Method, system, medium and equipment for estimating concentration of carbon monoxide at pulverized coal outlet

Abstract

The invention belongs to the technical field of industrial detection, and provides a pulverized coal outlet carbon monoxide concentration estimation method and system, a medium and equipment, and the method comprises the steps: obtaining a plurality of related data of a coal mill outlet; according to the multiple related data and the trained BP neural network model, obtaining a carbon monoxide concentration prediction result; performing carbon monoxide concentration prediction by using the trained BP neural network model; in a loss function of the BP neural network model, the sum of absolute values of corresponding weights of a plurality of related data is added as a penalty term, and each weight is dynamically adjusted according to a value of the corresponding related data, so that part of network weights become zero, the complexity of the model is reduced, the generalization ability is improved, and the feature selection effect is enhanced. And the model quality and the prediction precision are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial detection, and particularly relates to a method, system, medium and device for estimating the carbon monoxide concentration at the pulverized coal outlet. Background Art

[0002] Coal-fired power plants are an important part of the global energy supply. However, the safety issues during their production process have always been the focus of attention. Among them, the coal pulverizing system is a key part of the coal-fired power plant, and its safe operation is crucial for the stable operation of the entire power plant. In the coal pulverizing system, the coal mill is one of the core devices, and the change of the CO concentration at its outlet is directly related to the safe and stable operation of the system. Therefore, accurately predicting and warning the CO concentration at the coal mill outlet is of great significance for preventing explosion accidents. Traditional CO concentration prediction methods mainly rely on empirical formulas or simple mathematical models, and these methods often cannot accurately reflect complex actual working conditions, resulting in low prediction accuracy.

[0003] In recent years, with the development of artificial intelligence technology, deep learning technology has been widely applied to various prediction tasks, showing powerful prediction capabilities. However, when deep learning is currently applied to the monitoring of the carbon monoxide concentration at the coal mill outlet, only multiple characteristic parameters related to the CO concentration at the coal mill outlet are used to conduct conventional training on the neural network model, without better considering the influence of different types of characteristic parameters on the prediction results, which affects the neural network model and the prediction results. Summary of the Invention

[0004] In order to solve the above problems, the present invention proposes a method, system, medium and device for estimating the carbon monoxide concentration at the pulverized coal outlet. The present invention uses a trained BP neural network model to predict the carbon monoxide concentration; in the loss function of the BP neural network model, the sum of the absolute values of the weights corresponding to multiple relevant data is added as a penalty term, and each weight is dynamically adjusted according to the value of the corresponding relevant data, prompting some network weights to become zero, which helps to reduce the model complexity, improve the generalization ability, enhance the feature selection effect, and improve the model quality and prediction accuracy.

[0005] In order to achieve the above object, the present invention is realized by the following technical solutions:

[0006] In the first aspect, the present invention provides a method for estimating the carbon monoxide concentration at the pulverized coal outlet, including:

[0007] Obtain multiple relevant data at the outlet of the coal mill;

[0008] According to the multiple relevant data and the trained BP neural network model, obtain the carbon monoxide concentration prediction result;

[0009] Among them, in the loss function of the BP neural network model, the sum of the absolute values of the weights corresponding to multiple relevant data is added as a penalty term; each weight is dynamically adjusted according to the value of the corresponding relevant data.

[0010] Furthermore, the relevant data includes air temperature data, air volume data, powder volume data, and dynamic CO evolution data of coal quality.

[0011] Furthermore, in the BP neural network model, the ReLU activation function is used for air temperature data, the Sigmoid activation function is used for air volume data, the Tanh activation function is used for powder volume data, the Leaky ReLU activation function is used for dynamic CO evolution data of coal quality, and the Softmax activation function is used for the remaining neurons.

[0012] Furthermore, when the air temperature data exceeds the preset air temperature standard value, the weight corresponding to the air temperature data is increased; when the air volume data exceeds the preset air volume standard value, the weight corresponding to the air volume data is decreased; when the powder volume data exceeds the preset powder volume standard value, the weight corresponding to the powder volume data is increased; when the dynamic CO evolution data of coal quality exceeds the preset coal quality standard value, the weight corresponding to the dynamic CO evolution data of coal quality is decreased.

[0013] Furthermore, the ReLU function is selected for the hidden layer neurons in the BP neural network model; starting from the input layer, the output of the neurons is calculated layer by layer; for each neuron, its input is the weighted sum of the outputs of all neurons in the previous layer, plus the bias term, and then the output is obtained through the activation function.

