A coal quality on-line detection method and system

Abstract

The present application relates to the technical field of coal detection, and particularly relates to a coal quality online detection method and system. The method comprises: obtaining values of each dimension at each time during historical coal detection; determining noise possibility of each dimension at each time; determining correction values of each dimension at each time; determining feature importance of each dimension at each time; determining adjusted output of each gate of an LSTM model in each dimension at each time; and training a coal quality detection model constructed by the LSTM model based on the adjusted output, so as to realize online detection of coal quality. The present application dynamically distinguishes noise possibility in each dimension data at each time based on statistical features such as correlation mean value of a target time and a target dimension, mean value and standard deviation, and corrects the noise, thereby significantly improving noise interference problems caused by complex environmental influences such as high dust, high humidity, vibration and electromagnetic interference on sensor signals.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of coal detection, and in particular to a coal quality online detection method and system. BACKGROUND

[0002] In coal quality online detection, the physical and chemical properties of coal samples (such as ash content, sulfur content, and calorific value) directly affect the subsequent combustion efficiency, pollution emission, and resource assessment. Therefore, there is an urgent need for high-precision and real-time online detection means. Currently, multi-modal sensors (such as infrared spectrometers, gamma / neutron detectors, temperature sensors, etc.) can be used for non-destructive sensing of coal state and form time series data for modeling and prediction. To model the long-term dependence and complex dynamic changes in these data, the Long Short-Term Memory network (LSTM) as an improved recurrent neural network (RNN) model, with its unique gating mechanism, can effectively alleviate the gradient vanishing problem in traditional neural networks.

[0003] However, the coal transmission site is often in a high-dust, high-humidity, vibration, and electromagnetic interference environment, which seriously interferes with the signal quality of various sensors. The information entropy of the output of some sensors is increased after being disturbed, and the data fluctuates violently, making it difficult for traditional LSTM models to accurately distinguish between noise and effective features, and thus misidentifying the interference signal as an effective feature. At the same time, there is a complex nonlinear coupling relationship between coal quality parameters and multi-source sensor data, and the discrimination ability of each dimension under different working conditions is significantly different. The LSTM model defaults to equal weight input when processing such multi-modal data, lacks dynamic recognition ability for "key dimensions", and is prone to overfitting or weak generalization performance, which reduces the accuracy of subsequent coal quality online detection. SUMMARY

[0004] To solve the problem of poor model robustness caused by environmental interference on site and the inability to dynamically identify key feature dimensions, which reduces the accuracy of subsequent coal quality online detection, the present application provides a coal quality online detection method and system.

[0005] In the first aspect, the present application provides a coal quality online detection method, which adopts the following technical solution:

[0006] The application discloses a coal quality online detection method, which comprises the following steps: acquiring values of each dimension at each time in historical coal detection; taking any time as a target time and any dimension as a target dimension; determining a noise possibility of the target time in the target dimension according to a mean value of correlation between values of the target time in the target dimension and values in other dimensions, a value of the target time in the target dimension, a mean value of the target time in the target dimension, and a standard deviation of the target time in the target dimension; correcting the value of the target time in the target dimension based on the noise possibility to obtain a corrected value of the target time in the target dimension; determining a feature importance of the target time in the target dimension according to a maximum value of information entropy and variance of the corrected value of the target time in the target dimension and information entropy and variance of the corrected value of each dimension in the target time; determining an adjusted output of each gate of an LSTM model in the target dimension of the target time according to a projection of an original output of each gate of the LSTM model in the target dimension of the target time, the feature importance, the corrected value of the target time in the target dimension, and a mean value of the corrected value of the target time in the target dimension; and training a coal quality detection model constructed by the LSTM model based on the adjusted output to realize online detection of coal quality.

