Coal Preparation Whole Process Monitoring and Decision-Making Method and System Based on Internet of Things Sensing

Abstract

The present invention provides a method and system for monitoring and decision-making for the entire coal preparation process based on Internet of Things sensing. First, a real-time sensing data set of multiple monitoring nodes in the coal preparation process is obtained, covering equipment vibration, medium density, and process flow data. Then, feature extraction is performed on it to obtain the equipment status, medium dynamics, and process stability features of each monitoring node. Then, based on a pre-trained monitoring and decision-making model, abnormal correlation analysis is performed on these features to generate an abnormal probability distribution and process adjustment parameters. According to the abnormal probability distribution, a priority processing queue is determined, and an optimization strategy for the coal preparation process is generated according to the process adjustment parameters. Finally, the optimization strategy for the coal preparation process is sent to the edge execution terminal to trigger the coal preparation process adjustment operation, realizing effective monitoring and decision-making optimization for the entire coal preparation process and improving the stability and efficiency of the coal preparation process.

Description

Technical Field

[0001] The present invention relates to the technical field of the Internet of Things, and in particular, to a method and system for monitoring and decision-making for the whole process of coal preparation based on Internet of Things sensing. Background Art

[0002] In the field of coal washing and processing, the efficient and stable operation of the coal preparation process is of crucial significance for improving coal quality, reducing production costs, and realizing the rational utilization of resources. Although the traditional regular detection method of equipment parameters can provide certain data support, the detection frequency is limited and it cannot reflect the dynamic changes of the coal preparation process in real time. If a device fails or the process experiences abnormal fluctuations between two detections, it is difficult to detect in time and take effective measures for adjustment.

[0003] With the development of Internet of Things technology, although some coal preparation plants have begun to attempt to introduce Internet of Things sensors to collect equipment data, the existing data collection and analysis methods still have deficiencies. On the one hand, the existing data collection is often limited to some key devices or links, lacking comprehensive coverage of multiple monitoring nodes in the whole process of coal preparation, resulting in incomplete data and unable to provide sufficient basis for subsequent decision-making. On the other hand, in terms of data analysis, most use simple threshold comparison or statistical analysis methods, which can only judge whether a single parameter exceeds the normal range and cannot deeply explore the potential laws and abnormal correlations behind the data. For complex abnormal situations in the coal preparation process, the existing analysis methods are difficult to accurately identify and locate, and even less able to generate scientific and reasonable process optimization strategies. Summary of the Invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, embodiments of the present invention provide a method for monitoring and decision-making for the whole process of coal preparation based on Internet of Things sensing, and the method includes:

[0005] Obtain a real-time sensing data set of multiple monitoring nodes in the coal preparation process, where the real-time sensing data set includes equipment vibration data, medium density data, and process flow data;

[0006] Perform feature extraction processing on the real-time sensing data set to obtain the equipment state feature, medium dynamic feature, and process stability feature of each monitoring node;

[0007] Based on a pre-trained monitoring and decision-making model, perform abnormal correlation analysis processing on the equipment state feature, medium dynamic feature, and process stability feature to generate an abnormal probability distribution and process adjustment parameters for each monitoring node;

[0008] Determine a priority processing queue according to the abnormal probability distribution, and generate a coal preparation process optimization strategy based on the process adjustment parameters;

[0009] Send the optimized coal preparation process strategy to the corresponding edge execution terminal to trigger the coal preparation process adjustment operation.

[0010] In another aspect, an embodiment of the present invention further provides a coal preparation full-process monitoring and decision-making system based on Internet of Things sensing, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0011] Based on the above aspects, the embodiment of the present invention first obtains the real-time sensing data set of multiple monitoring nodes in the coal preparation process. On this basis, feature extraction processing is performed on the real-time sensing data set to refine the device status features, medium dynamic features and process stability features of each monitoring node. Further, based on the pre-trained monitoring and decision-making model, abnormal correlation analysis processing is performed on the extracted above features, comprehensively considering the correlation relationship between equipment, medium and process, and can accurately generate the abnormal probability distribution and process adjustment parameters of each monitoring node. Finally, the generated optimized coal preparation process strategy is sent to the corresponding edge execution terminal to trigger the coal preparation process adjustment operation, realizing the seamless connection between monitoring and decision-making and actual execution, not only improving the automation and intelligence level of the coal preparation process, reducing the requirements and costs of manual intervention, but also being able to discover and solve abnormal situations in the coal preparation process in real time and accurately, effectively improving the coal preparation efficiency and quality, and reducing resource waste and environmental pollution. Description of the Drawings

[0012] Figure 1 It is a schematic execution flow diagram of the coal preparation full-process monitoring and decision-making method based on Internet of Things sensing provided by an embodiment of the present invention.

[0013] Figure 2 It is a schematic diagram of exemplary hardware and software components of the coal preparation full-process monitoring and decision-making system based on Internet of Things sensing provided by an embodiment of the present invention. Detailed Embodiments

[0014] The present invention will be specifically described below in conjunction with the accompanying drawings of the specification. Figure 1 It is a schematic flow diagram of the coal preparation full-process monitoring and decision-making method based on Internet of Things sensing provided by an embodiment of the present invention. The coal preparation full-process monitoring and decision-making method based on Internet of Things sensing will be introduced in detail below.

[0015] Step S110: Obtain the real-time sensing data set of multiple monitoring nodes in the coal preparation process. The real-time sensing data set includes equipment vibration data, medium density data and process flow data.

[0016] In this embodiment, in order to comprehensively monitor the operating status of the coal preparation process, multiple monitoring nodes can be set at each key position in the coal preparation plant. These monitoring nodes are distributed on different devices and pipelines, and each monitoring node is equipped with a corresponding sensor for real-time collection of equipment vibration data, medium density data, and process flow data. For example, vibration sensors are installed on equipment such as crushers, vibrating screens, and flotation machines to detect the vibration of the equipment during operation through the sensors and collect equipment vibration data. Density sensors are installed at positions such as medium tanks and pipelines to measure the density of the medium and obtain medium density data. Flow sensors are installed in places such as water pipelines and coal flow conveying channels to monitor the flow rate of the fluid in real time and obtain process flow data.

[0017] The above sensors transmit the collected data to the data collection center at regular time intervals (such as every second, every minute, etc.), thus forming a real-time sensing data set. Suppose there are 20 monitoring nodes set in the coal preparation plant. The equipment vibration data collected by each monitoring node within 1 minute is a sequence containing 120 data points, the medium density data is a sequence containing 60 data points, and the process flow data is a sequence containing 90 data points. Then the finally obtained real-time sensing data set is a multi-dimensional data set, containing various types of data of 20 monitoring nodes at different time points.

[0018] Step S120: Perform feature extraction processing on the real-time sensing data set to obtain the equipment status features, medium dynamic features, and process stability features of each monitoring node.

[0019] In this embodiment, the obtained real-time sensing data set is raw and unprocessed data. In order to better analyze the operating status of the coal preparation process, it is necessary to perform feature extraction processing on these data. For different types of data, different methods are used to extract the corresponding features.

[0020] Step S121: Perform time series data cleaning processing on the equipment vibration data, remove the data segments containing abnormal fluctuation intervals, and generate a standardized vibration data sequence.

[0021] In this embodiment, the equipment vibration data is a time series data sequence. Due to various reasons, such as equipment failures and external interferences, there may be abnormal fluctuation intervals. In order to ensure the accuracy of subsequent analysis, it is necessary to process these abnormal fluctuation intervals.

[0022] Step S1211: Use an adaptive threshold algorithm to detect the outliers in the vibration data and generate a vibration data sequence containing outlier marks.

[0023] In this embodiment, the adaptive threshold algorithm is an algorithm that dynamically adjusts the threshold according to the distribution of data. For the device vibration data sequence, first calculate the mean and standard deviation of the device vibration data. The mean is the average level of the device vibration data, and the standard deviation reflects the degree of dispersion of the device vibration data. For example, for a device vibration data sequence containing 100 data points, add these 100 data points and then divide by 100 to obtain the mean. By calculating the square of the difference between each data point and the mean, then finding the average of these squared values, and finally taking the square root to obtain the standard deviation. Then determine the threshold range according to a preset multiple (such as 3 times the standard deviation), and the data points outside this threshold range are outlier points. Compare each data point with the threshold range. If the data point exceeds the threshold range, mark it as an outlier point, thereby generating a vibration data sequence containing outlier point marks.

[0024] Step S1212: Identify the interval where the number of consecutive outlier points exceeds the threshold according to the outlier point marks, and generate the position information of the abnormal fluctuation segment.

[0025] In this embodiment, set a threshold for the number of consecutive outlier points, for example, 5. In the vibration data sequence containing outlier point marks, search for consecutive outlier points. When it is detected that the number of consecutive outlier points exceeds the threshold, mark this interval as an abnormal fluctuation segment and record its position information. For example, in a data sequence, from the 20th data point to the 26th data point are all outlier points, and the number of consecutive outlier points is 7, which exceeds the threshold of 5. Then mark this interval (from the 20th data point to the 26th data point) as an abnormal fluctuation segment, and record its starting position (the 20th data point) and ending position (the 26th data point), generating the position information of the abnormal fluctuation segment.

[0026] Step S1213: Use the generative repair model to reconstruct the vibration data corresponding to the position information of the abnormal fluctuation segment, and generate a repaired vibration data sequence.

[0027] In this embodiment, the generative repair model is a model that can learn the characteristics of the normal data distribution. In the training stage, this generative repair model will learn a large amount of normal device vibration data to understand the characteristics and laws of the normal data. After obtaining the position information of the abnormal fluctuation segment, the generative repair model predicts the data values that should appear at the position of the abnormal fluctuation segment according to the characteristics of the normal data distribution. For example, the generative repair model can analyze the normal data before and after the abnormal fluctuation segment, and generate reasonable data to replace the abnormal fluctuation segment according to the trends and characteristics of these normal data. In this way, the vibration data corresponding to the position information of the abnormal fluctuation segment is reconstructed to generate a repaired vibration data sequence.

[0028] Step S1214: Calculate the difference degree between the repaired vibration data sequence and the original vibration data in the frequency-domain energy distribution, and generate a frequency-domain difference degree index.

[0029] In this embodiment, in order to evaluate the quality of the repaired vibration data sequence, it is necessary to calculate the difference degree between it and the original vibration data in the frequency-domain energy distribution. First, perform Fourier transforms on the repaired vibration data sequence and the original vibration data to convert the time-domain data to the frequency domain and obtain spectrograms. In the spectrogram, different frequencies correspond to different vibration energies. Then, compare the energy distribution differences between the two at each frequency. The difference degree of the frequency-domain energy distribution can be obtained by calculating the absolute value of the energy difference at each frequency and then summing these differences. Normalize this difference degree to generate a frequency-domain difference degree index. For example, divide the difference degree by a preset reference value to obtain a frequency-domain difference degree index between 0 and 1.

[0030] Step S1215: When the frequency-domain difference degree index exceeds a preset safety threshold, trigger a sensor calibration instruction for the corresponding monitoring node.

[0031] In this embodiment, a safety threshold is preset, for example, 0.2. When the calculated frequency-domain difference degree index exceeds this safety threshold, it indicates that the difference in the frequency-domain energy distribution between the repaired vibration data and the original data is relatively large, and there may be a problem with the sensor. At this time, trigger a sensor calibration instruction for the corresponding monitoring node. This sensor calibration instruction will be sent to the device where the sensor is located, and the sensor is required to perform a calibration operation.