[0014] Furthermore, the loss function of the BP neural network is:

[0015]

[0016] Among them, Loss is the current loss; Original·Loss is the original loss; W i is the weight parameter in the model, i is a constant; n is the number of weights; λ is the hyperparameter of the regularization strength.

[0017] Furthermore, the batch normalization technique is adopted to standardize the input of each layer. For the data of each mini-batch, its mean and variance are calculated. Each sample is subtracted by the mean and then divided by the standard deviation, so that the distribution of the data becomes a mean of 0 and a variance of 1 in this mini-batch; two learnable parameter scaling factors and offset factors are introduced to restore the representation ability of the model.

[0018] In the second aspect, the present invention also provides a system for estimating the carbon monoxide concentration at the coal powder outlet, including:

[0019] The data acquisition module is configured to: obtain multiple relevant data at the outlet of the coal mill;

[0020] The prediction module is configured to: obtain the carbon monoxide concentration prediction result according to the multiple relevant data and the trained BP neural network model;

[0021] Wherein, in the loss function of the BP neural network model, the sum of the absolute values of the weights corresponding to the multiple relevant data is added as a penalty term; each weight is dynamically adjusted according to the value of the corresponding relevant data.

[0022] In a third aspect, the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the method for estimating the carbon monoxide concentration at the pulverized coal outlet described in the first aspect are implemented.

[0023] In a fourth aspect, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and capable of running on the processor, and when the processor executes the program, the steps of the method for estimating the carbon monoxide concentration at the pulverized coal outlet described in the first aspect are implemented.

[0024] In a fifth aspect, the present invention also provides a computer program product, the computer program product includes a computer program, and when the computer program is executed by a processor, the steps of the method for estimating the carbon monoxide concentration at the pulverized coal outlet described in the first aspect are implemented.

[0025] Compared with the prior art, the beneficial effects of the present invention are:

[0026] 1. In the present invention, the trained BP neural network model is used to predict the carbon monoxide concentration; in the loss function of the BP neural network model, the sum of the absolute values of the weights corresponding to the multiple relevant data is added as a penalty term, and each weight is dynamically adjusted according to the value of the corresponding relevant data, which prompts some network weights to become zero, helps to reduce the model complexity, improve the generalization ability, enhance the feature selection effect, and improve the model quality and prediction accuracy, etc.

[0027] 2. In the present invention, for the data of the air temperature detector, the ReLU activation function is selected; the ReLU function can effectively handle the non-linear relationship in the temperature data, improve the reaction sensitivity of the model to temperature changes, and thus improve the prediction accuracy. The data of the air volume detector is suitable for using the Sigmoid activation function; the Sigmoid function can map the air volume data to the interval (0, 1), which helps to capture the influence of air volume fluctuations on the CO concentration and enhance the stability of the model. The data of the powder volume detector selects the Tanh activation function; when processing the powder volume data, the Tanh function can better handle the situation of a large data range, avoid the problem of gradient disappearance, and improve the training efficiency and prediction accuracy of the model. The data of the coal quality CO dynamic precipitation detector is suitable for using the LeakyReLU activation function; Leaky ReLU can allow a small amount of negative values to pass through while maintaining the high gradient characteristics, adapt to the complexity of the dynamic changes of coal quality CO, and improve the robustness of the model. For the neurons of other remaining influencing factors, the Softmax activation function is selected; the Softmax function can convert multi-classification data into a probability distribution, which is suitable for dealing with the comprehensive influence of multiple factors and improving the model's ability to identify different factors.