[0007] The beneficial effects are as follows: the noise possibility in the data of each time and each dimension is dynamically determined and the noise is corrected based on the mean value of correlation between the target time and the target dimension and statistical characteristics such as the mean value and the standard deviation, so that the noise interference problem caused by the influence of complex environments such as high dust, high humidity, vibration and electromagnetic interference on the sensor signal is significantly improved; the feature importance is measured by using the information entropy and the variance of the corrected value, the dynamic identification and adaptive weight adjustment of the key features in different dimensions and different time periods are realized, the limitation of the default equal weight input of the traditional LSTM model to the multi-modal data is solved, and the expression ability of the complex nonlinear coupling relationship of the coal quality parameters is enhanced; the gating mechanism of the model is dynamically adjusted by combining the feature importance and the adjusted output of each gate of the LSTM model, the model focuses more on the effective features and the key dimensions, the robustness and generalization ability of the model are enhanced, the overfitting risk is reduced, and the accuracy of subsequent online detection of the coal quality is effectively improved.

[0008] Further, the dimensions include infrared spectrum dimension, temperature dimension, pressure dimension, vibration dimension and humidity dimension.

[0009] Further, the values of each dimension at each time in the historical coal detection are values after maximum-minimum normalization processing.

[0010] Further, the noise possibility satisfies:

[0011] ; in the formula, is the noise possibility of the k-th time point in dimension d, is the average of the correlation of the value of the time period to which the k-th time point belongs in dimension d with the values in the remaining dimensions, is the value of the k-th time point in dimension d, is the average of the value of the time period to which the k-th time point belongs in dimension d, is the value of the k-th time point in dimension d, is the average of the value of the time period to which the k-th time point belongs in dimension d, is the standard deviation of the value of the time period to which the k-th time point belongs in dimension d, is a hyperparameter, is a max-max normalization function.

[0012] The beneficial effects are that by introducing the average of the correlation as a weight, combined with the standard deviation of the target time point and dimension, the degree to which each data point deviates from the trend of its time period can be quantified, and the noise determination probability of data points with high correlation with the remaining dimensions is reduced, so that effective signals and noise can be more accurately distinguished; the introduction of the standard deviation σ ensures that the calculation of the noise possibility not only depends on the deviation of the data value, but also considers the fluctuation range of the data and the noise tolerance, improving the adaptability and robustness of the model to different working conditions and environmental fluctuations.

[0013] Further, the correlation uses a Pearson correlation coefficient.

[0014] Further, the correction value satisfies:

[0015] ; in the formula, is the correction value of the k-th time point in dimension d, is the value of the k-th time point in dimension d, is the value of the k-th time point in dimension d, is the noise possibility of the k-th time point in dimension d, is the average of the value of the time period to which the k-th time point belongs in dimension d.

[0016] ​​​​​​​​​​​​​​​The beneficial effects are as follows: by dynamically adjusting the contribution ratio of the original observation value and the mean of the time period through the noise probability weight, it is possible to selectively and smoothly correct abnormal data that may be affected by noise to the stable time period mean, thereby reducing the impact of noise on the model input; for data points with low noise probability, the corrected value is closer to the original observation value, ensuring the preservation of effective signals and feature integrity, and avoiding the loss of important information caused by the uniform processing of all data by traditional simple filtering.

[0017] Furthermore, the importance of the features satisfies:

[0018] In the formula, For the first At any moment in dimension Feature importance, For the first The time period to which each moment belongs is in the dimension The information entropy of the corrected value, For the first The time period to which each moment belongs is in the dimension The variance of the correction value in For the first The maximum value of the information entropy of the corrected values ​​in each dimension for the time period to which a given moment belongs. For the first The maximum value among the variances of the correction values ​​in each dimension for the time period to which a given time point belongs. The number of dimensions.

[0019] The beneficial effects are as follows: By combining information entropy and variance, the complexity and volatility of data for each dimension within the target time period are comprehensively measured, enabling the scientific and dynamic identification of the key feature dimensions most discriminative for coal quality detection; normalization is achieved by using the maximum values ​​of information entropy and variance for each dimension, eliminating the influence of different units and data ranges, ensuring a reasonable distribution of feature importance across all dimensions, and guaranteeing that the model can fairly and effectively utilize multidimensional data; through the normalization of feature importance weights, the model can highlight the role of key dimensions and weaken the dependence on redundant or noisy dimensions, thereby improving the LSTM model's ability to capture important signals in complex nonlinear coupling environments, which helps to improve prediction accuracy and generalization performance.