[0032] Step S1216: After performing a calibration operation on the target sensor according to the sensor calibration instruction, re-collect vibration data and replace the abnormal fluctuation segment to generate a standardized vibration data sequence.

[0033] In this embodiment, after receiving the sensor calibration instruction, perform a calibration operation on the target sensor. The calibration operation may include adjusting the parameters of the sensor, checking the connection of the sensor, etc. After calibration, re-collect vibration data. Replace the original abnormal fluctuation segment with the re-collected vibration data to obtain a new vibration data sequence. Perform standardization processing on this new vibration data sequence, for example, normalize the data so that its mean is 0 and the standard deviation is 1, to generate a standardized vibration data sequence.

[0034] Step S122: Analyze the spectral energy distribution of the standardized vibration data sequence, extract the distribution characteristics of the vibration energy within a preset frequency band and the main frequency offset characteristics, and combine the distribution characteristics and the offset characteristics into device state characteristics.

[0035] In this embodiment, after obtaining the standardized vibration data sequence, it is necessary to perform spectral energy distribution analysis on it to extract features related to the equipment state.

[0036] Step S1221: Perform spectral analysis on the standardized vibration data sequence to convert the time-domain data to the frequency domain.

[0037] In this embodiment, the fast Fourier transform (FFT) is used to perform spectral analysis on the standardized vibration data sequence. The fast Fourier transform is an efficient algorithm that can convert the vibration data in the time domain to the frequency domain to obtain a spectrogram. In the spectrogram, the horizontal axis represents the frequency, and the vertical axis represents the vibration energy. Through spectral analysis, the vibration data that is difficult to analyze in the time domain can be converted to the frequency domain, making it easier to observe and analyze the vibration energy distribution at different frequencies.

[0038] Step S1222: Set a preset frequency band according to the normal operating frequency range of the equipment.

[0039] In this embodiment, different equipment has different frequency ranges during normal operation. For example, if the normal operating frequency range of a certain crusher is between 10 Hz and 50 Hz, then this range is set as the preset frequency band. The setting of the preset frequency band is determined according to the design parameters and actual operating experience of the equipment, and it can reflect the main frequency components of the equipment in the normal working state.

[0040] Step S1223: Extract the distribution characteristics of the vibration energy within the preset frequency band.

[0041] In this embodiment, find the area corresponding to the preset frequency band in the spectrogram and count the vibration energy at different frequencies within this area. The preset frequency band can be divided into several small frequency intervals, and the total vibration energy within each frequency interval is calculated. For example, the preset frequency band from 10 Hz to 50 Hz is divided into 10 frequency intervals, each interval is 4 Hz, and the total vibration energy within each interval is calculated respectively. These energy sums are used as the distribution characteristics of the vibration energy within the preset frequency band. These distribution characteristics can reflect the vibration energy distribution of the equipment within the normal frequency range and are of great significance for judging the operating state of the equipment.

[0042] Step S1224: Extract the main frequency offset characteristics.

[0043] In this embodiment, the main frequency refers to the frequency with the largest vibration energy in the spectrogram. When the device is operating normally, the main frequency is usually in a relatively stable position. When a fault or anomaly occurs in the device, the main frequency may shift. By comparing the position of the main frequency in the current spectrogram with the main frequency position when the device is operating normally, the difference between the two is calculated to obtain the main frequency shift feature. For example, when the main frequency of the device during normal operation is 20 Hz and the main frequency in the current spectrogram is 22 Hz, the value of the main frequency shift feature is 2 Hz. The main frequency shift feature can reflect the change in the operating state of the device and is one of the important indicators for judging whether the device is abnormal.

[0044] Step S1225: Combine the distribution feature and the offset feature into a device state feature.

[0045] In this embodiment, the distribution feature of the vibration energy in the preset frequency band and the main frequency shift feature extracted are combined to form a device state feature. The distribution feature and the offset feature can be arranged in a certain order to form a feature vector. For example, the sum of the energies in the 10 frequency intervals of the distribution feature are arranged in sequence, and then the value of the main frequency shift feature is placed at the end to form a feature vector containing 11 elements. This feature vector can comprehensively reflect the operating state of the device.

[0046] Step S123: Perform density gradient change analysis on the medium density data, identify density mutation points and steady-state density intervals, extract density change trend features and steady-state maintenance features, and combine the trend features and the steady-state maintenance features into a medium dynamic feature.

[0047] In this embodiment, the medium density data reflects the density change of the medium during the coal preparation process. To better understand the dynamic characteristics of the medium, it is necessary to perform density gradient change analysis on the medium density data and extract relevant features.

[0048] Step S1231: Use a multi-scale sliding window algorithm to extract local change features and global trend features of the density data.

[0049] In this embodiment, the multi-scale sliding window algorithm is an algorithm that can analyze data at different scales. First, sliding windows of different sizes are selected. For example, the small window size is 5 data points and the large window size is 20 data points. These windows are slid over the density data sequence in turn, and the statistical features of the data in each window are calculated. For the small window, the local change of the data is mainly concerned, such as calculating the mean, standard deviation, etc. of the data in the window to obtain local change features. For the large window, the global trend of the data is concerned, such as calculating the slope of the data in the window to obtain global trend features. Through the multi-scale sliding window algorithm, the change of the density data can be analyzed from different angles, and more comprehensive feature information can be extracted.

[0050] Step S1232: Based on the local change features and global trend features, identify the step change intervals of the density values, and generate the position and amplitude information of the density mutation points.

[0051] In this embodiment, according to the extracted local change features and global trend features, it is judged whether the density value has a step change. When the standard deviation in the local change features suddenly increases and the slope in the global trend features changes significantly, it indicates that the density value may have a step change. Find these step change intervals, record their starting positions and ending positions as the position information of the density mutation points. Calculate the density difference before and after the step change as the amplitude information of the density mutation points. For example, in a density data sequence, a step change occurs between the 30th data point and the 35th data point. The starting position is the 30th data point, the ending position is the 35th data point. The density value before the step change is 1.2 g / cm³, and the density value after the step change is 1.5 g / cm³. Then the position information of the density mutation point is from the 30th data point to the 35th data point, and the amplitude information is 0.3 g / cm³.

[0052] Step S1233: Divide the intervals outside the positions of the density mutation points into steady-state density segments to generate the distribution of steady-state density intervals.

[0053] In this embodiment, after excluding the intervals at the positions of the density mutation points, the remaining density data is divided into steady-state density segments. Specifically, the fluctuation degree of the data can be calculated to judge whether it is a steady-state density segment. For example, calculate the absolute value of the difference between each data point and its adjacent data point. If these differences are all less than a preset threshold, it indicates that the density data in this interval is relatively stable, and it is divided into a steady-state density segment. Record the starting position and ending position of each steady-state density segment to generate the distribution of steady-state density intervals. For example, in a density data sequence, except that the 30th data point to the 35th data point is the position of the density mutation point, the density data from the 1st data point to the 29th data point has small fluctuations and is divided into a steady-state density segment. The density data from the 36th data point to the 100th data point is also relatively stable and is divided into another steady-state density segment. Then the distribution of steady-state density intervals includes two intervals: from the 1st data point to the 29th data point and from the 36th data point to the 100th data point.

[0054] Step S1234: Perform autocorrelation analysis on each steady-state density interval to extract the periodic component and random component of the density fluctuation.

[0055] In this embodiment, autocorrelation analysis is a method for analyzing the self-correlation of data. For each data in the steady-state density interval, its autocorrelation function is calculated. The autocorrelation function reflects the correlation of data at different time intervals. By analyzing the graph of the autocorrelation function, the periodic component and random component of the density fluctuation can be extracted. If the autocorrelation function shows periodic fluctuations, it means that there are periodic components in the density fluctuation; if the autocorrelation function decays rapidly in a short time interval, it means that there are more random components in the density fluctuation. For example, when performing autocorrelation analysis on the data in a steady-state density interval, it is found that the autocorrelation function has a peak every 5 time intervals, indicating that there is a periodic component with a period of 5 in the density fluctuation.

[0056] Step S1235: Generate medium dynamic characteristics according to the density mutation point position and amplitude information and the steady-state density interval distribution.

[0057] In this embodiment, the position and amplitude information of the density mutation point are integrated with the distribution of the steady-state density interval to generate the dynamic characteristics of the medium. The position information, amplitude information of the density mutation point, and the starting position and ending position of the steady-state density interval can be arranged in a certain order to form a feature vector. For example, the position information and amplitude information of the density mutation point are arranged first, and then the starting position and ending position of each steady-state density interval are arranged in sequence to form a feature vector containing multiple elements, which can fully reflect the dynamic changes of the medium density.

[0058] Step S124: Perform sliding window statistical calculation on the process flow data, extract flow fluctuation amplitude characteristics and periodic fluctuation regularity characteristics, and combine the fluctuation amplitude characteristics and periodic regularity characteristics into process stability characteristics.

[0059] In this embodiment, the process flow data reflects the flow change of the fluid in the coal preparation process. In order to evaluate the stability of the process, it is necessary to perform sliding window statistical calculation on the process flow data and extract relevant features.

[0060] For example, step S1241: dynamically adjust the sliding window length according to the flow change rate to generate an adaptive window length parameter.

[0061] In this embodiment, the flow rate of change refers to the change amount of flow data per unit time. First, calculate the rate of change of the flow data. The rate of change can be obtained by calculating the difference between adjacent data points and then dividing by the time interval. When the flow rate of change is large, it indicates that the flow changes rapidly, and a smaller sliding window is needed to capture detailed information; when the flow rate of change is small, it indicates that the flow changes slowly, and a larger sliding window can be used to obtain more macroscopic information. According to the magnitude of the flow rate of change, dynamically adjust the length of the sliding window. For example, when the flow rate of change is greater than a preset threshold, set the length of the sliding window to 5 data points; when the flow rate of change is less than the threshold, set the length of the sliding window to 10 data points. Take the adjusted length of the sliding window as the adaptive window length parameter.

[0062] Step S1242: Based on the adaptive window length parameter, divide the flow data into windows to generate multiple sets of window statistics.

[0063] In this embodiment, according to the adaptive window length parameter, divide the flow data into windows. Slide the sliding window sequentially on the flow data sequence, and each window contains the number of data points equal to the adaptive window length. For the data within each window, calculate its statistics, such as mean, standard deviation, maximum value, minimum value, etc. Record the statistics of each window to form a set of window statistics. Since the sliding window will sequentially cover the entire flow data sequence, multiple sets of window statistics will be generated. For example, if the adaptive window length is 5 data points and the flow data sequence contains 100 data points, then 96 sets of window statistics will be generated (100 - 5 + 1 = 96).

[0064] Step S1243: Calculate the skewness value and kurtosis value for each set of window statistics to generate the flow asymmetry feature.

[0065] In this embodiment, the skewness value and kurtosis value are statistics used to describe the distribution pattern of data. The skewness value reflects the symmetry of the data distribution, and the kurtosis value reflects the peakedness of the data distribution. For the data in each set of window statistics, calculate its skewness value and kurtosis value. The skewness value can be obtained by calculating the ratio of the third central moment to the cube of the standard deviation, and the kurtosis value can be obtained by calculating the ratio of the fourth central moment to the fourth power of the standard deviation and then subtracting 3. Record the skewness value and kurtosis value of each window to form the flow asymmetry feature, which can reflect the distribution pattern differences of the flow data in different windows.

[0066] Step S1244: Perform a difference calculation on the skewness value and kurtosis value of consecutive windows to generate the flow change acceleration feature.