[0028] 3. In the present invention, when the air temperature data exceeds the preset air temperature standard value, the weight corresponding to the air temperature data needs to be increased, as the air temperature has a greater impact on the outlet temperature of the coal mill. Correspondingly, the weights corresponding to the air volume data, powder volume data, and dynamic CO evolution data of coal quality need to be proportionally reduced. By increasing the weight of the air temperature data, the model will respond more sensitively to changes in the air temperature. This means that in the case of abnormal air temperature, the model can identify potential problems or risks more quickly, and thus take timely measures for adjustment or warning. When the air volume data exceeds the preset air volume standard value, the weight corresponding to the air volume data needs to be reduced. The change in the air volume is affected by other factors such as fan failures, so it is necessary to reduce its impact on the model. Correspondingly, the weights corresponding to the air temperature data, powder volume data, and dynamic CO evolution data of coal quality need to be proportionally increased. By reducing the weight of the air volume data, the negative impact of abnormal air volume data on the model prediction results can be reduced. This helps to improve the stability and reliability of the model. When the powder volume data exceeds the preset powder volume standard value, the weight corresponding to the powder volume data needs to be increased, as the change in the powder volume directly affects the working efficiency and outlet temperature of the coal mill, so more attention needs to be given. Correspondingly, the weights corresponding to the air temperature data, air volume data, and dynamic CO evolution data of coal quality need to be proportionally reduced. Since the powder volume is one of the key factors affecting the outlet temperature of the coal mill, increasing its weight in the model can improve the overall prediction accuracy. This helps to reduce errors and make the monitoring results more reliable. When the dynamic CO evolution data of coal quality exceeds the preset coal quality standard value, the weight corresponding to the dynamic CO evolution data of coal quality needs to be reduced. The change in the dynamic CO evolution data of coal quality is affected by various factors such as coal type changes and combustion condition changes, so it is necessary to reduce its impact on the model. Correspondingly, the weights corresponding to the air temperature data, air volume data, and powder volume data need to be proportionally increased. By reducing the weight of the dynamic CO evolution data of coal quality, the negative impact of abnormal dynamic CO data of coal quality on the model prediction results can be reduced. This helps to improve the stability and reliability of the model. Dynamically adjusting the weights enables the model to better adapt to different working environments and data changes, thereby improving the robustness and adaptability of the model. Description of the Drawings

[0029] The schematic diagrams forming a part of this embodiment are used to provide a further understanding of this embodiment. The illustrative embodiments and descriptions thereof are used to explain this embodiment and do not constitute an improper limitation of this embodiment.

[0030] Figure 1 It is a flowchart of the monitoring method according to Embodiment 1 of the present invention;

[0031] Figure 2 It is a front structural sectional view of the monitoring system according to Embodiment 2 of the present invention;

[0032] Among them, 1. box body; 2. universal wheel; 3. electric telescopic column; 4. mounting seat; 5. electric rotating shaft; 6. air temperature detector; 7. air volume detector; 8. powder volume detector; 9. coal quality CO dynamic evolution detector. Detailed implementation mode

[0033] The present invention will be further described below in conjunction with the drawings and embodiments.

[0034] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which this application belongs.

[0035] As recorded in the background art, the pulverizing system is a key part of a coal-fired power plant, and its safe operation is crucial for the stable operation of the entire power plant. In the pulverizing system, the coal mill is one of the core devices, and the change in the CO concentration at its outlet is directly related to the safe and stable operation of the system. Therefore, accurately predicting and warning the CO concentration at the outlet of the coal mill is of great significance for preventing explosion accidents. Traditional CO concentration prediction methods mainly rely on empirical formulas or simple mathematical models, which often cannot accurately reflect complex actual working conditions, resulting in low prediction accuracy; when deep learning is applied to the monitoring of the CO concentration at the outlet of the coal mill, only multiple characteristic parameters related to the CO concentration at the outlet of the coal mill are used to conduct conventional training on the neural network model, without better considering the influence of different types of characteristic parameters on the prediction results, affecting the neural network model and the prediction results.

[0036] To improve the above problems, this embodiment provides a method, system, medium and device for estimating the CO concentration at the coal powder outlet. By combining deep learning with online monitoring data, the CO concentration at the outlet of the coal mill can be predicted more accurately, thereby providing a more accurate basis for adjusting the operating parameters of the coal-fired power plant. This system can monitor the operating status of the coal mill in real time and issue an early warning in a timely manner when an abnormal situation is detected, which helps the operator take measures quickly and avoid potential safety accidents. By using an improved BP neural network, momentum method and learning rate adaptive method, the number of training times and time of the model can be reduced, the consumption of computing resources can be reduced, and the system efficiency can be improved. By optimizing the algorithm to improve the performance of the BP neural network, the problem of the network falling into a local optimal solution is effectively avoided, ensuring the global optimality of the prediction model. Accurate CO concentration prediction helps to maintain the stable operation of the pulverizing system and reduce the risks of production interruption and equipment damage caused by abnormal CO concentration.