[0020] Furthermore, the adjusted output satisfies:

[0021] In the formula, For the LSTM model The door in the Dimensions of a moment The adjusted output For the LSTM model the original output of the gate at the first time point in the dimension the projection of the original output of the gate at the first time point in the dimension the feature importance of the first time point in the dimension the correction value of the first time point in the dimension the average of the correction value of the time period to which the first time point belongs in the dimension is a learning rate parameter dynamically adjusted according to the model training process by using an adaptive learning rate algorithm, is a function.

[0022] The beneficial effects are that: by combining the feature importance with the correction value deviation and applying the weighted adjustment of the dynamic learning rate, the original gate output is effectively corrected, the model is more sensitive to the effective changes of the key dimensions, thereby enhancing the ability of the LSTM to capture complex nonlinear time series features; the feature importance weight enables the gate mechanism to dynamically amplify or suppress the information flow according to the importance of each dimension, which helps to distinguish key features from noise, strengthens the selective forgetting and memory function of the model, and avoids the performance degradation caused by the traditional LSTM treating all input dimensions equally; the adjusted output is mapped by the activation function to ensure that the gate output changes between 0 and 1, smoothly and stably controls the information flow, which helps to prevent gradient explosion or disappearance and improves the stability of model training.

[0023] Further, the training LSTM model constructs a coal quality detection model, including: taking the coal quality parameters corresponding to the values of each time point in each dimension of the historical coal detection as the labels of the LSTM model for constructing the coal quality detection model, taking the adjusted output of each dimension of each gate of the LSTM model at the target time point as a vector as the final output of each gate of the LSTM model at the target time point, and training the coal quality detection model of the LSTM model, wherein the coal quality parameters include ash content, sulfur content, calorific value, volatile matter, and moisture content.

[0024] In a second aspect, the present application provides a coal quality online detection system, which adopts the following technical solution:

[0025] A coal quality online detection system, comprising: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned coal quality online detection method is realized.

[0026] ​​​​​​​By adopting the technical scheme, the coal quality online detection method is generated into a computer program and stored in a memory to be loaded and executed by a processor, so that a terminal device is manufactured according to the memory and the processor, and use is facilitated.

[0027] The present application has the following technical effects:

[0028] By comprehensively evaluating the correlation, mean value and standard deviation of the data in each dimension at each time, noise data generated in environments such as high dust, high humidity, vibration and electromagnetic interference can be accurately identified and corrected, and the effectiveness and stability of the sensor signal can be significantly improved; based on the "feature importance" measurement of information entropy and variance, the model can automatically adjust the input weight of each dimension under different working conditions, highlight the discrimination ability of the key dimension, and avoid the feature redundancy or omission problem caused by the traditional LSTM weight input; the noise is pre-corrected and combined with the dynamic feature importance, so that the model focuses more on the real and effective quality features during the training process, thereby weakening the overfitting tendency caused by random interference and improving the generalization performance of the model; the LSTM coal quality detection model constructed by the present application can more accurately capture the nonlinear coupling relationship between the coal quality parameters and the multi-source sensor data, realize real-time and high-precision online quality evaluation, and meet the demand for fast and reliable detection in industrial field. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 is a method flowchart in a coal quality online detection method according to an embodiment of the present application. DETAILED DESCRIPTION

[0030] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0031] An embodiment of the present application discloses a coal quality online detection method, referring to Figure 1 , comprising steps S01-S06:

[0032] S01: Obtain the values in each dimension at each time during historical coal detection.

[0033] It should be noted that near-infrared spectrometers, temperature sensors, pressure sensors, vibration sensors and humidity sensors are deployed along the coal conveying belt of the coal mine, data is collected at intervals of 1 second for example, time synchronization is ensured, and data is normalized to map each feature value to The interval is used to obtain the median value of each dimension at each time point during historical coal testing; and the coal quality parameters (such as ash content, sulfur content, calorific value, etc.) corresponding to the values ​​of each dimension at each time point during historical coal testing are obtained through laboratory analysis.

[0034] Specifically, the dimensions include the infrared spectrum dimension, temperature dimension, pressure dimension, vibration dimension, and humidity dimension.