[0067] In this embodiment, in order to further analyze the change of traffic flow, the difference calculation is performed on the skewness value and kurtosis value of consecutive windows. The difference calculation refers to calculating the difference between adjacent windows. For the skewness value, calculate the difference between the skewness values of adjacent windows; for the kurtosis value, calculate the difference between the kurtosis values of adjacent windows. These differences are used as the traffic flow change acceleration features. The traffic flow change acceleration features can reflect the change of the traffic flow change rate between adjacent windows. If the value of the traffic flow change acceleration feature is large, it indicates that the traffic flow change rate is changing rapidly and the traffic flow stability is poor; on the contrary, if the value of the traffic flow change acceleration feature is small, it indicates that the traffic flow change rate is relatively stable and the traffic flow stability is good. For example, for two adjacent windows, the skewness value of the previous window is 0.2 and the skewness value of the next window is 0.5, then the difference result of the skewness value is 0.5 - 0.2 = 0.3; the kurtosis value of the previous window is 1.5 and the kurtosis value of the next window is 1.8, then the difference result of the kurtosis value is 1.8 - 1.5 = 0.3. The above difference results are used as part of the traffic flow change acceleration features. By performing difference calculation on consecutive windows in the entire traffic flow data sequence, a series of traffic flow change acceleration feature values can be obtained, thereby forming a new feature sequence for more in-depth analysis of the traffic flow change characteristics.

[0068] Step S1245: Identify the type of fluctuation pattern according to the traffic flow asymmetry feature and the traffic flow change acceleration feature, and generate the process stability feature.

[0069] In this embodiment, the fluctuation mode type of the flow rate is identified by integrating the characteristics of flow rate asymmetry and the acceleration of flow rate change. Among them, several common fluctuation mode types can be predefined, such as stable fluctuation mode, periodic fluctuation mode, random fluctuation mode, etc. For the flow rate asymmetry characteristic, if the skewness value is close to 0, it indicates that the distribution of the flow rate data is relatively symmetric; if the skewness value is positive, it indicates that the data distribution is skewed to the right; if the skewness value is negative, it indicates that the data distribution is skewed to the left. Combining the magnitude of the kurtosis value, the degree of peakedness of the data distribution is judged. For the acceleration characteristic of flow rate change, if its value fluctuates within a small range, it indicates that the change rate of the flow rate is relatively stable; if its value shows periodic changes, there may be a periodic fluctuation mode; if its value changes randomly and fluctuates greatly, it may be a random fluctuation mode. For example, when the flow rate asymmetry characteristic shows that the skewness value is close to 0 and the kurtosis value is moderate, and at the same time the value of the flow rate change acceleration characteristic fluctuates within a small range, it is judged as a stable fluctuation mode. According to the identified fluctuation mode type, it is converted into the corresponding numerical code, which is concatenated with the flow rate asymmetry characteristic and the flow rate change acceleration characteristic to generate the process stability characteristic. For example, the stable fluctuation mode is encoded as 1, the periodic fluctuation mode is encoded as 2, and the random fluctuation mode is encoded as 3. The encoded values are arranged in order with the flow rate asymmetry characteristic and the flow rate change acceleration characteristic to form a feature vector containing multiple elements, and this feature vector is the process stability characteristic, which can comprehensively reflect the stability of the flow rate in the coal preparation process.

[0070] Step S130: Based on the pre-trained monitoring decision model, perform abnormal correlation analysis on the device state characteristics, medium dynamic characteristics, and process stability characteristics to generate the abnormal probability distribution and process adjustment parameters of each monitoring node.

[0071] In this embodiment, after obtaining the device state characteristics, medium dynamic characteristics, and process stability characteristics of each monitoring node, it is necessary to use the pre-trained monitoring decision model to perform abnormal correlation analysis on these characteristics to generate the abnormal probability distribution and process adjustment parameters of each monitoring node.

[0072] Step S131: Input the device state characteristics, medium dynamic characteristics, and process stability characteristics into the time series feature extraction unit of the monitoring decision model to extract the time series dependence characteristics of each monitoring node within a continuous time window.

[0073] In this embodiment, the time-series feature extraction unit of the monitoring decision model may adopt a recurrent neural network structure such as a long short-term memory network (LSTM). The device state features, medium dynamic features, and process stability features of each monitoring node are arranged in chronological order to form a multi-dimensional feature sequence. For example, the device state feature of each monitoring node is a feature vector containing 11 elements, the medium dynamic feature is a feature vector containing 8 elements, and the process stability feature is a feature vector containing 10 elements. Then, these three feature vectors are concatenated in order to form a feature vector containing 29 elements. These feature vectors of multiple consecutive time windows are input into the LSTM network. The LSTM network can capture the time-series dependence relationships in the data. Through the internal memory units and gating mechanisms, it processes the input feature sequence and extracts the time-series dependence features of each monitoring node within the consecutive time windows. For example, the LSTM network updates the hidden state of the current time window based on the input of the current time window and the hidden state of the previous time window, and this hidden state contains the time-series dependence information of the monitoring node within the consecutive time windows. After being processed by the LSTM network, a time-series dependence feature sequence of each monitoring node is output. Each element in this time-series dependence feature sequence is a feature vector containing multiple dimensions, reflecting the time-series features of the monitoring node at different time points.

[0074] Step S132: Input the time-series dependence features into the cross-modal attention unit of the monitoring decision model, perform dynamic weight allocation on the feature interaction relationships among the device state, medium dynamics, and process stability, and generate a cross-modal correlation feature vector.

[0075] In this embodiment, the role of the cross-modal attention unit is to interact and fuse the features of three different modalities: device status, medium dynamics, and process stability. The temporal dependence features output by the temporal feature extraction unit are input into the cross-modal attention unit. The cross-modal attention unit calculates the importance weights of each modality feature and dynamically weights the features of different modalities according to these weights. Specifically, it can be implemented through an attention mechanism. First, the temporal dependence features are projected into different feature spaces respectively to obtain the projection vectors of each modality feature. Then, the similarity between these projection vectors is calculated, for example, the similarity score is obtained through the dot product operation. The similarity score is normalized to obtain the attention weights of each modality feature. According to these attention weights, the features of different modalities are weighted and summed to obtain the cross-modal correlation feature vector. For example, the attention weight of the device status feature is 0.4, the attention weight of the medium dynamics feature is 0.3, and the attention weight of the process stability feature is 0.3. The corresponding temporal dependence features are weighted and summed according to these weights to obtain a new feature vector, which is the cross-modal correlation feature vector. It integrates the information of different modality features and highlights the important feature information at the current moment.

[0076] Step S133: Input the cross-modal correlation feature vector into the multi-layer perceptron network of the monitoring decision model, and output the abnormal probability distribution curves of different abnormal types for each monitoring node within multiple consecutive time windows.

[0077] In this embodiment, the multi-layer perceptron network (MLP) is a feed-forward neural network composed of multiple fully connected layers. Among them, the cross-modal association feature vector is input into the multi-layer perceptron network. The multi-layer perceptron network will perform non-linear transformation and feature extraction on the input feature vector. After receiving the cross-modal association feature vector at the input layer, through the processing of the hidden layer, the features are further abstracted and transformed. The neurons in the hidden layer will perform weighted summation on the input features and perform non-linear transformation through an activation function (such as the ReLU function) to enhance the expression ability of the model. Finally, at the output layer, the abnormal probability distribution curves of different abnormal types of each monitoring node within multiple consecutive time windows are output. Assuming that 5 abnormal types are predefined, then there will be 5 neurons in the output layer, corresponding to these 5 abnormal types respectively. The output value of each neuron represents the abnormal probability of the corresponding abnormal type of the monitoring node within the current time window. By analyzing the outputs of multiple consecutive time windows, the abnormal probability distribution curves of each monitoring node under different abnormal types can be obtained. For example, for a certain monitoring node, within 10 consecutive time windows, the abnormal probabilities of abnormal type 1 are 0.1, 0.15, 0.2, 0.25, 0.3, 0.35, 0.4, 0.45, 0.5, 0.55 in sequence. Connecting these probability values forms the abnormal probability distribution curve of this monitoring node under abnormal type 1.

[0078] Step S134: Input the cross-modal association feature vector into the regression analysis module of the monitoring and decision-making model, and output the medium density adjustment parameter, flow control parameter, and vibration warning parameter of the corresponding monitoring node.

[0079] In this embodiment, the role of the regression analysis module is to predict the medium density adjustment parameter, flow control parameter, and vibration warning parameter of the corresponding monitoring node according to the cross-modal association feature vector. The regression analysis module can adopt methods such as linear regression and polynomial regression. Taking the cross-modal association feature vector as the input, through the weight parameters obtained by training the regression model, perform linear combination and transformation on the input feature vector to obtain the predicted parameter values. For example, for the medium density adjustment parameter, the regression model will multiply each feature element in the cross-modal association feature vector by the corresponding weight coefficient, then add these products together, and then add a bias term to obtain the predicted medium density adjustment parameter value. The same method is used to calculate the flow control parameter and the vibration warning parameter. Assuming that the cross-modal association feature vector has 20 elements, for the regression model of the medium density adjustment parameter, there will be 20 weight coefficients and a bias term. Through training on a large amount of historical data, determine the values of these weight coefficients and bias terms so that the model can accurately predict the corresponding parameter values according to the input cross-modal association feature vector, and finally output the medium density adjustment parameter, flow control parameter, and vibration warning parameter of the corresponding monitoring node.

[0080] Step S135: According to the abnormal types exceeding the preset threshold in the probability distribution, match the corresponding medium density adjustment parameters, flow control parameters, and vibration warning parameters, and generate a set of process adjustment parameters associated with the abnormal types.

[0081] In this embodiment, an abnormal probability threshold is preset, for example, 0.3. In the abnormal probability distribution curve of each monitoring node, find the abnormal types with abnormal probabilities exceeding this threshold. For each abnormal type exceeding the threshold, match the corresponding parameter values from the medium density adjustment parameters, flow control parameters, and vibration warning parameters output by the regression analysis module. For example, if the abnormal probability of abnormal type 2 exceeds the threshold, then select the medium density adjustment parameter, flow control parameter, and vibration warning parameter corresponding to abnormal type 2 from the output parameters. Combine these matched parameter values together to form a set of process adjustment parameters associated with the abnormal type. Each abnormal type corresponds to such a set of parameters, and these sets of parameters contain specific adjustment parameters for different abnormal types.

[0082] Step S140: Determine a priority processing queue according to the abnormal probability distribution, and generate a coal preparation process optimization strategy based on the process adjustment parameters.

[0083] In this embodiment, after obtaining the abnormal probability distribution and process adjustment parameters of each monitoring node, it is necessary to determine a priority processing queue according to the abnormal probability distribution, and at the same time generate a coal preparation process optimization strategy based on the process adjustment parameters.

[0084] Step S141: Perform a time integral calculation on the abnormal probability distribution curve of each monitoring node to obtain an abnormal energy integral representing the cumulative risk value.

[0085] In this embodiment, the abnormal probability distribution curve represents the change of abnormal probabilities of different abnormal types of each monitoring node in multiple consecutive time windows. To evaluate the cumulative risk of each monitoring node, a time integral calculation is performed on the abnormal probability distribution curve. A numerical integration method such as the trapezoidal integration method can be used. Divide the abnormal probability distribution curve on the time axis into several small intervals. For each small interval, calculate the product of its corresponding abnormal probability value and the time interval, and then add these products together to obtain the abnormal energy integral. For example, for the abnormal probability distribution curve of a monitoring node in 10 consecutive time windows, multiply the abnormal probability value of each time window by the time interval (assuming the time interval is 1 minute), and then add these 10 products together to obtain the abnormal energy integral of this monitoring node. The larger the value of the abnormal energy integral, the higher the cumulative risk of this monitoring node during this period.