[0037] Embodiment 1:

[0038] As Figure 1As shown in the figure, this embodiment provides a method for estimating the carbon monoxide concentration at the pulverized coal outlet, including the following steps:

[0039] S1. Collect the on-line monitoring data of the coal mill site;

[0040] S2. Select the key features related to the CO concentration at the coal mill outlet from the on-line detected data;

[0041] S3. Establish a BP neural network model with the extracted key features as the input and the CO concentration at the coal mill outlet as the output;

[0042] S4. Improve the process memory matrix through an algorithm;

[0043] S5. Use the test data set to evaluate the trained model, and calculate its prediction error and accuracy;

[0044] S6. Based on the trained BP neural network model, establish a CO early warning diagnosis system.

[0045] In some specific embodiments, the on-line monitoring data in step S1 is the air temperature data, air volume data, powder volume data, and dynamic CO precipitation data of coal quality at the coal mill site.

[0046] In some specific embodiments, for the air temperature data, air volume data, powder volume data, and dynamic CO precipitation data of coal quality, calculate the Z-score of each data point, and regard the data points with |Z|>3 as abnormal and process them to check whether there are abnormal values deviating from the normal range; for abnormal values, use methods such as deletion, replacement, or correction for processing, and use the interpolation algorithm to fill in the missing values in the data set. Use Z-score standardization to scale each item of data so that it is at the same order of magnitude. Combine all the processed data into a data set, ensure that the timestamps are aligned, divide the data set into a training set and a test set with a ratio of 80:20, and divide the training set into a training subset and a validation subset for model selection and tuning.

[0047] Specifically, for the stored historical data, it is necessary to be sorted and archived for subsequent analysis. Check whether there are outliers in the dataset that deviate significantly from the normal range, such as too high or too low temperature, sudden increase or decrease in air volume, etc. For outliers, methods such as deletion, replacement, or correction can be used. For the missing values in the dataset, depending on the specific situation, records containing missing values can be deleted, statistical methods such as mean / median / mode can be used to fill in the missing values, or more complex interpolation algorithms can be used for filling. Check whether there are duplicate records in the dataset and delete the duplicates to ensure the uniqueness of the data. Convert data with different dimensions to the same dimension for subsequent analysis and modeling. Optional standardization methods include Z-score standardization, Min-Max standardization, etc.

[0048] In some specific embodiments, in step S2, the key features related to the CO concentration at the outlet of the coal mill are selected from the online detected data as the air temperature, air volume, powder volume, and dynamic CO precipitation of coal quality.

[0049] Specifically, standardize the data with different dimensions to make them at the same order of magnitude for subsequent analysis. For example, standardize the air volume (unit: t / h) and powder volume (unit: t / h) so that their numerical ranges are between [0, 1].

[0050] In some specific embodiments, in step S3, a BP neural network model is established with the extracted key features as the input and the CO concentration at the outlet of the coal mill as the output; including using the Min-Max standardization method for the data of air temperature, air volume, powder volume, dynamic CO precipitation of coal quality, and the CO concentration at the outlet of the coal mill to make the data at the same order of magnitude, randomly selecting 20% of the data for model testing, and 80% of the data for training.

[0051] In some specific embodiments, in step S3, with the extracted key features as the input and the CO concentration at the outlet of the coal mill as the output, for the data of the air temperature detector, select the ReLU activation function; for the data of the air volume detector, use the Sigmoid activation function; for the data of the powder volume detector, select the Tanh activation function; for the data of the dynamic CO precipitation detector of coal quality, it is suitable to use the Leaky ReLU activation function; for the remaining neurons, select the Softmax activation function.

[0052] Specifically, normalize the data of input features (such as air temperature, air volume, powder volume, dynamic evolution of coal quality CO) and output target (CO concentration at the outlet of the coal mill) so that they are at the same order of magnitude, which is convenient for network training. The Min-Max normalization method can be used to convert all data into the range of [0, 1]. Generally, 20% of the data is randomly selected for model verification, and 80% of the data is used for training. This can evaluate the performance of the model on the data not involved in training and ensure the generalization ability of the model. Set the number of hidden layers of the network and the number of neurons in each layer, and set one hidden layer. For example, a hidden layer containing 10 neurons can be set. Select a suitable activation function for each neuron. Optional activation functions include ReLU, Sigmoid, Tanh, etc. For hidden layer neurons, the ReLU function can be selected because it has good gradient characteristics and can alleviate the problem of gradient disappearance. Starting from the input layer, calculate the output of neurons layer by layer. For each neuron, its input is the weighted sum of the outputs of all neurons in the previous layer, plus the bias term, and then the output is obtained through the activation function.