[0035] Specifically, the values ​​in each dimension at each moment during the historical coal detection are values ​​after max-min normalization processing.

[0036] S02: Determine the probability of noise in each dimension at each time point.

[0037] It should be noted that there are close physical connections between the data from various sensors at the coal testing site. Changes in coal quality often cause coordinated changes in multiple dimensions of data, while noise typically manifests as abnormal fluctuations in a single dimension. This is due to factors such as dust, water mist, vibration, and electromagnetic interference at the site, and their impact range and mode of action are localized and targeted. For example, electromagnetic interference usually only affects sensors with electromagnetic induction characteristics (such as pressure sensors and temperature sensors with electrical signal outputs) and will not simultaneously interfere with optical sensors such as infrared spectrometers. Dust adhering to the lens of a near-infrared spectrometer will only cause abnormal spectral signals and will not directly affect the vibration data of coal transportation collected by vibration sensors. These interference sources are often only targeted at specific types of sensors or the detection of a certain physical quantity, thus causing data fluctuations in a single dimension. Therefore, this step determines the probability of noise in each dimension at each time point by quantifying the correlation of data in each dimension.

[0038] Let any time be designated as the target time, and any dimension as the target dimension. Based on the mean of the correlation between the value of the target time period in the target dimension and the value in the other dimensions, the value of the target time in the target dimension, the mean of the target time period in the target dimension, and the standard deviation of the value of the target time period in the target dimension, determine the noise probability of the target time in the target dimension.

[0039] The implementers can set the number of times within the time period to which the target time belongs, for example, 15 times, depending on the specific implementation situation.

[0040] Specifically, the noise probability satisfies:

[0041] ;

[0042] In the formula, For the first At any moment in dimension The possibility of noise, For the first the average of the values of the time period to which the i-th time point belongs in dimension , the average of the values of the i-th time point in dimension , the average of the values of the time period to which the i-th time point belongs in dimension , the average of the values of the i-th time point in dimension , the standard deviation of the values of the time period to which the i-th time point belongs in dimension , the standard deviation of the values of the i-th time point in dimension , is a hyperparameter, is a max-max normalization function.

[0043] The hyperparameter can be set by the implementer according to the specific implementation, for example, 0.001, and the hyperparameter is present to prevent from making the formula meaningless.

[0044] Specifically, the correlation adopts a Pearson correlation coefficient.

[0045] reflects the degree of coordinated change of the data in the dimension and the data in the remaining dimensions (the Pearson correlation coefficient is used to measure linear correlation, the higher the average, the more consistent the change trend of the dimension and the remaining dimensions; otherwise, the weaker the correlation between the dimension and the remaining dimensions), the greater the value, the more independent the data change of the dimension from the remaining dimensions, the less consistent with the multi-dimensional coordinated fluctuation characteristics caused by coal quality change, then is greater; otherwise. represents the standardized deviation degree of the data of the i-th time point in dimension relative to the average of the time period (normalized by standard deviation), reflecting the abnormality degree of the i-th time point in dimension in the time period, the average and the standard deviation of the time period depict the normal fluctuation range, the greater the deviation degree, the more the data of the i-th time point exceeds the conventional fluctuation mode, the greater the value, the higher the deviation degree of the i-th time point from the historical fluctuation range, the more likely it is a single-dimensional abnormality caused by local interference (such as dust, electromagnetic interference), then is greater; otherwise.

[0046] S03: Determine the correction value of each time point in each dimension.

[0047] ​​​​​​​It should be noted that the noise in the coal detection data is usually represented as single-dimensional abnormal fluctuation, and the real coal quality change will cause multi-dimensional coordinated change. Therefore, the original data is weighted and corrected by the noise possibility in this step, while retaining the real signal and suppressing abnormal values caused by local interference, so that the data is more in line with the physical law of coal quality change.

[0048] Based on the noise possibility, the value of the target time in the target dimension is corrected to obtain a corrected value of the target time in the target dimension.

[0049] Specifically, the corrected value satisfies:

[0050]

[0051] In the formula, is the corrected value of the i th time in the dimension j, is the value of the i th time in the dimension j, is the noise possibility of the i th time in the dimension j, is the average value of the value of the time period to which the i th time belongs in the dimension j.