[0086] Step S142: Sort all the monitoring nodes in descending order according to the abnormal energy integral to generate a preliminary priority sequence.

[0087] In this embodiment, the abnormal energy integrals of all the monitoring nodes are compared, and the monitoring nodes are sorted in descending order. For example, there are 5 monitoring nodes, and their abnormal energy integrals are 10, 8, 6, 4, and 2 respectively. After sorting in descending order, the preliminary priority sequence is Monitoring Node 1, Monitoring Node 2, Monitoring Node 3, Monitoring Node 4, and Monitoring Node 5. This preliminary priority sequence reflects the relative importance of each monitoring node in terms of cumulative risk. The monitoring node with a higher cumulative risk is ranked higher and needs to be processed first.

[0088] Step S143: Identify the groups of monitoring nodes with upstream and downstream relationships in the technological process in the preliminary priority sequence, and perform risk superposition calculation on the node groups with equipment interlock dependencies to generate the combined risk value of the node groups.

[0089] In this embodiment, in the coal preparation process, different monitoring nodes may have upstream and downstream relationships in the technological process and equipment interlock dependencies. By analyzing the coal preparation technological process, the groups of monitoring nodes with these relationships in the preliminary priority sequence are identified. For example, Monitoring Node A and Monitoring Node B have an upstream and downstream relationship, and the output of Monitoring Node A is the input of Monitoring Node B; Monitoring Node C and Monitoring Node D have an equipment interlock dependency, and the operating state of one device will affect the operation of the other device. For the node groups with equipment interlock dependencies, the abnormal energy integrals of these nodes are superimposed and calculated to obtain the combined risk value of the node groups. For example, the abnormal energy integral of Monitoring Node C is 5, and the abnormal energy integral of Monitoring Node D is 3, then the combined risk value of the node group is 5 + 3 = 8. The combined risk value of the node group reflects the overall cumulative risk situation of this node group.

[0090] Step S144: Dynamically adjust the preliminary priority sequence based on the combined risk value of the node groups to generate a mixed priority queue containing independent nodes and node groups.

[0091] In this embodiment, the preliminary priority sequence is adjusted according to the combined risk value of the node group. The combined risk value of the node group is compared with the abnormal energy integral of the independent node, and the priorities are rearranged. For example, the preliminary priority sequence is Monitoring Node 1, Monitoring Node 2, Monitoring Node 3, Monitoring Node 4, Monitoring Node 5, where Monitoring Node 3 and Monitoring Node 4 form a node group, the combined risk value of the node group is 12, the abnormal energy integral of Monitoring Node 1 is 10, and the abnormal energy integral of Monitoring Node 2 is 8. Then the adjusted hybrid priority queue is the node group (Monitoring Node 3 and Monitoring Node 4), Monitoring Node 1, Monitoring Node 2, Monitoring Node 5. The hybrid priority queue comprehensively considers the cumulative risk situations of independent nodes and node groups, and more accurately reflects the objects that need to be processed preferentially.

[0092] Step S145: Add a time sensitivity label to each entry in the hybrid priority queue, and the time sensitivity label is automatically generated according to the slope change rate of the abnormal probability distribution curve.

[0093] In this embodiment, the time sensitivity label is used to indicate the sensitivity of the abnormal situation of each entry (independent node or node group) to time. The time sensitivity label is generated by analyzing the slope change rate of the abnormal probability distribution curve. If the slope change rate of the abnormal probability distribution curve is large, it means that the abnormal situation develops rapidly and is more sensitive to time; if the slope change rate is small, it means that the abnormal situation develops slowly and has a lower sensitivity to time. The slope change rate can be divided into several levels, such as high, medium, and low levels, which correspond to different time sensitivity labels respectively. For example, for the abnormal probability distribution curve of a certain monitoring node, calculate its slope change rate between adjacent time windows. If the slope change rate exceeds a preset high threshold, then add a "high time sensitivity" label to this monitoring node; if the slope change rate is between the high threshold and the low threshold, add a "medium time sensitivity" label; if the slope change rate is lower than the low threshold, add a "low time sensitivity" label. Add corresponding time sensitivity labels to each entry in the hybrid priority queue so as to take different measures according to the time sensitivity in subsequent processing.

[0094] Step S146: Analyze the medium density compensation value in the process adjustment parameters, calculate the deviation between the current density measurement value and the target density interval, and generate an opening control instruction sequence for the density regulating valve.

[0095] In this embodiment, the process adjustment parameters include the medium density compensation value. First, obtain the measured value of the current medium density and the target density range. The target density range is a reasonable density range preset according to the requirements of the coal preparation process, such as 1.3 g / cm³ to 1.5 g / cm³. Calculate the deviation between the current density measurement value and the target density range. If the current density measurement value is less than the lower limit of the target density range, it means that the medium density needs to be increased; if the current density measurement value is greater than the upper limit of the target density range, it means that the medium density needs to be decreased. Generate an opening control instruction sequence for the density regulating valve according to the magnitude of the deviation and the medium density compensation value. For example, the current density measurement value is 1.2 g / cm³, the target density range is 1.3 g / cm³ to 1.5 g / cm³, the deviation is 0.1 g / cm³, and the medium density compensation value is 0.05 g / cm³. According to the preset control rules, determine that the opening of the density regulating valve needs to be increased by a certain proportion, such as 20%, and generate the corresponding opening control instruction sequence to control the opening degree of the density regulating valve to adjust the medium density.

[0096] Step S147: Analyze the flow control threshold in the process adjustment parameters, and generate an increment adjustment instruction or a deceleration adjustment instruction for the pump speed according to the real-time flow fluctuation direction.

[0097] In this embodiment, the process adjustment parameters include the flow control threshold. Monitor the flow fluctuation situation in the coal preparation process in real time and determine the flow fluctuation direction. If the real-time flow is greater than the flow control threshold, it means that the flow is too large and the pump speed needs to be decreased; if the real-time flow is less than the flow control threshold, it means that the flow is too small and the pump speed needs to be increased. Generate an increment adjustment instruction or a deceleration adjustment instruction for the pump speed according to the flow fluctuation direction and the preset adjustment rules. For example, the flow control threshold is 100 cubic meters per hour, and the real-time flow is 120 cubic meters per hour, indicating that the flow is too large. According to the adjustment rules, determine that the pump speed needs to be decreased by 10%, and generate the corresponding deceleration adjustment instruction to control the running speed of the pump to adjust the flow.

[0098] Step S148: Analyze the vibration warning threshold in the process adjustment parameters, and generate a hierarchical response instruction when the vibration energy of the equipment exceeds the threshold, including a load balancing instruction in the primary warning state and an equipment switching instruction in the high-level alarm state.

[0099] In this embodiment, the process adjustment parameters include a vibration warning threshold. The vibration energy of the device is monitored in real time and compared with the vibration warning threshold. When the vibration energy of the device exceeds the threshold, it is determined whether it is in the primary warning state or the high-level alarm state according to the degree to which the vibration energy exceeds the threshold. If the vibration energy only slightly exceeds the threshold, it is in the primary warning state, and a load balancing instruction is generated. The load balancing instruction may include adjusting the load distribution of the device, such as reducing the input material volume of the device to reduce the vibration level of the device. If the vibration energy greatly exceeds the threshold, it is in the high-level alarm state, and a device switching instruction is generated. The device switching instruction is used to switch the currently operating device to a standby device to avoid device damage and affect the normal operation of the coal preparation process. For example, the vibration warning threshold is 50 vibration units, and the current device vibration energy is 60 vibration units, which is in the primary warning state, and a load balancing instruction is generated, requiring the input material volume of the device to be reduced by 20%; if the current device vibration energy is 80 vibration units, it is in the high-level alarm state, and a device switching instruction is generated to switch the current device to a standby device.

[0100] Step S149: Package the opening control instruction, the rotational speed adjustment instruction, and the classification response instruction according to the execution priority and the timing constraint relationship to generate a coal preparation process optimization policy package including a multiple verification mechanism.

[0101] In this embodiment, the generated opening control instruction, rotational speed adjustment instruction, and hierarchical response instruction are encapsulated into a strategy. They are sorted according to the execution priority and timing constraint relationship of these instructions. For example, the opening control instruction may need to be executed first to adjust the medium density, and then the rotational speed adjustment instruction and hierarchical response instruction are executed according to the adjusted situation. During the encapsulation process, a multiple verification mechanism is added to ensure the accuracy and reliability of the instructions. The multiple verification mechanism can include instruction format verification, instruction parameter range verification, and instruction execution order verification, etc. For instruction format verification, it is checked whether the opening control instruction, rotational speed adjustment instruction, and hierarchical response instruction conform to the preset format specifications. For example, the opening control instruction may be specified in the format of "opening adjustment: X%", where X is the specific opening adjustment percentage. The verification process checks whether the instruction follows this format. If it does not, the instruction is determined to be an invalid instruction. For instruction parameter range verification, the parameters in the instruction are checked to ensure they are within a reasonable range. For example, the opening adjustment percentage in the opening control instruction should be between 0% and 100%, and the rotational speed adjustment ratio in the rotational speed adjustment instruction should also be within the adjustable range allowed by the equipment. If the parameter exceeds the range, the instruction is corrected or determined to be an invalid instruction. Instruction execution order verification is to check whether the execution order of the instructions is correct according to the actual requirements of the coal preparation process and the operating logic of the equipment. For example, the medium density must be adjusted first (executing the opening control instruction), and then the flow rate is adjusted according to the new medium density situation (executing the rotational speed adjustment instruction). If the instruction order is incorrect, it is re-sorted and then encapsulated again.

[0102] After the verification of the instructions is completed, these instructions are sorted in an orderly manner according to the execution priority and timing constraint relationship. The instructions with higher priority and need to be executed first are arranged in the front, and the instructions with lower priority and can be executed later are arranged in the back. At the same time, information such as execution timestamps and execution conditions is added to each instruction to form a complete instruction sequence. Then, this instruction sequence is encapsulated into a coal preparation process optimization strategy package. The coal preparation process optimization strategy package can be stored in a set file format, such as XML format, for easy transmission and parsing. In addition to the instruction sequence, some additional metadata can be added to the coal preparation process optimization strategy package, such as the version number of the strategy package, generation time, applicable coal preparation process stage, etc., to facilitate subsequent management and use.

[0103] Step S1410: Embed an execution feedback monitoring logic in the coal preparation process optimization strategy package to capture the instruction response status returned by the edge execution terminal in real time and trigger a strategy dynamic correction mechanism.

[0104] In this embodiment, in order to ensure that the coal preparation process optimization strategy package can be effectively executed and adjusted according to the actual situation, an execution feedback monitoring logic is embedded in the strategy package. After the strategy package is sent to the edge execution terminal, the edge execution terminal will perform corresponding operations according to the instruction sequence in the strategy package. During the execution process, the edge execution terminal will collect the execution status of the instructions in real time and feedback the instruction response status to the strategy management system. The instruction response status can include information such as whether the instruction is successfully executed, whether an exception occurs during the execution process, and whether the execution result meets the expectation.

[0105] The strategy management system will capture these instruction response statuses in real time. For example, after the edge execution terminal executes the opening control instruction, it will feedback whether the valve opening has been adjusted to the specified percentage; after executing the speed adjustment instruction, it will feedback whether the pump speed has been adjusted as required. If the captured instruction response status indicates that the instruction execution fails or the execution result does not meet the expectation, the strategy management system will trigger the strategy dynamic correction mechanism.