[0053] For the data of the air temperature detector, select the ReLU activation function; the ReLU function can effectively handle the non-linear relationship in the temperature data, improve the response sensitivity of the model to temperature changes, and thus improve the prediction accuracy. The data of the air volume detector is suitable for using the Sigmoid activation function; the Sigmoid function can map the air volume data to the interval (0, 1), which helps to capture the impact of air volume fluctuations on the CO concentration and enhance the stability of the model. The data of the powder volume detector selects the Tanh activation function; when dealing with the powder volume data, the Tanh function can better handle the situation of a large data range, avoid the problem of gradient disappearance, and improve the training efficiency and prediction accuracy of the model. The data of the dynamic evolution detector of coal quality CO is suitable for using the Leaky ReLU activation function; Leaky ReLU can allow a small amount of negative values to pass through while maintaining high gradient characteristics, adapt to the complexity of the dynamic changes of coal quality CO, and enhance the robustness of the model. For neurons of other remaining influencing factors, select the Softmax activation function; the Softmax function can convert multi-classification data into a probability distribution, which is suitable for dealing with the comprehensive influence of multiple factors and improving the recognition ability of the model for different factors.

[0054] In some specific embodiments, in step S4, the algorithm for improving the process memory matrix by the algorithm can be optionally Adam (Adaptive Moment Estimation).

[0055] Introduce L1 regularization to the BP neural network, and its mathematical expression is:

[0056]

[0057] Among them, Loss is the current loss; Original·Loss is the original loss; W i is the weight parameter in the model, i is a constant, and optional i=1, 2, 3, 4 represent wind temperature, air volume, powder volume and dynamic precipitation of coal quality CO respectively; n is the number of weights; λ is a hyperparameter of regularization strength.

[0058] Batch normalization technology is used to standardize the input of each layer. For each mini-batch of data, its mean and variance are calculated, the mean is subtracted from each sample, and then divided by the standard deviation, so that the distribution of data in the mini-batch becomes a mean of 0 and a variance of 1. Two learnable parameters, the scaling factor gamma and the offset factor beta, are introduced to restore the representation ability of the model.

[0059] Optionally, the weight parameters of wind temperature data, air volume data, powder volume data, and coal quality CO dynamic precipitation data are set to 0.2, 0.2, 0.2, and 0.4 respectively; the gamma value of wind temperature data is 0.1, and the beta value is 0.01; the gamma value of air volume data is 0.5, and the beta value is 0.05; the gamma value of powder volume data is 0.3, and the beta value is 0.02; the gamma value of coal quality CO dynamic precipitation data is 0.05, and the beta value is 0.01. Set the wind temperature data hyperparameter value to 65℃, and the hyperparameter limit range is 60℃~70℃; set the air volume data hyperparameter value to 50t / h, and the hyperparameter limit range is 20t / h~71.9t / h; set the powder volume data hyperparameter value to 60t / h, and the hyperparameter limit range is 19.95t / h~71.9t / h; set the coal quality CO dynamic precipitation data hyperparameter value to 50ppm, and the hyperparameter limit range is 0ppm~5000ppm. Hyperparameters are tuned using a Bayesian optimization learning algorithm.

[0060] Since CO mainly comes from coal quality, its weight parameter value is relatively large: when each data is within the corresponding numerical range, by giving a higher weight to the dynamic precipitation data of CO from coal quality, the system can detect abnormal conditions earlier, issue early warning signals in time, and effectively prevent accidents; wind temperature, air volume, and powder volume are all important factors affecting the CO concentration at the outlet of the coal mill. By setting their weights to 0.2, it can be ensured that these key factors are balanced in the model to avoid excessive deviation of the prediction results by a certain factor).

[0061] When the air temperature data exceeds the preset air temperature standard value, the weight corresponding to the air temperature data needs to be increased. The air temperature has a greater impact on the outlet temperature of the coal mill. Correspondingly, the weights corresponding to the air volume data, powder volume data, and dynamic CO evolution data of coal quality need to be proportionally reduced. By increasing the weight of the air temperature data, the model will respond more sensitively to changes in the air temperature. This means that in the case of abnormal air temperature, the model can identify potential problems or risks more quickly, so as to take measures for adjustment or warning in a timely manner.

