for the full-condition energy efficiency evaluation of pulverizing systems.

[0075] 3. Three-dimensional energy efficiency evaluation index system: To address the limitations of existing energy efficiency evaluation indicators for pulverizing systems, which are often singular and have limited coverage, this chapter constructs a three-dimensional energy efficiency evaluation indicator system encompassing energy consumption, reliability, and environmental impact. This system achieves a comprehensive characterization of system energy efficiency through multi-dimensional collaborative evaluation. The system defines Net Energy Efficiency Ratio (NER), Health Index (HI), and Carbon Emission Factor (CF) as core indicators, employing the Analytic Hierarchy Process (AHP) for weight allocation. This results in a 67% increase in evaluation coverage compared to traditional single-indicator evaluations, significantly improving the comprehensiveness and accuracy of the assessment.

[0076] 3.1 Definition of core indicators and determination of benchmark values: Energy consumption dimension – with Net Energy Efficiency Ratio (NER) as the core indicator, defined as the ratio of effective energy output to total input energy, calculated as: NER = Effective Energy Output / Total Input Energy. Its benchmark value references industry-leading data for platform-type complete wheat milling equipment at 70%-90% design capacity utilization, specifically 44-46 kWh / t of electricity consumption and 76%-78% flour yield. Within this range, the system's energy efficiency is stable and can serve as a target reference for energy efficiency optimization.

[0077] Reliability dimension – The health index HI is used to quantify the equipment operating status. The calculation formula integrates the relative entropy index in the coal mill fault early warning: HI=1 - Σ(feature deviation weight x wear coefficient); the benchmark value is set to ≥0.85, and needs to be dynamically adjusted in combination with the equipment fault early warning time (e.g., rolling bearing fault early warning ≥48 hours in advance) and the probability of fault occurrence (annual fault count ≤2 times / unit) to ensure long-term stable operation of the equipment.

[0078] Environmental impact dimension: Characterized by carbon emission factor CF, calculated as: CF = CO2 emissions / pulverized coal production (unit: tCO2 / t); the baseline value is set at ≤0.12 tCO2 / t, and is adjusted according to the type of coal used in the pulverizing system (such as lignite and bituminous coal) and the difference in combustion efficiency to achieve a synergistic assessment of energy consumption and environmental protection.

[0079] 3.2 Indicator System Structure and Weight Allocation: The three-dimensional indicator system achieves comprehensive evaluation through a hierarchical structure. The bottom layer consists of basic parameters (such as power consumption, failure frequency, and CO2 concentration), the middle layer consists of core indicators (NER, HI, and CF), and the top layer is the comprehensive energy efficiency index (EEI). The weight allocation adopts the AHP method, with a weight of 0.5 for the energy consumption dimension (NER), 0.3 for the reliability dimension (HI), and 0.2 for the environmental impact dimension (CF), forming the following comprehensive evaluation model: EEI = 0.5 x NER + 0.3 x HI + 0.2 x CF.

[0080] Example 3 To achieve the above embodiments, such as Figure 3 As shown, this embodiment also provides a pulverizing system energy efficiency evaluation device 10, which includes a multimodal data acquisition module 100, a three-dimensional evaluation system construction module 200, a physical collaborative modeling module 300, and a migration optimization module 400.

[0081] The multimodal data acquisition module 100 is used to acquire multi-dimensional feature data of the pulverizing system, including acoustic signals, vibration signals, operating parameters and coal quality characteristics. The three-dimensional evaluation system construction module 200 is used to construct a three-dimensional energy efficiency evaluation system including energy consumption, reliability and environmental impact based on the multi-dimensional feature data. The three-dimensional energy efficiency evaluation system allocates the weights of the energy consumption dimension, reliability dimension and environmental impact dimension through the analytic hierarchy process. The physical co-modeling module 300 is used to establish a mechanism-data hybrid dynamic model. It embeds the energy conservation equation as a hard constraint into an improved Transformer network to achieve co-modeling of physical mechanisms and data characteristics. The transfer optimization module 400 is used to dynamically optimize the hybrid dynamic model using a transfer learning framework. It minimizes the feature distribution differences under different coal types and working conditions through adversarial training, and implements consistency regularization in combination with unlabeled samples to improve the model's generalization ability.

[0082] Furthermore, the aforementioned multimodal data acquisition module 100 is also used for: Acoustic signals are acquired using a 4-channel microphone array, with two microphones arranged at each end of the coal mill, a sampling radius of 0.5m, and a sampling frequency of 44.1kHz. The acoustic signal is subjected to anti-aliasing filtering and spectral subtraction noise reduction in sequence to eliminate environmental noise and retain the characteristic sound waves of the equipment operation.

[0083] Furthermore, the aforementioned three-dimensional evaluation system construction module 200 is also used for: The weights of the three-dimensional energy efficiency evaluation system were assigned using the Analytic Hierarchy Process (AHP), with energy consumption dimension having a weight of 0.5, reliability dimension having a weight of 0.3, and environmental impact dimension having a weight of 0.2. The weights were then calculated using the formula... Calculate the overall energy efficiency index.

[0084] Furthermore, the aforementioned physical collaborative modeling module 300 is also used for: The energy conservation equation As an improved Transformer network with hard-constrained embedding, in which For electrical energy input, To carry away heat from the pulverized coal, For heat dissipation loss, This is the grinding work.

[0085] Furthermore, the migration optimization module 400 described above is also used for: Adversarial training was used to converge the feature space distance between the source and target domains to 0.03, and consistency regularization was implemented by introducing 40% unlabeled samples to improve the model's generalization ability.