[0106] The strategy dynamic correction mechanism will adjust the strategy package according to the instruction response status and the actual situation of the current coal preparation process. If the execution fails due to unreasonable instruction parameters, appropriate parameters will be recalculated to generate new instructions. For example, if the medium density does not reach the target range after the opening control instruction is executed, the opening adjustment percentage will be recalculated based on the current density measurement value and the target density range to generate a new opening control instruction. If the instruction execution is abnormal due to equipment failure, corresponding fault handling instructions will be generated, such as equipment maintenance instructions, replacement of standby equipment instructions, etc. At the same time, the instruction sequence and execution order in the strategy package will be updated to ensure that subsequent instructions can be correctly executed. The corrected strategy package will be sent to the edge execution terminal again to continue executing the optimization strategy to achieve the continuous optimization of the coal preparation process.

[0107] Step S150: Send the coal preparation process optimization strategy to the corresponding edge execution terminal to trigger the coal preparation process adjustment operation.

[0108] In this embodiment, after generating the coal preparation process optimization strategy package with a multiple verification mechanism and embedded execution feedback monitoring logic, it needs to be sent to the corresponding edge execution terminal. The edge execution terminal is usually an intelligent terminal device installed near each device in the coal preparation plant, which can receive the strategy package and execute the instructions therein to realize the adjustment operation of the coal preparation process.

[0109] First, according to the set of process adjustment parameters associated with the mixed priority queue and the exception type, determine the strategy package corresponding to each edge execution terminal. For example, for the edge execution terminal responsible for medium density regulation, the strategy package containing the opening control instruction will be sent to it; for the edge execution terminal responsible for flow control, the strategy package containing the speed adjustment instruction will be sent to it.

[0110] Then, the policy package is transmitted to the corresponding edge execution terminal through network communication. The network communication can use a wired network (such as Ethernet) or a wireless network (such as Wi-Fi, 4G / 5G, etc.), and the specific choice depends on the actual network environment and device configuration of the coal preparation plant. During the transmission process, encryption technology is used to encrypt the policy package to ensure the security and integrity of the data and prevent the policy package from being tampered with or leaked during transmission.

[0111] After receiving the policy package, the edge execution terminal decrypts and parses it. First, verify the digital signature of the policy package to ensure that the policy package comes from a legitimate policy management system. Then, execute the corresponding instructions in sequence according to the instruction sequence in the policy package. For example, first execute the opening control instruction to adjust the opening of the density regulating valve; then execute the speed adjustment instruction to adjust the speed of the pump; execute the hierarchical response instruction according to the equipment vibration condition to perform load balancing or equipment switching operations. During the execution process, the edge execution terminal will real-time feedback the execution status of the instruction for the policy management system to perform real-time monitoring and dynamic adjustment.

[0112] Step S210: Obtain the historical sensing data set and the corresponding historical anomaly label set of multiple monitoring nodes in the historical coal preparation process.

[0113] In this embodiment, in order to train the pre-trained monitoring decision model, it is necessary to collect the historical sensing data set and the corresponding historical anomaly label set of multiple monitoring nodes in the historical coal preparation process. The collection of the historical sensing data set can be carried out through the data storage system of the coal preparation plant. The coal preparation plant usually stores the sensing data of each monitoring node for a long time for subsequent analysis and processing. Extract the historical sensing data within a certain time period (such as the past year) from the data storage system, including equipment vibration data, medium density data, and process flow data. These data will be sorted according to the monitoring nodes and time sequence to form a multi-dimensional data set.

[0114] Meanwhile, it is also necessary to collect the corresponding historical anomaly label set. The historical anomaly label set is a mark of whether there are anomalies in the historical sensing data. These labels can be generated through manual annotation or automatic detection algorithms. Manual annotation is to analyze and judge the historical sensing data by professional technicians, marking the time periods and anomaly types with anomalies. The automatic detection algorithm is to use some preset rules or machine learning models to monitor the historical sensing data in real time, and automatically generate anomaly labels when anomalies are detected. For example, for device vibration data, when the vibration energy exceeds a certain preset threshold, it is marked as "device vibration anomaly"; for medium density data, when the density value exceeds the normal range, it is marked as "medium density anomaly". Associate these historical anomaly labels with the corresponding historical sensing data to form a historical anomaly label set.

[0115] Step S220: Perform the same feature extraction process as the real-time data processing on the historical sensing data set to generate historical device state features, historical medium dynamic features, and historical process stability features.

[0116] In this embodiment, in order to make the training data have the same feature representation as the real-time data, the same feature extraction process as the real-time data processing is performed on the historical sensing data set. The specific processing process is the same as the processing process of the real-time sensing data in steps S120 - S124.

[0117] For historical device vibration data, first perform time series data cleaning. Use the adaptive threshold algorithm to detect outliers, identify abnormal fluctuation intervals where the number of consecutive outliers exceeds the threshold, use the generative repair model to reconstruct the abnormal fluctuation segments, calculate the difference degree of the frequency domain energy distribution between the repaired data and the original data, and trigger the sensor calibration instruction when the difference degree exceeds the preset safety threshold. After calibration, re-collect data to replace the abnormal fluctuation segments to generate a standardized vibration data sequence. Then perform spectral energy distribution analysis on the standardized vibration data sequence, extract the distribution features of vibration energy in the preset frequency band and the main frequency offset features, and combine these features into historical device state features.

[0118] For historical medium density data, use the multi-scale sliding window algorithm to extract local change features and global trend features, identify density mutation points and steady-state density intervals based on these features, perform autocorrelation analysis on each steady-state density interval, extract the periodic and random components of density fluctuations, and generate historical medium dynamic features according to the position and amplitude information of the density mutation points and the distribution of the steady-state density intervals.

[0119] For historical process flow data, dynamically adjust the sliding window length according to the flow change rate, divide the flow data into windows, calculate the skewness value and kurtosis value of each window statistic set to generate flow asymmetry features, perform differential calculations on the skewness value and kurtosis value of consecutive windows to generate flow change acceleration features, and identify the fluctuation mode type based on the flow asymmetry features and flow change acceleration features to generate historical process stability features.

[0120] Step S230: Construct a feature fusion module including a temporal feature extraction unit and a cross-modal attention unit, and input the historical device state features, historical medium dynamic features, and historical process stability features into the temporal feature extraction unit for modeling the time dimension dependency relationship to generate a feature sequence with time continuity.

[0121] In this embodiment, in order to better fuse the historical device state features, historical medium dynamic features, and historical process stability features, a feature fusion module including a temporal feature extraction unit and a cross-modal attention unit is constructed.

[0122] The temporal feature extraction unit can adopt a recurrent neural network structure such as a long short-term memory network (LSTM) or a gated recurrent unit (GRU). Arrange the historical device state features, historical medium dynamic features, and historical process stability features in chronological order to form a multi-dimensional feature sequence. For example, the feature vector at each time point is composed of the concatenation of the historical device state features, historical medium dynamic features, and historical process stability features. Suppose the historical device state features have 11 dimensions, the historical medium dynamic features have 8 dimensions, and the historical process stability features have 10 dimensions, then the feature vector at each time point is a vector containing 29 dimensions.

[0123] Input this multi-dimensional feature sequence into the temporal feature extraction unit. Taking LSTM as an example, the LSTM network updates the hidden state at the current time point based on the input feature vector at the current time point and the hidden state at the previous time point. The hidden state contains the time dimension dependency information of the historical data. By continuously processing the input at each time point, the LSTM network can learn the time continuity and dependency relationship in the feature sequence, and generate a feature sequence with time continuity, where each element is a feature vector containing multiple dimensions, reflecting the feature information that combines the device state, medium dynamics, and process stability at different time points.

[0124] Step S240: Perform interactive weight assignment on the device state features, medium dynamic features, and process stability features in the feature sequence through the cross-modal attention unit to generate a cross-modal correlation feature vector.

[0125] In this embodiment, the role of the cross-modal attention unit is to perform interactive weight allocation on the device state features, medium dynamic features, and process stability features in the feature sequence to highlight the importance of different modal features at different time points.

[0126] First, the feature sequence with temporal continuity output by the temporal feature extraction unit is input into the cross-modal attention unit. The cross-modal attention unit decomposes the feature sequence and separately extracts the device state feature subsequence, medium dynamic feature subsequence, and process stability feature subsequence.

[0127] Then, calculate the importance weights of each modal feature at each time point. The specific method is to project the subsequence of each modal feature into different feature spaces through the attention mechanism to obtain projection vectors. Calculate the similarity between these projection vectors, for example, obtain the similarity score through dot product operation. Normalize the similarity score to obtain the attention weight of each modal feature at each time point.

[0128] According to these attention weights, dynamically weight the features of different modalities. For each time point, multiply the device state feature, medium dynamic feature, and process stability feature by their corresponding attention weights respectively, and then splice the weighted features to obtain the cross-modal correlation feature vector at this time point. For example, at a certain time point, the attention weight of the device state feature is 0.4, the attention weight of the medium dynamic feature is 0.3, and the attention weight of the process stability feature is 0.3. Weight and splice the corresponding feature vectors according to these weights to obtain the cross-modal correlation feature vector at this time point. By processing the entire feature sequence, a series of cross-modal correlation feature vectors are obtained. These vectors integrate the information of different modal features and highlight the important feature information at different time points.

[0129] Step S250: Invoke a multi-layer perceptron network to perform abnormal probability prediction processing on the cross-modal correlation feature vector, and output the predicted abnormal probability distribution of each monitoring node within multiple consecutive time windows.

[0130] In this embodiment, the multi-layer perceptron network (MLP) is a commonly used feedforward neural network, which consists of an input layer, a hidden layer, and an output layer. Input the cross-modal correlation feature vector output by the cross-modal attention unit into the multi-layer perceptron network.

[0131] The input layer receives the cross-modal correlation feature vector, and the hidden layer performs non-linear transformation and feature extraction on the input features. The neurons in the hidden layer perform weighted summation on the input feature vectors and perform non-linear transformation through an activation function (such as the ReLU function) to enhance the expression ability of the model. Multiple hidden layers can be set, and the number of neurons in each hidden layer can be adjusted according to the actual situation.

[0132] The number of neurons in the output layer is determined according to the number of anomaly types. Suppose 5 anomaly types are predefined, then there will be 5 neurons in the output layer, corresponding to these 5 anomaly types respectively. The output value of each neuron represents the predicted anomaly probability of the corresponding anomaly type for the monitoring node within the current time window. By processing the cross-modal correlation feature vectors of multiple consecutive time windows, the multi-layer perceptron network will output the predicted anomaly probability distributions of different anomaly types for each monitoring node within multiple consecutive time windows. For example, for a certain monitoring node, within 10 consecutive time windows, the predicted anomaly probabilities of anomaly type 1 are 0.1, 0.15, 0.2, 0.25, 0.3, 0.35, 0.4, 0.45, 0.5, 0.55 in sequence. Connecting these probability values forms the predicted anomaly probability distribution of this monitoring node under anomaly type 1.

[0133] Step S260: Synchronously call the regression analysis module to perform process parameter prediction processing on the cross-modal correlation feature vectors, and output the predicted process parameters corresponding to each monitoring node. The predicted process parameters include medium density adjustment parameters, flow control parameters, and vibration warning parameters.

[0134] In this embodiment, the regression analysis module is used to predict the process parameters corresponding to each monitoring node according to the cross-modal correlation feature vectors, including medium density adjustment parameters, flow control parameters, and vibration warning parameters.