[0062] When the air volume data exceeds the preset air volume standard value, the weight corresponding to the air volume data needs to be reduced. The change in the air volume is affected by other factors, such as fan failure, etc., so it is necessary to reduce its impact on the model. Correspondingly, the weights corresponding to the air temperature data, powder volume data, and dynamic CO evolution data of coal quality need to be proportionally increased. By reducing the weight of the air volume data, the negative impact of abnormal air volume data on the model prediction result can be reduced. This helps to improve the stability and reliability of the model.

[0063] When the powder volume data exceeds the preset powder volume standard value, the weight corresponding to the powder volume data needs to be increased. The change in the powder volume directly affects the working efficiency and outlet temperature of the coal mill, so more attention needs to be paid. Correspondingly, the weights corresponding to the air temperature data, air volume data, and dynamic CO evolution data of coal quality need to be proportionally reduced. Since the powder volume is one of the key factors affecting the outlet temperature of the coal mill, increasing its weight in the model can improve the overall prediction accuracy. This helps to reduce errors and make the monitoring results more reliable.

[0064] When the dynamic CO evolution data of coal quality exceeds the preset coal quality standard value, the weight corresponding to the dynamic CO evolution data of coal quality needs to be reduced. The change in the dynamic CO evolution data of coal quality is affected by various factors, such as coal type change, combustion condition change, etc., so it is necessary to reduce its impact on the model. Correspondingly, the weights corresponding to the air temperature data, air volume data, and powder volume data need to be proportionally increased. By reducing the weight of the dynamic CO evolution data of coal quality, the negative impact of abnormal dynamic CO data of coal quality on the model prediction result can be reduced. This helps to improve the stability and reliability of the model. Dynamically adjusting the weights enables the model to better adapt to different working environments and data changes, thereby improving the robustness and adaptability of the model.

[0065] Specifically, Adam combines the advantages of AdaGrad and RMSProp and further improves them. It not only considers the first moment estimate of the gradient (i.e., the momentum term), but also the second moment estimate (i.e., root mean square propagation). The momentum term can help accelerate convergence, especially in cases where the data is sparse or the parameter space has a high dimension. Root mean square propagation, on the other hand, can make a more refined adjustment to the learning rate, making the parameter updates more stable. Adam also performs bias correction on these two moment estimates, further improving the stability and performance of the algorithm. Combining the process memory matrix with the learning rate adaptation method can further enhance the performance and prediction accuracy of the BP neural network model. When using Adam as the optimization algorithm, the learning rate can be dynamically adjusted according to the historical gradient information recorded in the process memory matrix. Specifically, the gradient information in the process memory matrix can be used as part of the second moment estimate in the Adam algorithm or to assist in calculating the momentum term, thus making the adjustment of the learning rate more accurate and reasonable. This combination method can, on the basis of retaining the original advantages of the process memory matrix, make full use of the advantages of the learning rate adaptation method to further improve the learning efficiency and prediction ability of the model.

[0066] In the BP neural network, L1 regularization promotes some network weights to become zero by adding the sum of the absolute values of the weights as a penalty term to the loss function. This method helps reduce the model complexity, improve the generalization ability, and enhance the feature selection effect. The batch normalization technique stabilizes the distribution of data during training, accelerates the convergence speed, and improves the generalization ability of the model by normalizing the input of each layer. The specific operations include calculating the mean and variance of each batch of data and performing normalization. The improved BP neural network can accurately and efficiently identify the state of the coal mill, reduce the number of model training times and training time, effectively inhibit the risk of the BP neural network falling into local optimum, and improve the prediction accuracy of the CO concentration at the outlet of the coal mill.

[0067] The air temperature detector is mainly used to measure the temperature of the air, and its data usually has large fluctuations, especially in different seasons or weather conditions. Due to the large fluctuations in air temperature data, the gamma value should be set to a small positive number 0.1 to avoid over-amplifying the data and causing model instability. The beta value should be set to a decimal number close to zero, 0.01, to keep the central position of the data relatively stable.

[0068] The air volume detector is used to measure the volume of air passing through a certain cross-section per unit time, and its data may be affected by various factors, such as the operating state of the equipment, environmental conditions, etc. The fluctuations in air volume data may not be as large as those in air temperature data. Therefore, the gamma value is a medium-sized positive number 0.5 to appropriately adjust the scale of the data. The beta value should also be set to a small positive number 0.05 to keep the data stable.