[0086] Furthermore, device 10 also includes: The operating condition feature adjustment module is used to dynamically adjust the multimodal feature weights based on operating condition features, specifically including: When the moisture content of the coal is detected to be higher than the benchmark value, the acoustic feature weight is automatically increased to 0.62 by the working condition attention weight; A temporal attention mechanism is used to rank the feature importance of vibration and acoustic signals, and parameters that are strongly correlated with energy efficiency are extracted first.

[0087] The present invention discloses an energy efficiency assessment device for a pulverizing system, which can realize a comprehensive energy efficiency assessment of the pulverizing system in three dimensions: energy consumption, reliability, and environmental impact, improve the accuracy and dynamic adaptability of the assessment model, and significantly enhance the timeliness of fault early warning and carbon emission control capabilities.

[0088] To implement the methods of the above embodiments, the present invention also provides a computer device, such as... Figure 4 As shown, the computer device 600 includes a memory 601 and a processor 602; wherein, the processor 602 reads executable program code stored in the memory 601 to run a program corresponding to the executable program code, so as to implement the various steps of the method described above.

[0089] To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in the foregoing embodiments.

[0090] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0091] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for evaluating the energy efficiency of a milling system, characterized in that, include: S1, acquire multi-dimensional feature data of the pulverizing system, the multi-dimensional feature data including acoustic signals, vibration signals, operating parameters and coal quality characteristics; S2, Based on the multi-dimensional feature data, a three-dimensional energy efficiency evaluation system including energy consumption, reliability and environmental impact is constructed. The three-dimensional energy efficiency evaluation system allocates the weights of the energy consumption dimension, reliability dimension and environmental impact dimension through the analytic hierarchy process. S3. Establish a mechanism-data hybrid dynamic model, embed the energy conservation equation as a hard constraint into the improved Transformer network, and realize the collaborative modeling of physical mechanisms and data characteristics. S4. The hybrid dynamic model is dynamically optimized using a transfer learning framework. The feature distribution differences under different coal types are minimized through adversarial training, and consistency regularization is implemented in combination with unlabeled samples to improve the model's generalization ability.

2. The method as described in claim 1, characterized in that, S1 includes: S11, Acoustic signals are acquired through a 4-channel microphone array. The microphone array has 2 microphones at each end of the coal mill, with a sampling radius of 0.5m and a sampling frequency of 44.1kHz. S12, the acoustic signal is sequentially subjected to anti-aliasing filtering and spectral subtraction noise reduction processing to eliminate environmental noise and retain the characteristic sound waves of the equipment operation.

3. The method as described in claim 1, characterized in that, S2 includes: S21, the weights of the three-dimensional energy efficiency evaluation system are allocated using the Analytic Hierarchy Process (AHP), with the energy consumption dimension having a weight of 0.5, the reliability dimension having a weight of 0.3, and the environmental impact dimension having a weight of 0.2, and this is done using the formula... Calculate the overall energy efficiency index.

4. The method as described in claim 1, characterized in that, The S3 further includes: S31, the energy conservation equation As an improved Transformer network with hard-constrained embedding, in which For electrical energy input, To carry away heat from the pulverized coal, For heat dissipation loss, This refers to the grinding work.

5. The method as described in claim 1, characterized in that, The S4 includes: S41 uses adversarial training to bring the feature space distance between the source and target domains to 0.03, and introduces 40% unlabeled samples to implement consistency regularization to improve the model's generalization ability.

6. The method as described in claim 1, characterized in that, The method further includes: S5 dynamically adjusts the multimodal feature weights based on operating condition characteristics, specifically including: S51, when the moisture content of coal is detected to be higher than the benchmark value, the acoustic feature weight is automatically increased to 0.62 through the working condition attention weight; S52 employs a temporal attention mechanism to rank the feature importance of vibration and acoustic signals, prioritizing the extraction of parameters strongly correlated with energy efficiency.

7. An energy efficiency evaluation device for a flour milling system, characterized in that, include: A multimodal data acquisition module is used to acquire multi-dimensional feature data of the pulverizing system, including acoustic signals, vibration signals, operating parameters, and coal quality characteristics. The three-dimensional evaluation system construction module is used to construct a three-dimensional energy efficiency evaluation system that includes energy consumption, reliability and environmental impact based on the multi-dimensional feature data. The three-dimensional energy efficiency evaluation system allocates the weights of the energy consumption dimension, reliability dimension and environmental impact dimension through the analytic hierarchy process. The physical co-modeling module is used to establish a mechanism-data hybrid dynamic model. It embeds the energy conservation equation as a hard constraint into an improved Transformer network to achieve co-modeling of physical mechanisms and data characteristics. The transfer optimization module is used to dynamically optimize the hybrid dynamic model using a transfer learning framework. It minimizes the differences in feature distribution under different coal types and working conditions through adversarial training, and implements consistency regularization in combination with unlabeled samples to improve the model's generalization ability.

8. The apparatus as claimed in claim 7, characterized in that, The device further includes: The operating condition feature adjustment module is used to dynamically adjust the multimodal feature weights based on operating condition features, specifically including: When the moisture content of the coal is detected to be higher than the benchmark value, the acoustic feature weight is automatically increased to 0.62 by the working condition attention weight; A temporal attention mechanism is used to rank the feature importance of vibration and acoustic signals, and parameters that are strongly correlated with energy efficiency are extracted first.

9. A computer device, characterized in that, Including processor and memory; The processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the energy efficiency evaluation method for a pulverizing system as described in any one of claims 1-6.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements a method for evaluating the energy efficiency of a pulverizing system as described in any one of claims 1-6.

Images