[0135] The regression analysis module can adopt methods such as linear regression, polynomial regression, or support vector regression. Taking linear regression as an example, first, the parameters of the regression model, that is, the weight coefficients and the bias term, need to be determined. By training a large amount of historical data and using optimization algorithms such as the least squares method to solve these parameters, the regression model can accurately predict the corresponding process parameters according to the input cross-modal correlation feature vectors.

[0136] Input the cross-modal correlation feature vectors into the regression analysis module. The regression model will multiply the input feature vectors by the corresponding weight coefficients, then add up these products, and then add the bias term to obtain the predicted process parameter values. For example, for the regression model of the medium density adjustment parameter, assuming that the cross-modal correlation feature vector has 20 elements, then there will be 20 weight coefficients and one bias term. Multiply each element of the cross-modal correlation feature vector by the corresponding weight coefficient, then add up these products, and then add the bias term to obtain the predicted medium density adjustment parameter value. The same method is used to calculate the flow control parameters and the vibration warning parameters. Finally, the regression analysis module will output the predicted process parameters corresponding to each monitoring node.

[0137] Step S270: Construct a first loss function based on the difference between the predicted anomaly probability distribution and the historical anomaly label set, construct a second loss function based on the difference between the predicted process parameters and the historical actual adjustment parameters, and complete the iterative training of the monitoring decision model by jointly optimizing the first loss function and the second loss function.

[0138] In this embodiment, in order to train the monitoring decision model, it is necessary to construct a loss function to measure the difference between the model prediction result and the true label. A first loss function and a second loss function are respectively constructed, and then the iterative training of the model is completed by jointly optimizing these two loss functions.

[0139] For example, step S271: Sum the cross-entropy of the probability value of each monitoring node in the predicted anomaly probability distribution and the corresponding historical anomaly label in the historical anomaly label set to generate a first loss function.

[0140] In this embodiment, the predicted anomaly probability distribution is the predicted anomaly probability of each monitoring node output by the multi-layer perceptron network for different anomaly types in multiple consecutive time windows. The historical anomaly label set is the marking of whether there is an anomaly in the historical sensing data, represented in binary form (such as marking 1 for the existence of an anomaly and 0 for the non-existence of an anomaly).

[0141] For each monitoring node and each anomaly type, calculate the cross-entropy between the predicted anomaly probability value and the historical anomaly label. Cross-entropy is a metric that measures the difference between two probability distributions. For the predicted anomaly probability p of a monitoring node for a certain anomaly type in a certain time window and the corresponding historical anomaly label y (y is 0 or 1), the calculation formula for cross-entropy is: if y = 1, then the cross-entropy is -log(p); if y = 0, then the cross-entropy is -log(1 - p).

[0142] Sum the cross-entropy of each monitoring node for all anomaly types in all time windows to obtain the first loss function. For example, there are 10 monitoring nodes, each monitoring node has 5 anomaly types, and each monitoring node has predicted anomaly probabilities and historical anomaly labels for 10 time windows. Then, 10×5×10 cross-entropy values need to be calculated, and the sum of these values is used to obtain the value of the first loss function. The first loss function measures the accuracy of the model in predicting anomaly probabilities. The smaller the value of the loss function, the closer the model's prediction result is to the true label.

[0143] Step S272: Calculate the difference between the medium density adjustment parameter in the predicted process parameters and the medium density adjustment parameter in the historical actual adjustment parameters to generate a medium density adjustment difference value.

[0144] In this embodiment, the predicted process parameters are the predicted medium density adjustment parameters, predicted flow control parameters, and predicted vibration warning parameters corresponding to each monitoring node output by the regression analysis module. The historical actual adjustment parameters are the medium density adjustment parameters, flow control parameters, and vibration warning parameters actually carried out for abnormal conditions in the historical coal preparation process.

[0145] For the medium density adjustment parameters, calculate the difference between the predicted medium density adjustment parameters and the medium density adjustment parameters of the historical actual adjustment parameters. For example, if the predicted medium density adjustment parameter is 0.05 g / cm³ and the medium density adjustment parameter of the historical actual adjustment parameter is 0.06 g / cm³, then the medium density adjustment difference value is 0.05 - 0.06 = -0.01 g / cm³. By calculating the difference values of the medium density adjustment parameters for each monitoring node, a series of medium density adjustment difference values are obtained.

[0146] Step S273: Calculate the difference between the flow control parameter in the predicted process parameters and the flow control parameter of the historical actual adjustment parameter to generate a flow control difference value.

[0147] In this embodiment, similar to the processing of the medium density adjustment parameters, for the flow control parameters, calculate the difference between the predicted flow control parameters and the flow control parameters of the historical actual adjustment parameters. For example, if the predicted flow control parameter is an increase of 10 cubic meters per hour and the flow control parameter of the historical actual adjustment parameter is an increase of 12 cubic meters per hour, then the flow control difference value is 10 - 12 = -2 cubic meters per hour. Calculate the difference values of the flow control parameters for each monitoring node to obtain a series of flow control difference values.

[0148] Step S274: Calculate the difference between the vibration warning parameter in the predicted process parameters and the vibration warning parameter of the historical actual adjustment parameter to generate a vibration warning difference value.

[0149] In this embodiment, according to the same logic, calculate the difference between the vibration warning parameter in the predicted process parameters and the vibration warning parameter of the historical actual adjustment parameter. The vibration warning parameter may be related values such as the threshold setting of vibration energy. For example, if the vibration energy threshold corresponding to the predicted vibration warning parameter is 60 vibration units and the vibration energy threshold corresponding to the historical actual adjustment parameter is 65 vibration units, then the vibration warning difference value is 60 - 65 = -5 vibration units. Perform such calculations for each monitoring node to obtain a series of vibration warning difference values.

[0150] Step S275: Divide the medium density adjustment difference value by the historical reference parameter value of the historical medium density adjustment parameter to generate a standardized medium density adjustment difference value.

[0151] In this embodiment, in order to eliminate the influence of the dimension and numerical range between different parameters, it is necessary to standardize the medium density adjustment difference value. The historical reference parameter value of the historical medium density adjustment parameter is a representative reference value obtained based on a large amount of historical data statistics. For example, through statistical analysis of all medium density adjustment parameters in the past year, the historical reference parameter value is obtained as 0.1 g / cm³. Divide the medium density adjustment difference value of each monitoring node calculated previously by this historical reference parameter value to obtain the standardized medium density adjustment difference value. Suppose the medium density adjustment difference value of a certain monitoring node is -0.02 g / cm³, then the standardized medium density adjustment difference value is -0.02 ÷ 0.1 = -0.2. The standardized difference value can better reflect the relative deviation between the predicted value and the actual value, facilitating subsequent comprehensive comparison and calculation.

[0152] Step S276: Divide the flow control difference value by the historical reference parameter value of the historical flow control parameter to generate a standardized flow control difference value.

[0153] In this embodiment, the flow control difference value is also standardized. The historical reference parameter value of the historical flow control parameter is a reference value determined based on historical flow control data. For example, through statistics of historical data, the historical reference parameter value of the historical flow control parameter is obtained as 20 cubic meters per hour. Divide the flow control difference value of each monitoring node by this historical reference parameter value to obtain the standardized flow control difference value. If the flow control difference value of a certain monitoring node is -3 cubic meters per hour, then the standardized flow control difference value is -3 ÷ 20 = -0.15. The purpose of this is to unify the flow control difference value to a relatively comparable scale, avoiding the influence of the large difference in the numerical range of the flow data of different monitoring nodes on the final evaluation result.

[0154] Step S277: Divide the vibration warning difference value by the historical reference parameter value of the historical vibration warning parameter to generate a standardized vibration warning difference value.

[0155] In this embodiment, the vibration warning difference value is standardized. The historical reference parameter value of the historical vibration warning parameter is a representative value statistically obtained from the historical vibration warning data. For example, through the analysis of historical data, the historical reference parameter value of the historical vibration warning parameter is determined to be 10 vibration units. Divide the vibration warning difference value of each monitoring node by the historical reference parameter value to obtain the standardized vibration warning difference value. Suppose the vibration warning difference value of a certain monitoring node is -2 vibration units, then the standardized vibration warning difference value is -2÷10 = -0.2. The standardized vibration warning difference value can more accurately reflect the deviation degree of the predicted vibration warning parameter from the historical actual situation, which helps to more reasonably consider the factors of vibration warning in the comprehensive evaluation.

[0156] Step S278: Perform weighted summation on the standardized medium density adjustment difference value, the standardized flow control difference value, and the standardized vibration warning difference value according to a preset weight coefficient to generate a second loss function.

[0157] In this embodiment, the preset weight coefficient is preset according to the importance of different process parameters in the coal preparation process. For example, according to experience and actual production conditions, the weight coefficient of the medium density adjustment parameter is set to 0.4, the weight coefficient of the flow control parameter is set to 0.3, and the weight coefficient of the vibration warning parameter is set to 0.3. For each monitoring node, multiply its standardized medium density adjustment difference value by the corresponding weight coefficient, multiply the standardized flow control difference value by the corresponding weight coefficient, multiply the standardized vibration warning difference value by the corresponding weight coefficient, and then add these three products to obtain the weighted sum of this monitoring node. For example, if the standardized medium density adjustment difference value of a certain monitoring node is -0.2, the standardized flow control difference value is -0.15, and the standardized vibration warning difference value is -0.2, then the weighted sum of this monitoring node is (-0.2)×0.4 + (-0.15)×0.3 + (-0.2)×0.3 = -0.185. Add the weighted sums of all monitoring nodes to obtain the value of the second loss function. The second loss function measures the accuracy of the model in predicting process parameters. The smaller the loss function value, the closer the prediction of the model for the process parameters is to the historical actual adjustment parameters.

[0158] Step S279: Add the first loss function and the second loss function according to a preset proportional coefficient to generate a combined loss function.

[0159] In this embodiment, the preset proportionality coefficient is preset according to the degree of emphasis on the abnormal probability prediction and the process parameter prediction. For example, the proportionality coefficient of the abnormal probability prediction is set to 0.6, and the proportionality coefficient of the process parameter prediction is set to 0.4. Multiply the value of the first loss function by the proportionality coefficient of the abnormal probability prediction, multiply the value of the second loss function by the proportionality coefficient of the process parameter prediction, and then add these two products to obtain the value of the combined loss function. Suppose the value of the first loss function is 0.3 and the value of the second loss function is 0.2, then the value of the combined loss function is 0.3×0.6 + 0.2×0.4 = 0.26. The combined loss function comprehensively considers the performance of the model in two aspects: abnormal probability prediction and process parameter prediction. By minimizing the combined loss function, the model can achieve better prediction effects in these two aspects.

[0160] Step S2710: Calculate the gradient of the combined loss function through the backpropagation algorithm, and update the weight parameters of the cross-modal attention unit and the multi-layer perceptron network in the monitoring decision model to complete the iterative training.

[0161] In this embodiment, the backpropagation algorithm is a commonly used neural network training algorithm, which is used to calculate the gradient of the loss function with respect to the model weight parameters and update the weight parameters according to the gradient. First, according to the value of the combined loss function, use the chain rule to calculate the gradient of the combined loss function with respect to each weight parameter of the cross-modal attention unit and the multi-layer perceptron network in the monitoring decision model. The gradient represents the rate of change of the loss function in the direction of this weight parameter. Through the gradient, it can be known how to adjust the weight parameter to reduce the value of the loss function.