[0069] The powder quantity detector is used to measure the weight or flow rate of powdery materials, and its data may be affected by various factors such as material properties and equipment accuracy. The volatility of the powder quantity data may be relatively large, but it may also have a certain stability due to the limitations of material properties. Therefore, the gamma value is set to a moderate positive number 0.3 to balance the adjustment of the data scale. The beta value should be set to a decimal number 0.02 close to zero to keep the central position of the data stable.

[0070] The coal quality CO dynamic evolution detector is used to measure the dynamic evolution of CO during the combustion of coal, and its data may contain complex non-linear relationships. Since the coal quality CO data may have relatively large volatility and complexity, the gamma value should be set to a small positive number 0.05 to avoid over-amplifying the noise in the data. The beta value should also be set to a decimal number 0.01 close to zero to keep the central position of the data stable.

[0071] In some specific embodiments, in step S5, the trained model is evaluated using the test data set, and calculating its prediction error and accuracy includes: evaluating the trained BP neural network model using the test data set, and calculating the error between the predicted value and the measured target value of the model through the mean square error.

[0072] In some specific embodiments, in step S5, evaluating the trained model using the test data set and calculating its prediction error and accuracy further includes setting an error threshold, and calculating the accuracy of the model through the error threshold.

[0073] Specifically, the trained BP neural network model is evaluated using the test data set, and the error between the predicted value of the model and the actual target value (CO concentration at the outlet of the coal mill) is calculated. Common error metrics include mean square error (MSE), mean absolute error (MAE), etc. If the predicted CO concentration is 50 ppm and the actual value is 48 ppm, the error of this sample is 2 ppm. By statistically analyzing the errors of all test samples, the overall prediction error level can be obtained. An error threshold (such as ±5 ppm) can be set. If the prediction error of the model is within this threshold range, the prediction is considered accurate. Then calculate the proportion of accurately predicted test samples as the accuracy of the model. According to the performance of the model on the test set, consider adjusting the structure of the BP neural network, including increasing or decreasing the number of hidden layers, changing the number of neurons in each layer, etc. For example, if it is found that the prediction error of the model is relatively large and the training time is relatively long, one can try to add a hidden layer or appropriately reduce the number of neurons in each layer to improve the learning ability of the model.

[0074] Example 2:

[0075] As Figure 2 shown, in order to cooperate with the implementation of the method in Embodiment 1, this embodiment provides a carbon monoxide concentration monitoring system for the outlet of a coal mill, including:

[0076] A box body 1, with a plurality of universal wheels 2 provided at the bottom of the box body 1, and a chamber is provided inside the box body 1;

[0077] An electric telescopic column 3, which is arranged on the top of the box body 1. The bottom mounting end of the electric telescopic column 3 is fixedly connected to the top of the box body 1, and the electric telescopic column 3 coincides with the axis of the box body;

[0078] A mounting seat 4, the bottom of the mounting seat 4 is rotationally connected to the telescopic end of the top of the electric telescopic column 3 through an electric rotating shaft 5. A wind temperature detector 6, an air volume detector 7, a powder volume detector 8, and a coal quality CO dynamic precipitation detector 9 are provided on the top of the mounting seat 4. The wind temperature detector 6, the air volume detector 7, the powder volume detector 8, and the coal quality CO dynamic precipitation detector 9 transmit data to a data processing center through a wireless transmission method.

[0079] Specifically, in the data processing center, the trained BP neural network model is used to process and analyze the collected data to predict the outlet CO concentration value of the coal mill. Whether there is a potential safety hazard is judged according to the predicted CO concentration value. If the CO concentration exceeds the preset threshold, the system will issue a warning signal to remind the operator to take measures for processing in time.

[0080] The working method of the data processing center is the same as that of the coal powder outlet carbon monoxide concentration estimation method in Embodiment 1, and will not be elaborated here.

[0081] Embodiment 3:

[0082] This embodiment provides a coal powder outlet carbon monoxide concentration estimation system, including:

[0083] A data acquisition module, configured to: obtain a plurality of relevant data at the outlet of the coal mill;

[0084] A prediction module, configured to: obtain a carbon monoxide concentration prediction result according to a plurality of relevant data and the trained BP neural network model;

[0085] Among them, in the loss function of the BP neural network model, the sum of the absolute values of the weights corresponding to a plurality of relevant data is added as a penalty term; each weight is dynamically adjusted according to the value of the corresponding relevant data.

[0086] The working method of the system is the same as that of the coal powder outlet carbon monoxide concentration estimation method in Embodiment 1, and will not be elaborated here.