[0052] wherein, represents the retention degree of the original value in the correction, which reflects the possibility that the data at this time belongs to the real signal. When is small (low noise possibility), tends to 1, the original value accounts for a high proportion, which indicates that the data is more likely to be caused by coal quality change, and the corrected value is close to the original collection result. represents the weight of the time period average in the correction, which reflects the interference degree of the noise on the data. When is large (high noise possibility), tends to 1, the time period average accounts for a high proportion, which indicates that the data is more likely to be caused by local interference (such as dust, electromagnetic interference), and the corrected value is close to the normal fluctuation range (average).

[0053] S04: Determine the feature importance of each time in each dimension.

[0054] ​​​​​​​​​It should be noted that in the multi-dimensional data of coal testing, different dimensions contribute differently to coal quality parameters (such as ash content and calorific value). Furthermore, during coal mining, different coal seams and mining faces exhibit natural differences in ash content, sulfur content, and calorific value. For example, coal seams mined in the morning may be rich in kaolinite (the source of ash), and the infrared spectral dimension is significantly more sensitive to ash changes during this period. In the afternoon, when switching to a pyrite-rich layer, sulfur-related spectral features become dominant, and the focus of the infrared spectral dimension shifts from ash to sulfur. Therefore, the sensitivity of each dimension changes dynamically over time when coal quality changes. While LSTM models possess the ability to model long-term dependencies, their standard structure defaults to equal weighting of all input dimensions, which can easily introduce high noise or low-contribution signals, leading to problems such as mislearning and overfitting. Therefore, it is necessary to dynamically evaluate the importance of each feature before the data enters the LSTM model, thereby guiding the model to focus on the feature dimensions most sensitive to the current coal quality state. Therefore, this step quantifies the information richness and volatility of data in each dimension, dynamically assesses the importance of features in different dimensions at each time point, enhances the subsequent model's focus on key features, and strengthens the adaptability and robustness of the detection system to environmental disturbances.

[0055] The feature importance of the target time in the target dimension is determined by the information entropy and variance of the corrected value of the time period to which the target time belongs in the target dimension, and the maximum value of the information entropy and variance of the corrected value of the time period to which the target time belongs in each dimension.

[0056] Specifically, the importance of the features satisfies:

[0057] ;

[0058] In the formula, For the first At any moment in dimension Feature importance, For the first The time period to which each moment belongs is in the dimension The information entropy of the corrected value, For the first The time period to which each moment belongs is in the dimension The variance of the correction value in For the first The maximum value of the information entropy of the corrected values ​​in each dimension for the time period to which a given moment belongs. For the first The maximum value among the variances of the correction values ​​in each dimension for the time period to which a given time point belongs. The number of dimensions.

[0059] in, Indicates passage After normalization, the first The time period to which each moment belongs is in the dimension The uncertainty of this dimension reflects the diversity of its fluctuations within a time window. A larger value indicates a more uniform distribution and higher state diversity of the data in this dimension, potentially containing more high-entropy information related to coal quality changes. This suggests a stronger discriminative ability for this dimension, and it should be given a higher importance weight. The larger it is, the greater it becomes; conversely, the smaller it is, the greater it becomes. Indicates passage After normalization, the first The time period to which each moment belongs is in the dimension The degree of data dispersion reflects the intensity of fluctuation of the value of that dimension within a time window. The larger the value, the more drastic the change in that dimension during that period, potentially capturing dynamic signals related to coal quality changes, such as fluctuations in particle size, moisture, and ash content. Therefore, its weight should be increased. The larger it is, the smaller it is; conversely, the smaller it is, the larger it is. Finally, by dividing by... Ensure that the sum of the feature importance of all dimensions is 1.

[0060] S05: Determine the adjusted output of each gate of the LSTM model at each time step for each dimension.