[0162] For example, for a certain weight parameter in the cross-modal attention unit, the gradient of the combined loss function with respect to this weight parameter is calculated to be 0.01 through the backpropagation algorithm. Then, according to the principle of the gradient descent method, a preset learning rate (such as 0.001) is used to update this weight parameter. Subtract the value of the learning rate multiplied by the gradient from this weight parameter, that is, the new weight parameter = the original weight parameter - the learning rate × the gradient. Suppose the original weight parameter is 0.5, then the updated weight parameter is 0.5 - 0.001×0.01 = 0.49999.

[0163] Such update operations are performed on all weight parameters in the cross-modal attention unit and the multi-layer perceptron network. After completing one update of the weight parameters, re-enter the training data, calculate the value of the combined loss function again, and repeat the above process of gradient calculation and weight parameter update until the value of the combined loss function converges to a smaller value or reaches the preset number of iterations. At this time, the iterative training of the monitoring decision model is completed.

[0164] Step S310: Real-time collect the process adjustment execution result data and the adjusted sensing data set fed back by the edge execution terminal.

[0165] In this embodiment, after the coal preparation process optimization strategy is sent to the edge execution terminal and executed, it is necessary to real-time collect the process adjustment execution result data and the adjusted sensing data set fed back by the edge execution terminal. After the edge execution terminal executes the instructions in the policy package, it will record the execution situation of the instructions, such as whether the valve opening has been adjusted according to the instructions, whether the pump speed has reached the specified value, etc. These information constitute the process adjustment execution result data. At the same time, the edge execution terminal will continue to collect the sensing data of each monitoring node, including equipment vibration data, medium density data, and process flow data. These data are the sensing data set after the process adjustment operation.

[0166] For example, the edge execution terminal will feedback the actual opening percentage of the density regulating valve, compare it with the opening required by the opening control instruction in the policy package, and record the execution result of the opening adjustment. For the pump, it will feedback the actual speed value and compare it with the speed required by the speed adjustment instruction. In terms of collecting sensing data, the data of the equipment vibration sensor, medium density sensor, and process flow sensor will be collected at a certain time interval (such as every second, every minute, etc.) to form the adjusted sensing data set. These feedback data and sensing data will be transmitted back to the policy management system in real time through network communication for subsequent analysis and processing.

[0167] Step S320: Perform feature extraction processing on the adjusted sensing data set to obtain updated equipment state features, medium dynamic features, and process stability features.

[0168] In this embodiment, the feature extraction processing is performed on the collected adjusted sensing data set, and the processing process is the same as the processing process of the real-time sensing data in steps S120 - S124.

[0169] For the adjusted equipment vibration data, first use the adaptive threshold algorithm to detect outliers, identify the abnormal fluctuation interval where the number of consecutive outliers exceeds the threshold, use the generative repair model to reconstruct the abnormal fluctuation segment, calculate the difference degree of the energy distribution in the frequency domain between the repaired data and the original data, and trigger the sensor calibration instruction when the difference degree exceeds the preset safety threshold. After calibration, re-collect the data to replace the abnormal fluctuation segment to generate a standardized vibration data sequence. Then perform spectral energy distribution analysis on the standardized vibration data sequence, extract the distribution characteristics of the vibration energy in the preset frequency band and the main frequency offset characteristics, and merge these characteristics into the updated equipment state characteristics.

[0170] For the adjusted medium density data, the multi-scale sliding window algorithm is used to extract local change features and global trend features. Based on these features, density mutation points and steady-state density intervals are identified. Autocorrelation analysis is performed on each steady-state density interval to extract the periodic and random components of density fluctuations. According to the density mutation point positions, amplitude information, and the distribution of steady-state density intervals, updated medium dynamic features are generated.

[0171] For the adjusted process flow rate data, the sliding window length is dynamically adjusted according to the flow rate change rate. The flow rate data is windowed, and the skewness and kurtosis values of the statistical quantity set of each window are calculated to generate flow asymmetry features. The difference between the skewness and kurtosis values of consecutive windows is calculated to generate flow change acceleration features. According to the flow asymmetry features and flow change acceleration features, the fluctuation pattern type is identified, and updated process stability features are generated.

[0172] Step S330: Input the updated device state features, medium dynamic features, and process stability features into the current monitoring decision model to generate new abnormal probability prediction results and process parameter prediction results.

[0173] In this embodiment, the updated device state features, medium dynamic features, and process stability features are input into the currently trained monitoring decision model. First, these features enter the time series feature extraction unit of the monitoring decision model, such as an LSTM network, to extract the time series dependence features of each monitoring node within a continuous time window. Then, the time series dependence features are input into the cross-modal attention unit to dynamically allocate weights to the feature interaction relationships of device state, medium dynamics, and process stability, generating a cross-modal correlation feature vector.

[0174] Next, the cross-modal correlation feature vector is input into a multi-layer perceptron network and a regression analysis module respectively. The multi-layer perceptron network outputs new abnormal probability prediction results of different abnormal types for each monitoring node within multiple continuous time windows, similar to the processing process in step S250. The regression analysis module outputs new predicted process parameters for the corresponding monitoring nodes, including medium density adjustment parameters, flow control parameters, and vibration warning parameters, similar to the processing process in step S260. Through these steps, new abnormal probability prediction results and process parameter prediction results are generated using the updated feature data to evaluate the state after the coal preparation process adjustment and further optimization strategies.

[0175] Step S340: Construct a model drift detection index based on the deviation degrees of the new abnormal probability prediction results and process parameter prediction results from historical actual data.

[0176] In this embodiment, in order to detect whether the monitoring decision-making model drifts, it is necessary to construct a model drift detection index. First, determine the historical actual data, including the historical abnormal label set and the historical actual adjustment parameters. For the new abnormal probability prediction results, compare the predicted abnormal probability of each monitoring node under each abnormal type with the corresponding value in the historical abnormal label set. The difference between the predicted abnormal probability and the historical abnormal label can be calculated, for example, using methods such as mean square error. For each monitoring node, calculate the sum of the difference values under all abnormal types to obtain the deviation degree of the monitoring node in terms of abnormal probability prediction.

[0177] For the new process parameter prediction results, compare the predicted medium density adjustment parameter, flow control parameter, and vibration warning parameter with the historical actual adjustment parameters respectively. Similarly, methods such as mean square error can be used to calculate the difference between the predicted value and the actual value of each parameter. Weighted sum the difference values of the medium density adjustment parameter, flow control parameter, and vibration warning parameter according to a certain weight to obtain the deviation degree of the monitoring node in terms of process parameter prediction.

[0178] Weighted sum the deviation degree in terms of abnormal probability prediction and the deviation degree in terms of process parameter prediction according to a preset ratio to obtain the comprehensive deviation degree of each monitoring node. Average the comprehensive deviation degrees of all monitoring nodes to obtain the model drift detection index. For example, set the weight for abnormal probability prediction to be 0.6 and the weight for process parameter prediction to be 0.4. The deviation degree of a certain monitoring node in terms of abnormal probability prediction is 0.1, and the deviation degree in terms of process parameter prediction is 0.05. Then the comprehensive deviation degree of this monitoring node is 0.1×0.6 + 0.05×0.4 = 0.08. After averaging the comprehensive deviation degrees of all monitoring nodes, obtain the value of the model drift detection index.

[0179] Step S350: When the model drift detection index exceeds the preset threshold, trigger the model extended training process, and mix and sample the updated device state features, medium dynamic features, and process stability features with the historical training data to generate an enhanced training data set.

[0180] In this embodiment, preset a model drift detection threshold, such as 0.1. When the calculated model drift detection index exceeds this model drift detection threshold, it indicates that the prediction performance of the monitoring decision-making model has significantly decreased, and it may be that the operating state of the coal preparation process has changed, and it is necessary to trigger the model extended training process.

[0181] In the model extension training process, the updated device status features, medium dynamic features, and process stability features are mixed and sampled with historical training data. The purpose of mixed sampling is to add new feature data while retaining the information of historical data, so that the training data set can better reflect the actual situation of the current coal preparation process. For example, a random sampling method can be used to extract a certain proportion (such as 30%) of the data from the updated feature data and merge it with the historical training data. At the same time, it is necessary to ensure that the label information (historical anomaly label set and historical actual adjustment parameters) of the data is correctly corresponded and merged during the mixed sampling process to form a new enhanced training data set. This enhanced training data set will be used to further train the monitoring decision model to improve the adaptability and prediction accuracy of the model.

[0182] Step S360: Fine-tune the parameters of the cross-modal attention unit and the multi-layer perceptron network in the monitoring decision model through the enhanced training data set, and keep the parameters of the time series feature extraction unit unchanged to maintain the stability of time modeling.

[0183] In this embodiment, the enhanced training data set is used to fine-tune the parameters of the monitoring decision model. Considering that the time series feature extraction unit (such as the LSTM network) has learned the time dimension dependence of the data, in order to maintain the stability of time modeling, the parameters of the time series feature extraction unit are kept unchanged.

[0184] For the cross-modal attention unit and the multi-layer perceptron network, a training method similar to steps S270 - S2710 is used for parameter fine-tuning. First, input the enhanced training data set into the monitoring decision model, calculate the value of the joint loss function, and the calculation method of the joint loss function is the same as before, comprehensively considering the accuracy of anomaly probability prediction and process parameter prediction. Then, calculate the gradient of the joint loss function with respect to the weight parameters of the cross-modal attention unit and the multi-layer perceptron network through the backpropagation algorithm.

[0185] According to the gradient and the preset learning rate, update the weight parameters of the cross-modal attention unit and the multi-layer perceptron network. For example, for a certain weight parameter in the cross-modal attention unit, the calculated gradient is 0.005, the learning rate is 0.0001, and the original weight parameter is 0.3. Then the updated weight parameter is 0.3 - 0.0001×0.005 = 0.299995. Repeat this process, continuously adjust the weight parameters until the value of the joint loss function converges to a smaller value or reaches the preset number of iterations, and complete the parameter fine-tuning of the cross-modal attention unit and the multi-layer perceptron network.

[0186] Step S370: Encrypt and transmit the fine-tuned model parameters to all edge execution terminals to complete the model hot update.

[0187] In this embodiment, after the parameter fine-tuning of the monitoring decision-making model is completed, it is necessary to transmit the fine-tuned model parameters to all edge execution terminals to achieve the hot update of the model. First, the fine-tuned model parameters are encrypted to ensure the security and integrity of the data. Symmetric encryption algorithms (such as the AES algorithm) or asymmetric encryption algorithms (such as the RSA algorithm) can be used to encrypt the model parameters.

[0188] Then, the encrypted model parameters are transmitted to all edge execution terminals through network communication. Wired networks (such as Ethernet) or wireless networks (such as Wi-Fi, 4G / 5G, etc.) can be used for network communication, and the specific choice depends on the actual network environment and device configuration of the coal preparation plant. During the transmission process, reliable data transmission should be ensured, and methods such as retransmission mechanisms are used to handle possible network failures.

[0189] After the edge execution terminal receives the encrypted model parameters, it uses the corresponding decryption key for decryption. After decryption, the new model parameters are used to replace the original model parameters to complete the hot update of the model. During the update process, the normal operation of the edge execution terminal should be ensured to avoid interruptions or abnormalities in the coal preparation process due to model updates. After the update is completed, the edge execution terminal can use the new model parameters for anomaly analysis and process adjustment to improve the monitoring and decision-making capabilities of the coal preparation process.

[0190] Figure 2 FIG. shows a schematic diagram of exemplary hardware and software components of a coal preparation whole-process monitoring and decision-making system 100 based on Internet of Things sensing that can implement the idea of the present application. For example, a processor 120 can be used on the coal preparation whole-process monitoring and decision-making system 100 based on Internet of Things sensing and is used to execute the functions in the present application.