[0087] Embodiment 4:

[0088] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored on the memory and capable of running on the processor. When the processor executes the program, the steps of the method for estimating the carbon monoxide concentration at the pulverized coal outlet described in Embodiment 1 are implemented.

[0089] Embodiment 5:

[0090] This embodiment provides a computer program product. The computer program product includes a computer program. When the computer program is executed by a processor, the steps of the method for estimating the carbon monoxide concentration at the pulverized coal outlet described in Embodiment 1 are implemented.

[0091] The foregoing is only the preferred embodiment of this embodiment and is not intended to limit this embodiment. For those skilled in the art, this embodiment can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of this embodiment shall be included within the protection scope of this embodiment.

Claims

1. A method for estimating carbon monoxide concentration at a pulverized coal outlet, characterized in that: include: Obtain multiple relevant data of coal mill outlet; According to multiple relevant data and the trained BP neural network model, the carbon monoxide concentration prediction result is obtained; Among them, the absolute value sum of weights corresponding to multiple relevant data is added as a penalty term in the loss function of the BP neural network model; each weight is dynamically adjusted according to the value of the corresponding relevant data.

2. The method for estimating carbon monoxide concentration at a pulverized coal outlet according to claim 1, characterized in that: The relevant data include wind temperature data, wind volume data, powder volume data and coal quality CO dynamic precipitation data.

3. The method for estimating carbon monoxide concentration at a pulverized coal outlet according to claim 2, characterized in that: In the BP neural network model, the wind temperature data uses the ReLU activation function, the wind volume data uses the Sigmoid activation function, the powder volume data uses the Tanh activation function, the coal quality CO dynamic precipitation data uses the Leaky ReLU activation function, and the remaining neurons use the Softmax activation function.

4. The method for estimating the carbon monoxide concentration at the outlet of pulverized coal as claimed in claim 2, characterized in that: When the wind temperature data exceeds the preset wind temperature standard value, the weight corresponding to the wind temperature data is increased; when the wind volume data exceeds the preset wind volume standard value, the weight corresponding to the wind volume data is reduced; When the powder quantity data exceeds the preset powder quantity standard value, the weight corresponding to the powder quantity data is increased; when the coal quality CO dynamic precipitation data exceeds the preset coal quality standard value, the weight corresponding to the coal quality CO dynamic precipitation data is reduced.

5. The method for estimating carbon monoxide concentration at a pulverized coal outlet according to claim 1, characterized in that: The hidden layer neurons in the BP neural network model select the ReLU function; starting from the input layer, the output of the neurons is calculated layer by layer; for each neuron, its input is the weighted sum of the outputs of all neurons in the previous layer, plus a bias term, and then the output is obtained through the activation function.

6. The method for estimating carbon monoxide concentration at a pulverized coal outlet according to claim 1, characterized in that: The loss function of the BP neural network is: Among them, Loss is the current loss; Original·Loss is the original loss; W i is the weight parameter in the model, i is a constant; n is the number of weights; λ is a hyperparameter of the regularization strength.

7. The method for estimating carbon monoxide concentration at a pulverized coal outlet according to claim 1, characterized in that: Batch normalization technology is used to standardize the input of each layer. For each mini-batch of data, its mean and variance are calculated, the mean is subtracted from each sample, and then divided by the standard deviation, so that the distribution of data in the mini-batch becomes a mean of 0 and a variance of 1. Two learnable parameters, scaling factor and offset factor, are introduced to restore the representation ability of the model.

8. A system for estimating carbon monoxide concentration at a pulverized coal outlet, characterized in that: include: The data acquisition module is configured to: obtain a plurality of relevant data of the coal mill outlet; The prediction module is configured to: obtain a prediction result of carbon monoxide concentration according to a plurality of relevant data and a trained BP neural network model; Among them, the absolute value sum of weights corresponding to multiple relevant data is added as a penalty term in the loss function of the BP neural network model; each weight is dynamically adjusted according to the value of the corresponding relevant data.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that: When the processor executes the program, the steps of the method for estimating the carbon monoxide concentration at the coal powder outlet as described in any one of claims 1-7 are implemented.

10. A computer program product, characterized in that The computer program product comprises a computer program, and when the computer program is executed by a processor, the steps of the method for estimating the carbon monoxide concentration at a pulverized coal outlet as claimed in any one of claims 1 to 7 are implemented.

Images