[0061] It should be noted that in traditional LSTM structures, the states of the forget gate, input gate, and output gate are calculated from the input features and the hidden state. This approach, by default, treats all input dimensions with equal weight, making it difficult to distinguish the differences in the contributions of different sensor features to coal quality changes at a given moment. This can lead to high-noise dimensions misleading the gating state, thus affecting the updating and output of the memory cells. Therefore, to enhance LSTM's attention to high-discriminative dimensions, this step introduces a gating adjustment term based on feature importance into the LSTM's gating mechanism. This dynamically amplifies the influence of key dimension features on the gating state and suppresses interference from low-discriminative features.

[0062] Based on the projection of the original output of each gate of the LSTM model onto the target dimension at the target time (by mapping the standard forget gate output to the feature dimension space through linear transformation), the importance of the features, the correction value in the target dimension at the target time, and the mean of the correction values ​​in the target dimension for the time period to which the target time belongs, the adjusted output of each gate of the LSTM model in the target dimension at the target time is determined.

[0063] Specifically, the adjusted output satisfies:

[0064] ;

[0065] In the formula, For the LSTM model The door in Dimensions of a moment The adjusted output For the LSTM model The original output of the gate is in the first... Dimensions of a moment The projection, For the first At any moment in dimension Feature importance, For the first At any moment in dimension The correction value in For the first The time period to which each moment belongs is in the dimension The mean of the corrected values ​​in To utilize the adaptive learning rate algorithm to dynamically adjust the learning rate parameter according to the model training process, for function.

[0066] Implementers can set an adaptive learning rate algorithm based on the specific implementation situation, such as the Adam algorithm.

[0067] in, Indicates the LSTM number of The original output of each gate (forget gate / input gate / output gate) is in dimension The linear projection value reflects the fundamental influence of this dimension feature on the gated state in the traditional LSTM structure. This value is generated by matrix operations between the input features and hidden states of the LSTM, and its magnitude directly determines the degree of control of the gate signal on the memory unit. Reflects dimensions In the The strength of the ability to discriminate coal quality changes at any given time point is determined by the value. A larger value indicates that this dimension contains more significant information about coal quality changes in the current time period, and its contribution to the gating state should be increased. The larger it is, the greater it becomes; conversely, the smaller it is, the greater it becomes. Indicates the first At any moment in dimension The degree of deviation of the corrected value from the mean of the corresponding period reflects whether there is a sudden change or abnormal fluctuation in coal quality at the current moment. The larger the value, the more likely there may be a critical event such as a sudden change in coal quality, and therefore the response to this change should be enhanced. The larger it is, the greater it becomes; conversely, the smaller it is, the greater it becomes. It is used to control the adjustment range of feature importance and deviation on the gating, and can be dynamically adjusted according to the training process to avoid model oscillation or over-correction.

[0068] S06: Based on the adjusted output, train the LSTM model to construct a coal quality detection model to achieve online detection of coal quality.

[0069] Specifically, the training LSTM model builds a coal quality detection model, including:

[0070] The coal quality parameters corresponding to the values of each dimension at each time in the historical coal detection are acquired in advance as labels for the LSTM model to build the coal quality detection model, the adjusted output of each dimension of each gate of the LSTM model at the target time is constituted as a vector as the final output of each gate of the LSTM model at the target time, and the coal quality detection model of the LSTM model is trained, wherein the coal quality parameters include ash content, sulfur content, calorific value, volatile matter and moisture content.

[0071] After the model training is completed, the trained coal quality detection model is used to perform online coal quality detection on the coal to be detected, that is, the multi-modal sensor data of the coal to be detected is input into the trained coal quality detection model, the trained coal quality detection model is used to perform real-time prediction on the coal quality parameters of the coal to be detected, the online coal quality detection is completed, and the LSTM model is trained and optimized according to the actual situation to improve the adaptability of the model to new coal quality conditions and environmental changes, thereby continuously ensuring the accuracy of the online coal quality detection (for example, when it is detected that the ash content prediction error caused by the mining of a new coal type is more than 1.5% for three times in succession, or the online detection cycle reaches 10 times, new multi-modal data and corresponding coal quality parameters are acquired to retrain the model).

[0072] The embodiment of the present application also discloses a coal quality online detection system, comprising a processor and a memory, and the memory stores computer program instructions, which realize a coal quality online detection method according to the present application when executed by the processor.