[0191] The coal preparation whole-process monitoring and decision-making system 100 based on Internet of Things sensing can be a general-purpose server or a special-purpose server, both of which can be used to implement the coal preparation whole-process monitoring and decision-making method based on Internet of Things sensing in the present application. Although only one server is shown in the present application, for convenience, the functions described in the present application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0192] For example, the whole-process monitoring and decision-making system 100 for coal preparation based on Internet of Things sensing may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as disks, ROM, or RAM, or any combination thereof. Exemplarily, the whole-process monitoring and decision-making system 100 for coal preparation based on Internet of Things sensing may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present application can be implemented according to these program instructions. The whole-process monitoring and decision-making system 100 for coal preparation based on Internet of Things sensing further includes an input / output (I / O) interface 150 between the computer and other input / output devices.

[0193] For ease of explanation, only one processor is described in the whole-process monitoring and decision-making system 100 for coal preparation based on Internet of Things sensing. However, it should be noted that the whole-process monitoring and decision-making system 100 for coal preparation in the present application may also include multiple processors. Therefore, the steps performed by one processor described in the present application may also be jointly performed or separately performed by multiple processors. For example, if the processor of the whole-process monitoring and decision-making system 100 for coal preparation based on Internet of Things sensing performs step A and step B, it should be understood that step A and step B may also be jointly performed by two different processors or separately performed in one processor. For example, the first processor performs step A, the second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0194] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the above-mentioned whole-process monitoring and decision-making method for coal preparation based on Internet of Things sensing is implemented.

[0195] It should be noted that, in order to simplify the description of the present invention disclosure and thus help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, sometimes multiple features are merged into one embodiment, drawing, or description thereof.

Claims

1. A full-process monitoring and decision-making method for coal preparation based on Internet of Things sensing, characterized in that, The method includes: Obtaining a real-time sensing data set of multiple monitoring nodes in the coal preparation process, where the real-time sensing data set includes equipment vibration data, medium density data, and process flow data; Performing feature extraction processing on the real-time sensing data set to obtain equipment state features, medium dynamic features, and process stability features of each monitoring node; Based on a pre-trained monitoring decision model, performing abnormal correlation analysis processing on the equipment state features, medium dynamic features, and process stability features to generate an abnormal probability distribution and process adjustment parameters for each monitoring node; Determining a priority processing queue according to the abnormal probability distribution, and generating a coal preparation process optimization strategy based on the process adjustment parameters; Sending the coal preparation process optimization strategy to the corresponding edge execution terminal to trigger a coal preparation process adjustment operation; The performing abnormal correlation analysis processing on the equipment state features, medium dynamic features, and process stability features based on the pre-trained monitoring decision model to generate an abnormal probability distribution and process adjustment parameters for each monitoring node includes: Inputting the equipment state features, medium dynamic features, and process stability features into the time series feature extraction unit of the monitoring decision model to extract the time series dependence features of each monitoring node within a continuous time window; Inputting the time series dependence features into the cross-modal attention unit of the monitoring decision model to perform dynamic weight allocation on the feature interaction relationships of equipment state, medium dynamic, and process stability, and generating a cross-modal correlation feature vector; Inputting the cross-modal correlation feature vector into the multi-layer perceptron network of the monitoring decision model to output the abnormal probability distribution curves of different abnormal types of each monitoring node within multiple continuous time windows; Inputting the cross-modal correlation feature vector into the regression analysis module of the monitoring decision model to output the medium density adjustment parameters, flow control parameters, and vibration warning parameters of the corresponding monitoring node; According to the abnormal types exceeding the preset threshold in the probability distribution, matching the corresponding medium density adjustment parameters, flow control parameters, and vibration warning parameters to generate a set of process adjustment parameters associated with the abnormal types.

2. The coal preparation full-process monitoring and decision-making method based on Internet of Things sensing according to claim 1, characterized in that, The performing feature extraction processing on the real-time sensing data set to obtain equipment state features, medium dynamic features, and process stability features of each monitoring node includes: Performing time series data cleaning processing on the equipment vibration data, removing data segments containing abnormal fluctuation intervals, and generating a standardized vibration data sequence; Performing spectrum energy distribution analysis on the standardized vibration data sequence, extracting the distribution features of vibration energy within a preset frequency band and the main frequency offset features, and combining the distribution features and the main frequency offset features into equipment state features; Performing density gradient change analysis on the medium density data, identifying density mutation points and steady-state density intervals, extracting density change trend features and steady-state maintenance features, and combining the density change trend features and the steady-state maintenance features into medium dynamic features; Perform sliding window statistical calculations on the process flow rate data, extract the characteristics of the flow rate fluctuation amplitude and the characteristics of the periodic fluctuation law, and combine the flow rate fluctuation amplitude characteristics and the periodic fluctuation law characteristics into process stability characteristics.

3. The coal preparation full-process monitoring and decision-making method based on Internet of Things sensing according to claim 2, characterized in that, The training method of the pre-trained monitoring decision model includes: Obtain the historical sensing data set of multiple monitoring nodes in the historical coal preparation process and the corresponding historical anomaly label set; Perform the same feature extraction processing on the historical sensing data set as the real-time data processing to generate historical equipment state characteristics, historical medium dynamic characteristics, and historical process stability characteristics; Construct a feature fusion module including a time series feature extraction unit and a cross-modal attention unit, input the historical equipment state characteristics, historical medium dynamic characteristics, and historical process stability characteristics into the time series feature extraction unit for modeling the time dimension dependence relationship, and generate a feature sequence with time continuity; Perform interactive weight assignment on the historical equipment state characteristics, historical medium dynamic characteristics, and historical process stability characteristics in the feature sequence through the cross-modal attention unit to generate a cross-modal correlation feature vector; Call a multi-layer perceptron network to perform anomaly probability prediction processing on the cross-modal correlation feature vector, and output the predicted anomaly probability distribution of each monitoring node in multiple consecutive time windows; Synchronously call a regression analysis module to perform process parameter prediction processing on the cross-modal correlation feature vector, and output the predicted process parameters corresponding to each monitoring node. The predicted process parameters include medium density adjustment parameters, flow control parameters, and vibration warning parameters; Construct a first loss function based on the difference between the predicted anomaly probability distribution and the historical anomaly label set, construct a second loss function based on the difference between the predicted process parameters and the historical actual adjustment parameters, and complete the iterative training of the monitoring decision model by jointly optimizing the first loss function and the second loss function.

4. The coal preparation full-process monitoring and decision-making method based on Internet of Things sensing according to claim 1, characterized in that The determination of the priority processing queue according to the anomaly probability distribution includes: Perform time integration calculation on the anomaly probability distribution curve of each monitoring node to obtain the anomaly energy integral representing the cumulative risk value; Sort all monitoring nodes in descending order according to the anomaly energy integral to generate a preliminary priority sequence; Identify the monitoring node groups with upstream and downstream relationships in the process flow in the preliminary priority sequence, and perform risk superposition calculation on the node groups with equipment linkage dependencies to generate the combined risk value of the node groups; Dynamically adjust the preliminary priority sequence based on the combined risk value of the node groups to generate a mixed priority queue including independent nodes and node groups; Add a time sensitivity label to each entry in the mixed priority queue, and the time sensitivity label is automatically generated according to the slope change rate of the anomaly probability distribution curve.

5. The coal preparation full-process monitoring and decision-making method based on Internet of Things sensing according to claim 1, characterized in that The generation of the coal preparation process optimization strategy based on the process adjustment parameters includes: Analyze the medium density compensation value in the process adjustment parameters, calculate the deviation between the current density measurement value and the target density interval, and generate an opening control instruction sequence for the density regulating valve; Analyze the flow control threshold in the process adjustment parameters, and generate an incremental adjustment instruction or a deceleration adjustment instruction for the pump speed according to the real-time flow fluctuation direction; Analyze the vibration warning threshold in the process adjustment parameters, and generate a hierarchical response instruction when the vibration energy of the equipment exceeds the threshold, including a load balancing instruction in the primary warning state and an equipment switching instruction in the advanced alarm state; Encapsulate the opening control instruction, the speed adjustment instruction, and the hierarchical response instruction according to the execution priority and the timing constraint relationship to generate a coal preparation process optimization strategy package containing a multiple verification mechanism; Embed an execution feedback monitoring logic in the coal preparation process optimization strategy package to capture the instruction response status returned by the edge execution terminal in real time and trigger a strategy dynamic correction mechanism.

6. The coal preparation full-process monitoring and decision-making method based on Internet of Things sensing according to claim 3, wherein The method further includes: Collect the process adjustment execution result data and the adjusted sensing data set feedback by the edge execution terminal in real time; Extract features from the adjusted sensing data set to obtain updated equipment state features, medium dynamic features, and process stability features; Input the updated equipment state features, medium dynamic features, and process stability features into the current monitoring decision model to generate new abnormal probability prediction results and process parameter prediction results; Construct a model drift detection index based on the deviation degree between the new abnormal probability prediction results and process parameter prediction results and the historical actual data; When the model drift detection index exceeds the preset threshold, trigger a model extension training process, and perform hybrid sampling on the updated equipment state features, medium dynamic features, and process stability features and the historical training data to generate an enhanced training data set; Fine-tune the parameters of the cross-modal attention unit and the multi-layer perceptron network in the monitoring decision model through the enhanced training data set, and keep the parameters of the time series feature extraction unit unchanged to maintain the time modeling stability; Encrypt and transmit the fine-tuned model parameters to all edge execution terminals to complete the model hot update.

7. The coal preparation full-process monitoring and decision-making method based on Internet of Things sensing according to claim 2, wherein The time series data cleaning process for the equipment vibration data, removing the data segments containing abnormal fluctuation intervals, and generating a standardized vibration data sequence, includes: Adopt an adaptive threshold algorithm to detect the outliers in the vibration data and generate a vibration data sequence containing outlier marks; Identify the intervals where the number of consecutive outliers exceeds the threshold according to the outlier marks to generate the position information of abnormal fluctuation segments; Adopt a generative repair model to reconstruct the vibration data corresponding to the position information of the abnormal fluctuation segments to generate a repaired vibration data sequence; Calculate the difference degree of the frequency domain energy distribution between the repaired vibration data sequence and the original vibration data to generate a frequency domain difference degree index; When the frequency domain difference degree index exceeds the preset safety threshold, trigger the sensor calibration instruction of the corresponding monitoring node; After performing the calibration operation on the target sensor according to the sensor calibration instruction, re-collect the vibration data and replace the abnormal fluctuation segments to generate a standardized vibration data sequence.

8. The coal preparation full-process monitoring and decision-making method based on Internet of Things sensing according to claim 2, characterized in that Performing density gradient change analysis on the medium density data, identifying density mutation points and steady-state density intervals, extracting density change trend features and steady-state maintenance features, and combining the density change trend features and steady-state maintenance features into medium dynamic features, including: Using a multi-scale sliding window algorithm to extract local change features and global trend features of density data; Identifying the step change interval of density values based on the local change features and global trend features, and generating density mutation point position and amplitude information; Dividing the steady-state density segments in the interval outside the density mutation point position to generate the steady-state density interval distribution; Performing autocorrelation analysis on each steady-state density interval to extract the periodic component and random component of density fluctuation; Generating medium dynamic features according to the density mutation point position and amplitude information and the steady-state density interval distribution.

9. A coal preparation whole-process monitoring and decision-making system based on Internet of Things sensing, characterized in that, Including a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the coal preparation full-process monitoring and decision-making method based on Internet of Things sensing described in any one of the above claims 1-8.

Images