[0073] The above system also includes a communication bus and a communication interface and other components well known to those skilled in the art, the settings and functions of which are known in the art, and thus will not be described here.

[0074] The above are preferred embodiments of the present application, and are not intended to limit the protection scope of the present application, therefore: any equivalent changes made on the structure, shape and principle of the present application should be covered within the protection scope of the present application.

Claims

1. A method for on-line detection of coal quality, characterized in that, Comprising: acquiring values of each dimension at each moment in the historical coal detection; each dimension comprises an infrared spectrum dimension, a temperature dimension, a pressure dimension, a vibration dimension, and a humidity dimension; each dimension data is acquired through a sensor; any moment is recorded as a target moment, and any dimension is recorded as a target dimension; The noise possibility of the target time in the target dimension is determined according to a mean of a correlation between a value of a time period to which the target time belongs in the target dimension and values in the remaining dimensions, a value of the target time in the target dimension, a mean of the time period to which the target time belongs in the target dimension, and a standard deviation of the value of the time period to which the target time belongs in the target dimension, and comprises: , is the noise possibility of the i-th time in the dimension , is the mean of the correlation between the value of the time period to which the i-th time belongs in the dimension , is the value of the i-th time in the dimension , is the value of the i-th time in the dimension , , , is the mean and the standard deviation of the value of the time period to which the i-th time belongs in the dimension , , is a hyperparameter, is a max-max normalization function; Based on the probability of noise, the value of the target time in the target dimension is corrected to obtain the corrected value of the target time in the target dimension, including: , For the first At any moment in dimension The correction value in the target dimension; based on the information entropy and variance of the correction value of the time period to which the target time belongs in the target dimension, and the maximum value of the information entropy and variance of the correction value of the time period to which the target time belongs in each dimension, the feature importance of the target time in the target dimension is determined; based on The projection of the raw output of each gate of the model onto the target dimension at the target time, the feature importance, the correction value of the target time in the target dimension, and the mean of the correction values ​​of the time period to which the target time belongs in the target dimension are used to determine the model. The adjusted output of each gate of the model at the target time dimension includes: , for The model's first The door in Dimensions of a moment The adjusted output for The model's first The original output of the gate is in the first... Dimensions of a moment The projection, For the first At any moment in dimension The importance of features For the first The time period to which each moment belongs is in the dimension The mean of the corrected values ​​in To utilize the adaptive learning rate algorithm to dynamically adjust the learning rate parameter according to the model training process, for function; Based on the adjusted output, training A coal quality detection model is constructed based on the model to realize online detection of coal quality.

2. The coal quality on-line detection method according to claim 1, characterized in that, the values of each dimension at each moment in the historical coal detection are values after maximum-minimum normalization processing.

3. The coal quality on-line detection method according to claim 1 or 2, characterized in that, the correlation adopts a Pearson correlation coefficient.

4. The coal quality on-line detection method according to claim 1, characterized in that, the feature importance satisfies: ； In the formula, For the first At any moment in dimension The importance of features For the first The time period to which each moment belongs is in the dimension The information entropy of the corrected value, For the first The time period to which each moment belongs is in the dimension The variance of the correction value in For the first The maximum value of the information entropy of the corrected values ​​in each dimension for the time period to which a given time point belongs. For the first The maximum value among the variances of the correction values ​​in each dimension for the time period to which a given time point belongs. The number of dimensions.

5. The coal quality on-line detection method according to claim 1, characterized in that, The training Model construction A coal quality detection model is constructed, comprising: The pre-acquired historical coal detection time value corresponding to the coal quality parameter in each dimension at each time is taken as The label of the coal quality detection model is constructed The adjusted output of each gate of the model in each dimension at the target time constitutes a vector, which is taken as The final input of each gate of the model at the target time is taken as The coal quality detection model of the model is trained, and the coal quality parameters include ash content, sulfur content, calorific value, volatile matter and moisture content.

6. An on-line coal quality detection system, characterized in that, Comprising: a processor and a memory, the memory storing computer program instructions which, when executed by the processor, implement the coal quality online detection method according to any one of claims 1-5.

Images